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) 1996-2017, 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 MultiXactOffsetControlLock and
196 * MultiXactMemberControlLock to guard accesses to the two sets of SLRU
197 * buffers. For concurrency's sake, we avoid holding more than one of these
198 * locks at a time.)
199 */
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 startOff, MultiXactId endOff,
371 MultiXactOffset startMemb, MultiXactOffset endMemb);
372
373
374 /*
375 * MultiXactIdCreate
376 * Construct a MultiXactId representing two TransactionIds.
377 *
378 * The two XIDs must be different, or be requesting different statuses.
379 *
380 * NB - we don't worry about our local MultiXactId cache here, because that
381 * is handled by the lower-level routines.
382 */
383 MultiXactId
MultiXactIdCreate(TransactionId xid1,MultiXactStatus status1,TransactionId xid2,MultiXactStatus status2)384 MultiXactIdCreate(TransactionId xid1, MultiXactStatus status1,
385 TransactionId xid2, MultiXactStatus status2)
386 {
387 MultiXactId newMulti;
388 MultiXactMember members[2];
389
390 AssertArg(TransactionIdIsValid(xid1));
391 AssertArg(TransactionIdIsValid(xid2));
392
393 Assert(!TransactionIdEquals(xid1, xid2) || (status1 != status2));
394
395 /* MultiXactIdSetOldestMember() must have been called already. */
396 Assert(MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]));
397
398 /*
399 * Note: unlike MultiXactIdExpand, we don't bother to check that both XIDs
400 * are still running. In typical usage, xid2 will be our own XID and the
401 * caller just did a check on xid1, so it'd be wasted effort.
402 */
403
404 members[0].xid = xid1;
405 members[0].status = status1;
406 members[1].xid = xid2;
407 members[1].status = status2;
408
409 newMulti = MultiXactIdCreateFromMembers(2, members);
410
411 debug_elog3(DEBUG2, "Create: %s",
412 mxid_to_string(newMulti, 2, members));
413
414 return newMulti;
415 }
416
417 /*
418 * MultiXactIdExpand
419 * Add a TransactionId to a pre-existing MultiXactId.
420 *
421 * If the TransactionId is already a member of the passed MultiXactId with the
422 * same status, just return it as-is.
423 *
424 * Note that we do NOT actually modify the membership of a pre-existing
425 * MultiXactId; instead we create a new one. This is necessary to avoid
426 * a race condition against code trying to wait for one MultiXactId to finish;
427 * see notes in heapam.c.
428 *
429 * NB - we don't worry about our local MultiXactId cache here, because that
430 * is handled by the lower-level routines.
431 *
432 * Note: It is critical that MultiXactIds that come from an old cluster (i.e.
433 * one upgraded by pg_upgrade from a cluster older than this feature) are not
434 * passed in.
435 */
436 MultiXactId
MultiXactIdExpand(MultiXactId multi,TransactionId xid,MultiXactStatus status)437 MultiXactIdExpand(MultiXactId multi, TransactionId xid, MultiXactStatus status)
438 {
439 MultiXactId newMulti;
440 MultiXactMember *members;
441 MultiXactMember *newMembers;
442 int nmembers;
443 int i;
444 int j;
445
446 AssertArg(MultiXactIdIsValid(multi));
447 AssertArg(TransactionIdIsValid(xid));
448
449 /* MultiXactIdSetOldestMember() must have been called already. */
450 Assert(MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]));
451
452 debug_elog5(DEBUG2, "Expand: received multi %u, xid %u status %s",
453 multi, xid, mxstatus_to_string(status));
454
455 /*
456 * Note: we don't allow for old multis here. The reason is that the only
457 * caller of this function does a check that the multixact is no longer
458 * running.
459 */
460 nmembers = GetMultiXactIdMembers(multi, &members, false, false);
461
462 if (nmembers < 0)
463 {
464 MultiXactMember member;
465
466 /*
467 * The MultiXactId is obsolete. This can only happen if all the
468 * MultiXactId members stop running between the caller checking and
469 * passing it to us. It would be better to return that fact to the
470 * caller, but it would complicate the API and it's unlikely to happen
471 * too often, so just deal with it by creating a singleton MultiXact.
472 */
473 member.xid = xid;
474 member.status = status;
475 newMulti = MultiXactIdCreateFromMembers(1, &member);
476
477 debug_elog4(DEBUG2, "Expand: %u has no members, create singleton %u",
478 multi, newMulti);
479 return newMulti;
480 }
481
482 /*
483 * If the TransactionId is already a member of the MultiXactId with the
484 * same status, just return the existing MultiXactId.
485 */
486 for (i = 0; i < nmembers; i++)
487 {
488 if (TransactionIdEquals(members[i].xid, xid) &&
489 (members[i].status == status))
490 {
491 debug_elog4(DEBUG2, "Expand: %u is already a member of %u",
492 xid, multi);
493 pfree(members);
494 return multi;
495 }
496 }
497
498 /*
499 * Determine which of the members of the MultiXactId are still of
500 * interest. This is any running transaction, and also any transaction
501 * that grabbed something stronger than just a lock and was committed. (An
502 * update that aborted is of no interest here; and having more than one
503 * update Xid in a multixact would cause errors elsewhere.)
504 *
505 * Removing dead members is not just an optimization: freezing of tuples
506 * whose Xmax are multis depends on this behavior.
507 *
508 * Note we have the same race condition here as above: j could be 0 at the
509 * end of the loop.
510 */
511 newMembers = (MultiXactMember *)
512 palloc(sizeof(MultiXactMember) * (nmembers + 1));
513
514 for (i = 0, j = 0; i < nmembers; i++)
515 {
516 if (TransactionIdIsInProgress(members[i].xid) ||
517 (ISUPDATE_from_mxstatus(members[i].status) &&
518 TransactionIdDidCommit(members[i].xid)))
519 {
520 newMembers[j].xid = members[i].xid;
521 newMembers[j++].status = members[i].status;
522 }
523 }
524
525 newMembers[j].xid = xid;
526 newMembers[j++].status = status;
527 newMulti = MultiXactIdCreateFromMembers(j, newMembers);
528
529 pfree(members);
530 pfree(newMembers);
531
532 debug_elog3(DEBUG2, "Expand: returning new multi %u", newMulti);
533
534 return newMulti;
535 }
536
537 /*
538 * MultiXactIdIsRunning
539 * Returns whether a MultiXactId is "running".
540 *
541 * We return true if at least one member of the given MultiXactId is still
542 * running. Note that a "false" result is certain not to change,
543 * because it is not legal to add members to an existing MultiXactId.
544 *
545 * Caller is expected to have verified that the multixact does not come from
546 * a pg_upgraded share-locked tuple.
547 */
548 bool
MultiXactIdIsRunning(MultiXactId multi,bool isLockOnly)549 MultiXactIdIsRunning(MultiXactId multi, bool isLockOnly)
550 {
551 MultiXactMember *members;
552 int nmembers;
553 int i;
554
555 debug_elog3(DEBUG2, "IsRunning %u?", multi);
556
557 /*
558 * "false" here means we assume our callers have checked that the given
559 * multi cannot possibly come from a pg_upgraded database.
560 */
561 nmembers = GetMultiXactIdMembers(multi, &members, false, isLockOnly);
562
563 if (nmembers <= 0)
564 {
565 debug_elog2(DEBUG2, "IsRunning: no members");
566 return false;
567 }
568
569 /*
570 * Checking for myself is cheap compared to looking in shared memory;
571 * return true if any live subtransaction of the current top-level
572 * transaction is a member.
573 *
574 * This is not needed for correctness, it's just a fast path.
575 */
576 for (i = 0; i < nmembers; i++)
577 {
578 if (TransactionIdIsCurrentTransactionId(members[i].xid))
579 {
580 debug_elog3(DEBUG2, "IsRunning: I (%d) am running!", i);
581 pfree(members);
582 return true;
583 }
584 }
585
586 /*
587 * This could be made faster by having another entry point in procarray.c,
588 * walking the PGPROC array only once for all the members. But in most
589 * cases nmembers should be small enough that it doesn't much matter.
590 */
591 for (i = 0; i < nmembers; i++)
592 {
593 if (TransactionIdIsInProgress(members[i].xid))
594 {
595 debug_elog4(DEBUG2, "IsRunning: member %d (%u) is running",
596 i, members[i].xid);
597 pfree(members);
598 return true;
599 }
600 }
601
602 pfree(members);
603
604 debug_elog3(DEBUG2, "IsRunning: %u is not running", multi);
605
606 return false;
607 }
608
609 /*
610 * MultiXactIdSetOldestMember
611 * Save the oldest MultiXactId this transaction could be a member of.
612 *
613 * We set the OldestMemberMXactId for a given transaction the first time it's
614 * going to do some operation that might require a MultiXactId (tuple lock,
615 * update or delete). We need to do this even if we end up using a
616 * TransactionId instead of a MultiXactId, because there is a chance that
617 * another transaction would add our XID to a MultiXactId.
618 *
619 * The value to set is the next-to-be-assigned MultiXactId, so this is meant to
620 * be called just before doing any such possibly-MultiXactId-able operation.
621 */
622 void
MultiXactIdSetOldestMember(void)623 MultiXactIdSetOldestMember(void)
624 {
625 if (!MultiXactIdIsValid(OldestMemberMXactId[MyBackendId]))
626 {
627 MultiXactId nextMXact;
628
629 /*
630 * You might think we don't need to acquire a lock here, since
631 * fetching and storing of TransactionIds is probably atomic, but in
632 * fact we do: suppose we pick up nextMXact and then lose the CPU for
633 * a long time. Someone else could advance nextMXact, and then
634 * another someone else could compute an OldestVisibleMXactId that
635 * would be after the value we are going to store when we get control
636 * back. Which would be wrong.
637 *
638 * Note that a shared lock is sufficient, because it's enough to stop
639 * someone from advancing nextMXact; and nobody else could be trying
640 * to write to our OldestMember entry, only reading (and we assume
641 * storing it is atomic.)
642 */
643 LWLockAcquire(MultiXactGenLock, LW_SHARED);
644
645 /*
646 * We have to beware of the possibility that nextMXact is in the
647 * wrapped-around state. We don't fix the counter itself here, but we
648 * must be sure to store a valid value in our array entry.
649 */
650 nextMXact = MultiXactState->nextMXact;
651 if (nextMXact < FirstMultiXactId)
652 nextMXact = FirstMultiXactId;
653
654 OldestMemberMXactId[MyBackendId] = nextMXact;
655
656 LWLockRelease(MultiXactGenLock);
657
658 debug_elog4(DEBUG2, "MultiXact: setting OldestMember[%d] = %u",
659 MyBackendId, nextMXact);
660 }
661 }
662
663 /*
664 * MultiXactIdSetOldestVisible
665 * Save the oldest MultiXactId this transaction considers possibly live.
666 *
667 * We set the OldestVisibleMXactId for a given transaction the first time
668 * it's going to inspect any MultiXactId. Once we have set this, we are
669 * guaranteed that the checkpointer won't truncate off SLRU data for
670 * MultiXactIds at or after our OldestVisibleMXactId.
671 *
672 * The value to set is the oldest of nextMXact and all the valid per-backend
673 * OldestMemberMXactId[] entries. Because of the locking we do, we can be
674 * certain that no subsequent call to MultiXactIdSetOldestMember can set
675 * an OldestMemberMXactId[] entry older than what we compute here. Therefore
676 * there is no live transaction, now or later, that can be a member of any
677 * MultiXactId older than the OldestVisibleMXactId we compute here.
678 */
679 static void
MultiXactIdSetOldestVisible(void)680 MultiXactIdSetOldestVisible(void)
681 {
682 if (!MultiXactIdIsValid(OldestVisibleMXactId[MyBackendId]))
683 {
684 MultiXactId oldestMXact;
685 int i;
686
687 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
688
689 /*
690 * We have to beware of the possibility that nextMXact is in the
691 * wrapped-around state. We don't fix the counter itself here, but we
692 * must be sure to store a valid value in our array entry.
693 */
694 oldestMXact = MultiXactState->nextMXact;
695 if (oldestMXact < FirstMultiXactId)
696 oldestMXact = FirstMultiXactId;
697
698 for (i = 1; i <= MaxOldestSlot; i++)
699 {
700 MultiXactId thisoldest = OldestMemberMXactId[i];
701
702 if (MultiXactIdIsValid(thisoldest) &&
703 MultiXactIdPrecedes(thisoldest, oldestMXact))
704 oldestMXact = thisoldest;
705 }
706
707 OldestVisibleMXactId[MyBackendId] = oldestMXact;
708
709 LWLockRelease(MultiXactGenLock);
710
711 debug_elog4(DEBUG2, "MultiXact: setting OldestVisible[%d] = %u",
712 MyBackendId, oldestMXact);
713 }
714 }
715
716 /*
717 * ReadNextMultiXactId
718 * Return the next MultiXactId to be assigned, but don't allocate it
719 */
720 MultiXactId
ReadNextMultiXactId(void)721 ReadNextMultiXactId(void)
722 {
723 MultiXactId mxid;
724
725 /* XXX we could presumably do this without a lock. */
726 LWLockAcquire(MultiXactGenLock, LW_SHARED);
727 mxid = MultiXactState->nextMXact;
728 LWLockRelease(MultiXactGenLock);
729
730 if (mxid < FirstMultiXactId)
731 mxid = FirstMultiXactId;
732
733 return mxid;
734 }
735
736 /*
737 * MultiXactIdCreateFromMembers
738 * Make a new MultiXactId from the specified set of members
739 *
740 * Make XLOG, SLRU and cache entries for a new MultiXactId, recording the
741 * given TransactionIds as members. Returns the newly created MultiXactId.
742 *
743 * NB: the passed members[] array will be sorted in-place.
744 */
745 MultiXactId
MultiXactIdCreateFromMembers(int nmembers,MultiXactMember * members)746 MultiXactIdCreateFromMembers(int nmembers, MultiXactMember *members)
747 {
748 MultiXactId multi;
749 MultiXactOffset offset;
750 xl_multixact_create xlrec;
751
752 debug_elog3(DEBUG2, "Create: %s",
753 mxid_to_string(InvalidMultiXactId, nmembers, members));
754
755 /*
756 * See if the same set of members already exists in our cache; if so, just
757 * re-use that MultiXactId. (Note: it might seem that looking in our
758 * cache is insufficient, and we ought to search disk to see if a
759 * duplicate definition already exists. But since we only ever create
760 * MultiXacts containing our own XID, in most cases any such MultiXacts
761 * were in fact created by us, and so will be in our cache. There are
762 * corner cases where someone else added us to a MultiXact without our
763 * knowledge, but it's not worth checking for.)
