1/*-------------------------------------------------------------------------
2 *
3 * clog.c
4 * PostgreSQL transaction-commit-log manager
5 *
6 * This module replaces the old "pg_log" access code, which treated pg_log
7 * essentially like a relation, in that it went through the regular buffer
8 * manager. The problem with that was that there wasn't any good way to
9 * recycle storage space for transactions so old that they'll never be
10 * looked up again. Now we use specialized access code so that the commit
11 * log can be broken into relatively small, independent segments.
12 *
13 * XLOG interactions: this module generates an XLOG record whenever a new
14 * CLOG page is initialized to zeroes. Other writes of CLOG come from
15 * recording of transaction commit or abort in xact.c, which generates its
16 * own XLOG records for these events and will re-perform the status update
17 * on redo; so we need make no additional XLOG entry here. For synchronous
18 * transaction commits, the XLOG is guaranteed flushed through the XLOG commit
19 * record before we are called to log a commit, so the WAL rule "write xlog
20 * before data" is satisfied automatically. However, for async commits we
21 * must track the latest LSN affecting each CLOG page, so that we can flush
22 * XLOG that far and satisfy the WAL rule. We don't have to worry about this
23 * for aborts (whether sync or async), since the post-crash assumption would
24 * be that such transactions failed anyway.
25 *
26 * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
27 * Portions Copyright (c) 1994, Regents of the University of California
28 *
29 * src/backend/access/transam/clog.c
30 *
31 *-------------------------------------------------------------------------
32 */
33#include "postgres.h"
34
35#include "access/clog.h"
36#include "access/slru.h"
37#include "access/transam.h"
38#include "access/xlog.h"
39#include "access/xloginsert.h"
40#include "access/xlogutils.h"
41#include "miscadmin.h"
42#include "pgstat.h"
43#include "pg_trace.h"
44#include "storage/proc.h"
45
46/*
47 * Defines for CLOG page sizes. A page is the same BLCKSZ as is used
48 * everywhere else in Postgres.
49 *
50 * Note: because TransactionIds are 32 bits and wrap around at 0xFFFFFFFF,
51 * CLOG page numbering also wraps around at 0xFFFFFFFF/CLOG_XACTS_PER_PAGE,
52 * and CLOG segment numbering at
53 * 0xFFFFFFFF/CLOG_XACTS_PER_PAGE/SLRU_PAGES_PER_SEGMENT. We need take no
54 * explicit notice of that fact in this module, except when comparing segment
55 * and page numbers in TruncateCLOG (see CLOGPagePrecedes).
56 */
57
58/* We need two bits per xact, so four xacts fit in a byte */
59#define CLOG_BITS_PER_XACT 2
60#define CLOG_XACTS_PER_BYTE 4
61#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
62#define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
63
64#define TransactionIdToPage(xid) ((xid) / (TransactionId) CLOG_XACTS_PER_PAGE)
65#define TransactionIdToPgIndex(xid) ((xid) % (TransactionId) CLOG_XACTS_PER_PAGE)
66#define TransactionIdToByte(xid) (TransactionIdToPgIndex(xid) / CLOG_XACTS_PER_BYTE)
67#define TransactionIdToBIndex(xid) ((xid) % (TransactionId) CLOG_XACTS_PER_BYTE)
68
69/* We store the latest async LSN for each group of transactions */
70#define CLOG_XACTS_PER_LSN_GROUP 32 /* keep this a power of 2 */
71#define CLOG_LSNS_PER_PAGE (CLOG_XACTS_PER_PAGE / CLOG_XACTS_PER_LSN_GROUP)
72
73#define GetLSNIndex(slotno, xid) ((slotno) * CLOG_LSNS_PER_PAGE + \
74 ((xid) % (TransactionId) CLOG_XACTS_PER_PAGE) / CLOG_XACTS_PER_LSN_GROUP)
75
76/*
77 * The number of subtransactions below which we consider to apply clog group
78 * update optimization. Testing reveals that the number higher than this can
79 * hurt performance.
