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) 1996-2018, 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 */
86 static SlruCtlData ClogCtlData;
87
88 #define ClogCtl (&ClogCtlData)
89
90
91 static int ZeroCLOGPage(int pageno, bool writeXlog);
92 static bool CLOGPagePrecedes(int page1, int page2);
93 static void WriteZeroPageXlogRec(int pageno);
94 static void WriteTruncateXlogRec(int pageno, TransactionId oldestXact,
95 Oid oldestXidDb);
96 static void TransactionIdSetPageStatus(TransactionId xid, int nsubxids,
97 TransactionId *subxids, XidStatus status,
98 XLogRecPtr lsn, int pageno,
99 bool all_xact_same_page);
100 static void TransactionIdSetStatusBit(TransactionId xid, XidStatus status,
101 XLogRecPtr lsn, int slotno);
102 static void set_status_by_pages(int nsubxids, TransactionId *subxids,
103 XidStatus status, XLogRecPtr lsn);
104 static bool TransactionGroupUpdateXidStatus(TransactionId xid,
105 XidStatus status, XLogRecPtr lsn, int pageno);
106 static 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 */
162 void
TransactionIdSetTreeStatus(TransactionId xid,int nsubxids,TransactionId * subxids,XidStatus status,XLogRecPtr lsn)163 TransactionIdSetTreeStatus(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 */
236 static void
set_status_by_pages(int nsubxids,TransactionId * subxids,XidStatus status,XLogRecPtr lsn)237 set_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 */
272 static void
TransactionIdSetPageStatus(TransactionId xid,int nsubxids,TransactionId * subxids,XidStatus status,XLogRecPtr lsn,int pageno,bool all_xact_same_page)273 TransactionIdSetPageStatus(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 * If we can immediately acquire CLogControlLock, we update the status
304 * of our own XID and release the lock. If not, try use group XID
305 * update. If that doesn't work out, fall back to waiting for the
306 * lock to perform an update for this transaction only.
307 */
308 if (LWLockConditionalAcquire(CLogControlLock, LW_EXCLUSIVE))
309 {
310 /* Got the lock without waiting! Do the update. */
311 TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
312 lsn, pageno);
313 LWLockRelease(CLogControlLock);
314 return;
315 }
316 else if (TransactionGroupUpdateXidStatus(xid, status, lsn, pageno))
317 {
318 /* Group update mechanism has done the work. */
319 return;
320 }
321
322 /* Fall through only if update isn't done yet. */
323 }
324
325 /* Group update not applicable, or couldn't accept this page number. */
326 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
327 TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
328 lsn, pageno);
329 LWLockRelease(CLogControlLock);
330 }
331
332 /*
333 * Record the final state of transaction entry in the commit log
334 *
335 * We don't do any locking here; caller must handle that.
336 */
337 static void
TransactionIdSetPageStatusInternal(TransactionId xid,int nsubxids,TransactionId * subxids,XidStatus status,XLogRecPtr lsn,int pageno)338 TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids,
339 TransactionId *subxids, XidStatus status,
340 XLogRecPtr lsn, int pageno)
341 {
342 int slotno;
343 int i;
344
345 Assert(status == TRANSACTION_STATUS_COMMITTED ||
346 status == TRANSACTION_STATUS_ABORTED ||
347 (status == TRANSACTION_STATUS_SUB_COMMITTED && !TransactionIdIsValid(xid)));
348 Assert(LWLockHeldByMeInMode(CLogControlLock, LW_EXCLUSIVE));
349
350 /*
351 * If we're doing an async commit (ie, lsn is valid), then we must wait
352 * for any active write on the page slot to complete. Otherwise our
353 * update could reach disk in that write, which will not do since we
354 * mustn't let it reach disk until we've done the appropriate WAL flush.
355 * But when lsn is invalid, it's OK to scribble on a page while it is
356 * write-busy, since we don't care if the update reaches disk sooner than
357 * we think.
358 */
359 slotno = SimpleLruReadPage(ClogCtl, pageno, XLogRecPtrIsInvalid(lsn), xid);
360
361 /*
362 * Set the main transaction id, if any.
363 *
364 * If we update more than one xid on this page while it is being written
365 * out, we might find that some of the bits go to disk and others don't.
366 * If we are updating commits on the page with the top-level xid that
367 * could break atomicity, so we subcommit the subxids first before we mark
368 * the top-level commit.