764 */
765 multi = mXactCacheGetBySet(nmembers, members);
766 if (MultiXactIdIsValid(multi))
767 {
768 debug_elog2(DEBUG2, "Create: in cache!");
769 return multi;
770 }
771
772 /* Verify that there is a single update Xid among the given members. */
773 {
774 int i;
775 bool has_update = false;
776
777 for (i = 0; i < nmembers; i++)
778 {
779 if (ISUPDATE_from_mxstatus(members[i].status))
780 {
781 if (has_update)
782 elog(ERROR, "new multixact has more than one updating member");
783 has_update = true;
784 }
785 }
786 }
787
788 /*
789 * Assign the MXID and offsets range to use, and make sure there is space
790 * in the OFFSETs and MEMBERs files. NB: this routine does
791 * START_CRIT_SECTION().
792 *
793 * Note: unlike MultiXactIdCreate and MultiXactIdExpand, we do not check
794 * that we've called MultiXactIdSetOldestMember here. This is because
795 * this routine is used in some places to create new MultiXactIds of which
796 * the current backend is not a member, notably during freezing of multis
797 * in vacuum. During vacuum, in particular, it would be unacceptable to
798 * keep OldestMulti set, in case it runs for long.
799 */
800 multi = GetNewMultiXactId(nmembers, &offset);
801
802 /* Make an XLOG entry describing the new MXID. */
803 xlrec.mid = multi;
804 xlrec.moff = offset;
805 xlrec.nmembers = nmembers;
806
807 /*
808 * XXX Note: there's a lot of padding space in MultiXactMember. We could
809 * find a more compact representation of this Xlog record -- perhaps all
810 * the status flags in one XLogRecData, then all the xids in another one?
811 * Not clear that it's worth the trouble though.
812 */
813 XLogBeginInsert();
814 XLogRegisterData((char *) (&xlrec), SizeOfMultiXactCreate);
815 XLogRegisterData((char *) members, nmembers * sizeof(MultiXactMember));
816
817 (void) XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_CREATE_ID);
818
819 /* Now enter the information into the OFFSETs and MEMBERs logs */
820 RecordNewMultiXact(multi, offset, nmembers, members);
821
822 /* Done with critical section */
823 END_CRIT_SECTION();
824
825 /* Store the new MultiXactId in the local cache, too */
826 mXactCachePut(multi, nmembers, members);
827
828 debug_elog2(DEBUG2, "Create: all done");
829
830 return multi;
831 }
832
833 /*
834 * RecordNewMultiXact
835 * Write info about a new multixact into the offsets and members files
836 *
837 * This is broken out of MultiXactIdCreateFromMembers so that xlog replay can
838 * use it.
839 */
840 static void
RecordNewMultiXact(MultiXactId multi,MultiXactOffset offset,int nmembers,MultiXactMember * members)841 RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
842 int nmembers, MultiXactMember *members)
843 {
844 int pageno;
845 int prev_pageno;
846 int entryno;
847 int slotno;
848 MultiXactOffset *offptr;
849 int i;
850
851 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
852
853 pageno = MultiXactIdToOffsetPage(multi);
854 entryno = MultiXactIdToOffsetEntry(multi);
855
856 /*
857 * Note: we pass the MultiXactId to SimpleLruReadPage as the "transaction"
858 * to complain about if there's any I/O error. This is kinda bogus, but
859 * since the errors will always give the full pathname, it should be clear
860 * enough that a MultiXactId is really involved. Perhaps someday we'll
861 * take the trouble to generalize the slru.c error reporting code.
862 */
863 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
864 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
865 offptr += entryno;
866
867 *offptr = offset;
868
869 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
870
871 /* Exchange our lock */
872 LWLockRelease(MultiXactOffsetControlLock);
873
874 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
875
876 prev_pageno = -1;
877
878 for (i = 0; i < nmembers; i++, offset++)
879 {
880 TransactionId *memberptr;
881 uint32 *flagsptr;
882 uint32 flagsval;
883 int bshift;
884 int flagsoff;
885 int memberoff;
886
887 Assert(members[i].status <= MultiXactStatusUpdate);
888
889 pageno = MXOffsetToMemberPage(offset);
890 memberoff = MXOffsetToMemberOffset(offset);
891 flagsoff = MXOffsetToFlagsOffset(offset);
892 bshift = MXOffsetToFlagsBitShift(offset);
893
894 if (pageno != prev_pageno)
895 {
896 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
897 prev_pageno = pageno;
898 }
899
900 memberptr = (TransactionId *)
901 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
902
903 *memberptr = members[i].xid;
904
905 flagsptr = (uint32 *)
906 (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
907
908 flagsval = *flagsptr;
909 flagsval &= ~(((1 << MXACT_MEMBER_BITS_PER_XACT) - 1) << bshift);
910 flagsval |= (members[i].status << bshift);
911 *flagsptr = flagsval;
912
913 MultiXactMemberCtl->shared->page_dirty[slotno] = true;
914 }
915
916 LWLockRelease(MultiXactMemberControlLock);
917 }
918
919 /*
920 * GetNewMultiXactId
921 * Get the next MultiXactId.
922 *
923 * Also, reserve the needed amount of space in the "members" area. The
924 * starting offset of the reserved space is returned in *offset.
925 *
926 * This may generate XLOG records for expansion of the offsets and/or members
927 * files. Unfortunately, we have to do that while holding MultiXactGenLock
928 * to avoid race conditions --- the XLOG record for zeroing a page must appear
929 * before any backend can possibly try to store data in that page!
930 *
931 * We start a critical section before advancing the shared counters. The
932 * caller must end the critical section after writing SLRU data.
933 */
934 static MultiXactId
GetNewMultiXactId(int nmembers,MultiXactOffset * offset)935 GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
936 {
937 MultiXactId result;
938 MultiXactOffset nextOffset;
939
940 debug_elog3(DEBUG2, "GetNew: for %d xids", nmembers);
941
942 /* safety check, we should never get this far in a HS standby */
943 if (RecoveryInProgress())
944 elog(ERROR, "cannot assign MultiXactIds during recovery");
945
946 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
947
948 /* Handle wraparound of the nextMXact counter */
949 if (MultiXactState->nextMXact < FirstMultiXactId)
950 MultiXactState->nextMXact = FirstMultiXactId;
951
952 /* Assign the MXID */
953 result = MultiXactState->nextMXact;
954
955 /*----------
956 * Check to see if it's safe to assign another MultiXactId. This protects
957 * against catastrophic data loss due to multixact wraparound. The basic
958 * rules are:
959 *
960 * If we're past multiVacLimit or the safe threshold for member storage
961 * space, or we don't know what the safe threshold for member storage is,
962 * start trying to force autovacuum cycles.
963 * If we're past multiWarnLimit, start issuing warnings.
964 * If we're past multiStopLimit, refuse to create new MultiXactIds.
965 *
966 * Note these are pretty much the same protections in GetNewTransactionId.
967 *----------
968 */
969 if (!MultiXactIdPrecedes(result, MultiXactState->multiVacLimit))
970 {
971 /*
972 * For safety's sake, we release MultiXactGenLock while sending
973 * signals, warnings, etc. This is not so much because we care about
974 * preserving concurrency in this situation, as to avoid any
975 * possibility of deadlock while doing get_database_name(). First,
976 * copy all the shared values we'll need in this path.
977 */
978 MultiXactId multiWarnLimit = MultiXactState->multiWarnLimit;
979 MultiXactId multiStopLimit = MultiXactState->multiStopLimit;
980 MultiXactId multiWrapLimit = MultiXactState->multiWrapLimit;
981 Oid oldest_datoid = MultiXactState->oldestMultiXactDB;
982
983 LWLockRelease(MultiXactGenLock);
984
985 if (IsUnderPostmaster &&
986 !MultiXactIdPrecedes(result, multiStopLimit))
987 {
988 char *oldest_datname = get_database_name(oldest_datoid);
989
990 /*
991 * Immediately kick autovacuum into action as we're already in
992 * ERROR territory.
993 */
994 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
995
996 /* complain even if that DB has disappeared */
997 if (oldest_datname)
998 ereport(ERROR,
999 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1000 errmsg("database is not accepting commands that generate new MultiXactIds to avoid wraparound data loss in database \"%s\"",
1001 oldest_datname),
1002 errhint("Execute a database-wide VACUUM in that database.\n"
1003 "You might also need to commit or roll back old prepared transactions.")));
1004 else
1005 ereport(ERROR,
1006 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1007 errmsg("database is not accepting commands that generate new MultiXactIds to avoid wraparound data loss in database with OID %u",
1008 oldest_datoid),
1009 errhint("Execute a database-wide VACUUM in that database.\n"
1010 "You might also need to commit or roll back old prepared transactions.")));
1011 }
1012
1013 /*
1014 * To avoid swamping the postmaster with signals, we issue the autovac
1015 * request only once per 64K multis generated. This still gives
1016 * plenty of chances before we get into real trouble.
1017 */
1018 if (IsUnderPostmaster && (result % 65536) == 0)
1019 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1020
1021 if (!MultiXactIdPrecedes(result, multiWarnLimit))
1022 {
1023 char *oldest_datname = get_database_name(oldest_datoid);
1024
1025 /* complain even if that DB has disappeared */
1026 if (oldest_datname)
1027 ereport(WARNING,
1028 (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
1029 "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
1030 multiWrapLimit - result,
1031 oldest_datname,
1032 multiWrapLimit - result),
1033 errhint("Execute a database-wide VACUUM in that database.\n"
1034 "You might also need to commit or roll back old prepared transactions.")));
1035 else
1036 ereport(WARNING,
1037 (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
1038 "database with OID %u must be vacuumed before %u more MultiXactIds are used",
1039 multiWrapLimit - result,
1040 oldest_datoid,
1041 multiWrapLimit - result),
1042 errhint("Execute a database-wide VACUUM in that database.\n"
1043 "You might also need to commit or roll back old prepared transactions.")));
1044 }
1045
1046 /* Re-acquire lock and start over */
1047 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1048 result = MultiXactState->nextMXact;
1049 if (result < FirstMultiXactId)
1050 result = FirstMultiXactId;
1051 }
1052
1053 /* Make sure there is room for the MXID in the file. */
1054 ExtendMultiXactOffset(result);
1055
1056 /*
1057 * Reserve the members space, similarly to above. Also, be careful not to
1058 * return zero as the starting offset for any multixact. See
1059 * GetMultiXactIdMembers() for motivation.
1060 */
1061 nextOffset = MultiXactState->nextOffset;
1062 if (nextOffset == 0)
1063 {
1064 *offset = 1;
1065 nmembers++; /* allocate member slot 0 too */
1066 }
1067 else
1068 *offset = nextOffset;
1069
1070 /*----------
1071 * Protect against overrun of the members space as well, with the
1072 * following rules:
1073 *
1074 * If we're past offsetStopLimit, refuse to generate more multis.
1075 * If we're close to offsetStopLimit, emit a warning.
1076 *
1077 * Arbitrarily, we start emitting warnings when we're 20 segments or less
1078 * from offsetStopLimit.
1079 *
1080 * Note we haven't updated the shared state yet, so if we fail at this
1081 * point, the multixact ID we grabbed can still be used by the next guy.
1082 *
1083 * Note that there is no point in forcing autovacuum runs here: the
1084 * multixact freeze settings would have to be reduced for that to have any
1085 * effect.
1086 *----------
1087 */
1088 #define OFFSET_WARN_SEGMENTS 20
1089 if (MultiXactState->oldestOffsetKnown &&
1090 MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit, nextOffset,
1091 nmembers))
1092 {
1093 /* see comment in the corresponding offsets wraparound case */
1094 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1095
1096 ereport(ERROR,
1097 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1098 errmsg("multixact \"members\" limit exceeded"),
1099 errdetail_plural("This command would create a multixact with %u members, but the remaining space is only enough for %u member.",
1100 "This command would create a multixact with %u members, but the remaining space is only enough for %u members.",
1101 MultiXactState->offsetStopLimit - nextOffset - 1,
1102 nmembers,
1103 MultiXactState->offsetStopLimit - nextOffset - 1),
1104 errhint("Execute a database-wide VACUUM in database with OID %u with reduced vacuum_multixact_freeze_min_age and vacuum_multixact_freeze_table_age settings.",
1105 MultiXactState->oldestMultiXactDB)));
1106 }
1107
1108 /*
1109 * Check whether we should kick autovacuum into action, to prevent members
1110 * wraparound. NB we use a much larger window to trigger autovacuum than
1111 * just the warning limit. The warning is just a measure of last resort -
1112 * this is in line with GetNewTransactionId's behaviour.
1113 */
1114 if (!MultiXactState->oldestOffsetKnown ||
1115 (MultiXactState->nextOffset - MultiXactState->oldestOffset
1116 > MULTIXACT_MEMBER_SAFE_THRESHOLD))
1117 {
1118 /*
1119 * To avoid swamping the postmaster with signals, we issue the autovac
1120 * request only when crossing a segment boundary. With default
1121 * compilation settings that's roughly after 50k members. This still
1122 * gives plenty of chances before we get into real trouble.
1123 */
1124 if ((MXOffsetToMemberPage(nextOffset) / SLRU_PAGES_PER_SEGMENT) !=
1125 (MXOffsetToMemberPage(nextOffset + nmembers) / SLRU_PAGES_PER_SEGMENT))
1126 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
1127 }
1128
1129 if (MultiXactState->oldestOffsetKnown &&
1130 MultiXactOffsetWouldWrap(MultiXactState->offsetStopLimit,
1131 nextOffset,
1132 nmembers + MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT * OFFSET_WARN_SEGMENTS))
1133 ereport(WARNING,
1134 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1135 errmsg_plural("database with OID %u must be vacuumed before %d more multixact member is used",
1136 "database with OID %u must be vacuumed before %d more multixact members are used",
1137 MultiXactState->offsetStopLimit - nextOffset + nmembers,
1138 MultiXactState->oldestMultiXactDB,
1139 MultiXactState->offsetStopLimit - nextOffset + nmembers),
1140 errhint("Execute a database-wide VACUUM in that database with reduced vacuum_multixact_freeze_min_age and vacuum_multixact_freeze_table_age settings.")));
1141
1142 ExtendMultiXactMember(nextOffset, nmembers);
1143
1144 /*
1145 * Critical section from here until caller has written the data into the
1146 * just-reserved SLRU space; we don't want to error out with a partly
1147 * written MultiXact structure. (In particular, failing to write our
1148 * start offset after advancing nextMXact would effectively corrupt the
1149 * previous MultiXact.)
1150 */
1151 START_CRIT_SECTION();
1152
1153 /*
1154 * Advance counters. As in GetNewTransactionId(), this must not happen
1155 * until after file extension has succeeded!
1156 *
1157 * We don't care about MultiXactId wraparound here; it will be handled by
1158 * the next iteration. But note that nextMXact may be InvalidMultiXactId
1159 * or the first value on a segment-beginning page after this routine
1160 * exits, so anyone else looking at the variable must be prepared to deal
1161 * with either case. Similarly, nextOffset may be zero, but we won't use
1162 * that as the actual start offset of the next multixact.