80 */
81#define THRESHOLD_SUBTRANS_CLOG_OPT 5
82
83/*
84 * Link to shared-memory data structures for CLOG control
85 */
86static SlruCtlData ClogCtlData;
87
88#define ClogCtl (&ClogCtlData)
89
90
91static int ZeroCLOGPage(int pageno, bool writeXlog);
92static bool CLOGPagePrecedes(int page1, int page2);
93static void WriteZeroPageXlogRec(int pageno);
94static void WriteTruncateXlogRec(int pageno, TransactionId oldestXact,
95 Oid oldestXidDb);
96static void TransactionIdSetPageStatus(TransactionId xid, int nsubxids,
97 TransactionId *subxids, XidStatus status,
98 XLogRecPtr lsn, int pageno,
99 bool all_xact_same_page);
100static void TransactionIdSetStatusBit(TransactionId xid, XidStatus status,
101 XLogRecPtr lsn, int slotno);
102static void set_status_by_pages(int nsubxids, TransactionId *subxids,
103 XidStatus status, XLogRecPtr lsn);
104static bool TransactionGroupUpdateXidStatus(TransactionId xid,
105 XidStatus status, XLogRecPtr lsn, int pageno);
106static void TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids,
107 TransactionId *subxids, XidStatus status,
108 XLogRecPtr lsn, int pageno);
109
110
111/*
112 * TransactionIdSetTreeStatus
113 *
114 * Record the final state of transaction entries in the commit log for
115 * a transaction and its subtransaction tree. Take care to ensure this is
116 * efficient, and as atomic as possible.
117 *
118 * xid is a single xid to set status for. This will typically be
119 * the top level transactionid for a top level commit or abort. It can
120 * also be a subtransaction when we record transaction aborts.
121 *
122 * subxids is an array of xids of length nsubxids, representing subtransactions
123 * in the tree of xid. In various cases nsubxids may be zero.
124 *
125 * lsn must be the WAL location of the commit record when recording an async
126 * commit. For a synchronous commit it can be InvalidXLogRecPtr, since the
127 * caller guarantees the commit record is already flushed in that case. It
128 * should be InvalidXLogRecPtr for abort cases, too.
129 *
130 * In the commit case, atomicity is limited by whether all the subxids are in
131 * the same CLOG page as xid. If they all are, then the lock will be grabbed
132 * only once, and the status will be set to committed directly. Otherwise
133 * we must
134 * 1. set sub-committed all subxids that are not on the same page as the
135 * main xid
136 * 2. atomically set committed the main xid and the subxids on the same page
137 * 3. go over the first bunch again and set them committed
138 * Note that as far as concurrent checkers are concerned, main transaction
139 * commit as a whole is still atomic.
140 *
141 * Example:
142 * TransactionId t commits and has subxids t1, t2, t3, t4
143 * t is on page p1, t1 is also on p1, t2 and t3 are on p2, t4 is on p3
144 * 1. update pages2-3:
145 * page2: set t2,t3 as sub-committed
146 * page3: set t4 as sub-committed
147 * 2. update page1:
148 * set t1 as sub-committed,
149 * then set t as committed,
150 then set t1 as committed
151 * 3. update pages2-3:
152 * page2: set t2,t3 as committed
153 * page3: set t4 as committed
154 *
155 * NB: this is a low-level routine and is NOT the preferred entry point
156 * for most uses; functions in transam.c are the intended callers.
157 *
158 * XXX Think about issuing FADVISE_WILLNEED on pages that we will need,
159 * but aren't yet in cache, as well as hinting pages not to fall out of
160 * cache yet.
161 */
162void
163TransactionIdSetTreeStatus(TransactionId xid, int nsubxids,
164 TransactionId *subxids, XidStatus status, XLogRecPtr lsn)
165{
166 int pageno = TransactionIdToPage(xid); /* get page of parent */
167 int i;
168
169 Assert(status == TRANSACTION_STATUS_COMMITTED ||
170 status == TRANSACTION_STATUS_ABORTED);
171
172 /*
173 * See how many subxids, if any, are on the same page as the parent, if
174 * any.
175 */
176 for (i = 0; i < nsubxids; i++)
177 {
178 if (TransactionIdToPage(subxids[i]) != pageno)
179 break;
180 }
181
182 /*
183 * Do all items fit on a single page?
184 */
185 if (i == nsubxids)
186 {
187 /*
188 * Set the parent and all subtransactions in a single call
189 */
190 TransactionIdSetPageStatus(xid, nsubxids, subxids, status, lsn,
191 pageno, true);
192 }
193 else
194 {
195 int nsubxids_on_first_page = i;
196
197 /*
198 * If this is a commit then we care about doing this correctly (i.e.
199 * using the subcommitted intermediate status). By here, we know
200 * we're updating more than one page of clog, so we must mark entries
201 * that are *not* on the first page so that they show as subcommitted
202 * before we then return to update the status to fully committed.
203 *
204 * To avoid touching the first page twice, skip marking subcommitted
205 * for the subxids on that first page.
206 */
207 if (status == TRANSACTION_STATUS_COMMITTED)
208 set_status_by_pages(nsubxids - nsubxids_on_first_page,
209 subxids + nsubxids_on_first_page,
210 TRANSACTION_STATUS_SUB_COMMITTED, lsn);
211
212 /*
213 * Now set the parent and subtransactions on same page as the parent,
214 * if any
215 */
216 pageno = TransactionIdToPage(xid);
217 TransactionIdSetPageStatus(xid, nsubxids_on_first_page, subxids, status,
218 lsn, pageno, false);
219
220 /*
221 * Now work through the rest of the subxids one clog page at a time,
222 * starting from the second page onwards, like we did above.