369 */
370 if (TransactionIdIsValid(xid))
371 {
372 /* Subtransactions first, if needed ... */
373 if (status == TRANSACTION_STATUS_COMMITTED)
374 {
375 for (i = 0; i < nsubxids; i++)
376 {
377 Assert(ClogCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
378 TransactionIdSetStatusBit(subxids[i],
379 TRANSACTION_STATUS_SUB_COMMITTED,
380 lsn, slotno);
381 }
382 }
383
384 /* ... then the main transaction */
385 TransactionIdSetStatusBit(xid, status, lsn, slotno);
386 }
387
388 /* Set the subtransactions */
389 for (i = 0; i < nsubxids; i++)
390 {
391 Assert(ClogCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
392 TransactionIdSetStatusBit(subxids[i], status, lsn, slotno);
393 }
394
395 ClogCtl->shared->page_dirty[slotno] = true;
396 }
397
398 /*
399 * When we cannot immediately acquire CLogControlLock in exclusive mode at
400 * commit time, add ourselves to a list of processes that need their XIDs
401 * status update. The first process to add itself to the list will acquire
402 * CLogControlLock in exclusive mode and set transaction status as required
403 * on behalf of all group members. This avoids a great deal of contention
404 * around CLogControlLock when many processes are trying to commit at once,
405 * since the lock need not be repeatedly handed off from one committing
406 * process to the next.
407 *
408 * Returns true when transaction status has been updated in clog; returns
409 * false if we decided against applying the optimization because the page
410 * number we need to update differs from those processes already waiting.
411 */
412 static bool
TransactionGroupUpdateXidStatus(TransactionId xid,XidStatus status,XLogRecPtr lsn,int pageno)413 TransactionGroupUpdateXidStatus(TransactionId xid, XidStatus status,
414 XLogRecPtr lsn, int pageno)
415 {
416 volatile PROC_HDR *procglobal = ProcGlobal;
417 PGPROC *proc = MyProc;
418 uint32 nextidx;
419 uint32 wakeidx;
420
421 /* We should definitely have an XID whose status needs to be updated. */
422 Assert(TransactionIdIsValid(xid));
423
424 /*
425 * Add ourselves to the list of processes needing a group XID status
426 * update.
427 */
428 proc->clogGroupMember = true;
429 proc->clogGroupMemberXid = xid;
430 proc->clogGroupMemberXidStatus = status;
431 proc->clogGroupMemberPage = pageno;
432 proc->clogGroupMemberLsn = lsn;
433
434 nextidx = pg_atomic_read_u32(&procglobal->clogGroupFirst);
435
436 while (true)
437 {
438 /*
439 * Add the proc to list, if the clog page where we need to update the
440 * current transaction status is same as group leader's clog page.
441 *
442 * There is a race condition here, which is that after doing the below
443 * check and before adding this proc's clog update to a group, the
444 * group leader might have already finished the group update for this
445 * page and becomes group leader of another group. This will lead to a
446 * situation where a single group can have different clog page
447 * updates. This isn't likely and will still work, just maybe a bit
448 * less efficiently.
449 */
450 if (nextidx != INVALID_PGPROCNO &&
451 ProcGlobal->allProcs[nextidx].clogGroupMemberPage != proc->clogGroupMemberPage)
452 {
453 /*
454 * Ensure that this proc is not a member of any clog group that
455 * needs an XID status update.
456 */
457 proc->clogGroupMember = false;
458 pg_atomic_write_u32(&proc->clogGroupNext, INVALID_PGPROCNO);
459 return false;
460 }
461
462 pg_atomic_write_u32(&proc->clogGroupNext, nextidx);
463
464 if (pg_atomic_compare_exchange_u32(&procglobal->clogGroupFirst,
465 &nextidx,
466 (uint32) proc->pgprocno))
467 break;
468 }
469
470 /*
471 * If the list was not empty, the leader will update the status of our
472 * XID. It is impossible to have followers without a leader because the
473 * first process that has added itself to the list will always have
474 * nextidx as INVALID_PGPROCNO.
475 */
476 if (nextidx != INVALID_PGPROCNO)
477 {
478 int extraWaits = 0;
479
480 /* Sleep until the leader updates our XID status. */
481 pgstat_report_wait_start(WAIT_EVENT_CLOG_GROUP_UPDATE);
482 for (;;)
483 {
484 /* acts as a read barrier */
485 PGSemaphoreLock(proc->sem);
486 if (!proc->clogGroupMember)
487 break;
488 extraWaits++;
489 }
490 pgstat_report_wait_end();
491
492 Assert(pg_atomic_read_u32(&proc->clogGroupNext) == INVALID_PGPROCNO);
493
494 /* Fix semaphore count for any absorbed wakeups */
495 while (extraWaits-- > 0)
496 PGSemaphoreUnlock(proc->sem);
497 return true;
498 }
499
500 /* We are the leader. Acquire the lock on behalf of everyone. */
501 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
502
503 /*
504 * Now that we've got the lock, clear the list of processes waiting for
505 * group XID status update, saving a pointer to the head of the list.
