Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * multixact.c
4 : * PostgreSQL multi-transaction-log manager
5 : *
6 : * The pg_multixact manager is a pg_xact-like manager that stores an array of
7 : * MultiXactMember for each MultiXactId. It is a fundamental part of the
8 : * shared-row-lock implementation. Each MultiXactMember is comprised of a
9 : * TransactionId and a set of flag bits. The name is a bit historical:
10 : * originally, a MultiXactId consisted of more than one TransactionId (except
11 : * in rare corner cases), hence "multi". Nowadays, however, it's perfectly
12 : * legitimate to have MultiXactIds that only include a single Xid.
13 : *
14 : * The meaning of the flag bits is opaque to this module, but they are mostly
15 : * used in heapam.c to identify lock modes that each of the member transactions
16 : * is holding on any given tuple. This module just contains support to store
17 : * and retrieve the arrays.
18 : *
19 : * We use two SLRU areas, one for storing the offsets at which the data
20 : * starts for each MultiXactId in the other one. This trick allows us to
21 : * store variable length arrays of TransactionIds. (We could alternatively
22 : * use one area containing counts and TransactionIds, with valid MultiXactId
23 : * values pointing at slots containing counts; but that way seems less robust
24 : * since it would get completely confused if someone inquired about a bogus
25 : * MultiXactId that pointed to an intermediate slot containing an XID.)
26 : *
27 : * XLOG interactions: this module generates a record whenever a new OFFSETs or
28 : * MEMBERs page is initialized to zeroes, as well as an
29 : * XLOG_MULTIXACT_CREATE_ID record whenever a new MultiXactId is defined.
30 : * This module ignores the WAL rule "write xlog before data," because it
31 : * suffices that actions recording a MultiXactId in a heap xmax do follow that
32 : * rule. The only way for the MXID to be referenced from any data page is for
33 : * heap_lock_tuple() or heap_update() to have put it there, and each generates
34 : * an XLOG record that must follow ours. The normal LSN interlock between the
35 : * data page and that XLOG record will ensure that our XLOG record reaches
36 : * disk first. If the SLRU members/offsets data reaches disk sooner than the
37 : * XLOG records, we do not care; after recovery, no xmax will refer to it. On
38 : * the flip side, to ensure that all referenced entries _do_ reach disk, this
39 : * module's XLOG records completely rebuild the data entered since the last
40 : * checkpoint. We flush and sync all dirty OFFSETs and MEMBERs pages to disk
41 : * before each checkpoint is considered complete.
42 : *
43 : * Like clog.c, and unlike subtrans.c, we have to preserve state across
44 : * crashes and ensure that MXID and offset numbering increases monotonically
45 : * across a crash. We do this in the same way as it's done for transaction
46 : * IDs: the WAL record is guaranteed to contain evidence of every MXID we
47 : * could need to worry about, and we just make sure that at the end of
48 : * replay, the next-MXID and next-offset counters are at least as large as
49 : * anything we saw during replay.
50 : *
51 : * We are able to remove segments no longer necessary by carefully tracking
52 : * each table's used values: during vacuum, any multixact older than a certain
53 : * value is removed; the cutoff value is stored in pg_class. The minimum value
54 : * across all tables in each database is stored in pg_database, and the global
55 : * minimum across all databases is part of pg_control and is kept in shared
56 : * memory. Whenever that minimum is advanced, the SLRUs are truncated.
57 : *
58 : * When new multixactid values are to be created, care is taken that the
59 : * counter does not fall within the wraparound horizon considering the global
60 : * minimum value.
61 : *
62 : * Portions Copyright (c) 1996-2025, 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 4716 : MultiXactIdToOffsetPage(MultiXactId multi)
114 : {
115 4716 : return multi / MULTIXACT_OFFSETS_PER_PAGE;
116 : }
117 :
118 : static inline int
119 3346 : MultiXactIdToOffsetEntry(MultiXactId multi)
120 : {
121 3346 : return multi % MULTIXACT_OFFSETS_PER_PAGE;
122 : }
123 :
124 : static inline int64
125 0 : MultiXactIdToOffsetSegment(MultiXactId multi)
126 : {
127 0 : 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 5890 : MXOffsetToMemberPage(MultiXactOffset offset)
174 : {
175 5890 : return offset / MULTIXACT_MEMBERS_PER_PAGE;
176 : }
177 :
178 : static inline int64
179 0 : MXOffsetToMemberSegment(MultiXactOffset offset)
180 : {
181 0 : 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 6978 : MXOffsetToFlagsOffset(MultiXactOffset offset)
187 : {
188 6978 : MultiXactOffset group = offset / MULTIXACT_MEMBERS_PER_MEMBERGROUP;
189 6978 : int grouponpg = group % MULTIXACT_MEMBERGROUPS_PER_PAGE;
190 6978 : int byteoff = grouponpg * MULTIXACT_MEMBERGROUP_SIZE;
191 :
192 6978 : return byteoff;
193 : }
194 :
195 : static inline int
196 2908 : MXOffsetToFlagsBitShift(MultiXactOffset offset)
197 : {
198 2908 : int member_in_group = offset % MULTIXACT_MEMBERS_PER_MEMBERGROUP;
199 2908 : int bshift = member_in_group * MXACT_MEMBER_BITS_PER_XACT;
200 :
201 2908 : return bshift;
202 : }
203 :
204 : /* Location (byte offset within page) of TransactionId of given member */
205 : static inline int
206 2370 : MXOffsetToMemberOffset(MultiXactOffset offset)
207 : {
208 2370 : int member_in_group = offset % MULTIXACT_MEMBERS_PER_MEMBERGROUP;
209 :
210 2370 : return MXOffsetToFlagsOffset(offset) +
211 2370 : 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 0 : PreviousMultiXactId(MultiXactId multi)
222 : {
223 0 : 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 : * This is used to sleep until a multixact offset is written when we want
280 : * to create the next one.
281 : */
282 : ConditionVariable nextoff_cv;
283 :
284 : /*
285 : * Per-backend data starts here. We have two arrays stored in the area
286 : * immediately following the MultiXactStateData struct. Each is indexed by
287 : * ProcNumber.
288 : *
289 : * In both arrays, there's a slot for all normal backends
290 : * (0..MaxBackends-1) followed by a slot for max_prepared_xacts prepared
291 : * transactions.
292 : *
293 : * OldestMemberMXactId[k] is the oldest MultiXactId each backend's current
294 : * transaction(s) could possibly be a member of, or InvalidMultiXactId
295 : * when the backend has no live transaction that could possibly be a
296 : * member of a MultiXact. Each backend sets its entry to the current
297 : * nextMXact counter just before first acquiring a shared lock in a given
298 : * transaction, and clears it at transaction end. (This works because only
299 : * during or after acquiring a shared lock could an XID possibly become a
300 : * member of a MultiXact, and that MultiXact would have to be created
301 : * during or after the lock acquisition.)
302 : *
303 : * OldestVisibleMXactId[k] is the oldest MultiXactId each backend's
304 : * current transaction(s) think is potentially live, or InvalidMultiXactId
305 : * when not in a transaction or not in a transaction that's paid any
306 : * attention to MultiXacts yet. This is computed when first needed in a
307 : * given transaction, and cleared at transaction end. We can compute it
308 : * as the minimum of the valid OldestMemberMXactId[] entries at the time
309 : * we compute it (using nextMXact if none are valid). Each backend is
310 : * required not to attempt to access any SLRU data for MultiXactIds older
311 : * than its own OldestVisibleMXactId[] setting; this is necessary because
312 : * the relevant SLRU data can be concurrently truncated away.
313 : *
314 : * The oldest valid value among all of the OldestMemberMXactId[] and
315 : * OldestVisibleMXactId[] entries is considered by vacuum as the earliest
316 : * possible value still having any live member transaction -- OldestMxact.
317 : * Any value older than that is typically removed from tuple headers, or
318 : * "frozen" via being replaced with a new xmax. VACUUM can sometimes even
319 : * remove an individual MultiXact xmax whose value is >= its OldestMxact
320 : * cutoff, though typically only when no individual member XID is still
321 : * running. See FreezeMultiXactId for full details.
322 : *
323 : * Whenever VACUUM advances relminmxid, then either its OldestMxact cutoff
324 : * or the oldest extant Multi remaining in the table is used as the new
325 : * pg_class.relminmxid value (whichever is earlier). The minimum of all
326 : * relminmxid values in each database is stored in pg_database.datminmxid.
327 : * In turn, the minimum of all of those values is stored in pg_control.
328 : * This is used as the truncation point for pg_multixact when unneeded
329 : * segments get removed by vac_truncate_clog() during vacuuming.
330 : */
331 : MultiXactId perBackendXactIds[FLEXIBLE_ARRAY_MEMBER];
332 : } MultiXactStateData;
333 :
334 : /*
335 : * Size of OldestMemberMXactId and OldestVisibleMXactId arrays.
336 : */
337 : #define MaxOldestSlot (MaxBackends + max_prepared_xacts)
338 :
339 : /* Pointers to the state data in shared memory */
340 : static MultiXactStateData *MultiXactState;
341 : static MultiXactId *OldestMemberMXactId;
342 : static MultiXactId *OldestVisibleMXactId;
343 :
344 :
345 : /*
346 : * Definitions for the backend-local MultiXactId cache.
347 : *
348 : * We use this cache to store known MultiXacts, so we don't need to go to
349 : * SLRU areas every time.
350 : *
351 : * The cache lasts for the duration of a single transaction, the rationale
352 : * for this being that most entries will contain our own TransactionId and
353 : * so they will be uninteresting by the time our next transaction starts.
354 : * (XXX not clear that this is correct --- other members of the MultiXact
355 : * could hang around longer than we did. However, it's not clear what a
356 : * better policy for flushing old cache entries would be.) FIXME actually
357 : * this is plain wrong now that multixact's may contain update Xids.
358 : *
359 : * We allocate the cache entries in a memory context that is deleted at
360 : * transaction end, so we don't need to do retail freeing of entries.
361 : */
362 : typedef struct mXactCacheEnt
363 : {
364 : MultiXactId multi;
365 : int nmembers;
366 : dlist_node node;
367 : MultiXactMember members[FLEXIBLE_ARRAY_MEMBER];
368 : } mXactCacheEnt;
369 :
370 : #define MAX_CACHE_ENTRIES 256
371 : static dclist_head MXactCache = DCLIST_STATIC_INIT(MXactCache);
372 : static MemoryContext MXactContext = NULL;
373 :
374 : #ifdef MULTIXACT_DEBUG
375 : #define debug_elog2(a,b) elog(a,b)
376 : #define debug_elog3(a,b,c) elog(a,b,c)
377 : #define debug_elog4(a,b,c,d) elog(a,b,c,d)
378 : #define debug_elog5(a,b,c,d,e) elog(a,b,c,d,e)
379 : #define debug_elog6(a,b,c,d,e,f) elog(a,b,c,d,e,f)
380 : #else
381 : #define debug_elog2(a,b)
382 : #define debug_elog3(a,b,c)
383 : #define debug_elog4(a,b,c,d)
384 : #define debug_elog5(a,b,c,d,e)
385 : #define debug_elog6(a,b,c,d,e,f)
386 : #endif
387 :
388 : /* internal MultiXactId management */
389 : static void MultiXactIdSetOldestVisible(void);
390 : static void RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
391 : int nmembers, MultiXactMember *members);
392 : static MultiXactId GetNewMultiXactId(int nmembers, MultiXactOffset *offset);
393 :
394 : /* MultiXact cache management */
395 : static int mxactMemberComparator(const void *arg1, const void *arg2);
396 : static MultiXactId mXactCacheGetBySet(int nmembers, MultiXactMember *members);
397 : static int mXactCacheGetById(MultiXactId multi, MultiXactMember **members);
398 : static void mXactCachePut(MultiXactId multi, int nmembers,
399 : MultiXactMember *members);
400 :
401 : static char *mxstatus_to_string(MultiXactStatus status);
402 :
403 : /* management of SLRU infrastructure */
404 : static int ZeroMultiXactOffsetPage(int64 pageno, bool writeXlog);
405 : static int ZeroMultiXactMemberPage(int64 pageno, bool writeXlog);
406 : static bool MultiXactOffsetPagePrecedes(int64 page1, int64 page2);
407 : static bool MultiXactMemberPagePrecedes(int64 page1, int64 page2);
408 : static bool MultiXactOffsetPrecedes(MultiXactOffset offset1,
409 : MultiXactOffset offset2);
410 : static void ExtendMultiXactOffset(MultiXactId multi);
411 : static void ExtendMultiXactMember(MultiXactOffset offset, int nmembers);
412 : static bool MultiXactOffsetWouldWrap(MultiXactOffset boundary,
413 : MultiXactOffset start, uint32 distance);
414 : static bool SetOffsetVacuumLimit(bool is_startup);
415 : static bool find_multixact_start(MultiXactId multi, MultiXactOffset *result);
416 : static void WriteMZeroPageXlogRec(int64 pageno, uint8 info);
417 : static void WriteMTruncateXlogRec(Oid oldestMultiDB,
418 : MultiXactId startTruncOff,
419 : MultiXactId endTruncOff,
420 : MultiXactOffset startTruncMemb,
421 : MultiXactOffset endTruncMemb);
422 :
423 :
424 : /*
425 : * MultiXactIdCreate
426 : * Construct a MultiXactId representing two TransactionIds.
427 : *
428 : * The two XIDs must be different, or be requesting different statuses.
429 : *
430 : * NB - we don't worry about our local MultiXactId cache here, because that
431 : * is handled by the lower-level routines.
432 : */
433 : MultiXactId
434 2064 : MultiXactIdCreate(TransactionId xid1, MultiXactStatus status1,
435 : TransactionId xid2, MultiXactStatus status2)
436 : {
437 : MultiXactId newMulti;
438 : MultiXactMember members[2];
439 :
440 : Assert(TransactionIdIsValid(xid1));
441 : Assert(TransactionIdIsValid(xid2));
442 :
443 : Assert(!TransactionIdEquals(xid1, xid2) || (status1 != status2));
444 :
445 : /* MultiXactIdSetOldestMember() must have been called already. */
446 : Assert(MultiXactIdIsValid(OldestMemberMXactId[MyProcNumber]));
447 :
448 : /*
449 : * Note: unlike MultiXactIdExpand, we don't bother to check that both XIDs
450 : * are still running. In typical usage, xid2 will be our own XID and the
451 : * caller just did a check on xid1, so it'd be wasted effort.
452 : */
453 :
454 2064 : members[0].xid = xid1;
455 2064 : members[0].status = status1;
456 2064 : members[1].xid = xid2;
457 2064 : members[1].status = status2;
458 :
459 2064 : newMulti = MultiXactIdCreateFromMembers(2, members);
460 :
461 : debug_elog3(DEBUG2, "Create: %s",
462 : mxid_to_string(newMulti, 2, members));
463 :
464 2064 : return newMulti;
465 : }
466 :
467 : /*
468 : * MultiXactIdExpand
469 : * Add a TransactionId to a pre-existing MultiXactId.
470 : *
471 : * If the TransactionId is already a member of the passed MultiXactId with the
472 : * same status, just return it as-is.
473 : *
474 : * Note that we do NOT actually modify the membership of a pre-existing
475 : * MultiXactId; instead we create a new one. This is necessary to avoid
476 : * a race condition against code trying to wait for one MultiXactId to finish;
477 : * see notes in heapam.c.
478 : *
479 : * NB - we don't worry about our local MultiXactId cache here, because that
480 : * is handled by the lower-level routines.
481 : *
482 : * Note: It is critical that MultiXactIds that come from an old cluster (i.e.
483 : * one upgraded by pg_upgrade from a cluster older than this feature) are not
484 : * passed in.
485 : */
486 : MultiXactId
487 194 : MultiXactIdExpand(MultiXactId multi, TransactionId xid, MultiXactStatus status)
488 : {
489 : MultiXactId newMulti;
490 : MultiXactMember *members;
491 : MultiXactMember *newMembers;
492 : int nmembers;
493 : int i;
494 : int j;
495 :
496 : Assert(MultiXactIdIsValid(multi));
497 : Assert(TransactionIdIsValid(xid));
498 :
499 : /* MultiXactIdSetOldestMember() must have been called already. */
500 : Assert(MultiXactIdIsValid(OldestMemberMXactId[MyProcNumber]));
501 :
502 : debug_elog5(DEBUG2, "Expand: received multi %u, xid %u status %s",
503 : multi, xid, mxstatus_to_string(status));
504 :
505 : /*
506 : * Note: we don't allow for old multis here. The reason is that the only
507 : * caller of this function does a check that the multixact is no longer
508 : * running.
509 : */
510 194 : nmembers = GetMultiXactIdMembers(multi, &members, false, false);
511 :
512 194 : if (nmembers < 0)
513 : {
514 : MultiXactMember member;
515 :
516 : /*
517 : * The MultiXactId is obsolete. This can only happen if all the
518 : * MultiXactId members stop running between the caller checking and
519 : * passing it to us. It would be better to return that fact to the
520 : * caller, but it would complicate the API and it's unlikely to happen
521 : * too often, so just deal with it by creating a singleton MultiXact.
522 : */
523 0 : member.xid = xid;
524 0 : member.status = status;
525 0 : newMulti = MultiXactIdCreateFromMembers(1, &member);
526 :
527 : debug_elog4(DEBUG2, "Expand: %u has no members, create singleton %u",
528 : multi, newMulti);
529 0 : return newMulti;
530 : }
531 :
532 : /*
533 : * If the TransactionId is already a member of the MultiXactId with the
534 : * same status, just return the existing MultiXactId.
535 : */
536 604 : for (i = 0; i < nmembers; i++)
537 : {
538 410 : if (TransactionIdEquals(members[i].xid, xid) &&
539 108 : (members[i].status == status))
540 : {
541 : debug_elog4(DEBUG2, "Expand: %u is already a member of %u",
542 : xid, multi);
543 0 : pfree(members);
544 0 : return multi;
545 : }
546 : }
547 :
548 : /*
549 : * Determine which of the members of the MultiXactId are still of
550 : * interest. This is any running transaction, and also any transaction
551 : * that grabbed something stronger than just a lock and was committed. (An
552 : * update that aborted is of no interest here; and having more than one
553 : * update Xid in a multixact would cause errors elsewhere.)
