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) 1996-2020, 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 */
187 static SlruCtlData MultiXactOffsetCtlData;
188 static 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 MultiXactOffsetSLRULock and
196 * MultiXactMemberSLRULock 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 */
200 typedef 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 */
291 static MultiXactStateData *MultiXactState;
292 static MultiXactId *OldestMemberMXactId;
293 static 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 */
313 typedef 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
322 static dlist_head MXactCache = DLIST_STATIC_INIT(MXactCache);
323 static int MXactCacheMembers = 0;
324 static 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 */
341 static void MultiXactIdSetOldestVisible(void);
342 static void RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
343 int nmembers, MultiXactMember *members);
344 static MultiXactId GetNewMultiXactId(int nmembers, MultiXactOffset *offset);
345
346 /* MultiXact cache management */
347 static int mxactMemberComparator(const void *arg1, const void *arg2);
348 static MultiXactId mXactCacheGetBySet(int nmembers, MultiXactMember *members);
349 static int mXactCacheGetById(MultiXactId multi, MultiXactMember **members);
350 static void mXactCachePut(MultiXactId multi, int nmembers,
351 MultiXactMember *members);
352
353 static char *mxstatus_to_string(MultiXactStatus status);
354
355 /* management of SLRU infrastructure */
356 static int ZeroMultiXactOffsetPage(int pageno, bool writeXlog);
357 static int ZeroMultiXactMemberPage(int pageno, bool writeXlog);
358 static bool MultiXactOffsetPagePrecedes(int page1, int page2);
359 static bool MultiXactMemberPagePrecedes(int page1, int page2);
360 static bool MultiXactOffsetPrecedes(MultiXactOffset offset1,
361 MultiXactOffset offset2);
362 static void ExtendMultiXactOffset(MultiXactId multi);
363 static void ExtendMultiXactMember(MultiXactOffset offset, int nmembers);
364 static bool MultiXactOffsetWouldWrap(MultiXactOffset boundary,
365 MultiXactOffset start, uint32 distance);
366 static bool SetOffsetVacuumLimit(bool is_startup);
367 static bool find_multixact_start(MultiXactId multi, MultiXactOffset *result);
368 static void WriteMZeroPageXlogRec(int pageno, uint8 info);
369 static void WriteMTruncateXlogRec(Oid oldestMultiDB,
370 MultiXactId startTruncOff,
371 MultiXactId endTruncOff,
372 MultiXactOffset startTruncMemb,
373 MultiXactOffset endTruncMemb);
374
375
376 /*
377 * MultiXactIdCreate
378 * Construct a MultiXactId representing two TransactionIds.
379 *
380 * The two XIDs must be different, or be requesting different statuses.
381 *
382 * NB - we don't worry about our local MultiXactId cache here, because that
383 * is handled by the lower-level routines.
384 */
385 MultiXactId
MultiXactIdCreate(TransactionId xid1,MultiXactStatus status1,TransactionId xid2,MultiXactStatus status2)386 MultiXactIdCreate(TransactionId xid1, MultiXactStatus status1,
387 TransactionId xid2, MultiXactStatus status2)
388 {
389 MultiXactId newMulti;
390 MultiXactMember members[2];
391
392 AssertArg(TransactionIdIsValid(xid1));
393 AssertArg(TransactionIdIsValid(xid2));
394
395 Assert(!TransactionIdEquals(xid1, xid2) || (status1 != status2));
396
397 /* MultiXactIdSetOldestMember() must have been called already. */
398 Assert(MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]));
399
400 /*
401 * Note: unlike MultiXactIdExpand, we don't bother to check that both XIDs
402 * are still running. In typical usage, xid2 will be our own XID and the
403 * caller just did a check on xid1, so it'd be wasted effort.
404 */
405
406 members[0].xid = xid1;
407 members[0].status = status1;
408 members[1].xid = xid2;
409 members[1].status = status2;
410
411 newMulti = MultiXactIdCreateFromMembers(2, members);
412
413 debug_elog3(DEBUG2, "Create: %s",
414 mxid_to_string(newMulti, 2, members));
415
416 return newMulti;
417 }
418
419 /*
420 * MultiXactIdExpand
421 * Add a TransactionId to a pre-existing MultiXactId.
422 *
423 * If the TransactionId is already a member of the passed MultiXactId with the
424 * same status, just return it as-is.
425 *
426 * Note that we do NOT actually modify the membership of a pre-existing
427 * MultiXactId; instead we create a new one. This is necessary to avoid
428 * a race condition against code trying to wait for one MultiXactId to finish;
429 * see notes in heapam.c.
430 *
431 * NB - we don't worry about our local MultiXactId cache here, because that
432 * is handled by the lower-level routines.
433 *
434 * Note: It is critical that MultiXactIds that come from an old cluster (i.e.
435 * one upgraded by pg_upgrade from a cluster older than this feature) are not
436 * passed in.
437 */
438 MultiXactId
MultiXactIdExpand(MultiXactId multi,TransactionId xid,MultiXactStatus status)439 MultiXactIdExpand(MultiXactId multi, TransactionId xid, MultiXactStatus status)
440 {
441 MultiXactId newMulti;
442 MultiXactMember *members;
443 MultiXactMember *newMembers;
444 int nmembers;
445 int i;
446 int j;
447
448 AssertArg(MultiXactIdIsValid(multi));
449 AssertArg(TransactionIdIsValid(xid));
450
451 /* MultiXactIdSetOldestMember() must have been called already. */
452 Assert(MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]));
453
454 debug_elog5(DEBUG2, "Expand: received multi %u, xid %u status %s",
455 multi, xid, mxstatus_to_string(status));
456
457 /*
458 * Note: we don't allow for old multis here. The reason is that the only
459 * caller of this function does a check that the multixact is no longer
460 * running.
461 */
462 nmembers = GetMultiXactIdMembers(multi, &members, false, false);
463
464 if (nmembers < 0)
465 {
466 MultiXactMember member;
467
468 /*
469 * The MultiXactId is obsolete. This can only happen if all the
470 * MultiXactId members stop running between the caller checking and
471 * passing it to us. It would be better to return that fact to the
472 * caller, but it would complicate the API and it's unlikely to happen
473 * too often, so just deal with it by creating a singleton MultiXact.
474 */
475 member.xid = xid;
476 member.status = status;
477 newMulti = MultiXactIdCreateFromMembers(1, &member);
478
479 debug_elog4(DEBUG2, "Expand: %u has no members, create singleton %u",
480 multi, newMulti);
481 return newMulti;
482 }
483
484 /*
485 * If the TransactionId is already a member of the MultiXactId with the
486 * same status, just return the existing MultiXactId.
487 */
488 for (i = 0; i < nmembers; i++)
489 {
490 if (TransactionIdEquals(members[i].xid, xid) &&
491 (members[i].status == status))
492 {
493 debug_elog4(DEBUG2, "Expand: %u is already a member of %u",
494 xid, multi);
495 pfree(members);
496 return multi;
497 }
498 }
499
500 /*
501 * Determine which of the members of the MultiXactId are still of
502 * interest. This is any running transaction, and also any transaction
503 * that grabbed something stronger than just a lock and was committed. (An
504 * update that aborted is of no interest here; and having more than one
505 * update Xid in a multixact would cause errors elsewhere.)
506 *
507 * Removing dead members is not just an optimization: freezing of tuples
508 * whose Xmax are multis depends on this behavior.
509 *
510 * Note we have the same race condition here as above: j could be 0 at the
511 * end of the loop.
512 */
513 newMembers = (MultiXactMember *)
514 palloc(sizeof(MultiXactMember) * (nmembers + 1));
515
516 for (i = 0, j = 0; i < nmembers; i++)
517 {
518 if (TransactionIdIsInProgress(members[i].xid) ||
519 (ISUPDATE_from_mxstatus(members[i].status) &&
520 TransactionIdDidCommit(members[i].xid)))
521 {
522 newMembers[j].xid = members[i].xid;
523 newMembers[j++].status = members[i].status;
524 }
525 }
526
527 newMembers[j].xid = xid;
528 newMembers[j++].status = status;
529 newMulti = MultiXactIdCreateFromMembers(j, newMembers);
530
531 pfree(members);
532 pfree(newMembers);
533
534 debug_elog3(DEBUG2, "Expand: returning new multi %u", newMulti);
535
536 return newMulti;
537 }
538
539 /*
540 * MultiXactIdIsRunning
541 * Returns whether a MultiXactId is "running".
542 *
543 * We return true if at least one member of the given MultiXactId is still
544 * running. Note that a "false" result is certain not to change,
545 * because it is not legal to add members to an existing MultiXactId.
546 *
547 * Caller is expected to have verified that the multixact does not come from
548 * a pg_upgraded share-locked tuple.
549 */
550 bool
MultiXactIdIsRunning(MultiXactId multi,bool isLockOnly)551 MultiXactIdIsRunning(MultiXactId multi, bool isLockOnly)
552 {
553 MultiXactMember *members;
554 int nmembers;
555 int i;
556
557 debug_elog3(DEBUG2, "IsRunning %u?", multi);
558
559 /*
560 * "false" here means we assume our callers have checked that the given
561 * multi cannot possibly come from a pg_upgraded database.
562 */
563 nmembers = GetMultiXactIdMembers(multi, &members, false, isLockOnly);
564
565 if (nmembers <= 0)
566 {
567 debug_elog2(DEBUG2, "IsRunning: no members");
568 return false;
569 }
570
571 /*
572 * Checking for myself is cheap compared to looking in shared memory;
573 * return true if any live subtransaction of the current top-level
574 * transaction is a member.
575 *
576 * This is not needed for correctness, it's just a fast path.
577 */
578 for (i = 0; i < nmembers; i++)
579 {
580 if (TransactionIdIsCurrentTransactionId(members[i].xid))
581 {
582 debug_elog3(DEBUG2, "IsRunning: I (%d) am running!", i);
583 pfree(members);
584 return true;
585 }
586 }
587
588 /*
589 * This could be made faster by having another entry point in procarray.c,
590 * walking the PGPROC array only once for all the members. But in most
591 * cases nmembers should be small enough that it doesn't much matter.
592 */
593 for (i = 0; i < nmembers; i++)
594 {
595 if (TransactionIdIsInProgress(members[i].xid))
596 {
597 debug_elog4(DEBUG2, "IsRunning: member %d (%u) is running",
598 i, members[i].xid);
599 pfree(members);
600 return true;
601 }
602 }
603
604 pfree(members);
605
606 debug_elog3(DEBUG2, "IsRunning: %u is not running", multi);
607
608 return false;
609 }
610
611 /*
612 * MultiXactIdSetOldestMember
613 * Save the oldest MultiXactId this transaction could be a member of.
614 *
615 * We set the OldestMemberMXactId for a given transaction the first time it's
616 * going to do some operation that might require a MultiXactId (tuple lock,
617 * update or delete). We need to do this even if we end up using a
618 * TransactionId instead of a MultiXactId, because there is a chance that
619 * another transaction would add our XID to a MultiXactId.
620 *
621 * The value to set is the next-to-be-assigned MultiXactId, so this is meant to
622 * be called just before doing any such possibly-MultiXactId-able operation.
623 */
624 void
MultiXactIdSetOldestMember(void)625 MultiXactIdSetOldestMember(void)
626 {
627 if (!MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]))
628 {
629 MultiXactId nextMXact;
630
631 /*
632 * You might think we don't need to acquire a lock here, since
633 * fetching and storing of TransactionIds is probably atomic, but in
634 * fact we do: suppose we pick up nextMXact and then lose the CPU for
635 * a long time. Someone else could advance nextMXact, and then
636 * another someone else could compute an OldestVisibleMXactId that
637 * would be after the value we are going to store when we get control
638 * back. Which would be wrong.
639 *
640 * Note that a shared lock is sufficient, because it's enough to stop
641 * someone from advancing nextMXact; and nobody else could be trying
642 * to write to our OldestMember entry, only reading (and we assume
643 * storing it is atomic.)
644 */
645 LWLockAcquire(MultiXactGenLock, LW_SHARED);
646
647 /*
648 * We have to beware of the possibility that nextMXact is in the
649 * wrapped-around state. We don't fix the counter itself here, but we
650 * must be sure to store a valid value in our array entry.
651 */
652 nextMXact = MultiXactState->nextMXact;
653 if (nextMXact < FirstMultiXactId)
654 nextMXact = FirstMultiXactId;
655
656 OldestMemberMXactId[MyBackendId] = nextMXact;
657
658 LWLockRelease(MultiXactGenLock);
659
660 debug_elog4(DEBUG2, "MultiXact: setting OldestMember[%d] = %u",
661 MyBackendId, nextMXact);
662 }
663 }
664
665 /*
666 * MultiXactIdSetOldestVisible
667 * Save the oldest MultiXactId this transaction considers possibly live.
668 *
669 * We set the OldestVisibleMXactId for a given transaction the first time
670 * it's going to inspect any MultiXactId. Once we have set this, we are
671 * guaranteed that the checkpointer won't truncate off SLRU data for
672 * MultiXactIds at or after our OldestVisibleMXactId.
673 *
674 * The value to set is the oldest of nextMXact and all the valid per-backend
675 * OldestMemberMXactId[] entries. Because of the locking we do, we can be
676 * certain that no subsequent call to MultiXactIdSetOldestMember can set
677 * an OldestMemberMXactId[] entry older than what we compute here. Therefore
678 * there is no live transaction, now or later, that can be a member of any
679 * MultiXactId older than the OldestVisibleMXactId we compute here.
680 */
681 static void
MultiXactIdSetOldestVisible(void)682 MultiXactIdSetOldestVisible(void)
683 {
684 if (!MultiXactIdIsValid(OldestVisibleMXactId[MyBackendId]))
685 {
686 MultiXactId oldestMXact;
687 int i;
688
689 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
690
691 /*
692 * We have to beware of the possibility that nextMXact is in the
693 * wrapped-around state. We don't fix the counter itself here, but we
694 * must be sure to store a valid value in our array entry.
695 */
696 oldestMXact = MultiXactState->nextMXact;
697 if (oldestMXact < FirstMultiXactId)
698 oldestMXact = FirstMultiXactId;
699
700 for (i = 1; i <= MaxOldestSlot; i++)
701 {
702 MultiXactId thisoldest = OldestMemberMXactId[i];
703
704 if (MultiXactIdIsValid(thisoldest) &&
705 MultiXactIdPrecedes(thisoldest, oldestMXact))
706 oldestMXact = thisoldest;
707 }
708
709 OldestVisibleMXactId[MyBackendId] = oldestMXact;
710
711 LWLockRelease(MultiXactGenLock);
712
713 debug_elog4(DEBUG2, "MultiXact: setting OldestVisible[%d] = %u",
714 MyBackendId, oldestMXact);
715 }
716 }
717
718 /*
719 * ReadNextMultiXactId
720 * Return the next MultiXactId to be assigned, but don't allocate it
721 */
722 MultiXactId
ReadNextMultiXactId(void)723 ReadNextMultiXactId(void)
724 {
725 MultiXactId mxid;
726
727 /* XXX we could presumably do this without a lock. */
728 LWLockAcquire(MultiXactGenLock, LW_SHARED);
729 mxid = MultiXactState->nextMXact;
730 LWLockRelease(MultiXactGenLock);
731
732 if (mxid < FirstMultiXactId)
733 mxid = FirstMultiXactId;
734
735 return mxid;
736 }
737
738 /*
739 * MultiXactIdCreateFromMembers
740 * Make a new MultiXactId from the specified set of members
741 *
742 * Make XLOG, SLRU and cache entries for a new MultiXactId, recording the
743 * given TransactionIds as members. Returns the newly created MultiXactId.