506 * Trying to pop elements one at a time could lead to an ABA problem.
507 */
508 nextidx = pg_atomic_exchange_u32(&procglobal->clogGroupFirst,
509 INVALID_PGPROCNO);
510
511 /* Remember head of list so we can perform wakeups after dropping lock. */
512 wakeidx = nextidx;
513
514 /* Walk the list and update the status of all XIDs. */
515 while (nextidx != INVALID_PGPROCNO)
516 {
517 PGPROC *proc = &ProcGlobal->allProcs[nextidx];
518 PGXACT *pgxact = &ProcGlobal->allPgXact[nextidx];
519
520 /*
521 * Transactions with more than THRESHOLD_SUBTRANS_CLOG_OPT sub-XIDs
522 * should not use group XID status update mechanism.
523 */
524 Assert(pgxact->nxids <= THRESHOLD_SUBTRANS_CLOG_OPT);
525
526 TransactionIdSetPageStatusInternal(proc->clogGroupMemberXid,
527 pgxact->nxids,
528 proc->subxids.xids,
529 proc->clogGroupMemberXidStatus,
530 proc->clogGroupMemberLsn,
531 proc->clogGroupMemberPage);
532
533 /* Move to next proc in list. */
534 nextidx = pg_atomic_read_u32(&proc->clogGroupNext);
535 }
536
537 /* We're done with the lock now. */
538 LWLockRelease(CLogControlLock);
539
540 /*
541 * Now that we've released the lock, go back and wake everybody up. We
542 * don't do this under the lock so as to keep lock hold times to a
543 * minimum.
544 */
545 while (wakeidx != INVALID_PGPROCNO)
546 {
547 PGPROC *proc = &ProcGlobal->allProcs[wakeidx];
548
549 wakeidx = pg_atomic_read_u32(&proc->clogGroupNext);
550 pg_atomic_write_u32(&proc->clogGroupNext, INVALID_PGPROCNO);
551
552 /* ensure all previous writes are visible before follower continues. */
553 pg_write_barrier();
554
555 proc->clogGroupMember = false;
556
557 if (proc != MyProc)
558 PGSemaphoreUnlock(proc->sem);
559 }
560
561 return true;
562 }
563
564 /*
565 * Sets the commit status of a single transaction.
566 *
567 * Must be called with CLogControlLock held
568 */
569 static void
TransactionIdSetStatusBit(TransactionId xid,XidStatus status,XLogRecPtr lsn,int slotno)570 TransactionIdSetStatusBit(TransactionId xid, XidStatus status, XLogRecPtr lsn, int slotno)
571 {
572 int byteno = TransactionIdToByte(xid);
573 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
574 char *byteptr;
575 char byteval;
576 char curval;
577
578 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
579 curval = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
580
581 /*
582 * When replaying transactions during recovery we still need to perform
583 * the two phases of subcommit and then commit. However, some transactions
584 * are already correctly marked, so we just treat those as a no-op which
585 * allows us to keep the following Assert as restrictive as possible.
586 */
587 if (InRecovery && status == TRANSACTION_STATUS_SUB_COMMITTED &&
588 curval == TRANSACTION_STATUS_COMMITTED)
589 return;
590
591 /*
592 * Current state change should be from 0 or subcommitted to target state
593 * or we should already be there when replaying changes during recovery.
594 */
595 Assert(curval == 0 ||
596 (curval == TRANSACTION_STATUS_SUB_COMMITTED &&
597 status != TRANSACTION_STATUS_IN_PROGRESS) ||
598 curval == status);
599
600 /* note this assumes exclusive access to the clog page */
601 byteval = *byteptr;
602 byteval &= ~(((1 << CLOG_BITS_PER_XACT) - 1) << bshift);
603 byteval |= (status << bshift);
604 *byteptr = byteval;
605
606 /*
607 * Update the group LSN if the transaction completion LSN is higher.
608 *
609 * Note: lsn will be invalid when supplied during InRecovery processing,
610 * so we don't need to do anything special to avoid LSN updates during
611 * recovery. After recovery completes the next clog change will set the
612 * LSN correctly.
613 */
614 if (!XLogRecPtrIsInvalid(lsn))
615 {
616 int lsnindex = GetLSNIndex(slotno, xid);
617
618 if (ClogCtl->shared->group_lsn[lsnindex] < lsn)
619 ClogCtl->shared->group_lsn[lsnindex] = lsn;
620 }
621 }
622
623 /*
624 * Interrogate the state of a transaction in the commit log.