223 */
224 set_status_by_pages(nsubxids - nsubxids_on_first_page,
225 subxids + nsubxids_on_first_page,
226 status, lsn);
227 }
228}
229
230/*
231 * Helper for TransactionIdSetTreeStatus: set the status for a bunch of
232 * transactions, chunking in the separate CLOG pages involved. We never
233 * pass the whole transaction tree to this function, only subtransactions
234 * that are on different pages to the top level transaction id.
235 */
236static void
237set_status_by_pages(int nsubxids, TransactionId *subxids,
238 XidStatus status, XLogRecPtr lsn)
239{
240 int pageno = TransactionIdToPage(subxids[0]);
241 int offset = 0;
242 int i = 0;
243
244 Assert(nsubxids > 0); /* else the pageno fetch above is unsafe */
245
246 while (i < nsubxids)
247 {
248 int num_on_page = 0;
249 int nextpageno;
250
251 do
252 {
253 nextpageno = TransactionIdToPage(subxids[i]);
254 if (nextpageno != pageno)
255 break;
256 num_on_page++;
257 i++;
258 } while (i < nsubxids);
259
260 TransactionIdSetPageStatus(InvalidTransactionId,
261 num_on_page, subxids + offset,
262 status, lsn, pageno, false);
263 offset = i;
264 pageno = nextpageno;
265 }
266}
267
268/*
269 * Record the final state of transaction entries in the commit log for all
270 * entries on a single page. Atomic only on this page.
271 */
272static void
273TransactionIdSetPageStatus(TransactionId xid, int nsubxids,
274 TransactionId *subxids, XidStatus status,
275 XLogRecPtr lsn, int pageno,
276 bool all_xact_same_page)
277{
278 /* Can't use group update when PGPROC overflows. */
279 StaticAssertStmt(THRESHOLD_SUBTRANS_CLOG_OPT <= PGPROC_MAX_CACHED_SUBXIDS,
280 "group clog threshold less than PGPROC cached subxids");
281
282 /*
283 * When there is contention on CLogControlLock, we try to group multiple
284 * updates; a single leader process will perform transaction status
285 * updates for multiple backends so that the number of times
286 * CLogControlLock needs to be acquired is reduced.
287 *
288 * For this optimization to be safe, the XID in MyPgXact and the subxids
289 * in MyProc must be the same as the ones for which we're setting the
290 * status. Check that this is the case.
291 *
292 * For this optimization to be efficient, we shouldn't have too many
293 * sub-XIDs and all of the XIDs for which we're adjusting clog should be
294 * on the same page. Check those conditions, too.
295 */
296 if (all_xact_same_page && xid == MyPgXact->xid &&
297 nsubxids <= THRESHOLD_SUBTRANS_CLOG_OPT &&
298 nsubxids == MyPgXact->nxids &&
299 memcmp(subxids, MyProc->subxids.xids,
300 nsubxids * sizeof(TransactionId)) == 0)
301 {
302 /*
303 * We don't try to do group update optimization if a process has
304 * overflowed the subxids array in its PGPROC, since in that case we
305 * don't have a complete list of XIDs for it.
306 */
307 Assert(THRESHOLD_SUBTRANS_CLOG_OPT <= PGPROC_MAX_CACHED_SUBXIDS);
308
309 /*
310 * If we can immediately acquire CLogControlLock, we update the status
311 * of our own XID and release the lock. If not, try use group XID
312 * update. If that doesn't work out, fall back to waiting for the
313 * lock to perform an update for this transaction only.
314 */
315 if (LWLockConditionalAcquire(CLogControlLock, LW_EXCLUSIVE))
316 {
317 /* Got the lock without waiting! Do the update. */
318 TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
319 lsn, pageno);
320 LWLockRelease(CLogControlLock);
321 return;
322 }
323 else if (TransactionGroupUpdateXidStatus(xid, status, lsn, pageno))
324 {
325 /* Group update mechanism has done the work. */
326 return;
327 }
328
329 /* Fall through only if update isn't done yet. */
330 }
331
332 /* Group update not applicable, or couldn't accept this page number. */
333 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
334 TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
335 lsn, pageno);
336 LWLockRelease(CLogControlLock);
337}
338
339/*
340 * Record the final state of transaction entry in the commit log
341 *
342 * We don't do any locking here; caller must handle that.