744 *
745 * NB: the passed members[] array will be sorted in-place.
746 */
747 MultiXactId
MultiXactIdCreateFromMembers(int nmembers,MultiXactMember * members)748 MultiXactIdCreateFromMembers(int nmembers, MultiXactMember *members)
749 {
750 MultiXactId multi;
751 MultiXactOffset offset;
752 xl_multixact_create xlrec;
753
754 debug_elog3(DEBUG2, "Create: %s",
755 mxid_to_string(InvalidMultiXactId, nmembers, members));
756
757 /*
758 * See if the same set of members already exists in our cache; if so, just
759 * re-use that MultiXactId. (Note: it might seem that looking in our
760 * cache is insufficient, and we ought to search disk to see if a
761 * duplicate definition already exists. But since we only ever create
762 * MultiXacts containing our own XID, in most cases any such MultiXacts
763 * were in fact created by us, and so will be in our cache. There are
764 * corner cases where someone else added us to a MultiXact without our
765 * knowledge, but it's not worth checking for.)
766 */
767 multi = mXactCacheGetBySet(nmembers, members);
768 if (MultiXactIdIsValid(multi))
769 {
770 debug_elog2(DEBUG2, "Create: in cache!");
771 return multi;
772 }
773
774 /* Verify that there is a single update Xid among the given members. */
775 {
776 int i;
777 bool has_update = false;
778
779 for (i = 0; i < nmembers; i++)
780 {
781 if (ISUPDATE_from_mxstatus(members[i].status))
782 {
783 if (has_update)
784 elog(ERROR, "new multixact has more than one updating member");
785 has_update = true;
786 }
787 }
788 }
789
790 /*
791 * Assign the MXID and offsets range to use, and make sure there is space
792 * in the OFFSETs and MEMBERs files. NB: this routine does
793 * START_CRIT_SECTION().
794 *
795 * Note: unlike MultiXactIdCreate and MultiXactIdExpand, we do not check
796 * that we've called MultiXactIdSetOldestMember here. This is because
797 * this routine is used in some places to create new MultiXactIds of which
798 * the current backend is not a member, notably during freezing of multis
799 * in vacuum. During vacuum, in particular, it would be unacceptable to
800 * keep OldestMulti set, in case it runs for long.
801 */
802 multi = GetNewMultiXactId(nmembers, &offset);
803
804 /* Make an XLOG entry describing the new MXID. */
805 xlrec.mid = multi;
806 xlrec.moff = offset;
807 xlrec.nmembers = nmembers;
808
809 /*
810 * XXX Note: there's a lot of padding space in MultiXactMember. We could
811 * find a more compact representation of this Xlog record -- perhaps all
812 * the status flags in one XLogRecData, then all the xids in another one?
813 * Not clear that it's worth the trouble though.
814 */
815 XLogBeginInsert();
816 XLogRegisterData((char *) (&xlrec), SizeOfMultiXactCreate);
817 XLogRegisterData((char *) members, nmembers * sizeof(MultiXactMember));
818
819 (void) XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_CREATE_ID);
820
821 /* Now enter the information into the OFFSETs and MEMBERs logs */
822 RecordNewMultiXact(multi, offset, nmembers, members);
823
824 /* Done with critical section */
825 END_CRIT_SECTION();
826
827 /* Store the new MultiXactId in the local cache, too */
828 mXactCachePut(multi, nmembers, members);
829
830 debug_elog2(DEBUG2, "Create: all done");
831
832 return multi;
833 }
834
835 /*
836 * RecordNewMultiXact
837 * Write info about a new multixact into the offsets and members files
838 *
839 * This is broken out of MultiXactIdCreateFromMembers so that xlog replay can
840 * use it.
841 */
842 static void
RecordNewMultiXact(MultiXactId multi,MultiXactOffset offset,int nmembers,MultiXactMember * members)843 RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
844 int nmembers, MultiXactMember *members)
845 {
846 int pageno;
847 int prev_pageno;
848 int entryno;
849 int slotno;
850 MultiXactOffset *offptr;
851 int i;
852
853 LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
854
855 pageno = MultiXactIdToOffsetPage(multi);
856 entryno = MultiXactIdToOffsetEntry(multi);
857
858 /*
859 * Note: we pass the MultiXactId to SimpleLruReadPage as the "transaction"
860 * to complain about if there's any I/O error. This is kinda bogus, but
861 * since the errors will always give the full pathname, it should be clear
862 * enough that a MultiXactId is really involved. Perhaps someday we'll
863 * take the trouble to generalize the slru.c error reporting code.
864 */
865 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
866 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
867 offptr += entryno;
868
869 *offptr = offset;
870
871 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
872
873 /* Exchange our lock */
874 LWLockRelease(MultiXactOffsetSLRULock);
875
876 LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
877
878 prev_pageno = -1;
879
880 for (i = 0; i < nmembers; i++, offset++)
881 {
882 TransactionId *memberptr;
883 uint32 *flagsptr;
884 uint32 flagsval;
885 int bshift;
886 int flagsoff;
887 int memberoff;
888
889 Assert(members[i].status <= MultiXactStatusUpdate);
890
891 pageno = MXOffsetToMemberPage(offset);
892 memberoff = MXOffsetToMemberOffset(offset);
893 flagsoff = MXOffsetToFlagsOffset(offset);
894 bshift = MXOffsetToFlagsBitShift(offset);
895
896 if (pageno != prev_pageno)
897 {
898 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
899 prev_pageno = pageno;
900 }
901
902 memberptr = (TransactionId *)
903 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
904
905 *memberptr = members[i].xid;
906
907 flagsptr = (uint32 *)
908 (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
909
910 flagsval = *flagsptr;
911 flagsval &= ~(((1 << MXACT_MEMBER_BITS_PER_XACT) - 1) << bshift);
912 flagsval |= (members[i].status << bshift);
913 *flagsptr = flagsval;
914
915 MultiXactMemberCtl->shared->page_dirty[slotno] = true;
916 }
917
918 LWLockRelease(MultiXactMemberSLRULock);
919 }
920
921 /*
922 * GetNewMultiXactId
923 * Get the next MultiXactId.
924 *
925 * Also, reserve the needed amount of space in the "members" area. The
926 * starting offset of the reserved space is returned in *offset.
927 *
928 * This may generate XLOG records for expansion of the offsets and/or members
929 * files. Unfortunately, we have to do that while holding MultiXactGenLock
930 * to avoid race conditions --- the XLOG record for zeroing a page must appear
931 * before any backend can possibly try to store data in that page!
932 *
933 * We start a critical section before advancing the shared counters. The
934 * caller must end the critical section after writing SLRU data.
935 */
936 static MultiXactId
GetNewMultiXactId(int nmembers,MultiXactOffset * offset)937 GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
938 {
939 MultiXactId result;
940 MultiXactOffset nextOffset;
941
942 debug_elog3(DEBUG2, "GetNew: for %d xids", nmembers);
943
944 /* safety check, we should never get this far in a HS standby */
945 if (RecoveryInProgress())
946 elog(ERROR, "cannot assign MultiXactIds during recovery");
947
948 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
949
950 /* Handle wraparound of the nextMXact counter */
951 if (MultiXactState->nextMXact < FirstMultiXactId)
952 MultiXactState->nextMXact = FirstMultiXactId;
953
954 /* Assign the MXID */
955 result = MultiXactState->nextMXact;
956
957 /*----------
958 * Check to see if it's safe to assign another MultiXactId. This protects
959 * against catastrophic data loss due to multixact wraparound. The basic
960 * rules are:
961 *
962 * If we're past multiVacLimit or the safe threshold for member storage
963 * space, or we don't know what the safe threshold for member storage is,
964 * start trying to force autovacuum cycles.
965 * If we're past multiWarnLimit, start issuing warnings.
966 * If we're past multiStopLimit, refuse to create new MultiXactIds.
967 *
968 * Note these are pretty much the same protections in GetNewTransactionId.
969 *----------
970 */
971 if (!MultiXactIdPrecedes(result, MultiXactState->multiVacLimit))
972 {
973 /*
974 * For safety's sake, we release MultiXactGenLock while sending
975 * signals, warnings, etc. This is not so much because we care about
976 * preserving concurrency in this situation, as to avoid any
977 * possibility of deadlock while doing get_database_name(). First,
978 * copy all the shared values we'll need in this path.
979 */
980 MultiXactId multiWarnLimit = MultiXactState->multiWarnLimit;
981 MultiXactId multiStopLimit = MultiXactState->multiStopLimit;
982 MultiXactId multiWrapLimit = MultiXactState->multiWrapLimit;
983 Oid oldest_datoid = MultiXactState->oldestMultiXactDB;
984
985 LWLockRelease(MultiXactGenLock);
986
987 if (IsUnderPostmaster &&
988 !MultiXactIdPrecedes(result, multiStopLimit))
989 {
990 char *oldest_datname = get_database_name(oldest_datoid);
991
992 /*
993 * Immediately kick autovacuum into action as we're already in
994 * ERROR territory.
995 */
996 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
997
998 /* complain even if that DB has disappeared */
999 if (oldest_datname)
1000 ereport(ERROR,
1001 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1002 errmsg("database is not accepting commands that generate new MultiXactIds to avoid wraparound data loss in database \"%s\"",
1003 oldest_datname),
1004 errhint("Execute a database-wide VACUUM in that database.\n"
1005 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1006 else
1007 ereport(ERROR,
1008 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1009 errmsg("database is not accepting commands that generate new MultiXactIds to avoid wraparound data loss in database with OID %u",
1010 oldest_datoid),
1011 errhint("Execute a database-wide VACUUM in that database.\n"
1012 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1013 }
1014
1015 /*
1016 * To avoid swamping the postmaster with signals, we issue the autovac
1017 * request only once per 64K multis generated. This still gives
1018 * plenty of chances before we get into real trouble.
1019 */
1020 if (IsUnderPostmaster && (result % 65536) == 0)
1021 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1022
1023 if (!MultiXactIdPrecedes(result, multiWarnLimit))
1024 {
1025 char *oldest_datname = get_database_name(oldest_datoid);
1026
1027 /* complain even if that DB has disappeared */
1028 if (oldest_datname)
1029 ereport(WARNING,
1030 (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
1031 "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
1032 multiWrapLimit - result,
1033 oldest_datname,
1034 multiWrapLimit - result),
1035 errhint("Execute a database-wide VACUUM in that database.\n"
1036 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1037 else
1038 ereport(WARNING,
1039 (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
1040 "database with OID %u must be vacuumed before %u more MultiXactIds are used",
1041 multiWrapLimit - result,
1042 oldest_datoid,
1043 multiWrapLimit - result),
1044 errhint("Execute a database-wide VACUUM in that database.\n"
1045 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1046 }
1047
1048 /* Re-acquire lock and start over */
1049 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1050 result = MultiXactState->nextMXact;
1051 if (result < FirstMultiXactId)
1052 result = FirstMultiXactId;
1053 }
1054
1055 /* Make sure there is room for the MXID in the file. */
1056 ExtendMultiXactOffset(result);
1057
1058 /*
1059 * Reserve the members space, similarly to above. Also, be careful not to
1060 * return zero as the starting offset for any multixact. See
1061 * GetMultiXactIdMembers() for motivation.
1062 */
1063 nextOffset = MultiXactState->nextOffset;
1064 if (nextOffset == 0)
1065 {
1066 *offset = 1;
1067 nmembers++; /* allocate member slot 0 too */
1068 }
1069 else
1070 *offset = nextOffset;
1071
1072 /*----------
1073 * Protect against overrun of the members space as well, with the
1074 * following rules:
1075 *
1076 * If we're past offsetStopLimit, refuse to generate more multis.
1077 * If we're close to offsetStopLimit, emit a warning.
1078 *
1079 * Arbitrarily, we start emitting warnings when we're 20 segments or less
1080 * from offsetStopLimit.
1081 *
1082 * Note we haven't updated the shared state yet, so if we fail at this
1083 * point, the multixact ID we grabbed can still be used by the next guy.
1084 *
1085 * Note that there is no point in forcing autovacuum runs here: the
1086 * multixact freeze settings would have to be reduced for that to have any
1087 * effect.
1088 *----------
1089 */
1090 #define OFFSET_WARN_SEGMENTS 20
1091 if (MultiXactState->oldestOffsetKnown &&
1092 MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit, nextOffset,
1093 nmembers))
1094 {
1095 /* see comment in the corresponding offsets wraparound case */
1096 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1097
1098 ereport(ERROR,
1099 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1100 errmsg("multixact \"members\" limit exceeded"),
1101 errdetail_plural("This command would create a multixact with %u members, but the remaining space is only enough for %u member.",
1102 "This command would create a multixact with %u members, but the remaining space is only enough for %u members.",
1103 MultiXactState->offsetStopLimit - nextOffset - 1,
1104 nmembers,
1105 MultiXactState->offsetStopLimit - nextOffset - 1),
1106 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.",
1107 MultiXactState->oldestMultiXactDB)));
1108 }
1109
1110 /*
1111 * Check whether we should kick autovacuum into action, to prevent members
1112 * wraparound. NB we use a much larger window to trigger autovacuum than
1113 * just the warning limit. The warning is just a measure of last resort -
1114 * this is in line with GetNewTransactionId's behaviour.
1115 */
1116 if (!MultiXactState->oldestOffsetKnown ||
1117 (MultiXactState->nextOffset - MultiXactState->oldestOffset
1118 > MULTIXACT_MEMBER_SAFE_THRESHOLD))
1119 {
1120 /*
1121 * To avoid swamping the postmaster with signals, we issue the autovac
1122 * request only when crossing a segment boundary. With default
1123 * compilation settings that's roughly after 50k members. This still
1124 * gives plenty of chances before we get into real trouble.