554 : *
555 : * Removing dead members is not just an optimization: freezing of tuples
556 : * whose Xmax are multis depends on this behavior.
557 : *
558 : * Note we have the same race condition here as above: j could be 0 at the
559 : * end of the loop.
560 : */
561 : newMembers = (MultiXactMember *)
562 194 : palloc(sizeof(MultiXactMember) * (nmembers + 1));
563 :
564 604 : for (i = 0, j = 0; i < nmembers; i++)
565 : {
566 410 : if (TransactionIdIsInProgress(members[i].xid) ||
567 94 : (ISUPDATE_from_mxstatus(members[i].status) &&
568 12 : TransactionIdDidCommit(members[i].xid)))
569 : {
570 328 : newMembers[j].xid = members[i].xid;
571 328 : newMembers[j++].status = members[i].status;
572 : }
573 : }
574 :
575 194 : newMembers[j].xid = xid;
576 194 : newMembers[j++].status = status;
577 194 : newMulti = MultiXactIdCreateFromMembers(j, newMembers);
578 :
579 194 : pfree(members);
580 194 : pfree(newMembers);
581 :
582 : debug_elog3(DEBUG2, "Expand: returning new multi %u", newMulti);
583 :
584 194 : return newMulti;
585 : }
586 :
587 : /*
588 : * MultiXactIdIsRunning
589 : * Returns whether a MultiXactId is "running".
590 : *
591 : * We return true if at least one member of the given MultiXactId is still
592 : * running. Note that a "false" result is certain not to change,
593 : * because it is not legal to add members to an existing MultiXactId.
594 : *
595 : * Caller is expected to have verified that the multixact does not come from
596 : * a pg_upgraded share-locked tuple.
597 : */
598 : bool
599 1970 : MultiXactIdIsRunning(MultiXactId multi, bool isLockOnly)
600 : {
601 : MultiXactMember *members;
602 : int nmembers;
603 : int i;
604 :
605 : debug_elog3(DEBUG2, "IsRunning %u?", multi);
606 :
607 : /*
608 : * "false" here means we assume our callers have checked that the given
609 : * multi cannot possibly come from a pg_upgraded database.
610 : */
611 1970 : nmembers = GetMultiXactIdMembers(multi, &members, false, isLockOnly);
612 :
613 1970 : if (nmembers <= 0)
614 : {
615 : debug_elog2(DEBUG2, "IsRunning: no members");
616 1336 : return false;
617 : }
618 :
619 : /*
620 : * Checking for myself is cheap compared to looking in shared memory;
621 : * return true if any live subtransaction of the current top-level
622 : * transaction is a member.
623 : *
624 : * This is not needed for correctness, it's just a fast path.
625 : */
626 1468 : for (i = 0; i < nmembers; i++)
627 : {
628 1146 : if (TransactionIdIsCurrentTransactionId(members[i].xid))
629 : {
630 : debug_elog3(DEBUG2, "IsRunning: I (%d) am running!", i);
631 312 : pfree(members);
632 312 : return true;
633 : }
634 : }
635 :
636 : /*
637 : * This could be made faster by having another entry point in procarray.c,
638 : * walking the PGPROC array only once for all the members. But in most
639 : * cases nmembers should be small enough that it doesn't much matter.
640 : */
641 586 : for (i = 0; i < nmembers; i++)
642 : {
643 494 : if (TransactionIdIsInProgress(members[i].xid))
644 : {
645 : debug_elog4(DEBUG2, "IsRunning: member %d (%u) is running",
646 : i, members[i].xid);
647 230 : pfree(members);
648 230 : return true;
649 : }
650 : }
651 :
652 92 : pfree(members);
653 :
654 : debug_elog3(DEBUG2, "IsRunning: %u is not running", multi);
655 :
656 92 : return false;
657 : }
658 :
659 : /*
660 : * MultiXactIdSetOldestMember
661 : * Save the oldest MultiXactId this transaction could be a member of.
662 : *
663 : * We set the OldestMemberMXactId for a given transaction the first time it's
664 : * going to do some operation that might require a MultiXactId (tuple lock,
665 : * update or delete). We need to do this even if we end up using a
666 : * TransactionId instead of a MultiXactId, because there is a chance that
667 : * another transaction would add our XID to a MultiXactId.
668 : *
669 : * The value to set is the next-to-be-assigned MultiXactId, so this is meant to
670 : * be called just before doing any such possibly-MultiXactId-able operation.
671 : */
672 : void
673 3719718 : MultiXactIdSetOldestMember(void)
674 : {
675 3719718 : if (!MultiXactIdIsValid(OldestMemberMXactId[MyProcNumber]))
676 : {
677 : MultiXactId nextMXact;
678 :
679 : /*
680 : * You might think we don't need to acquire a lock here, since
681 : * fetching and storing of TransactionIds is probably atomic, but in
682 : * fact we do: suppose we pick up nextMXact and then lose the CPU for
683 : * a long time. Someone else could advance nextMXact, and then
684 : * another someone else could compute an OldestVisibleMXactId that
685 : * would be after the value we are going to store when we get control
686 : * back. Which would be wrong.
687 : *
688 : * Note that a shared lock is sufficient, because it's enough to stop
689 : * someone from advancing nextMXact; and nobody else could be trying
690 : * to write to our OldestMember entry, only reading (and we assume
691 : * storing it is atomic.)
692 : */
693 139552 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
694 :
695 : /*
696 : * We have to beware of the possibility that nextMXact is in the
697 : * wrapped-around state. We don't fix the counter itself here, but we
698 : * must be sure to store a valid value in our array entry.
699 : */
700 139552 : nextMXact = MultiXactState->nextMXact;
701 139552 : if (nextMXact < FirstMultiXactId)
702 0 : nextMXact = FirstMultiXactId;
703 :
704 139552 : OldestMemberMXactId[MyProcNumber] = nextMXact;
705 :
706 139552 : LWLockRelease(MultiXactGenLock);
707 :
708 : debug_elog4(DEBUG2, "MultiXact: setting OldestMember[%d] = %u",
709 : MyProcNumber, nextMXact);
710 : }
711 3719718 : }
712 :
713 : /*
714 : * MultiXactIdSetOldestVisible
715 : * Save the oldest MultiXactId this transaction considers possibly live.
716 : *
717 : * We set the OldestVisibleMXactId for a given transaction the first time
718 : * it's going to inspect any MultiXactId. Once we have set this, we are
719 : * guaranteed that SLRU data for MultiXactIds >= our own OldestVisibleMXactId
720 : * won't be truncated away.
721 : *
722 : * The value to set is the oldest of nextMXact and all the valid per-backend
723 : * OldestMemberMXactId[] entries. Because of the locking we do, we can be
724 : * certain that no subsequent call to MultiXactIdSetOldestMember can set
725 : * an OldestMemberMXactId[] entry older than what we compute here. Therefore
726 : * there is no live transaction, now or later, that can be a member of any
727 : * MultiXactId older than the OldestVisibleMXactId we compute here.
728 : */
729 : static void
730 1770 : MultiXactIdSetOldestVisible(void)
731 : {
732 1770 : if (!MultiXactIdIsValid(OldestVisibleMXactId[MyProcNumber]))
733 : {
734 : MultiXactId oldestMXact;
735 : int i;
736 :
737 424 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
738 :
739 : /*
740 : * We have to beware of the possibility that nextMXact is in the
741 : * wrapped-around state. We don't fix the counter itself here, but we
742 : * must be sure to store a valid value in our array entry.
743 : */
744 424 : oldestMXact = MultiXactState->nextMXact;
745 424 : if (oldestMXact < FirstMultiXactId)
746 0 : oldestMXact = FirstMultiXactId;
747 :
748 57136 : for (i = 0; i < MaxOldestSlot; i++)
749 : {
750 56712 : MultiXactId thisoldest = OldestMemberMXactId[i];
751 :
752 57188 : if (MultiXactIdIsValid(thisoldest) &&
753 476 : MultiXactIdPrecedes(thisoldest, oldestMXact))
754 230 : oldestMXact = thisoldest;
755 : }
756 :
757 424 : OldestVisibleMXactId[MyProcNumber] = oldestMXact;
758 :
759 424 : LWLockRelease(MultiXactGenLock);
760 :
761 : debug_elog4(DEBUG2, "MultiXact: setting OldestVisible[%d] = %u",
762 : MyProcNumber, oldestMXact);
763 : }
764 1770 : }
765 :
766 : /*
767 : * ReadNextMultiXactId
768 : * Return the next MultiXactId to be assigned, but don't allocate it
769 : */
770 : MultiXactId
771 428194 : ReadNextMultiXactId(void)
772 : {
773 : MultiXactId mxid;
774 :
775 : /* XXX we could presumably do this without a lock. */
776 428194 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
777 428194 : mxid = MultiXactState->nextMXact;
778 428194 : LWLockRelease(MultiXactGenLock);
779 :
780 428194 : if (mxid < FirstMultiXactId)
781 0 : mxid = FirstMultiXactId;
782 :
783 428194 : return mxid;
784 : }
785 :
786 : /*
787 : * ReadMultiXactIdRange
788 : * Get the range of IDs that may still be referenced by a relation.
789 : */
790 : void
791 2860 : ReadMultiXactIdRange(MultiXactId *oldest, MultiXactId *next)
792 : {
793 2860 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
794 2860 : *oldest = MultiXactState->oldestMultiXactId;
795 2860 : *next = MultiXactState->nextMXact;
796 2860 : LWLockRelease(MultiXactGenLock);
797 :
798 2860 : if (*oldest < FirstMultiXactId)
799 0 : *oldest = FirstMultiXactId;
800 2860 : if (*next < FirstMultiXactId)
801 0 : *next = FirstMultiXactId;
802 2860 : }
803 :
804 :
805 : /*
806 : * MultiXactIdCreateFromMembers
807 : * Make a new MultiXactId from the specified set of members
808 : *
809 : * Make XLOG, SLRU and cache entries for a new MultiXactId, recording the
810 : * given TransactionIds as members. Returns the newly created MultiXactId.
811 : *
812 : * NB: the passed members[] array will be sorted in-place.
813 : */
814 : MultiXactId
815 2260 : MultiXactIdCreateFromMembers(int nmembers, MultiXactMember *members)
816 : {
817 : MultiXactId multi;
818 : MultiXactOffset offset;
819 : xl_multixact_create xlrec;
820 :
821 : debug_elog3(DEBUG2, "Create: %s",
822 : mxid_to_string(InvalidMultiXactId, nmembers, members));
823 :
824 : /*
825 : * See if the same set of members already exists in our cache; if so, just
826 : * re-use that MultiXactId. (Note: it might seem that looking in our
827 : * cache is insufficient, and we ought to search disk to see if a
828 : * duplicate definition already exists. But since we only ever create
829 : * MultiXacts containing our own XID, in most cases any such MultiXacts
830 : * were in fact created by us, and so will be in our cache. There are
831 : * corner cases where someone else added us to a MultiXact without our
832 : * knowledge, but it's not worth checking for.)
833 : */
834 2260 : multi = mXactCacheGetBySet(nmembers, members);
835 2260 : if (MultiXactIdIsValid(multi))
836 : {
837 : debug_elog2(DEBUG2, "Create: in cache!");
838 1700 : return multi;
839 : }
840 :
841 : /* Verify that there is a single update Xid among the given members. */
842 : {
843 : int i;
844 560 : bool has_update = false;
845 :
846 1802 : for (i = 0; i < nmembers; i++)
847 : {
848 1242 : if (ISUPDATE_from_mxstatus(members[i].status))
849 : {
850 264 : if (has_update)
851 0 : elog(ERROR, "new multixact has more than one updating member: %s",
852 : mxid_to_string(InvalidMultiXactId, nmembers, members));
853 264 : has_update = true;
854 : }
855 : }
856 : }
857 :
858 : /* Load the injection point before entering the critical section */
859 560 : INJECTION_POINT_LOAD("multixact-create-from-members");
860 :
861 : /*
862 : * Assign the MXID and offsets range to use, and make sure there is space
863 : * in the OFFSETs and MEMBERs files. NB: this routine does
864 : * START_CRIT_SECTION().
865 : *
866 : * Note: unlike MultiXactIdCreate and MultiXactIdExpand, we do not check
867 : * that we've called MultiXactIdSetOldestMember here. This is because
868 : * this routine is used in some places to create new MultiXactIds of which
869 : * the current backend is not a member, notably during freezing of multis
870 : * in vacuum. During vacuum, in particular, it would be unacceptable to
871 : * keep OldestMulti set, in case it runs for long.
872 : */
873 560 : multi = GetNewMultiXactId(nmembers, &offset);
874 :
875 560 : INJECTION_POINT_CACHED("multixact-create-from-members", NULL);
876 :
877 : /* Make an XLOG entry describing the new MXID. */
878 560 : xlrec.mid = multi;
879 560 : xlrec.moff = offset;
880 560 : xlrec.nmembers = nmembers;
881 :
882 : /*
883 : * XXX Note: there's a lot of padding space in MultiXactMember. We could
884 : * find a more compact representation of this Xlog record -- perhaps all
885 : * the status flags in one XLogRecData, then all the xids in another one?
886 : * Not clear that it's worth the trouble though.
887 : */
888 560 : XLogBeginInsert();
889 560 : XLogRegisterData(&xlrec, SizeOfMultiXactCreate);
890 560 : XLogRegisterData(members, nmembers * sizeof(MultiXactMember));
891 :
892 560 : (void) XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_CREATE_ID);
893 :
894 : /* Now enter the information into the OFFSETs and MEMBERs logs */
895 560 : RecordNewMultiXact(multi, offset, nmembers, members);
896 :
897 : /* Done with critical section */
898 560 : END_CRIT_SECTION();
899 :
900 : /* Store the new MultiXactId in the local cache, too */
901 560 : mXactCachePut(multi, nmembers, members);
902 :
903 : debug_elog2(DEBUG2, "Create: all done");
904 :
905 560 : return multi;
906 : }
907 :
908 : /*
909 : * RecordNewMultiXact
910 : * Write info about a new multixact into the offsets and members files
911 : *
912 : * This is broken out of MultiXactIdCreateFromMembers so that xlog replay can
913 : * use it.
914 : */
915 : static void
916 564 : RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
917 : int nmembers, MultiXactMember *members)
918 : {
919 : int64 pageno;
920 : int64 prev_pageno;
921 : int entryno;
922 : int slotno;
923 : MultiXactOffset *offptr;
924 : int i;
925 : LWLock *lock;
926 564 : LWLock *prevlock = NULL;
927 :
928 564 : pageno = MultiXactIdToOffsetPage(multi);
929 564 : entryno = MultiXactIdToOffsetEntry(multi);
930 :
931 564 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
932 564 : LWLockAcquire(lock, LW_EXCLUSIVE);
933 :
934 : /*
935 : * Note: we pass the MultiXactId to SimpleLruReadPage as the "transaction"
936 : * to complain about if there's any I/O error. This is kinda bogus, but
937 : * since the errors will always give the full pathname, it should be clear
938 : * enough that a MultiXactId is really involved. Perhaps someday we'll
939 : * take the trouble to generalize the slru.c error reporting code.
940 : */
941 564 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
942 564 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
943 564 : offptr += entryno;
944 :
945 564 : *offptr = offset;
946 :
947 564 : MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
948 :
949 : /* Release MultiXactOffset SLRU lock. */
950 564 : LWLockRelease(lock);
951 :
952 : /*
953 : * If anybody was waiting to know the offset of this multixact ID we just
954 : * wrote, they can read it now, so wake them up.
955 : */
956 564 : ConditionVariableBroadcast(&MultiXactState->nextoff_cv);
957 :
958 564 : prev_pageno = -1;
959 :
960 1814 : for (i = 0; i < nmembers; i++, offset++)
961 : {
962 : TransactionId *memberptr;
963 : uint32 *flagsptr;
964 : uint32 flagsval;
965 : int bshift;
966 : int flagsoff;
967 : int memberoff;
968 :
969 : Assert(members[i].status <= MultiXactStatusUpdate);
970 :
971 1250 : pageno = MXOffsetToMemberPage(offset);
972 1250 : memberoff = MXOffsetToMemberOffset(offset);
973 1250 : flagsoff = MXOffsetToFlagsOffset(offset);
974 1250 : bshift = MXOffsetToFlagsBitShift(offset);
975 :
976 1250 : if (pageno != prev_pageno)
977 : {
978 : /*
979 : * MultiXactMember SLRU page is changed so check if this new page
980 : * fall into the different SLRU bank then release the old bank's
981 : * lock and acquire lock on the new bank.
982 : */
983 564 : lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
984 564 : if (lock != prevlock)
985 : {
986 564 : if (prevlock != NULL)
987 0 : LWLockRelease(prevlock);
988 :
989 564 : LWLockAcquire(lock, LW_EXCLUSIVE);
990 564 : prevlock = lock;
991 : }
992 564 : slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
993 564 : prev_pageno = pageno;
994 : }
995 :
996 1250 : memberptr = (TransactionId *)
997 1250 : (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
998 :
999 1250 : *memberptr = members[i].xid;
1000 :
1001 1250 : flagsptr = (uint32 *)
1002 1250 : (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
1003 :
1004 1250 : flagsval = *flagsptr;
1005 1250 : flagsval &= ~(((1 << MXACT_MEMBER_BITS_PER_XACT) - 1) << bshift);
1006 1250 : flagsval |= (members[i].status << bshift);
1007 1250 : *flagsptr = flagsval;
1008 :
1009 1250 : MultiXactMemberCtl->shared->page_dirty[slotno] = true;
1010 : }
1011 :
1012 564 : if (prevlock != NULL)
1013 564 : LWLockRelease(prevlock);
1014 564 : }
1015 :
1016 : /*
1017 : * GetNewMultiXactId
1018 : * Get the next MultiXactId.
1019 : *
1020 : * Also, reserve the needed amount of space in the "members" area. The
1021 : * starting offset of the reserved space is returned in *offset.
