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) 1996-2019, 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 "catalog/pg_type.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 "postmaster/autovacuum.h"
86#include "storage/lmgr.h"
87#include "storage/pmsignal.h"
88#include "storage/proc.h"
89#include "storage/procarray.h"
90#include "utils/builtins.h"
91#include "utils/memutils.h"
92#include "utils/snapmgr.h"
93
94
95/*
96 * Defines for MultiXactOffset page sizes. A page is the same BLCKSZ as is
97 * used everywhere else in Postgres.
98 *
99 * Note: because MultiXactOffsets are 32 bits and wrap around at 0xFFFFFFFF,
100 * MultiXact page numbering also wraps around at
101 * 0xFFFFFFFF/MULTIXACT_OFFSETS_PER_PAGE, and segment numbering at
102 * 0xFFFFFFFF/MULTIXACT_OFFSETS_PER_PAGE/SLRU_PAGES_PER_SEGMENT. We need
103 * take no explicit notice of that fact in this module, except when comparing
104 * segment and page numbers in TruncateMultiXact (see
105 * MultiXactOffsetPagePrecedes).
106 */
107
108/* We need four bytes per offset */
109#define MULTIXACT_OFFSETS_PER_PAGE (BLCKSZ / sizeof(MultiXactOffset))
110
111#define MultiXactIdToOffsetPage(xid) \
112 ((xid) / (MultiXactOffset) MULTIXACT_OFFSETS_PER_PAGE)
113#define MultiXactIdToOffsetEntry(xid) \
114 ((xid) % (MultiXactOffset) MULTIXACT_OFFSETS_PER_PAGE)
115#define MultiXactIdToOffsetSegment(xid) (MultiXactIdToOffsetPage(xid) / SLRU_PAGES_PER_SEGMENT)
116
117/*
118 * The situation for members is a bit more complex: we store one byte of
119 * additional flag bits for each TransactionId. To do this without getting
120 * into alignment issues, we store four bytes of flags, and then the
121 * corresponding 4 Xids. Each such 5-word (20-byte) set we call a "group", and
122 * are stored as a whole in pages. Thus, with 8kB BLCKSZ, we keep 409 groups
123 * per page. This wastes 12 bytes per page, but that's OK -- simplicity (and
124 * performance) trumps space efficiency here.
125 *
126 * Note that the "offset" macros work with byte offset, not array indexes, so
127 * arithmetic must be done using "char *" pointers.
128 */
129/* We need eight bits per xact, so one xact fits in a byte */
130#define MXACT_MEMBER_BITS_PER_XACT 8
131#define MXACT_MEMBER_FLAGS_PER_BYTE 1
132#define MXACT_MEMBER_XACT_BITMASK ((1 << MXACT_MEMBER_BITS_PER_XACT) - 1)
133
134/* how many full bytes of flags are there in a group? */
135#define MULTIXACT_FLAGBYTES_PER_GROUP 4
136#define MULTIXACT_MEMBERS_PER_MEMBERGROUP \
137 (MULTIXACT_FLAGBYTES_PER_GROUP * MXACT_MEMBER_FLAGS_PER_BYTE)
138/* size in bytes of a complete group */
139#define MULTIXACT_MEMBERGROUP_SIZE \
140 (sizeof(TransactionId) * MULTIXACT_MEMBERS_PER_MEMBERGROUP + MULTIXACT_FLAGBYTES_PER_GROUP)
141#define MULTIXACT_MEMBERGROUPS_PER_PAGE (BLCKSZ / MULTIXACT_MEMBERGROUP_SIZE)
142#define MULTIXACT_MEMBERS_PER_PAGE \
143 (MULTIXACT_MEMBERGROUPS_PER_PAGE * MULTIXACT_MEMBERS_PER_MEMBERGROUP)
144
145/*
146 * Because the number of items per page is not a divisor of the last item
147 * number (member 0xFFFFFFFF), the last segment does not use the maximum number
148 * of pages, and moreover the last used page therein does not use the same
149 * number of items as previous pages. (Another way to say it is that the
150 * 0xFFFFFFFF member is somewhere in the middle of the last page, so the page
151 * has some empty space after that item.)
152 *
153 * This constant is the number of members in the last page of the last segment.
154 */
155#define MAX_MEMBERS_IN_LAST_MEMBERS_PAGE \
156 ((uint32) ((0xFFFFFFFF % MULTIXACT_MEMBERS_PER_PAGE) + 1))
157
158/* page in which a member is to be found */
159#define MXOffsetToMemberPage(xid) ((xid) / (TransactionId) MULTIXACT_MEMBERS_PER_PAGE)
160#define MXOffsetToMemberSegment(xid) (MXOffsetToMemberPage(xid) / SLRU_PAGES_PER_SEGMENT)
161
162/* Location (byte offset within page) of flag word for a given member */
163#define MXOffsetToFlagsOffset(xid) \
164 ((((xid) / (TransactionId) MULTIXACT_MEMBERS_PER_MEMBERGROUP) % \
165 (TransactionId) MULTIXACT_MEMBERGROUPS_PER_PAGE) * \
166 (TransactionId) MULTIXACT_MEMBERGROUP_SIZE)
167#define MXOffsetToFlagsBitShift(xid) \
168 (((xid) % (TransactionId) MULTIXACT_MEMBERS_PER_MEMBERGROUP) * \
169 MXACT_MEMBER_BITS_PER_XACT)
170
171/* Location (byte offset within page) of TransactionId of given member */
172#define MXOffsetToMemberOffset(xid) \
173 (MXOffsetToFlagsOffset(xid) + MULTIXACT_FLAGBYTES_PER_GROUP + \
174 ((xid) % MULTIXACT_MEMBERS_PER_MEMBERGROUP) * sizeof(TransactionId))
175
176/* Multixact members wraparound thresholds. */
177#define MULTIXACT_MEMBER_SAFE_THRESHOLD (MaxMultiXactOffset / 2)
178#define MULTIXACT_MEMBER_DANGER_THRESHOLD \
179 (MaxMultiXactOffset - MaxMultiXactOffset / 4)
180
181#define PreviousMultiXactId(xid) \
182 ((xid) == FirstMultiXactId ? MaxMultiXactId : (xid) - 1)
183
184/*
185 * Links to shared-memory data structures for MultiXact control
186 */
187static SlruCtlData MultiXactOffsetCtlData;
188static SlruCtlData MultiXactMemberCtlData;
189
190#define MultiXactOffsetCtl (&MultiXactOffsetCtlData)
191#define MultiXactMemberCtl (&MultiXactMemberCtlData)
192
193/*
194 * MultiXact state shared across all backends. All this state is protected
195 * by MultiXactGenLock. (We also use MultiXactOffsetControlLock and
196 * MultiXactMemberControlLock to guard accesses to the two sets of SLRU
197 * buffers. For concurrency's sake, we avoid holding more than one of these
198 * locks at a time.)
199 */
200typedef struct MultiXactStateData
201{
202 /* next-to-be-assigned MultiXactId */
203 MultiXactId nextMXact;
204
205 /* next-to-be-assigned offset */
206 MultiXactOffset nextOffset;
207
208 /* Have we completed multixact startup? */
209 bool finishedStartup;
210
211 /*
212 * Oldest multixact that is still potentially referenced by a relation.
213 * Anything older than this should not be consulted. These values are
214 * updated by vacuum.
215 */
216 MultiXactId oldestMultiXactId;
217 Oid oldestMultiXactDB;
218
219 /*
220 * Oldest multixact offset that is potentially referenced by a multixact
221 * referenced by a relation. We don't always know this value, so there's
222 * a flag here to indicate whether or not we currently do.
223 */
224 MultiXactOffset oldestOffset;
225 bool oldestOffsetKnown;
226
227 /* support for anti-wraparound measures */
228 MultiXactId multiVacLimit;
229 MultiXactId multiWarnLimit;
230 MultiXactId multiStopLimit;
231 MultiXactId multiWrapLimit;
232
233 /* support for members anti-wraparound measures */
234 MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
235
236 /*
237 * Per-backend data starts here. We have two arrays stored in the area
238 * immediately following the MultiXactStateData struct. Each is indexed by
239 * BackendId.
240 *
241 * In both arrays, there's a slot for all normal backends (1..MaxBackends)
242 * followed by a slot for max_prepared_xacts prepared transactions. Valid
243 * BackendIds start from 1; element zero of each array is never used.
244 *
245 * OldestMemberMXactId[k] is the oldest MultiXactId each backend's current
246 * transaction(s) could possibly be a member of, or InvalidMultiXactId
247 * when the backend has no live transaction that could possibly be a
248 * member of a MultiXact. Each backend sets its entry to the current
249 * nextMXact counter just before first acquiring a shared lock in a given
250 * transaction, and clears it at transaction end. (This works because only
251 * during or after acquiring a shared lock could an XID possibly become a
252 * member of a MultiXact, and that MultiXact would have to be created
253 * during or after the lock acquisition.)
254 *
255 * OldestVisibleMXactId[k] is the oldest MultiXactId each backend's
256 * current transaction(s) think is potentially live, or InvalidMultiXactId
257 * when not in a transaction or not in a transaction that's paid any
258 * attention to MultiXacts yet. This is computed when first needed in a
259 * given transaction, and cleared at transaction end. We can compute it
260 * as the minimum of the valid OldestMemberMXactId[] entries at the time
261 * we compute it (using nextMXact if none are valid). Each backend is
262 * required not to attempt to access any SLRU data for MultiXactIds older
263 * than its own OldestVisibleMXactId[] setting; this is necessary because
264 * the checkpointer could truncate away such data at any instant.
265 *
266 * The oldest valid value among all of the OldestMemberMXactId[] and
267 * OldestVisibleMXactId[] entries is considered by vacuum as the earliest
268 * possible value still having any live member transaction. Subtracting
269 * vacuum_multixact_freeze_min_age from that value we obtain the freezing
270 * point for multixacts for that table. Any value older than that is
271 * removed from tuple headers (or "frozen"; see FreezeMultiXactId. Note
272 * that multis that have member xids that are older than the cutoff point
273 * for xids must also be frozen, even if the multis themselves are newer
274 * than the multixid cutoff point). Whenever a full table vacuum happens,
275 * the freezing point so computed is used as the new pg_class.relminmxid
276 * value. The minimum of all those values in a database is stored as
277 * pg_database.datminmxid. In turn, the minimum of all of those values is
278 * stored in pg_control and used as truncation point for pg_multixact. At
279 * checkpoint or restartpoint, unneeded segments are removed.
280 */
281 MultiXactId perBackendXactIds[FLEXIBLE_ARRAY_MEMBER];
282} MultiXactStateData;
283
284/*
285 * Last element of OldestMemberMXactID and OldestVisibleMXactId arrays.
286 * Valid elements are (1..MaxOldestSlot); element 0 is never used.
287 */
288#define MaxOldestSlot (MaxBackends + max_prepared_xacts)
289
290/* Pointers to the state data in shared memory */
291static MultiXactStateData *MultiXactState;
292static MultiXactId *OldestMemberMXactId;
293static MultiXactId *OldestVisibleMXactId;
294
295
296/*
297 * Definitions for the backend-local MultiXactId cache.
298 *
299 * We use this cache to store known MultiXacts, so we don't need to go to
300 * SLRU areas every time.
301 *
302 * The cache lasts for the duration of a single transaction, the rationale
303 * for this being that most entries will contain our own TransactionId and
304 * so they will be uninteresting by the time our next transaction starts.
305 * (XXX not clear that this is correct --- other members of the MultiXact
306 * could hang around longer than we did. However, it's not clear what a
307 * better policy for flushing old cache entries would be.) FIXME actually
308 * this is plain wrong now that multixact's may contain update Xids.
309 *
310 * We allocate the cache entries in a memory context that is deleted at
311 * transaction end, so we don't need to do retail freeing of entries.
312 */
313typedef struct mXactCacheEnt
314{
315 MultiXactId multi;
316 int nmembers;
317 dlist_node node;
318 MultiXactMember members[FLEXIBLE_ARRAY_MEMBER];
319} mXactCacheEnt;
320
321#define MAX_CACHE_ENTRIES 256
322static dlist_head MXactCache = DLIST_STATIC_INIT(MXactCache);
323static int MXactCacheMembers = 0;
324static MemoryContext MXactContext = NULL;
325
326#ifdef MULTIXACT_DEBUG
327#define debug_elog2(a,b) elog(a,b)
328#define debug_elog3(a,b,c) elog(a,b,c)
329#define debug_elog4(a,b,c,d) elog(a,b,c,d)
330#define debug_elog5(a,b,c,d,e) elog(a,b,c,d,e)
331#define debug_elog6(a,b,c,d,e,f) elog(a,b,c,d,e,f)
332#else
333#define debug_elog2(a,b)
334#define debug_elog3(a,b,c)
335#define debug_elog4(a,b,c,d)
336#define debug_elog5(a,b,c,d,e)
337#define debug_elog6(a,b,c,d,e,f)
338#endif
339
340/* internal MultiXactId management */
341static void MultiXactIdSetOldestVisible(void);
342static void RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
343 int nmembers, MultiXactMember *members);
344static MultiXactId GetNewMultiXactId(int nmembers, MultiXactOffset *offset);
345
346/* MultiXact cache management */
347static int mxactMemberComparator(const void *arg1, const void *arg2);
348static MultiXactId mXactCacheGetBySet(int nmembers, MultiXactMember *members);
349static int mXactCacheGetById(MultiXactId multi, MultiXactMember **members);
350static void mXactCachePut(MultiXactId multi, int nmembers,
351 MultiXactMember *members);
352
353static char *mxstatus_to_string(MultiXactStatus status);
354
355/* management of SLRU infrastructure */
356static int ZeroMultiXactOffsetPage(int pageno, bool writeXlog);
357static int ZeroMultiXactMemberPage(int pageno, bool writeXlog);
358static bool MultiXactOffsetPagePrecedes(int page1, int page2);
359static bool MultiXactMemberPagePrecedes(int page1, int page2);
360static bool MultiXactOffsetPrecedes(MultiXactOffset offset1,
361 MultiXactOffset offset2);
362static void ExtendMultiXactOffset(MultiXactId multi);
363static void ExtendMultiXactMember(MultiXactOffset offset, int nmembers);
364static bool MultiXactOffsetWouldWrap(MultiXactOffset boundary,
365 MultiXactOffset start, uint32 distance);
366static bool SetOffsetVacuumLimit(bool is_startup);
367static bool find_multixact_start(MultiXactId multi, MultiXactOffset *result);
368static void WriteMZeroPageXlogRec(int pageno, uint8 info);
369static void WriteMTruncateXlogRec(Oid oldestMultiDB,
370 MultiXactId startOff, MultiXactId endOff,
371 MultiXactOffset startMemb, MultiXactOffset endMemb);
372
373
374/*
375 * MultiXactIdCreate
376 * Construct a MultiXactId representing two TransactionIds.
377 *
378 * The two XIDs must be different, or be requesting different statuses.
379 *
380 * NB - we don't worry about our local MultiXactId cache here, because that
381 * is handled by the lower-level routines.
382 */
383MultiXactId
384MultiXactIdCreate(TransactionId xid1, MultiXactStatus status1,
385 TransactionId xid2, MultiXactStatus status2)
386{
387 MultiXactId newMulti;
388 MultiXactMember members[2];
389
390 AssertArg(TransactionIdIsValid(xid1));
391 AssertArg(TransactionIdIsValid(xid2));
392
393 Assert(!TransactionIdEquals(xid1, xid2) || (status1 != status2));
394
395 /* MultiXactIdSetOldestMember() must have been called already. */
396 Assert(MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]));
397
398 /*
399 * Note: unlike MultiXactIdExpand, we don't bother to check that both XIDs
400 * are still running. In typical usage, xid2 will be our own XID and the
401 * caller just did a check on xid1, so it'd be wasted effort.