1125 */
1126 if ((MXOffsetToMemberPage(nextOffset) / SLRU_PAGES_PER_SEGMENT) !=
1127 (MXOffsetToMemberPage(nextOffset + nmembers) / SLRU_PAGES_PER_SEGMENT))
1128 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1129 }
1130
1131 if (MultiXactState->oldestOffsetKnown &&
1132 MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit,
1133 nextOffset,
1134 nmembers + MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT * OFFSET_WARN_SEGMENTS))
1135 ereport(WARNING,
1136 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1137 errmsg_plural("database with OID %u must be vacuumed before %d more multixact member is used",
1138 "database with OID %u must be vacuumed before %d more multixact members are used",
1139 MultiXactState->offsetStopLimit - nextOffset + nmembers,
1140 MultiXactState->oldestMultiXactDB,
1141 MultiXactState->offsetStopLimit - nextOffset + nmembers),
1142 errhint("Execute a database-wide VACUUM in that database with reduced vacuum_multixact_freeze_min_age and vacuum_multixact_freeze_table_age settings.")));
1143
1144 ExtendMultiXactMember(nextOffset, nmembers);
1145
1146 /*
1147 * Critical section from here until caller has written the data into the
1148 * just-reserved SLRU space; we don't want to error out with a partly
1149 * written MultiXact structure. (In particular, failing to write our
1150 * start offset after advancing nextMXact would effectively corrupt the
1151 * previous MultiXact.)
1152 */
1153 START_CRIT_SECTION();
1154
1155 /*
1156 * Advance counters. As in GetNewTransactionId(), this must not happen
1157 * until after file extension has succeeded!
1158 *
1159 * We don't care about MultiXactId wraparound here; it will be handled by
1160 * the next iteration. But note that nextMXact may be InvalidMultiXactId
1161 * or the first value on a segment-beginning page after this routine
1162 * exits, so anyone else looking at the variable must be prepared to deal
1163 * with either case. Similarly, nextOffset may be zero, but we won't use
1164 * that as the actual start offset of the next multixact.
1165 */
1166 (MultiXactState->nextMXact)++;
1167
1168 MultiXactState->nextOffset += nmembers;
1169
1170 LWLockRelease(MultiXactGenLock);
1171
1172 debug_elog4(DEBUG2, "GetNew: returning %u offset %u", result, *offset);
1173 return result;
1174 }
1175
1176 /*
1177 * GetMultiXactIdMembers
1178 * Return the set of MultiXactMembers that make up a MultiXactId
1179 *
1180 * Return value is the number of members found, or -1 if there are none,
1181 * and *members is set to a newly palloc'ed array of members. It's the
1182 * caller's responsibility to free it when done with it.
1183 *
1184 * from_pgupgrade must be passed as true if and only if only the multixact
1185 * corresponds to a value from a tuple that was locked in a 9.2-or-older
1186 * installation and later pg_upgrade'd (that is, the infomask is
1187 * HEAP_LOCKED_UPGRADED). In this case, we know for certain that no members
1188 * can still be running, so we return -1 just like for an empty multixact
1189 * without any further checking. It would be wrong to try to resolve such a
1190 * multixact: either the multixact is within the current valid multixact
1191 * range, in which case the returned result would be bogus, or outside that
1192 * range, in which case an error would be raised.
1193 *
1194 * In all other cases, the passed multixact must be within the known valid
1195 * range, that is, greater to or equal than oldestMultiXactId, and less than
1196 * nextMXact. Otherwise, an error is raised.
1197 *
1198 * onlyLock must be set to true if caller is certain that the given multi
1199 * is used only to lock tuples; can be false without loss of correctness,
1200 * but passing a true means we can return quickly without checking for
1201 * old updates.
1202 */
1203 int
GetMultiXactIdMembers(MultiXactId multi,MultiXactMember ** members,bool from_pgupgrade,bool onlyLock)1204 GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
1205 bool from_pgupgrade, bool onlyLock)
1206 {
1207 int pageno;
1208 int prev_pageno;
1209 int entryno;
1210 int slotno;
1211 MultiXactOffset *offptr;
1212 MultiXactOffset offset;
1213 int length;
1214 int truelength;
1215 int i;
1216 MultiXactId oldestMXact;
1217 MultiXactId nextMXact;
1218 MultiXactId tmpMXact;
1219 MultiXactOffset nextOffset;
1220 MultiXactMember *ptr;
1221
1222 debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
1223
1224 if (!MultiXactIdIsValid(multi) || from_pgupgrade)
1225 {
1226 *members = NULL;
1227 return -1;
1228 }
1229
1230 /* See if the MultiXactId is in the local cache */
1231 length = mXactCacheGetById(multi, members);
1232 if (length >= 0)
1233 {
1234 debug_elog3(DEBUG2, "GetMembers: found %s in the cache",
1235 mxid_to_string(multi, length, *members));
1236 return length;
1237 }
1238
1239 /* Set our OldestVisibleMXactId[] entry if we didn't already */
1240 MultiXactIdSetOldestVisible();
1241
1242 /*
1243 * If we know the multi is used only for locking and not for updates, then
1244 * we can skip checking if the value is older than our oldest visible
1245 * multi. It cannot possibly still be running.
1246 */
1247 if (onlyLock &&
1248 MultiXactIdPrecedes(multi, OldestVisibleMXactId[MyBackendId]))
1249 {
1250 debug_elog2(DEBUG2, "GetMembers: a locker-only multi is too old");
1251 *members = NULL;
1252 return -1;
1253 }
1254
1255 /*
1256 * We check known limits on MultiXact before resorting to the SLRU area.
1257 *
1258 * An ID older than MultiXactState->oldestMultiXactId cannot possibly be
1259 * useful; it has already been removed, or will be removed shortly, by
1260 * truncation. If one is passed, an error is raised.
1261 *
1262 * Also, an ID >= nextMXact shouldn't ever be seen here; if it is seen, it
1263 * implies undetected ID wraparound has occurred. This raises a hard
1264 * error.
1265 *
1266 * Shared lock is enough here since we aren't modifying any global state.
1267 * Acquire it just long enough to grab the current counter values. We may
1268 * need both nextMXact and nextOffset; see below.
1269 */
1270 LWLockAcquire(MultiXactGenLock, LW_SHARED);
1271
1272 oldestMXact = MultiXactState->oldestMultiXactId;
1273 nextMXact = MultiXactState->nextMXact;
1274 nextOffset = MultiXactState->nextOffset;
1275
1276 LWLockRelease(MultiXactGenLock);
1277
1278 if (MultiXactIdPrecedes(multi, oldestMXact))
1279 ereport(ERROR,
1280 (errcode(ERRCODE_INTERNAL_ERROR),
1281 errmsg("MultiXactId %u does no longer exist -- apparent wraparound",
1282 multi)));
1283
1284 if (!MultiXactIdPrecedes(multi, nextMXact))
1285 ereport(ERROR,
1286 (errcode(ERRCODE_INTERNAL_ERROR),
1287 errmsg("MultiXactId %u has not been created yet -- apparent wraparound",
1288 multi)));
1289
1290 /*
1291 * Find out the offset at which we need to start reading MultiXactMembers
1292 * and the number of members in the multixact. We determine the latter as
1293 * the difference between this multixact's starting offset and the next
1294 * one's. However, there are some corner cases to worry about:
1295 *
1296 * 1. This multixact may be the latest one created, in which case there is
1297 * no next one to look at. In this case the nextOffset value we just
1298 * saved is the correct endpoint.
1299 *
1300 * 2. The next multixact may still be in process of being filled in: that
1301 * is, another process may have done GetNewMultiXactId but not yet written
1302 * the offset entry for that ID. In that scenario, it is guaranteed that
1303 * the offset entry for that multixact exists (because GetNewMultiXactId
1304 * won't release MultiXactGenLock until it does) but contains zero
1305 * (because we are careful to pre-zero offset pages). Because
1306 * GetNewMultiXactId will never return zero as the starting offset for a
1307 * multixact, when we read zero as the next multixact's offset, we know we
1308 * have this case. We sleep for a bit and try again.
1309 *
1310 * 3. Because GetNewMultiXactId increments offset zero to offset one to
1311 * handle case #2, there is an ambiguity near the point of offset
1312 * wraparound. If we see next multixact's offset is one, is that our
1313 * multixact's actual endpoint, or did it end at zero with a subsequent
1314 * increment? We handle this using the knowledge that if the zero'th
1315 * member slot wasn't filled, it'll contain zero, and zero isn't a valid
1316 * transaction ID so it can't be a multixact member. Therefore, if we
1317 * read a zero from the members array, just ignore it.
1318 *
1319 * This is all pretty messy, but the mess occurs only in infrequent corner
1320 * cases, so it seems better than holding the MultiXactGenLock for a long
1321 * time on every multixact creation.
1322 */
1323 retry:
1324 LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
1325
1326 pageno = MultiXactIdToOffsetPage(multi);
1327 entryno = MultiXactIdToOffsetEntry(multi);
1328
1329 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
1330 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1331 offptr += entryno;
1332 offset = *offptr;
1333
1334 Assert(offset != 0);
1335
1336 /*
1337 * Use the same increment rule as GetNewMultiXactId(), that is, don't
1338 * handle wraparound explicitly until needed.
1339 */
1340 tmpMXact = multi + 1;
1341
1342 if (nextMXact == tmpMXact)
1343 {
1344 /* Corner case 1: there is no next multixact */
1345 length = nextOffset - offset;
1346 }
1347 else
1348 {
1349 MultiXactOffset nextMXOffset;
1350
1351 /* handle wraparound if needed */
1352 if (tmpMXact < FirstMultiXactId)
1353 tmpMXact = FirstMultiXactId;
1354
1355 prev_pageno = pageno;
1356
1357 pageno = MultiXactIdToOffsetPage(tmpMXact);
1358 entryno = MultiXactIdToOffsetEntry(tmpMXact);
1359
1360 if (pageno != prev_pageno)
1361 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
1362
1363 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1364 offptr += entryno;
1365 nextMXOffset = *offptr;
1366
1367 if (nextMXOffset == 0)
1368 {
1369 /* Corner case 2: next multixact is still being filled in */
1370 LWLockRelease(MultiXactOffsetSLRULock);
1371 CHECK_FOR_INTERRUPTS();
1372 pg_usleep(1000L);
1373 goto retry;
1374 }
1375
1376 length = nextMXOffset - offset;
1377 }
1378
1379 LWLockRelease(MultiXactOffsetSLRULock);
1380
1381 ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
1382
1383 /* Now get the members themselves. */
1384 LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
1385
1386 truelength = 0;
1387 prev_pageno = -1;
1388 for (i = 0; i < length; i++, offset++)
1389 {
1390 TransactionId *xactptr;
1391 uint32 *flagsptr;
1392 int flagsoff;
1393 int bshift;
1394 int memberoff;
1395
1396 pageno = MXOffsetToMemberPage(offset);
1397 memberoff = MXOffsetToMemberOffset(offset);
1398
1399 if (pageno != prev_pageno)
1400 {
1401 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
1402 prev_pageno = pageno;
1403 }
1404
1405 xactptr = (TransactionId *)
1406 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
1407
1408 if (!TransactionIdIsValid(*xactptr))
1409 {
1410 /* Corner case 3: we must be looking at unused slot zero */
1411 Assert(offset == 0);
1412 continue;
1413 }
1414
1415 flagsoff = MXOffsetToFlagsOffset(offset);
1416 bshift = MXOffsetToFlagsBitShift(offset);
1417 flagsptr = (uint32 *) (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
1418
1419 ptr[truelength].xid = *xactptr;
1420 ptr[truelength].status = (*flagsptr >> bshift) & MXACT_MEMBER_XACT_BITMASK;
1421 truelength++;
1422 }
1423
1424 LWLockRelease(MultiXactMemberSLRULock);
1425
1426 /* A multixid with zero members should not happen */
1427 Assert(truelength > 0);
1428
1429 /*
1430 * Copy the result into the local cache.
1431 */
1432 mXactCachePut(multi, truelength, ptr);
1433
1434 debug_elog3(DEBUG2, "GetMembers: no cache for %s",
1435 mxid_to_string(multi, truelength, ptr));
1436 *members = ptr;
1437 return truelength;
1438 }
1439
1440 /*
1441 * mxactMemberComparator
1442 * qsort comparison function for MultiXactMember
1443 *
1444 * We can't use wraparound comparison for XIDs because that does not respect
1445 * the triangle inequality! Any old sort order will do.
1446 */
1447 static int
mxactMemberComparator(const void * arg1,const void * arg2)1448 mxactMemberComparator(const void *arg1, const void *arg2)
1449 {
1450 MultiXactMember member1 = *(const MultiXactMember *) arg1;
1451 MultiXactMember member2 = *(const MultiXactMember *) arg2;
1452
1453 if (member1.xid > member2.xid)
1454 return 1;
1455 if (member1.xid < member2.xid)
1456 return -1;
1457 if (member1.status > member2.status)
1458 return 1;
1459 if (member1.status < member2.status)
1460 return -1;
1461 return 0;
1462 }
1463
1464 /*
1465 * mXactCacheGetBySet
1466 * returns a MultiXactId from the cache based on the set of
1467 * TransactionIds that compose it, or InvalidMultiXactId if
1468 * none matches.
1469 *
1470 * This is helpful, for example, if two transactions want to lock a huge
1471 * table. By using the cache, the second will use the same MultiXactId
1472 * for the majority of tuples, thus keeping MultiXactId usage low (saving
1473 * both I/O and wraparound issues).
1474 *
1475 * NB: the passed members array will be sorted in-place.
1476 */
1477 static MultiXactId
mXactCacheGetBySet(int nmembers,MultiXactMember * members)1478 mXactCacheGetBySet(int nmembers, MultiXactMember *members)
1479 {
1480 dlist_iter iter;
1481
1482 debug_elog3(DEBUG2, "CacheGet: looking for %s",
1483 mxid_to_string(InvalidMultiXactId, nmembers, members));
1484
1485 /* sort the array so comparison is easy */
1486 qsort(members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1487
1488 dlist_foreach(iter, &MXactCache)
1489 {
1490 mXactCacheEnt *entry = dlist_container(mXactCacheEnt, node, iter.cur);
1491
1492 if (entry->nmembers != nmembers)
1493 continue;
1494
1495 /*
1496 * We assume the cache entries are sorted, and that the unused bits in
1497 * "status" are zeroed.
1498 */
1499 if (memcmp(members, entry->members, nmembers * sizeof(MultiXactMember)) == 0)
1500 {
1501 debug_elog3(DEBUG2, "CacheGet: found %u", entry->multi);
1502 dlist_move_head(&MXactCache, iter.cur);
1503 return entry->multi;
1504 }
1505 }
1506
1507 debug_elog2(DEBUG2, "CacheGet: not found :-(");
1508 return InvalidMultiXactId;
1509 }
1510
1511 /*
1512 * mXactCacheGetById
1513 * returns the composing MultiXactMember set from the cache for a
1514 * given MultiXactId, if present.
1515 *
1516 * If successful, *xids is set to the address of a palloc'd copy of the
1517 * MultiXactMember set. Return value is number of members, or -1 on failure.