343 */
344static void
345TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids,
346 TransactionId *subxids, XidStatus status,
347 XLogRecPtr lsn, int pageno)
348{
349 int slotno;
350 int i;
351
352 Assert(status == TRANSACTION_STATUS_COMMITTED ||
353 status == TRANSACTION_STATUS_ABORTED ||
354 (status == TRANSACTION_STATUS_SUB_COMMITTED && !TransactionIdIsValid(xid)));
355 Assert(LWLockHeldByMeInMode(CLogControlLock, LW_EXCLUSIVE));
356
357 /*
358 * If we're doing an async commit (ie, lsn is valid), then we must wait
359 * for any active write on the page slot to complete. Otherwise our
360 * update could reach disk in that write, which will not do since we
361 * mustn't let it reach disk until we've done the appropriate WAL flush.
362 * But when lsn is invalid, it's OK to scribble on a page while it is
363 * write-busy, since we don't care if the update reaches disk sooner than
364 * we think.
365 */
366 slotno = SimpleLruReadPage(ClogCtl, pageno, XLogRecPtrIsInvalid(lsn), xid);
367
368 /*
369 * Set the main transaction id, if any.
370 *
371 * If we update more than one xid on this page while it is being written
372 * out, we might find that some of the bits go to disk and others don't.
373 * If we are updating commits on the page with the top-level xid that
374 * could break atomicity, so we subcommit the subxids first before we mark
375 * the top-level commit.
376 */
377 if (TransactionIdIsValid(xid))
378 {
379 /* Subtransactions first, if needed ... */
380 if (status == TRANSACTION_STATUS_COMMITTED)
381 {
382 for (i = 0; i < nsubxids; i++)
383 {
384 Assert(ClogCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
385 TransactionIdSetStatusBit(subxids[i],
386 TRANSACTION_STATUS_SUB_COMMITTED,
387 lsn, slotno);
388 }
389 }
390
391 /* ... then the main transaction */
392 TransactionIdSetStatusBit(xid, status, lsn, slotno);
393 }
394
395 /* Set the subtransactions */
396 for (i = 0; i < nsubxids; i++)
397 {
398 Assert(ClogCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
399 TransactionIdSetStatusBit(subxids[i], status, lsn, slotno);
400 }
401
402 ClogCtl->shared->page_dirty[slotno] = true;
403}
404
405/*
406 * When we cannot immediately acquire CLogControlLock in exclusive mode at
407 * commit time, add ourselves to a list of processes that need their XIDs
408 * status update. The first process to add itself to the list will acquire
409 * CLogControlLock in exclusive mode and set transaction status as required
410 * on behalf of all group members. This avoids a great deal of contention
411 * around CLogControlLock when many processes are trying to commit at once,
412 * since the lock need not be repeatedly handed off from one committing
413 * process to the next.
414 *
415 * Returns true when transaction status has been updated in clog; returns
416 * false if we decided against applying the optimization because the page
417 * number we need to update differs from those processes already waiting.
418 */
419static bool
420TransactionGroupUpdateXidStatus(TransactionId xid, XidStatus status,
421 XLogRecPtr lsn, int pageno)
422{
423 volatile PROC_HDR *procglobal = ProcGlobal;
424 PGPROC *proc = MyProc;
425 uint32 nextidx;
426 uint32 wakeidx;
427
428 /* We should definitely have an XID whose status needs to be updated. */
429 Assert(TransactionIdIsValid(xid));
430
431 /*
432 * Add ourselves to the list of processes needing a group XID status
433 * update.
434 */
435 proc->clogGroupMember = true;
436 proc->clogGroupMemberXid = xid;
437 proc->clogGroupMemberXidStatus = status;
438 proc->clogGroupMemberPage = pageno;
439 proc->clogGroupMemberLsn = lsn;
440
441 nextidx = pg_atomic_read_u32(&procglobal->clogGroupFirst);
442
443 while (true)
444 {
445 /*
446 * Add the proc to list, if the clog page where we need to update the
447 * current transaction status is same as group leader's clog page.
448 *
449 * There is a race condition here, which is that after doing the below
450 * check and before adding this proc's clog update to a group, the
451 * group leader might have already finished the group update for this
452 * page and becomes group leader of another group. This will lead to a
453 * situation where a single group can have different clog page
454 * updates. This isn't likely and will still work, just maybe a bit
455 * less efficiently.
456 */
457 if (nextidx != INVALID_PGPROCNO &&
458 ProcGlobal->allProcs[nextidx].clogGroupMemberPage != proc->clogGroupMemberPage)
459 {
460 proc->clogGroupMember = false;
461 return false;
462 }
463
464 pg_atomic_write_u32(&proc->clogGroupNext, nextidx);
465
466 if (pg_atomic_compare_exchange_u32(&procglobal->clogGroupFirst,
467 &nextidx,
468 (uint32) proc->pgprocno))
469 break;
470 }
471
472 /*
473 * If the list was not empty, the leader will update the status of our
474 * XID. It is impossible to have followers without a leader because the
475 * first process that has added itself to the list will always have
476 * nextidx as INVALID_PGPROCNO.