1163 */
1164 (MultiXactState->nextMXact)++;
1165
1166 MultiXactState->nextOffset += nmembers;
1167
1168 LWLockRelease(MultiXactGenLock);
1169
1170 debug_elog4(DEBUG2, "GetNew: returning %u offset %u", result, *offset);
1171 return result;
1172 }
1173
1174 /*
1175 * GetMultiXactIdMembers
1176 * Return the set of MultiXactMembers that make up a MultiXactId
1177 *
1178 * Return value is the number of members found, or -1 if there are none,
1179 * and *members is set to a newly palloc'ed array of members. It's the
1180 * caller's responsibility to free it when done with it.
1181 *
1182 * from_pgupgrade must be passed as true if and only if only the multixact
1183 * corresponds to a value from a tuple that was locked in a 9.2-or-older
1184 * installation and later pg_upgrade'd (that is, the infomask is
1185 * HEAP_LOCKED_UPGRADED). In this case, we know for certain that no members
1186 * can still be running, so we return -1 just like for an empty multixact
1187 * without any further checking. It would be wrong to try to resolve such a
1188 * multixact: either the multixact is within the current valid multixact
1189 * range, in which case the returned result would be bogus, or outside that
1190 * range, in which case an error would be raised.
1191 *
1192 * In all other cases, the passed multixact must be within the known valid
1193 * range, that is, greater to or equal than oldestMultiXactId, and less than
1194 * nextMXact. Otherwise, an error is raised.
1195 *
1196 * onlyLock must be set to true if caller is certain that the given multi
1197 * is used only to lock tuples; can be false without loss of correctness,
1198 * but passing a true means we can return quickly without checking for
1199 * old updates.
1200 */
1201 int
GetMultiXactIdMembers(MultiXactId multi,MultiXactMember ** members,bool from_pgupgrade,bool onlyLock)1202 GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
1203 bool from_pgupgrade, bool onlyLock)
1204 {
1205 int pageno;
1206 int prev_pageno;
1207 int entryno;
1208 int slotno;
1209 MultiXactOffset *offptr;
1210 MultiXactOffset offset;
1211 int length;
1212 int truelength;
1213 int i;
1214 MultiXactId oldestMXact;
1215 MultiXactId nextMXact;
1216 MultiXactId tmpMXact;
1217 MultiXactOffset nextOffset;
1218 MultiXactMember *ptr;
1219
1220 debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
1221
1222 if (!MultiXactIdIsValid(multi) || from_pgupgrade)
1223 {
1224 *members = NULL;
1225 return -1;
1226 }
1227
1228 /* See if the MultiXactId is in the local cache */
1229 length = mXactCacheGetById(multi, members);
1230 if (length >= 0)
1231 {
1232 debug_elog3(DEBUG2, "GetMembers: found %s in the cache",
1233 mxid_to_string(multi, length, *members));
1234 return length;
1235 }
1236
1237 /* Set our OldestVisibleMXactId[] entry if we didn't already */
1238 MultiXactIdSetOldestVisible();
1239
1240 /*
1241 * If we know the multi is used only for locking and not for updates, then
1242 * we can skip checking if the value is older than our oldest visible
1243 * multi. It cannot possibly still be running.
1244 */
1245 if (onlyLock &&
1246 MultiXactIdPrecedes(multi, OldestVisibleMXactId[MyBackendId]))
1247 {
1248 debug_elog2(DEBUG2, "GetMembers: a locker-only multi is too old");
1249 *members = NULL;
1250 return -1;
1251 }
1252
1253 /*
1254 * We check known limits on MultiXact before resorting to the SLRU area.
1255 *
1256 * An ID older than MultiXactState->oldestMultiXactId cannot possibly be
1257 * useful; it has already been removed, or will be removed shortly, by
1258 * truncation. If one is passed, an error is raised.
1259 *
1260 * Also, an ID >= nextMXact shouldn't ever be seen here; if it is seen, it
1261 * implies undetected ID wraparound has occurred. This raises a hard
1262 * error.
1263 *
1264 * Shared lock is enough here since we aren't modifying any global state.
1265 * Acquire it just long enough to grab the current counter values. We may
1266 * need both nextMXact and nextOffset; see below.
1267 */
1268 LWLockAcquire(MultiXactGenLock, LW_SHARED);
1269
1270 oldestMXact = MultiXactState->oldestMultiXactId;
1271 nextMXact = MultiXactState->nextMXact;
1272 nextOffset = MultiXactState->nextOffset;
1273
1274 LWLockRelease(MultiXactGenLock);
1275
1276 if (MultiXactIdPrecedes(multi, oldestMXact))
1277 ereport(ERROR,
1278 (errcode(ERRCODE_INTERNAL_ERROR),
1279 errmsg("MultiXactId %u does no longer exist -- apparent wraparound",
1280 multi)));
1281
1282 if (!MultiXactIdPrecedes(multi, nextMXact))
1283 ereport(ERROR,
1284 (errcode(ERRCODE_INTERNAL_ERROR),
1285 errmsg("MultiXactId %u has not been created yet -- apparent wraparound",
1286 multi)));
1287
1288 /*
1289 * Find out the offset at which we need to start reading MultiXactMembers
1290 * and the number of members in the multixact. We determine the latter as
1291 * the difference between this multixact's starting offset and the next
1292 * one's. However, there are some corner cases to worry about:
1293 *
1294 * 1. This multixact may be the latest one created, in which case there is
1295 * no next one to look at. In this case the nextOffset value we just
1296 * saved is the correct endpoint.
1297 *
1298 * 2. The next multixact may still be in process of being filled in: that
1299 * is, another process may have done GetNewMultiXactId but not yet written
1300 * the offset entry for that ID. In that scenario, it is guaranteed that
1301 * the offset entry for that multixact exists (because GetNewMultiXactId
1302 * won't release MultiXactGenLock until it does) but contains zero
1303 * (because we are careful to pre-zero offset pages). Because
1304 * GetNewMultiXactId will never return zero as the starting offset for a
1305 * multixact, when we read zero as the next multixact's offset, we know we
1306 * have this case. We sleep for a bit and try again.
1307 *
1308 * 3. Because GetNewMultiXactId increments offset zero to offset one to
1309 * handle case #2, there is an ambiguity near the point of offset
1310 * wraparound. If we see next multixact's offset is one, is that our
1311 * multixact's actual endpoint, or did it end at zero with a subsequent
1312 * increment? We handle this using the knowledge that if the zero'th
1313 * member slot wasn't filled, it'll contain zero, and zero isn't a valid
1314 * transaction ID so it can't be a multixact member. Therefore, if we
1315 * read a zero from the members array, just ignore it.
1316 *
1317 * This is all pretty messy, but the mess occurs only in infrequent corner
1318 * cases, so it seems better than holding the MultiXactGenLock for a long
1319 * time on every multixact creation.
1320 */
1321 retry:
1322 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
1323
1324 pageno = MultiXactIdToOffsetPage(multi);
1325 entryno = MultiXactIdToOffsetEntry(multi);
1326
1327 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
1328 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1329 offptr += entryno;
1330 offset = *offptr;
1331
1332 Assert(offset != 0);
1333
1334 /*
1335 * Use the same increment rule as GetNewMultiXactId(), that is, don't
1336 * handle wraparound explicitly until needed.
1337 */
1338 tmpMXact = multi + 1;
1339
1340 if (nextMXact == tmpMXact)
1341 {
1342 /* Corner case 1: there is no next multixact */
1343 length = nextOffset - offset;
1344 }
1345 else
1346 {
1347 MultiXactOffset nextMXOffset;
1348
1349 /* handle wraparound if needed */
1350 if (tmpMXact < FirstMultiXactId)
1351 tmpMXact = FirstMultiXactId;
1352
1353 prev_pageno = pageno;
1354
1355 pageno = MultiXactIdToOffsetPage(tmpMXact);
1356 entryno = MultiXactIdToOffsetEntry(tmpMXact);
1357
1358 if (pageno != prev_pageno)
1359 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
1360
1361 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
1362 offptr += entryno;
1363 nextMXOffset = *offptr;
1364
1365 if (nextMXOffset == 0)
1366 {
1367 /* Corner case 2: next multixact is still being filled in */
1368 LWLockRelease(MultiXactOffsetControlLock);
1369 CHECK_FOR_INTERRUPTS();
1370 pg_usleep(1000L);
1371 goto retry;
1372 }
1373
1374 length = nextMXOffset - offset;
1375 }
1376
1377 LWLockRelease(MultiXactOffsetControlLock);
1378
1379 ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
1380
1381 /* Now get the members themselves. */
1382 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
1383
1384 truelength = 0;
1385 prev_pageno = -1;
1386 for (i = 0; i < length; i++, offset++)
1387 {
1388 TransactionId *xactptr;
1389 uint32 *flagsptr;
1390 int flagsoff;
1391 int bshift;
1392 int memberoff;
1393
1394 pageno = MXOffsetToMemberPage(offset);
1395 memberoff = MXOffsetToMemberOffset(offset);
1396
1397 if (pageno != prev_pageno)
1398 {
1399 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
1400 prev_pageno = pageno;
1401 }
1402
1403 xactptr = (TransactionId *)
1404 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
1405
1406 if (!TransactionIdIsValid(*xactptr))
1407 {
1408 /* Corner case 3: we must be looking at unused slot zero */
1409 Assert(offset == 0);
1410 continue;
1411 }
1412
1413 flagsoff = MXOffsetToFlagsOffset(offset);
1414 bshift = MXOffsetToFlagsBitShift(offset);
1415 flagsptr = (uint32 *) (MultiXactMemberCtl->shared->page_buffer[slotno] + flagsoff);
1416
1417 ptr[truelength].xid = *xactptr;
1418 ptr[truelength].status = (*flagsptr >> bshift) & MXACT_MEMBER_XACT_BITMASK;
1419 truelength++;
1420 }
1421
1422 LWLockRelease(MultiXactMemberControlLock);
1423
1424 /* A multixid with zero members should not happen */
1425 Assert(truelength > 0);
1426
1427 /*
1428 * Copy the result into the local cache.
1429 */
1430 mXactCachePut(multi, truelength, ptr);
1431
1432 debug_elog3(DEBUG2, "GetMembers: no cache for %s",
1433 mxid_to_string(multi, truelength, ptr));
1434 *members = ptr;
1435 return truelength;
1436 }
1437
1438 /*
1439 * mxactMemberComparator
1440 * qsort comparison function for MultiXactMember
1441 *
1442 * We can't use wraparound comparison for XIDs because that does not respect
1443 * the triangle inequality! Any old sort order will do.
1444 */
1445 static int
mxactMemberComparator(const void * arg1,const void * arg2)1446 mxactMemberComparator(const void *arg1, const void *arg2)
1447 {
1448 MultiXactMember member1 = *(const MultiXactMember *) arg1;
1449 MultiXactMember member2 = *(const MultiXactMember *) arg2;
1450
1451 if (member1.xid > member2.xid)
1452 return 1;
1453 if (member1.xid < member2.xid)
1454 return -1;
1455 if (member1.status > member2.status)
1456 return 1;
1457 if (member1.status < member2.status)
1458 return -1;
1459 return 0;
1460 }
1461
1462 /*
1463 * mXactCacheGetBySet
1464 * returns a MultiXactId from the cache based on the set of
1465 * TransactionIds that compose it, or InvalidMultiXactId if
1466 * none matches.
1467 *
1468 * This is helpful, for example, if two transactions want to lock a huge
1469 * table. By using the cache, the second will use the same MultiXactId
1470 * for the majority of tuples, thus keeping MultiXactId usage low (saving
1471 * both I/O and wraparound issues).
1472 *
1473 * NB: the passed members array will be sorted in-place.
1474 */
1475 static MultiXactId
mXactCacheGetBySet(int nmembers,MultiXactMember * members)1476 mXactCacheGetBySet(int nmembers, MultiXactMember *members)
1477 {
1478 dlist_iter iter;
1479
1480 debug_elog3(DEBUG2, "CacheGet: looking for %s",
1481 mxid_to_string(InvalidMultiXactId, nmembers, members));
1482
1483 /* sort the array so comparison is easy */
1484 qsort(members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1485
1486 dlist_foreach(iter, &MXactCache)
1487 {
1488 mXactCacheEnt *entry = dlist_container(mXactCacheEnt, node, iter.cur);
1489
1490 if (entry->nmembers != nmembers)
1491 continue;
1492
1493 /*
1494 * We assume the cache entries are sorted, and that the unused bits in
1495 * "status" are zeroed.
1496 */
1497 if (memcmp(members, entry->members, nmembers * sizeof(MultiXactMember)) == 0)
1498 {
1499 debug_elog3(DEBUG2, "CacheGet: found %u", entry->multi);
1500 dlist_move_head(&MXactCache, iter.cur);
1501 return entry->multi;
1502 }
1503 }
1504
1505 debug_elog2(DEBUG2, "CacheGet: not found :-(");
1506 return InvalidMultiXactId;
1507 }
1508
1509 /*
1510 * mXactCacheGetById
1511 * returns the composing MultiXactMember set from the cache for a
1512 * given MultiXactId, if present.
1513 *
1514 * If successful, *xids is set to the address of a palloc'd copy of the
1515 * MultiXactMember set. Return value is number of members, or -1 on failure.
1516 */
1517 static int
mXactCacheGetById(MultiXactId multi,MultiXactMember ** members)1518 mXactCacheGetById(MultiXactId multi, MultiXactMember **members)
1519 {
1520 dlist_iter iter;
1521
1522 debug_elog3(DEBUG2, "CacheGet: looking for %u", multi);
1523
1524 dlist_foreach(iter, &MXactCache)
1525 {
1526 mXactCacheEnt *entry = dlist_container(mXactCacheEnt, node, iter.cur);
1527
1528 if (entry->multi == multi)
1529 {
1530 MultiXactMember *ptr;
1531 Size size;
1532
1533 size = sizeof(MultiXactMember) * entry->nmembers;
1534 ptr = (MultiXactMember *) palloc(size);
1535
1536 memcpy(ptr, entry->members, size);
1537
1538 debug_elog3(DEBUG2, "CacheGet: found %s",
1539 mxid_to_string(multi,
1540 entry->nmembers,
1541 entry->members));
1542
1543 /*
1544 * Note we modify the list while not using a modifiable iterator.
1545 * This is acceptable only because we exit the iteration
1546 * immediately afterwards.
1547 */
1548 dlist_move_head(&MXactCache, iter.cur);
1549
1550 *members = ptr;
1551 return entry->nmembers;
1552 }
1553 }
1554
1555 debug_elog2(DEBUG2, "CacheGet: not found");
1556 return -1;
1557 }
1558
1559 /*
1560 * mXactCachePut
1561 * Add a new MultiXactId and its composing set into the local cache.