1022 : *
1023 : * This may generate XLOG records for expansion of the offsets and/or members
1024 : * files. Unfortunately, we have to do that while holding MultiXactGenLock
1025 : * to avoid race conditions --- the XLOG record for zeroing a page must appear
1026 : * before any backend can possibly try to store data in that page!
1027 : *
1028 : * We start a critical section before advancing the shared counters. The
1029 : * caller must end the critical section after writing SLRU data.
1030 : */
1031 : static MultiXactId
1032 560 : GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
1033 : {
1034 : MultiXactId result;
1035 : MultiXactOffset nextOffset;
1036 :
1037 : debug_elog3(DEBUG2, "GetNew: for %d xids", nmembers);
1038 :
1039 : /* safety check, we should never get this far in a HS standby */
1040 560 : if (RecoveryInProgress())
1041 0 : elog(ERROR, "cannot assign MultiXactIds during recovery");
1042 :
1043 560 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1044 :
1045 : /* Handle wraparound of the nextMXact counter */
1046 560 : if (MultiXactState->nextMXact < FirstMultiXactId)
1047 0 : MultiXactState->nextMXact = FirstMultiXactId;
1048 :
1049 : /* Assign the MXID */
1050 560 : result = MultiXactState->nextMXact;
1051 :
1052 : /*----------
1053 : * Check to see if it's safe to assign another MultiXactId. This protects
1054 : * against catastrophic data loss due to multixact wraparound. The basic
1055 : * rules are:
1056 : *
1057 : * If we're past multiVacLimit or the safe threshold for member storage
1058 : * space, or we don't know what the safe threshold for member storage is,
1059 : * start trying to force autovacuum cycles.
1060 : * If we're past multiWarnLimit, start issuing warnings.
1061 : * If we're past multiStopLimit, refuse to create new MultiXactIds.
1062 : *
1063 : * Note these are pretty much the same protections in GetNewTransactionId.
1064 : *----------
1065 : */
1066 560 : if (!MultiXactIdPrecedes(result, MultiXactState->multiVacLimit))
1067 : {
1068 : /*
1069 : * For safety's sake, we release MultiXactGenLock while sending
1070 : * signals, warnings, etc. This is not so much because we care about
1071 : * preserving concurrency in this situation, as to avoid any
1072 : * possibility of deadlock while doing get_database_name(). First,
1073 : * copy all the shared values we'll need in this path.
1074 : */
1075 0 : MultiXactId multiWarnLimit = MultiXactState->multiWarnLimit;
1076 0 : MultiXactId multiStopLimit = MultiXactState->multiStopLimit;
1077 0 : MultiXactId multiWrapLimit = MultiXactState->multiWrapLimit;
1078 0 : Oid oldest_datoid = MultiXactState->oldestMultiXactDB;
1079 :
1080 0 : LWLockRelease(MultiXactGenLock);
1081 :
1082 0 : if (IsUnderPostmaster &&
1083 0 : !MultiXactIdPrecedes(result, multiStopLimit))
1084 : {
1085 0 : char *oldest_datname = get_database_name(oldest_datoid);
1086 :
1087 : /*
1088 : * Immediately kick autovacuum into action as we're already in
1089 : * ERROR territory.
1090 : */
1091 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1092 :
1093 : /* complain even if that DB has disappeared */
1094 0 : if (oldest_datname)
1095 0 : ereport(ERROR,
1096 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1097 : errmsg("database is not accepting commands that assign new MultiXactIds to avoid wraparound data loss in database \"%s\"",
1098 : oldest_datname),
1099 : errhint("Execute a database-wide VACUUM in that database.\n"
1100 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1101 : else
1102 0 : ereport(ERROR,
1103 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1104 : errmsg("database is not accepting commands that assign new MultiXactIds to avoid wraparound data loss in database with OID %u",
1105 : oldest_datoid),
1106 : errhint("Execute a database-wide VACUUM in that database.\n"
1107 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1108 : }
1109 :
1110 : /*
1111 : * To avoid swamping the postmaster with signals, we issue the autovac
1112 : * request only once per 64K multis generated. This still gives
1113 : * plenty of chances before we get into real trouble.
1114 : */
1115 0 : if (IsUnderPostmaster && (result % 65536) == 0)
1116 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1117 :
1118 0 : if (!MultiXactIdPrecedes(result, multiWarnLimit))
1119 : {
1120 0 : char *oldest_datname = get_database_name(oldest_datoid);
1121 :
1122 : /* complain even if that DB has disappeared */
1123 0 : if (oldest_datname)
1124 0 : ereport(WARNING,
1125 : (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
1126 : "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
1127 : multiWrapLimit - result,
1128 : oldest_datname,
1129 : multiWrapLimit - result),
1130 : errhint("Execute a database-wide VACUUM in that database.\n"
1131 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1132 : else
1133 0 : ereport(WARNING,
1134 : (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
1135 : "database with OID %u must be vacuumed before %u more MultiXactIds are used",
1136 : multiWrapLimit - result,
1137 : oldest_datoid,
1138 : multiWrapLimit - result),
1139 : errhint("Execute a database-wide VACUUM in that database.\n"
1140 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1141 : }
1142 :
1143 : /* Re-acquire lock and start over */
1144 0 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1145 0 : result = MultiXactState->nextMXact;
1146 0 : if (result < FirstMultiXactId)
1147 0 : result = FirstMultiXactId;
1148 : }
1149 :
1150 : /* Make sure there is room for the MXID in the file. */
1151 560 : ExtendMultiXactOffset(result);
1152 :
1153 : /*
1154 : * Reserve the members space, similarly to above. Also, be careful not to
1155 : * return zero as the starting offset for any multixact. See
1156 : * GetMultiXactIdMembers() for motivation.
1157 : */
1158 560 : nextOffset = MultiXactState->nextOffset;
1159 560 : if (nextOffset == 0)
1160 : {
1161 22 : *offset = 1;
1162 22 : nmembers++; /* allocate member slot 0 too */
1163 : }
1164 : else
1165 538 : *offset = nextOffset;
1166 :
1167 : /*----------
1168 : * Protect against overrun of the members space as well, with the
1169 : * following rules:
1170 : *
1171 : * If we're past offsetStopLimit, refuse to generate more multis.
1172 : * If we're close to offsetStopLimit, emit a warning.
1173 : *
1174 : * Arbitrarily, we start emitting warnings when we're 20 segments or less
1175 : * from offsetStopLimit.
1176 : *
1177 : * Note we haven't updated the shared state yet, so if we fail at this
1178 : * point, the multixact ID we grabbed can still be used by the next guy.
1179 : *
1180 : * Note that there is no point in forcing autovacuum runs here: the
1181 : * multixact freeze settings would have to be reduced for that to have any
1182 : * effect.
1183 : *----------
1184 : */
1185 : #define OFFSET_WARN_SEGMENTS 20
1186 1120 : if (MultiXactState->oldestOffsetKnown &&
1187 560 : MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit, nextOffset,
1188 : nmembers))
1189 : {
1190 : /* see comment in the corresponding offsets wraparound case */
1191 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1192 :
1193 0 : ereport(ERROR,
1194 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1195 : errmsg("multixact \"members\" limit exceeded"),
1196 : errdetail_plural("This command would create a multixact with %u members, but the remaining space is only enough for %u member.",
1197 : "This command would create a multixact with %u members, but the remaining space is only enough for %u members.",
1198 : MultiXactState->offsetStopLimit - nextOffset - 1,
1199 : nmembers,
1200 : MultiXactState->offsetStopLimit - nextOffset - 1),
1201 : 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.",
1202 : MultiXactState->oldestMultiXactDB)));
1203 : }
1204 :
1205 : /*
1206 : * Check whether we should kick autovacuum into action, to prevent members
1207 : * wraparound. NB we use a much larger window to trigger autovacuum than
1208 : * just the warning limit. The warning is just a measure of last resort -
1209 : * this is in line with GetNewTransactionId's behaviour.
1210 : */
1211 560 : if (!MultiXactState->oldestOffsetKnown ||
1212 560 : (MultiXactState->nextOffset - MultiXactState->oldestOffset
1213 560 : > MULTIXACT_MEMBER_SAFE_THRESHOLD))
1214 : {
1215 : /*
1216 : * To avoid swamping the postmaster with signals, we issue the autovac
1217 : * request only when crossing a segment boundary. With default
1218 : * compilation settings that's roughly after 50k members. This still
1219 : * gives plenty of chances before we get into real trouble.
1220 : */
1221 0 : if ((MXOffsetToMemberPage(nextOffset) / SLRU_PAGES_PER_SEGMENT) !=
1222 0 : (MXOffsetToMemberPage(nextOffset + nmembers) / SLRU_PAGES_PER_SEGMENT))
1223 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1224 : }
1225 :
1226 1120 : if (MultiXactState->oldestOffsetKnown &&
1227 560 : MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit,
1228 : nextOffset,
1229 : nmembers + MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT * OFFSET_WARN_SEGMENTS))
1230 0 : ereport(WARNING,
1231 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1232 : errmsg_plural("database with OID %u must be vacuumed before %d more multixact member is used",
1233 : "database with OID %u must be vacuumed before %d more multixact members are used",
1234 : MultiXactState->offsetStopLimit - nextOffset + nmembers,
1235 : MultiXactState->oldestMultiXactDB,
1236 : MultiXactState->offsetStopLimit - nextOffset + nmembers),
1237 : errhint("Execute a database-wide VACUUM in that database with reduced \"vacuum_multixact_freeze_min_age\" and \"vacuum_multixact_freeze_table_age\" settings.")));
1238 :
1239 560 : ExtendMultiXactMember(nextOffset, nmembers);
1240 :
1241 : /*
1242 : * Critical section from here until caller has written the data into the
1243 : * just-reserved SLRU space; we don't want to error out with a partly
1244 : * written MultiXact structure. (In particular, failing to write our
1245 : * start offset after advancing nextMXact would effectively corrupt the
1246 : * previous MultiXact.)
1247 : */
1248 560 : START_CRIT_SECTION();
1249 :
1250 : /*
1251 : * Advance counters. As in GetNewTransactionId(), this must not happen
1252 : * until after file extension has succeeded!
1253 : *
1254 : * We don't care about MultiXactId wraparound here; it will be handled by
1255 : * the next iteration. But note that nextMXact may be InvalidMultiXactId
1256 : * or the first value on a segment-beginning page after this routine
1257 : * exits, so anyone else looking at the variable must be prepared to deal
1258 : * with either case. Similarly, nextOffset may be zero, but we won't use
1259 : * that as the actual start offset of the next multixact.
1260 : */
1261 560 : (MultiXactState->nextMXact)++;
1262 :
1263 560 : MultiXactState->nextOffset += nmembers;
1264 :
1265 560 : LWLockRelease(MultiXactGenLock);
1266 :
1267 : debug_elog4(DEBUG2, "GetNew: returning %u offset %u", result, *offset);
1268 560 : return result;
1269 : }
1270 :
1271 : /*
1272 : * GetMultiXactIdMembers
1273 : * Return the set of MultiXactMembers that make up a MultiXactId
1274 : *
1275 : * Return value is the number of members found, or -1 if there are none,
1276 : * and *members is set to a newly palloc'ed array of members. It's the
1277 : * caller's responsibility to free it when done with it.
1278 : *
1279 : * from_pgupgrade must be passed as true if and only if only the multixact
1280 : * corresponds to a value from a tuple that was locked in a 9.2-or-older
1281 : * installation and later pg_upgrade'd (that is, the infomask is
1282 : * HEAP_LOCKED_UPGRADED). In this case, we know for certain that no members
1283 : * can still be running, so we return -1 just like for an empty multixact
1284 : * without any further checking. It would be wrong to try to resolve such a
1285 : * multixact: either the multixact is within the current valid multixact
1286 : * range, in which case the returned result would be bogus, or outside that
1287 : * range, in which case an error would be raised.
1288 : *
1289 : * In all other cases, the passed multixact must be within the known valid
1290 : * range, that is, greater than or equal to oldestMultiXactId, and less than
1291 : * nextMXact. Otherwise, an error is raised.
1292 : *
1293 : * isLockOnly must be set to true if caller is certain that the given multi
1294 : * is used only to lock tuples; can be false without loss of correctness,
1295 : * but passing a true means we can return quickly without checking for
1296 : * old updates.
1297 : */
1298 : int
1299 5958 : GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
1300 : bool from_pgupgrade, bool isLockOnly)
1301 : {
1302 : int64 pageno;
1303 : int64 prev_pageno;
1304 : int entryno;
1305 : int slotno;
1306 : MultiXactOffset *offptr;
1307 : MultiXactOffset offset;
1308 : int length;
1309 : int truelength;
1310 : MultiXactId oldestMXact;
1311 : MultiXactId nextMXact;
1312 : MultiXactId tmpMXact;
1313 : MultiXactOffset nextOffset;
1314 : MultiXactMember *ptr;
1315 : LWLock *lock;
1316 5958 : bool slept = false;
1317 :
1318 : debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
1319 :
1320 5958 : if (!MultiXactIdIsValid(multi) || from_pgupgrade)
1321 : {
1322 0 : *members = NULL;
1323 0 : return -1;
1324 : }
1325 :
1326 : /* See if the MultiXactId is in the local cache */
1327 5958 : length = mXactCacheGetById(multi, members);
1328 5958 : if (length >= 0)
1329 : {
1330 : debug_elog3(DEBUG2, "GetMembers: found %s in the cache",
1331 : mxid_to_string(multi, length, *members));
1332 4188 : return length;
1333 : }
1334 :
1335 : /* Set our OldestVisibleMXactId[] entry if we didn't already */
1336 1770 : MultiXactIdSetOldestVisible();
1337 :
1338 : /*
1339 : * If we know the multi is used only for locking and not for updates, then
1340 : * we can skip checking if the value is older than our oldest visible
1341 : * multi. It cannot possibly still be running.
1342 : */
1343 3270 : if (isLockOnly &&
1344 1500 : MultiXactIdPrecedes(multi, OldestVisibleMXactId[MyProcNumber]))
1345 : {
1346 : debug_elog2(DEBUG2, "GetMembers: a locker-only multi is too old");
1347 1338 : *members = NULL;
1348 1338 : return -1;
1349 : }
1350 :
1351 : /*
1352 : * We check known limits on MultiXact before resorting to the SLRU area.
1353 : *
1354 : * An ID older than MultiXactState->oldestMultiXactId cannot possibly be
1355 : * useful; it has already been removed, or will be removed shortly, by
1356 : * truncation. If one is passed, an error is raised.
1357 : *
1358 : * Also, an ID >= nextMXact shouldn't ever be seen here; if it is seen, it
1359 : * implies undetected ID wraparound has occurred. This raises a hard
1360 : * error.
1361 : *
1362 : * Shared lock is enough here since we aren't modifying any global state.
1363 : * Acquire it just long enough to grab the current counter values. We may
1364 : * need both nextMXact and nextOffset; see below.
1365 : */
1366 432 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
1367 :
1368 432 : oldestMXact = MultiXactState->oldestMultiXactId;
1369 432 : nextMXact = MultiXactState->nextMXact;
1370 432 : nextOffset = MultiXactState->nextOffset;
1371 :
1372 432 : LWLockRelease(MultiXactGenLock);
1373 :
1374 432 : if (MultiXactIdPrecedes(multi, oldestMXact))
1375 0 : ereport(ERROR,
1376 : (errcode(ERRCODE_INTERNAL_ERROR),
1377 : errmsg("MultiXactId %u does no longer exist -- apparent wraparound",
1378 : multi)));
1379 :
1380 432 : if (!MultiXactIdPrecedes(multi, nextMXact))
1381 0 : ereport(ERROR,
1382 : (errcode(ERRCODE_INTERNAL_ERROR),
1383 : errmsg("MultiXactId %u has not been created yet -- apparent wraparound",
1384 : multi)));
1385 :
1386 : /*
1387 : * Find out the offset at which we need to start reading MultiXactMembers
1388 : * and the number of members in the multixact. We determine the latter as
1389 : * the difference between this multixact's starting offset and the next
1390 : * one's. However, there are some corner cases to worry about:
1391 : *
1392 : * 1. This multixact may be the latest one created, in which case there is
1393 : * no next one to look at. In this case the nextOffset value we just
1394 : * saved is the correct endpoint.
1395 : *
1396 : * 2. The next multixact may still be in process of being filled in: that
1397 : * is, another process may have done GetNewMultiXactId but not yet written
1398 : * the offset entry for that ID. In that scenario, it is guaranteed that
1399 : * the offset entry for that multixact exists (because GetNewMultiXactId
1400 : * won't release MultiXactGenLock until it does) but contains zero
1401 : * (because we are careful to pre-zero offset pages). Because
1402 : * GetNewMultiXactId will never return zero as the starting offset for a
1403 : * multixact, when we read zero as the next multixact's offset, we know we
1404 : * have this case. We handle this by sleeping on the condition variable
1405 : * we have just for this; the process in charge will signal the CV as soon
1406 : * as it has finished writing the multixact offset.
1407 : *
1408 : * 3. Because GetNewMultiXactId increments offset zero to offset one to
1409 : * handle case #2, there is an ambiguity near the point of offset
1410 : * wraparound. If we see next multixact's offset is one, is that our
1411 : * multixact's actual endpoint, or did it end at zero with a subsequent
1412 : * increment? We handle this using the knowledge that if the zero'th
1413 : * member slot wasn't filled, it'll contain zero, and zero isn't a valid
1414 : * transaction ID so it can't be a multixact member. Therefore, if we
1415 : * read a zero from the members array, just ignore it.
1416 : *
1417 : * This is all pretty messy, but the mess occurs only in infrequent corner
1418 : * cases, so it seems better than holding the MultiXactGenLock for a long
1419 : * time on every multixact creation.