402 */
403
404 members[0].xid = xid1;
405 members[0].status = status1;
406 members[1].xid = xid2;
407 members[1].status = status2;
408
409 newMulti = MultiXactIdCreateFromMembers(2, members);
410
411 debug_elog3(DEBUG2, "Create: %s",
412 mxid_to_string(newMulti, 2, members));
413
414 return newMulti;
415}
416
417/*
418 * MultiXactIdExpand
419 * Add a TransactionId to a pre-existing MultiXactId.
420 *
421 * If the TransactionId is already a member of the passed MultiXactId with the
422 * same status, just return it as-is.
423 *
424 * Note that we do NOT actually modify the membership of a pre-existing
425 * MultiXactId; instead we create a new one. This is necessary to avoid
426 * a race condition against code trying to wait for one MultiXactId to finish;
427 * see notes in heapam.c.
428 *
429 * NB - we don't worry about our local MultiXactId cache here, because that
430 * is handled by the lower-level routines.
431 *
432 * Note: It is critical that MultiXactIds that come from an old cluster (i.e.
433 * one upgraded by pg_upgrade from a cluster older than this feature) are not
434 * passed in.
435 */
436MultiXactId
437MultiXactIdExpand(MultiXactId multi, TransactionId xid, MultiXactStatus status)
438{
439 MultiXactId newMulti;
440 MultiXactMember *members;
441 MultiXactMember *newMembers;
442 int nmembers;
443 int i;
444 int j;
445
446 AssertArg(MultiXactIdIsValid(multi));
447 AssertArg(TransactionIdIsValid(xid));
448
449 /* MultiXactIdSetOldestMember() must have been called already. */
450 Assert(MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]));
451
452 debug_elog5(DEBUG2, "Expand: received multi %u, xid %u status %s",
453 multi, xid, mxstatus_to_string(status));
454
455 /*
456 * Note: we don't allow for old multis here. The reason is that the only
457 * caller of this function does a check that the multixact is no longer
458 * running.
459 */
460 nmembers = GetMultiXactIdMembers(multi, &members, false, false);
461
462 if (nmembers < 0)
463 {
464 MultiXactMember member;
465
466 /*
467 * The MultiXactId is obsolete. This can only happen if all the
468 * MultiXactId members stop running between the caller checking and
469 * passing it to us. It would be better to return that fact to the
470 * caller, but it would complicate the API and it's unlikely to happen
471 * too often, so just deal with it by creating a singleton MultiXact.
472 */
473 member.xid = xid;
474 member.status = status;
475 newMulti = MultiXactIdCreateFromMembers(1, &member);
476
477 debug_elog4(DEBUG2, "Expand: %u has no members, create singleton %u",
478 multi, newMulti);
479 return newMulti;
480 }
481
482 /*
483 * If the TransactionId is already a member of the MultiXactId with the
484 * same status, just return the existing MultiXactId.
485 */
486 for (i = 0; i < nmembers; i++)
487 {
488 if (TransactionIdEquals(members[i].xid, xid) &&
489 (members[i].status == status))
490 {
491 debug_elog4(DEBUG2, "Expand: %u is already a member of %u",
492 xid, multi);
493 pfree(members);
494 return multi;
495 }
496 }
497
498 /*
499 * Determine which of the members of the MultiXactId are still of
500 * interest. This is any running transaction, and also any transaction
501 * that grabbed something stronger than just a lock and was committed. (An
502 * update that aborted is of no interest here; and having more than one
503 * update Xid in a multixact would cause errors elsewhere.)
504 *
505 * Removing dead members is not just an optimization: freezing of tuples
506 * whose Xmax are multis depends on this behavior.
507 *
508 * Note we have the same race condition here as above: j could be 0 at the
509 * end of the loop.
510 */
511 newMembers = (MultiXactMember *)
512 palloc(sizeof(MultiXactMember) * (nmembers + 1));
513
514 for (i = 0, j = 0; i < nmembers; i++)
515 {
516 if (TransactionIdIsInProgress(members[i].xid) ||
517 (ISUPDATE_from_mxstatus(members[i].status) &&
518 TransactionIdDidCommit(members[i].xid)))
519 {
520 newMembers[j].xid = members[i].xid;
521 newMembers[j++].status = members[i].status;
522 }
523 }
524
525 newMembers[j].xid = xid;
526 newMembers[j++].status = status;
527 newMulti = MultiXactIdCreateFromMembers(j, newMembers);
528
529 pfree(members);
530 pfree(newMembers);
531
532 debug_elog3(DEBUG2, "Expand: returning new multi %u", newMulti);
533
534 return newMulti;
535}
536
537/*
538 * MultiXactIdIsRunning
539 * Returns whether a MultiXactId is "running".
540 *
541 * We return true if at least one member of the given MultiXactId is still
542 * running. Note that a "false" result is certain not to change,
543 * because it is not legal to add members to an existing MultiXactId.
544 *
545 * Caller is expected to have verified that the multixact does not come from
546 * a pg_upgraded share-locked tuple.
547 */
548bool
549MultiXactIdIsRunning(MultiXactId multi, bool isLockOnly)
550{
551 MultiXactMember *members;
552 int nmembers;
553 int i;
554
555 debug_elog3(DEBUG2, "IsRunning %u?", multi);
556
557 /*
558 * "false" here means we assume our callers have checked that the given
559 * multi cannot possibly come from a pg_upgraded database.
560 */
561 nmembers = GetMultiXactIdMembers(multi, &members, false, isLockOnly);
562
563 if (nmembers <= 0)
564 {
565 debug_elog2(DEBUG2, "IsRunning: no members");
566 return false;
567 }
568
569 /*
570 * Checking for myself is cheap compared to looking in shared memory;
571 * return true if any live subtransaction of the current top-level
572 * transaction is a member.
573 *
574 * This is not needed for correctness, it's just a fast path.
575 */
576 for (i = 0; i < nmembers; i++)
577 {
578 if (TransactionIdIsCurrentTransactionId(members[i].xid))
579 {
580 debug_elog3(DEBUG2, "IsRunning: I (%d) am running!", i);
581 pfree(members);
582 return true;
583 }
584 }
585
586 /*
587 * This could be made faster by having another entry point in procarray.c,
588 * walking the PGPROC array only once for all the members. But in most
589 * cases nmembers should be small enough that it doesn't much matter.
590 */
591 for (i = 0; i < nmembers; i++)
592 {
593 if (TransactionIdIsInProgress(members[i].xid))
594 {
595 debug_elog4(DEBUG2, "IsRunning: member %d (%u) is running",
596 i, members[i].xid);
597 pfree(members);
598 return true;
599 }
600 }
601
602 pfree(members);
603
604 debug_elog3(DEBUG2, "IsRunning: %u is not running", multi);
605
606 return false;
607}
608
609/*
610 * MultiXactIdSetOldestMember
611 * Save the oldest MultiXactId this transaction could be a member of.
612 *
613 * We set the OldestMemberMXactId for a given transaction the first time it's
614 * going to do some operation that might require a MultiXactId (tuple lock,
615 * update or delete). We need to do this even if we end up using a
616 * TransactionId instead of a MultiXactId, because there is a chance that
617 * another transaction would add our XID to a MultiXactId.
618 *
619 * The value to set is the next-to-be-assigned MultiXactId, so this is meant to
620 * be called just before doing any such possibly-MultiXactId-able operation.
621 */
622void
623MultiXactIdSetOldestMember(void)
624{
625 if (!MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]))
626 {
627 MultiXactId nextMXact;
628
629 /*
630 * You might think we don't need to acquire a lock here, since
631 * fetching and storing of TransactionIds is probably atomic, but in
632 * fact we do: suppose we pick up nextMXact and then lose the CPU for
633 * a long time. Someone else could advance nextMXact, and then
634 * another someone else could compute an OldestVisibleMXactId that
635 * would be after the value we are going to store when we get control
636 * back. Which would be wrong.
637 *
638 * Note that a shared lock is sufficient, because it's enough to stop
639 * someone from advancing nextMXact; and nobody else could be trying
640 * to write to our OldestMember entry, only reading (and we assume
641 * storing it is atomic.)
642 */
643 LWLockAcquire(MultiXactGenLock, LW_SHARED);
644
645 /*
646 * We have to beware of the possibility that nextMXact is in the
647 * wrapped-around state. We don't fix the counter itself here, but we
648 * must be sure to store a valid value in our array entry.
649 */
650 nextMXact = MultiXactState->nextMXact;
651 if (nextMXact < FirstMultiXactId)
652 nextMXact = FirstMultiXactId;
653
654 OldestMemberMXactId[MyBackendId] = nextMXact;
655
656 LWLockRelease(MultiXactGenLock);
657
658 debug_elog4(DEBUG2, "MultiXact: setting OldestMember[%d] = %u",
659 MyBackendId, nextMXact);
660 }
661}
662
663/*
664 * MultiXactIdSetOldestVisible
665 * Save the oldest MultiXactId this transaction considers possibly live.
666 *
667 * We set the OldestVisibleMXactId for a given transaction the first time
668 * it's going to inspect any MultiXactId. Once we have set this, we are
669 * guaranteed that the checkpointer won't truncate off SLRU data for
670 * MultiXactIds at or after our OldestVisibleMXactId.
671 *
672 * The value to set is the oldest of nextMXact and all the valid per-backend
673 * OldestMemberMXactId[] entries. Because of the locking we do, we can be
674 * certain that no subsequent call to MultiXactIdSetOldestMember can set
675 * an OldestMemberMXactId[] entry older than what we compute here. Therefore
676 * there is no live transaction, now or later, that can be a member of any
677 * MultiXactId older than the OldestVisibleMXactId we compute here.
678 */
679static void
680MultiXactIdSetOldestVisible(void)
681{
682 if (!MultiXactIdIsValid(OldestVisibleMXactId[MyBackendId]))
683 {
684 MultiXactId oldestMXact;
685 int i;
686
687 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
688
689 /*
690 * We have to beware of the possibility that nextMXact is in the
691 * wrapped-around state. We don't fix the counter itself here, but we
692 * must be sure to store a valid value in our array entry.
693 */
694 oldestMXact = MultiXactState->nextMXact;
695 if (oldestMXact < FirstMultiXactId)
696 oldestMXact = FirstMultiXactId;
697
698 for (i = 1; i <= MaxOldestSlot; i++)
699 {
700 MultiXactId thisoldest = OldestMemberMXactId[i];
701
702 if (MultiXactIdIsValid(thisoldest) &&
703 MultiXactIdPrecedes(thisoldest, oldestMXact))
704 oldestMXact = thisoldest;
705 }
706
707 OldestVisibleMXactId[MyBackendId] = oldestMXact;
708
709 LWLockRelease(MultiXactGenLock);
710
711 debug_elog4(DEBUG2, "MultiXact: setting OldestVisible[%d] = %u",
712 MyBackendId, oldestMXact);
713 }
714}
715
716/*
717 * ReadNextMultiXactId
718 * Return the next MultiXactId to be assigned, but don't allocate it
719 */
720MultiXactId
721ReadNextMultiXactId(void)
722{
723 MultiXactId mxid;
724
725 /* XXX we could presumably do this without a lock. */
726 LWLockAcquire(MultiXactGenLock, LW_SHARED);
727 mxid = MultiXactState->nextMXact;
728 LWLockRelease(MultiXactGenLock);
729
730 if (mxid < FirstMultiXactId)
731 mxid = FirstMultiXactId;
732
733 return mxid;
734}
735
736/*
737 * MultiXactIdCreateFromMembers
738 * Make a new MultiXactId from the specified set of members
739 *
740 * Make XLOG, SLRU and cache entries for a new MultiXactId, recording the
741 * given TransactionIds as members. Returns the newly created MultiXactId.
742 *
743 * NB: the passed members[] array will be sorted in-place.
744 */
745MultiXactId
746MultiXactIdCreateFromMembers(int nmembers, MultiXactMember *members)
747{
748 MultiXactId multi;
749 MultiXactOffset offset;
750 xl_multixact_create xlrec;
751
752 debug_elog3(DEBUG2, "Create: %s",
753 mxid_to_string(InvalidMultiXactId, nmembers, members));
754
755 /*
756 * See if the same set of members already exists in our cache; if so, just
757 * re-use that MultiXactId. (Note: it might seem that looking in our
758 * cache is insufficient, and we ought to search disk to see if a
759 * duplicate definition already exists. But since we only ever create
760 * MultiXacts containing our own XID, in most cases any such MultiXacts
761 * were in fact created by us, and so will be in our cache. There are
762 * corner cases where someone else added us to a MultiXact without our
763 * knowledge, but it's not worth checking for.)
764 */
765 multi = mXactCacheGetBySet(nmembers, members);
766 if (MultiXactIdIsValid(multi))
767 {
768 debug_elog2(DEBUG2, "Create: in cache!");
769 return multi;
770 }
771
772 /* Verify that there is a single update Xid among the given members. */
773 {
774 int i;
775 bool has_update = false;
776
777 for (i = 0; i < nmembers; i++)
778 {
779 if (ISUPDATE_from_mxstatus(members[i].status))
780 {
781 if (has_update)
782 elog(ERROR, "new multixact has more than one updating member");
783 has_update = true;
784 }
785 }
786 }
787
788 /*
789 * Assign the MXID and offsets range to use, and make sure there is space
790 * in the OFFSETs and MEMBERs files. NB: this routine does
791 * START_CRIT_SECTION().
792 *
793 * Note: unlike MultiXactIdCreate and MultiXactIdExpand, we do not check
794 * that we've called MultiXactIdSetOldestMember here. This is because
795 * this routine is used in some places to create new MultiXactIds of which
796 * the current backend is not a member, notably during freezing of multis
797 * in vacuum. During vacuum, in particular, it would be unacceptable to
798 * keep OldestMulti set, in case it runs for long.
799 */
800 multi = GetNewMultiXactId(nmembers, &offset);
801
802 /* Make an XLOG entry describing the new MXID. */
803 xlrec.mid = multi;
804 xlrec.moff = offset;
805 xlrec.nmembers = nmembers;
806
807 /*
808 * XXX Note: there's a lot of padding space in MultiXactMember. We could
809 * find a more compact representation of this Xlog record -- perhaps all
810 * the status flags in one XLogRecData, then all the xids in another one?
811 * Not clear that it's worth the trouble though.
812 */
813 XLogBeginInsert();
814 XLogRegisterData((char *) (&xlrec), SizeOfMultiXactCreate);
815 XLogRegisterData((char *) members, nmembers * sizeof(MultiXactMember));
816
817 (void) XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_CREATE_ID);
818
819 /* Now enter the information into the OFFSETs and MEMBERs logs */
820 RecordNewMultiXact(multi, offset, nmembers, members);
821
822 /* Done with critical section */
823 END_CRIT_SECTION();
824
825 /* Store the new MultiXactId in the local cache, too */
826 mXactCachePut(multi, nmembers, members);
827
828 debug_elog2(DEBUG2, "Create: all done");
829
830 return multi;
831}
832
833/*
834 * RecordNewMultiXact
835 * Write info about a new multixact into the offsets and members files
836 *
837 * This is broken out of MultiXactIdCreateFromMembers so that xlog replay can
838 * use it.
839 */
840static void
841RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
842 int nmembers, MultiXactMember *members)
843{
844 int pageno;
845 int prev_pageno;
846 int entryno;
847 int slotno;
848 MultiXactOffset *offptr;
849 int i;
850
851 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
852
853 pageno = MultiXactIdToOffsetPage(multi);
854 entryno = MultiXactIdToOffsetEntry(multi);
855
856 /*
857 * Note: we pass the MultiXactId to SimpleLruReadPage as the "transaction"
858 * to complain about if there's any I/O error. This is kinda bogus, but
859 * since the errors will always give the full pathname, it should be clear
860 * enough that a MultiXactId is really involved. Perhaps someday we'll
861 * take the trouble to generalize the slru.c error reporting code.