1518 */
1519 static int
mXactCacheGetById(MultiXactId multi,MultiXactMember ** members)1520 mXactCacheGetById(MultiXactId multi, MultiXactMember **members)
1521 {
1522 dlist_iter iter;
1523
1524 debug_elog3(DEBUG2, "CacheGet: looking for %u", multi);
1525
1526 dlist_foreach(iter, &MXactCache)
1527 {
1528 mXactCacheEnt *entry = dlist_container(mXactCacheEnt, node, iter.cur);
1529
1530 if (entry->multi == multi)
1531 {
1532 MultiXactMember *ptr;
1533 Size size;
1534
1535 size = sizeof(MultiXactMember) * entry->nmembers;
1536 ptr = (MultiXactMember *) palloc(size);
1537
1538 memcpy(ptr, entry->members, size);
1539
1540 debug_elog3(DEBUG2, "CacheGet: found %s",
1541 mxid_to_string(multi,
1542 entry->nmembers,
1543 entry->members));
1544
1545 /*
1546 * Note we modify the list while not using a modifiable iterator.
1547 * This is acceptable only because we exit the iteration
1548 * immediately afterwards.
1549 */
1550 dlist_move_head(&MXactCache, iter.cur);
1551
1552 *members = ptr;
1553 return entry->nmembers;
1554 }
1555 }
1556
1557 debug_elog2(DEBUG2, "CacheGet: not found");
1558 return -1;
1559 }
1560
1561 /*
1562 * mXactCachePut
1563 * Add a new MultiXactId and its composing set into the local cache.
1564 */
1565 static void
mXactCachePut(MultiXactId multi,int nmembers,MultiXactMember * members)1566 mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
1567 {
1568 mXactCacheEnt *entry;
1569
1570 debug_elog3(DEBUG2, "CachePut: storing %s",
1571 mxid_to_string(multi, nmembers, members));
1572
1573 if (MXactContext == NULL)
1574 {
1575 /* The cache only lives as long as the current transaction */
1576 debug_elog2(DEBUG2, "CachePut: initializing memory context");
1577 MXactContext = AllocSetContextCreate(TopTransactionContext,
1578 "MultiXact cache context",
1579 ALLOCSET_SMALL_SIZES);
1580 }
1581
1582 entry = (mXactCacheEnt *)
1583 MemoryContextAlloc(MXactContext,
1584 offsetof(mXactCacheEnt, members) +
1585 nmembers * sizeof(MultiXactMember));
1586
1587 entry->multi = multi;
1588 entry->nmembers = nmembers;
1589 memcpy(entry->members, members, nmembers * sizeof(MultiXactMember));
1590
1591 /* mXactCacheGetBySet assumes the entries are sorted, so sort them */
1592 qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1593
1594 dlist_push_head(&MXactCache, &entry->node);
1595 if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
1596 {
1597 dlist_node *node;
1598 mXactCacheEnt *entry;
1599
1600 node = dlist_tail_node(&MXactCache);
1601 dlist_delete(node);
1602 MXactCacheMembers--;
1603
1604 entry = dlist_container(mXactCacheEnt, node, node);
1605 debug_elog3(DEBUG2, "CachePut: pruning cached multi %u",
1606 entry->multi);
1607
1608 pfree(entry);
1609 }
1610 }
1611
1612 static char *
mxstatus_to_string(MultiXactStatus status)1613 mxstatus_to_string(MultiXactStatus status)
1614 {
1615 switch (status)
1616 {
1617 case MultiXactStatusForKeyShare:
1618 return "keysh";
1619 case MultiXactStatusForShare:
1620 return "sh";
1621 case MultiXactStatusForNoKeyUpdate:
1622 return "fornokeyupd";
1623 case MultiXactStatusForUpdate:
1624 return "forupd";
1625 case MultiXactStatusNoKeyUpdate:
1626 return "nokeyupd";
1627 case MultiXactStatusUpdate:
1628 return "upd";
1629 default:
1630 elog(ERROR, "unrecognized multixact status %d", status);
1631 return "";
1632 }
1633 }
1634
1635 char *
mxid_to_string(MultiXactId multi,int nmembers,MultiXactMember * members)1636 mxid_to_string(MultiXactId multi, int nmembers, MultiXactMember *members)
1637 {
1638 static char *str = NULL;
1639 StringInfoData buf;
1640 int i;
1641
1642 if (str != NULL)
1643 pfree(str);
1644
1645 initStringInfo(&buf);
1646
1647 appendStringInfo(&buf, "%u %d[%u (%s)", multi, nmembers, members[0].xid,
1648 mxstatus_to_string(members[0].status));
1649
1650 for (i = 1; i < nmembers; i++)
1651 appendStringInfo(&buf, ", %u (%s)", members[i].xid,
1652 mxstatus_to_string(members[i].status));
1653
1654 appendStringInfoChar(&buf, ']');
1655 str = MemoryContextStrdup(TopMemoryContext, buf.data);
1656 pfree(buf.data);
1657 return str;
1658 }
1659
1660 /*
1661 * AtEOXact_MultiXact
1662 * Handle transaction end for MultiXact
1663 *
1664 * This is called at top transaction commit or abort (we don't care which).
1665 */
1666 void
AtEOXact_MultiXact(void)1667 AtEOXact_MultiXact(void)
1668 {
1669 /*
1670 * Reset our OldestMemberMXactId and OldestVisibleMXactId values, both of
1671 * which should only be valid while within a transaction.
1672 *
1673 * We assume that storing a MultiXactId is atomic and so we need not take
1674 * MultiXactGenLock to do this.
1675 */
1676 OldestMemberMXactId[MyBackendId] = InvalidMultiXactId;
1677 OldestVisibleMXactId[MyBackendId] = InvalidMultiXactId;
1678
1679 /*
1680 * Discard the local MultiXactId cache. Since MXactContext was created as
1681 * a child of TopTransactionContext, we needn't delete it explicitly.
1682 */
1683 MXactContext = NULL;
1684 dlist_init(&MXactCache);
1685 MXactCacheMembers = 0;
1686 }
1687
1688 /*
1689 * AtPrepare_MultiXact
1690 * Save multixact state at 2PC transaction prepare
1691 *
1692 * In this phase, we only store our OldestMemberMXactId value in the two-phase
1693 * state file.
1694 */
1695 void
AtPrepare_MultiXact(void)1696 AtPrepare_MultiXact(void)
1697 {
1698 MultiXactId myOldestMember = OldestMemberMXactId[MyBackendId];
1699
1700 if (MultiXactIdIsValid(myOldestMember))
1701 RegisterTwoPhaseRecord(TWOPHASE_RM_MULTIXACT_ID, 0,
1702 &myOldestMember, sizeof(MultiXactId));
1703 }
1704
1705 /*
1706 * PostPrepare_MultiXact
1707 * Clean up after successful PREPARE TRANSACTION
1708 */
1709 void
PostPrepare_MultiXact(TransactionId xid)1710 PostPrepare_MultiXact(TransactionId xid)
1711 {
1712 MultiXactId myOldestMember;
1713
1714 /*
1715 * Transfer our OldestMemberMXactId value to the slot reserved for the
1716 * prepared transaction.
1717 */
1718 myOldestMember = OldestMemberMXactId[MyBackendId];
1719 if (MultiXactIdIsValid(myOldestMember))
1720 {
1721 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid, false);
1722
1723 /*
1724 * Even though storing MultiXactId is atomic, acquire lock to make
1725 * sure others see both changes, not just the reset of the slot of the
1726 * current backend. Using a volatile pointer might suffice, but this
1727 * isn't a hot spot.
1728 */
1729 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1730
1731 OldestMemberMXactId[dummyBackendId] = myOldestMember;
1732 OldestMemberMXactId[MyBackendId] = InvalidMultiXactId;
1733
1734 LWLockRelease(MultiXactGenLock);
1735 }
1736
1737 /*
1738 * We don't need to transfer OldestVisibleMXactId value, because the
1739 * transaction is not going to be looking at any more multixacts once it's
1740 * prepared.
1741 *
1742 * We assume that storing a MultiXactId is atomic and so we need not take
1743 * MultiXactGenLock to do this.
1744 */
1745 OldestVisibleMXactId[MyBackendId] = InvalidMultiXactId;
1746
1747 /*
1748 * Discard the local MultiXactId cache like in AtEOX_MultiXact
1749 */
1750 MXactContext = NULL;
1751 dlist_init(&MXactCache);
1752 MXactCacheMembers = 0;
1753 }
1754
1755 /*
1756 * multixact_twophase_recover
1757 * Recover the state of a prepared transaction at startup
1758 */
1759 void
multixact_twophase_recover(TransactionId xid,uint16 info,void * recdata,uint32 len)1760 multixact_twophase_recover(TransactionId xid, uint16 info,
1761 void *recdata, uint32 len)
1762 {
1763 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid, false);
1764 MultiXactId oldestMember;
1765
1766 /*
1767 * Get the oldest member XID from the state file record, and set it in the
1768 * OldestMemberMXactId slot reserved for this prepared transaction.
1769 */
1770 Assert(len == sizeof(MultiXactId));
1771 oldestMember = *((MultiXactId *) recdata);
1772
1773 OldestMemberMXactId[dummyBackendId] = oldestMember;
1774 }
1775
1776 /*
1777 * multixact_twophase_postcommit
1778 * Similar to AtEOX_MultiXact but for COMMIT PREPARED
1779 */
1780 void
multixact_twophase_postcommit(TransactionId xid,uint16 info,void * recdata,uint32 len)1781 multixact_twophase_postcommit(TransactionId xid, uint16 info,
1782 void *recdata, uint32 len)
1783 {
1784 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid, true);
1785
1786 Assert(len == sizeof(MultiXactId));
1787
1788 OldestMemberMXactId[dummyBackendId] = InvalidMultiXactId;
1789 }
1790
1791 /*
1792 * multixact_twophase_postabort
1793 * This is actually just the same as the COMMIT case.
1794 */
1795 void
multixact_twophase_postabort(TransactionId xid,uint16 info,void * recdata,uint32 len)1796 multixact_twophase_postabort(TransactionId xid, uint16 info,
1797 void *recdata, uint32 len)
1798 {
1799 multixact_twophase_postcommit(xid, info, recdata, len);
1800 }
1801
1802 /*
1803 * Initialization of shared memory for MultiXact. We use two SLRU areas,
1804 * thus double memory. Also, reserve space for the shared MultiXactState
1805 * struct and the per-backend MultiXactId arrays (two of those, too).
1806 */
1807 Size
MultiXactShmemSize(void)1808 MultiXactShmemSize(void)
1809 {
1810 Size size;
1811
1812 /* We need 2*MaxOldestSlot + 1 perBackendXactIds[] entries */
1813 #define SHARED_MULTIXACT_STATE_SIZE \
1814 add_size(offsetof(MultiXactStateData, perBackendXactIds) + sizeof(MultiXactId), \
1815 mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
1816
1817 size = SHARED_MULTIXACT_STATE_SIZE;
1818 size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
1819 size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
1820
1821 return size;
1822 }
1823
1824 void
MultiXactShmemInit(void)1825 MultiXactShmemInit(void)
1826 {
1827 bool found;
1828
1829 debug_elog2(DEBUG2, "Shared Memory Init for MultiXact");
1830
1831 MultiXactOffsetCtl->PagePrecedes = MultiXactOffsetPagePrecedes;
1832 MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
1833
1834 SimpleLruInit(MultiXactOffsetCtl,
1835 "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
1836 MultiXactOffsetSLRULock, "pg_multixact/offsets",
1837 LWTRANCHE_MULTIXACTOFFSET_BUFFER);
1838 SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
1839 SimpleLruInit(MultiXactMemberCtl,
1840 "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
1841 MultiXactMemberSLRULock, "pg_multixact/members",
1842 LWTRANCHE_MULTIXACTMEMBER_BUFFER);
1843 /* doesn't call SimpleLruTruncate() or meet criteria for unit tests */
1844
1845 /* Initialize our shared state struct */
1846 MultiXactState = ShmemInitStruct("Shared MultiXact State",
1847 SHARED_MULTIXACT_STATE_SIZE,
1848 &found);
1849 if (!IsUnderPostmaster)
1850 {
1851 Assert(!found);
1852
1853 /* Make sure we zero out the per-backend state */
1854 MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
1855 }
1856 else
1857 Assert(found);
1858
1859 /*
1860 * Set up array pointers. Note that perBackendXactIds[0] is wasted space
1861 * since we only use indexes 1..MaxOldestSlot in each array.
1862 */
1863 OldestMemberMXactId = MultiXactState->perBackendXactIds;
1864 OldestVisibleMXactId = OldestMemberMXactId + MaxOldestSlot;
1865 }
1866
1867 /*
1868 * This func must be called ONCE on system install. It creates the initial
1869 * MultiXact segments. (The MultiXacts directories are assumed to have been
1870 * created by initdb, and MultiXactShmemInit must have been called already.)
1871 */
1872 void
BootStrapMultiXact(void)1873 BootStrapMultiXact(void)
1874 {
1875 int slotno;
1876
1877 LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
1878
1879 /* Create and zero the first page of the offsets log */
1880 slotno = ZeroMultiXactOffsetPage(0, false);
1881
1882 /* Make sure it's written out */
1883 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
1884 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
1885
1886 LWLockRelease(MultiXactOffsetSLRULock);
1887
1888 LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
1889
1890 /* Create and zero the first page of the members log */
1891 slotno = ZeroMultiXactMemberPage(0, false);
1892
1893 /* Make sure it's written out */
1894 SimpleLruWritePage(MultiXactMemberCtl, slotno);
1895 Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
1896
1897 LWLockRelease(MultiXactMemberSLRULock);
1898 }
1899
1900 /*
1901 * Initialize (or reinitialize) a page of MultiXactOffset to zeroes.
1902 * If writeXlog is true, also emit an XLOG record saying we did this.
1903 *
1904 * The page is not actually written, just set up in shared memory.
1905 * The slot number of the new page is returned.
1906 *
1907 * Control lock must be held at entry, and will be held at exit.