1420 : */
1421 432 : retry:
1422 434 : pageno = MultiXactIdToOffsetPage(multi);
1423 434 : entryno = MultiXactIdToOffsetEntry(multi);
1424 :
1425 : /* Acquire the bank lock for the page we need. */
1426 434 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
1427 434 : LWLockAcquire(lock, LW_EXCLUSIVE);
1428 :
1429 434 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
1430 434 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1431 434 : offptr += entryno;
1432 434 : offset = *offptr;
1433 :
1434 : Assert(offset != 0);
1435 :
1436 : /*
1437 : * Use the same increment rule as GetNewMultiXactId(), that is, don't
1438 : * handle wraparound explicitly until needed.
1439 : */
1440 434 : tmpMXact = multi + 1;
1441 :
1442 434 : if (nextMXact == tmpMXact)
1443 : {
1444 : /* Corner case 1: there is no next multixact */
1445 374 : length = nextOffset - offset;
1446 : }
1447 : else
1448 : {
1449 : MultiXactOffset nextMXOffset;
1450 :
1451 : /* handle wraparound if needed */
1452 60 : if (tmpMXact < FirstMultiXactId)
1453 0 : tmpMXact = FirstMultiXactId;
1454 :
1455 60 : prev_pageno = pageno;
1456 :
1457 60 : pageno = MultiXactIdToOffsetPage(tmpMXact);
1458 60 : entryno = MultiXactIdToOffsetEntry(tmpMXact);
1459 :
1460 60 : if (pageno != prev_pageno)
1461 : {
1462 : LWLock *newlock;
1463 :
1464 : /*
1465 : * Since we're going to access a different SLRU page, if this page
1466 : * falls under a different bank, release the old bank's lock and
1467 : * acquire the lock of the new bank.
1468 : */
1469 0 : newlock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
1470 0 : if (newlock != lock)
1471 : {
1472 0 : LWLockRelease(lock);
1473 0 : LWLockAcquire(newlock, LW_EXCLUSIVE);
1474 0 : lock = newlock;
1475 : }
1476 0 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
1477 : }
1478 :
1479 60 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1480 60 : offptr += entryno;
1481 60 : nextMXOffset = *offptr;
1482 :
1483 60 : if (nextMXOffset == 0)
1484 : {
1485 : /* Corner case 2: next multixact is still being filled in */
1486 2 : LWLockRelease(lock);
1487 2 : CHECK_FOR_INTERRUPTS();
1488 :
1489 2 : INJECTION_POINT("multixact-get-members-cv-sleep", NULL);
1490 :
1491 2 : ConditionVariableSleep(&MultiXactState->nextoff_cv,
1492 : WAIT_EVENT_MULTIXACT_CREATION);
1493 2 : slept = true;
1494 2 : goto retry;
1495 : }
1496 :
1497 58 : length = nextMXOffset - offset;
1498 : }
1499 :
1500 432 : LWLockRelease(lock);
1501 432 : lock = NULL;
1502 :
1503 : /*
1504 : * If we slept above, clean up state; it's no longer needed.
1505 : */
1506 432 : if (slept)
1507 2 : ConditionVariableCancelSleep();
1508 :
1509 432 : ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
1510 :
1511 432 : truelength = 0;
1512 432 : prev_pageno = -1;
1513 1530 : for (int i = 0; i < length; i++, offset++)
1514 : {
1515 : TransactionId *xactptr;
1516 : uint32 *flagsptr;
1517 : int flagsoff;
1518 : int bshift;
1519 : int memberoff;
1520 :
1521 1098 : pageno = MXOffsetToMemberPage(offset);
1522 1098 : memberoff = MXOffsetToMemberOffset(offset);
1523 :
1524 1098 : if (pageno != prev_pageno)
1525 : {
1526 : LWLock *newlock;
1527 :
1528 : /*
1529 : * Since we're going to access a different SLRU page, if this page
1530 : * falls under a different bank, release the old bank's lock and
1531 : * acquire the lock of the new bank.
1532 : */
1533 432 : newlock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
1534 432 : if (newlock != lock)
1535 : {
1536 432 : if (lock)
1537 0 : LWLockRelease(lock);
1538 432 : LWLockAcquire(newlock, LW_EXCLUSIVE);
1539 432 : lock = newlock;
1540 : }
1541 :
1542 432 : slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
1543 432 : prev_pageno = pageno;
1544 : }
1545 :
1546 1098 : xactptr = (TransactionId *)
1547 1098 : (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
1548 :
1549 1098 : if (!TransactionIdIsValid(*xactptr))
1550 : {
1551 : /* Corner case 3: we must be looking at unused slot zero */
1552 : Assert(offset == 0);
1553 0 : continue;
1554 : }
1555 :
1556 1098 : flagsoff = MXOffsetToFlagsOffset(offset);
1557 1098 : bshift = MXOffsetToFlagsBitShift(offset);
1558 1098 : flagsptr = (uint32 *) (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
1559 :
1560 1098 : ptr[truelength].xid = *xactptr;
1561 1098 : ptr[truelength].status = (*flagsptr >> bshift) & MXACT_MEMBER_XACT_BITMASK;
1562 1098 : truelength++;
1563 : }
1564 :
1565 432 : LWLockRelease(lock);
1566 :
1567 : /* A multixid with zero members should not happen */
1568 : Assert(truelength > 0);
1569 :
1570 : /*
1571 : * Copy the result into the local cache.
1572 : */
1573 432 : mXactCachePut(multi, truelength, ptr);
1574 :
1575 : debug_elog3(DEBUG2, "GetMembers: no cache for %s",
1576 : mxid_to_string(multi, truelength, ptr));
1577 432 : *members = ptr;
1578 432 : return truelength;
1579 : }
1580 :
1581 : /*
1582 : * mxactMemberComparator
1583 : * qsort comparison function for MultiXactMember
1584 : *
1585 : * We can't use wraparound comparison for XIDs because that does not respect
1586 : * the triangle inequality! Any old sort order will do.
1587 : */
1588 : static int
1589 3760 : mxactMemberComparator(const void *arg1, const void *arg2)
1590 : {
1591 3760 : MultiXactMember member1 = *(const MultiXactMember *) arg1;
1592 3760 : MultiXactMember member2 = *(const MultiXactMember *) arg2;
1593 :
1594 3760 : if (member1.xid > member2.xid)
1595 54 : return 1;
1596 3706 : if (member1.xid < member2.xid)
1597 3360 : return -1;
1598 346 : if (member1.status > member2.status)
1599 4 : return 1;
1600 342 : if (member1.status < member2.status)
1601 342 : return -1;
1602 0 : return 0;
1603 : }
1604 :
1605 : /*
1606 : * mXactCacheGetBySet
1607 : * returns a MultiXactId from the cache based on the set of
1608 : * TransactionIds that compose it, or InvalidMultiXactId if
1609 : * none matches.
1610 : *
1611 : * This is helpful, for example, if two transactions want to lock a huge
1612 : * table. By using the cache, the second will use the same MultiXactId
1613 : * for the majority of tuples, thus keeping MultiXactId usage low (saving
1614 : * both I/O and wraparound issues).
1615 : *
1616 : * NB: the passed members array will be sorted in-place.
1617 : */
1618 : static MultiXactId
1619 2260 : mXactCacheGetBySet(int nmembers, MultiXactMember *members)
1620 : {
1621 : dlist_iter iter;
1622 :
1623 : debug_elog3(DEBUG2, "CacheGet: looking for %s",
1624 : mxid_to_string(InvalidMultiXactId, nmembers, members));
1625 :
1626 : /* sort the array so comparison is easy */
1627 2260 : qsort(members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1628 :
1629 2742 : dclist_foreach(iter, &MXactCache)
1630 : {
1631 2182 : mXactCacheEnt *entry = dclist_container(mXactCacheEnt, node,
1632 : iter.cur);
1633 :
1634 2182 : if (entry->nmembers != nmembers)
1635 242 : continue;
1636 :
1637 : /*
1638 : * We assume the cache entries are sorted, and that the unused bits in
1639 : * "status" are zeroed.
1640 : */
1641 1940 : if (memcmp(members, entry->members, nmembers * sizeof(MultiXactMember)) == 0)
1642 : {
1643 : debug_elog3(DEBUG2, "CacheGet: found %u", entry->multi);
1644 1700 : dclist_move_head(&MXactCache, iter.cur);
1645 1700 : return entry->multi;
1646 : }
1647 : }
1648 :
1649 : debug_elog2(DEBUG2, "CacheGet: not found :-(");
1650 560 : return InvalidMultiXactId;
1651 : }
1652 :
1653 : /*
1654 : * mXactCacheGetById
1655 : * returns the composing MultiXactMember set from the cache for a
1656 : * given MultiXactId, if present.
1657 : *
1658 : * If successful, *xids is set to the address of a palloc'd copy of the
1659 : * MultiXactMember set. Return value is number of members, or -1 on failure.
1660 : */
1661 : static int
1662 5958 : mXactCacheGetById(MultiXactId multi, MultiXactMember **members)
1663 : {
1664 : dlist_iter iter;
1665 :
1666 : debug_elog3(DEBUG2, "CacheGet: looking for %u", multi);
1667 :
1668 6786 : dclist_foreach(iter, &MXactCache)
1669 : {
1670 5016 : mXactCacheEnt *entry = dclist_container(mXactCacheEnt, node,
1671 : iter.cur);
1672 :
1673 5016 : if (entry->multi == multi)
1674 : {
1675 : MultiXactMember *ptr;
1676 : Size size;
1677 :
1678 4188 : size = sizeof(MultiXactMember) * entry->nmembers;
1679 4188 : ptr = (MultiXactMember *) palloc(size);
1680 :
1681 4188 : memcpy(ptr, entry->members, size);
1682 :
1683 : debug_elog3(DEBUG2, "CacheGet: found %s",
1684 : mxid_to_string(multi,
1685 : entry->nmembers,
1686 : entry->members));
1687 :
1688 : /*
1689 : * Note we modify the list while not using a modifiable iterator.
1690 : * This is acceptable only because we exit the iteration
1691 : * immediately afterwards.
1692 : */
1693 4188 : dclist_move_head(&MXactCache, iter.cur);
1694 :
1695 4188 : *members = ptr;
1696 4188 : return entry->nmembers;
1697 : }
1698 : }
1699 :
1700 : debug_elog2(DEBUG2, "CacheGet: not found");
1701 1770 : return -1;
1702 : }
1703 :
1704 : /*
1705 : * mXactCachePut
1706 : * Add a new MultiXactId and its composing set into the local cache.
1707 : */
1708 : static void
1709 992 : mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
1710 : {
1711 : mXactCacheEnt *entry;
1712 :
1713 : debug_elog3(DEBUG2, "CachePut: storing %s",
1714 : mxid_to_string(multi, nmembers, members));
1715 :
1716 992 : if (MXactContext == NULL)
1717 : {
1718 : /* The cache only lives as long as the current transaction */
1719 : debug_elog2(DEBUG2, "CachePut: initializing memory context");
1720 696 : MXactContext = AllocSetContextCreate(TopTransactionContext,
1721 : "MultiXact cache context",
1722 : ALLOCSET_SMALL_SIZES);
1723 : }
1724 :
1725 : entry = (mXactCacheEnt *)
1726 992 : MemoryContextAlloc(MXactContext,
1727 992 : offsetof(mXactCacheEnt, members) +
1728 : nmembers * sizeof(MultiXactMember));
1729 :
1730 992 : entry->multi = multi;
1731 992 : entry->nmembers = nmembers;
1732 992 : memcpy(entry->members, members, nmembers * sizeof(MultiXactMember));
1733 :
1734 : /* mXactCacheGetBySet assumes the entries are sorted, so sort them */
1735 992 : qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1736 :
1737 992 : dclist_push_head(&MXactCache, &entry->node);
1738 992 : if (dclist_count(&MXactCache) > MAX_CACHE_ENTRIES)
1739 : {
1740 : dlist_node *node;
1741 :
1742 0 : node = dclist_tail_node(&MXactCache);
1743 0 : dclist_delete_from(&MXactCache, node);
1744 :
1745 0 : entry = dclist_container(mXactCacheEnt, node, node);
1746 : debug_elog3(DEBUG2, "CachePut: pruning cached multi %u",
1747 : entry->multi);
1748 :
1749 0 : pfree(entry);
1750 : }
1751 992 : }
1752 :
1753 : static char *
1754 0 : mxstatus_to_string(MultiXactStatus status)
1755 : {
1756 0 : switch (status)
1757 : {
1758 0 : case MultiXactStatusForKeyShare:
1759 0 : return "keysh";
1760 0 : case MultiXactStatusForShare:
1761 0 : return "sh";
1762 0 : case MultiXactStatusForNoKeyUpdate:
1763 0 : return "fornokeyupd";
1764 0 : case MultiXactStatusForUpdate:
1765 0 : return "forupd";
1766 0 : case MultiXactStatusNoKeyUpdate:
1767 0 : return "nokeyupd";
1768 0 : case MultiXactStatusUpdate:
1769 0 : return "upd";
1770 0 : default:
1771 0 : elog(ERROR, "unrecognized multixact status %d", status);
1772 : return "";
1773 : }
1774 : }
1775 :
1776 : char *
1777 0 : mxid_to_string(MultiXactId multi, int nmembers, MultiXactMember *members)
1778 : {
1779 : static char *str = NULL;
1780 : StringInfoData buf;
1781 : int i;
1782 :
1783 0 : if (str != NULL)
1784 0 : pfree(str);
1785 :
1786 0 : initStringInfo(&buf);
1787 :
1788 0 : appendStringInfo(&buf, "%u %d[%u (%s)", multi, nmembers, members[0].xid,
1789 : mxstatus_to_string(members[0].status));
1790 :
1791 0 : for (i = 1; i < nmembers; i++)
1792 0 : appendStringInfo(&buf, ", %u (%s)", members[i].xid,
1793 0 : mxstatus_to_string(members[i].status));
1794 :
1795 0 : appendStringInfoChar(&buf, ']');
1796 0 : str = MemoryContextStrdup(TopMemoryContext, buf.data);
1797 0 : pfree(buf.data);
1798 0 : return str;
1799 : }
1800 :
1801 : /*
1802 : * AtEOXact_MultiXact
1803 : * Handle transaction end for MultiXact
1804 : *
1805 : * This is called at top transaction commit or abort (we don't care which).
1806 : */
1807 : void
1808 1067740 : AtEOXact_MultiXact(void)
1809 : {
1810 : /*
1811 : * Reset our OldestMemberMXactId and OldestVisibleMXactId values, both of
1812 : * which should only be valid while within a transaction.
1813 : *
1814 : * We assume that storing a MultiXactId is atomic and so we need not take
1815 : * MultiXactGenLock to do this.
1816 : */
1817 1067740 : OldestMemberMXactId[MyProcNumber] = InvalidMultiXactId;
1818 1067740 : OldestVisibleMXactId[MyProcNumber] = InvalidMultiXactId;
1819 :
1820 : /*
1821 : * Discard the local MultiXactId cache. Since MXactContext was created as
1822 : * a child of TopTransactionContext, we needn't delete it explicitly.
1823 : */
1824 1067740 : MXactContext = NULL;
1825 1067740 : dclist_init(&MXactCache);
1826 1067740 : }
1827 :
1828 : /*
1829 : * AtPrepare_MultiXact
1830 : * Save multixact state at 2PC transaction prepare
1831 : *
1832 : * In this phase, we only store our OldestMemberMXactId value in the two-phase
1833 : * state file.
1834 : */
1835 : void
1836 564 : AtPrepare_MultiXact(void)
1837 : {
1838 564 : MultiXactId myOldestMember = OldestMemberMXactId[MyProcNumber];
1839 :
1840 564 : if (MultiXactIdIsValid(myOldestMember))
1841 94 : RegisterTwoPhaseRecord(TWOPHASE_RM_MULTIXACT_ID, 0,
1842 : &myOldestMember, sizeof(MultiXactId));
1843 564 : }
1844 :
1845 : /*
1846 : * PostPrepare_MultiXact
1847 : * Clean up after successful PREPARE TRANSACTION
1848 : */
1849 : void
1850 564 : PostPrepare_MultiXact(TransactionId xid)
1851 : {
1852 : MultiXactId myOldestMember;
1853 :
1854 : /*
1855 : * Transfer our OldestMemberMXactId value to the slot reserved for the
1856 : * prepared transaction.
1857 : */
1858 564 : myOldestMember = OldestMemberMXactId[MyProcNumber];
1859 564 : if (MultiXactIdIsValid(myOldestMember))
1860 : {
1861 94 : ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(xid, false);
1862 :
1863 : /*
1864 : * Even though storing MultiXactId is atomic, acquire lock to make
1865 : * sure others see both changes, not just the reset of the slot of the
1866 : * current backend. Using a volatile pointer might suffice, but this
1867 : * isn't a hot spot.
1868 : */
1869 94 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1870 :
1871 94 : OldestMemberMXactId[dummyProcNumber] = myOldestMember;
1872 94 : OldestMemberMXactId[MyProcNumber] = InvalidMultiXactId;
1873 :
1874 94 : LWLockRelease(MultiXactGenLock);
1875 : }
1876 :
1877 : /*
1878 : * We don't need to transfer OldestVisibleMXactId value, because the
1879 : * transaction is not going to be looking at any more multixacts once it's
1880 : * prepared.
1881 : *
1882 : * We assume that storing a MultiXactId is atomic and so we need not take
1883 : * MultiXactGenLock to do this.
1884 : */
1885 564 : OldestVisibleMXactId[MyProcNumber] = InvalidMultiXactId;
1886 :
1887 : /*
1888 : * Discard the local MultiXactId cache like in AtEOXact_MultiXact.
1889 : */
1890 564 : MXactContext = NULL;
1891 564 : dclist_init(&MXactCache);
1892 564 : }
1893 :
1894 : /*
1895 : * multixact_twophase_recover
1896 : * Recover the state of a prepared transaction at startup
1897 : */
1898 : void
1899 16 : multixact_twophase_recover(TransactionId xid, uint16 info,
1900 : void *recdata, uint32 len)
1901 : {
1902 16 : ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(xid, false);
1903 : MultiXactId oldestMember;
1904 :
1905 : /*
1906 : * Get the oldest member XID from the state file record, and set it in the
1907 : * OldestMemberMXactId slot reserved for this prepared transaction.