477 */
478 if (nextidx != INVALID_PGPROCNO)
479 {
480 int extraWaits = 0;
481
482 /* Sleep until the leader updates our XID status. */
483 pgstat_report_wait_start(WAIT_EVENT_CLOG_GROUP_UPDATE);
484 for (;;)
485 {
486 /* acts as a read barrier */
487 PGSemaphoreLock(proc->sem);
488 if (!proc->clogGroupMember)
489 break;
490 extraWaits++;
491 }
492 pgstat_report_wait_end();
493
494 Assert(pg_atomic_read_u32(&proc->clogGroupNext) == INVALID_PGPROCNO);
495
496 /* Fix semaphore count for any absorbed wakeups */
497 while (extraWaits-- > 0)
498 PGSemaphoreUnlock(proc->sem);
499 return true;
500 }
501
502 /* We are the leader. Acquire the lock on behalf of everyone. */
503 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
504
505 /*
506 * Now that we've got the lock, clear the list of processes waiting for
507 * group XID status update, saving a pointer to the head of the list.
508 * Trying to pop elements one at a time could lead to an ABA problem.
509 */
510 nextidx = pg_atomic_exchange_u32(&procglobal->clogGroupFirst,
511 INVALID_PGPROCNO);
512
513 /* Remember head of list so we can perform wakeups after dropping lock. */
514 wakeidx = nextidx;
515
516 /* Walk the list and update the status of all XIDs. */
517 while (nextidx != INVALID_PGPROCNO)
518 {
519 PGPROC *proc = &ProcGlobal->allProcs[nextidx];
520 PGXACT *pgxact = &ProcGlobal->allPgXact[nextidx];
521
522 /*
523 * Overflowed transactions should not use group XID status update
524 * mechanism.
525 */
526 Assert(!pgxact->overflowed);
527
528 TransactionIdSetPageStatusInternal(proc->clogGroupMemberXid,
529 pgxact->nxids,
530 proc->subxids.xids,
531 proc->clogGroupMemberXidStatus,
532 proc->clogGroupMemberLsn,
533 proc->clogGroupMemberPage);
534
535 /* Move to next proc in list. */
536 nextidx = pg_atomic_read_u32(&proc->clogGroupNext);
537 }
538
539 /* We're done with the lock now. */
540 LWLockRelease(CLogControlLock);
541
542 /*
543 * Now that we've released the lock, go back and wake everybody up. We
544 * don't do this under the lock so as to keep lock hold times to a
545 * minimum.
546 */
547 while (wakeidx != INVALID_PGPROCNO)
548 {
549 PGPROC *proc = &ProcGlobal->allProcs[wakeidx];
550
551 wakeidx = pg_atomic_read_u32(&proc->clogGroupNext);
552 pg_atomic_write_u32(&proc->clogGroupNext, INVALID_PGPROCNO);
553
554 /* ensure all previous writes are visible before follower continues. */
555 pg_write_barrier();
556
557 proc->clogGroupMember = false;
558
559 if (proc != MyProc)
560 PGSemaphoreUnlock(proc->sem);
561 }
562
563 return true;
564}
565
566/*
567 * Sets the commit status of a single transaction.
568 *
569 * Must be called with CLogControlLock held
570 */
571static void
572TransactionIdSetStatusBit(TransactionId xid, XidStatus status, XLogRecPtr lsn, int slotno)
573{
574 int byteno = TransactionIdToByte(xid);
575 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
576 char *byteptr;
577 char byteval;
578 char curval;
579
580 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
581 curval = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
582
583 /*
584 * When replaying transactions during recovery we still need to perform
585 * the two phases of subcommit and then commit. However, some transactions
586 * are already correctly marked, so we just treat those as a no-op which
587 * allows us to keep the following Assert as restrictive as possible.
588 */
589 if (InRecovery && status == TRANSACTION_STATUS_SUB_COMMITTED &&
590 curval == TRANSACTION_STATUS_COMMITTED)
591 return;
592
593 /*
594 * Current state change should be from 0 or subcommitted to target state
595 * or we should already be there when replaying changes during recovery.
596 */
597 Assert(curval == 0 ||
598 (curval == TRANSACTION_STATUS_SUB_COMMITTED &&
599 status != TRANSACTION_STATUS_IN_PROGRESS) ||
600 curval == status);
601
602 /* note this assumes exclusive access to the clog page */
603 byteval = *byteptr;
604 byteval &= ~(((1 << CLOG_BITS_PER_XACT) - 1) << bshift);
605 byteval |= (status << bshift);
606 *byteptr = byteval;
607
608 /*
609 * Update the group LSN if the transaction completion LSN is higher.