625 *
626 * Aside from the actual commit status, this function returns (into *lsn)
627 * an LSN that is late enough to be able to guarantee that if we flush up to
628 * that LSN then we will have flushed the transaction's commit record to disk.
629 * The result is not necessarily the exact LSN of the transaction's commit
630 * record! For example, for long-past transactions (those whose clog pages
631 * already migrated to disk), we'll return InvalidXLogRecPtr. Also, because
632 * we group transactions on the same clog page to conserve storage, we might
633 * return the LSN of a later transaction that falls into the same group.
634 *
635 * NB: this is a low-level routine and is NOT the preferred entry point
636 * for most uses; TransactionLogFetch() in transam.c is the intended caller.
637 */
638 XidStatus
TransactionIdGetStatus(TransactionId xid,XLogRecPtr * lsn)639 TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
640 {
641 int pageno = TransactionIdToPage(xid);
642 int byteno = TransactionIdToByte(xid);
643 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
644 int slotno;
645 int lsnindex;
646 char *byteptr;
647 XidStatus status;
648
649 /* lock is acquired by SimpleLruReadPage_ReadOnly */
650
651 slotno = SimpleLruReadPage_ReadOnly(ClogCtl, pageno, xid);
652 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
653
654 status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
655
656 lsnindex = GetLSNIndex(slotno, xid);
657 *lsn = ClogCtl->shared->group_lsn[lsnindex];
658
659 LWLockRelease(CLogControlLock);
660
661 return status;
662 }
663
664 /*
665 * Number of shared CLOG buffers.
666 *
667 * On larger multi-processor systems, it is possible to have many CLOG page
668 * requests in flight at one time which could lead to disk access for CLOG
669 * page if the required page is not found in memory. Testing revealed that we
670 * can get the best performance by having 128 CLOG buffers, more than that it
671 * doesn't improve performance.
672 *
673 * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
674 * a good idea, because it would increase the minimum amount of shared memory
675 * required to start, which could be a problem for people running very small
676 * configurations. The following formula seems to represent a reasonable
677 * compromise: people with very low values for shared_buffers will get fewer
678 * CLOG buffers as well, and everyone else will get 128.
679 */
680 Size
CLOGShmemBuffers(void)681 CLOGShmemBuffers(void)
682 {
683 return Min(128, Max(4, NBuffers / 512));
684 }
685
686 /*
687 * Initialization of shared memory for CLOG
688 */
689 Size
CLOGShmemSize(void)690 CLOGShmemSize(void)
691 {
692 return SimpleLruShmemSize(CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE);
693 }
694
695 void
CLOGShmemInit(void)696 CLOGShmemInit(void)
697 {
698 ClogCtl->PagePrecedes = CLOGPagePrecedes;
699 SimpleLruInit(ClogCtl, "clog", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
700 CLogControlLock, "pg_xact", LWTRANCHE_CLOG_BUFFERS);
701 SlruPagePrecedesUnitTests(ClogCtl, CLOG_XACTS_PER_PAGE);
702 }
703
704 /*
705 * This func must be called ONCE on system install. It creates
706 * the initial CLOG segment. (The CLOG directory is assumed to
707 * have been created by initdb, and CLOGShmemInit must have been
708 * called already.)
709 */
710 void
BootStrapCLOG(void)711 BootStrapCLOG(void)
712 {
713 int slotno;
714
715 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
716
717 /* Create and zero the first page of the commit log */
718 slotno = ZeroCLOGPage(0, false);
719
720 /* Make sure it's written out */
721 SimpleLruWritePage(ClogCtl, slotno);
722 Assert(!ClogCtl->shared->page_dirty[slotno]);
723
724 LWLockRelease(CLogControlLock);
725 }
726
727 /*
728 * Initialize (or reinitialize) a page of CLOG to zeroes.
729 * If writeXlog is true, also emit an XLOG record saying we did this.
730 *
731 * The page is not actually written, just set up in shared memory.
732 * The slot number of the new page is returned.
733 *
734 * Control lock must be held at entry, and will be held at exit.
735 */
736 static int
ZeroCLOGPage(int pageno,bool writeXlog)737 ZeroCLOGPage(int pageno, bool writeXlog)
738 {
739 int slotno;
740
741 slotno = SimpleLruZeroPage(ClogCtl, pageno);
742
743 if (writeXlog)
744 WriteZeroPageXlogRec(pageno);
745
746 return slotno;
747 }
748
749 /*
750 * This must be called ONCE during postmaster or standalone-backend startup,
751 * after StartupXLOG has initialized ShmemVariableCache->nextXid.