862 */
863 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
864 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
865 offptr += entryno;
866
867 *offptr = offset;
868
869 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
870
871 /* Exchange our lock */
872 LWLockRelease(MultiXactOffsetControlLock);
873
874 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
875
876 prev_pageno = -1;
877
878 for (i = 0; i < nmembers; i++, offset++)
879 {
880 TransactionId *memberptr;
881 uint32 *flagsptr;
882 uint32 flagsval;
883 int bshift;
884 int flagsoff;
885 int memberoff;
886
887 Assert(members[i].status <= MultiXactStatusUpdate);
888
889 pageno = MXOffsetToMemberPage(offset);
890 memberoff = MXOffsetToMemberOffset(offset);
891 flagsoff = MXOffsetToFlagsOffset(offset);
892 bshift = MXOffsetToFlagsBitShift(offset);
893
894 if (pageno != prev_pageno)
895 {
896 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
897 prev_pageno = pageno;
898 }
899
900 memberptr = (TransactionId *)
901 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
902
903 *memberptr = members[i].xid;
904
905 flagsptr = (uint32 *)
906 (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
907
908 flagsval = *flagsptr;
909 flagsval &= ~(((1 << MXACT_MEMBER_BITS_PER_XACT) - 1) << bshift);
910 flagsval |= (members[i].status << bshift);
911 *flagsptr = flagsval;
912
913 MultiXactMemberCtl->shared->page_dirty[slotno] = true;
914 }
915
916 LWLockRelease(MultiXactMemberControlLock);
917}
918
919/*
920 * GetNewMultiXactId
921 * Get the next MultiXactId.
922 *
923 * Also, reserve the needed amount of space in the "members" area. The
924 * starting offset of the reserved space is returned in *offset.
925 *
926 * This may generate XLOG records for expansion of the offsets and/or members
927 * files. Unfortunately, we have to do that while holding MultiXactGenLock
928 * to avoid race conditions --- the XLOG record for zeroing a page must appear
929 * before any backend can possibly try to store data in that page!
930 *
931 * We start a critical section before advancing the shared counters. The
932 * caller must end the critical section after writing SLRU data.
933 */
934static MultiXactId
935GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
936{
937 MultiXactId result;
938 MultiXactOffset nextOffset;
939
940 debug_elog3(DEBUG2, "GetNew: for %d xids", nmembers);
941
942 /* safety check, we should never get this far in a HS standby */
943 if (RecoveryInProgress())
944 elog(ERROR, "cannot assign MultiXactIds during recovery");
945
946 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
947
948 /* Handle wraparound of the nextMXact counter */
949 if (MultiXactState->nextMXact < FirstMultiXactId)
950 MultiXactState->nextMXact = FirstMultiXactId;
951
952 /* Assign the MXID */
953 result = MultiXactState->nextMXact;
954
955 /*----------
956 * Check to see if it's safe to assign another MultiXactId. This protects
957 * against catastrophic data loss due to multixact wraparound. The basic
958 * rules are:
959 *
960 * If we're past multiVacLimit or the safe threshold for member storage
961 * space, or we don't know what the safe threshold for member storage is,
962 * start trying to force autovacuum cycles.
963 * If we're past multiWarnLimit, start issuing warnings.
964 * If we're past multiStopLimit, refuse to create new MultiXactIds.
965 *
966 * Note these are pretty much the same protections in GetNewTransactionId.
967 *----------
968 */
969 if (!MultiXactIdPrecedes(result, MultiXactState->multiVacLimit))
970 {
971 /*
972 * For safety's sake, we release MultiXactGenLock while sending
973 * signals, warnings, etc. This is not so much because we care about
974 * preserving concurrency in this situation, as to avoid any
975 * possibility of deadlock while doing get_database_name(). First,
976 * copy all the shared values we'll need in this path.
977 */
978 MultiXactId multiWarnLimit = MultiXactState->multiWarnLimit;
979 MultiXactId multiStopLimit = MultiXactState->multiStopLimit;
980 MultiXactId multiWrapLimit = MultiXactState->multiWrapLimit;
981 Oid oldest_datoid = MultiXactState->oldestMultiXactDB;
982
983 LWLockRelease(MultiXactGenLock);
984
985 if (IsUnderPostmaster &&
986 !MultiXactIdPrecedes(result, multiStopLimit))
987 {
988 char *oldest_datname = get_database_name(oldest_datoid);
989
990 /*
991 * Immediately kick autovacuum into action as we're already in
992 * ERROR territory.
993 */
994 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
995
996 /* complain even if that DB has disappeared */
997 if (oldest_datname)
998 ereport(ERROR,
999 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1000 errmsg("database is not accepting commands that generate new MultiXactIds to avoid wraparound data loss in database \"%s\"",
1001 oldest_datname),
1002 errhint("Execute a database-wide VACUUM in that database.\n"
1003 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1004 else
1005 ereport(ERROR,
1006 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1007 errmsg("database is not accepting commands that generate new MultiXactIds to avoid wraparound data loss in database with OID %u",
1008 oldest_datoid),
1009 errhint("Execute a database-wide VACUUM in that database.\n"
1010 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1011 }
1012
1013 /*
1014 * To avoid swamping the postmaster with signals, we issue the autovac
1015 * request only once per 64K multis generated. This still gives
1016 * plenty of chances before we get into real trouble.
1017 */
1018 if (IsUnderPostmaster && (result % 65536) == 0)
1019 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1020
1021 if (!MultiXactIdPrecedes(result, multiWarnLimit))
1022 {
1023 char *oldest_datname = get_database_name(oldest_datoid);
1024
1025 /* complain even if that DB has disappeared */
1026 if (oldest_datname)
1027 ereport(WARNING,
1028 (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
1029 "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
1030 multiWrapLimit - result,
1031 oldest_datname,
1032 multiWrapLimit - result),
1033 errhint("Execute a database-wide VACUUM in that database.\n"
1034 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1035 else
1036 ereport(WARNING,
1037 (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
1038 "database with OID %u must be vacuumed before %u more MultiXactIds are used",
1039 multiWrapLimit - result,
1040 oldest_datoid,
1041 multiWrapLimit - result),
1042 errhint("Execute a database-wide VACUUM in that database.\n"
1043 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1044 }
1045
1046 /* Re-acquire lock and start over */
1047 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1048 result = MultiXactState->nextMXact;
1049 if (result < FirstMultiXactId)
1050 result = FirstMultiXactId;
1051 }
1052
1053 /* Make sure there is room for the MXID in the file. */
1054 ExtendMultiXactOffset(result);
1055
1056 /*
1057 * Reserve the members space, similarly to above. Also, be careful not to
1058 * return zero as the starting offset for any multixact. See
1059 * GetMultiXactIdMembers() for motivation.
1060 */
1061 nextOffset = MultiXactState->nextOffset;
1062 if (nextOffset == 0)
1063 {
1064 *offset = 1;
1065 nmembers++; /* allocate member slot 0 too */
1066 }
1067 else
1068 *offset = nextOffset;
1069
1070 /*----------
1071 * Protect against overrun of the members space as well, with the
1072 * following rules:
1073 *
1074 * If we're past offsetStopLimit, refuse to generate more multis.
1075 * If we're close to offsetStopLimit, emit a warning.
1076 *
1077 * Arbitrarily, we start emitting warnings when we're 20 segments or less
1078 * from offsetStopLimit.
1079 *
1080 * Note we haven't updated the shared state yet, so if we fail at this
1081 * point, the multixact ID we grabbed can still be used by the next guy.
1082 *
1083 * Note that there is no point in forcing autovacuum runs here: the
1084 * multixact freeze settings would have to be reduced for that to have any
1085 * effect.
1086 *----------
1087 */
1088#define OFFSET_WARN_SEGMENTS 20
1089 if (MultiXactState->oldestOffsetKnown &&
1090 MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit, nextOffset,
1091 nmembers))
1092 {
1093 /* see comment in the corresponding offsets wraparound case */
1094 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1095
1096 ereport(ERROR,
1097 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1098 errmsg("multixact \"members\" limit exceeded"),
1099 errdetail_plural("This command would create a multixact with %u members, but the remaining space is only enough for %u member.",
1100 "This command would create a multixact with %u members, but the remaining space is only enough for %u members.",
1101 MultiXactState->offsetStopLimit - nextOffset - 1,
1102 nmembers,
1103 MultiXactState->offsetStopLimit - nextOffset - 1),
1104 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.",
1105 MultiXactState->oldestMultiXactDB)));
1106 }
1107
1108 /*
1109 * Check whether we should kick autovacuum into action, to prevent members
1110 * wraparound. NB we use a much larger window to trigger autovacuum than
1111 * just the warning limit. The warning is just a measure of last resort -
1112 * this is in line with GetNewTransactionId's behaviour.
1113 */
1114 if (!MultiXactState->oldestOffsetKnown ||
1115 (MultiXactState->nextOffset - MultiXactState->oldestOffset
1116 > MULTIXACT_MEMBER_SAFE_THRESHOLD))
1117 {
1118 /*
1119 * To avoid swamping the postmaster with signals, we issue the autovac
1120 * request only when crossing a segment boundary. With default
1121 * compilation settings that's roughly after 50k members. This still
1122 * gives plenty of chances before we get into real trouble.
1123 */
1124 if ((MXOffsetToMemberPage(nextOffset) / SLRU_PAGES_PER_SEGMENT) !=
1125 (MXOffsetToMemberPage(nextOffset + nmembers) / SLRU_PAGES_PER_SEGMENT))
1126 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1127 }
1128
1129 if (MultiXactState->oldestOffsetKnown &&
1130 MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit,
1131 nextOffset,
1132 nmembers + MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT * OFFSET_WARN_SEGMENTS))
1133 ereport(WARNING,
1134 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1135 errmsg_plural("database with OID %u must be vacuumed before %d more multixact member is used",
1136 "database with OID %u must be vacuumed before %d more multixact members are used",
1137 MultiXactState->offsetStopLimit - nextOffset + nmembers,
1138 MultiXactState->oldestMultiXactDB,
1139 MultiXactState->offsetStopLimit - nextOffset + nmembers),
1140 errhint("Execute a database-wide VACUUM in that database with reduced vacuum_multixact_freeze_min_age and vacuum_multixact_freeze_table_age settings.")));
1141
1142 ExtendMultiXactMember(nextOffset, nmembers);
1143
1144 /*
1145 * Critical section from here until caller has written the data into the
1146 * just-reserved SLRU space; we don't want to error out with a partly
1147 * written MultiXact structure. (In particular, failing to write our
1148 * start offset after advancing nextMXact would effectively corrupt the
1149 * previous MultiXact.)
1150 */
1151 START_CRIT_SECTION();
1152
1153 /*
1154 * Advance counters. As in GetNewTransactionId(), this must not happen
1155 * until after file extension has succeeded!
1156 *
1157 * We don't care about MultiXactId wraparound here; it will be handled by
1158 * the next iteration. But note that nextMXact may be InvalidMultiXactId
1159 * or the first value on a segment-beginning page after this routine
1160 * exits, so anyone else looking at the variable must be prepared to deal
1161 * with either case. Similarly, nextOffset may be zero, but we won't use
1162 * that as the actual start offset of the next multixact.
1163 */
1164 (MultiXactState->nextMXact)++;
1165
1166 MultiXactState->nextOffset += nmembers;
1167
1168 LWLockRelease(MultiXactGenLock);
1169
1170 debug_elog4(DEBUG2, "GetNew: returning %u offset %u", result, *offset);
1171 return result;
1172}
1173
1174/*
1175 * GetMultiXactIdMembers
1176 * Return the set of MultiXactMembers that make up a MultiXactId
1177 *
1178 * Return value is the number of members found, or -1 if there are none,
1179 * and *members is set to a newly palloc'ed array of members. It's the
1180 * caller's responsibility to free it when done with it.
1181 *
1182 * from_pgupgrade must be passed as true if and only if only the multixact
1183 * corresponds to a value from a tuple that was locked in a 9.2-or-older
1184 * installation and later pg_upgrade'd (that is, the infomask is
1185 * HEAP_LOCKED_UPGRADED). In this case, we know for certain that no members
1186 * can still be running, so we return -1 just like for an empty multixact
1187 * without any further checking. It would be wrong to try to resolve such a
1188 * multixact: either the multixact is within the current valid multixact
1189 * range, in which case the returned result would be bogus, or outside that
1190 * range, in which case an error would be raised.
1191 *
1192 * In all other cases, the passed multixact must be within the known valid
1193 * range, that is, greater to or equal than oldestMultiXactId, and less than
1194 * nextMXact. Otherwise, an error is raised.
1195 *
1196 * onlyLock must be set to true if caller is certain that the given multi
1197 * is used only to lock tuples; can be false without loss of correctness,
1198 * but passing a true means we can return quickly without checking for
1199 * old updates.
1200 */
1201int
1202GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
1203 bool from_pgupgrade, bool onlyLock)
1204{
1205 int pageno;
1206 int prev_pageno;
1207 int entryno;
1208 int slotno;
1209 MultiXactOffset *offptr;
1210 MultiXactOffset offset;
1211 int length;
1212 int truelength;
1213 int i;
1214 MultiXactId oldestMXact;
1215 MultiXactId nextMXact;
1216 MultiXactId tmpMXact;
1217 MultiXactOffset nextOffset;
1218 MultiXactMember *ptr;
1219
1220 debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
1221
1222 if (!MultiXactIdIsValid(multi) || from_pgupgrade)
1223 return -1;
1224
1225 /* See if the MultiXactId is in the local cache */
1226 length = mXactCacheGetById(multi, members);
1227 if (length >= 0)
1228 {
1229 debug_elog3(DEBUG2, "GetMembers: found %s in the cache",
1230 mxid_to_string(multi, length, *members));
1231 return length;
1232 }
1233
1234 /* Set our OldestVisibleMXactId[] entry if we didn't already */
1235 MultiXactIdSetOldestVisible();
1236
1237 /*
1238 * If we know the multi is used only for locking and not for updates, then
1239 * we can skip checking if the value is older than our oldest visible
1240 * multi. It cannot possibly still be running.
1241 */
1242 if (onlyLock &&
1243 MultiXactIdPrecedes(multi, OldestVisibleMXactId[MyBackendId]))
1244 {
1245 debug_elog2(DEBUG2, "GetMembers: a locker-only multi is too old");
1246 *members = NULL;
1247 return -1;
1248 }
1249
1250 /*
1251 * We check known limits on MultiXact before resorting to the SLRU area.
1252 *
1253 * An ID older than MultiXactState->oldestMultiXactId cannot possibly be
1254 * useful; it has already been removed, or will be removed shortly, by
1255 * truncation. If one is passed, an error is raised.
1256 *
1257 * Also, an ID >= nextMXact shouldn't ever be seen here; if it is seen, it
1258 * implies undetected ID wraparound has occurred. This raises a hard
1259 * error.
1260 *
1261 * Shared lock is enough here since we aren't modifying any global state.
1262 * Acquire it just long enough to grab the current counter values. We may
1263 * need both nextMXact and nextOffset; see below.