610 *
611 * Note: lsn will be invalid when supplied during InRecovery processing,
612 * so we don't need to do anything special to avoid LSN updates during
613 * recovery. After recovery completes the next clog change will set the
614 * LSN correctly.
615 */
616 if (!XLogRecPtrIsInvalid(lsn))
617 {
618 int lsnindex = GetLSNIndex(slotno, xid);
619
620 if (ClogCtl->shared->group_lsn[lsnindex] < lsn)
621 ClogCtl->shared->group_lsn[lsnindex] = lsn;
622 }
623}
624
625/*
626 * Interrogate the state of a transaction in the commit log.
627 *
628 * Aside from the actual commit status, this function returns (into *lsn)
629 * an LSN that is late enough to be able to guarantee that if we flush up to
630 * that LSN then we will have flushed the transaction's commit record to disk.
631 * The result is not necessarily the exact LSN of the transaction's commit
632 * record! For example, for long-past transactions (those whose clog pages
633 * already migrated to disk), we'll return InvalidXLogRecPtr. Also, because
634 * we group transactions on the same clog page to conserve storage, we might
635 * return the LSN of a later transaction that falls into the same group.
636 *
637 * NB: this is a low-level routine and is NOT the preferred entry point
638 * for most uses; TransactionLogFetch() in transam.c is the intended caller.
639 */
640XidStatus
641TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
642{
643 int pageno = TransactionIdToPage(xid);
644 int byteno = TransactionIdToByte(xid);
645 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
646 int slotno;
647 int lsnindex;
648 char *byteptr;
649 XidStatus status;
650
651 /* lock is acquired by SimpleLruReadPage_ReadOnly */
652
653 slotno = SimpleLruReadPage_ReadOnly(ClogCtl, pageno, xid);
654 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
655
656 status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
657
658 lsnindex = GetLSNIndex(slotno, xid);
659 *lsn = ClogCtl->shared->group_lsn[lsnindex];
660
661 LWLockRelease(CLogControlLock);
662
663 return status;
664}
665
666/*
667 * Number of shared CLOG buffers.
668 *
669 * On larger multi-processor systems, it is possible to have many CLOG page
670 * requests in flight at one time which could lead to disk access for CLOG
671 * page if the required page is not found in memory. Testing revealed that we
672 * can get the best performance by having 128 CLOG buffers, more than that it
673 * doesn't improve performance.
674 *
675 * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
676 * a good idea, because it would increase the minimum amount of shared memory
677 * required to start, which could be a problem for people running very small
678 * configurations. The following formula seems to represent a reasonable
679 * compromise: people with very low values for shared_buffers will get fewer
680 * CLOG buffers as well, and everyone else will get 128.
681 */
682Size
683CLOGShmemBuffers(void)
684{
685 return Min(128, Max(4, NBuffers / 512));
686}
687
688/*
689 * Initialization of shared memory for CLOG
690 */
691Size
692CLOGShmemSize(void)
693{
694 return SimpleLruShmemSize(CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE);
695}
696
697void
698CLOGShmemInit(void)
699{
700 ClogCtl->PagePrecedes = CLOGPagePrecedes;
701 SimpleLruInit(ClogCtl, "clog", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
702 CLogControlLock, "pg_xact", LWTRANCHE_CLOG_BUFFERS);
703}
704
705/*
706 * This func must be called ONCE on system install. It creates
707 * the initial CLOG segment. (The CLOG directory is assumed to
708 * have been created by initdb, and CLOGShmemInit must have been
709 * called already.)
710 */
711void
712BootStrapCLOG(void)
713{
714 int slotno;
715
716 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
717
718 /* Create and zero the first page of the commit log */
719 slotno = ZeroCLOGPage(0, false);
720
721 /* Make sure it's written out */
722 SimpleLruWritePage(ClogCtl, slotno);
723 Assert(!ClogCtl->shared->page_dirty[slotno]);
724
725 LWLockRelease(CLogControlLock);
726}
727
728/*
729 * Initialize (or reinitialize) a page of CLOG to zeroes.
730 * If writeXlog is true, also emit an XLOG record saying we did this.
731 *
732 * The page is not actually written, just set up in shared memory.
733 * The slot number of the new page is returned.
734 *
735 * Control lock must be held at entry, and will be held at exit.
736 */
737static int
738ZeroCLOGPage(int pageno, bool writeXlog)
739{
740 int slotno;
741
742 slotno = SimpleLruZeroPage(ClogCtl, pageno);
743
744 if (writeXlog)
745 WriteZeroPageXlogRec(pageno);
746
747 return slotno;
748}
749
750/*
751 * This must be called ONCE during postmaster or standalone-backend startup,
752 * after StartupXLOG has initialized ShmemVariableCache->nextFullXid.