1562 */
1563 static void
mXactCachePut(MultiXactId multi,int nmembers,MultiXactMember * members)1564 mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
1565 {
1566 mXactCacheEnt *entry;
1567
1568 debug_elog3(DEBUG2, "CachePut: storing %s",
1569 mxid_to_string(multi, nmembers, members));
1570
1571 if (MXactContext == NULL)
1572 {
1573 /* The cache only lives as long as the current transaction */
1574 debug_elog2(DEBUG2, "CachePut: initializing memory context");
1575 MXactContext = AllocSetContextCreate(TopTransactionContext,
1576 "MultiXact cache context",
1577 ALLOCSET_SMALL_SIZES);
1578 }
1579
1580 entry = (mXactCacheEnt *)
1581 MemoryContextAlloc(MXactContext,
1582 offsetof(mXactCacheEnt, members) +
1583 nmembers * sizeof(MultiXactMember));
1584
1585 entry->multi = multi;
1586 entry->nmembers = nmembers;
1587 memcpy(entry->members, members, nmembers * sizeof(MultiXactMember));
1588
1589 /* mXactCacheGetBySet assumes the entries are sorted, so sort them */
1590 qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
1591
1592 dlist_push_head(&MXactCache, &entry->node);
1593 if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
1594 {
1595 dlist_node *node;
1596 mXactCacheEnt *entry;
1597
1598 node = dlist_tail_node(&MXactCache);
1599 dlist_delete(node);
1600 MXactCacheMembers--;
1601
1602 entry = dlist_container(mXactCacheEnt, node, node);
1603 debug_elog3(DEBUG2, "CachePut: pruning cached multi %u",
1604 entry->multi);
1605
1606 pfree(entry);
1607 }
1608 }
1609
1610 static char *
mxstatus_to_string(MultiXactStatus status)1611 mxstatus_to_string(MultiXactStatus status)
1612 {
1613 switch (status)
1614 {
1615 case MultiXactStatusForKeyShare:
1616 return "keysh";
1617 case MultiXactStatusForShare:
1618 return "sh";
1619 case MultiXactStatusForNoKeyUpdate:
1620 return "fornokeyupd";
1621 case MultiXactStatusForUpdate:
1622 return "forupd";
1623 case MultiXactStatusNoKeyUpdate:
1624 return "nokeyupd";
1625 case MultiXactStatusUpdate:
1626 return "upd";
1627 default:
1628 elog(ERROR, "unrecognized multixact status %d", status);
1629 return "";
1630 }
1631 }
1632
1633 char *
mxid_to_string(MultiXactId multi,int nmembers,MultiXactMember * members)1634 mxid_to_string(MultiXactId multi, int nmembers, MultiXactMember *members)
1635 {
1636 static char *str = NULL;
1637 StringInfoData buf;
1638 int i;
1639
1640 if (str != NULL)
1641 pfree(str);
1642
1643 initStringInfo(&buf);
1644
1645 appendStringInfo(&buf, "%u %d[%u (%s)", multi, nmembers, members[0].xid,
1646 mxstatus_to_string(members[0].status));
1647
1648 for (i = 1; i < nmembers; i++)
1649 appendStringInfo(&buf, ", %u (%s)", members[i].xid,
1650 mxstatus_to_string(members[i].status));
1651
1652 appendStringInfoChar(&buf, ']');
1653 str = MemoryContextStrdup(TopMemoryContext, buf.data);
1654 pfree(buf.data);
1655 return str;
1656 }
1657
1658 /*
1659 * AtEOXact_MultiXact
1660 * Handle transaction end for MultiXact
1661 *
1662 * This is called at top transaction commit or abort (we don't care which).
1663 */
1664 void
AtEOXact_MultiXact(void)1665 AtEOXact_MultiXact(void)
1666 {
1667 /*
1668 * Reset our OldestMemberMXactId and OldestVisibleMXactId values, both of
1669 * which should only be valid while within a transaction.
1670 *
1671 * We assume that storing a MultiXactId is atomic and so we need not take
1672 * MultiXactGenLock to do this.
1673 */
1674 OldestMemberMXactId[MyBackendId] = InvalidMultiXactId;
1675 OldestVisibleMXactId[MyBackendId] = InvalidMultiXactId;
1676
1677 /*
1678 * Discard the local MultiXactId cache. Since MXactContext was created as
1679 * a child of TopTransactionContext, we needn't delete it explicitly.
1680 */
1681 MXactContext = NULL;
1682 dlist_init(&MXactCache);
1683 MXactCacheMembers = 0;
1684 }
1685
1686 /*
1687 * AtPrepare_MultiXact
1688 * Save multixact state at 2PC transaction prepare
1689 *
1690 * In this phase, we only store our OldestMemberMXactId value in the two-phase
1691 * state file.
1692 */
1693 void
AtPrepare_MultiXact(void)1694 AtPrepare_MultiXact(void)
1695 {
1696 MultiXactId myOldestMember = OldestMemberMXactId[MyBackendId];
1697
1698 if (MultiXactIdIsValid(myOldestMember))
1699 RegisterTwoPhaseRecord(TWOPHASE_RM_MULTIXACT_ID, 0,
1700 &myOldestMember, sizeof(MultiXactId));
1701 }
1702
1703 /*
1704 * PostPrepare_MultiXact
1705 * Clean up after successful PREPARE TRANSACTION
1706 */
1707 void
PostPrepare_MultiXact(TransactionId xid)1708 PostPrepare_MultiXact(TransactionId xid)
1709 {
1710 MultiXactId myOldestMember;
1711
1712 /*
1713 * Transfer our OldestMemberMXactId value to the slot reserved for the
1714 * prepared transaction.
1715 */
1716 myOldestMember = OldestMemberMXactId[MyBackendId];
1717 if (MultiXactIdIsValid(myOldestMember))
1718 {
1719 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid);
1720
1721 /*
1722 * Even though storing MultiXactId is atomic, acquire lock to make
1723 * sure others see both changes, not just the reset of the slot of the
1724 * current backend. Using a volatile pointer might suffice, but this
1725 * isn't a hot spot.
1726 */
1727 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
1728
1729 OldestMemberMXactId[dummyBackendId] = myOldestMember;
1730 OldestMemberMXactId[MyBackendId] = InvalidMultiXactId;
1731
1732 LWLockRelease(MultiXactGenLock);
1733 }
1734
1735 /*
1736 * We don't need to transfer OldestVisibleMXactId value, because the
1737 * transaction is not going to be looking at any more multixacts once it's
1738 * prepared.
1739 *
1740 * We assume that storing a MultiXactId is atomic and so we need not take
1741 * MultiXactGenLock to do this.
1742 */
1743 OldestVisibleMXactId[MyBackendId] = InvalidMultiXactId;
1744
1745 /*
1746 * Discard the local MultiXactId cache like in AtEOX_MultiXact
1747 */
1748 MXactContext = NULL;
1749 dlist_init(&MXactCache);
1750 MXactCacheMembers = 0;
1751 }
1752
1753 /*
1754 * multixact_twophase_recover
1755 * Recover the state of a prepared transaction at startup
1756 */
1757 void
multixact_twophase_recover(TransactionId xid,uint16 info,void * recdata,uint32 len)1758 multixact_twophase_recover(TransactionId xid, uint16 info,
1759 void *recdata, uint32 len)
1760 {
1761 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid);
1762 MultiXactId oldestMember;
1763
1764 /*
1765 * Get the oldest member XID from the state file record, and set it in the
1766 * OldestMemberMXactId slot reserved for this prepared transaction.
1767 */
1768 Assert(len == sizeof(MultiXactId));
1769 oldestMember = *((MultiXactId *) recdata);
1770
1771 OldestMemberMXactId[dummyBackendId] = oldestMember;
1772 }
1773
1774 /*
1775 * multixact_twophase_postcommit
1776 * Similar to AtEOX_MultiXact but for COMMIT PREPARED
1777 */
1778 void
multixact_twophase_postcommit(TransactionId xid,uint16 info,void * recdata,uint32 len)1779 multixact_twophase_postcommit(TransactionId xid, uint16 info,
1780 void *recdata, uint32 len)
1781 {
1782 BackendId dummyBackendId = TwoPhaseGetDummyBackendId(xid);
1783
1784 Assert(len == sizeof(MultiXactId));
1785
1786 OldestMemberMXactId[dummyBackendId] = InvalidMultiXactId;
1787 }
1788
1789 /*
1790 * multixact_twophase_postabort
1791 * This is actually just the same as the COMMIT case.
1792 */
1793 void
multixact_twophase_postabort(TransactionId xid,uint16 info,void * recdata,uint32 len)1794 multixact_twophase_postabort(TransactionId xid, uint16 info,
1795 void *recdata, uint32 len)
1796 {
1797 multixact_twophase_postcommit(xid, info, recdata, len);
1798 }
1799
1800 /*
1801 * Initialization of shared memory for MultiXact. We use two SLRU areas,
1802 * thus double memory. Also, reserve space for the shared MultiXactState
1803 * struct and the per-backend MultiXactId arrays (two of those, too).
1804 */
1805 Size
MultiXactShmemSize(void)1806 MultiXactShmemSize(void)
1807 {
1808 Size size;
1809
1810 /* We need 2*MaxOldestSlot + 1 perBackendXactIds[] entries */
1811 #define SHARED_MULTIXACT_STATE_SIZE \
1812 add_size(offsetof(MultiXactStateData, perBackendXactIds) + sizeof(MultiXactId), \
1813 mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
1814
1815 size = SHARED_MULTIXACT_STATE_SIZE;
1816 size = add_size(size, SimpleLruShmemSize(NUM_MXACTOFFSET_BUFFERS, 0));
1817 size = add_size(size, SimpleLruShmemSize(NUM_MXACTMEMBER_BUFFERS, 0));
1818
1819 return size;
1820 }
1821
1822 void
MultiXactShmemInit(void)1823 MultiXactShmemInit(void)
1824 {
1825 bool found;
1826
1827 debug_elog2(DEBUG2, "Shared Memory Init for MultiXact");
1828
1829 MultiXactOffsetCtl->PagePrecedes = MultiXactOffsetPagePrecedes;
1830 MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
1831
1832 SimpleLruInit(MultiXactOffsetCtl,
1833 "multixact_offset", NUM_MXACTOFFSET_BUFFERS, 0,
1834 MultiXactOffsetControlLock, "pg_multixact/offsets",
1835 LWTRANCHE_MXACTOFFSET_BUFFERS);
1836 SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
1837 SimpleLruInit(MultiXactMemberCtl,
1838 "multixact_member", NUM_MXACTMEMBER_BUFFERS, 0,
1839 MultiXactMemberControlLock, "pg_multixact/members",
1840 LWTRANCHE_MXACTMEMBER_BUFFERS);
1841 /* doesn't call SimpleLruTruncate() or meet criteria for unit tests */
1842
1843 /* Initialize our shared state struct */
1844 MultiXactState = ShmemInitStruct("Shared MultiXact State",
1845 SHARED_MULTIXACT_STATE_SIZE,
1846 &found);
1847 if (!IsUnderPostmaster)
1848 {
1849 Assert(!found);
1850
1851 /* Make sure we zero out the per-backend state */
1852 MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
1853 }
1854 else
1855 Assert(found);
1856
1857 /*
1858 * Set up array pointers. Note that perBackendXactIds[0] is wasted space
1859 * since we only use indexes 1..MaxOldestSlot in each array.
1860 */
1861 OldestMemberMXactId = MultiXactState->perBackendXactIds;
1862 OldestVisibleMXactId = OldestMemberMXactId + MaxOldestSlot;
1863 }
1864
1865 /*
1866 * This func must be called ONCE on system install. It creates the initial
1867 * MultiXact segments. (The MultiXacts directories are assumed to have been
1868 * created by initdb, and MultiXactShmemInit must have been called already.)
1869 */
1870 void
BootStrapMultiXact(void)1871 BootStrapMultiXact(void)
1872 {
1873 int slotno;
1874
1875 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
1876
1877 /* Create and zero the first page of the offsets log */
1878 slotno = ZeroMultiXactOffsetPage(0, false);
1879
1880 /* Make sure it's written out */
1881 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
1882 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
1883
1884 LWLockRelease(MultiXactOffsetControlLock);
1885
1886 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
1887
1888 /* Create and zero the first page of the members log */
1889 slotno = ZeroMultiXactMemberPage(0, false);
1890
1891 /* Make sure it's written out */
1892 SimpleLruWritePage(MultiXactMemberCtl, slotno);
1893 Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
1894
1895 LWLockRelease(MultiXactMemberControlLock);
1896 }
1897
1898 /*
1899 * Initialize (or reinitialize) a page of MultiXactOffset to zeroes.
1900 * If writeXlog is TRUE, also emit an XLOG record saying we did this.
1901 *
1902 * The page is not actually written, just set up in shared memory.
1903 * The slot number of the new page is returned.
1904 *
1905 * Control lock must be held at entry, and will be held at exit.
1906 */
1907 static int
ZeroMultiXactOffsetPage(int pageno,bool writeXlog)1908 ZeroMultiXactOffsetPage(int pageno, bool writeXlog)
1909 {
1910 int slotno;
1911
1912 slotno = SimpleLruZeroPage(MultiXactOffsetCtl, pageno);
1913
1914 if (writeXlog)
1915 WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_OFF_PAGE);
1916
1917 return slotno;
1918 }
1919
1920 /*
1921 * Ditto, for MultiXactMember
1922 */
1923 static int
ZeroMultiXactMemberPage(int pageno,bool writeXlog)1924 ZeroMultiXactMemberPage(int pageno, bool writeXlog)
1925 {
1926 int slotno;
1927
1928 slotno = SimpleLruZeroPage(MultiXactMemberCtl, pageno);
1929
1930 if (writeXlog)
1931 WriteMZeroPageXlogRec(pageno, XLOG_MULTIXACT_ZERO_MEM_PAGE);
1932
1933 return slotno;
1934 }
1935
1936 /*
1937 * MaybeExtendOffsetSlru
1938 * Extend the offsets SLRU area, if necessary
1939 *
1940 * After a binary upgrade from <= 9.2, the pg_multixact/offset SLRU area might
1941 * contain files that are shorter than necessary; this would occur if the old
1942 * installation had used multixacts beyond the first page (files cannot be
1943 * copied, because the on-disk representation is different). pg_upgrade would
1944 * update pg_control to set the next offset value to be at that position, so
1945 * that tuples marked as locked by such MultiXacts would be seen as visible
1946 * without having to consult multixact. However, trying to create and use a
1947 * new MultiXactId would result in an error because the page on which the new
1948 * value would reside does not exist. This routine is in charge of creating
1949 * such pages.
1950 */
1951 static void
MaybeExtendOffsetSlru(void)1952 MaybeExtendOffsetSlru(void)
1953 {
1954 int pageno;
1955
1956 pageno = MultiXactIdToOffsetPage(MultiXactState->nextMXact);
1957
1958 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
1959
1960 if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
1961 {
1962 int slotno;
1963
1964 /*
1965 * Fortunately for us, SimpleLruWritePage is already prepared to deal
1966 * with creating a new segment file even if the page we're writing is
1967 * not the first in it, so this is enough.