1264 */
1265 LWLockAcquire(MultiXactGenLock, LW_SHARED);
1266
1267 oldestMXact = MultiXactState->oldestMultiXactId;
1268 nextMXact = MultiXactState->nextMXact;
1269 nextOffset = MultiXactState->nextOffset;
1270
1271 LWLockRelease(MultiXactGenLock);
1272
1273 if (MultiXactIdPrecedes(multi, oldestMXact))
1274 {
1275 ereport(ERROR,
1276 (errcode(ERRCODE_INTERNAL_ERROR),
1277 errmsg("MultiXactId %u does no longer exist -- apparent wraparound",
1278 multi)));
1279 return -1;
1280 }
1281
1282 if (!MultiXactIdPrecedes(multi, nextMXact))
1283 ereport(ERROR,
1284 (errcode(ERRCODE_INTERNAL_ERROR),
1285 errmsg("MultiXactId %u has not been created yet -- apparent wraparound",
1286 multi)));
1287
1288 /*
1289 * Find out the offset at which we need to start reading MultiXactMembers
1290 * and the number of members in the multixact. We determine the latter as
1291 * the difference between this multixact's starting offset and the next
1292 * one's. However, there are some corner cases to worry about:
1293 *
1294 * 1. This multixact may be the latest one created, in which case there is
1295 * no next one to look at. In this case the nextOffset value we just
1296 * saved is the correct endpoint.
1297 *
1298 * 2. The next multixact may still be in process of being filled in: that
1299 * is, another process may have done GetNewMultiXactId but not yet written
1300 * the offset entry for that ID. In that scenario, it is guaranteed that
1301 * the offset entry for that multixact exists (because GetNewMultiXactId
1302 * won't release MultiXactGenLock until it does) but contains zero
1303 * (because we are careful to pre-zero offset pages). Because
1304 * GetNewMultiXactId will never return zero as the starting offset for a
1305 * multixact, when we read zero as the next multixact's offset, we know we
1306 * have this case. We sleep for a bit and try again.
1307 *
1308 * 3. Because GetNewMultiXactId increments offset zero to offset one to
1309 * handle case #2, there is an ambiguity near the point of offset
1310 * wraparound. If we see next multixact's offset is one, is that our
1311 * multixact's actual endpoint, or did it end at zero with a subsequent
1312 * increment? We handle this using the knowledge that if the zero'th
1313 * member slot wasn't filled, it'll contain zero, and zero isn't a valid
1314 * transaction ID so it can't be a multixact member. Therefore, if we
1315 * read a zero from the members array, just ignore it.
1316 *
1317 * This is all pretty messy, but the mess occurs only in infrequent corner
1318 * cases, so it seems better than holding the MultiXactGenLock for a long
1319 * time on every multixact creation.
1320 */
1321retry:
1322 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
1323
1324 pageno = MultiXactIdToOffsetPage(multi);
1325 entryno = MultiXactIdToOffsetEntry(multi);
1326
1327 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
1328 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1329 offptr += entryno;
1330 offset = *offptr;
1331
1332 Assert(offset != 0);
1333
1334 /*
1335 * Use the same increment rule as GetNewMultiXactId(), that is, don't
1336 * handle wraparound explicitly until needed.
1337 */
1338 tmpMXact = multi + 1;
1339
1340 if (nextMXact == tmpMXact)
1341 {
1342 /* Corner case 1: there is no next multixact */
1343 length = nextOffset - offset;
1344 }
1345 else
1346 {
1347 MultiXactOffset nextMXOffset;
1348
1349 /* handle wraparound if needed */
1350 if (tmpMXact < FirstMultiXactId)
1351 tmpMXact = FirstMultiXactId;
1352
1353 prev_pageno = pageno;
1354
1355 pageno = MultiXactIdToOffsetPage(tmpMXact);
1356 entryno = MultiXactIdToOffsetEntry(tmpMXact);
1357
1358 if (pageno != prev_pageno)
1359 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
1360
1361 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1362 offptr += entryno;
1363 nextMXOffset = *offptr;
1364
1365 if (nextMXOffset == 0)
1366 {
1367 /* Corner case 2: next multixact is still being filled in */
1368 LWLockRelease(MultiXactOffsetControlLock);
1369 CHECK_FOR_INTERRUPTS();
1370 pg_usleep(1000L);
1371 goto retry;
1372 }
1373
1374 length = nextMXOffset - offset;
1375 }
1376
1377 LWLockRelease(MultiXactOffsetControlLock);
1378
1379 ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
1380 *members = ptr;
1381
1382 /* Now get the members themselves. */
1383 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
1384
1385 truelength = 0;
1386 prev_pageno = -1;
1387 for (i = 0; i < length; i++, offset++)
1388 {
1389 TransactionId *xactptr;
1390 uint32 *flagsptr;
1391 int flagsoff;
1392 int bshift;
1393 int memberoff;
1394
1395 pageno = MXOffsetToMemberPage(offset);
1396 memberoff = MXOffsetToMemberOffset(offset);
1397
1398 if (pageno != prev_pageno)
1399 {
1400 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
1401 prev_pageno = pageno;
1402 }
1403
1404 xactptr = (TransactionId *)
1405 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
1406
1407 if (!TransactionIdIsValid(*xactptr))
1408 {
1409 /* Corner case 3: we must be looking at unused slot zero */
1410 Assert(offset == 0);
1411 continue;
1412 }
1413
1414 flagsoff = MXOffsetToFlagsOffset(offset);
1415 bshift = MXOffsetToFlagsBitShift(offset);
1416 flagsptr = (uint32 *) (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
1417
1418 ptr[truelength].xid = *xactptr;
1419 ptr[truelength].status = (*flagsptr >> bshift) & MXACT_MEMBER_XACT_BITMASK;
1420 truelength++;
1421 }
1422
1423 LWLockRelease(MultiXactMemberControlLock);
1424
1425 /*
1426 * Copy the result into the local cache.
1427 */
1428 mXactCachePut(multi, truelength, ptr);
1429
1430 debug_elog3(DEBUG2, "GetMembers: no cache for %s",
1431 mxid_to_string(multi, truelength, ptr));
1432 return truelength;
1433}
1434
1435/*
1436 * mxactMemberComparator
1437 * qsort comparison function for MultiXactMember
1438 *
1439 * We can't use wraparound comparison for XIDs because that does not respect
1440 * the triangle inequality! Any old sort order will do.
1441 */
1442static int
1443mxactMemberComparator(const void *arg1, const void *arg2)
1444{
1445 MultiXactMember member1 = *(const MultiXactMember *) arg1;
1446 MultiXactMember member2 = *(const MultiXactMember *) arg2;
1447
1448 if (member1.xid > member2.xid)
1449 return 1;
1450 if (member1.xid < member2.xid)
1451 return -1;
1452 if (member1.status > member2.status)
1453 return 1;
1454 if (member1.status < member2.status)
1455 return -1;
1456 return 0;
1457}
1458
1459/*
1460 * mXactCacheGetBySet
1461 * returns a MultiXactId from the cache based on the set of
1462 * TransactionIds that compose it, or InvalidMultiXactId if
1463 * none matches.
1464 *
1465 * This is helpful, for example, if two transactions want to lock a huge
1466 * table. By using the cache, the second will use the same MultiXactId
1467 * for the majority of tuples, thus keeping MultiXactId usage low (saving
1468 * both I/O and wraparound issues).
1469 *
1470 * NB: the passed members array will be sorted in-place.
1471 */
1472static MultiXactId
1473mXactCacheGetBySet(int nmembers, MultiXactMember *members)
1474{
1475 dlist_iter iter;
1476
1477 debug_elog3(DEBUG2, "CacheGet: looking for %s",
1478 mxid_to_string(InvalidMultiXactId, nmembers, members));
1479
1480 /* sort the array so comparison is easy */
1481 qsort(members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1482
1483 dlist_foreach(iter, &MXactCache)
1484 {
1485 mXactCacheEnt *entry = dlist_container(mXactCacheEnt, node, iter.cur);
1486
1487 if (entry->nmembers != nmembers)
1488 continue;
1489
1490 /*
1491 * We assume the cache entries are sorted, and that the unused bits in
1492 * "status" are zeroed.
1493 */
1494 if (memcmp(members, entry->members, nmembers * sizeof(MultiXactMember)) == 0)
1495 {
1496 debug_elog3(DEBUG2, "CacheGet: found %u", entry->multi);
1497 dlist_move_head(&MXactCache, iter.cur);
1498 return entry->multi;
1499 }
1500 }
1501
1502 debug_elog2(DEBUG2, "CacheGet: not found :-(");
1503 return InvalidMultiXactId;
1504}
1505
1506/*
1507 * mXactCacheGetById
1508 * returns the composing MultiXactMember set from the cache for a
1509 * given MultiXactId, if present.
1510 *
1511 * If successful, *xids is set to the address of a palloc'd copy of the
1512 * MultiXactMember set. Return value is number of members, or -1 on failure.
1513 */
1514static int
1515mXactCacheGetById(MultiXactId multi, MultiXactMember **members)
1516{
1517 dlist_iter iter;
1518
1519 debug_elog3(DEBUG2, "CacheGet: looking for %u", multi);
1520
1521 dlist_foreach(iter, &MXactCache)
1522 {
1523 mXactCacheEnt *entry = dlist_container(mXactCacheEnt, node, iter.cur);
1524
1525 if (entry->multi == multi)
1526 {
1527 MultiXactMember *ptr;
1528 Size size;
1529
1530 size = sizeof(MultiXactMember) * entry->nmembers;
1531 ptr = (MultiXactMember *) palloc(size);
1532 *members = ptr;
1533
1534 memcpy(ptr, entry->members, size);
1535
1536 debug_elog3(DEBUG2, "CacheGet: found %s",
1537 mxid_to_string(multi,
1538 entry->nmembers,
1539 entry->members));
1540
1541 /*
1542 * Note we modify the list while not using a modifiable iterator.
1543 * This is acceptable only because we exit the iteration
1544 * immediately afterwards.
1545 */
1546 dlist_move_head(&MXactCache, iter.cur);
1547
1548 return entry->nmembers;
1549 }
1550 }
1551
1552 debug_elog2(DEBUG2, "CacheGet: not found");
1553 return -1;
1554}
1555
1556/*
1557 * mXactCachePut
1558 * Add a new MultiXactId and its composing set into the local cache.
1559 */
1560static void
1561mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
1562{
1563 mXactCacheEnt *entry;
1564
1565 debug_elog3(DEBUG2, "CachePut: storing %s",
1566 mxid_to_string(multi, nmembers, members));
1567
1568 if (MXactContext == NULL)
1569 {
1570 /* The cache only lives as long as the current transaction */
1571 debug_elog2(DEBUG2, "CachePut: initializing memory context");
1572 MXactContext = AllocSetContextCreate(TopTransactionContext,
1573 "MultiXact cache context",
1574 ALLOCSET_SMALL_SIZES);
1575 }
1576
1577 entry = (mXactCacheEnt *)
1578 MemoryContextAlloc(MXactContext,
1579 offsetof(mXactCacheEnt, members) +
1580 nmembers * sizeof(MultiXactMember));
1581
1582 entry->multi = multi;
1583 entry->nmembers = nmembers;
1584 memcpy(entry->members, members, nmembers * sizeof(MultiXactMember));
1585
1586 /* mXactCacheGetBySet assumes the entries are sorted, so sort them */
1587 qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1588
1589 dlist_push_head(&MXactCache, &entry->node);
1590 if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
1591 {
1592 dlist_node *node;
1593 mXactCacheEnt *entry;
1594
1595 node = dlist_tail_node(&MXactCache);
1596 dlist_delete(node);
1597 MXactCacheMembers--;
1598
1599 entry = dlist_container(mXactCacheEnt, node, node);
1600 debug_elog3(DEBUG2, "CachePut: pruning cached multi %u",
1601 entry->multi);
1602
1603 pfree(entry);
1604 }
1605}
1606
1607static char *
1608mxstatus_to_string(MultiXactStatus status)
1609{
1610 switch (status)
1611 {
1612 case MultiXactStatusForKeyShare:
1613 return "keysh";
1614 case MultiXactStatusForShare:
1615 return "sh";
1616 case MultiXactStatusForNoKeyUpdate:
1617 return "fornokeyupd";
1618 case MultiXactStatusForUpdate:
1619 return "forupd";
1620 case MultiXactStatusNoKeyUpdate:
1621 return "nokeyupd";
1622 case MultiXactStatusUpdate:
1623 return "upd";
1624 default:
1625 elog(ERROR, "unrecognized multixact status %d", status);
1626 return "";
1627 }
1628}
1629
1630char *
1631mxid_to_string(MultiXactId multi, int nmembers, MultiXactMember *members)
1632{
1633 static char *str = NULL;
1634 StringInfoData buf;
1635 int i;
1636
1637 if (str != NULL)
1638 pfree(str);
1639
1640 initStringInfo(&buf);
1641
1642 appendStringInfo(&buf, "%u %d[%u (%s)", multi, nmembers, members[0].xid,
1643 mxstatus_to_string(members[0].status));
1644
1645 for (i = 1; i < nmembers; i++)
1646 appendStringInfo(&buf, ", %u (%s)", members[i].xid,
1647 mxstatus_to_string(members[i].status));
1648
1649 appendStringInfoChar(&buf, ']');
1650 str = MemoryContextStrdup(TopMemoryContext, buf.data);
1651 pfree(buf.data);
1652 return str;
1653}
1654
1655/*
1656 * AtEOXact_MultiXact
1657 * Handle transaction end for MultiXact
1658 *
1659 * This is called at top transaction commit or abort (we don't care which).
1660 */
1661void
1662AtEOXact_MultiXact(void)
1663{
1664 /*
1665 * Reset our OldestMemberMXactId and OldestVisibleMXactId values, both of
1666 * which should only be valid while within a transaction.
1667 *
1668 * We assume that storing a MultiXactId is atomic and so we need not take
1669 * MultiXactGenLock to do this.
1670 */
1671 OldestMemberMXactId[MyBackendId] = InvalidMultiXactId;
1672 OldestVisibleMXactId[MyBackendId] = InvalidMultiXactId;
1673
1674 /*
1675 * Discard the local MultiXactId cache. Since MXactContext was created as
1676 * a child of TopTransactionContext, we needn't delete it explicitly.
1677 */
1678 MXactContext = NULL;
1679 dlist_init(&MXactCache);
1680 MXactCacheMembers = 0;
1681}
1682
1683/*
1684 * AtPrepare_MultiXact
1685 * Save multixact state at 2PC transaction prepare
1686 *
1687 * In this phase, we only store our OldestMemberMXactId value in the two-phase
1688 * state file.
1689 */
1690void
1691AtPrepare_MultiXact(void)
1692{
1693 MultiXactId myOldestMember = OldestMemberMXactId[MyBackendId];
1694
1695 if (MultiXactIdIsValid(myOldestMember))
1696 RegisterTwoPhaseRecord(TWOPHASE_RM_MULTIXACT_ID, 0,
1697 &myOldestMember, sizeof(MultiXactId));
1698}
1699
1700/*
1701 * PostPrepare_MultiXact
1702 * Clean up after successful PREPARE TRANSACTION
1703 */
1704void
1705PostPrepare_MultiXact(TransactionId xid)
1706{
1707 MultiXactId myOldestMember;
1708
1709 /*
1710 * Transfer our OldestMemberMXactId value to the slot reserved for the
1711 * prepared transaction.
1712 */
1713 myOldestMember = OldestMemberMXactId[MyBackendId];
1714 if (MultiXactIdIsValid(myOldestMember))
1715 {
1716 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid, false);
1717
1718 /*
1719 * Even though storing MultiXactId is atomic, acquire lock to make
1720 * sure others see both changes, not just the reset of the slot of the
1721 * current backend. Using a volatile pointer might suffice, but this
1722 * isn't a hot spot.