753 */
754void
755StartupCLOG(void)
756{
757 TransactionId xid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
758 int pageno = TransactionIdToPage(xid);
759
760 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
761
762 /*
763 * Initialize our idea of the latest page number.
764 */
765 ClogCtl->shared->latest_page_number = pageno;
766
767 LWLockRelease(CLogControlLock);
768}
769
770/*
771 * This must be called ONCE at the end of startup/recovery.
772 */
773void
774TrimCLOG(void)
775{
776 TransactionId xid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
777 int pageno = TransactionIdToPage(xid);
778
779 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
780
781 /*
782 * Re-Initialize our idea of the latest page number.
783 */
784 ClogCtl->shared->latest_page_number = pageno;
785
786 /*
787 * Zero out the remainder of the current clog page. Under normal
788 * circumstances it should be zeroes already, but it seems at least
789 * theoretically possible that XLOG replay will have settled on a nextXID
790 * value that is less than the last XID actually used and marked by the
791 * previous database lifecycle (since subtransaction commit writes clog
792 * but makes no WAL entry). Let's just be safe. (We need not worry about
793 * pages beyond the current one, since those will be zeroed when first
794 * used. For the same reason, there is no need to do anything when
795 * nextFullXid is exactly at a page boundary; and it's likely that the
796 * "current" page doesn't exist yet in that case.)
797 */
798 if (TransactionIdToPgIndex(xid) != 0)
799 {
800 int byteno = TransactionIdToByte(xid);
801 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
802 int slotno;
803 char *byteptr;
804
805 slotno = SimpleLruReadPage(ClogCtl, pageno, false, xid);
806 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
807
808 /* Zero so-far-unused positions in the current byte */
809 *byteptr &= (1 << bshift) - 1;
810 /* Zero the rest of the page */
811 MemSet(byteptr + 1, 0, BLCKSZ - byteno - 1);
812
813 ClogCtl->shared->page_dirty[slotno] = true;
814 }
815
816 LWLockRelease(CLogControlLock);
817}
818
819/*
820 * This must be called ONCE during postmaster or standalone-backend shutdown
821 */
822void
823ShutdownCLOG(void)
824{
825 /* Flush dirty CLOG pages to disk */
826 TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(false);
827 SimpleLruFlush(ClogCtl, false);
828
829 /*
830 * fsync pg_xact to ensure that any files flushed previously are durably
831 * on disk.
832 */
833 fsync_fname("pg_xact", true);
834
835 TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(false);
836}
837
838/*
839 * Perform a checkpoint --- either during shutdown, or on-the-fly
840 */
841void
842CheckPointCLOG(void)
843{
844 /* Flush dirty CLOG pages to disk */
845 TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(true);
846 SimpleLruFlush(ClogCtl, true);
847
848 /*
849 * fsync pg_xact to ensure that any files flushed previously are durably
850 * on disk.
851 */
852 fsync_fname("pg_xact", true);
853
854 TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(true);
855}
856
857
858/*
859 * Make sure that CLOG has room for a newly-allocated XID.
860 *
861 * NB: this is called while holding XidGenLock. We want it to be very fast
862 * most of the time; even when it's not so fast, no actual I/O need happen
863 * unless we're forced to write out a dirty clog or xlog page to make room
864 * in shared memory.
865 */
866void
867ExtendCLOG(TransactionId newestXact)
868{
869 int pageno;
870
871 /*
872 * No work except at first XID of a page. But beware: just after
873 * wraparound, the first XID of page zero is FirstNormalTransactionId.
874 */
875 if (TransactionIdToPgIndex(newestXact) != 0 &&
876 !TransactionIdEquals(newestXact, FirstNormalTransactionId))
877 return;
878
879 pageno = TransactionIdToPage(newestXact);
880
881 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
882
883 /* Zero the page and make an XLOG entry about it */
884 ZeroCLOGPage(pageno, true);
885
886 LWLockRelease(CLogControlLock);
887}
888
889
890/*
891 * Remove all CLOG segments before the one holding the passed transaction ID
892 *
893 * Before removing any CLOG data, we must flush XLOG to disk, to ensure
894 * that any recently-emitted HEAP_FREEZE records have reached disk; otherwise
895 * a crash and restart might leave us with some unfrozen tuples referencing
896 * removed CLOG data. We choose to emit a special TRUNCATE XLOG record too.
897 * Replaying the deletion from XLOG is not critical, since the files could
898 * just as well be removed later, but doing so prevents a long-running hot
899 * standby server from acquiring an unreasonably bloated CLOG directory.
900 *
901 * Since CLOG segments hold a large number of transactions, the opportunity to
902 * actually remove a segment is fairly rare, and so it seems best not to do
903 * the XLOG flush unless we have confirmed that there is a removable segment.