1968 */
1969 slotno = ZeroMultiXactOffsetPage(pageno, false);
1970 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
1971 }
1972
1973 LWLockRelease(MultiXactOffsetControlLock);
1974 }
1975
1976 /*
1977 * This must be called ONCE during postmaster or standalone-backend startup.
1978 *
1979 * StartupXLOG has already established nextMXact/nextOffset by calling
1980 * MultiXactSetNextMXact and/or MultiXactAdvanceNextMXact, and the oldestMulti
1981 * info from pg_control and/or MultiXactAdvanceOldest, but we haven't yet
1982 * replayed WAL.
1983 */
1984 void
StartupMultiXact(void)1985 StartupMultiXact(void)
1986 {
1987 MultiXactId multi = MultiXactState->nextMXact;
1988 MultiXactOffset offset = MultiXactState->nextOffset;
1989 int pageno;
1990
1991 /*
1992 * Initialize offset's idea of the latest page number.
1993 */
1994 pageno = MultiXactIdToOffsetPage(multi);
1995 MultiXactOffsetCtl->shared->latest_page_number = pageno;
1996
1997 /*
1998 * Initialize member's idea of the latest page number.
1999 */
2000 pageno = MXOffsetToMemberPage(offset);
2001 MultiXactMemberCtl->shared->latest_page_number = pageno;
2002 }
2003
2004 /*
2005 * This must be called ONCE at the end of startup/recovery.
2006 */
2007 void
TrimMultiXact(void)2008 TrimMultiXact(void)
2009 {
2010 MultiXactId nextMXact;
2011 MultiXactOffset offset;
2012 MultiXactId oldestMXact;
2013 Oid oldestMXactDB;
2014 int pageno;
2015 int entryno;
2016 int flagsoff;
2017
2018 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2019 nextMXact = MultiXactState->nextMXact;
2020 offset = MultiXactState->nextOffset;
2021 oldestMXact = MultiXactState->oldestMultiXactId;
2022 oldestMXactDB = MultiXactState->oldestMultiXactDB;
2023 LWLockRelease(MultiXactGenLock);
2024
2025 /* Clean up offsets state */
2026 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
2027
2028 /*
2029 * (Re-)Initialize our idea of the latest page number for offsets.
2030 */
2031 pageno = MultiXactIdToOffsetPage(nextMXact);
2032 MultiXactOffsetCtl->shared->latest_page_number = pageno;
2033
2034 /*
2035 * Zero out the remainder of the current offsets page. See notes in
2036 * TrimCLOG() for background. Unlike CLOG, some WAL record covers every
2037 * pg_multixact SLRU mutation. Since, also unlike CLOG, we ignore the WAL
2038 * rule "write xlog before data," nextMXact successors may carry obsolete,
2039 * nonzero offset values. Zero those so case 2 of GetMultiXactIdMembers()
2040 * operates normally.
2041 */
2042 entryno = MultiXactIdToOffsetEntry(nextMXact);
2043 if (entryno != 0)
2044 {
2045 int slotno;
2046 MultiXactOffset *offptr;
2047
2048 slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, nextMXact);
2049 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2050 offptr += entryno;
2051
2052 MemSet(offptr, 0, BLCKSZ - (entryno * sizeof(MultiXactOffset)));
2053
2054 MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
2055 }
2056
2057 LWLockRelease(MultiXactOffsetControlLock);
2058
2059 /* And the same for members */
2060 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
2061
2062 /*
2063 * (Re-)Initialize our idea of the latest page number for members.
2064 */
2065 pageno = MXOffsetToMemberPage(offset);
2066 MultiXactMemberCtl->shared->latest_page_number = pageno;
2067
2068 /*
2069 * Zero out the remainder of the current members page. See notes in
2070 * TrimCLOG() for motivation.
2071 */
2072 flagsoff = MXOffsetToFlagsOffset(offset);
2073 if (flagsoff != 0)
2074 {
2075 int slotno;
2076 TransactionId *xidptr;
2077 int memberoff;
2078
2079 memberoff = MXOffsetToMemberOffset(offset);
2080 slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, offset);
2081 xidptr = (TransactionId *)
2082 (MultiXactMemberCtl->shared->page_buffer[slotno] + memberoff);
2083
2084 MemSet(xidptr, 0, BLCKSZ - memberoff);
2085
2086 /*
2087 * Note: we don't need to zero out the flag bits in the remaining
2088 * members of the current group, because they are always reset before
2089 * writing.
2090 */
2091
2092 MultiXactMemberCtl->shared->page_dirty[slotno] = true;
2093 }
2094
2095 LWLockRelease(MultiXactMemberControlLock);
2096
2097 /* signal that we're officially up */
2098 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2099 MultiXactState->finishedStartup = true;
2100 LWLockRelease(MultiXactGenLock);
2101
2102 /* Now compute how far away the next members wraparound is. */
2103 SetMultiXactIdLimit(oldestMXact, oldestMXactDB, true);
2104 }
2105
2106 /*
2107 * This must be called ONCE during postmaster or standalone-backend shutdown
2108 */
2109 void
ShutdownMultiXact(void)2110 ShutdownMultiXact(void)
2111 {
2112 /* Flush dirty MultiXact pages to disk */
2113 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(false);
2114 SimpleLruFlush(MultiXactOffsetCtl, false);
2115 SimpleLruFlush(MultiXactMemberCtl, false);
2116 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(false);
2117 }
2118
2119 /*
2120 * Get the MultiXact data to save in a checkpoint record
2121 */
2122 void
MultiXactGetCheckptMulti(bool is_shutdown,MultiXactId * nextMulti,MultiXactOffset * nextMultiOffset,MultiXactId * oldestMulti,Oid * oldestMultiDB)2123 MultiXactGetCheckptMulti(bool is_shutdown,
2124 MultiXactId *nextMulti,
2125 MultiXactOffset *nextMultiOffset,
2126 MultiXactId *oldestMulti,
2127 Oid *oldestMultiDB)
2128 {
2129 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2130 *nextMulti = MultiXactState->nextMXact;
2131 *nextMultiOffset = MultiXactState->nextOffset;
2132 *oldestMulti = MultiXactState->oldestMultiXactId;
2133 *oldestMultiDB = MultiXactState->oldestMultiXactDB;
2134 LWLockRelease(MultiXactGenLock);
2135
2136 debug_elog6(DEBUG2,
2137 "MultiXact: checkpoint is nextMulti %u, nextOffset %u, oldestMulti %u in DB %u",
2138 *nextMulti, *nextMultiOffset, *oldestMulti, *oldestMultiDB);
2139 }
2140
2141 /*
2142 * Perform a checkpoint --- either during shutdown, or on-the-fly
2143 */
2144 void
CheckPointMultiXact(void)2145 CheckPointMultiXact(void)
2146 {
2147 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_START(true);
2148
2149 /* Flush dirty MultiXact pages to disk */
2150 SimpleLruFlush(MultiXactOffsetCtl, true);
2151 SimpleLruFlush(MultiXactMemberCtl, true);
2152
2153 TRACE_POSTGRESQL_MULTIXACT_CHECKPOINT_DONE(true);
2154 }
2155
2156 /*
2157 * Set the next-to-be-assigned MultiXactId and offset
2158 *
2159 * This is used when we can determine the correct next ID/offset exactly
2160 * from a checkpoint record. Although this is only called during bootstrap
2161 * and XLog replay, we take the lock in case any hot-standby backends are
2162 * examining the values.
2163 */
2164 void
MultiXactSetNextMXact(MultiXactId nextMulti,MultiXactOffset nextMultiOffset)2165 MultiXactSetNextMXact(MultiXactId nextMulti,
2166 MultiXactOffset nextMultiOffset)
2167 {
2168 debug_elog4(DEBUG2, "MultiXact: setting next multi to %u offset %u",
2169 nextMulti, nextMultiOffset);
2170 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2171 MultiXactState->nextMXact = nextMulti;
2172 MultiXactState->nextOffset = nextMultiOffset;
2173 LWLockRelease(MultiXactGenLock);
2174
2175 /*
2176 * During a binary upgrade, make sure that the offsets SLRU is large
2177 * enough to contain the next value that would be created.
2178 *
2179 * We need to do this pretty early during the first startup in binary
2180 * upgrade mode: before StartupMultiXact() in fact, because this routine
2181 * is called even before that by StartupXLOG(). And we can't do it
2182 * earlier than at this point, because during that first call of this
2183 * routine we determine the MultiXactState->nextMXact value that
2184 * MaybeExtendOffsetSlru needs.
2185 */
2186 if (IsBinaryUpgrade)
2187 MaybeExtendOffsetSlru();
2188 }
2189
2190 /*
2191 * Determine the last safe MultiXactId to allocate given the currently oldest
2192 * datminmxid (ie, the oldest MultiXactId that might exist in any database
2193 * of our cluster), and the OID of the (or a) database with that value.
2194 *
2195 * is_startup is true when we are just starting the cluster, false when we
2196 * are updating state in a running cluster. This only affects log messages.
2197 */
2198 void
SetMultiXactIdLimit(MultiXactId oldest_datminmxid,Oid oldest_datoid,bool is_startup)2199 SetMultiXactIdLimit(MultiXactId oldest_datminmxid, Oid oldest_datoid,
2200 bool is_startup)
2201 {
2202 MultiXactId multiVacLimit;
2203 MultiXactId multiWarnLimit;
2204 MultiXactId multiStopLimit;
2205 MultiXactId multiWrapLimit;
2206 MultiXactId curMulti;
2207 bool needs_offset_vacuum;
2208
2209 Assert(MultiXactIdIsValid(oldest_datminmxid));
2210
2211 /*
2212 * We pretend that a wrap will happen halfway through the multixact ID
2213 * space, but that's not really true, because multixacts wrap differently
2214 * from transaction IDs. Note that, separately from any concern about
2215 * multixact IDs wrapping, we must ensure that multixact members do not
2216 * wrap. Limits for that are set in DetermineSafeOldestOffset, not here.
2217 */
2218 multiWrapLimit = oldest_datminmxid + (MaxMultiXactId >> 1);
2219 if (multiWrapLimit < FirstMultiXactId)
2220 multiWrapLimit += FirstMultiXactId;
2221
2222 /*
2223 * We'll refuse to continue assigning MultiXactIds once we get within 100
2224 * multi of data loss.
2225 *
2226 * Note: This differs from the magic number used in
2227 * SetTransactionIdLimit() since vacuum itself will never generate new
2228 * multis. XXX actually it does, if it needs to freeze old multis.
2229 */
2230 multiStopLimit = multiWrapLimit - 100;
2231 if (multiStopLimit < FirstMultiXactId)
2232 multiStopLimit -= FirstMultiXactId;
2233
2234 /*
2235 * We'll start complaining loudly when we get within 10M multis of the
2236 * stop point. This is kind of arbitrary, but if you let your gas gauge
2237 * get down to 1% of full, would you be looking for the next gas station?
2238 * We need to be fairly liberal about this number because there are lots
2239 * of scenarios where most transactions are done by automatic clients that
2240 * won't pay attention to warnings. (No, we're not gonna make this
2241 * configurable. If you know enough to configure it, you know enough to
2242 * not get in this kind of trouble in the first place.)
2243 */
2244 multiWarnLimit = multiStopLimit - 10000000;
2245 if (multiWarnLimit < FirstMultiXactId)
2246 multiWarnLimit -= FirstMultiXactId;
2247
2248 /*
2249 * We'll start trying to force autovacuums when oldest_datminmxid gets to
2250 * be more than autovacuum_multixact_freeze_max_age mxids old.
2251 *
2252 * Note: autovacuum_multixact_freeze_max_age is a PGC_POSTMASTER parameter
2253 * so that we don't have to worry about dealing with on-the-fly changes in
2254 * its value. See SetTransactionIdLimit.
2255 */
2256 multiVacLimit = oldest_datminmxid + autovacuum_multixact_freeze_max_age;
2257 if (multiVacLimit < FirstMultiXactId)
2258 multiVacLimit += FirstMultiXactId;
2259
2260 /* Grab lock for just long enough to set the new limit values */
2261 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2262 MultiXactState->oldestMultiXactId = oldest_datminmxid;
2263 MultiXactState->oldestMultiXactDB = oldest_datoid;
2264 MultiXactState->multiVacLimit = multiVacLimit;
2265 MultiXactState->multiWarnLimit = multiWarnLimit;
2266 MultiXactState->multiStopLimit = multiStopLimit;
2267 MultiXactState->multiWrapLimit = multiWrapLimit;
2268 curMulti = MultiXactState->nextMXact;
2269 LWLockRelease(MultiXactGenLock);
2270
2271 /* Log the info */
2272 ereport(DEBUG1,
2273 (errmsg("MultiXactId wrap limit is %u, limited by database with OID %u",
2274 multiWrapLimit, oldest_datoid)));
2275
2276 /*
2277 * Computing the actual limits is only possible once the data directory is
2278 * in a consistent state. There's no need to compute the limits while
2279 * still replaying WAL - no decisions about new multis are made even
2280 * though multixact creations might be replayed. So we'll only do further
2281 * checks after TrimMultiXact() has been called.
2282 */
2283 if (!MultiXactState->finishedStartup)
2284 return;
2285
2286 Assert(!InRecovery);
2287
2288 /* Set limits for offset vacuum. */
2289 needs_offset_vacuum = SetOffsetVacuumLimit(is_startup);
2290
2291 /*
2292 * If past the autovacuum force point, immediately signal an autovac
2293 * request. The reason for this is that autovac only processes one
2294 * database per invocation. Once it's finished cleaning up the oldest
2295 * database, it'll call here, and we'll signal the postmaster to start
2296 * another iteration immediately if there are still any old databases.
2297 */
2298 if ((MultiXactIdPrecedes(multiVacLimit, curMulti) ||
2299 needs_offset_vacuum) && IsUnderPostmaster)
2300 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
2301
2302 /* Give an immediate warning if past the wrap warn point */
2303 if (MultiXactIdPrecedes(multiWarnLimit, curMulti))
2304 {
2305 char *oldest_datname;
2306
2307 /*
2308 * We can be called when not inside a transaction, for example during
2309 * StartupXLOG(). In such a case we cannot do database access, so we
2310 * must just report the oldest DB's OID.
2311 *
2312 * Note: it's also possible that get_database_name fails and returns
2313 * NULL, for example because the database just got dropped. We'll
2314 * still warn, even though the warning might now be unnecessary.