1723 */
1724 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1725
1726 OldestMemberMXactId[dummyBackendId] = myOldestMember;
1727 OldestMemberMXactId[MyBackendId] = InvalidMultiXactId;
1728
1729 LWLockRelease(MultiXactGenLock);
1730 }
1731
1732 /*
1733 * We don't need to transfer OldestVisibleMXactId value, because the
1734 * transaction is not going to be looking at any more multixacts once it's
1735 * prepared.
1736 *
1737 * We assume that storing a MultiXactId is atomic and so we need not take
1738 * MultiXactGenLock to do this.
1739 */
1740 OldestVisibleMXactId[MyBackendId] = InvalidMultiXactId;
1741
1742 /*
1743 * Discard the local MultiXactId cache like in AtEOX_MultiXact
1744 */
1745 MXactContext = NULL;
1746 dlist_init(&MXactCache);
1747 MXactCacheMembers = 0;
1748}
1749
1750/*
1751 * multixact_twophase_recover
1752 * Recover the state of a prepared transaction at startup
1753 */
1754void
1755multixact_twophase_recover(TransactionId xid, uint16 info,
1756 void *recdata, uint32 len)
1757{
1758 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid, false);
1759 MultiXactId oldestMember;
1760
1761 /*
1762 * Get the oldest member XID from the state file record, and set it in the
1763 * OldestMemberMXactId slot reserved for this prepared transaction.
1764 */
1765 Assert(len == sizeof(MultiXactId));
1766 oldestMember = *((MultiXactId *) recdata);
1767
1768 OldestMemberMXactId[dummyBackendId] = oldestMember;
1769}
1770
1771/*
1772 * multixact_twophase_postcommit
1773 * Similar to AtEOX_MultiXact but for COMMIT PREPARED
1774 */
1775void
1776multixact_twophase_postcommit(TransactionId xid, uint16 info,
1777 void *recdata, uint32 len)
1778{
1779 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid, true);
1780
1781 Assert(len == sizeof(MultiXactId));
1782
1783 OldestMemberMXactId[dummyBackendId] = InvalidMultiXactId;
1784}
1785
1786/*
1787 * multixact_twophase_postabort
1788 * This is actually just the same as the COMMIT case.
1789 */
1790void
1791multixact_twophase_postabort(TransactionId xid, uint16 info,
1792 void *recdata, uint32 len)
1793{
1794 multixact_twophase_postcommit(xid, info, recdata, len);
1795}
1796
1797/*
1798 * Initialization of shared memory for MultiXact. We use two SLRU areas,
1799 * thus double memory. Also, reserve space for the shared MultiXactState
1800 * struct and the per-backend MultiXactId arrays (two of those, too).
1801 */
1802Size
1803MultiXactShmemSize(void)
1804{
1805 Size size;
1806
1807 /* We need 2*MaxOldestSlot + 1 perBackendXactIds[] entries */
1808#define SHARED_MULTIXACT_STATE_SIZE \
1809 add_size(offsetof(MultiXactStateData, perBackendXactIds) + sizeof(MultiXactId), \
1810 mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
1811
1812 size = SHARED_MULTIXACT_STATE_SIZE;
1813 size = add_size(size, SimpleLruShmemSize(NUM_MXACTOFFSET_BUFFERS, 0));
1814 size = add_size(size, SimpleLruShmemSize(NUM_MXACTMEMBER_BUFFERS, 0));
1815
1816 return size;
1817}
1818
1819void
1820MultiXactShmemInit(void)
1821{
1822 bool found;
1823
1824 debug_elog2(DEBUG2, "Shared Memory Init for MultiXact");
1825
1826 MultiXactOffsetCtl->PagePrecedes = MultiXactOffsetPagePrecedes;
1827 MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
1828
1829 SimpleLruInit(MultiXactOffsetCtl,
1830 "multixact_offset", NUM_MXACTOFFSET_BUFFERS, 0,
1831 MultiXactOffsetControlLock, "pg_multixact/offsets",
1832 LWTRANCHE_MXACTOFFSET_BUFFERS);
1833 SimpleLruInit(MultiXactMemberCtl,
1834 "multixact_member", NUM_MXACTMEMBER_BUFFERS, 0,
1835 MultiXactMemberControlLock, "pg_multixact/members",
1836 LWTRANCHE_MXACTMEMBER_BUFFERS);
1837
1838 /* Initialize our shared state struct */
1839 MultiXactState = ShmemInitStruct("Shared MultiXact State",
1840 SHARED_MULTIXACT_STATE_SIZE,
1841 &found);
1842 if (!IsUnderPostmaster)
1843 {
1844 Assert(!found);
1845
1846 /* Make sure we zero out the per-backend state */
1847 MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
1848 }
1849 else
1850 Assert(found);
1851
1852 /*
1853 * Set up array pointers. Note that perBackendXactIds[0] is wasted space
1854 * since we only use indexes 1..MaxOldestSlot in each array.
1855 */
1856 OldestMemberMXactId = MultiXactState->perBackendXactIds;
1857 OldestVisibleMXactId = OldestMemberMXactId + MaxOldestSlot;
1858}
1859
1860/*
1861 * This func must be called ONCE on system install. It creates the initial
1862 * MultiXact segments. (The MultiXacts directories are assumed to have been
1863 * created by initdb, and MultiXactShmemInit must have been called already.)
1864 */
1865void
1866BootStrapMultiXact(void)
1867{
1868 int slotno;
1869
1870 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
1871
1872 /* Create and zero the first page of the offsets log */
1873 slotno = ZeroMultiXactOffsetPage(0, false);
1874
1875 /* Make sure it's written out */
1876 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
1877 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
1878
1879 LWLockRelease(MultiXactOffsetControlLock);
1880
1881 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
1882
1883 /* Create and zero the first page of the members log */
1884 slotno = ZeroMultiXactMemberPage(0, false);
1885
1886 /* Make sure it's written out */
1887 SimpleLruWritePage(MultiXactMemberCtl, slotno);
1888 Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
1889
1890 LWLockRelease(MultiXactMemberControlLock);
1891}
1892
1893/*
1894 * Initialize (or reinitialize) a page of MultiXactOffset to zeroes.
1895 * If writeXlog is true, also emit an XLOG record saying we did this.
1896 *
1897 * The page is not actually written, just set up in shared memory.
1898 * The slot number of the new page is returned.
1899 *
1900 * Control lock must be held at entry, and will be held at exit.
1901 */
1902static int
1903ZeroMultiXactOffsetPage(int pageno, bool writeXlog)
1904{
1905 int slotno;
1906
1907 slotno = SimpleLruZeroPage(MultiXactOffsetCtl, pageno);
1908
1909 if (writeXlog)
1910 WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_OFF_PAGE);
1911
1912 return slotno;
1913}
1914
1915/*
1916 * Ditto, for MultiXactMember
1917 */
1918static int
1919ZeroMultiXactMemberPage(int pageno, bool writeXlog)
1920{
1921 int slotno;
1922
1923 slotno = SimpleLruZeroPage(MultiXactMemberCtl, pageno);
1924
1925 if (writeXlog)
1926 WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_MEM_PAGE);
1927
1928 return slotno;
1929}
1930
1931/*
1932 * MaybeExtendOffsetSlru
1933 * Extend the offsets SLRU area, if necessary
1934 *
1935 * After a binary upgrade from <= 9.2, the pg_multixact/offsets SLRU area might
1936 * contain files that are shorter than necessary; this would occur if the old
1937 * installation had used multixacts beyond the first page (files cannot be
1938 * copied, because the on-disk representation is different). pg_upgrade would
1939 * update pg_control to set the next offset value to be at that position, so
1940 * that tuples marked as locked by such MultiXacts would be seen as visible
1941 * without having to consult multixact. However, trying to create and use a
1942 * new MultiXactId would result in an error because the page on which the new
1943 * value would reside does not exist. This routine is in charge of creating
1944 * such pages.
1945 */
1946static void
1947MaybeExtendOffsetSlru(void)
1948{
1949 int pageno;
1950
1951 pageno = MultiXactIdToOffsetPage(MultiXactState->nextMXact);
1952
1953 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
1954
1955 if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
1956 {
1957 int slotno;
1958
1959 /*
1960 * Fortunately for us, SimpleLruWritePage is already prepared to deal
1961 * with creating a new segment file even if the page we're writing is
1962 * not the first in it, so this is enough.
1963 */
1964 slotno = ZeroMultiXactOffsetPage(pageno, false);
1965 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
1966 }
1967
1968 LWLockRelease(MultiXactOffsetControlLock);
1969}
1970
1971/*
1972 * This must be called ONCE during postmaster or standalone-backend startup.
1973 *
1974 * StartupXLOG has already established nextMXact/nextOffset by calling
1975 * MultiXactSetNextMXact and/or MultiXactAdvanceNextMXact, and the oldestMulti
1976 * info from pg_control and/or MultiXactAdvanceOldest, but we haven't yet
1977 * replayed WAL.
1978 */
1979void
1980StartupMultiXact(void)
1981{
1982 MultiXactId multi = MultiXactState->nextMXact;
1983 MultiXactOffset offset = MultiXactState->nextOffset;
1984 int pageno;
1985
1986 /*
1987 * Initialize offset's idea of the latest page number.
1988 */
1989 pageno = MultiXactIdToOffsetPage(multi);
1990 MultiXactOffsetCtl->shared->latest_page_number = pageno;
1991
1992 /*
1993 * Initialize member's idea of the latest page number.
1994 */
1995 pageno = MXOffsetToMemberPage(offset);
1996 MultiXactMemberCtl->shared->latest_page_number = pageno;
1997}
1998
1999/*
2000 * This must be called ONCE at the end of startup/recovery.
2001 */
2002void
2003TrimMultiXact(void)
2004{
2005 MultiXactId nextMXact;
2006 MultiXactOffset offset;
2007 MultiXactId oldestMXact;
2008 Oid oldestMXactDB;
2009 int pageno;
2010 int entryno;
2011 int flagsoff;
2012
2013 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2014 nextMXact = MultiXactState->nextMXact;
2015 offset = MultiXactState->nextOffset;
2016 oldestMXact = MultiXactState->oldestMultiXactId;
2017 oldestMXactDB = MultiXactState->oldestMultiXactDB;
2018 LWLockRelease(MultiXactGenLock);
2019
2020 /* Clean up offsets state */
2021 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
2022
2023 /*
2024 * (Re-)Initialize our idea of the latest page number for offsets.
2025 */
2026 pageno = MultiXactIdToOffsetPage(nextMXact);
2027 MultiXactOffsetCtl->shared->latest_page_number = pageno;
2028
2029 /*
2030 * Zero out the remainder of the current offsets page. See notes in
2031 * TrimCLOG() for background. Unlike CLOG, some WAL record covers every
2032 * pg_multixact SLRU mutation. Since, also unlike CLOG, we ignore the WAL
2033 * rule "write xlog before data," nextMXact successors may carry obsolete,
2034 * nonzero offset values. Zero those so case 2 of GetMultiXactIdMembers()
2035 * operates normally.
2036 */
2037 entryno = MultiXactIdToOffsetEntry(nextMXact);
2038 if (entryno != 0)
2039 {
2040 int slotno;
2041 MultiXactOffset *offptr;
2042
2043 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, nextMXact);
2044 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2045 offptr += entryno;
2046
2047 MemSet(offptr, 0, BLCKSZ - (entryno * sizeof(MultiXactOffset)));
2048
2049 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
2050 }
2051
2052 LWLockRelease(MultiXactOffsetControlLock);
2053
2054 /* And the same for members */
2055 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
2056
2057 /*
2058 * (Re-)Initialize our idea of the latest page number for members.
2059 */
2060 pageno = MXOffsetToMemberPage(offset);
2061 MultiXactMemberCtl->shared->latest_page_number = pageno;
2062
2063 /*
2064 * Zero out the remainder of the current members page. See notes in
2065 * TrimCLOG() for motivation.
2066 */
2067 flagsoff = MXOffsetToFlagsOffset(offset);
2068 if (flagsoff != 0)
2069 {
2070 int slotno;
2071 TransactionId *xidptr;
2072 int memberoff;
2073
2074 memberoff = MXOffsetToMemberOffset(offset);
2075 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, offset);
2076 xidptr = (TransactionId *)
2077 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
2078
2079 MemSet(xidptr, 0, BLCKSZ - memberoff);
2080
2081 /*
2082 * Note: we don't need to zero out the flag bits in the remaining
2083 * members of the current group, because they are always reset before
2084 * writing.
2085 */
2086
2087 MultiXactMemberCtl->shared->page_dirty[slotno] = true;
2088 }
2089
2090 LWLockRelease(MultiXactMemberControlLock);
2091
2092 /* signal that we're officially up */
2093 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2094 MultiXactState->finishedStartup = true;
2095 LWLockRelease(MultiXactGenLock);
2096
2097 /* Now compute how far away the next members wraparound is. */
2098 SetMultiXactIdLimit(oldestMXact, oldestMXactDB, true);
2099}
2100
2101/*
2102 * This must be called ONCE during postmaster or standalone-backend shutdown
2103 */
2104void
2105ShutdownMultiXact(void)
2106{
2107 /* Flush dirty MultiXact pages to disk */
2108 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(false);
2109 SimpleLruFlush(MultiXactOffsetCtl, false);
2110 SimpleLruFlush(MultiXactMemberCtl, false);
2111 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(false);
2112}
2113
2114/*
2115 * Get the MultiXact data to save in a checkpoint record
2116 */
2117void
2118MultiXactGetCheckptMulti(bool is_shutdown,
2119 MultiXactId *nextMulti,
2120 MultiXactOffset *nextMultiOffset,
2121 MultiXactId *oldestMulti,
2122 Oid *oldestMultiDB)
2123{
2124 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2125 *nextMulti = MultiXactState->nextMXact;
2126 *nextMultiOffset = MultiXactState->nextOffset;
2127 *oldestMulti = MultiXactState->oldestMultiXactId;
2128 *oldestMultiDB = MultiXactState->oldestMultiXactDB;
2129 LWLockRelease(MultiXactGenLock);
2130
2131 debug_elog6(DEBUG2,
2132 "MultiXact: checkpoint is nextMulti %u, nextOffset %u, oldestMulti %u in DB %u",
2133 *nextMulti, *nextMultiOffset, *oldestMulti, *oldestMultiDB);
2134}
2135
2136/*
2137 * Perform a checkpoint --- either during shutdown, or on-the-fly
2138 */
2139void
2140CheckPointMultiXact(void)
2141{
2142 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(true);
2143
2144 /* Flush dirty MultiXact pages to disk */
2145 SimpleLruFlush(MultiXactOffsetCtl, true);
2146 SimpleLruFlush(MultiXactMemberCtl, true);
2147
2148 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(true);
2149}
2150
2151/*
2152 * Set the next-to-be-assigned MultiXactId and offset
2153 *
2154 * This is used when we can determine the correct next ID/offset exactly
2155 * from a checkpoint record. Although this is only called during bootstrap
2156 * and XLog replay, we take the lock in case any hot-standby backends are
2157 * examining the values.
2158 */
2159void
2160MultiXactSetNextMXact(MultiXactId nextMulti,
2161 MultiXactOffset nextMultiOffset)
2162{
2163 debug_elog4(DEBUG2, "MultiXact: setting next multi to %u offset %u",
2164 nextMulti, nextMultiOffset);
2165 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2166 MultiXactState->nextMXact = nextMulti;
2167 MultiXactState->nextOffset = nextMultiOffset;
2168 LWLockRelease(MultiXactGenLock);
2169
2170 /*
2171 * During a binary upgrade, make sure that the offsets SLRU is large
2172 * enough to contain the next value that would be created.
2173 *
2174 * We need to do this pretty early during the first startup in binary
2175 * upgrade mode: before StartupMultiXact() in fact, because this routine
2176 * is called even before that by StartupXLOG(). And we can't do it
2177 * earlier than at this point, because during that first call of this
2178 * routine we determine the MultiXactState->nextMXact value that
2179 * MaybeExtendOffsetSlru needs.