1908 : */
1909 : Assert(len == sizeof(MultiXactId));
1910 16 : oldestMember = *((MultiXactId *) recdata);
1911 :
1912 16 : OldestMemberMXactId[dummyProcNumber] = oldestMember;
1913 16 : }
1914 :
1915 : /*
1916 : * multixact_twophase_postcommit
1917 : * Similar to AtEOXact_MultiXact but for COMMIT PREPARED
1918 : */
1919 : void
1920 106 : multixact_twophase_postcommit(TransactionId xid, uint16 info,
1921 : void *recdata, uint32 len)
1922 : {
1923 106 : ProcNumber dummyProcNumber = TwoPhaseGetDummyProcNumber(xid, true);
1924 :
1925 : Assert(len == sizeof(MultiXactId));
1926 :
1927 106 : OldestMemberMXactId[dummyProcNumber] = InvalidMultiXactId;
1928 106 : }
1929 :
1930 : /*
1931 : * multixact_twophase_postabort
1932 : * This is actually just the same as the COMMIT case.
1933 : */
1934 : void
1935 44 : multixact_twophase_postabort(TransactionId xid, uint16 info,
1936 : void *recdata, uint32 len)
1937 : {
1938 44 : multixact_twophase_postcommit(xid, info, recdata, len);
1939 44 : }
1940 :
1941 : /*
1942 : * Initialization of shared memory for MultiXact. We use two SLRU areas,
1943 : * thus double memory. Also, reserve space for the shared MultiXactState
1944 : * struct and the per-backend MultiXactId arrays (two of those, too).
1945 : */
1946 : Size
1947 3898 : MultiXactShmemSize(void)
1948 : {
1949 : Size size;
1950 :
1951 : /* We need 2*MaxOldestSlot perBackendXactIds[] entries */
1952 : #define SHARED_MULTIXACT_STATE_SIZE \
1953 : add_size(offsetof(MultiXactStateData, perBackendXactIds), \
1954 : mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
1955 :
1956 3898 : size = SHARED_MULTIXACT_STATE_SIZE;
1957 3898 : size = add_size(size, SimpleLruShmemSize(multixact_offset_buffers, 0));
1958 3898 : size = add_size(size, SimpleLruShmemSize(multixact_member_buffers, 0));
1959 :
1960 3898 : return size;
1961 : }
1962 :
1963 : void
1964 2096 : MultiXactShmemInit(void)
1965 : {
1966 : bool found;
1967 :
1968 : debug_elog2(DEBUG2, "Shared Memory Init for MultiXact");
1969 :
1970 2096 : MultiXactOffsetCtl->PagePrecedes = MultiXactOffsetPagePrecedes;
1971 2096 : MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
1972 :
1973 2096 : SimpleLruInit(MultiXactOffsetCtl,
1974 : "multixact_offset", multixact_offset_buffers, 0,
1975 : "pg_multixact/offsets", LWTRANCHE_MULTIXACTOFFSET_BUFFER,
1976 : LWTRANCHE_MULTIXACTOFFSET_SLRU,
1977 : SYNC_HANDLER_MULTIXACT_OFFSET,
1978 : false);
1979 : SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
1980 2096 : SimpleLruInit(MultiXactMemberCtl,
1981 : "multixact_member", multixact_member_buffers, 0,
1982 : "pg_multixact/members", LWTRANCHE_MULTIXACTMEMBER_BUFFER,
1983 : LWTRANCHE_MULTIXACTMEMBER_SLRU,
1984 : SYNC_HANDLER_MULTIXACT_MEMBER,
1985 : false);
1986 : /* doesn't call SimpleLruTruncate() or meet criteria for unit tests */
1987 :
1988 : /* Initialize our shared state struct */
1989 2096 : MultiXactState = ShmemInitStruct("Shared MultiXact State",
1990 2096 : SHARED_MULTIXACT_STATE_SIZE,
1991 : &found);
1992 2096 : if (!IsUnderPostmaster)
1993 : {
1994 : Assert(!found);
1995 :
1996 : /* Make sure we zero out the per-backend state */
1997 2096 : MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
1998 2096 : ConditionVariableInit(&MultiXactState->nextoff_cv);
1999 : }
2000 : else
2001 : Assert(found);
2002 :
2003 : /*
2004 : * Set up array pointers.
2005 : */
2006 2096 : OldestMemberMXactId = MultiXactState->perBackendXactIds;
2007 2096 : OldestVisibleMXactId = OldestMemberMXactId + MaxOldestSlot;
2008 2096 : }
2009 :
2010 : /*
2011 : * GUC check_hook for multixact_offset_buffers
2012 : */
2013 : bool
2014 2164 : check_multixact_offset_buffers(int *newval, void **extra, GucSource source)
2015 : {
2016 2164 : return check_slru_buffers("multixact_offset_buffers", newval);
2017 : }
2018 :
2019 : /*
2020 : * GUC check_hook for multixact_member_buffers
2021 : */
2022 : bool
2023 2164 : check_multixact_member_buffers(int *newval, void **extra, GucSource source)
2024 : {
2025 2164 : return check_slru_buffers("multixact_member_buffers", newval);
2026 : }
2027 :
2028 : /*
2029 : * This func must be called ONCE on system install. It creates the initial
2030 : * MultiXact segments. (The MultiXacts directories are assumed to have been
2031 : * created by initdb, and MultiXactShmemInit must have been called already.)
2032 : */
2033 : void
2034 98 : BootStrapMultiXact(void)
2035 : {
2036 : int slotno;
2037 : LWLock *lock;
2038 :
2039 98 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, 0);
2040 98 : LWLockAcquire(lock, LW_EXCLUSIVE);
2041 :
2042 : /* Create and zero the first page of the offsets log */
2043 98 : slotno = ZeroMultiXactOffsetPage(0, false);
2044 :
2045 : /* Make sure it's written out */
2046 98 : SimpleLruWritePage(MultiXactOffsetCtl, slotno);
2047 : Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
2048 :
2049 98 : LWLockRelease(lock);
2050 :
2051 98 : lock = SimpleLruGetBankLock(MultiXactMemberCtl, 0);
2052 98 : LWLockAcquire(lock, LW_EXCLUSIVE);
2053 :
2054 : /* Create and zero the first page of the members log */
2055 98 : slotno = ZeroMultiXactMemberPage(0, false);
2056 :
2057 : /* Make sure it's written out */
2058 98 : SimpleLruWritePage(MultiXactMemberCtl, slotno);
2059 : Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
2060 :
2061 98 : LWLockRelease(lock);
2062 98 : }
2063 :
2064 : /*
2065 : * Initialize (or reinitialize) a page of MultiXactOffset to zeroes.
2066 : * If writeXlog is true, also emit an XLOG record saying we did this.
2067 : *
2068 : * The page is not actually written, just set up in shared memory.
2069 : * The slot number of the new page is returned.
2070 : *
2071 : * Control lock must be held at entry, and will be held at exit.
2072 : */
2073 : static int
2074 122 : ZeroMultiXactOffsetPage(int64 pageno, bool writeXlog)
2075 : {
2076 : int slotno;
2077 :
2078 122 : slotno = SimpleLruZeroPage(MultiXactOffsetCtl, pageno);
2079 :
2080 122 : if (writeXlog)
2081 22 : WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_OFF_PAGE);
2082 :
2083 122 : return slotno;
2084 : }
2085 :
2086 : /*
2087 : * Ditto, for MultiXactMember
2088 : */
2089 : static int
2090 122 : ZeroMultiXactMemberPage(int64 pageno, bool writeXlog)
2091 : {
2092 : int slotno;
2093 :
2094 122 : slotno = SimpleLruZeroPage(MultiXactMemberCtl, pageno);
2095 :
2096 122 : if (writeXlog)
2097 22 : WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_MEM_PAGE);
2098 :
2099 122 : return slotno;
2100 : }
2101 :
2102 : /*
2103 : * MaybeExtendOffsetSlru
2104 : * Extend the offsets SLRU area, if necessary
2105 : *
2106 : * After a binary upgrade from <= 9.2, the pg_multixact/offsets SLRU area might
2107 : * contain files that are shorter than necessary; this would occur if the old
2108 : * installation had used multixacts beyond the first page (files cannot be
2109 : * copied, because the on-disk representation is different). pg_upgrade would
2110 : * update pg_control to set the next offset value to be at that position, so
2111 : * that tuples marked as locked by such MultiXacts would be seen as visible
2112 : * without having to consult multixact. However, trying to create and use a
2113 : * new MultiXactId would result in an error because the page on which the new
2114 : * value would reside does not exist. This routine is in charge of creating
2115 : * such pages.
2116 : */
2117 : static void
2118 88 : MaybeExtendOffsetSlru(void)
2119 : {
2120 : int64 pageno;
2121 : LWLock *lock;
2122 :
2123 88 : pageno = MultiXactIdToOffsetPage(MultiXactState->nextMXact);
2124 88 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
2125 :
2126 88 : LWLockAcquire(lock, LW_EXCLUSIVE);
2127 :
2128 88 : if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
2129 : {
2130 : int slotno;
2131 :
2132 : /*
2133 : * Fortunately for us, SimpleLruWritePage is already prepared to deal
2134 : * with creating a new segment file even if the page we're writing is
2135 : * not the first in it, so this is enough.
2136 : */
2137 0 : slotno = ZeroMultiXactOffsetPage(pageno, false);
2138 0 : SimpleLruWritePage(MultiXactOffsetCtl, slotno);
2139 : }
2140 :
2141 88 : LWLockRelease(lock);
2142 88 : }
2143 :
2144 : /*
2145 : * This must be called ONCE during postmaster or standalone-backend startup.
2146 : *
2147 : * StartupXLOG has already established nextMXact/nextOffset by calling
2148 : * MultiXactSetNextMXact and/or MultiXactAdvanceNextMXact, and the oldestMulti
2149 : * info from pg_control and/or MultiXactAdvanceOldest, but we haven't yet
2150 : * replayed WAL.
2151 : */
2152 : void
2153 1820 : StartupMultiXact(void)
2154 : {
2155 1820 : MultiXactId multi = MultiXactState->nextMXact;
2156 1820 : MultiXactOffset offset = MultiXactState->nextOffset;
2157 : int64 pageno;
2158 :
2159 : /*
2160 : * Initialize offset's idea of the latest page number.
2161 : */
2162 1820 : pageno = MultiXactIdToOffsetPage(multi);
2163 1820 : pg_atomic_write_u64(&MultiXactOffsetCtl->shared->latest_page_number,
2164 : pageno);
2165 :
2166 : /*
2167 : * Initialize member's idea of the latest page number.
2168 : */
2169 1820 : pageno = MXOffsetToMemberPage(offset);
2170 1820 : pg_atomic_write_u64(&MultiXactMemberCtl->shared->latest_page_number,
2171 : pageno);
2172 1820 : }
2173 :
2174 : /*
2175 : * This must be called ONCE at the end of startup/recovery.
2176 : */
2177 : void
2178 1700 : TrimMultiXact(void)
2179 : {
2180 : MultiXactId nextMXact;
2181 : MultiXactOffset offset;
2182 : MultiXactId oldestMXact;
2183 : Oid oldestMXactDB;
2184 : int64 pageno;
2185 : int entryno;
2186 : int flagsoff;
2187 :
2188 1700 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2189 1700 : nextMXact = MultiXactState->nextMXact;
2190 1700 : offset = MultiXactState->nextOffset;
2191 1700 : oldestMXact = MultiXactState->oldestMultiXactId;
2192 1700 : oldestMXactDB = MultiXactState->oldestMultiXactDB;
2193 1700 : LWLockRelease(MultiXactGenLock);
2194 :
2195 : /* Clean up offsets state */
2196 :
2197 : /*
2198 : * (Re-)Initialize our idea of the latest page number for offsets.
2199 : */
2200 1700 : pageno = MultiXactIdToOffsetPage(nextMXact);
2201 1700 : pg_atomic_write_u64(&MultiXactOffsetCtl->shared->latest_page_number,
2202 : pageno);
2203 :
2204 : /*
2205 : * Zero out the remainder of the current offsets page. See notes in
2206 : * TrimCLOG() for background. Unlike CLOG, some WAL record covers every
2207 : * pg_multixact SLRU mutation. Since, also unlike CLOG, we ignore the WAL
2208 : * rule "write xlog before data," nextMXact successors may carry obsolete,
2209 : * nonzero offset values. Zero those so case 2 of GetMultiXactIdMembers()
2210 : * operates normally.
2211 : */
2212 1700 : entryno = MultiXactIdToOffsetEntry(nextMXact);
2213 1700 : if (entryno != 0)
2214 : {
2215 : int slotno;
2216 : MultiXactOffset *offptr;
2217 1698 : LWLock *lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
2218 :
2219 1698 : LWLockAcquire(lock, LW_EXCLUSIVE);
2220 1698 : slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, nextMXact);
2221 1698 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2222 1698 : offptr += entryno;
2223 :
2224 1698 : MemSet(offptr, 0, BLCKSZ - (entryno * sizeof(MultiXactOffset)));
2225 :
2226 1698 : MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
2227 1698 : LWLockRelease(lock);
2228 : }
2229 :
2230 : /*
2231 : * And the same for members.
2232 : *
2233 : * (Re-)Initialize our idea of the latest page number for members.
2234 : */
2235 1700 : pageno = MXOffsetToMemberPage(offset);
2236 1700 : pg_atomic_write_u64(&MultiXactMemberCtl->shared->latest_page_number,
2237 : pageno);
2238 :
2239 : /*
2240 : * Zero out the remainder of the current members page. See notes in
2241 : * TrimCLOG() for motivation.
2242 : */
2243 1700 : flagsoff = MXOffsetToFlagsOffset(offset);
2244 1700 : if (flagsoff != 0)
2245 : {
2246 : int slotno;
2247 : TransactionId *xidptr;
2248 : int memberoff;
2249 22 : LWLock *lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
2250 :
2251 22 : LWLockAcquire(lock, LW_EXCLUSIVE);
2252 22 : memberoff = MXOffsetToMemberOffset(offset);
2253 22 : slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, offset);
2254 22 : xidptr = (TransactionId *)
2255 22 : (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
2256 :
2257 22 : MemSet(xidptr, 0, BLCKSZ - memberoff);
2258 :
2259 : /*
2260 : * Note: we don't need to zero out the flag bits in the remaining
2261 : * members of the current group, because they are always reset before
2262 : * writing.
2263 : */
2264 :
2265 22 : MultiXactMemberCtl->shared->page_dirty[slotno] = true;
2266 22 : LWLockRelease(lock);
2267 : }
2268 :
2269 : /* signal that we're officially up */
2270 1700 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2271 1700 : MultiXactState->finishedStartup = true;
2272 1700 : LWLockRelease(MultiXactGenLock);
2273 :
2274 : /* Now compute how far away the next members wraparound is. */
2275 1700 : SetMultiXactIdLimit(oldestMXact, oldestMXactDB, true);
2276 1700 : }
2277 :
2278 : /*
2279 : * Get the MultiXact data to save in a checkpoint record
2280 : */
2281 : void
2282 2950 : MultiXactGetCheckptMulti(bool is_shutdown,
2283 : MultiXactId *nextMulti,
2284 : MultiXactOffset *nextMultiOffset,
2285 : MultiXactId *oldestMulti,
2286 : Oid *oldestMultiDB)
2287 : {
2288 2950 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2289 2950 : *nextMulti = MultiXactState->nextMXact;
2290 2950 : *nextMultiOffset = MultiXactState->nextOffset;
2291 2950 : *oldestMulti = MultiXactState->oldestMultiXactId;
2292 2950 : *oldestMultiDB = MultiXactState->oldestMultiXactDB;
2293 2950 : LWLockRelease(MultiXactGenLock);
2294 :
2295 : debug_elog6(DEBUG2,
2296 : "MultiXact: checkpoint is nextMulti %u, nextOffset %u, oldestMulti %u in DB %u",
2297 : *nextMulti, *nextMultiOffset, *oldestMulti, *oldestMultiDB);
2298 2950 : }
2299 :
2300 : /*
2301 : * Perform a checkpoint --- either during shutdown, or on-the-fly
2302 : */
2303 : void
2304 3342 : CheckPointMultiXact(void)
2305 : {
2306 : TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(true);
2307 :
2308 : /*
2309 : * Write dirty MultiXact pages to disk. This may result in sync requests
2310 : * queued for later handling by ProcessSyncRequests(), as part of the
2311 : * checkpoint.
2312 : */
2313 3342 : SimpleLruWriteAll(MultiXactOffsetCtl, true);
2314 3342 : SimpleLruWriteAll(MultiXactMemberCtl, true);
2315 :
2316 : TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(true);
2317 3342 : }
2318 :
2319 : /*
2320 : * Set the next-to-be-assigned MultiXactId and offset
2321 : *
2322 : * This is used when we can determine the correct next ID/offset exactly
2323 : * from a checkpoint record. Although this is only called during bootstrap
2324 : * and XLog replay, we take the lock in case any hot-standby backends are
2325 : * examining the values.
2326 : */
2327 : void
2328 1988 : MultiXactSetNextMXact(MultiXactId nextMulti,
2329 : MultiXactOffset nextMultiOffset)
2330 : {
2331 : debug_elog4(DEBUG2, "MultiXact: setting next multi to %u offset %u",
2332 : nextMulti, nextMultiOffset);
2333 1988 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2334 1988 : MultiXactState->nextMXact = nextMulti;
2335 1988 : MultiXactState->nextOffset = nextMultiOffset;
2336 1988 : LWLockRelease(MultiXactGenLock);
2337 :
2338 : /*
2339 : * During a binary upgrade, make sure that the offsets SLRU is large
2340 : * enough to contain the next value that would be created.
2341 : *
2342 : * We need to do this pretty early during the first startup in binary
2343 : * upgrade mode: before StartupMultiXact() in fact, because this routine
2344 : * is called even before that by StartupXLOG(). And we can't do it
2345 : * earlier than at this point, because during that first call of this
2346 : * routine we determine the MultiXactState->nextMXact value that
2347 : * MaybeExtendOffsetSlru needs.