2315 */
2316 if (IsTransactionState())
2317 oldest_datname = get_database_name(oldest_datoid);
2318 else
2319 oldest_datname = NULL;
2320
2321 if (oldest_datname)
2322 ereport(WARNING,
2323 (errmsg_plural("database \"%s\" must be vacuumed before %u more MultiXactId is used",
2324 "database \"%s\" must be vacuumed before %u more MultiXactIds are used",
2325 multiWrapLimit - curMulti,
2326 oldest_datname,
2327 multiWrapLimit - curMulti),
2328 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
2329 "You might also need to commit or roll back old prepared transactions.")));
2330 else
2331 ereport(WARNING,
2332 (errmsg_plural("database with OID %u must be vacuumed before %u more MultiXactId is used",
2333 "database with OID %u must be vacuumed before %u more MultiXactIds are used",
2334 multiWrapLimit - curMulti,
2335 oldest_datoid,
2336 multiWrapLimit - curMulti),
2337 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
2338 "You might also need to commit or roll back old prepared transactions.")));
2339 }
2340 }
2341
2342 /*
2343 * Ensure the next-to-be-assigned MultiXactId is at least minMulti,
2344 * and similarly nextOffset is at least minMultiOffset.
2345 *
2346 * This is used when we can determine minimum safe values from an XLog
2347 * record (either an on-line checkpoint or an mxact creation log entry).
2348 * Although this is only called during XLog replay, we take the lock in case
2349 * any hot-standby backends are examining the values.
2350 */
2351 void
MultiXactAdvanceNextMXact(MultiXactId minMulti,MultiXactOffset minMultiOffset)2352 MultiXactAdvanceNextMXact(MultiXactId minMulti,
2353 MultiXactOffset minMultiOffset)
2354 {
2355 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2356 if (MultiXactIdPrecedes(MultiXactState->nextMXact, minMulti))
2357 {
2358 debug_elog3(DEBUG2, "MultiXact: setting next multi to %u", minMulti);
2359 MultiXactState->nextMXact = minMulti;
2360 }
2361 if (MultiXactOffsetPrecedes(MultiXactState->nextOffset, minMultiOffset))
2362 {
2363 debug_elog3(DEBUG2, "MultiXact: setting next offset to %u",
2364 minMultiOffset);
2365 MultiXactState->nextOffset = minMultiOffset;
2366 }
2367 LWLockRelease(MultiXactGenLock);
2368 }
2369
2370 /*
2371 * Update our oldestMultiXactId value, but only if it's more recent than what
2372 * we had.
2373 *
2374 * This may only be called during WAL replay.
2375 */
2376 void
MultiXactAdvanceOldest(MultiXactId oldestMulti,Oid oldestMultiDB)2377 MultiXactAdvanceOldest(MultiXactId oldestMulti, Oid oldestMultiDB)
2378 {
2379 Assert(InRecovery);
2380
2381 if (MultiXactIdPrecedes(MultiXactState->oldestMultiXactId, oldestMulti))
2382 SetMultiXactIdLimit(oldestMulti, oldestMultiDB, false);
2383 }
2384
2385 /*
2386 * Make sure that MultiXactOffset has room for a newly-allocated MultiXactId.
2387 *
2388 * NB: this is called while holding MultiXactGenLock. We want it to be very
2389 * fast most of the time; even when it's not so fast, no actual I/O need
2390 * happen unless we're forced to write out a dirty log or xlog page to make
2391 * room in shared memory.
2392 */
2393 static void
ExtendMultiXactOffset(MultiXactId multi)2394 ExtendMultiXactOffset(MultiXactId multi)
2395 {
2396 int pageno;
2397
2398 /*
2399 * No work except at first MultiXactId of a page. But beware: just after
2400 * wraparound, the first MultiXactId of page zero is FirstMultiXactId.
2401 */
2402 if (MultiXactIdToOffsetEntry(multi) != 0 &&
2403 multi != FirstMultiXactId)
2404 return;
2405
2406 pageno = MultiXactIdToOffsetPage(multi);
2407
2408 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
2409
2410 /* Zero the page and make an XLOG entry about it */
2411 ZeroMultiXactOffsetPage(pageno, true);
2412
2413 LWLockRelease(MultiXactOffsetControlLock);
2414 }
2415
2416 /*
2417 * Make sure that MultiXactMember has room for the members of a newly-
2418 * allocated MultiXactId.
2419 *
2420 * Like the above routine, this is called while holding MultiXactGenLock;
2421 * same comments apply.
2422 */
2423 static void
ExtendMultiXactMember(MultiXactOffset offset,int nmembers)2424 ExtendMultiXactMember(MultiXactOffset offset, int nmembers)
2425 {
2426 /*
2427 * It's possible that the members span more than one page of the members
2428 * file, so we loop to ensure we consider each page. The coding is not
2429 * optimal if the members span several pages, but that seems unusual
2430 * enough to not worry much about.
2431 */
2432 while (nmembers > 0)
2433 {
2434 int flagsoff;
2435 int flagsbit;
2436 uint32 difference;
2437
2438 /*
2439 * Only zero when at first entry of a page.
2440 */
2441 flagsoff = MXOffsetToFlagsOffset(offset);
2442 flagsbit = MXOffsetToFlagsBitShift(offset);
2443 if (flagsoff == 0 && flagsbit == 0)
2444 {
2445 int pageno;
2446
2447 pageno = MXOffsetToMemberPage(offset);
2448
2449 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
2450
2451 /* Zero the page and make an XLOG entry about it */
2452 ZeroMultiXactMemberPage(pageno, true);
2453
2454 LWLockRelease(MultiXactMemberControlLock);
2455 }
2456
2457 /*
2458 * Compute the number of items till end of current page. Careful: if
2459 * addition of unsigned ints wraps around, we're at the last page of
2460 * the last segment; since that page holds a different number of items
2461 * than other pages, we need to do it differently.
2462 */
2463 if (offset + MAX_MEMBERS_IN_LAST_MEMBERS_PAGE < offset)
2464 {
2465 /*
2466 * This is the last page of the last segment; we can compute the
2467 * number of items left to allocate in it without modulo
2468 * arithmetic.
2469 */
2470 difference = MaxMultiXactOffset - offset + 1;
2471 }
2472 else
2473 difference = MULTIXACT_MEMBERS_PER_PAGE - offset % MULTIXACT_MEMBERS_PER_PAGE;
2474
2475 /*
2476 * Advance to next page, taking care to properly handle the wraparound
2477 * case. OK if nmembers goes negative.
2478 */
2479 nmembers -= difference;
2480 offset += difference;
2481 }
2482 }
2483
2484 /*
2485 * GetOldestMultiXactId
2486 *
2487 * Return the oldest MultiXactId that's still possibly still seen as live by
2488 * any running transaction. Older ones might still exist on disk, but they no
2489 * longer have any running member transaction.
2490 *
2491 * It's not safe to truncate MultiXact SLRU segments on the value returned by
2492 * this function; however, it can be used by a full-table vacuum to set the
2493 * point at which it will be possible to truncate SLRU for that table.
2494 */
2495 MultiXactId
GetOldestMultiXactId(void)2496 GetOldestMultiXactId(void)
2497 {
2498 MultiXactId oldestMXact;
2499 MultiXactId nextMXact;
2500 int i;
2501
2502 /*
2503 * This is the oldest valid value among all the OldestMemberMXactId[] and
2504 * OldestVisibleMXactId[] entries, or nextMXact if none are valid.
2505 */
2506 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2507
2508 /*
2509 * We have to beware of the possibility that nextMXact is in the
2510 * wrapped-around state. We don't fix the counter itself here, but we
2511 * must be sure to use a valid value in our calculation.
2512 */
2513 nextMXact = MultiXactState->nextMXact;
2514 if (nextMXact < FirstMultiXactId)
2515 nextMXact = FirstMultiXactId;
2516
2517 oldestMXact = nextMXact;
2518 for (i = 1; i <= MaxOldestSlot; i++)
2519 {
2520 MultiXactId thisoldest;
2521
2522 thisoldest = OldestMemberMXactId[i];
2523 if (MultiXactIdIsValid(thisoldest) &&
2524 MultiXactIdPrecedes(thisoldest, oldestMXact))
2525 oldestMXact = thisoldest;
2526 thisoldest = OldestVisibleMXactId[i];
2527 if (MultiXactIdIsValid(thisoldest) &&
2528 MultiXactIdPrecedes(thisoldest, oldestMXact))
2529 oldestMXact = thisoldest;
2530 }
2531
2532 LWLockRelease(MultiXactGenLock);
2533
2534 return oldestMXact;
2535 }
2536
2537 /*
2538 * Determine how aggressively we need to vacuum in order to prevent member
2539 * wraparound.
2540 *
2541 * To do so determine what's the oldest member offset and install the limit
2542 * info in MultiXactState, where it can be used to prevent overrun of old data
2543 * in the members SLRU area.
2544 *
2545 * The return value is true if emergency autovacuum is required and false
2546 * otherwise.
2547 */
2548 static bool
SetOffsetVacuumLimit(bool is_startup)2549 SetOffsetVacuumLimit(bool is_startup)
2550 {
2551 MultiXactId oldestMultiXactId;
2552 MultiXactId nextMXact;
2553 MultiXactOffset oldestOffset = 0; /* placate compiler */
2554 MultiXactOffset prevOldestOffset;
2555 MultiXactOffset nextOffset;
2556 bool oldestOffsetKnown = false;
2557 bool prevOldestOffsetKnown;
2558 MultiXactOffset offsetStopLimit = 0;
2559 MultiXactOffset prevOffsetStopLimit;
2560
2561 /*
2562 * NB: Have to prevent concurrent truncation, we might otherwise try to
2563 * lookup an oldestMulti that's concurrently getting truncated away.
2564 */
2565 LWLockAcquire(MultiXactTruncationLock, LW_SHARED);
2566
2567 /* Read relevant fields from shared memory. */
2568 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2569 oldestMultiXactId = MultiXactState->oldestMultiXactId;
2570 nextMXact = MultiXactState->nextMXact;
2571 nextOffset = MultiXactState->nextOffset;
2572 prevOldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2573 prevOldestOffset = MultiXactState->oldestOffset;
2574 prevOffsetStopLimit = MultiXactState->offsetStopLimit;
2575 Assert(MultiXactState->finishedStartup);
2576 LWLockRelease(MultiXactGenLock);
2577
2578 /*
2579 * Determine the offset of the oldest multixact. Normally, we can read
2580 * the offset from the multixact itself, but there's an important special
2581 * case: if there are no multixacts in existence at all, oldestMXact
2582 * obviously can't point to one. It will instead point to the multixact
2583 * ID that will be assigned the next time one is needed.
2584 */
2585 if (oldestMultiXactId == nextMXact)
2586 {
2587 /*
2588 * When the next multixact gets created, it will be stored at the next
2589 * offset.
2590 */
2591 oldestOffset = nextOffset;
2592 oldestOffsetKnown = true;
2593 }
2594 else
2595 {
2596 /*
2597 * Figure out where the oldest existing multixact's offsets are
2598 * stored. Due to bugs in early release of PostgreSQL 9.3.X and 9.4.X,
2599 * the supposedly-earliest multixact might not really exist. We are
2600 * careful not to fail in that case.
2601 */
2602 oldestOffsetKnown =
2603 find_multixact_start(oldestMultiXactId, &oldestOffset);
2604
2605 if (oldestOffsetKnown)
2606 ereport(DEBUG1,
2607 (errmsg("oldest MultiXactId member is at offset %u",
2608 oldestOffset)));
2609 else
2610 ereport(LOG,
2611 (errmsg("MultiXact member wraparound protections are disabled because oldest checkpointed MultiXact %u does not exist on disk",
2612 oldestMultiXactId)));
2613 }
2614
2615 LWLockRelease(MultiXactTruncationLock);
2616
2617 /*
2618 * If we can, compute limits (and install them MultiXactState) to prevent
2619 * overrun of old data in the members SLRU area. We can only do so if the
2620 * oldest offset is known though.
2621 */
2622 if (oldestOffsetKnown)
2623 {
2624 /* move back to start of the corresponding segment */
2625 offsetStopLimit = oldestOffset - (oldestOffset %
2626 (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT));
2627
2628 /* always leave one segment before the wraparound point */
2629 offsetStopLimit -= (MULTIXACT_MEMBERS_PER_PAGE * SLRU_PAGES_PER_SEGMENT);
2630
2631 if (!prevOldestOffsetKnown && !is_startup)
2632 ereport(LOG,
2633 (errmsg("MultiXact member wraparound protections are now enabled")));
2634
2635 ereport(DEBUG1,
2636 (errmsg("MultiXact member stop limit is now %u based on MultiXact %u",
2637 offsetStopLimit, oldestMultiXactId)));
2638 }
2639 else if (prevOldestOffsetKnown)
2640 {
2641 /*
2642 * If we failed to get the oldest offset this time, but we have a
2643 * value from a previous pass through this function, use the old
2644 * values rather than automatically forcing an emergency autovacuum
2645 * cycle again.
2646 */
2647 oldestOffset = prevOldestOffset;
2648 oldestOffsetKnown = true;
2649 offsetStopLimit = prevOffsetStopLimit;
2650 }
2651
2652 /* Install the computed values */
2653 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
2654 MultiXactState->oldestOffset = oldestOffset;
2655 MultiXactState->oldestOffsetKnown = oldestOffsetKnown;
2656 MultiXactState->offsetStopLimit = offsetStopLimit;
2657 LWLockRelease(MultiXactGenLock);
2658
2659 /*
2660 * Do we need an emergency autovacuum? If we're not sure, assume yes.
2661 */
2662 return !oldestOffsetKnown ||
2663 (nextOffset - oldestOffset > MULTIXACT_MEMBER_SAFE_THRESHOLD);
2664 }
2665
2666 /*
2667 * Return whether adding "distance" to "start" would move past "boundary".
2668 *
2669 * We use this to determine whether the addition is "wrapping around" the
2670 * boundary point, hence the name. The reason we don't want to use the regular
2671 * 2^31-modulo arithmetic here is that we want to be able to use the whole of
2672 * the 2^32-1 space here, allowing for more multixacts that would fit
2673 * otherwise.
2674 */
2675 static bool
MultiXactOffsetWouldWrap(MultiXactOffset boundary,MultiXactOffset start,uint32 distance)2676 MultiXactOffsetWouldWrap(MultiXactOffset boundary, MultiXactOffset start,
2677 uint32 distance)
2678 {
2679 MultiXactOffset finish;
2680
2681 /*
2682 * Note that offset number 0 is not used (see GetMultiXactIdMembers), so
2683 * if the addition wraps around the UINT_MAX boundary, skip that value.
2684 */
2685 finish = start + distance;
2686 if (finish < start)
2687 finish++;
2688
2689 /*-----------------------------------------------------------------------
2690 * When the boundary is numerically greater than the starting point, any
2691 * value numerically between the two is not wrapped:
2692 *
2693 * <----S----B---->
2694 * [---) = F wrapped past B (and UINT_MAX)
2695 * [---) = F not wrapped
2696 * [----] = F wrapped past B
2697 *
2698 * When the boundary is numerically less than the starting point (i.e. the
2699 * UINT_MAX wraparound occurs somewhere in between) then all values in
2700 * between are wrapped:
2701 *
2702 * <----B----S---->
2703 * [---) = F not wrapped past B (but wrapped past UINT_MAX)
2704 * [---) = F wrapped past B (and UINT_MAX)
2705 * [----] = F not wrapped
2706 *-----------------------------------------------------------------------
2707 */
2708 if (start < boundary)
2709 return finish >= boundary || finish < start;
2710 else
2711 return finish >= boundary && finish < start;
2712 }
2713
2714 /*
2715 * Find the starting offset of the given MultiXactId.