2180 */
2181 if (IsBinaryUpgrade)
2182 MaybeExtendOffsetSlru();
2183}
2184
2185/*
2186 * Determine the last safe MultiXactId to allocate given the currently oldest
2187 * datminmxid (ie, the oldest MultiXactId that might exist in any database
2188 * of our cluster), and the OID of the (or a) database with that value.
2189 *
2190 * is_startup is true when we are just starting the cluster, false when we
2191 * are updating state in a running cluster. This only affects log messages.
2192 */
2193void
2194SetMultiXactIdLimit(MultiXactId oldest_datminmxid, Oid oldest_datoid,
2195 bool is_startup)
2196{
2197 MultiXactId multiVacLimit;
2198 MultiXactId multiWarnLimit;
2199 MultiXactId multiStopLimit;
2200 MultiXactId multiWrapLimit;
2201 MultiXactId curMulti;
2202 bool needs_offset_vacuum;
2203
2204 Assert(MultiXactIdIsValid(oldest_datminmxid));
2205
2206 /*
2207 * We pretend that a wrap will happen halfway through the multixact ID
2208 * space, but that's not really true, because multixacts wrap differently
2209 * from transaction IDs. Note that, separately from any concern about
2210 * multixact IDs wrapping, we must ensure that multixact members do not
2211 * wrap. Limits for that are set in SetOffsetVacuumLimit, not here.
2212 */
2213 multiWrapLimit = oldest_datminmxid + (MaxMultiXactId >> 1);
2214 if (multiWrapLimit < FirstMultiXactId)
2215 multiWrapLimit += FirstMultiXactId;
2216
2217 /*
2218 * We'll refuse to continue assigning MultiXactIds once we get within 100
2219 * multi of data loss.
2220 *
2221 * Note: This differs from the magic number used in
2222 * SetTransactionIdLimit() since vacuum itself will never generate new
2223 * multis. XXX actually it does, if it needs to freeze old multis.
2224 */
2225 multiStopLimit = multiWrapLimit - 100;
2226 if (multiStopLimit < FirstMultiXactId)
2227 multiStopLimit -= FirstMultiXactId;
2228
2229 /*
2230 * We'll start complaining loudly when we get within 10M multis of the
2231 * stop point. This is kind of arbitrary, but if you let your gas gauge
2232 * get down to 1% of full, would you be looking for the next gas station?
2233 * We need to be fairly liberal about this number because there are lots
2234 * of scenarios where most transactions are done by automatic clients that
2235 * won't pay attention to warnings. (No, we're not gonna make this
2236 * configurable. If you know enough to configure it, you know enough to
2237 * not get in this kind of trouble in the first place.)
2238 */
2239 multiWarnLimit = multiStopLimit - 10000000;
2240 if (multiWarnLimit < FirstMultiXactId)
2241 multiWarnLimit -= FirstMultiXactId;
2242
2243 /*
2244 * We'll start trying to force autovacuums when oldest_datminmxid gets to
2245 * be more than autovacuum_multixact_freeze_max_age mxids old.
2246 *
2247 * Note: autovacuum_multixact_freeze_max_age is a PGC_POSTMASTER parameter
2248 * so that we don't have to worry about dealing with on-the-fly changes in
2249 * its value. See SetTransactionIdLimit.
2250 */
2251 multiVacLimit = oldest_datminmxid + autovacuum_multixact_freeze_max_age;
2252 if (multiVacLimit < FirstMultiXactId)
2253 multiVacLimit += FirstMultiXactId;
2254
2255 /* Grab lock for just long enough to set the new limit values */
2256 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2257 MultiXactState->oldestMultiXactId = oldest_datminmxid;
2258 MultiXactState->oldestMultiXactDB = oldest_datoid;
2259 MultiXactState->multiVacLimit = multiVacLimit;
2260 MultiXactState->multiWarnLimit = multiWarnLimit;
2261 MultiXactState->multiStopLimit = multiStopLimit;
2262 MultiXactState->multiWrapLimit = multiWrapLimit;
2263 curMulti = MultiXactState->nextMXact;
2264 LWLockRelease(MultiXactGenLock);
2265
2266 /* Log the info */
2267 ereport(DEBUG1,
2268 (errmsg("MultiXactId wrap limit is %u, limited by database with OID %u",
2269 multiWrapLimit, oldest_datoid)));
2270
2271 /*
2272 * Computing the actual limits is only possible once the data directory is
2273 * in a consistent state. There's no need to compute the limits while
2274 * still replaying WAL - no decisions about new multis are made even
2275 * though multixact creations might be replayed. So we'll only do further
2276 * checks after TrimMultiXact() has been called.
2277 */
2278 if (!MultiXactState->finishedStartup)
2279 return;
2280
2281 Assert(!InRecovery);
2282
2283 /* Set limits for offset vacuum. */
2284 needs_offset_vacuum = SetOffsetVacuumLimit(is_startup);
2285
2286 /*
2287 * If past the autovacuum force point, immediately signal an autovac
2288 * request. The reason for this is that autovac only processes one
2289 * database per invocation. Once it's finished cleaning up the oldest
2290 * database, it'll call here, and we'll signal the postmaster to start
2291 * another iteration immediately if there are still any old databases.
2292 */
2293 if ((MultiXactIdPrecedes(multiVacLimit, curMulti) ||
2294 needs_offset_vacuum) && IsUnderPostmaster)
2295 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
2296
2297 /* Give an immediate warning if past the wrap warn point */
2298 if (MultiXactIdPrecedes(multiWarnLimit, curMulti))
2299 {
2300 char *oldest_datname;
2301
2302 /*
2303 * We can be called when not inside a transaction, for example during
2304 * StartupXLOG(). In such a case we cannot do database access, so we
2305 * must just report the oldest DB's OID.
2306 *
2307 * Note: it's also possible that get_database_name fails and returns
2308 * NULL, for example because the database just got dropped. We'll
2309 * still warn, even though the warning might now be unnecessary.
2310 */
2311 if (IsTransactionState())
2312 oldest_datname = get_database_name(oldest_datoid);
2313 else
2314 oldest_datname = NULL;
2315
2316 if (oldest_datname)
2317 ereport(WARNING,
2318 (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
2319 "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
2320 multiWrapLimit - curMulti,
2321 oldest_datname,
2322 multiWrapLimit - curMulti),
2323 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
2324 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
2325 else
2326 ereport(WARNING,
2327 (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
2328 "database with OID %u must be vacuumed before %u more MultiXactIds are used",
2329 multiWrapLimit - curMulti,
2330 oldest_datoid,
2331 multiWrapLimit - curMulti),
2332 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
2333 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
2334 }
2335}
2336
2337/*
2338 * Ensure the next-to-be-assigned MultiXactId is at least minMulti,
2339 * and similarly nextOffset is at least minMultiOffset.
2340 *
2341 * This is used when we can determine minimum safe values from an XLog
2342 * record (either an on-line checkpoint or an mxact creation log entry).
2343 * Although this is only called during XLog replay, we take the lock in case
2344 * any hot-standby backends are examining the values.
2345 */
2346void
2347MultiXactAdvanceNextMXact(MultiXactId minMulti,
2348 MultiXactOffset minMultiOffset)
2349{
2350 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2351 if (MultiXactIdPrecedes(MultiXactState->nextMXact, minMulti))
2352 {
2353 debug_elog3(DEBUG2, "MultiXact: setting next multi to %u", minMulti);
2354 MultiXactState->nextMXact = minMulti;
2355 }
2356 if (MultiXactOffsetPrecedes(MultiXactState->nextOffset, minMultiOffset))
2357 {
2358 debug_elog3(DEBUG2, "MultiXact: setting next offset to %u",
2359 minMultiOffset);
2360 MultiXactState->nextOffset = minMultiOffset;
2361 }
2362 LWLockRelease(MultiXactGenLock);
2363}
2364
2365/*
2366 * Update our oldestMultiXactId value, but only if it's more recent than what
2367 * we had.
2368 *
2369 * This may only be called during WAL replay.
2370 */
2371void
2372MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB)
2373{
2374 Assert(InRecovery);
2375
2376 if (MultiXactIdPrecedes(MultiXactState->oldestMultiXactId, oldestMulti))
2377 SetMultiXactIdLimit(oldestMulti, oldestMultiDB, false);
2378}
2379
2380/*
2381 * Make sure that MultiXactOffset has room for a newly-allocated MultiXactId.
2382 *
2383 * NB: this is called while holding MultiXactGenLock. We want it to be very
2384 * fast most of the time; even when it's not so fast, no actual I/O need
2385 * happen unless we're forced to write out a dirty log or xlog page to make
2386 * room in shared memory.
2387 */
2388static void
2389ExtendMultiXactOffset(MultiXactId multi)
2390{
2391 int pageno;
2392
2393 /*
2394 * No work except at first MultiXactId of a page. But beware: just after
2395 * wraparound, the first MultiXactId of page zero is FirstMultiXactId.
2396 */
2397 if (MultiXactIdToOffsetEntry(multi) != 0 &&
2398 multi != FirstMultiXactId)
2399 return;
2400
2401 pageno = MultiXactIdToOffsetPage(multi);
2402
2403 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
2404
2405 /* Zero the page and make an XLOG entry about it */
2406 ZeroMultiXactOffsetPage(pageno, true);
2407
2408 LWLockRelease(MultiXactOffsetControlLock);
2409}
2410
2411/*
2412 * Make sure that MultiXactMember has room for the members of a newly-
2413 * allocated MultiXactId.
2414 *
2415 * Like the above routine, this is called while holding MultiXactGenLock;
2416 * same comments apply.
2417 */
2418static void
2419ExtendMultiXactMember(MultiXactOffset offset, int nmembers)
2420{
2421 /*
2422 * It's possible that the members span more than one page of the members
2423 * file, so we loop to ensure we consider each page. The coding is not
2424 * optimal if the members span several pages, but that seems unusual
2425 * enough to not worry much about.
2426 */
2427 while (nmembers > 0)
2428 {
2429 int flagsoff;
2430 int flagsbit;
2431 uint32 difference;
2432
2433 /*
2434 * Only zero when at first entry of a page.
2435 */
2436 flagsoff = MXOffsetToFlagsOffset(offset);
2437 flagsbit = MXOffsetToFlagsBitShift(offset);
2438 if (flagsoff == 0 && flagsbit == 0)
2439 {
2440 int pageno;
2441
2442 pageno = MXOffsetToMemberPage(offset);
2443
2444 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
2445
2446 /* Zero the page and make an XLOG entry about it */
2447 ZeroMultiXactMemberPage(pageno, true);
2448
2449 LWLockRelease(MultiXactMemberControlLock);
2450 }
2451
2452 /*
2453 * Compute the number of items till end of current page. Careful: if
2454 * addition of unsigned ints wraps around, we're at the last page of
2455 * the last segment; since that page holds a different number of items
2456 * than other pages, we need to do it differently.
2457 */
2458 if (offset + MAX_MEMBERS_IN_LAST_MEMBERS_PAGE < offset)
2459 {
2460 /*
2461 * This is the last page of the last segment; we can compute the
2462 * number of items left to allocate in it without modulo
2463 * arithmetic.
2464 */
2465 difference = MaxMultiXactOffset - offset + 1;
2466 }
2467 else
2468 difference = MULTIXACT_MEMBERS_PER_PAGE - offset % MULTIXACT_MEMBERS_PER_PAGE;
2469
2470 /*
2471 * Advance to next page, taking care to properly handle the wraparound
2472 * case. OK if nmembers goes negative.
2473 */
2474 nmembers -= difference;
2475 offset += difference;
2476 }
2477}
2478
2479/*
2480 * GetOldestMultiXactId
2481 *
2482 * Return the oldest MultiXactId that's still possibly still seen as live by
2483 * any running transaction. Older ones might still exist on disk, but they no
2484 * longer have any running member transaction.
2485 *
2486 * It's not safe to truncate MultiXact SLRU segments on the value returned by
2487 * this function; however, it can be used by a full-table vacuum to set the
2488 * point at which it will be possible to truncate SLRU for that table.
2489 */
2490MultiXactId
2491GetOldestMultiXactId(void)
2492{
2493 MultiXactId oldestMXact;
2494 MultiXactId nextMXact;
2495 int i;
2496
2497 /*
2498 * This is the oldest valid value among all the OldestMemberMXactId[] and
2499 * OldestVisibleMXactId[] entries, or nextMXact if none are valid.
2500 */
2501 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2502
2503 /*
2504 * We have to beware of the possibility that nextMXact is in the
2505 * wrapped-around state. We don't fix the counter itself here, but we
2506 * must be sure to use a valid value in our calculation.
2507 */
2508 nextMXact = MultiXactState->nextMXact;
2509 if (nextMXact < FirstMultiXactId)
2510 nextMXact = FirstMultiXactId;
2511
2512 oldestMXact = nextMXact;
2513 for (i = 1; i <= MaxOldestSlot; i++)
2514 {
2515 MultiXactId thisoldest;
2516
2517 thisoldest = OldestMemberMXactId[i];
2518 if (MultiXactIdIsValid(thisoldest) &&
2519 MultiXactIdPrecedes(thisoldest, oldestMXact))
2520 oldestMXact = thisoldest;
2521 thisoldest = OldestVisibleMXactId[i];
2522 if (MultiXactIdIsValid(thisoldest) &&
2523 MultiXactIdPrecedes(thisoldest, oldestMXact))
2524 oldestMXact = thisoldest;
2525 }
2526
2527 LWLockRelease(MultiXactGenLock);
2528
2529 return oldestMXact;
2530}
2531
2532/*
2533 * Determine how aggressively we need to vacuum in order to prevent member
2534 * wraparound.
2535 *
2536 * To do so determine what's the oldest member offset and install the limit
2537 * info in MultiXactState, where it can be used to prevent overrun of old data
2538 * in the members SLRU area.
2539 *
2540 * The return value is true if emergency autovacuum is required and false
2541 * otherwise.
2542 */
2543static bool
2544SetOffsetVacuumLimit(bool is_startup)
2545{
2546 MultiXactId oldestMultiXactId;
2547 MultiXactId nextMXact;
2548 MultiXactOffset oldestOffset = 0; /* placate compiler */
2549 MultiXactOffset prevOldestOffset;
2550 MultiXactOffset nextOffset;
2551 bool oldestOffsetKnown = false;
2552 bool prevOldestOffsetKnown;
2553 MultiXactOffset offsetStopLimit = 0;
2554 MultiXactOffset prevOffsetStopLimit;
2555
2556 /*
2557 * NB: Have to prevent concurrent truncation, we might otherwise try to
2558 * lookup an oldestMulti that's concurrently getting truncated away.
2559 */
2560 LWLockAcquire(MultiXactTruncationLock, LW_SHARED);
2561
2562 /* Read relevant fields from shared memory. */
2563 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2564 oldestMultiXactId = MultiXactState->oldestMultiXactId;
2565 nextMXact = MultiXactState->nextMXact;
2566 nextOffset = MultiXactState->nextOffset;
2567 prevOldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2568 prevOldestOffset = MultiXactState->oldestOffset;
2569 prevOffsetStopLimit = MultiXactState->offsetStopLimit;
2570 Assert(MultiXactState->finishedStartup);
2571 LWLockRelease(MultiXactGenLock);
2572
2573 /*
2574 * Determine the offset of the oldest multixact. Normally, we can read
2575 * the offset from the multixact itself, but there's an important special
2576 * case: if there are no multixacts in existence at all, oldestMXact
2577 * obviously can't point to one. It will instead point to the multixact
2578 * ID that will be assigned the next time one is needed.