1908 */
1909 static int
ZeroMultiXactOffsetPage(int pageno,bool writeXlog)1910 ZeroMultiXactOffsetPage(int pageno, bool writeXlog)
1911 {
1912 int slotno;
1913
1914 slotno = SimpleLruZeroPage(MultiXactOffsetCtl, pageno);
1915
1916 if (writeXlog)
1917 WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_OFF_PAGE);
1918
1919 return slotno;
1920 }
1921
1922 /*
1923 * Ditto, for MultiXactMember
1924 */
1925 static int
ZeroMultiXactMemberPage(int pageno,bool writeXlog)1926 ZeroMultiXactMemberPage(int pageno, bool writeXlog)
1927 {
1928 int slotno;
1929
1930 slotno = SimpleLruZeroPage(MultiXactMemberCtl, pageno);
1931
1932 if (writeXlog)
1933 WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_MEM_PAGE);
1934
1935 return slotno;
1936 }
1937
1938 /*
1939 * MaybeExtendOffsetSlru
1940 * Extend the offsets SLRU area, if necessary
1941 *
1942 * After a binary upgrade from <= 9.2, the pg_multixact/offsets SLRU area might
1943 * contain files that are shorter than necessary; this would occur if the old
1944 * installation had used multixacts beyond the first page (files cannot be
1945 * copied, because the on-disk representation is different). pg_upgrade would
1946 * update pg_control to set the next offset value to be at that position, so
1947 * that tuples marked as locked by such MultiXacts would be seen as visible
1948 * without having to consult multixact. However, trying to create and use a
1949 * new MultiXactId would result in an error because the page on which the new
1950 * value would reside does not exist. This routine is in charge of creating
1951 * such pages.
1952 */
1953 static void
MaybeExtendOffsetSlru(void)1954 MaybeExtendOffsetSlru(void)
1955 {
1956 int pageno;
1957
1958 pageno = MultiXactIdToOffsetPage(MultiXactState->nextMXact);
1959
1960 LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
1961
1962 if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
1963 {
1964 int slotno;
1965
1966 /*
1967 * Fortunately for us, SimpleLruWritePage is already prepared to deal
1968 * with creating a new segment file even if the page we're writing is
1969 * not the first in it, so this is enough.
1970 */
1971 slotno = ZeroMultiXactOffsetPage(pageno, false);
1972 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
1973 }
1974
1975 LWLockRelease(MultiXactOffsetSLRULock);
1976 }
1977
1978 /*
1979 * This must be called ONCE during postmaster or standalone-backend startup.
1980 *
1981 * StartupXLOG has already established nextMXact/nextOffset by calling
1982 * MultiXactSetNextMXact and/or MultiXactAdvanceNextMXact, and the oldestMulti
1983 * info from pg_control and/or MultiXactAdvanceOldest, but we haven't yet
1984 * replayed WAL.
1985 */
1986 void
StartupMultiXact(void)1987 StartupMultiXact(void)
1988 {
1989 MultiXactId multi = MultiXactState->nextMXact;
1990 MultiXactOffset offset = MultiXactState->nextOffset;
1991 int pageno;
1992
1993 /*
1994 * Initialize offset's idea of the latest page number.
1995 */
1996 pageno = MultiXactIdToOffsetPage(multi);
1997 MultiXactOffsetCtl->shared->latest_page_number = pageno;
1998
1999 /*
2000 * Initialize member's idea of the latest page number.
2001 */
2002 pageno = MXOffsetToMemberPage(offset);
2003 MultiXactMemberCtl->shared->latest_page_number = pageno;
2004 }
2005
2006 /*
2007 * This must be called ONCE at the end of startup/recovery.
2008 */
2009 void
TrimMultiXact(void)2010 TrimMultiXact(void)
2011 {
2012 MultiXactId nextMXact;
2013 MultiXactOffset offset;
2014 MultiXactId oldestMXact;
2015 Oid oldestMXactDB;
2016 int pageno;
2017 int entryno;
2018 int flagsoff;
2019
2020 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2021 nextMXact = MultiXactState->nextMXact;
2022 offset = MultiXactState->nextOffset;
2023 oldestMXact = MultiXactState->oldestMultiXactId;
2024 oldestMXactDB = MultiXactState->oldestMultiXactDB;
2025 LWLockRelease(MultiXactGenLock);
2026
2027 /* Clean up offsets state */
2028 LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
2029
2030 /*
2031 * (Re-)Initialize our idea of the latest page number for offsets.
2032 */
2033 pageno = MultiXactIdToOffsetPage(nextMXact);
2034 MultiXactOffsetCtl->shared->latest_page_number = pageno;
2035
2036 /*
2037 * Zero out the remainder of the current offsets page. See notes in
2038 * TrimCLOG() for background. Unlike CLOG, some WAL record covers every
2039 * pg_multixact SLRU mutation. Since, also unlike CLOG, we ignore the WAL
2040 * rule "write xlog before data," nextMXact successors may carry obsolete,
2041 * nonzero offset values. Zero those so case 2 of GetMultiXactIdMembers()
2042 * operates normally.
2043 */
2044 entryno = MultiXactIdToOffsetEntry(nextMXact);
2045 if (entryno != 0)
2046 {
2047 int slotno;
2048 MultiXactOffset *offptr;
2049
2050 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, nextMXact);
2051 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2052 offptr += entryno;
2053
2054 MemSet(offptr, 0, BLCKSZ - (entryno * sizeof(MultiXactOffset)));
2055
2056 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
2057 }
2058
2059 LWLockRelease(MultiXactOffsetSLRULock);
2060
2061 /* And the same for members */
2062 LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
2063
2064 /*
2065 * (Re-)Initialize our idea of the latest page number for members.
2066 */
2067 pageno = MXOffsetToMemberPage(offset);
2068 MultiXactMemberCtl->shared->latest_page_number = pageno;
2069
2070 /*
2071 * Zero out the remainder of the current members page. See notes in
2072 * TrimCLOG() for motivation.
2073 */
2074 flagsoff = MXOffsetToFlagsOffset(offset);
2075 if (flagsoff != 0)
2076 {
2077 int slotno;
2078 TransactionId *xidptr;
2079 int memberoff;
2080
2081 memberoff = MXOffsetToMemberOffset(offset);
2082 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, offset);
2083 xidptr = (TransactionId *)
2084 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
2085
2086 MemSet(xidptr, 0, BLCKSZ - memberoff);
2087
2088 /*
2089 * Note: we don't need to zero out the flag bits in the remaining
2090 * members of the current group, because they are always reset before
2091 * writing.
2092 */
2093
2094 MultiXactMemberCtl->shared->page_dirty[slotno] = true;
2095 }
2096
2097 LWLockRelease(MultiXactMemberSLRULock);
2098
2099 /* signal that we're officially up */
2100 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2101 MultiXactState->finishedStartup = true;
2102 LWLockRelease(MultiXactGenLock);
2103
2104 /* Now compute how far away the next members wraparound is. */
2105 SetMultiXactIdLimit(oldestMXact, oldestMXactDB, true);
2106 }
2107
2108 /*
2109 * This must be called ONCE during postmaster or standalone-backend shutdown
2110 */
2111 void
ShutdownMultiXact(void)2112 ShutdownMultiXact(void)
2113 {
2114 /* Flush dirty MultiXact pages to disk */
2115 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(false);
2116 SimpleLruFlush(MultiXactOffsetCtl, false);
2117 SimpleLruFlush(MultiXactMemberCtl, false);
2118 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(false);
2119 }
2120
2121 /*
2122 * Get the MultiXact data to save in a checkpoint record
2123 */
2124 void
MultiXactGetCheckptMulti(bool is_shutdown,MultiXactId * nextMulti,MultiXactOffset * nextMultiOffset,MultiXactId * oldestMulti,Oid * oldestMultiDB)2125 MultiXactGetCheckptMulti(bool is_shutdown,
2126 MultiXactId *nextMulti,
2127 MultiXactOffset *nextMultiOffset,
2128 MultiXactId *oldestMulti,
2129 Oid *oldestMultiDB)
2130 {
2131 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2132 *nextMulti = MultiXactState->nextMXact;
2133 *nextMultiOffset = MultiXactState->nextOffset;
2134 *oldestMulti = MultiXactState->oldestMultiXactId;
2135 *oldestMultiDB = MultiXactState->oldestMultiXactDB;
2136 LWLockRelease(MultiXactGenLock);
2137
2138 debug_elog6(DEBUG2,
2139 "MultiXact: checkpoint is nextMulti %u, nextOffset %u, oldestMulti %u in DB %u",
2140 *nextMulti, *nextMultiOffset, *oldestMulti, *oldestMultiDB);
2141 }
2142
2143 /*
2144 * Perform a checkpoint --- either during shutdown, or on-the-fly
2145 */
2146 void
CheckPointMultiXact(void)2147 CheckPointMultiXact(void)
2148 {
2149 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(true);
2150
2151 /* Flush dirty MultiXact pages to disk */
2152 SimpleLruFlush(MultiXactOffsetCtl, true);
2153 SimpleLruFlush(MultiXactMemberCtl, true);
2154
2155 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(true);
2156 }
2157
2158 /*
2159 * Set the next-to-be-assigned MultiXactId and offset
2160 *
2161 * This is used when we can determine the correct next ID/offset exactly
2162 * from a checkpoint record. Although this is only called during bootstrap
2163 * and XLog replay, we take the lock in case any hot-standby backends are
2164 * examining the values.
2165 */
2166 void
MultiXactSetNextMXact(MultiXactId nextMulti,MultiXactOffset nextMultiOffset)2167 MultiXactSetNextMXact(MultiXactId nextMulti,
2168 MultiXactOffset nextMultiOffset)
2169 {
2170 debug_elog4(DEBUG2, "MultiXact: setting next multi to %u offset %u",
2171 nextMulti, nextMultiOffset);
2172 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2173 MultiXactState->nextMXact = nextMulti;
2174 MultiXactState->nextOffset = nextMultiOffset;
2175 LWLockRelease(MultiXactGenLock);
2176
2177 /*
2178 * During a binary upgrade, make sure that the offsets SLRU is large
2179 * enough to contain the next value that would be created.
2180 *
2181 * We need to do this pretty early during the first startup in binary
2182 * upgrade mode: before StartupMultiXact() in fact, because this routine
2183 * is called even before that by StartupXLOG(). And we can't do it
2184 * earlier than at this point, because during that first call of this
2185 * routine we determine the MultiXactState->nextMXact value that
2186 * MaybeExtendOffsetSlru needs.
2187 */
2188 if (IsBinaryUpgrade)
2189 MaybeExtendOffsetSlru();
2190 }
2191
2192 /*
2193 * Determine the last safe MultiXactId to allocate given the currently oldest
2194 * datminmxid (ie, the oldest MultiXactId that might exist in any database
2195 * of our cluster), and the OID of the (or a) database with that value.
2196 *
2197 * is_startup is true when we are just starting the cluster, false when we
2198 * are updating state in a running cluster. This only affects log messages.
2199 */
2200 void
SetMultiXactIdLimit(MultiXactId oldest_datminmxid,Oid oldest_datoid,bool is_startup)2201 SetMultiXactIdLimit(MultiXactId oldest_datminmxid, Oid oldest_datoid,
2202 bool is_startup)
2203 {
2204 MultiXactId multiVacLimit;
2205 MultiXactId multiWarnLimit;
2206 MultiXactId multiStopLimit;
2207 MultiXactId multiWrapLimit;
2208 MultiXactId curMulti;
2209 bool needs_offset_vacuum;
2210
2211 Assert(MultiXactIdIsValid(oldest_datminmxid));
2212
2213 /*
2214 * We pretend that a wrap will happen halfway through the multixact ID
2215 * space, but that's not really true, because multixacts wrap differently
2216 * from transaction IDs. Note that, separately from any concern about
2217 * multixact IDs wrapping, we must ensure that multixact members do not
2218 * wrap. Limits for that are set in SetOffsetVacuumLimit, not here.
2219 */
2220 multiWrapLimit = oldest_datminmxid + (MaxMultiXactId >> 1);
2221 if (multiWrapLimit < FirstMultiXactId)
2222 multiWrapLimit += FirstMultiXactId;
2223
2224 /*
2225 * We'll refuse to continue assigning MultiXactIds once we get within 100
2226 * multi of data loss.
2227 *
2228 * Note: This differs from the magic number used in
2229 * SetTransactionIdLimit() since vacuum itself will never generate new
2230 * multis. XXX actually it does, if it needs to freeze old multis.
2231 */
2232 multiStopLimit = multiWrapLimit - 100;
2233 if (multiStopLimit < FirstMultiXactId)
2234 multiStopLimit -= FirstMultiXactId;
2235
2236 /*
2237 * We'll start complaining loudly when we get within 10M multis of the
2238 * stop point. This is kind of arbitrary, but if you let your gas gauge
2239 * get down to 1% of full, would you be looking for the next gas station?
2240 * We need to be fairly liberal about this number because there are lots
2241 * of scenarios where most transactions are done by automatic clients that
2242 * won't pay attention to warnings. (No, we're not gonna make this
2243 * configurable. If you know enough to configure it, you know enough to
2244 * not get in this kind of trouble in the first place.)
2245 */
2246 multiWarnLimit = multiStopLimit - 10000000;
2247 if (multiWarnLimit < FirstMultiXactId)
2248 multiWarnLimit -= FirstMultiXactId;
2249
2250 /*
2251 * We'll start trying to force autovacuums when oldest_datminmxid gets to
2252 * be more than autovacuum_multixact_freeze_max_age mxids old.
2253 *
2254 * Note: autovacuum_multixact_freeze_max_age is a PGC_POSTMASTER parameter
2255 * so that we don't have to worry about dealing with on-the-fly changes in
2256 * its value. See SetTransactionIdLimit.
2257 */
2258 multiVacLimit = oldest_datminmxid + autovacuum_multixact_freeze_max_age;
2259 if (multiVacLimit < FirstMultiXactId)
2260 multiVacLimit += FirstMultiXactId;
2261
2262 /* Grab lock for just long enough to set the new limit values */
2263 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2264 MultiXactState->oldestMultiXactId = oldest_datminmxid;
2265 MultiXactState->oldestMultiXactDB = oldest_datoid;
2266 MultiXactState->multiVacLimit = multiVacLimit;
2267 MultiXactState->multiWarnLimit = multiWarnLimit;
2268 MultiXactState->multiStopLimit = multiStopLimit;
2269 MultiXactState->multiWrapLimit = multiWrapLimit;
2270 curMulti = MultiXactState->nextMXact;
2271 LWLockRelease(MultiXactGenLock);
2272
2273 /* Log the info */
2274 ereport(DEBUG1,
2275 (errmsg("MultiXactId wrap limit is %u, limited by database with OID %u",
2276 multiWrapLimit, oldest_datoid)));
2277
2278 /*
2279 * Computing the actual limits is only possible once the data directory is
2280 * in a consistent state. There's no need to compute the limits while
2281 * still replaying WAL - no decisions about new multis are made even
2282 * though multixact creations might be replayed. So we'll only do further
2283 * checks after TrimMultiXact() has been called.
2284 */
2285 if (!MultiXactState->finishedStartup)
2286 return;
2287
2288 Assert(!InRecovery);
2289
2290 /* Set limits for offset vacuum. */
2291 needs_offset_vacuum = SetOffsetVacuumLimit(is_startup);
2292
2293 /*
2294 * If past the autovacuum force point, immediately signal an autovac
2295 * request. The reason for this is that autovac only processes one
2296 * database per invocation. Once it's finished cleaning up the oldest
2297 * database, it'll call here, and we'll signal the postmaster to start
2298 * another iteration immediately if there are still any old databases.