2716 *
2717 * Returns false if the file containing the multi does not exist on disk.
2718 * Otherwise, returns true and sets *result to the starting member offset.
2719 *
2720 * This function does not prevent concurrent truncation, so if that's
2721 * required, the caller has to protect against that.
2722 */
2723 static bool
find_multixact_start(MultiXactId multi,MultiXactOffset * result)2724 find_multixact_start(MultiXactId multi, MultiXactOffset *result)
2725 {
2726 MultiXactOffset offset;
2727 int pageno;
2728 int entryno;
2729 int slotno;
2730 MultiXactOffset *offptr;
2731
2732 Assert(MultiXactState->finishedStartup);
2733
2734 pageno = MultiXactIdToOffsetPage(multi);
2735 entryno = MultiXactIdToOffsetEntry(multi);
2736
2737 /*
2738 * Flush out dirty data, so PhysicalPageExists can work correctly.
2739 * SimpleLruFlush() is a pretty big hammer for that. Alternatively we
2740 * could add an in-memory version of page exists, but find_multixact_start
2741 * is called infrequently, and it doesn't seem bad to flush buffers to
2742 * disk before truncation.
2743 */
2744 SimpleLruFlush(MultiXactOffsetCtl, true);
2745 SimpleLruFlush(MultiXactMemberCtl, true);
2746
2747 if (!SimpleLruDoesPhysicalPageExist(MultiXactOffsetCtl, pageno))
2748 return false;
2749
2750 /* lock is acquired by SimpleLruReadPage_ReadOnly */
2751 slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
2752 offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
2753 offptr += entryno;
2754 offset = *offptr;
2755 LWLockRelease(MultiXactOffsetControlLock);
2756
2757 *result = offset;
2758 return true;
2759 }
2760
2761 /*
2762 * Determine how many multixacts, and how many multixact members, currently
2763 * exist. Return false if unable to determine.
2764 */
2765 static bool
ReadMultiXactCounts(uint32 * multixacts,MultiXactOffset * members)2766 ReadMultiXactCounts(uint32 *multixacts, MultiXactOffset *members)
2767 {
2768 MultiXactOffset nextOffset;
2769 MultiXactOffset oldestOffset;
2770 MultiXactId oldestMultiXactId;
2771 MultiXactId nextMultiXactId;
2772 bool oldestOffsetKnown;
2773
2774 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2775 nextOffset = MultiXactState->nextOffset;
2776 oldestMultiXactId = MultiXactState->oldestMultiXactId;
2777 nextMultiXactId = MultiXactState->nextMXact;
2778 oldestOffset = MultiXactState->oldestOffset;
2779 oldestOffsetKnown = MultiXactState->oldestOffsetKnown;
2780 LWLockRelease(MultiXactGenLock);
2781
2782 if (!oldestOffsetKnown)
2783 return false;
2784
2785 *members = nextOffset - oldestOffset;
2786 *multixacts = nextMultiXactId - oldestMultiXactId;
2787 return true;
2788 }
2789
2790 /*
2791 * Multixact members can be removed once the multixacts that refer to them
2792 * are older than every datminxmid. autovacuum_multixact_freeze_max_age and
2793 * vacuum_multixact_freeze_table_age work together to make sure we never have
2794 * too many multixacts; we hope that, at least under normal circumstances,
2795 * this will also be sufficient to keep us from using too many offsets.
2796 * However, if the average multixact has many members, we might exhaust the
2797 * members space while still using few enough members that these limits fail
2798 * to trigger full table scans for relminmxid advancement. At that point,
2799 * we'd have no choice but to start failing multixact-creating operations
2800 * with an error.
2801 *
2802 * To prevent that, if more than a threshold portion of the members space is
2803 * used, we effectively reduce autovacuum_multixact_freeze_max_age and
2804 * to a value just less than the number of multixacts in use. We hope that
2805 * this will quickly trigger autovacuuming on the table or tables with the
2806 * oldest relminmxid, thus allowing datminmxid values to advance and removing
2807 * some members.
2808 *
2809 * As the fraction of the member space currently in use grows, we become
2810 * more aggressive in clamping this value. That not only causes autovacuum
2811 * to ramp up, but also makes any manual vacuums the user issues more
2812 * aggressive. This happens because vacuum_set_xid_limits() clamps the
2813 * freeze table and the minimum freeze age based on the effective
2814 * autovacuum_multixact_freeze_max_age this function returns. In the worst
2815 * case, we'll claim the freeze_max_age to zero, and every vacuum of any
2816 * table will try to freeze every multixact.
2817 *
2818 * It's possible that these thresholds should be user-tunable, but for now
2819 * we keep it simple.
2820 */
2821 int
MultiXactMemberFreezeThreshold(void)2822 MultiXactMemberFreezeThreshold(void)
2823 {
2824 MultiXactOffset members;
2825 uint32 multixacts;
2826 uint32 victim_multixacts;
2827 double fraction;
2828
2829 /* If we can't determine member space utilization, assume the worst. */
2830 if (!ReadMultiXactCounts(&multixacts, &members))
2831 return 0;
2832
2833 /* If member space utilization is low, no special action is required. */
2834 if (members <= MULTIXACT_MEMBER_SAFE_THRESHOLD)
2835 return autovacuum_multixact_freeze_max_age;
2836
2837 /*
2838 * Compute a target for relminmxid advancement. The number of multixacts
2839 * we try to eliminate from the system is based on how far we are past
2840 * MULTIXACT_MEMBER_SAFE_THRESHOLD.
2841 */
2842 fraction = (double) (members - MULTIXACT_MEMBER_SAFE_THRESHOLD) /
2843 (MULTIXACT_MEMBER_DANGER_THRESHOLD - MULTIXACT_MEMBER_SAFE_THRESHOLD);
2844 victim_multixacts = multixacts * fraction;
2845
2846 /* fraction could be > 1.0, but lowest possible freeze age is zero */
2847 if (victim_multixacts > multixacts)
2848 return 0;
2849 return multixacts - victim_multixacts;
2850 }
2851
2852 typedef struct mxtruncinfo
2853 {
2854 int earliestExistingPage;
2855 } mxtruncinfo;
2856
2857 /*
2858 * SlruScanDirectory callback
2859 * This callback determines the earliest existing page number.
2860 */
2861 static bool
SlruScanDirCbFindEarliest(SlruCtl ctl,char * filename,int segpage,void * data)2862 SlruScanDirCbFindEarliest(SlruCtl ctl, char *filename, int segpage, void *data)
2863 {
2864 mxtruncinfo *trunc = (mxtruncinfo *) data;
2865
2866 if (trunc->earliestExistingPage == -1 ||
2867 ctl->PagePrecedes(segpage, trunc->earliestExistingPage))
2868 {
2869 trunc->earliestExistingPage = segpage;
2870 }
2871
2872 return false; /* keep going */
2873 }
2874
2875
2876 /*
2877 * Delete members segments [oldest, newOldest)
2878 *
2879 * The members SLRU can, in contrast to the offsets one, be filled to almost
2880 * the full range at once. This means SimpleLruTruncate() can't trivially be
2881 * used - instead the to-be-deleted range is computed using the offsets
2882 * SLRU. C.f. TruncateMultiXact().
2883 */
2884 static void
PerformMembersTruncation(MultiXactOffset oldestOffset,MultiXactOffset newOldestOffset)2885 PerformMembersTruncation(MultiXactOffset oldestOffset, MultiXactOffset newOldestOffset)
2886 {
2887 const int maxsegment = MXOffsetToMemberSegment(MaxMultiXactOffset);
2888 int startsegment = MXOffsetToMemberSegment(oldestOffset);
2889 int endsegment = MXOffsetToMemberSegment(newOldestOffset);
2890 int segment = startsegment;
2891
2892 /*
2893 * Delete all the segments but the last one. The last segment can still
2894 * contain, possibly partially, valid data.
2895 */
2896 while (segment != endsegment)
2897 {
2898 elog(DEBUG2, "truncating multixact members segment %x", segment);
2899 SlruDeleteSegment(MultiXactMemberCtl, segment);
2900
2901 /* move to next segment, handling wraparound correctly */
2902 if (segment == maxsegment)
2903 segment = 0;
2904 else
2905 segment += 1;
2906 }
2907 }
2908
2909 /*
2910 * Delete offsets segments [oldest, newOldest)
2911 */
2912 static void
PerformOffsetsTruncation(MultiXactId oldestMulti,MultiXactId newOldestMulti)2913 PerformOffsetsTruncation(MultiXactId oldestMulti, MultiXactId newOldestMulti)
2914 {
2915 /*
2916 * We step back one multixact to avoid passing a cutoff page that hasn't
2917 * been created yet in the rare case that oldestMulti would be the first
2918 * item on a page and oldestMulti == nextMulti. In that case, if we
2919 * didn't subtract one, we'd trigger SimpleLruTruncate's wraparound
2920 * detection.
2921 */
2922 SimpleLruTruncate(MultiXactOffsetCtl,
2923 MultiXactIdToOffsetPage(PreviousMultiXactId(newOldestMulti)));
2924 }
2925
2926 /*
2927 * Remove all MultiXactOffset and MultiXactMember segments before the oldest
2928 * ones still of interest.
2929 *
2930 * This is only called on a primary as part of vacuum (via
2931 * vac_truncate_clog()). During recovery truncation is done by replaying
2932 * truncation WAL records logged here.
2933 *
2934 * newOldestMulti is the oldest currently required multixact, newOldestMultiDB
2935 * is one of the databases preventing newOldestMulti from increasing.
2936 */
2937 void
TruncateMultiXact(MultiXactId newOldestMulti,Oid newOldestMultiDB)2938 TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
2939 {
2940 MultiXactId oldestMulti;
2941 MultiXactId nextMulti;
2942 MultiXactOffset newOldestOffset;
2943 MultiXactOffset oldestOffset;
2944 MultiXactOffset nextOffset;
2945 mxtruncinfo trunc;
2946 MultiXactId earliest;
2947
2948 Assert(!RecoveryInProgress());
2949 Assert(MultiXactState->finishedStartup);
2950
2951 /*
2952 * We can only allow one truncation to happen at once. Otherwise parts of
2953 * members might vanish while we're doing lookups or similar. There's no
2954 * need to have an interlock with creating new multis or such, since those
2955 * are constrained by the limits (which only grow, never shrink).
2956 */
2957 LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
2958
2959 LWLockAcquire(MultiXactGenLock, LW_SHARED);
2960 nextMulti = MultiXactState->nextMXact;
2961 nextOffset = MultiXactState->nextOffset;
2962 oldestMulti = MultiXactState->oldestMultiXactId;
2963 LWLockRelease(MultiXactGenLock);
2964 Assert(MultiXactIdIsValid(oldestMulti));
2965
2966 /*
2967 * Make sure to only attempt truncation if there's values to truncate
2968 * away. In normal processing values shouldn't go backwards, but there's
2969 * some corner cases (due to bugs) where that's possible.
2970 */
2971 if (MultiXactIdPrecedesOrEquals(newOldestMulti, oldestMulti))
2972 {
2973 LWLockRelease(MultiXactTruncationLock);
2974 return;
2975 }
2976
2977 /*
2978 * Note we can't just plow ahead with the truncation; it's possible that
2979 * there are no segments to truncate, which is a problem because we are
2980 * going to attempt to read the offsets page to determine where to
2981 * truncate the members SLRU. So we first scan the directory to determine
2982 * the earliest offsets page number that we can read without error.
2983 *
2984 * When nextMXact is less than one segment away from multiWrapLimit,
2985 * SlruScanDirCbFindEarliest can find some early segment other than the
2986 * actual earliest. (MultiXactOffsetPagePrecedes(EARLIEST, LATEST)
2987 * returns false, because not all pairs of entries have the same answer.)
2988 * That can also arise when an earlier truncation attempt failed unlink()
2989 * or returned early from this function. The only consequence is
2990 * returning early, which wastes space that we could have liberated.
2991 *
2992 * NB: It's also possible that the page that oldestMulti is on has already
2993 * been truncated away, and we crashed before updating oldestMulti.
2994 */
2995 trunc.earliestExistingPage = -1;
2996 SlruScanDirectory(MultiXactOffsetCtl, SlruScanDirCbFindEarliest, &trunc);
2997 earliest = trunc.earliestExistingPage * MULTIXACT_OFFSETS_PER_PAGE;
2998 if (earliest < FirstMultiXactId)
2999 earliest = FirstMultiXactId;
3000
3001 /* If there's nothing to remove, we can bail out early. */
3002 if (MultiXactIdPrecedes(oldestMulti, earliest))
3003 {
3004 LWLockRelease(MultiXactTruncationLock);
3005 return;
3006 }
3007
3008 /*
3009 * First, compute the safe truncation point for MultiXactMember. This is
3010 * the starting offset of the oldest multixact.
3011 *
3012 * Hopefully, find_multixact_start will always work here, because we've
3013 * already checked that it doesn't precede the earliest MultiXact on disk.
3014 * But if it fails, don't truncate anything, and log a message.
3015 */
3016 if (oldestMulti == nextMulti)
3017 {
3018 /* there are NO MultiXacts */
3019 oldestOffset = nextOffset;
3020 }
3021 else if (!find_multixact_start(oldestMulti, &oldestOffset))
3022 {
3023 ereport(LOG,
3024 (errmsg("oldest MultiXact %u not found, earliest MultiXact %u, skipping truncation",
3025 oldestMulti, earliest)));
3026 LWLockRelease(MultiXactTruncationLock);
3027 return;
3028 }
3029
3030 /*
3031 * Secondly compute up to where to truncate. Lookup the corresponding
3032 * member offset for newOldestMulti for that.
3033 */
3034 if (newOldestMulti == nextMulti)
3035 {
3036 /* there are NO MultiXacts */
3037 newOldestOffset = nextOffset;
3038 }
3039 else if (!find_multixact_start(newOldestMulti, &newOldestOffset))
3040 {
3041 ereport(LOG,
3042 (errmsg("cannot truncate up to MultiXact %u because it does not exist on disk, skipping truncation",
3043 newOldestMulti)));
3044 LWLockRelease(MultiXactTruncationLock);
3045 return;
3046 }
3047
3048 elog(DEBUG1, "performing multixact truncation: "
3049 "offsets [%u, %u), offsets segments [%x, %x), "
3050 "members [%u, %u), members segments [%x, %x)",
3051 oldestMulti, newOldestMulti,
3052 MultiXactIdToOffsetSegment(oldestMulti),
3053 MultiXactIdToOffsetSegment(newOldestMulti),
3054 oldestOffset, newOldestOffset,
3055 MXOffsetToMemberSegment(oldestOffset),
3056 MXOffsetToMemberSegment(newOldestOffset));
3057
3058 /*
3059 * Do truncation, and the WAL logging of the truncation, in a critical
3060 * section. That way offsets/members cannot get out of sync anymore, i.e.