2579 */
2580 if (oldestMultiXactId == nextMXact)
2581 {
2582 /*
2583 * When the next multixact gets created, it will be stored at the next
2584 * offset.
2585 */
2586 oldestOffset = nextOffset;
2587 oldestOffsetKnown = true;
2588 }
2589 else
2590 {
2591 /*
2592 * Figure out where the oldest existing multixact's offsets are
2593 * stored. Due to bugs in early release of PostgreSQL 9.3.X and 9.4.X,
2594 * the supposedly-earliest multixact might not really exist. We are
2595 * careful not to fail in that case.
2596 */
2597 oldestOffsetKnown =
2598 find_multixact_start(oldestMultiXactId, &oldestOffset);
2599
2600 if (oldestOffsetKnown)
2601 ereport(DEBUG1,
2602 (errmsg("oldest MultiXactId member is at offset %u",
2603 oldestOffset)));
2604 else
2605 ereport(LOG,
2606 (errmsg("MultiXact member wraparound protections are disabled because oldest checkpointed MultiXact %u does not exist on disk",
2607 oldestMultiXactId)));
2608 }
2609
2610 LWLockRelease(MultiXactTruncationLock);
2611
2612 /*
2613 * If we can, compute limits (and install them MultiXactState) to prevent
2614 * overrun of old data in the members SLRU area. We can only do so if the
2615 * oldest offset is known though.
2616 */
2617 if (oldestOffsetKnown)
2618 {
2619 /* move back to start of the corresponding segment */
2620 offsetStopLimit = oldestOffset - (oldestOffset %
2621 (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT));
2622
2623 /* always leave one segment before the wraparound point */
2624 offsetStopLimit -= (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT);
2625
2626 if (!prevOldestOffsetKnown && !is_startup)
2627 ereport(LOG,
2628 (errmsg("MultiXact member wraparound protections are now enabled")));
2629
2630 ereport(DEBUG1,
2631 (errmsg("MultiXact member stop limit is now %u based on MultiXact %u",
2632 offsetStopLimit, oldestMultiXactId)));
2633 }
2634 else if (prevOldestOffsetKnown)
2635 {
2636 /*
2637 * If we failed to get the oldest offset this time, but we have a
2638 * value from a previous pass through this function, use the old
2639 * values rather than automatically forcing an emergency autovacuum
2640 * cycle again.
2641 */
2642 oldestOffset = prevOldestOffset;
2643 oldestOffsetKnown = true;
2644 offsetStopLimit = prevOffsetStopLimit;
2645 }
2646
2647 /* Install the computed values */
2648 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2649 MultiXactState->oldestOffset = oldestOffset;
2650 MultiXactState->oldestOffsetKnown = oldestOffsetKnown;
2651 MultiXactState->offsetStopLimit = offsetStopLimit;
2652 LWLockRelease(MultiXactGenLock);
2653
2654 /*
2655 * Do we need an emergency autovacuum? If we're not sure, assume yes.
2656 */
2657 return !oldestOffsetKnown ||
2658 (nextOffset - oldestOffset > MULTIXACT_MEMBER_SAFE_THRESHOLD);
2659}
2660
2661/*
2662 * Return whether adding "distance" to "start" would move past "boundary".
2663 *
2664 * We use this to determine whether the addition is "wrapping around" the
2665 * boundary point, hence the name. The reason we don't want to use the regular
2666 * 2^31-modulo arithmetic here is that we want to be able to use the whole of
2667 * the 2^32-1 space here, allowing for more multixacts that would fit
2668 * otherwise.
2669 */
2670static bool
2671MultiXactOffsetWouldWrap(MultiXactOffset boundary, MultiXactOffset start,
2672 uint32 distance)
2673{
2674 MultiXactOffset finish;
2675
2676 /*
2677 * Note that offset number 0 is not used (see GetMultiXactIdMembers), so
2678 * if the addition wraps around the UINT_MAX boundary, skip that value.
2679 */
2680 finish = start + distance;
2681 if (finish < start)
2682 finish++;
2683
2684 /*-----------------------------------------------------------------------
2685 * When the boundary is numerically greater than the starting point, any
2686 * value numerically between the two is not wrapped:
2687 *
2688 * <----S----B---->
2689 * [---) = F wrapped past B (and UINT_MAX)
2690 * [---) = F not wrapped
2691 * [----] = F wrapped past B
2692 *
2693 * When the boundary is numerically less than the starting point (i.e. the
2694 * UINT_MAX wraparound occurs somewhere in between) then all values in
2695 * between are wrapped:
2696 *
2697 * <----B----S---->
2698 * [---) = F not wrapped past B (but wrapped past UINT_MAX)
2699 * [---) = F wrapped past B (and UINT_MAX)
2700 * [----] = F not wrapped
2701 *-----------------------------------------------------------------------
2702 */
2703 if (start < boundary)
2704 return finish >= boundary || finish < start;
2705 else
2706 return finish >= boundary && finish < start;
2707}
2708
2709/*
2710 * Find the starting offset of the given MultiXactId.
2711 *
2712 * Returns false if the file containing the multi does not exist on disk.
2713 * Otherwise, returns true and sets *result to the starting member offset.
2714 *
2715 * This function does not prevent concurrent truncation, so if that's
2716 * required, the caller has to protect against that.
2717 */
2718static bool
2719find_multixact_start(MultiXactId multi, MultiXactOffset *result)
2720{
2721 MultiXactOffset offset;
2722 int pageno;
2723 int entryno;
2724 int slotno;
2725 MultiXactOffset *offptr;
2726
2727 Assert(MultiXactState->finishedStartup);
2728
2729 pageno = MultiXactIdToOffsetPage(multi);
2730 entryno = MultiXactIdToOffsetEntry(multi);
2731
2732 /*
2733 * Flush out dirty data, so PhysicalPageExists can work correctly.
2734 * SimpleLruFlush() is a pretty big hammer for that. Alternatively we
2735 * could add an in-memory version of page exists, but find_multixact_start
2736 * is called infrequently, and it doesn't seem bad to flush buffers to
2737 * disk before truncation.
2738 */
2739 SimpleLruFlush(MultiXactOffsetCtl, true);
2740 SimpleLruFlush(MultiXactMemberCtl, true);
2741
2742 if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
2743 return false;
2744
2745 /* lock is acquired by SimpleLruReadPage_ReadOnly */
2746 slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
2747 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2748 offptr += entryno;
2749 offset = *offptr;
2750 LWLockRelease(MultiXactOffsetControlLock);
2751
2752 *result = offset;
2753 return true;
2754}
2755
2756/*
2757 * Determine how many multixacts, and how many multixact members, currently
2758 * exist. Return false if unable to determine.
2759 */
2760static bool
2761ReadMultiXactCounts(uint32 *multixacts, MultiXactOffset *members)
2762{
2763 MultiXactOffset nextOffset;
2764 MultiXactOffset oldestOffset;
2765 MultiXactId oldestMultiXactId;
2766 MultiXactId nextMultiXactId;
2767 bool oldestOffsetKnown;
2768
2769 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2770 nextOffset = MultiXactState->nextOffset;
2771 oldestMultiXactId = MultiXactState->oldestMultiXactId;
2772 nextMultiXactId = MultiXactState->nextMXact;
2773 oldestOffset = MultiXactState->oldestOffset;
2774 oldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2775 LWLockRelease(MultiXactGenLock);
2776
2777 if (!oldestOffsetKnown)
2778 return false;
2779
2780 *members = nextOffset - oldestOffset;
2781 *multixacts = nextMultiXactId - oldestMultiXactId;
2782 return true;
2783}
2784
2785/*
2786 * Multixact members can be removed once the multixacts that refer to them
2787 * are older than every datminxmid. autovacuum_multixact_freeze_max_age and
2788 * vacuum_multixact_freeze_table_age work together to make sure we never have
2789 * too many multixacts; we hope that, at least under normal circumstances,
2790 * this will also be sufficient to keep us from using too many offsets.
2791 * However, if the average multixact has many members, we might exhaust the
2792 * members space while still using few enough members that these limits fail
2793 * to trigger full table scans for relminmxid advancement. At that point,
2794 * we'd have no choice but to start failing multixact-creating operations
2795 * with an error.
2796 *
2797 * To prevent that, if more than a threshold portion of the members space is
2798 * used, we effectively reduce autovacuum_multixact_freeze_max_age and
2799 * to a value just less than the number of multixacts in use. We hope that
2800 * this will quickly trigger autovacuuming on the table or tables with the
2801 * oldest relminmxid, thus allowing datminmxid values to advance and removing
2802 * some members.
2803 *
2804 * As the fraction of the member space currently in use grows, we become
2805 * more aggressive in clamping this value. That not only causes autovacuum
2806 * to ramp up, but also makes any manual vacuums the user issues more
2807 * aggressive. This happens because vacuum_set_xid_limits() clamps the
2808 * freeze table and the minimum freeze age based on the effective
2809 * autovacuum_multixact_freeze_max_age this function returns. In the worst
2810 * case, we'll claim the freeze_max_age to zero, and every vacuum of any
2811 * table will try to freeze every multixact.
2812 *
2813 * It's possible that these thresholds should be user-tunable, but for now
2814 * we keep it simple.
2815 */
2816int
2817MultiXactMemberFreezeThreshold(void)
2818{
2819 MultiXactOffset members;
2820 uint32 multixacts;
2821 uint32 victim_multixacts;
2822 double fraction;
2823
2824 /* If we can't determine member space utilization, assume the worst. */
2825 if (!ReadMultiXactCounts(&multixacts, &members))
2826 return 0;
2827
2828 /* If member space utilization is low, no special action is required. */
2829 if (members <= MULTIXACT_MEMBER_SAFE_THRESHOLD)
2830 return autovacuum_multixact_freeze_max_age;
2831
2832 /*
2833 * Compute a target for relminmxid advancement. The number of multixacts
2834 * we try to eliminate from the system is based on how far we are past
2835 * MULTIXACT_MEMBER_SAFE_THRESHOLD.
2836 */
2837 fraction = (double) (members - MULTIXACT_MEMBER_SAFE_THRESHOLD) /
2838 (MULTIXACT_MEMBER_DANGER_THRESHOLD - MULTIXACT_MEMBER_SAFE_THRESHOLD);
2839 victim_multixacts = multixacts * fraction;
2840
2841 /* fraction could be > 1.0, but lowest possible freeze age is zero */
2842 if (victim_multixacts > multixacts)
2843 return 0;
2844 return multixacts - victim_multixacts;
2845}
2846
2847typedef struct mxtruncinfo
2848{
2849 int earliestExistingPage;
2850} mxtruncinfo;
2851
2852/*
2853 * SlruScanDirectory callback
2854 * This callback determines the earliest existing page number.
2855 */
2856static bool
2857SlruScanDirCbFindEarliest(SlruCtl ctl, char *filename, int segpage, void *data)
2858{
2859 mxtruncinfo *trunc = (mxtruncinfo *) data;
2860
2861 if (trunc->earliestExistingPage == -1 ||
2862 ctl->PagePrecedes(segpage, trunc->earliestExistingPage))
2863 {
2864 trunc->earliestExistingPage = segpage;
2865 }
2866
2867 return false; /* keep going */
2868}
2869
2870
2871/*
2872 * Delete members segments [oldest, newOldest)
2873 *
2874 * The members SLRU can, in contrast to the offsets one, be filled to almost
2875 * the full range at once. This means SimpleLruTruncate() can't trivially be
2876 * used - instead the to-be-deleted range is computed using the offsets
2877 * SLRU. C.f. TruncateMultiXact().
2878 */
2879static void
2880PerformMembersTruncation(MultiXactOffset oldestOffset, MultiXactOffset newOldestOffset)
2881{
2882 const int maxsegment = MXOffsetToMemberSegment(MaxMultiXactOffset);
2883 int startsegment = MXOffsetToMemberSegment(oldestOffset);
2884 int endsegment = MXOffsetToMemberSegment(newOldestOffset);
2885 int segment = startsegment;
2886
2887 /*
2888 * Delete all the segments but the last one. The last segment can still
2889 * contain, possibly partially, valid data.
2890 */
2891 while (segment != endsegment)
2892 {
2893 elog(DEBUG2, "truncating multixact members segment %x", segment);
2894 SlruDeleteSegment(MultiXactMemberCtl, segment);
2895
2896 /* move to next segment, handling wraparound correctly */
2897 if (segment == maxsegment)
2898 segment = 0;
2899 else
2900 segment += 1;
2901 }
2902}
2903
2904/*
2905 * Delete offsets segments [oldest, newOldest)
2906 */
2907static void
2908PerformOffsetsTruncation(MultiXactId oldestMulti, MultiXactId newOldestMulti)
2909{
2910 /*
2911 * We step back one multixact to avoid passing a cutoff page that hasn't
2912 * been created yet in the rare case that oldestMulti would be the first
2913 * item on a page and oldestMulti == nextMulti. In that case, if we
2914 * didn't subtract one, we'd trigger SimpleLruTruncate's wraparound
2915 * detection.
2916 */
2917 SimpleLruTruncate(MultiXactOffsetCtl,
2918 MultiXactIdToOffsetPage(PreviousMultiXactId(newOldestMulti)));
2919}
2920
2921/*
2922 * Remove all MultiXactOffset and MultiXactMember segments before the oldest
2923 * ones still of interest.
2924 *
2925 * This is only called on a primary as part of vacuum (via
2926 * vac_truncate_clog()). During recovery truncation is done by replaying
2927 * truncation WAL records logged here.
2928 *
2929 * newOldestMulti is the oldest currently required multixact, newOldestMultiDB
2930 * is one of the databases preventing newOldestMulti from increasing.
2931 */
2932void
2933TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
2934{
2935 MultiXactId oldestMulti;
2936 MultiXactId nextMulti;
2937 MultiXactOffset newOldestOffset;
2938 MultiXactOffset oldestOffset;
2939 MultiXactOffset nextOffset;
2940 mxtruncinfo trunc;
2941 MultiXactId earliest;
2942
2943 Assert(!RecoveryInProgress());
2944 Assert(MultiXactState->finishedStartup);
2945
2946 /*
2947 * We can only allow one truncation to happen at once. Otherwise parts of
2948 * members might vanish while we're doing lookups or similar. There's no
2949 * need to have an interlock with creating new multis or such, since those
2950 * are constrained by the limits (which only grow, never shrink).
2951 */
2952 LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
2953
2954 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2955 nextMulti = MultiXactState->nextMXact;
2956 nextOffset = MultiXactState->nextOffset;
2957 oldestMulti = MultiXactState->oldestMultiXactId;
2958 LWLockRelease(MultiXactGenLock);
2959 Assert(MultiXactIdIsValid(oldestMulti));
2960
2961 /*
2962 * Make sure to only attempt truncation if there's values to truncate
2963 * away. In normal processing values shouldn't go backwards, but there's
2964 * some corner cases (due to bugs) where that's possible.
2965 */
2966 if (MultiXactIdPrecedesOrEquals(newOldestMulti, oldestMulti))
2967 {
2968 LWLockRelease(MultiXactTruncationLock);
2969 return;
2970 }
2971
2972 /*
2973 * Note we can't just plow ahead with the truncation; it's possible that
2974 * there are no segments to truncate, which is a problem because we are
2975 * going to attempt to read the offsets page to determine where to
2976 * truncate the members SLRU. So we first scan the directory to determine
2977 * the earliest offsets page number that we can read without error.
2978 *
2979 * NB: It's also possible that the page that oldestMulti is on has already
2980 * been truncated away, and we crashed before updating oldestMulti.