2348 : */
2349 1988 : if (IsBinaryUpgrade)
2350 88 : MaybeExtendOffsetSlru();
2351 1988 : }
2352 :
2353 : /*
2354 : * Determine the last safe MultiXactId to allocate given the currently oldest
2355 : * datminmxid (ie, the oldest MultiXactId that might exist in any database
2356 : * of our cluster), and the OID of the (or a) database with that value.
2357 : *
2358 : * is_startup is true when we are just starting the cluster, false when we
2359 : * are updating state in a running cluster. This only affects log messages.
2360 : */
2361 : void
2362 5952 : SetMultiXactIdLimit(MultiXactId oldest_datminmxid, Oid oldest_datoid,
2363 : bool is_startup)
2364 : {
2365 : MultiXactId multiVacLimit;
2366 : MultiXactId multiWarnLimit;
2367 : MultiXactId multiStopLimit;
2368 : MultiXactId multiWrapLimit;
2369 : MultiXactId curMulti;
2370 : bool needs_offset_vacuum;
2371 :
2372 : Assert(MultiXactIdIsValid(oldest_datminmxid));
2373 :
2374 : /*
2375 : * We pretend that a wrap will happen halfway through the multixact ID
2376 : * space, but that's not really true, because multixacts wrap differently
2377 : * from transaction IDs. Note that, separately from any concern about
2378 : * multixact IDs wrapping, we must ensure that multixact members do not
2379 : * wrap. Limits for that are set in SetOffsetVacuumLimit, not here.
2380 : */
2381 5952 : multiWrapLimit = oldest_datminmxid + (MaxMultiXactId >> 1);
2382 5952 : if (multiWrapLimit < FirstMultiXactId)
2383 0 : multiWrapLimit += FirstMultiXactId;
2384 :
2385 : /*
2386 : * We'll refuse to continue assigning MultiXactIds once we get within 3M
2387 : * multi of data loss. See SetTransactionIdLimit.
2388 : */
2389 5952 : multiStopLimit = multiWrapLimit - 3000000;
2390 5952 : if (multiStopLimit < FirstMultiXactId)
2391 0 : multiStopLimit -= FirstMultiXactId;
2392 :
2393 : /*
2394 : * We'll start complaining loudly when we get within 40M multis of data
2395 : * loss. This is kind of arbitrary, but if you let your gas gauge get
2396 : * down to 2% of full, would you be looking for the next gas station? We
2397 : * need to be fairly liberal about this number because there are lots of
2398 : * scenarios where most transactions are done by automatic clients that
2399 : * won't pay attention to warnings. (No, we're not gonna make this
2400 : * configurable. If you know enough to configure it, you know enough to
2401 : * not get in this kind of trouble in the first place.)
2402 : */
2403 5952 : multiWarnLimit = multiWrapLimit - 40000000;
2404 5952 : if (multiWarnLimit < FirstMultiXactId)
2405 0 : multiWarnLimit -= FirstMultiXactId;
2406 :
2407 : /*
2408 : * We'll start trying to force autovacuums when oldest_datminmxid gets to
2409 : * be more than autovacuum_multixact_freeze_max_age mxids old.
2410 : *
2411 : * Note: autovacuum_multixact_freeze_max_age is a PGC_POSTMASTER parameter
2412 : * so that we don't have to worry about dealing with on-the-fly changes in
2413 : * its value. See SetTransactionIdLimit.
2414 : */
2415 5952 : multiVacLimit = oldest_datminmxid + autovacuum_multixact_freeze_max_age;
2416 5952 : if (multiVacLimit < FirstMultiXactId)
2417 0 : multiVacLimit += FirstMultiXactId;
2418 :
2419 : /* Grab lock for just long enough to set the new limit values */
2420 5952 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2421 5952 : MultiXactState->oldestMultiXactId = oldest_datminmxid;
2422 5952 : MultiXactState->oldestMultiXactDB = oldest_datoid;
2423 5952 : MultiXactState->multiVacLimit = multiVacLimit;
2424 5952 : MultiXactState->multiWarnLimit = multiWarnLimit;
2425 5952 : MultiXactState->multiStopLimit = multiStopLimit;
2426 5952 : MultiXactState->multiWrapLimit = multiWrapLimit;
2427 5952 : curMulti = MultiXactState->nextMXact;
2428 5952 : LWLockRelease(MultiXactGenLock);
2429 :
2430 : /* Log the info */
2431 5952 : ereport(DEBUG1,
2432 : (errmsg_internal("MultiXactId wrap limit is %u, limited by database with OID %u",
2433 : multiWrapLimit, oldest_datoid)));
2434 :
2435 : /*
2436 : * Computing the actual limits is only possible once the data directory is
2437 : * in a consistent state. There's no need to compute the limits while
2438 : * still replaying WAL - no decisions about new multis are made even
2439 : * though multixact creations might be replayed. So we'll only do further
2440 : * checks after TrimMultiXact() has been called.
2441 : */
2442 5952 : if (!MultiXactState->finishedStartup)
2443 1920 : return;
2444 :
2445 : Assert(!InRecovery);
2446 :
2447 : /* Set limits for offset vacuum. */
2448 4032 : needs_offset_vacuum = SetOffsetVacuumLimit(is_startup);
2449 :
2450 : /*
2451 : * If past the autovacuum force point, immediately signal an autovac
2452 : * request. The reason for this is that autovac only processes one
2453 : * database per invocation. Once it's finished cleaning up the oldest
2454 : * database, it'll call here, and we'll signal the postmaster to start
2455 : * another iteration immediately if there are still any old databases.
2456 : */
2457 4032 : if ((MultiXactIdPrecedes(multiVacLimit, curMulti) ||
2458 0 : needs_offset_vacuum) && IsUnderPostmaster)
2459 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
2460 :
2461 : /* Give an immediate warning if past the wrap warn point */
2462 4032 : if (MultiXactIdPrecedes(multiWarnLimit, curMulti))
2463 : {
2464 : char *oldest_datname;
2465 :
2466 : /*
2467 : * We can be called when not inside a transaction, for example during
2468 : * StartupXLOG(). In such a case we cannot do database access, so we
2469 : * must just report the oldest DB's OID.
2470 : *
2471 : * Note: it's also possible that get_database_name fails and returns
2472 : * NULL, for example because the database just got dropped. We'll
2473 : * still warn, even though the warning might now be unnecessary.
2474 : */
2475 0 : if (IsTransactionState())
2476 0 : oldest_datname = get_database_name(oldest_datoid);
2477 : else
2478 0 : oldest_datname = NULL;
2479 :
2480 0 : if (oldest_datname)
2481 0 : ereport(WARNING,
2482 : (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
2483 : "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
2484 : multiWrapLimit - curMulti,
2485 : oldest_datname,
2486 : multiWrapLimit - curMulti),
2487 : errhint("To avoid MultiXactId assignment failures, execute a database-wide VACUUM in that database.\n"
2488 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
2489 : else
2490 0 : ereport(WARNING,
2491 : (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
2492 : "database with OID %u must be vacuumed before %u more MultiXactIds are used",
2493 : multiWrapLimit - curMulti,
2494 : oldest_datoid,
2495 : multiWrapLimit - curMulti),
2496 : errhint("To avoid MultiXactId assignment failures, execute a database-wide VACUUM in that database.\n"
2497 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
2498 : }
2499 : }
2500 :
2501 : /*
2502 : * Ensure the next-to-be-assigned MultiXactId is at least minMulti,
2503 : * and similarly nextOffset is at least minMultiOffset.
2504 : *
2505 : * This is used when we can determine minimum safe values from an XLog
2506 : * record (either an on-line checkpoint or an mxact creation log entry).
2507 : * Although this is only called during XLog replay, we take the lock in case
2508 : * any hot-standby backends are examining the values.
2509 : */
2510 : void
2511 1354 : MultiXactAdvanceNextMXact(MultiXactId minMulti,
2512 : MultiXactOffset minMultiOffset)
2513 : {
2514 1354 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2515 1354 : if (MultiXactIdPrecedes(MultiXactState->nextMXact, minMulti))
2516 : {
2517 : debug_elog3(DEBUG2, "MultiXact: setting next multi to %u", minMulti);
2518 4 : MultiXactState->nextMXact = minMulti;
2519 : }
2520 1354 : if (MultiXactOffsetPrecedes(MultiXactState->nextOffset, minMultiOffset))
2521 : {
2522 : debug_elog3(DEBUG2, "MultiXact: setting next offset to %u",
2523 : minMultiOffset);
2524 4 : MultiXactState->nextOffset = minMultiOffset;
2525 : }
2526 1354 : LWLockRelease(MultiXactGenLock);
2527 1354 : }
2528 :
2529 : /*
2530 : * Update our oldestMultiXactId value, but only if it's more recent than what
2531 : * we had.
2532 : *
2533 : * This may only be called during WAL replay.
2534 : */
2535 : void
2536 1418 : MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB)
2537 : {
2538 : Assert(InRecovery);
2539 :
2540 1418 : if (MultiXactIdPrecedes(MultiXactState->oldestMultiXactId, oldestMulti))
2541 0 : SetMultiXactIdLimit(oldestMulti, oldestMultiDB, false);
2542 1418 : }
2543 :
2544 : /*
2545 : * Make sure that MultiXactOffset has room for a newly-allocated MultiXactId.
2546 : *
2547 : * NB: this is called while holding MultiXactGenLock. We want it to be very
2548 : * fast most of the time; even when it's not so fast, no actual I/O need
2549 : * happen unless we're forced to write out a dirty log or xlog page to make
2550 : * room in shared memory.
2551 : */
2552 : static void
2553 560 : ExtendMultiXactOffset(MultiXactId multi)
2554 : {
2555 : int64 pageno;
2556 : LWLock *lock;
2557 :
2558 : /*
2559 : * No work except at first MultiXactId of a page. But beware: just after
2560 : * wraparound, the first MultiXactId of page zero is FirstMultiXactId.
2561 : */
2562 560 : if (MultiXactIdToOffsetEntry(multi) != 0 &&
2563 : multi != FirstMultiXactId)
2564 538 : return;
2565 :
2566 22 : pageno = MultiXactIdToOffsetPage(multi);
2567 22 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
2568 :
2569 22 : LWLockAcquire(lock, LW_EXCLUSIVE);
2570 :
2571 : /* Zero the page and make an XLOG entry about it */
2572 22 : ZeroMultiXactOffsetPage(pageno, true);
2573 :
2574 22 : LWLockRelease(lock);
2575 : }
2576 :
2577 : /*
2578 : * Make sure that MultiXactMember has room for the members of a newly-
2579 : * allocated MultiXactId.
2580 : *
2581 : * Like the above routine, this is called while holding MultiXactGenLock;
2582 : * same comments apply.
2583 : */
2584 : static void
2585 560 : ExtendMultiXactMember(MultiXactOffset offset, int nmembers)
2586 : {
2587 : /*
2588 : * It's possible that the members span more than one page of the members
2589 : * file, so we loop to ensure we consider each page. The coding is not
2590 : * optimal if the members span several pages, but that seems unusual
2591 : * enough to not worry much about.
2592 : */
2593 1120 : while (nmembers > 0)
2594 : {
2595 : int flagsoff;
2596 : int flagsbit;
2597 : uint32 difference;
2598 :
2599 : /*
2600 : * Only zero when at first entry of a page.
2601 : */
2602 560 : flagsoff = MXOffsetToFlagsOffset(offset);
2603 560 : flagsbit = MXOffsetToFlagsBitShift(offset);
2604 560 : if (flagsoff == 0 && flagsbit == 0)
2605 : {
2606 : int64 pageno;
2607 : LWLock *lock;
2608 :
2609 22 : pageno = MXOffsetToMemberPage(offset);
2610 22 : lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
2611 :
2612 22 : LWLockAcquire(lock, LW_EXCLUSIVE);
2613 :
2614 : /* Zero the page and make an XLOG entry about it */
2615 22 : ZeroMultiXactMemberPage(pageno, true);
2616 :
2617 22 : LWLockRelease(lock);
2618 : }
2619 :
2620 : /*
2621 : * Compute the number of items till end of current page. Careful: if
2622 : * addition of unsigned ints wraps around, we're at the last page of
2623 : * the last segment; since that page holds a different number of items
2624 : * than other pages, we need to do it differently.
2625 : */
2626 560 : if (offset + MAX_MEMBERS_IN_LAST_MEMBERS_PAGE < offset)
2627 : {
2628 : /*
2629 : * This is the last page of the last segment; we can compute the
2630 : * number of items left to allocate in it without modulo
2631 : * arithmetic.
2632 : */
2633 0 : difference = MaxMultiXactOffset - offset + 1;
2634 : }
2635 : else
2636 560 : difference = MULTIXACT_MEMBERS_PER_PAGE - offset % MULTIXACT_MEMBERS_PER_PAGE;
2637 :
2638 : /*
2639 : * Advance to next page, taking care to properly handle the wraparound
2640 : * case. OK if nmembers goes negative.
2641 : */
2642 560 : nmembers -= difference;
2643 560 : offset += difference;
2644 : }
2645 560 : }
2646 :
2647 : /*
2648 : * GetOldestMultiXactId
2649 : *
2650 : * Return the oldest MultiXactId that's still possibly still seen as live by
2651 : * any running transaction. Older ones might still exist on disk, but they no
2652 : * longer have any running member transaction.
2653 : *
2654 : * It's not safe to truncate MultiXact SLRU segments on the value returned by
2655 : * this function; however, it can be set as the new relminmxid for any table
2656 : * that VACUUM knows has no remaining MXIDs < the same value. It is only safe
2657 : * to truncate SLRUs when no table can possibly still have a referencing MXID.
2658 : */
2659 : MultiXactId
2660 305376 : GetOldestMultiXactId(void)
2661 : {
2662 : MultiXactId oldestMXact;
2663 : MultiXactId nextMXact;
2664 : int i;
2665 :
2666 : /*
2667 : * This is the oldest valid value among all the OldestMemberMXactId[] and
2668 : * OldestVisibleMXactId[] entries, or nextMXact if none are valid.
2669 : */
2670 305376 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2671 :
2672 : /*
2673 : * We have to beware of the possibility that nextMXact is in the
2674 : * wrapped-around state. We don't fix the counter itself here, but we
2675 : * must be sure to use a valid value in our calculation.
2676 : */
2677 305376 : nextMXact = MultiXactState->nextMXact;
2678 305376 : if (nextMXact < FirstMultiXactId)
2679 0 : nextMXact = FirstMultiXactId;
2680 :
2681 305376 : oldestMXact = nextMXact;
2682 38174154 : for (i = 0; i < MaxOldestSlot; i++)
2683 : {
2684 : MultiXactId thisoldest;
2685 :
2686 37868778 : thisoldest = OldestMemberMXactId[i];
2687 37936722 : if (MultiXactIdIsValid(thisoldest) &&
2688 67944 : MultiXactIdPrecedes(thisoldest, oldestMXact))
2689 22 : oldestMXact = thisoldest;
2690 37868778 : thisoldest = OldestVisibleMXactId[i];
2691 37869028 : if (MultiXactIdIsValid(thisoldest) &&
2692 250 : MultiXactIdPrecedes(thisoldest, oldestMXact))
2693 4 : oldestMXact = thisoldest;
2694 : }
2695 :
2696 305376 : LWLockRelease(MultiXactGenLock);
2697 :
2698 305376 : return oldestMXact;
2699 : }
2700 :
2701 : /*
2702 : * Determine how aggressively we need to vacuum in order to prevent member
2703 : * wraparound.
2704 : *
2705 : * To do so determine what's the oldest member offset and install the limit
2706 : * info in MultiXactState, where it can be used to prevent overrun of old data
2707 : * in the members SLRU area.
2708 : *
2709 : * The return value is true if emergency autovacuum is required and false
2710 : * otherwise.
2711 : */
2712 : static bool
2713 4032 : SetOffsetVacuumLimit(bool is_startup)
2714 : {
2715 : MultiXactId oldestMultiXactId;
2716 : MultiXactId nextMXact;
2717 4032 : MultiXactOffset oldestOffset = 0; /* placate compiler */
2718 : MultiXactOffset prevOldestOffset;
2719 : MultiXactOffset nextOffset;
2720 4032 : bool oldestOffsetKnown = false;
2721 : bool prevOldestOffsetKnown;
2722 4032 : MultiXactOffset offsetStopLimit = 0;
2723 : MultiXactOffset prevOffsetStopLimit;
2724 :
2725 : /*
2726 : * NB: Have to prevent concurrent truncation, we might otherwise try to
2727 : * lookup an oldestMulti that's concurrently getting truncated away.
2728 : */
2729 4032 : LWLockAcquire(MultiXactTruncationLock, LW_SHARED);
2730 :
2731 : /* Read relevant fields from shared memory. */
2732 4032 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2733 4032 : oldestMultiXactId = MultiXactState->oldestMultiXactId;
2734 4032 : nextMXact = MultiXactState->nextMXact;
2735 4032 : nextOffset = MultiXactState->nextOffset;
2736 4032 : prevOldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2737 4032 : prevOldestOffset = MultiXactState->oldestOffset;
2738 4032 : prevOffsetStopLimit = MultiXactState->offsetStopLimit;
2739 : Assert(MultiXactState->finishedStartup);
2740 4032 : LWLockRelease(MultiXactGenLock);
2741 :
2742 : /*
2743 : * Determine the offset of the oldest multixact. Normally, we can read
2744 : * the offset from the multixact itself, but there's an important special
2745 : * case: if there are no multixacts in existence at all, oldestMXact
2746 : * obviously can't point to one. It will instead point to the multixact
2747 : * ID that will be assigned the next time one is needed.
2748 : */
2749 4032 : if (oldestMultiXactId == nextMXact)
2750 : {
2751 : /*
2752 : * When the next multixact gets created, it will be stored at the next
2753 : * offset.
2754 : */
2755 4004 : oldestOffset = nextOffset;
2756 4004 : oldestOffsetKnown = true;
2757 : }
2758 : else
2759 : {
2760 : /*
2761 : * Figure out where the oldest existing multixact's offsets are
2762 : * stored. Due to bugs in early release of PostgreSQL 9.3.X and 9.4.X,
2763 : * the supposedly-earliest multixact might not really exist. We are
2764 : * careful not to fail in that case.