2299 */
2300 if ((MultiXactIdPrecedes(multiVacLimit, curMulti) ||
2301 needs_offset_vacuum) && IsUnderPostmaster)
2302 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
2303
2304 /* Give an immediate warning if past the wrap warn point */
2305 if (MultiXactIdPrecedes(multiWarnLimit, curMulti))
2306 {
2307 char *oldest_datname;
2308
2309 /*
2310 * We can be called when not inside a transaction, for example during
2311 * StartupXLOG(). In such a case we cannot do database access, so we
2312 * must just report the oldest DB's OID.
2313 *
2314 * Note: it's also possible that get_database_name fails and returns
2315 * NULL, for example because the database just got dropped. We'll
2316 * still warn, even though the warning might now be unnecessary.
2317 */
2318 if (IsTransactionState())
2319 oldest_datname = get_database_name(oldest_datoid);
2320 else
2321 oldest_datname = NULL;
2322
2323 if (oldest_datname)
2324 ereport(WARNING,
2325 (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
2326 "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
2327 multiWrapLimit - curMulti,
2328 oldest_datname,
2329 multiWrapLimit - curMulti),
2330 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
2331 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
2332 else
2333 ereport(WARNING,
2334 (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
2335 "database with OID %u must be vacuumed before %u more MultiXactIds are used",
2336 multiWrapLimit - curMulti,
2337 oldest_datoid,
2338 multiWrapLimit - curMulti),
2339 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
2340 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
2341 }
2342 }
2343
2344 /*
2345 * Ensure the next-to-be-assigned MultiXactId is at least minMulti,
2346 * and similarly nextOffset is at least minMultiOffset.
2347 *
2348 * This is used when we can determine minimum safe values from an XLog
2349 * record (either an on-line checkpoint or an mxact creation log entry).
2350 * Although this is only called during XLog replay, we take the lock in case
2351 * any hot-standby backends are examining the values.
2352 */
2353 void
MultiXactAdvanceNextMXact(MultiXactId minMulti,MultiXactOffset minMultiOffset)2354 MultiXactAdvanceNextMXact(MultiXactId minMulti,
2355 MultiXactOffset minMultiOffset)
2356 {
2357 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2358 if (MultiXactIdPrecedes(MultiXactState->nextMXact, minMulti))
2359 {
2360 debug_elog3(DEBUG2, "MultiXact: setting next multi to %u", minMulti);
2361 MultiXactState->nextMXact = minMulti;
2362 }
2363 if (MultiXactOffsetPrecedes(MultiXactState->nextOffset, minMultiOffset))
2364 {
2365 debug_elog3(DEBUG2, "MultiXact: setting next offset to %u",
2366 minMultiOffset);
2367 MultiXactState->nextOffset = minMultiOffset;
2368 }
2369 LWLockRelease(MultiXactGenLock);
2370 }
2371
2372 /*
2373 * Update our oldestMultiXactId value, but only if it's more recent than what
2374 * we had.
2375 *
2376 * This may only be called during WAL replay.
2377 */
2378 void
MultiXactAdvanceOldest(MultiXactId oldestMulti,Oid oldestMultiDB)2379 MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB)
2380 {
2381 Assert(InRecovery);
2382
2383 if (MultiXactIdPrecedes(MultiXactState->oldestMultiXactId, oldestMulti))
2384 SetMultiXactIdLimit(oldestMulti, oldestMultiDB, false);
2385 }
2386
2387 /*
2388 * Make sure that MultiXactOffset has room for a newly-allocated MultiXactId.
2389 *
2390 * NB: this is called while holding MultiXactGenLock. We want it to be very
2391 * fast most of the time; even when it's not so fast, no actual I/O need
2392 * happen unless we're forced to write out a dirty log or xlog page to make
2393 * room in shared memory.
2394 */
2395 static void
ExtendMultiXactOffset(MultiXactId multi)2396 ExtendMultiXactOffset(MultiXactId multi)
2397 {
2398 int pageno;
2399
2400 /*
2401 * No work except at first MultiXactId of a page. But beware: just after
2402 * wraparound, the first MultiXactId of page zero is FirstMultiXactId.
2403 */
2404 if (MultiXactIdToOffsetEntry(multi) != 0 &&
2405 multi != FirstMultiXactId)
2406 return;
2407
2408 pageno = MultiXactIdToOffsetPage(multi);
2409
2410 LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
2411
2412 /* Zero the page and make an XLOG entry about it */
2413 ZeroMultiXactOffsetPage(pageno, true);
2414
2415 LWLockRelease(MultiXactOffsetSLRULock);
2416 }
2417
2418 /*
2419 * Make sure that MultiXactMember has room for the members of a newly-
2420 * allocated MultiXactId.
2421 *
2422 * Like the above routine, this is called while holding MultiXactGenLock;
2423 * same comments apply.
2424 */
2425 static void
ExtendMultiXactMember(MultiXactOffset offset,int nmembers)2426 ExtendMultiXactMember(MultiXactOffset offset, int nmembers)
2427 {
2428 /*
2429 * It's possible that the members span more than one page of the members
2430 * file, so we loop to ensure we consider each page. The coding is not
2431 * optimal if the members span several pages, but that seems unusual
2432 * enough to not worry much about.
2433 */
2434 while (nmembers > 0)
2435 {
2436 int flagsoff;
2437 int flagsbit;
2438 uint32 difference;
2439
2440 /*
2441 * Only zero when at first entry of a page.
2442 */
2443 flagsoff = MXOffsetToFlagsOffset(offset);
2444 flagsbit = MXOffsetToFlagsBitShift(offset);
2445 if (flagsoff == 0 && flagsbit == 0)
2446 {
2447 int pageno;
2448
2449 pageno = MXOffsetToMemberPage(offset);
2450
2451 LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
2452
2453 /* Zero the page and make an XLOG entry about it */
2454 ZeroMultiXactMemberPage(pageno, true);
2455
2456 LWLockRelease(MultiXactMemberSLRULock);
2457 }
2458
2459 /*
2460 * Compute the number of items till end of current page. Careful: if
2461 * addition of unsigned ints wraps around, we're at the last page of
2462 * the last segment; since that page holds a different number of items
2463 * than other pages, we need to do it differently.
2464 */
2465 if (offset + MAX_MEMBERS_IN_LAST_MEMBERS_PAGE < offset)
2466 {
2467 /*
2468 * This is the last page of the last segment; we can compute the
2469 * number of items left to allocate in it without modulo
2470 * arithmetic.
2471 */
2472 difference = MaxMultiXactOffset - offset + 1;
2473 }
2474 else
2475 difference = MULTIXACT_MEMBERS_PER_PAGE - offset % MULTIXACT_MEMBERS_PER_PAGE;
2476
2477 /*
2478 * Advance to next page, taking care to properly handle the wraparound
2479 * case. OK if nmembers goes negative.
2480 */
2481 nmembers -= difference;
2482 offset += difference;
2483 }
2484 }
2485
2486 /*
2487 * GetOldestMultiXactId
2488 *
2489 * Return the oldest MultiXactId that's still possibly still seen as live by
2490 * any running transaction. Older ones might still exist on disk, but they no
2491 * longer have any running member transaction.
2492 *
2493 * It's not safe to truncate MultiXact SLRU segments on the value returned by
2494 * this function; however, it can be used by a full-table vacuum to set the
2495 * point at which it will be possible to truncate SLRU for that table.
2496 */
2497 MultiXactId
GetOldestMultiXactId(void)2498 GetOldestMultiXactId(void)
2499 {
2500 MultiXactId oldestMXact;
2501 MultiXactId nextMXact;
2502 int i;
2503
2504 /*
2505 * This is the oldest valid value among all the OldestMemberMXactId[] and
2506 * OldestVisibleMXactId[] entries, or nextMXact if none are valid.
2507 */
2508 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2509
2510 /*
2511 * We have to beware of the possibility that nextMXact is in the
2512 * wrapped-around state. We don't fix the counter itself here, but we
2513 * must be sure to use a valid value in our calculation.
2514 */
2515 nextMXact = MultiXactState->nextMXact;
2516 if (nextMXact < FirstMultiXactId)
2517 nextMXact = FirstMultiXactId;
2518
2519 oldestMXact = nextMXact;
2520 for (i = 1; i <= MaxOldestSlot; i++)
2521 {
2522 MultiXactId thisoldest;
2523
2524 thisoldest = OldestMemberMXactId[i];
2525 if (MultiXactIdIsValid(thisoldest) &&
2526 MultiXactIdPrecedes(thisoldest, oldestMXact))
2527 oldestMXact = thisoldest;
2528 thisoldest = OldestVisibleMXactId[i];
2529 if (MultiXactIdIsValid(thisoldest) &&
2530 MultiXactIdPrecedes(thisoldest, oldestMXact))
2531 oldestMXact = thisoldest;
2532 }
2533
2534 LWLockRelease(MultiXactGenLock);
2535
2536 return oldestMXact;
2537 }
2538
2539 /*
2540 * Determine how aggressively we need to vacuum in order to prevent member
2541 * wraparound.
2542 *
2543 * To do so determine what's the oldest member offset and install the limit
2544 * info in MultiXactState, where it can be used to prevent overrun of old data
2545 * in the members SLRU area.
2546 *
2547 * The return value is true if emergency autovacuum is required and false
2548 * otherwise.
2549 */
2550 static bool
SetOffsetVacuumLimit(bool is_startup)2551 SetOffsetVacuumLimit(bool is_startup)
2552 {
2553 MultiXactId oldestMultiXactId;
2554 MultiXactId nextMXact;
2555 MultiXactOffset oldestOffset = 0; /* placate compiler */
2556 MultiXactOffset prevOldestOffset;
2557 MultiXactOffset nextOffset;
2558 bool oldestOffsetKnown = false;
2559 bool prevOldestOffsetKnown;
2560 MultiXactOffset offsetStopLimit = 0;
2561 MultiXactOffset prevOffsetStopLimit;
2562
2563 /*
2564 * NB: Have to prevent concurrent truncation, we might otherwise try to
2565 * lookup an oldestMulti that's concurrently getting truncated away.
2566 */
2567 LWLockAcquire(MultiXactTruncationLock, LW_SHARED);
2568
2569 /* Read relevant fields from shared memory. */
2570 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2571 oldestMultiXactId = MultiXactState->oldestMultiXactId;
2572 nextMXact = MultiXactState->nextMXact;
2573 nextOffset = MultiXactState->nextOffset;
2574 prevOldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2575 prevOldestOffset = MultiXactState->oldestOffset;
2576 prevOffsetStopLimit = MultiXactState->offsetStopLimit;
2577 Assert(MultiXactState->finishedStartup);
2578 LWLockRelease(MultiXactGenLock);
2579
2580 /*
2581 * Determine the offset of the oldest multixact. Normally, we can read
2582 * the offset from the multixact itself, but there's an important special
2583 * case: if there are no multixacts in existence at all, oldestMXact
2584 * obviously can't point to one. It will instead point to the multixact
2585 * ID that will be assigned the next time one is needed.
2586 */
2587 if (oldestMultiXactId == nextMXact)
2588 {
2589 /*
2590 * When the next multixact gets created, it will be stored at the next
2591 * offset.
2592 */
2593 oldestOffset = nextOffset;
2594 oldestOffsetKnown = true;
2595 }
2596 else
2597 {
2598 /*
2599 * Figure out where the oldest existing multixact's offsets are
2600 * stored. Due to bugs in early release of PostgreSQL 9.3.X and 9.4.X,
2601 * the supposedly-earliest multixact might not really exist. We are
2602 * careful not to fail in that case.
2603 */
2604 oldestOffsetKnown =
2605 find_multixact_start(oldestMultiXactId, &oldestOffset);
2606
2607 if (oldestOffsetKnown)
2608 ereport(DEBUG1,
2609 (errmsg("oldest MultiXactId member is at offset %u",
2610 oldestOffset)));
2611 else
2612 ereport(LOG,
2613 (errmsg("MultiXact member wraparound protections are disabled because oldest checkpointed MultiXact %u does not exist on disk",
2614 oldestMultiXactId)));
2615 }
2616
2617 LWLockRelease(MultiXactTruncationLock);
2618
2619 /*
2620 * If we can, compute limits (and install them MultiXactState) to prevent
2621 * overrun of old data in the members SLRU area. We can only do so if the
2622 * oldest offset is known though.
2623 */
2624 if (oldestOffsetKnown)
2625 {
2626 /* move back to start of the corresponding segment */
2627 offsetStopLimit = oldestOffset - (oldestOffset %
2628 (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT));
2629
2630 /* always leave one segment before the wraparound point */
2631 offsetStopLimit -= (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT);
2632
2633 if (!prevOldestOffsetKnown && !is_startup)
2634 ereport(LOG,
2635 (errmsg("MultiXact member wraparound protections are now enabled")));
2636
2637 ereport(DEBUG1,
2638 (errmsg("MultiXact member stop limit is now %u based on MultiXact %u",
2639 offsetStopLimit, oldestMultiXactId)));
2640 }
2641 else if (prevOldestOffsetKnown)
2642 {
2643 /*
2644 * If we failed to get the oldest offset this time, but we have a
2645 * value from a previous pass through this function, use the old
2646 * values rather than automatically forcing an emergency autovacuum
2647 * cycle again.
2648 */
2649 oldestOffset = prevOldestOffset;
2650 oldestOffsetKnown = true;
2651 offsetStopLimit = prevOffsetStopLimit;
2652 }
2653
2654 /* Install the computed values */
2655 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2656 MultiXactState->oldestOffset = oldestOffset;
2657 MultiXactState->oldestOffsetKnown = oldestOffsetKnown;
2658 MultiXactState->offsetStopLimit = offsetStopLimit;
2659 LWLockRelease(MultiXactGenLock);
2660
2661 /*
2662 * Do we need an emergency autovacuum? If we're not sure, assume yes.
2663 */
2664 return !oldestOffsetKnown ||
2665 (nextOffset - oldestOffset > MULTIXACT_MEMBER_SAFE_THRESHOLD);
2666 }
2667
2668 /*
2669 * Return whether adding "distance" to "start" would move past "boundary".
2670 *
2671 * We use this to determine whether the addition is "wrapping around" the
2672 * boundary point, hence the name. The reason we don't want to use the regular
2673 * 2^31-modulo arithmetic here is that we want to be able to use the whole of
2674 * the 2^32-1 space here, allowing for more multixacts than would fit
2675 * otherwise.
2676 */
2677 static bool
MultiXactOffsetWouldWrap(MultiXactOffset boundary,MultiXactOffset start,uint32 distance)2678 MultiXactOffsetWouldWrap(MultiXactOffset boundary, MultiXactOffset start,
2679 uint32 distance)
2680 {
2681 MultiXactOffset finish;
2682
2683 /*
2684 * Note that offset number 0 is not used (see GetMultiXactIdMembers), so
2685 * if the addition wraps around the UINT_MAX boundary, skip that value.