3061 * once consistent the newOldestMulti will always exist in members, even
3062 * if we crashed in the wrong moment.
3063 */
3064 START_CRIT_SECTION();
3065
3066 /*
3067 * Prevent checkpoints from being scheduled concurrently. This is critical
3068 * because otherwise a truncation record might not be replayed after a
3069 * crash/basebackup, even though the state of the data directory would
3070 * require it.
3071 */
3072 Assert(!MyPgXact->delayChkpt);
3073 MyPgXact->delayChkpt = true;
3074
3075 /* WAL log truncation */
3076 WriteMTruncateXlogRec(newOldestMultiDB,
3077 oldestMulti, newOldestMulti,
3078 oldestOffset, newOldestOffset);
3079
3080 /*
3081 * Update in-memory limits before performing the truncation, while inside
3082 * the critical section: Have to do it before truncation, to prevent
3083 * concurrent lookups of those values. Has to be inside the critical
3084 * section as otherwise a future call to this function would error out,
3085 * while looking up the oldest member in offsets, if our caller crashes
3086 * before updating the limits.
3087 */
3088 LWLockAcquire(MultiXactGenLock, LW_EXCLUSIVE);
3089 MultiXactState->oldestMultiXactId = newOldestMulti;
3090 MultiXactState->oldestMultiXactDB = newOldestMultiDB;
3091 LWLockRelease(MultiXactGenLock);
3092
3093 /* First truncate members */
3094 PerformMembersTruncation(oldestOffset, newOldestOffset);
3095
3096 /* Then offsets */
3097 PerformOffsetsTruncation(oldestMulti, newOldestMulti);
3098
3099 MyPgXact->delayChkpt = false;
3100
3101 END_CRIT_SECTION();
3102 LWLockRelease(MultiXactTruncationLock);
3103 }
3104
3105 /*
3106 * Decide whether a MultiXactOffset page number is "older" for truncation
3107 * purposes. Analogous to CLOGPagePrecedes().
3108 *
3109 * Offsetting the values is optional, because MultiXactIdPrecedes() has
3110 * translational symmetry.
3111 */
3112 static bool
MultiXactOffsetPagePrecedes(int page1,int page2)3113 MultiXactOffsetPagePrecedes(int page1, int page2)
3114 {
3115 MultiXactId multi1;
3116 MultiXactId multi2;
3117
3118 multi1 = ((MultiXactId) page1) * MULTIXACT_OFFSETS_PER_PAGE;
3119 multi1 += FirstMultiXactId + 1;
3120 multi2 = ((MultiXactId) page2) * MULTIXACT_OFFSETS_PER_PAGE;
3121 multi2 += FirstMultiXactId + 1;
3122
3123 return (MultiXactIdPrecedes(multi1, multi2) &&
3124 MultiXactIdPrecedes(multi1,
3125 multi2 + MULTIXACT_OFFSETS_PER_PAGE - 1));
3126 }
3127
3128 /*
3129 * Decide whether a MultiXactMember page number is "older" for truncation
3130 * purposes. There is no "invalid offset number" so use the numbers verbatim.
3131 */
3132 static bool
MultiXactMemberPagePrecedes(int page1,int page2)3133 MultiXactMemberPagePrecedes(int page1, int page2)
3134 {
3135 MultiXactOffset offset1;
3136 MultiXactOffset offset2;
3137
3138 offset1 = ((MultiXactOffset) page1) * MULTIXACT_MEMBERS_PER_PAGE;
3139 offset2 = ((MultiXactOffset) page2) * MULTIXACT_MEMBERS_PER_PAGE;
3140
3141 return (MultiXactOffsetPrecedes(offset1, offset2) &&
3142 MultiXactOffsetPrecedes(offset1,
3143 offset2 + MULTIXACT_MEMBERS_PER_PAGE - 1));
3144 }
3145
3146 /*
3147 * Decide which of two MultiXactIds is earlier.
3148 *
3149 * XXX do we need to do something special for InvalidMultiXactId?
3150 * (Doesn't look like it.)
3151 */
3152 bool
MultiXactIdPrecedes(MultiXactId multi1,MultiXactId multi2)3153 MultiXactIdPrecedes(MultiXactId multi1, MultiXactId multi2)
3154 {
3155 int32 diff = (int32) (multi1 - multi2);
3156
3157 return (diff < 0);
3158 }
3159
3160 /*
3161 * MultiXactIdPrecedesOrEquals -- is multi1 logically <= multi2?
3162 *
3163 * XXX do we need to do something special for InvalidMultiXactId?
3164 * (Doesn't look like it.)
3165 */
3166 bool
MultiXactIdPrecedesOrEquals(MultiXactId multi1,MultiXactId multi2)3167 MultiXactIdPrecedesOrEquals(MultiXactId multi1, MultiXactId multi2)
3168 {
3169 int32 diff = (int32) (multi1 - multi2);
3170
3171 return (diff <= 0);
3172 }
3173
3174
3175 /*
3176 * Decide which of two offsets is earlier.
3177 */
3178 static bool
MultiXactOffsetPrecedes(MultiXactOffset offset1,MultiXactOffset offset2)3179 MultiXactOffsetPrecedes(MultiXactOffset offset1, MultiXactOffset offset2)
3180 {
3181 int32 diff = (int32) (offset1 - offset2);
3182
3183 return (diff < 0);
3184 }
3185
3186 /*
3187 * Write an xlog record reflecting the zeroing of either a MEMBERs or
3188 * OFFSETs page (info shows which)
3189 */
3190 static void
WriteMZeroPageXlogRec(int pageno,uint8 info)3191 WriteMZeroPageXlogRec(int pageno, uint8 info)
3192 {
3193 XLogBeginInsert();
3194 XLogRegisterData((char *) (&pageno), sizeof(int));
3195 (void) XLogInsert(RM_MULTIXACT_ID, info);
3196 }
3197
3198 /*
3199 * Write a TRUNCATE xlog record
3200 *
3201 * We must flush the xlog record to disk before returning --- see notes in
3202 * TruncateCLOG().
3203 */
3204 static void
WriteMTruncateXlogRec(Oid oldestMultiDB,MultiXactId startTruncOff,MultiXactId endTruncOff,MultiXactOffset startTruncMemb,MultiXactOffset endTruncMemb)3205 WriteMTruncateXlogRec(Oid oldestMultiDB,
3206 MultiXactId startTruncOff, MultiXactId endTruncOff,
3207 MultiXactOffset startTruncMemb, MultiXactOffset endTruncMemb)
3208 {
3209 XLogRecPtr recptr;
3210 xl_multixact_truncate xlrec;
3211
3212 xlrec.oldestMultiDB = oldestMultiDB;
3213
3214 xlrec.startTruncOff = startTruncOff;
3215 xlrec.endTruncOff = endTruncOff;
3216
3217 xlrec.startTruncMemb = startTruncMemb;
3218 xlrec.endTruncMemb = endTruncMemb;
3219
3220 XLogBeginInsert();
3221 XLogRegisterData((char *) (&xlrec), SizeOfMultiXactTruncate);
3222 recptr = XLogInsert(RM_MULTIXACT_ID, XLOG_MULTIXACT_TRUNCATE_ID);
3223 XLogFlush(recptr);
3224 }
3225
3226 /*
3227 * MULTIXACT resource manager's routines
3228 */
3229 void
multixact_redo(XLogReaderState * record)3230 multixact_redo(XLogReaderState *record)
3231 {
3232 uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
3233
3234 /* Backup blocks are not used in multixact records */
3235 Assert(!XLogRecHasAnyBlockRefs(record));
3236
3237 if (info == XLOG_MULTIXACT_ZERO_OFF_PAGE)
3238 {
3239 int pageno;
3240 int slotno;
3241
3242 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
3243
3244 LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
3245
3246 slotno = ZeroMultiXactOffsetPage(pageno, false);
3247 SimpleLruWritePage(MultiXactOffsetCtl, slotno);
3248 Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
3249
3250 LWLockRelease(MultiXactOffsetControlLock);
3251 }
3252 else if (info == XLOG_MULTIXACT_ZERO_MEM_PAGE)
3253 {
3254 int pageno;
3255 int slotno;
3256
3257 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
3258
3259 LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
3260
3261 slotno = ZeroMultiXactMemberPage(pageno, false);
3262 SimpleLruWritePage(MultiXactMemberCtl, slotno);
3263 Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
3264
3265 LWLockRelease(MultiXactMemberControlLock);
3266 }
3267 else if (info == XLOG_MULTIXACT_CREATE_ID)
3268 {
3269 xl_multixact_create *xlrec =
3270 (xl_multixact_create *) XLogRecGetData(record);
3271 TransactionId max_xid;
3272 int i;
3273
3274 /* Store the data back into the SLRU files */
3275 RecordNewMultiXact(xlrec->mid, xlrec->moff, xlrec->nmembers,
3276 xlrec->members);
3277
3278 /* Make sure nextMXact/nextOffset are beyond what this record has */
3279 MultiXactAdvanceNextMXact(xlrec->mid + 1,
3280 xlrec->moff + xlrec->nmembers);
3281
3282 /*
3283 * Make sure nextXid is beyond any XID mentioned in the record. This
3284 * should be unnecessary, since any XID found here ought to have other
3285 * evidence in the XLOG, but let's be safe.
3286 */
3287 max_xid = XLogRecGetXid(record);
3288 for (i = 0; i < xlrec->nmembers; i++)
3289 {
3290 if (TransactionIdPrecedes(max_xid, xlrec->members[i].xid))
3291 max_xid = xlrec->members[i].xid;
3292 }
3293
3294 /*
3295 * We don't expect anyone else to modify nextXid, hence startup
3296 * process doesn't need to hold a lock while checking this. We still
3297 * acquire the lock to modify it, though.
3298 */
3299 if (TransactionIdFollowsOrEquals(max_xid,
3300 ShmemVariableCache->nextXid))
3301 {
3302 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
3303 ShmemVariableCache->nextXid = max_xid;
3304 TransactionIdAdvance(ShmemVariableCache->nextXid);
3305 LWLockRelease(XidGenLock);
3306 }
3307 }
3308 else if (info == XLOG_MULTIXACT_TRUNCATE_ID)
3309 {
3310 xl_multixact_truncate xlrec;
3311 int pageno;
3312
3313 memcpy(&xlrec, XLogRecGetData(record),
3314 SizeOfMultiXactTruncate);
3315
3316 elog(DEBUG1, "replaying multixact truncation: "
3317 "offsets [%u, %u), offsets segments [%x, %x), "
3318 "members [%u, %u), members segments [%x, %x)",
3319 xlrec.startTruncOff, xlrec.endTruncOff,
3320 MultiXactIdToOffsetSegment(xlrec.startTruncOff),
3321 MultiXactIdToOffsetSegment(xlrec.endTruncOff),
3322 xlrec.startTruncMemb, xlrec.endTruncMemb,
3323 MXOffsetToMemberSegment(xlrec.startTruncMemb),
3324 MXOffsetToMemberSegment(xlrec.endTruncMemb));
3325
3326 /* should not be required, but more than cheap enough */
3327 LWLockAcquire(MultiXactTruncationLock, LW_EXCLUSIVE);
3328
3329 /*
3330 * Advance the horizon values, so they're current at the end of
3331 * recovery.
3332 */
3333 SetMultiXactIdLimit(xlrec.endTruncOff, xlrec.oldestMultiDB, false);
3334
3335 PerformMembersTruncation(xlrec.startTruncMemb, xlrec.endTruncMemb);
3336
3337 /*
3338 * During XLOG replay, latest_page_number isn't necessarily set up
3339 * yet; insert a suitable value to bypass the sanity test in
3340 * SimpleLruTruncate.
3341 */
3342 pageno = MultiXactIdToOffsetPage(xlrec.endTruncOff);
3343 MultiXactOffsetCtl->shared->latest_page_number = pageno;
3344 PerformOffsetsTruncation(xlrec.startTruncOff, xlrec.endTruncOff);
3345
3346 LWLockRelease(MultiXactTruncationLock);
3347 }
3348 else
3349 elog(PANIC, "multixact_redo: unknown op code %u", info);
3350 }
3351
3352 Datum
pg_get_multixact_members(PG_FUNCTION_ARGS)3353 pg_get_multixact_members(PG_FUNCTION_ARGS)
3354 {
3355 typedef struct
3356 {
3357 MultiXactMember *members;
3358 int nmembers;
3359 int iter;
3360 } mxact;
3361 MultiXactId mxid = PG_GETARG_UINT32(0);
3362 mxact *multi;
3363 FuncCallContext *funccxt;
3364
3365 if (mxid < FirstMultiXactId)
3366 ereport(ERROR,
3367 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3368 errmsg("invalid MultiXactId: %u", mxid)));
3369
3370 if (SRF_IS_FIRSTCALL())
3371 {
3372 MemoryContext oldcxt;
3373 TupleDesc tupdesc;
3374
3375 funccxt = SRF_FIRSTCALL_INIT();
3376 oldcxt = MemoryContextSwitchTo(funccxt->multi_call_memory_ctx);
3377
3378 multi = palloc(sizeof(mxact));
3379 /* no need to allow for old values here */
3380 multi->nmembers = GetMultiXactIdMembers(mxid, &multi->members, false,
3381 false);
3382 multi->iter = 0;
3383
3384 tupdesc = CreateTemplateTupleDesc(2, false);
3385 TupleDescInitEntry(tupdesc, (AttrNumber) 1, "xid",
3386 XIDOID, -1, 0);
3387 TupleDescInitEntry(tupdesc, (AttrNumber) 2, "mode",
3388 TEXTOID, -1, 0);
3389
3390 funccxt->attinmeta = TupleDescGetAttInMetadata(tupdesc);
3391 funccxt->user_fctx = multi;
3392
3393 MemoryContextSwitchTo(oldcxt);
3394 }
3395
3396 funccxt = SRF_PERCALL_SETUP();
3397 multi = (mxact *) funccxt->user_fctx;
3398
3399 while (multi->iter < multi->nmembers)
3400 {
3401 HeapTuple tuple;
3402 char *values[2];
3403
3404 values[0] = psprintf("%u", multi->members[multi->iter].xid);
3405 values[1] = mxstatus_to_string(multi->members[multi->iter].status);
3406
3407 tuple = BuildTupleFromCStrings(funccxt->attinmeta, values);
3408
3409 multi->iter++;
3410 pfree(values[0]);
3411 SRF_RETURN_NEXT(funccxt, HeapTupleGetDatum(tuple));
3412 }
3413
3414 if (multi->nmembers > 0)
3415 pfree(multi->members);
3416 pfree(multi);
3417
3418 SRF_RETURN_DONE(funccxt);
3419 }
3420