2765 : */
2766 : oldestOffsetKnown =
2767 28 : find_multixact_start(oldestMultiXactId, &oldestOffset);
2768 :
2769 28 : if (oldestOffsetKnown)
2770 28 : ereport(DEBUG1,
2771 : (errmsg_internal("oldest MultiXactId member is at offset %u",
2772 : oldestOffset)));
2773 : else
2774 0 : ereport(LOG,
2775 : (errmsg("MultiXact member wraparound protections are disabled because oldest checkpointed MultiXact %u does not exist on disk",
2776 : oldestMultiXactId)));
2777 : }
2778 :
2779 4032 : LWLockRelease(MultiXactTruncationLock);
2780 :
2781 : /*
2782 : * If we can, compute limits (and install them MultiXactState) to prevent
2783 : * overrun of old data in the members SLRU area. We can only do so if the
2784 : * oldest offset is known though.
2785 : */
2786 4032 : if (oldestOffsetKnown)
2787 : {
2788 : /* move back to start of the corresponding segment */
2789 4032 : offsetStopLimit = oldestOffset - (oldestOffset %
2790 : (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT));
2791 :
2792 : /* always leave one segment before the wraparound point */
2793 4032 : offsetStopLimit -= (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT);
2794 :
2795 4032 : if (!prevOldestOffsetKnown && !is_startup)
2796 0 : ereport(LOG,
2797 : (errmsg("MultiXact member wraparound protections are now enabled")));
2798 :
2799 4032 : ereport(DEBUG1,
2800 : (errmsg_internal("MultiXact member stop limit is now %u based on MultiXact %u",
2801 : offsetStopLimit, oldestMultiXactId)));
2802 : }
2803 0 : else if (prevOldestOffsetKnown)
2804 : {
2805 : /*
2806 : * If we failed to get the oldest offset this time, but we have a
2807 : * value from a previous pass through this function, use the old
2808 : * values rather than automatically forcing an emergency autovacuum
2809 : * cycle again.
2810 : */
2811 0 : oldestOffset = prevOldestOffset;
2812 0 : oldestOffsetKnown = true;
2813 0 : offsetStopLimit = prevOffsetStopLimit;
2814 : }
2815 :
2816 : /* Install the computed values */
2817 4032 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2818 4032 : MultiXactState->oldestOffset = oldestOffset;
2819 4032 : MultiXactState->oldestOffsetKnown = oldestOffsetKnown;
2820 4032 : MultiXactState->offsetStopLimit = offsetStopLimit;
2821 4032 : LWLockRelease(MultiXactGenLock);
2822 :
2823 : /*
2824 : * Do we need an emergency autovacuum? If we're not sure, assume yes.
2825 : */
2826 8064 : return !oldestOffsetKnown ||
2827 4032 : (nextOffset - oldestOffset > MULTIXACT_MEMBER_SAFE_THRESHOLD);
2828 : }
2829 :
2830 : /*
2831 : * Return whether adding "distance" to "start" would move past "boundary".
2832 : *
2833 : * We use this to determine whether the addition is "wrapping around" the
2834 : * boundary point, hence the name. The reason we don't want to use the regular
2835 : * 2^31-modulo arithmetic here is that we want to be able to use the whole of
2836 : * the 2^32-1 space here, allowing for more multixacts than would fit
2837 : * otherwise.
2838 : */
2839 : static bool
2840 1120 : MultiXactOffsetWouldWrap(MultiXactOffset boundary, MultiXactOffset start,
2841 : uint32 distance)
2842 : {
2843 : MultiXactOffset finish;
2844 :
2845 : /*
2846 : * Note that offset number 0 is not used (see GetMultiXactIdMembers), so
2847 : * if the addition wraps around the UINT_MAX boundary, skip that value.
2848 : */
2849 1120 : finish = start + distance;
2850 1120 : if (finish < start)
2851 0 : finish++;
2852 :
2853 : /*-----------------------------------------------------------------------
2854 : * When the boundary is numerically greater than the starting point, any
2855 : * value numerically between the two is not wrapped:
2856 : *
2857 : * <----S----B---->
2858 : * [---) = F wrapped past B (and UINT_MAX)
2859 : * [---) = F not wrapped
2860 : * [----] = F wrapped past B
2861 : *
2862 : * When the boundary is numerically less than the starting point (i.e. the
2863 : * UINT_MAX wraparound occurs somewhere in between) then all values in
2864 : * between are wrapped:
2865 : *
2866 : * <----B----S---->
2867 : * [---) = F not wrapped past B (but wrapped past UINT_MAX)
2868 : * [---) = F wrapped past B (and UINT_MAX)
2869 : * [----] = F not wrapped
2870 : *-----------------------------------------------------------------------
2871 : */
2872 1120 : if (start < boundary)
2873 1120 : return finish >= boundary || finish < start;
2874 : else
2875 0 : return finish >= boundary && finish < start;
2876 : }
2877 :
2878 : /*
2879 : * Find the starting offset of the given MultiXactId.
2880 : *
2881 : * Returns false if the file containing the multi does not exist on disk.
2882 : * Otherwise, returns true and sets *result to the starting member offset.
2883 : *
2884 : * This function does not prevent concurrent truncation, so if that's
2885 : * required, the caller has to protect against that.
2886 : */
2887 : static bool
2888 28 : find_multixact_start(MultiXactId multi, MultiXactOffset *result)
2889 : {
2890 : MultiXactOffset offset;
2891 : int64 pageno;
2892 : int entryno;
2893 : int slotno;
2894 : MultiXactOffset *offptr;
2895 :
2896 : Assert(MultiXactState->finishedStartup);
2897 :
2898 28 : pageno = MultiXactIdToOffsetPage(multi);
2899 28 : entryno = MultiXactIdToOffsetEntry(multi);
2900 :
2901 : /*
2902 : * Write out dirty data, so PhysicalPageExists can work correctly.
2903 : */
2904 28 : SimpleLruWriteAll(MultiXactOffsetCtl, true);
2905 28 : SimpleLruWriteAll(MultiXactMemberCtl, true);
2906 :
2907 28 : if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
2908 0 : return false;
2909 :
2910 : /* lock is acquired by SimpleLruReadPage_ReadOnly */
2911 28 : slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
2912 28 : offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2913 28 : offptr += entryno;
2914 28 : offset = *offptr;
2915 28 : LWLockRelease(SimpleLruGetBankLock(MultiXactOffsetCtl, pageno));
2916 :
2917 28 : *result = offset;
2918 28 : return true;
2919 : }
2920 :
2921 : /*
2922 : * Determine how many multixacts, and how many multixact members, currently
2923 : * exist. Return false if unable to determine.
2924 : */
2925 : static bool
2926 239888 : ReadMultiXactCounts(uint32 *multixacts, MultiXactOffset *members)
2927 : {
2928 : MultiXactOffset nextOffset;
2929 : MultiXactOffset oldestOffset;
2930 : MultiXactId oldestMultiXactId;
2931 : MultiXactId nextMultiXactId;
2932 : bool oldestOffsetKnown;
2933 :
2934 239888 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
2935 239888 : nextOffset = MultiXactState->nextOffset;
2936 239888 : oldestMultiXactId = MultiXactState->oldestMultiXactId;
2937 239888 : nextMultiXactId = MultiXactState->nextMXact;
2938 239888 : oldestOffset = MultiXactState->oldestOffset;
2939 239888 : oldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2940 239888 : LWLockRelease(MultiXactGenLock);
2941 :
2942 239888 : if (!oldestOffsetKnown)
2943 0 : return false;
2944 :
2945 239888 : *members = nextOffset - oldestOffset;
2946 239888 : *multixacts = nextMultiXactId - oldestMultiXactId;
2947 239888 : return true;
2948 : }
2949 :
2950 : /*
2951 : * Multixact members can be removed once the multixacts that refer to them
2952 : * are older than every datminmxid. autovacuum_multixact_freeze_max_age and
2953 : * vacuum_multixact_freeze_table_age work together to make sure we never have
2954 : * too many multixacts; we hope that, at least under normal circumstances,
2955 : * this will also be sufficient to keep us from using too many offsets.
2956 : * However, if the average multixact has many members, we might exhaust the
2957 : * members space while still using few enough members that these limits fail
2958 : * to trigger relminmxid advancement by VACUUM. At that point, we'd have no
2959 : * choice but to start failing multixact-creating operations with an error.
2960 : *
2961 : * To prevent that, if more than a threshold portion of the members space is
2962 : * used, we effectively reduce autovacuum_multixact_freeze_max_age and
2963 : * to a value just less than the number of multixacts in use. We hope that
2964 : * this will quickly trigger autovacuuming on the table or tables with the
2965 : * oldest relminmxid, thus allowing datminmxid values to advance and removing
2966 : * some members.
2967 : *
2968 : * As the fraction of the member space currently in use grows, we become
2969 : * more aggressive in clamping this value. That not only causes autovacuum
2970 : * to ramp up, but also makes any manual vacuums the user issues more
2971 : * aggressive. This happens because vacuum_get_cutoffs() will clamp the
2972 : * freeze table and the minimum freeze age cutoffs based on the effective
2973 : * autovacuum_multixact_freeze_max_age this function returns. In the worst
2974 : * case, we'll claim the freeze_max_age to zero, and every vacuum of any
2975 : * table will freeze every multixact.
2976 : */
2977 : int
2978 239888 : MultiXactMemberFreezeThreshold(void)
2979 : {
2980 : MultiXactOffset members;
2981 : uint32 multixacts;
2982 : uint32 victim_multixacts;
2983 : double fraction;
2984 : int result;
2985 :
2986 : /* If we can't determine member space utilization, assume the worst. */
2987 239888 : if (!ReadMultiXactCounts(&multixacts, &members))
2988 0 : return 0;
2989 :
2990 : /* If member space utilization is low, no special action is required. */
2991 239888 : if (members <= MULTIXACT_MEMBER_SAFE_THRESHOLD)
2992 239888 : return autovacuum_multixact_freeze_max_age;
2993 :
2994 : /*
2995 : * Compute a target for relminmxid advancement. The number of multixacts
2996 : * we try to eliminate from the system is based on how far we are past
2997 : * MULTIXACT_MEMBER_SAFE_THRESHOLD.
2998 : */
2999 0 : fraction = (double) (members - MULTIXACT_MEMBER_SAFE_THRESHOLD) /
3000 : (MULTIXACT_MEMBER_DANGER_THRESHOLD - MULTIXACT_MEMBER_SAFE_THRESHOLD);
3001 0 : victim_multixacts = multixacts * fraction;
3002 :
3003 : /* fraction could be > 1.0, but lowest possible freeze age is zero */
3004 0 : if (victim_multixacts > multixacts)
3005 0 : return 0;
3006 0 : result = multixacts - victim_multixacts;
3007 :
3008 : /*
3009 : * Clamp to autovacuum_multixact_freeze_max_age, so that we never make
3010 : * autovacuum less aggressive than it would otherwise be.
3011 : */
3012 0 : return Min(result, autovacuum_multixact_freeze_max_age);
3013 : }
3014 :
3015 : typedef struct mxtruncinfo
3016 : {
3017 : int64 earliestExistingPage;
3018 : } mxtruncinfo;
3019 :
3020 : /*
3021 : * SlruScanDirectory callback
3022 : * This callback determines the earliest existing page number.
3023 : */
3024 : static bool
3025 0 : SlruScanDirCbFindEarliest(SlruCtl ctl, char *filename, int64 segpage, void *data)
3026 : {
3027 0 : mxtruncinfo *trunc = (mxtruncinfo *) data;
3028 :
3029 0 : if (trunc->earliestExistingPage == -1 ||
3030 0 : ctl->PagePrecedes(segpage, trunc->earliestExistingPage))
3031 : {
3032 0 : trunc->earliestExistingPage = segpage;
3033 : }
3034 :
3035 0 : return false; /* keep going */
3036 : }
3037 :
3038 :
3039 : /*
3040 : * Delete members segments [oldest, newOldest)
3041 : *
3042 : * The members SLRU can, in contrast to the offsets one, be filled to almost
3043 : * the full range at once. This means SimpleLruTruncate() can't trivially be
3044 : * used - instead the to-be-deleted range is computed using the offsets
3045 : * SLRU. C.f. TruncateMultiXact().
3046 : */
3047 : static void
3048 0 : PerformMembersTruncation(MultiXactOffset oldestOffset, MultiXactOffset newOldestOffset)
3049 : {
3050 0 : const int64 maxsegment = MXOffsetToMemberSegment(MaxMultiXactOffset);
3051 0 : int64 startsegment = MXOffsetToMemberSegment(oldestOffset);
3052 0 : int64 endsegment = MXOffsetToMemberSegment(newOldestOffset);
3053 0 : int64 segment = startsegment;
3054 :
3055 : /*
3056 : * Delete all the segments but the last one. The last segment can still
3057 : * contain, possibly partially, valid data.
3058 : */
3059 0 : while (segment != endsegment)
3060 : {
3061 0 : elog(DEBUG2, "truncating multixact members segment %" PRIx64,
3062 : segment);
3063 0 : SlruDeleteSegment(MultiXactMemberCtl, segment);
3064 :
3065 : /* move to next segment, handling wraparound correctly */
3066 0 : if (segment == maxsegment)
3067 0 : segment = 0;
3068 : else
3069 0 : segment += 1;
3070 : }
3071 0 : }
3072 :
3073 : /*
3074 : * Delete offsets segments [oldest, newOldest)
3075 : */
3076 : static void
3077 0 : PerformOffsetsTruncation(MultiXactId oldestMulti, MultiXactId newOldestMulti)
3078 : {
3079 : /*
3080 : * We step back one multixact to avoid passing a cutoff page that hasn't
3081 : * been created yet in the rare case that oldestMulti would be the first
3082 : * item on a page and oldestMulti == nextMulti. In that case, if we
3083 : * didn't subtract one, we'd trigger SimpleLruTruncate's wraparound
3084 : * detection.
3085 : */
3086 0 : SimpleLruTruncate(MultiXactOffsetCtl,
3087 : MultiXactIdToOffsetPage(PreviousMultiXactId(newOldestMulti)));
3088 0 : }
3089 :
3090 : /*
3091 : * Remove all MultiXactOffset and MultiXactMember segments before the oldest
3092 : * ones still of interest.
3093 : *
3094 : * This is only called on a primary as part of vacuum (via
3095 : * vac_truncate_clog()). During recovery truncation is done by replaying
3096 : * truncation WAL records logged here.
3097 : *
3098 : * newOldestMulti is the oldest currently required multixact, newOldestMultiDB
3099 : * is one of the databases preventing newOldestMulti from increasing.
3100 : */
3101 : void
3102 2332 : TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
3103 : {
3104 : MultiXactId oldestMulti;
3105 : MultiXactId nextMulti;
3106 : MultiXactOffset newOldestOffset;
3107 : MultiXactOffset oldestOffset;
3108 : MultiXactOffset nextOffset;
3109 : mxtruncinfo trunc;
3110 : MultiXactId earliest;
3111 :
3112 : Assert(!RecoveryInProgress());
3113 : Assert(MultiXactState->finishedStartup);
3114 :
3115 : /*
3116 : * We can only allow one truncation to happen at once. Otherwise parts of
3117 : * members might vanish while we're doing lookups or similar. There's no
3118 : * need to have an interlock with creating new multis or such, since those
3119 : * are constrained by the limits (which only grow, never shrink).
3120 : */
3121 2332 : LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
3122 :
3123 2332 : LWLockAcquire(MultiXactGenLock, LW_SHARED);
3124 2332 : nextMulti = MultiXactState->nextMXact;
3125 2332 : nextOffset = MultiXactState->nextOffset;
3126 2332 : oldestMulti = MultiXactState->oldestMultiXactId;
3127 2332 : LWLockRelease(MultiXactGenLock);
3128 : Assert(MultiXactIdIsValid(oldestMulti));
3129 :
3130 : /*
3131 : * Make sure to only attempt truncation if there's values to truncate
3132 : * away. In normal processing values shouldn't go backwards, but there's
3133 : * some corner cases (due to bugs) where that's possible.
3134 : */
3135 2332 : if (MultiXactIdPrecedesOrEquals(newOldestMulti, oldestMulti))
3136 : {
3137 2332 : LWLockRelease(MultiXactTruncationLock);
3138 2332 : return;
3139 : }
3140 :
3141 : /*
3142 : * Note we can't just plow ahead with the truncation; it's possible that
3143 : * there are no segments to truncate, which is a problem because we are
3144 : * going to attempt to read the offsets page to determine where to
3145 : * truncate the members SLRU. So we first scan the directory to determine
3146 : * the earliest offsets page number that we can read without error.
3147 : *
3148 : * When nextMXact is less than one segment away from multiWrapLimit,
3149 : * SlruScanDirCbFindEarliest can find some early segment other than the
3150 : * actual earliest. (MultiXactOffsetPagePrecedes(EARLIEST, LATEST)
3151 : * returns false, because not all pairs of entries have the same answer.)
3152 : * That can also arise when an earlier truncation attempt failed unlink()
3153 : * or returned early from this function. The only consequence is
3154 : * returning early, which wastes space that we could have liberated.
3155 : *
3156 : * NB: It's also possible that the page that oldestMulti is on has already
3157 : * been truncated away, and we crashed before updating oldestMulti.
3158 : */
3159 0 : trunc.earliestExistingPage = -1;
3160 0 : SlruScanDirectory(MultiXactOffsetCtl, SlruScanDirCbFindEarliest, &trunc);
3161 0 : earliest = trunc.earliestExistingPage * MULTIXACT_OFFSETS_PER_PAGE;
3162 0 : if (earliest < FirstMultiXactId)
3163 0 : earliest = FirstMultiXactId;
3164 :
3165 : /* If there's nothing to remove, we can bail out early. */
3166 0 : if (MultiXactIdPrecedes(oldestMulti, earliest))
3167 : {
3168 0 : LWLockRelease(MultiXactTruncationLock);
3169 0 : return;
3170 : }
3171 :
3172 : /*
3173 : * First, compute the safe truncation point for MultiXactMember. This is
3174 : * the starting offset of the oldest multixact.