904 */
905void
906TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
907{
908 int cutoffPage;
909
910 /*
911 * The cutoff point is the start of the segment containing oldestXact. We
912 * pass the *page* containing oldestXact to SimpleLruTruncate.
913 */
914 cutoffPage = TransactionIdToPage(oldestXact);
915
916 /* Check to see if there's any files that could be removed */
917 if (!SlruScanDirectory(ClogCtl, SlruScanDirCbReportPresence, &cutoffPage))
918 return; /* nothing to remove */
919
920 /*
921 * Advance oldestClogXid before truncating clog, so concurrent xact status
922 * lookups can ensure they don't attempt to access truncated-away clog.
923 *
924 * It's only necessary to do this if we will actually truncate away clog
925 * pages.
926 */
927 AdvanceOldestClogXid(oldestXact);
928
929 /*
930 * Write XLOG record and flush XLOG to disk. We record the oldest xid
931 * we're keeping information about here so we can ensure that it's always
932 * ahead of clog truncation in case we crash, and so a standby finds out
933 * the new valid xid before the next checkpoint.
934 */
935 WriteTruncateXlogRec(cutoffPage, oldestXact, oldestxid_datoid);
936
937 /* Now we can remove the old CLOG segment(s) */
938 SimpleLruTruncate(ClogCtl, cutoffPage);
939}
940
941
942/*
943 * Decide which of two CLOG page numbers is "older" for truncation purposes.
944 *
945 * We need to use comparison of TransactionIds here in order to do the right
946 * thing with wraparound XID arithmetic. However, if we are asked about
947 * page number zero, we don't want to hand InvalidTransactionId to
948 * TransactionIdPrecedes: it'll get weird about permanent xact IDs. So,
949 * offset both xids by FirstNormalTransactionId to avoid that.
950 */
951static bool
952CLOGPagePrecedes(int page1, int page2)
953{
954 TransactionId xid1;
955 TransactionId xid2;
956
957 xid1 = ((TransactionId) page1) * CLOG_XACTS_PER_PAGE;
958 xid1 += FirstNormalTransactionId;
959 xid2 = ((TransactionId) page2) * CLOG_XACTS_PER_PAGE;
960 xid2 += FirstNormalTransactionId;
961
962 return TransactionIdPrecedes(xid1, xid2);
963}
964
965
966/*
967 * Write a ZEROPAGE xlog record
968 */
969static void
970WriteZeroPageXlogRec(int pageno)
971{
972 XLogBeginInsert();
973 XLogRegisterData((char *) (&pageno), sizeof(int));
974 (void) XLogInsert(RM_CLOG_ID, CLOG_ZEROPAGE);
975}
976
977/*
978 * Write a TRUNCATE xlog record
979 *
980 * We must flush the xlog record to disk before returning --- see notes
981 * in TruncateCLOG().
982 */
983static void
984WriteTruncateXlogRec(int pageno, TransactionId oldestXact, Oid oldestXactDb)
985{
986 XLogRecPtr recptr;
987 xl_clog_truncate xlrec;
988
989 xlrec.pageno = pageno;
990 xlrec.oldestXact = oldestXact;
991 xlrec.oldestXactDb = oldestXactDb;
992
993 XLogBeginInsert();
994 XLogRegisterData((char *) (&xlrec), sizeof(xl_clog_truncate));
995 recptr = XLogInsert(RM_CLOG_ID, CLOG_TRUNCATE);
996 XLogFlush(recptr);
997}
998
999/*
1000 * CLOG resource manager's routines
1001 */
1002void
1003clog_redo(XLogReaderState *record)
1004{
1005 uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
1006
1007 /* Backup blocks are not used in clog records */
1008 Assert(!XLogRecHasAnyBlockRefs(record));
1009
1010 if (info == CLOG_ZEROPAGE)
1011 {
1012 int pageno;
1013 int slotno;
1014
1015 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
1016
1017 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
1018
1019 slotno = ZeroCLOGPage(pageno, false);
1020 SimpleLruWritePage(ClogCtl, slotno);
1021 Assert(!ClogCtl->shared->page_dirty[slotno]);
1022
1023 LWLockRelease(CLogControlLock);
1024 }
1025 else if (info == CLOG_TRUNCATE)
1026 {
1027 xl_clog_truncate xlrec;
1028
1029 memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_clog_truncate));
1030
1031 /*
1032 * During XLOG replay, latest_page_number isn't set up yet; insert a
1033 * suitable value to bypass the sanity test in SimpleLruTruncate.
1034 */
1035 ClogCtl->shared->latest_page_number = xlrec.pageno;
1036
1037 AdvanceOldestClogXid(xlrec.oldestXact);
1038
1039 SimpleLruTruncate(ClogCtl, xlrec.pageno);
1040 }
1041 else
1042 elog(PANIC, "clog_redo: unknown op code %u", info);
1043}
1044