2686 */
2687 finish = start + distance;
2688 if (finish < start)
2689 finish++;
2690
2691 /*-----------------------------------------------------------------------
2692 * When the boundary is numerically greater than the starting point, any
2693 * value numerically between the two is not wrapped:
2694 *
2695 * <----S----B---->
2696 * [---) = F wrapped past B (and UINT_MAX)
2697 * [---) = F not wrapped
2698 * [----] = F wrapped past B
2699 *
2700 * When the boundary is numerically less than the starting point (i.e. the
2701 * UINT_MAX wraparound occurs somewhere in between) then all values in
2702 * between are wrapped:
2703 *
2704 * <----B----S---->
2705 * [---) = F not wrapped past B (but wrapped past UINT_MAX)
2706 * [---) = F wrapped past B (and UINT_MAX)
2707 * [----] = F not wrapped
2708 *-----------------------------------------------------------------------
2709 */
2710 if (start < boundary)
2711 return finish >= boundary || finish < start;
2712 else
2713 return finish >= boundary && finish < start;
2714 }
2715
2716 /*
2717 * Find the starting offset of the given MultiXactId.
2718 *
2719 * Returns false if the file containing the multi does not exist on disk.
2720 * Otherwise, returns true and sets *result to the starting member offset.
2721 *
2722 * This function does not prevent concurrent truncation, so if that's
2723 * required, the caller has to protect against that.
2724 */
2725 static bool
find_multixact_start(MultiXactId multi,MultiXactOffset * result)2726 find_multixact_start(MultiXactId multi, MultiXactOffset *result)
2727 {
2728 MultiXactOffset offset;
2729 int pageno;
2730 int entryno;
2731 int slotno;
2732 MultiXactOffset *offptr;
2733
2734 Assert(MultiXactState->finishedStartup);
2735
2736 pageno = MultiXactIdToOffsetPage(multi);
2737 entryno = MultiXactIdToOffsetEntry(multi);
2738
2739 /*
2740 * Flush out dirty data, so PhysicalPageExists can work correctly.
2741 * SimpleLruFlush() is a pretty big hammer for that. Alternatively we
2742 * could add an in-memory version of page exists, but find_multixact_start
2743 * is called infrequently, and it doesn't seem bad to flush buffers to
2744 * disk before truncation.
2745 */
2746 SimpleLruFlush(MultiXactOffsetCtl, true);
2747 SimpleLruFlush(MultiXactMemberCtl, true);
2748
2749 if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
2750 return false;
2751
2752 /* lock is acquired by SimpleLruReadPage_ReadOnly */
2753 slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
2754 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2755 offptr += entryno;
2756 offset = *offptr;
2757 LWLockRelease(MultiXactOffsetSLRULock);
2758
2759 *result = offset;
2760 return true;
2761 }
2762
2763 /*
2764 * Determine how many multixacts, and how many multixact members, currently
2765 * exist. Return false if unable to determine.
2766 */
2767 static bool
ReadMultiXactCounts(uint32 * multixacts,MultiXactOffset * members)2768 ReadMultiXactCounts(uint32 *multixacts, MultiXactOffset *members)
2769 {
2770 MultiXactOffset nextOffset;
2771 MultiXactOffset oldestOffset;
2772 MultiXactId oldestMultiXactId;
2773 MultiXactId nextMultiXactId;
2774 bool oldestOffsetKnown;
2775
2776 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2777 nextOffset = MultiXactState->nextOffset;
2778 oldestMultiXactId = MultiXactState->oldestMultiXactId;
2779 nextMultiXactId = MultiXactState->nextMXact;
2780 oldestOffset = MultiXactState->oldestOffset;
2781 oldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2782 LWLockRelease(MultiXactGenLock);
2783
2784 if (!oldestOffsetKnown)
2785 return false;
2786
2787 *members = nextOffset - oldestOffset;
2788 *multixacts = nextMultiXactId - oldestMultiXactId;
2789 return true;
2790 }
2791
2792 /*
2793 * Multixact members can be removed once the multixacts that refer to them
2794 * are older than every datminmxid. autovacuum_multixact_freeze_max_age and
2795 * vacuum_multixact_freeze_table_age work together to make sure we never have
2796 * too many multixacts; we hope that, at least under normal circumstances,
2797 * this will also be sufficient to keep us from using too many offsets.
2798 * However, if the average multixact has many members, we might exhaust the
2799 * members space while still using few enough members that these limits fail
2800 * to trigger full table scans for relminmxid advancement. At that point,
2801 * we'd have no choice but to start failing multixact-creating operations
2802 * with an error.
2803 *
2804 * To prevent that, if more than a threshold portion of the members space is
2805 * used, we effectively reduce autovacuum_multixact_freeze_max_age and
2806 * to a value just less than the number of multixacts in use. We hope that
2807 * this will quickly trigger autovacuuming on the table or tables with the
2808 * oldest relminmxid, thus allowing datminmxid values to advance and removing
2809 * some members.
2810 *
2811 * As the fraction of the member space currently in use grows, we become
2812 * more aggressive in clamping this value. That not only causes autovacuum
2813 * to ramp up, but also makes any manual vacuums the user issues more
2814 * aggressive. This happens because vacuum_set_xid_limits() clamps the
2815 * freeze table and the minimum freeze age based on the effective
2816 * autovacuum_multixact_freeze_max_age this function returns. In the worst
2817 * case, we'll claim the freeze_max_age to zero, and every vacuum of any
2818 * table will try to freeze every multixact.
2819 *
2820 * It's possible that these thresholds should be user-tunable, but for now
2821 * we keep it simple.
2822 */
2823 int
MultiXactMemberFreezeThreshold(void)2824 MultiXactMemberFreezeThreshold(void)
2825 {
2826 MultiXactOffset members;
2827 uint32 multixacts;
2828 uint32 victim_multixacts;
2829 double fraction;
2830
2831 /* If we can't determine member space utilization, assume the worst. */
2832 if (!ReadMultiXactCounts(&multixacts, &members))
2833 return 0;
2834
2835 /* If member space utilization is low, no special action is required. */
2836 if (members <= MULTIXACT_MEMBER_SAFE_THRESHOLD)
2837 return autovacuum_multixact_freeze_max_age;
2838
2839 /*
2840 * Compute a target for relminmxid advancement. The number of multixacts
2841 * we try to eliminate from the system is based on how far we are past
2842 * MULTIXACT_MEMBER_SAFE_THRESHOLD.
2843 */
2844 fraction = (double) (members - MULTIXACT_MEMBER_SAFE_THRESHOLD) /
2845 (MULTIXACT_MEMBER_DANGER_THRESHOLD - MULTIXACT_MEMBER_SAFE_THRESHOLD);
2846 victim_multixacts = multixacts * fraction;
2847
2848 /* fraction could be > 1.0, but lowest possible freeze age is zero */
2849 if (victim_multixacts > multixacts)
2850 return 0;
2851 return multixacts - victim_multixacts;
2852 }
2853
2854 typedef struct mxtruncinfo
2855 {
2856 int earliestExistingPage;
2857 } mxtruncinfo;
2858
2859 /*
2860 * SlruScanDirectory callback
2861 * This callback determines the earliest existing page number.
2862 */
2863 static bool
SlruScanDirCbFindEarliest(SlruCtl ctl,char * filename,int segpage,void * data)2864 SlruScanDirCbFindEarliest(SlruCtl ctl, char *filename, int segpage, void *data)
2865 {
2866 mxtruncinfo *trunc = (mxtruncinfo *) data;
2867
2868 if (trunc->earliestExistingPage == -1 ||
2869 ctl->PagePrecedes(segpage, trunc->earliestExistingPage))
2870 {
2871 trunc->earliestExistingPage = segpage;
2872 }
2873
2874 return false; /* keep going */
2875 }
2876
2877
2878 /*
2879 * Delete members segments [oldest, newOldest)
2880 *
2881 * The members SLRU can, in contrast to the offsets one, be filled to almost
2882 * the full range at once. This means SimpleLruTruncate() can't trivially be
2883 * used - instead the to-be-deleted range is computed using the offsets
2884 * SLRU. C.f. TruncateMultiXact().
2885 */
2886 static void
PerformMembersTruncation(MultiXactOffset oldestOffset,MultiXactOffset newOldestOffset)2887 PerformMembersTruncation(MultiXactOffset oldestOffset, MultiXactOffset newOldestOffset)
2888 {
2889 const int maxsegment = MXOffsetToMemberSegment(MaxMultiXactOffset);
2890 int startsegment = MXOffsetToMemberSegment(oldestOffset);
2891 int endsegment = MXOffsetToMemberSegment(newOldestOffset);
2892 int segment = startsegment;
2893
2894 /*
2895 * Delete all the segments but the last one. The last segment can still
2896 * contain, possibly partially, valid data.
2897 */
2898 while (segment != endsegment)
2899 {
2900 elog(DEBUG2, "truncating multixact members segment %x", segment);
2901 SlruDeleteSegment(MultiXactMemberCtl, segment);
2902
2903 /* move to next segment, handling wraparound correctly */
2904 if (segment == maxsegment)
2905 segment = 0;
2906 else
2907 segment += 1;
2908 }
2909 }
2910
2911 /*
2912 * Delete offsets segments [oldest, newOldest)
2913 */
2914 static void
PerformOffsetsTruncation(MultiXactId oldestMulti,MultiXactId newOldestMulti)2915 PerformOffsetsTruncation(MultiXactId oldestMulti, MultiXactId newOldestMulti)
2916 {
2917 /*
2918 * We step back one multixact to avoid passing a cutoff page that hasn't
2919 * been created yet in the rare case that oldestMulti would be the first
2920 * item on a page and oldestMulti == nextMulti. In that case, if we
2921 * didn't subtract one, we'd trigger SimpleLruTruncate's wraparound
2922 * detection.
2923 */
2924 SimpleLruTruncate(MultiXactOffsetCtl,
2925 MultiXactIdToOffsetPage(PreviousMultiXactId(newOldestMulti)));
2926 }
2927
2928 /*
2929 * Remove all MultiXactOffset and MultiXactMember segments before the oldest
2930 * ones still of interest.
2931 *
2932 * This is only called on a primary as part of vacuum (via
2933 * vac_truncate_clog()). During recovery truncation is done by replaying
2934 * truncation WAL records logged here.
2935 *
2936 * newOldestMulti is the oldest currently required multixact, newOldestMultiDB
2937 * is one of the databases preventing newOldestMulti from increasing.
2938 */
2939 void
TruncateMultiXact(MultiXactId newOldestMulti,Oid newOldestMultiDB)2940 TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
2941 {
2942 MultiXactId oldestMulti;
2943 MultiXactId nextMulti;
2944 MultiXactOffset newOldestOffset;
2945 MultiXactOffset oldestOffset;
2946 MultiXactOffset nextOffset;
2947 mxtruncinfo trunc;
2948 MultiXactId earliest;
2949
2950 Assert(!RecoveryInProgress());
2951 Assert(MultiXactState->finishedStartup);
2952
2953 /*
2954 * We can only allow one truncation to happen at once. Otherwise parts of
2955 * members might vanish while we're doing lookups or similar. There's no
2956 * need to have an interlock with creating new multis or such, since those
2957 * are constrained by the limits (which only grow, never shrink).
2958 */
2959 LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
2960
2961 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2962 nextMulti = MultiXactState->nextMXact;
2963 nextOffset = MultiXactState->nextOffset;
2964 oldestMulti = MultiXactState->oldestMultiXactId;
2965 LWLockRelease(MultiXactGenLock);
2966 Assert(MultiXactIdIsValid(oldestMulti));
2967
2968 /*
2969 * Make sure to only attempt truncation if there's values to truncate
2970 * away. In normal processing values shouldn't go backwards, but there's
2971 * some corner cases (due to bugs) where that's possible.
2972 */
2973 if (MultiXactIdPrecedesOrEquals(newOldestMulti, oldestMulti))
2974 {
2975 LWLockRelease(MultiXactTruncationLock);
2976 return;
2977 }
2978
2979 /*
2980 * Note we can't just plow ahead with the truncation; it's possible that
2981 * there are no segments to truncate, which is a problem because we are
2982 * going to attempt to read the offsets page to determine where to
2983 * truncate the members SLRU. So we first scan the directory to determine
2984 * the earliest offsets page number that we can read without error.
2985 *
2986 * When nextMXact is less than one segment away from multiWrapLimit,
2987 * SlruScanDirCbFindEarliest can find some early segment other than the
2988 * actual earliest. (MultiXactOffsetPagePrecedes(EARLIEST, LATEST)
2989 * returns false, because not all pairs of entries have the same answer.)
2990 * That can also arise when an earlier truncation attempt failed unlink()
2991 * or returned early from this function. The only consequence is
2992 * returning early, which wastes space that we could have liberated.
2993 *
2994 * NB: It's also possible that the page that oldestMulti is on has already
2995 * been truncated away, and we crashed before updating oldestMulti.
2996 */
2997 trunc.earliestExistingPage = -1;
2998 SlruScanDirectory(MultiXactOffsetCtl, SlruScanDirCbFindEarliest, &trunc);
2999 earliest = trunc.earliestExistingPage * MULTIXACT_OFFSETS_PER_PAGE;
3000 if (earliest < FirstMultiXactId)
3001 earliest = FirstMultiXactId;
3002
3003 /* If there's nothing to remove, we can bail out early. */
3004 if (MultiXactIdPrecedes(oldestMulti, earliest))
3005 {
3006 LWLockRelease(MultiXactTruncationLock);
3007 return;
3008 }
3009
3010 /*
3011 * First, compute the safe truncation point for MultiXactMember. This is
3012 * the starting offset of the oldest multixact.
3013 *
3014 * Hopefully, find_multixact_start will always work here, because we've
3015 * already checked that it doesn't precede the earliest MultiXact on disk.
3016 * But if it fails, don't truncate anything, and log a message.
3017 */
3018 if (oldestMulti == nextMulti)
3019 {
3020 /* there are NO MultiXacts */
3021 oldestOffset = nextOffset;
3022 }
3023 else if (!find_multixact_start(oldestMulti, &oldestOffset))
3024 {
3025 ereport(LOG,
3026 (errmsg("oldest MultiXact %u not found, earliest MultiXact %u, skipping truncation",
3027 oldestMulti, earliest)));
3028 LWLockRelease(MultiXactTruncationLock);
3029 return;
3030 }
3031
3032 /*
3033 * Secondly compute up to where to truncate. Lookup the corresponding
3034 * member offset for newOldestMulti for that.