3175 : *
3176 : * Hopefully, find_multixact_start will always work here, because we've
3177 : * already checked that it doesn't precede the earliest MultiXact on disk.
3178 : * But if it fails, don't truncate anything, and log a message.
3179 : */
3180 0 : if (oldestMulti == nextMulti)
3181 : {
3182 : /* there are NO MultiXacts */
3183 0 : oldestOffset = nextOffset;
3184 : }
3185 0 : else if (!find_multixact_start(oldestMulti, &oldestOffset))
3186 : {
3187 0 : ereport(LOG,
3188 : (errmsg("oldest MultiXact %u not found, earliest MultiXact %u, skipping truncation",
3189 : oldestMulti, earliest)));
3190 0 : LWLockRelease(MultiXactTruncationLock);
3191 0 : return;
3192 : }
3193 :
3194 : /*
3195 : * Secondly compute up to where to truncate. Lookup the corresponding
3196 : * member offset for newOldestMulti for that.
3197 : */
3198 0 : if (newOldestMulti == nextMulti)
3199 : {
3200 : /* there are NO MultiXacts */
3201 0 : newOldestOffset = nextOffset;
3202 : }
3203 0 : else if (!find_multixact_start(newOldestMulti, &newOldestOffset))
3204 : {
3205 0 : ereport(LOG,
3206 : (errmsg("cannot truncate up to MultiXact %u because it does not exist on disk, skipping truncation",
3207 : newOldestMulti)));
3208 0 : LWLockRelease(MultiXactTruncationLock);
3209 0 : return;
3210 : }
3211 :
3212 0 : elog(DEBUG1, "performing multixact truncation: "
3213 : "offsets [%u, %u), offsets segments [%" PRIx64 ", %" PRIx64 "), "
3214 : "members [%u, %u), members segments [%" PRIx64 ", %" PRIx64 ")",
3215 : oldestMulti, newOldestMulti,
3216 : MultiXactIdToOffsetSegment(oldestMulti),
3217 : MultiXactIdToOffsetSegment(newOldestMulti),
3218 : oldestOffset, newOldestOffset,
3219 : MXOffsetToMemberSegment(oldestOffset),
3220 : MXOffsetToMemberSegment(newOldestOffset));
3221 :
3222 : /*
3223 : * Do truncation, and the WAL logging of the truncation, in a critical
3224 : * section. That way offsets/members cannot get out of sync anymore, i.e.
3225 : * once consistent the newOldestMulti will always exist in members, even
3226 : * if we crashed in the wrong moment.
3227 : */
3228 0 : START_CRIT_SECTION();
3229 :
3230 : /*
3231 : * Prevent checkpoints from being scheduled concurrently. This is critical
3232 : * because otherwise a truncation record might not be replayed after a
3233 : * crash/basebackup, even though the state of the data directory would
3234 : * require it.
3235 : */
3236 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
3237 0 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
3238 :
3239 : /* WAL log truncation */
3240 0 : WriteMTruncateXlogRec(newOldestMultiDB,
3241 : oldestMulti, newOldestMulti,
3242 : oldestOffset, newOldestOffset);
3243 :
3244 : /*
3245 : * Update in-memory limits before performing the truncation, while inside
3246 : * the critical section: Have to do it before truncation, to prevent
3247 : * concurrent lookups of those values. Has to be inside the critical
3248 : * section as otherwise a future call to this function would error out,
3249 : * while looking up the oldest member in offsets, if our caller crashes
3250 : * before updating the limits.
3251 : */
3252 0 : LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
3253 0 : MultiXactState->oldestMultiXactId = newOldestMulti;
3254 0 : MultiXactState->oldestMultiXactDB = newOldestMultiDB;
3255 0 : LWLockRelease(MultiXactGenLock);
3256 :
3257 : /* First truncate members */
3258 0 : PerformMembersTruncation(oldestOffset, newOldestOffset);
3259 :
3260 : /* Then offsets */
3261 0 : PerformOffsetsTruncation(oldestMulti, newOldestMulti);
3262 :
3263 0 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
3264 :
3265 0 : END_CRIT_SECTION();
3266 0 : LWLockRelease(MultiXactTruncationLock);
3267 : }
3268 :
3269 : /*
3270 : * Decide whether a MultiXactOffset page number is "older" for truncation
3271 : * purposes. Analogous to CLOGPagePrecedes().
3272 : *
3273 : * Offsetting the values is optional, because MultiXactIdPrecedes() has
3274 : * translational symmetry.
3275 : */
3276 : static bool
3277 0 : MultiXactOffsetPagePrecedes(int64 page1, int64 page2)
3278 : {
3279 : MultiXactId multi1;
3280 : MultiXactId multi2;
3281 :
3282 0 : multi1 = ((MultiXactId) page1) * MULTIXACT_OFFSETS_PER_PAGE;
3283 0 : multi1 += FirstMultiXactId + 1;
3284 0 : multi2 = ((MultiXactId) page2) * MULTIXACT_OFFSETS_PER_PAGE;
3285 0 : multi2 += FirstMultiXactId + 1;
3286 :
3287 0 : return (MultiXactIdPrecedes(multi1, multi2) &&
3288 0 : MultiXactIdPrecedes(multi1,
3289 : multi2 + MULTIXACT_OFFSETS_PER_PAGE - 1));
3290 : }
3291 :
3292 : /*
3293 : * Decide whether a MultiXactMember page number is "older" for truncation
3294 : * purposes. There is no "invalid offset number" so use the numbers verbatim.
3295 : */
3296 : static bool
3297 0 : MultiXactMemberPagePrecedes(int64 page1, int64 page2)
3298 : {
3299 : MultiXactOffset offset1;
3300 : MultiXactOffset offset2;
3301 :
3302 0 : offset1 = ((MultiXactOffset) page1) * MULTIXACT_MEMBERS_PER_PAGE;
3303 0 : offset2 = ((MultiXactOffset) page2) * MULTIXACT_MEMBERS_PER_PAGE;
3304 :
3305 0 : return (MultiXactOffsetPrecedes(offset1, offset2) &&
3306 0 : MultiXactOffsetPrecedes(offset1,
3307 : offset2 + MULTIXACT_MEMBERS_PER_PAGE - 1));
3308 : }
3309 :
3310 : /*
3311 : * Decide which of two MultiXactIds is earlier.
3312 : *
3313 : * XXX do we need to do something special for InvalidMultiXactId?
3314 : * (Doesn't look like it.)
3315 : */
3316 : bool
3317 2509420 : MultiXactIdPrecedes(MultiXactId multi1, MultiXactId multi2)
3318 : {
3319 2509420 : int32 diff = (int32) (multi1 - multi2);
3320 :
3321 2509420 : return (diff < 0);
3322 : }
3323 :
3324 : /*
3325 : * MultiXactIdPrecedesOrEquals -- is multi1 logically <= multi2?
3326 : *
3327 : * XXX do we need to do something special for InvalidMultiXactId?
3328 : * (Doesn't look like it.)
3329 : */
3330 : bool
3331 13370 : MultiXactIdPrecedesOrEquals(MultiXactId multi1, MultiXactId multi2)
3332 : {
3333 13370 : int32 diff = (int32) (multi1 - multi2);
3334 :
3335 13370 : return (diff <= 0);
3336 : }
3337 :
3338 :
3339 : /*
3340 : * Decide which of two offsets is earlier.
3341 : */
3342 : static bool
3343 1354 : MultiXactOffsetPrecedes(MultiXactOffset offset1, MultiXactOffset offset2)
3344 : {
3345 1354 : int32 diff = (int32) (offset1 - offset2);
3346 :
3347 1354 : return (diff < 0);
3348 : }
3349 :
3350 : /*
3351 : * Write an xlog record reflecting the zeroing of either a MEMBERs or
3352 : * OFFSETs page (info shows which)
3353 : */
3354 : static void
3355 44 : WriteMZeroPageXlogRec(int64 pageno, uint8 info)
3356 : {
3357 44 : XLogBeginInsert();
3358 44 : XLogRegisterData(&pageno, sizeof(pageno));
3359 44 : (void) XLogInsert(RM_MULTIXACT_ID, info);
3360 44 : }
3361 :
3362 : /*
3363 : * Write a TRUNCATE xlog record
3364 : *
3365 : * We must flush the xlog record to disk before returning --- see notes in
3366 : * TruncateCLOG().
3367 : */
3368 : static void
3369 0 : WriteMTruncateXlogRec(Oid oldestMultiDB,
3370 : MultiXactId startTruncOff, MultiXactId endTruncOff,
3371 : MultiXactOffset startTruncMemb, MultiXactOffset endTruncMemb)
3372 : {
3373 : XLogRecPtr recptr;
3374 : xl_multixact_truncate xlrec;
3375 :
3376 0 : xlrec.oldestMultiDB = oldestMultiDB;
3377 :
3378 0 : xlrec.startTruncOff = startTruncOff;
3379 0 : xlrec.endTruncOff = endTruncOff;
3380 :
3381 0 : xlrec.startTruncMemb = startTruncMemb;
3382 0 : xlrec.endTruncMemb = endTruncMemb;
3383 :
3384 0 : XLogBeginInsert();
3385 0 : XLogRegisterData(&xlrec, SizeOfMultiXactTruncate);
3386 0 : recptr = XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_TRUNCATE_ID);
3387 0 : XLogFlush(recptr);
3388 0 : }
3389 :
3390 : /*
3391 : * MULTIXACT resource manager's routines
3392 : */
3393 : void
3394 8 : multixact_redo(XLogReaderState *record)
3395 : {
3396 8 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
3397 :
3398 : /* Backup blocks are not used in multixact records */
3399 : Assert(!XLogRecHasAnyBlockRefs(record));
3400 :
3401 8 : if (info == XLOG_MULTIXACT_ZERO_OFF_PAGE)
3402 : {
3403 : int64 pageno;
3404 : int slotno;
3405 : LWLock *lock;
3406 :
3407 2 : memcpy(&pageno, XLogRecGetData(record), sizeof(pageno));
3408 :
3409 2 : lock = SimpleLruGetBankLock(MultiXactOffsetCtl, pageno);
3410 2 : LWLockAcquire(lock, LW_EXCLUSIVE);
3411 :
3412 2 : slotno = ZeroMultiXactOffsetPage(pageno, false);
3413 2 : SimpleLruWritePage(MultiXactOffsetCtl, slotno);
3414 : Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
3415 :
3416 2 : LWLockRelease(lock);
3417 : }
3418 6 : else if (info == XLOG_MULTIXACT_ZERO_MEM_PAGE)
3419 : {
3420 : int64 pageno;
3421 : int slotno;
3422 : LWLock *lock;
3423 :
3424 2 : memcpy(&pageno, XLogRecGetData(record), sizeof(pageno));
3425 :
3426 2 : lock = SimpleLruGetBankLock(MultiXactMemberCtl, pageno);
3427 2 : LWLockAcquire(lock, LW_EXCLUSIVE);
3428 :
3429 2 : slotno = ZeroMultiXactMemberPage(pageno, false);
3430 2 : SimpleLruWritePage(MultiXactMemberCtl, slotno);
3431 : Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
3432 :
3433 2 : LWLockRelease(lock);
3434 : }
3435 4 : else if (info == XLOG_MULTIXACT_CREATE_ID)
3436 : {
3437 4 : xl_multixact_create *xlrec =
3438 4 : (xl_multixact_create *) XLogRecGetData(record);
3439 : TransactionId max_xid;
3440 : int i;
3441 :
3442 : /* Store the data back into the SLRU files */
3443 4 : RecordNewMultiXact(xlrec->mid, xlrec->moff, xlrec->nmembers,
3444 4 : xlrec->members);
3445 :
3446 : /* Make sure nextMXact/nextOffset are beyond what this record has */
3447 4 : MultiXactAdvanceNextMXact(xlrec->mid + 1,
3448 4 : xlrec->moff + xlrec->nmembers);
3449 :
3450 : /*
3451 : * Make sure nextXid is beyond any XID mentioned in the record. This
3452 : * should be unnecessary, since any XID found here ought to have other
3453 : * evidence in the XLOG, but let's be safe.
3454 : */
3455 4 : max_xid = XLogRecGetXid(record);
3456 12 : for (i = 0; i < xlrec->nmembers; i++)
3457 : {
3458 8 : if (TransactionIdPrecedes(max_xid, xlrec->members[i].xid))
3459 0 : max_xid = xlrec->members[i].xid;
3460 : }
3461 :
3462 4 : AdvanceNextFullTransactionIdPastXid(max_xid);
3463 : }
3464 0 : else if (info == XLOG_MULTIXACT_TRUNCATE_ID)
3465 : {
3466 : xl_multixact_truncate xlrec;
3467 : int64 pageno;
3468 :
3469 0 : memcpy(&xlrec, XLogRecGetData(record),
3470 : SizeOfMultiXactTruncate);
3471 :
3472 0 : elog(DEBUG1, "replaying multixact truncation: "
3473 : "offsets [%u, %u), offsets segments [%" PRIx64 ", %" PRIx64 "), "
3474 : "members [%u, %u), members segments [%" PRIx64 ", %" PRIx64 ")",
3475 : xlrec.startTruncOff, xlrec.endTruncOff,
3476 : MultiXactIdToOffsetSegment(xlrec.startTruncOff),
3477 : MultiXactIdToOffsetSegment(xlrec.endTruncOff),
3478 : xlrec.startTruncMemb, xlrec.endTruncMemb,
3479 : MXOffsetToMemberSegment(xlrec.startTruncMemb),
3480 : MXOffsetToMemberSegment(xlrec.endTruncMemb));
3481 :
3482 : /* should not be required, but more than cheap enough */
3483 0 : LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
3484 :
3485 : /*
3486 : * Advance the horizon values, so they're current at the end of
3487 : * recovery.
3488 : */
3489 0 : SetMultiXactIdLimit(xlrec.endTruncOff, xlrec.oldestMultiDB, false);
3490 :
3491 0 : PerformMembersTruncation(xlrec.startTruncMemb, xlrec.endTruncMemb);
3492 :
3493 : /*
3494 : * During XLOG replay, latest_page_number isn't necessarily set up
3495 : * yet; insert a suitable value to bypass the sanity test in
3496 : * SimpleLruTruncate.
3497 : */
3498 0 : pageno = MultiXactIdToOffsetPage(xlrec.endTruncOff);
3499 0 : pg_atomic_write_u64(&MultiXactOffsetCtl->shared->latest_page_number,
3500 : pageno);
3501 0 : PerformOffsetsTruncation(xlrec.startTruncOff, xlrec.endTruncOff);
3502 :
3503 0 : LWLockRelease(MultiXactTruncationLock);
3504 : }
3505 : else
3506 0 : elog(PANIC, "multixact_redo: unknown op code %u", info);
3507 8 : }
3508 :
3509 : Datum
3510 0 : pg_get_multixact_members(PG_FUNCTION_ARGS)
3511 : {
3512 : typedef struct
3513 : {
3514 : MultiXactMember *members;
3515 : int nmembers;
3516 : int iter;
3517 : } mxact;
3518 0 : MultiXactId mxid = PG_GETARG_TRANSACTIONID(0);
3519 : mxact *multi;
3520 : FuncCallContext *funccxt;
3521 :
3522 0 : if (mxid < FirstMultiXactId)
3523 0 : ereport(ERROR,
3524 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3525 : errmsg("invalid MultiXactId: %u", mxid)));
3526 :
3527 0 : if (SRF_IS_FIRSTCALL())
3528 : {
3529 : MemoryContext oldcxt;
3530 : TupleDesc tupdesc;
3531 :
3532 0 : funccxt = SRF_FIRSTCALL_INIT();
3533 0 : oldcxt = MemoryContextSwitchTo(funccxt->multi_call_memory_ctx);
3534 :
3535 0 : multi = palloc(sizeof(mxact));
3536 : /* no need to allow for old values here */
3537 0 : multi->nmembers = GetMultiXactIdMembers(mxid, &multi->members, false,
3538 : false);
3539 0 : multi->iter = 0;
3540 :
3541 0 : if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
3542 0 : elog(ERROR, "return type must be a row type");
3543 0 : funccxt->tuple_desc = tupdesc;
3544 0 : funccxt->attinmeta = TupleDescGetAttInMetadata(tupdesc);
3545 0 : funccxt->user_fctx = multi;
3546 :
3547 0 : MemoryContextSwitchTo(oldcxt);
3548 : }
3549 :
3550 0 : funccxt = SRF_PERCALL_SETUP();
3551 0 : multi = (mxact *) funccxt->user_fctx;
3552 :
3553 0 : while (multi->iter < multi->nmembers)
3554 : {
3555 : HeapTuple tuple;
3556 : char *values[2];
3557 :
3558 0 : values[0] = psprintf("%u", multi->members[multi->iter].xid);
3559 0 : values[1] = mxstatus_to_string(multi->members[multi->iter].status);
3560 :
3561 0 : tuple = BuildTupleFromCStrings(funccxt->attinmeta, values);
3562 :
3563 0 : multi->iter++;
3564 0 : pfree(values[0]);
3565 0 : SRF_RETURN_NEXT(funccxt, HeapTupleGetDatum(tuple));
3566 : }
3567 :
3568 0 : SRF_RETURN_DONE(funccxt);
3569 : }
3570 :
3571 : /*
3572 : * Entrypoint for sync.c to sync offsets files.
3573 : */
3574 : int
3575 0 : multixactoffsetssyncfiletag(const FileTag *ftag, char *path)
3576 : {
3577 0 : return SlruSyncFileTag(MultiXactOffsetCtl, ftag, path);
3578 : }
3579 :
3580 : /*
3581 : * Entrypoint for sync.c to sync members files.
3582 : */
3583 : int
3584 0 : multixactmemberssyncfiletag(const FileTag *ftag, char *path)
3585 : {
3586 0 : return SlruSyncFileTag(MultiXactMemberCtl, ftag, path);
3587 : }
|