2981 */
2982 trunc.earliestExistingPage = -1;
2983 SlruScanDirectory(MultiXactOffsetCtl, SlruScanDirCbFindEarliest, &trunc);
2984 earliest = trunc.earliestExistingPage * MULTIXACT_OFFSETS_PER_PAGE;
2985 if (earliest < FirstMultiXactId)
2986 earliest = FirstMultiXactId;
2987
2988 /* If there's nothing to remove, we can bail out early. */
2989 if (MultiXactIdPrecedes(oldestMulti, earliest))
2990 {
2991 LWLockRelease(MultiXactTruncationLock);
2992 return;
2993 }
2994
2995 /*
2996 * First, compute the safe truncation point for MultiXactMember. This is
2997 * the starting offset of the oldest multixact.
2998 *
2999 * Hopefully, find_multixact_start will always work here, because we've
3000 * already checked that it doesn't precede the earliest MultiXact on disk.
3001 * But if it fails, don't truncate anything, and log a message.
3002 */
3003 if (oldestMulti == nextMulti)
3004 {
3005 /* there are NO MultiXacts */
3006 oldestOffset = nextOffset;
3007 }
3008 else if (!find_multixact_start(oldestMulti, &oldestOffset))
3009 {
3010 ereport(LOG,
3011 (errmsg("oldest MultiXact %u not found, earliest MultiXact %u, skipping truncation",
3012 oldestMulti, earliest)));
3013 LWLockRelease(MultiXactTruncationLock);
3014 return;
3015 }
3016
3017 /*
3018 * Secondly compute up to where to truncate. Lookup the corresponding
3019 * member offset for newOldestMulti for that.
3020 */
3021 if (newOldestMulti == nextMulti)
3022 {
3023 /* there are NO MultiXacts */
3024 newOldestOffset = nextOffset;
3025 }
3026 else if (!find_multixact_start(newOldestMulti, &newOldestOffset))
3027 {
3028 ereport(LOG,
3029 (errmsg("cannot truncate up to MultiXact %u because it does not exist on disk, skipping truncation",
3030 newOldestMulti)));
3031 LWLockRelease(MultiXactTruncationLock);
3032 return;
3033 }
3034
3035 elog(DEBUG1, "performing multixact truncation: "
3036 "offsets [%u, %u), offsets segments [%x, %x), "
3037 "members [%u, %u), members segments [%x, %x)",
3038 oldestMulti, newOldestMulti,
3039 MultiXactIdToOffsetSegment(oldestMulti),
3040 MultiXactIdToOffsetSegment(newOldestMulti),
3041 oldestOffset, newOldestOffset,
3042 MXOffsetToMemberSegment(oldestOffset),
3043 MXOffsetToMemberSegment(newOldestOffset));
3044
3045 /*
3046 * Do truncation, and the WAL logging of the truncation, in a critical
3047 * section. That way offsets/members cannot get out of sync anymore, i.e.
3048 * once consistent the newOldestMulti will always exist in members, even
3049 * if we crashed in the wrong moment.
3050 */
3051 START_CRIT_SECTION();
3052
3053 /*
3054 * Prevent checkpoints from being scheduled concurrently. This is critical
3055 * because otherwise a truncation record might not be replayed after a
3056 * crash/basebackup, even though the state of the data directory would
3057 * require it.
3058 */
3059 Assert(!MyPgXact->delayChkpt);
3060 MyPgXact->delayChkpt = true;
3061
3062 /* WAL log truncation */
3063 WriteMTruncateXlogRec(newOldestMultiDB,
3064 oldestMulti, newOldestMulti,
3065 oldestOffset, newOldestOffset);
3066
3067 /*
3068 * Update in-memory limits before performing the truncation, while inside
3069 * the critical section: Have to do it before truncation, to prevent
3070 * concurrent lookups of those values. Has to be inside the critical
3071 * section as otherwise a future call to this function would error out,
3072 * while looking up the oldest member in offsets, if our caller crashes
3073 * before updating the limits.
3074 */
3075 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
3076 MultiXactState->oldestMultiXactId = newOldestMulti;
3077 MultiXactState->oldestMultiXactDB = newOldestMultiDB;
3078 LWLockRelease(MultiXactGenLock);
3079
3080 /* First truncate members */
3081 PerformMembersTruncation(oldestOffset, newOldestOffset);
3082
3083 /* Then offsets */
3084 PerformOffsetsTruncation(oldestMulti, newOldestMulti);
3085
3086 MyPgXact->delayChkpt = false;
3087
3088 END_CRIT_SECTION();
3089 LWLockRelease(MultiXactTruncationLock);
3090}
3091
3092/*
3093 * Decide which of two MultiXactOffset page numbers is "older" for truncation
3094 * purposes.
3095 *
3096 * We need to use comparison of MultiXactId here in order to do the right
3097 * thing with wraparound. However, if we are asked about page number zero, we
3098 * don't want to hand InvalidMultiXactId to MultiXactIdPrecedes: it'll get
3099 * weird. So, offset both multis by FirstMultiXactId to avoid that.
3100 * (Actually, the current implementation doesn't do anything weird with
3101 * InvalidMultiXactId, but there's no harm in leaving this code like this.)
3102 */
3103static bool
3104MultiXactOffsetPagePrecedes(int page1, int page2)
3105{
3106 MultiXactId multi1;
3107 MultiXactId multi2;
3108
3109 multi1 = ((MultiXactId) page1) * MULTIXACT_OFFSETS_PER_PAGE;
3110 multi1 += FirstMultiXactId;
3111 multi2 = ((MultiXactId) page2) * MULTIXACT_OFFSETS_PER_PAGE;
3112 multi2 += FirstMultiXactId;
3113
3114 return MultiXactIdPrecedes(multi1, multi2);
3115}
3116
3117/*
3118 * Decide which of two MultiXactMember page numbers is "older" for truncation
3119 * purposes. There is no "invalid offset number" so use the numbers verbatim.
3120 */
3121static bool
3122MultiXactMemberPagePrecedes(int page1, int page2)
3123{
3124 MultiXactOffset offset1;
3125 MultiXactOffset offset2;
3126
3127 offset1 = ((MultiXactOffset) page1) * MULTIXACT_MEMBERS_PER_PAGE;
3128 offset2 = ((MultiXactOffset) page2) * MULTIXACT_MEMBERS_PER_PAGE;
3129
3130 return MultiXactOffsetPrecedes(offset1, offset2);
3131}
3132
3133/*
3134 * Decide which of two MultiXactIds is earlier.
3135 *
3136 * XXX do we need to do something special for InvalidMultiXactId?
3137 * (Doesn't look like it.)
3138 */
3139bool
3140MultiXactIdPrecedes(MultiXactId multi1, MultiXactId multi2)
3141{
3142 int32 diff = (int32) (multi1 - multi2);
3143
3144 return (diff < 0);
3145}
3146
3147/*
3148 * MultiXactIdPrecedesOrEquals -- is multi1 logically <= multi2?
3149 *
3150 * XXX do we need to do something special for InvalidMultiXactId?
3151 * (Doesn't look like it.)
3152 */
3153bool
3154MultiXactIdPrecedesOrEquals(MultiXactId multi1, MultiXactId multi2)
3155{
3156 int32 diff = (int32) (multi1 - multi2);
3157
3158 return (diff <= 0);
3159}
3160
3161
3162/*
3163 * Decide which of two offsets is earlier.
3164 */
3165static bool
3166MultiXactOffsetPrecedes(MultiXactOffset offset1, MultiXactOffset offset2)
3167{
3168 int32 diff = (int32) (offset1 - offset2);
3169
3170 return (diff < 0);
3171}
3172
3173/*
3174 * Write an xlog record reflecting the zeroing of either a MEMBERs or
3175 * OFFSETs page (info shows which)
3176 */
3177static void
3178WriteMZeroPageXlogRec(int pageno, uint8 info)
3179{
3180 XLogBeginInsert();
3181 XLogRegisterData((char *) (&pageno), sizeof(int));
3182 (void) XLogInsert(RM_MULTIXACT_ID, info);
3183}
3184
3185/*
3186 * Write a TRUNCATE xlog record
3187 *
3188 * We must flush the xlog record to disk before returning --- see notes in
3189 * TruncateCLOG().
3190 */
3191static void
3192WriteMTruncateXlogRec(Oid oldestMultiDB,
3193 MultiXactId startTruncOff, MultiXactId endTruncOff,
3194 MultiXactOffset startTruncMemb, MultiXactOffset endTruncMemb)
3195{
3196 XLogRecPtr recptr;
3197 xl_multixact_truncate xlrec;
3198
3199 xlrec.oldestMultiDB = oldestMultiDB;
3200
3201 xlrec.startTruncOff = startTruncOff;
3202 xlrec.endTruncOff = endTruncOff;
3203
3204 xlrec.startTruncMemb = startTruncMemb;
3205 xlrec.endTruncMemb = endTruncMemb;
3206
3207 XLogBeginInsert();
3208 XLogRegisterData((char *) (&xlrec), SizeOfMultiXactTruncate);
3209 recptr = XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_TRUNCATE_ID);
3210 XLogFlush(recptr);
3211}
3212
3213/*
3214 * MULTIXACT resource manager's routines
3215 */
3216void
3217multixact_redo(XLogReaderState *record)
3218{
3219 uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
3220
3221 /* Backup blocks are not used in multixact records */
3222 Assert(!XLogRecHasAnyBlockRefs(record));
3223
3224 if (info == XLOG_MULTIXACT_ZERO_OFF_PAGE)
3225 {
3226 int pageno;
3227 int slotno;
3228
3229 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
3230
3231 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
3232
3233 slotno = ZeroMultiXactOffsetPage(pageno, false);
3234 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
3235 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
3236
3237 LWLockRelease(MultiXactOffsetControlLock);
3238 }
3239 else if (info == XLOG_MULTIXACT_ZERO_MEM_PAGE)
3240 {
3241 int pageno;
3242 int slotno;
3243
3244 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
3245
3246 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
3247
3248 slotno = ZeroMultiXactMemberPage(pageno, false);
3249 SimpleLruWritePage(MultiXactMemberCtl, slotno);
3250 Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
3251
3252 LWLockRelease(MultiXactMemberControlLock);
3253 }
3254 else if (info == XLOG_MULTIXACT_CREATE_ID)
3255 {
3256 xl_multixact_create *xlrec =
3257 (xl_multixact_create *) XLogRecGetData(record);
3258 TransactionId max_xid;
3259 int i;
3260
3261 /* Store the data back into the SLRU files */
3262 RecordNewMultiXact(xlrec->mid, xlrec->moff, xlrec->nmembers,
3263 xlrec->members);
3264
3265 /* Make sure nextMXact/nextOffset are beyond what this record has */
3266 MultiXactAdvanceNextMXact(xlrec->mid + 1,
3267 xlrec->moff + xlrec->nmembers);
3268
3269 /*
3270 * Make sure nextFullXid is beyond any XID mentioned in the record.
3271 * This should be unnecessary, since any XID found here ought to have
3272 * other evidence in the XLOG, but let's be safe.
3273 */
3274 max_xid = XLogRecGetXid(record);
3275 for (i = 0; i < xlrec->nmembers; i++)
3276 {
3277 if (TransactionIdPrecedes(max_xid, xlrec->members[i].xid))
3278 max_xid = xlrec->members[i].xid;
3279 }
3280
3281 AdvanceNextFullTransactionIdPastXid(max_xid);
3282 }
3283 else if (info == XLOG_MULTIXACT_TRUNCATE_ID)
3284 {
3285 xl_multixact_truncate xlrec;
3286 int pageno;
3287
3288 memcpy(&xlrec, XLogRecGetData(record),
3289 SizeOfMultiXactTruncate);
3290
3291 elog(DEBUG1, "replaying multixact truncation: "
3292 "offsets [%u, %u), offsets segments [%x, %x), "
3293 "members [%u, %u), members segments [%x, %x)",
3294 xlrec.startTruncOff, xlrec.endTruncOff,
3295 MultiXactIdToOffsetSegment(xlrec.startTruncOff),
3296 MultiXactIdToOffsetSegment(xlrec.endTruncOff),
3297 xlrec.startTruncMemb, xlrec.endTruncMemb,
3298 MXOffsetToMemberSegment(xlrec.startTruncMemb),
3299 MXOffsetToMemberSegment(xlrec.endTruncMemb));
3300
3301 /* should not be required, but more than cheap enough */
3302 LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
3303
3304 /*
3305 * Advance the horizon values, so they're current at the end of
3306 * recovery.
3307 */
3308 SetMultiXactIdLimit(xlrec.endTruncOff, xlrec.oldestMultiDB, false);
3309
3310 PerformMembersTruncation(xlrec.startTruncMemb, xlrec.endTruncMemb);
3311
3312 /*
3313 * During XLOG replay, latest_page_number isn't necessarily set up
3314 * yet; insert a suitable value to bypass the sanity test in
3315 * SimpleLruTruncate.
3316 */
3317 pageno = MultiXactIdToOffsetPage(xlrec.endTruncOff);
3318 MultiXactOffsetCtl->shared->latest_page_number = pageno;
3319 PerformOffsetsTruncation(xlrec.startTruncOff, xlrec.endTruncOff);
3320
3321 LWLockRelease(MultiXactTruncationLock);
3322 }
3323 else
3324 elog(PANIC, "multixact_redo: unknown op code %u", info);
3325}
3326
3327Datum
3328pg_get_multixact_members(PG_FUNCTION_ARGS)
3329{
3330 typedef struct
3331 {
3332 MultiXactMember *members;
3333 int nmembers;
3334 int iter;
3335 } mxact;
3336 MultiXactId mxid = PG_GETARG_UINT32(0);
3337 mxact *multi;
3338 FuncCallContext *funccxt;
3339
3340 if (mxid < FirstMultiXactId)
3341 ereport(ERROR,
3342 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3343 errmsg("invalid MultiXactId: %u", mxid)));
3344
3345 if (SRF_IS_FIRSTCALL())
3346 {
3347 MemoryContext oldcxt;
3348 TupleDesc tupdesc;
3349
3350 funccxt = SRF_FIRSTCALL_INIT();
3351 oldcxt = MemoryContextSwitchTo(funccxt->multi_call_memory_ctx);
3352
3353 multi = palloc(sizeof(mxact));
3354 /* no need to allow for old values here */
3355 multi->nmembers = GetMultiXactIdMembers(mxid, &multi->members, false,
3356 false);
3357 multi->iter = 0;
3358
3359 tupdesc = CreateTemplateTupleDesc(2);
3360 TupleDescInitEntry(tupdesc, (AttrNumber) 1, "xid",
3361 XIDOID, -1, 0);
3362 TupleDescInitEntry(tupdesc, (AttrNumber) 2, "mode",
3363 TEXTOID, -1, 0);
3364
3365 funccxt->attinmeta = TupleDescGetAttInMetadata(tupdesc);
3366 funccxt->user_fctx = multi;
3367
3368 MemoryContextSwitchTo(oldcxt);
3369 }
3370
3371 funccxt = SRF_PERCALL_SETUP();
3372 multi = (mxact *) funccxt->user_fctx;
3373
3374 while (multi->iter < multi->nmembers)
3375 {
3376 HeapTuple tuple;
3377 char *values[2];
3378
3379 values[0] = psprintf("%u", multi->members[multi->iter].xid);
3380 values[1] = mxstatus_to_string(multi->members[multi->iter].status);
3381
3382 tuple = BuildTupleFromCStrings(funccxt->attinmeta, values);
3383
3384 multi->iter++;
3385 pfree(values[0]);
3386 SRF_RETURN_NEXT(funccxt, HeapTupleGetDatum(tuple));
3387 }
3388
3389 if (multi->nmembers > 0)
3390 pfree(multi->members);
3391 pfree(multi);
3392
3393 SRF_RETURN_DONE(funccxt);
3394}
3395