3035 */
3036 if (newOldestMulti == nextMulti)
3037 {
3038 /* there are NO MultiXacts */
3039 newOldestOffset = nextOffset;
3040 }
3041 else if (!find_multixact_start(newOldestMulti, &newOldestOffset))
3042 {
3043 ereport(LOG,
3044 (errmsg("cannot truncate up to MultiXact %u because it does not exist on disk, skipping truncation",
3045 newOldestMulti)));
3046 LWLockRelease(MultiXactTruncationLock);
3047 return;
3048 }
3049
3050 elog(DEBUG1, "performing multixact truncation: "
3051 "offsets [%u, %u), offsets segments [%x, %x), "
3052 "members [%u, %u), members segments [%x, %x)",
3053 oldestMulti, newOldestMulti,
3054 MultiXactIdToOffsetSegment(oldestMulti),
3055 MultiXactIdToOffsetSegment(newOldestMulti),
3056 oldestOffset, newOldestOffset,
3057 MXOffsetToMemberSegment(oldestOffset),
3058 MXOffsetToMemberSegment(newOldestOffset));
3059
3060 /*
3061 * Do truncation, and the WAL logging of the truncation, in a critical
3062 * section. That way offsets/members cannot get out of sync anymore, i.e.
3063 * once consistent the newOldestMulti will always exist in members, even
3064 * if we crashed in the wrong moment.
3065 */
3066 START_CRIT_SECTION();
3067
3068 /*
3069 * Prevent checkpoints from being scheduled concurrently. This is critical
3070 * because otherwise a truncation record might not be replayed after a
3071 * crash/basebackup, even though the state of the data directory would
3072 * require it.
3073 */
3074 Assert(!MyProc->delayChkpt);
3075 MyProc->delayChkpt = true;
3076
3077 /* WAL log truncation */
3078 WriteMTruncateXlogRec(newOldestMultiDB,
3079 oldestMulti, newOldestMulti,
3080 oldestOffset, newOldestOffset);
3081
3082 /*
3083 * Update in-memory limits before performing the truncation, while inside
3084 * the critical section: Have to do it before truncation, to prevent
3085 * concurrent lookups of those values. Has to be inside the critical
3086 * section as otherwise a future call to this function would error out,
3087 * while looking up the oldest member in offsets, if our caller crashes
3088 * before updating the limits.
3089 */
3090 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
3091 MultiXactState->oldestMultiXactId = newOldestMulti;
3092 MultiXactState->oldestMultiXactDB = newOldestMultiDB;
3093 LWLockRelease(MultiXactGenLock);
3094
3095 /* First truncate members */
3096 PerformMembersTruncation(oldestOffset, newOldestOffset);
3097
3098 /* Then offsets */
3099 PerformOffsetsTruncation(oldestMulti, newOldestMulti);
3100
3101 MyProc->delayChkpt = false;
3102
3103 END_CRIT_SECTION();
3104 LWLockRelease(MultiXactTruncationLock);
3105 }
3106
3107 /*
3108 * Decide whether a MultiXactOffset page number is "older" for truncation
3109 * purposes. Analogous to CLOGPagePrecedes().
3110 *
3111 * Offsetting the values is optional, because MultiXactIdPrecedes() has
3112 * translational symmetry.
3113 */
3114 static bool
MultiXactOffsetPagePrecedes(int page1,int page2)3115 MultiXactOffsetPagePrecedes(int page1, int page2)
3116 {
3117 MultiXactId multi1;
3118 MultiXactId multi2;
3119
3120 multi1 = ((MultiXactId) page1) * MULTIXACT_OFFSETS_PER_PAGE;
3121 multi1 += FirstMultiXactId + 1;
3122 multi2 = ((MultiXactId) page2) * MULTIXACT_OFFSETS_PER_PAGE;
3123 multi2 += FirstMultiXactId + 1;
3124
3125 return (MultiXactIdPrecedes(multi1, multi2) &&
3126 MultiXactIdPrecedes(multi1,
3127 multi2 + MULTIXACT_OFFSETS_PER_PAGE - 1));
3128 }
3129
3130 /*
3131 * Decide whether a MultiXactMember page number is "older" for truncation
3132 * purposes. There is no "invalid offset number" so use the numbers verbatim.
3133 */
3134 static bool
MultiXactMemberPagePrecedes(int page1,int page2)3135 MultiXactMemberPagePrecedes(int page1, int page2)
3136 {
3137 MultiXactOffset offset1;
3138 MultiXactOffset offset2;
3139
3140 offset1 = ((MultiXactOffset) page1) * MULTIXACT_MEMBERS_PER_PAGE;
3141 offset2 = ((MultiXactOffset) page2) * MULTIXACT_MEMBERS_PER_PAGE;
3142
3143 return (MultiXactOffsetPrecedes(offset1, offset2) &&
3144 MultiXactOffsetPrecedes(offset1,
3145 offset2 + MULTIXACT_MEMBERS_PER_PAGE - 1));
3146 }
3147
3148 /*
3149 * Decide which of two MultiXactIds is earlier.
3150 *
3151 * XXX do we need to do something special for InvalidMultiXactId?
3152 * (Doesn't look like it.)
3153 */
3154 bool
MultiXactIdPrecedes(MultiXactId multi1,MultiXactId multi2)3155 MultiXactIdPrecedes(MultiXactId multi1, MultiXactId multi2)
3156 {
3157 int32 diff = (int32) (multi1 - multi2);
3158
3159 return (diff < 0);
3160 }
3161
3162 /*
3163 * MultiXactIdPrecedesOrEquals -- is multi1 logically <= multi2?
3164 *
3165 * XXX do we need to do something special for InvalidMultiXactId?
3166 * (Doesn't look like it.)
3167 */
3168 bool
MultiXactIdPrecedesOrEquals(MultiXactId multi1,MultiXactId multi2)3169 MultiXactIdPrecedesOrEquals(MultiXactId multi1, MultiXactId multi2)
3170 {
3171 int32 diff = (int32) (multi1 - multi2);
3172
3173 return (diff <= 0);
3174 }
3175
3176
3177 /*
3178 * Decide which of two offsets is earlier.
3179 */
3180 static bool
MultiXactOffsetPrecedes(MultiXactOffset offset1,MultiXactOffset offset2)3181 MultiXactOffsetPrecedes(MultiXactOffset offset1, MultiXactOffset offset2)
3182 {
3183 int32 diff = (int32) (offset1 - offset2);
3184
3185 return (diff < 0);
3186 }
3187
3188 /*
3189 * Write an xlog record reflecting the zeroing of either a MEMBERs or
3190 * OFFSETs page (info shows which)
3191 */
3192 static void
WriteMZeroPageXlogRec(int pageno,uint8 info)3193 WriteMZeroPageXlogRec(int pageno, uint8 info)
3194 {
3195 XLogBeginInsert();
3196 XLogRegisterData((char *) (&pageno), sizeof(int));
3197 (void) XLogInsert(RM_MULTIXACT_ID, info);
3198 }
3199
3200 /*
3201 * Write a TRUNCATE xlog record
3202 *
3203 * We must flush the xlog record to disk before returning --- see notes in
3204 * TruncateCLOG().
3205 */
3206 static void
WriteMTruncateXlogRec(Oid oldestMultiDB,MultiXactId startTruncOff,MultiXactId endTruncOff,MultiXactOffset startTruncMemb,MultiXactOffset endTruncMemb)3207 WriteMTruncateXlogRec(Oid oldestMultiDB,
3208 MultiXactId startTruncOff, MultiXactId endTruncOff,
3209 MultiXactOffset startTruncMemb, MultiXactOffset endTruncMemb)
3210 {
3211 XLogRecPtr recptr;
3212 xl_multixact_truncate xlrec;
3213
3214 xlrec.oldestMultiDB = oldestMultiDB;
3215
3216 xlrec.startTruncOff = startTruncOff;
3217 xlrec.endTruncOff = endTruncOff;
3218
3219 xlrec.startTruncMemb = startTruncMemb;
3220 xlrec.endTruncMemb = endTruncMemb;
3221
3222 XLogBeginInsert();
3223 XLogRegisterData((char *) (&xlrec), SizeOfMultiXactTruncate);
3224 recptr = XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_TRUNCATE_ID);
3225 XLogFlush(recptr);
3226 }
3227
3228 /*
3229 * MULTIXACT resource manager's routines
3230 */
3231 void
multixact_redo(XLogReaderState * record)3232 multixact_redo(XLogReaderState *record)
3233 {
3234 uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
3235
3236 /* Backup blocks are not used in multixact records */
3237 Assert(!XLogRecHasAnyBlockRefs(record));
3238
3239 if (info == XLOG_MULTIXACT_ZERO_OFF_PAGE)
3240 {
3241 int pageno;
3242 int slotno;
3243
3244 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
3245
3246 LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
3247
3248 slotno = ZeroMultiXactOffsetPage(pageno, false);
3249 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
3250 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
3251
3252 LWLockRelease(MultiXactOffsetSLRULock);
3253 }
3254 else if (info == XLOG_MULTIXACT_ZERO_MEM_PAGE)
3255 {
3256 int pageno;
3257 int slotno;
3258
3259 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
3260
3261 LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
3262
3263 slotno = ZeroMultiXactMemberPage(pageno, false);
3264 SimpleLruWritePage(MultiXactMemberCtl, slotno);
3265 Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
3266
3267 LWLockRelease(MultiXactMemberSLRULock);
3268 }
3269 else if (info == XLOG_MULTIXACT_CREATE_ID)
3270 {
3271 xl_multixact_create *xlrec =
3272 (xl_multixact_create *) XLogRecGetData(record);
3273 TransactionId max_xid;
3274 int i;
3275
3276 /* Store the data back into the SLRU files */
3277 RecordNewMultiXact(xlrec->mid, xlrec->moff, xlrec->nmembers,
3278 xlrec->members);
3279
3280 /* Make sure nextMXact/nextOffset are beyond what this record has */
3281 MultiXactAdvanceNextMXact(xlrec->mid + 1,
3282 xlrec->moff + xlrec->nmembers);
3283
3284 /*
3285 * Make sure nextFullXid is beyond any XID mentioned in the record.
3286 * This should be unnecessary, since any XID found here ought to have
3287 * other evidence in the XLOG, but let's be safe.
3288 */
3289 max_xid = XLogRecGetXid(record);
3290 for (i = 0; i < xlrec->nmembers; i++)
3291 {
3292 if (TransactionIdPrecedes(max_xid, xlrec->members[i].xid))
3293 max_xid = xlrec->members[i].xid;
3294 }
3295
3296 AdvanceNextFullTransactionIdPastXid(max_xid);
3297 }
3298 else if (info == XLOG_MULTIXACT_TRUNCATE_ID)
3299 {
3300 xl_multixact_truncate xlrec;
3301 int pageno;
3302
3303 memcpy(&xlrec, XLogRecGetData(record),
3304 SizeOfMultiXactTruncate);
3305
3306 elog(DEBUG1, "replaying multixact truncation: "
3307 "offsets [%u, %u), offsets segments [%x, %x), "
3308 "members [%u, %u), members segments [%x, %x)",
3309 xlrec.startTruncOff, xlrec.endTruncOff,
3310 MultiXactIdToOffsetSegment(xlrec.startTruncOff),
3311 MultiXactIdToOffsetSegment(xlrec.endTruncOff),
3312 xlrec.startTruncMemb, xlrec.endTruncMemb,
3313 MXOffsetToMemberSegment(xlrec.startTruncMemb),
3314 MXOffsetToMemberSegment(xlrec.endTruncMemb));
3315
3316 /* should not be required, but more than cheap enough */
3317 LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
3318
3319 /*
3320 * Advance the horizon values, so they're current at the end of
3321 * recovery.
3322 */
3323 SetMultiXactIdLimit(xlrec.endTruncOff, xlrec.oldestMultiDB, false);
3324
3325 PerformMembersTruncation(xlrec.startTruncMemb, xlrec.endTruncMemb);
3326
3327 /*
3328 * During XLOG replay, latest_page_number isn't necessarily set up
3329 * yet; insert a suitable value to bypass the sanity test in
3330 * SimpleLruTruncate.
3331 */
3332 pageno = MultiXactIdToOffsetPage(xlrec.endTruncOff);
3333 MultiXactOffsetCtl->shared->latest_page_number = pageno;
3334 PerformOffsetsTruncation(xlrec.startTruncOff, xlrec.endTruncOff);
3335
3336 LWLockRelease(MultiXactTruncationLock);
3337 }
3338 else
3339 elog(PANIC, "multixact_redo: unknown op code %u", info);
3340 }
3341
3342 Datum
pg_get_multixact_members(PG_FUNCTION_ARGS)3343 pg_get_multixact_members(PG_FUNCTION_ARGS)
3344 {
3345 typedef struct
3346 {
3347 MultiXactMember *members;
3348 int nmembers;
3349 int iter;
3350 } mxact;
3351 MultiXactId mxid = PG_GETARG_UINT32(0);
3352 mxact *multi;
3353 FuncCallContext *funccxt;
3354
3355 if (mxid < FirstMultiXactId)
3356 ereport(ERROR,
3357 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3358 errmsg("invalid MultiXactId: %u", mxid)));
3359
3360 if (SRF_IS_FIRSTCALL())
3361 {
3362 MemoryContext oldcxt;
3363 TupleDesc tupdesc;
3364
3365 funccxt = SRF_FIRSTCALL_INIT();
3366 oldcxt = MemoryContextSwitchTo(funccxt->multi_call_memory_ctx);
3367
3368 multi = palloc(sizeof(mxact));
3369 /* no need to allow for old values here */
3370 multi->nmembers = GetMultiXactIdMembers(mxid, &multi->members, false,
3371 false);
3372 multi->iter = 0;
3373
3374 tupdesc = CreateTemplateTupleDesc(2);
3375 TupleDescInitEntry(tupdesc, (AttrNumber) 1, "xid",
3376 XIDOID, -1, 0);
3377 TupleDescInitEntry(tupdesc, (AttrNumber) 2, "mode",
3378 TEXTOID, -1, 0);
3379
3380 funccxt->attinmeta = TupleDescGetAttInMetadata(tupdesc);
3381 funccxt->user_fctx = multi;
3382
3383 MemoryContextSwitchTo(oldcxt);
3384 }
3385
3386 funccxt = SRF_PERCALL_SETUP();
3387 multi = (mxact *) funccxt->user_fctx;
3388
3389 while (multi->iter < multi->nmembers)
3390 {
3391 HeapTuple tuple;
3392 char *values[2];
3393
3394 values[0] = psprintf("%u", multi->members[multi->iter].xid);
3395 values[1] = mxstatus_to_string(multi->members[multi->iter].status);
3396
3397 tuple = BuildTupleFromCStrings(funccxt->attinmeta, values);
3398
3399 multi->iter++;
3400 pfree(values[0]);
3401 SRF_RETURN_NEXT(funccxt, HeapTupleGetDatum(tuple));
3402 }
3403
3404 SRF_RETURN_DONE(funccxt);
3405 }
3406