752 */
753 void
StartupCLOG(void)754 StartupCLOG(void)
755 {
756 TransactionId xid = ShmemVariableCache->nextXid;
757 int pageno = TransactionIdToPage(xid);
758
759 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
760
761 /*
762 * Initialize our idea of the latest page number.
763 */
764 ClogCtl->shared->latest_page_number = pageno;
765
766 LWLockRelease(CLogControlLock);
767 }
768
769 /*
770 * This must be called ONCE at the end of startup/recovery.
771 */
772 void
TrimCLOG(void)773 TrimCLOG(void)
774 {
775 TransactionId xid = ShmemVariableCache->nextXid;
776 int pageno = TransactionIdToPage(xid);
777
778 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
779
780 /*
781 * Re-Initialize our idea of the latest page number.
782 */
783 ClogCtl->shared->latest_page_number = pageno;
784
785 /*
786 * Zero out the remainder of the current clog page. Under normal
787 * circumstances it should be zeroes already, but it seems at least
788 * theoretically possible that XLOG replay will have settled on a nextXID
789 * value that is less than the last XID actually used and marked by the
790 * previous database lifecycle (since subtransaction commit writes clog
791 * but makes no WAL entry). Let's just be safe. (We need not worry about
792 * pages beyond the current one, since those will be zeroed when first
793 * used. For the same reason, there is no need to do anything when
794 * nextXid is exactly at a page boundary; and it's likely that the
795 * "current" page doesn't exist yet in that case.)
796 */
797 if (TransactionIdToPgIndex(xid) != 0)
798 {
799 int byteno = TransactionIdToByte(xid);
800 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
801 int slotno;
802 char *byteptr;
803
804 slotno = SimpleLruReadPage(ClogCtl, pageno, false, xid);
805 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
806
807 /* Zero so-far-unused positions in the current byte */
808 *byteptr &= (1 << bshift) - 1;
809 /* Zero the rest of the page */
810 MemSet(byteptr + 1, 0, BLCKSZ - byteno - 1);
811
812 ClogCtl->shared->page_dirty[slotno] = true;
813 }
814
815 LWLockRelease(CLogControlLock);
816 }
817
818 /*
819 * This must be called ONCE during postmaster or standalone-backend shutdown
820 */
821 void
ShutdownCLOG(void)822 ShutdownCLOG(void)
823 {
824 /* Flush dirty CLOG pages to disk */
825 TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(false);
826 SimpleLruFlush(ClogCtl, false);
827
828 /*
829 * fsync pg_xact to ensure that any files flushed previously are durably
830 * on disk.
831 */
832 fsync_fname("pg_xact", true);
833
834 TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(false);
835 }
836
837 /*
838 * Perform a checkpoint --- either during shutdown, or on-the-fly
839 */
840 void
CheckPointCLOG(void)841 CheckPointCLOG(void)
842 {
843 /* Flush dirty CLOG pages to disk */
844 TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(true);
845 SimpleLruFlush(ClogCtl, true);
846 TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(true);
847 }
848
849
850 /*
851 * Make sure that CLOG has room for a newly-allocated XID.
852 *
853 * NB: this is called while holding XidGenLock. We want it to be very fast
854 * most of the time; even when it's not so fast, no actual I/O need happen
855 * unless we're forced to write out a dirty clog or xlog page to make room
856 * in shared memory.
857 */
858 void
ExtendCLOG(TransactionId newestXact)859 ExtendCLOG(TransactionId newestXact)
860 {
861 int pageno;
862
863 /*
864 * No work except at first XID of a page. But beware: just after
865 * wraparound, the first XID of page zero is FirstNormalTransactionId.
866 */
867 if (TransactionIdToPgIndex(newestXact) != 0 &&
868 !TransactionIdEquals(newestXact, FirstNormalTransactionId))
869 return;
870
871 pageno = TransactionIdToPage(newestXact);
872
873 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
874
875 /* Zero the page and make an XLOG entry about it */
876 ZeroCLOGPage(pageno, true);
877
878 LWLockRelease(CLogControlLock);
879 }
880
881
882 /*
883 * Remove all CLOG segments before the one holding the passed transaction ID
884 *
885 * Before removing any CLOG data, we must flush XLOG to disk, to ensure
886 * that any recently-emitted HEAP_FREEZE records have reached disk; otherwise
887 * a crash and restart might leave us with some unfrozen tuples referencing
888 * removed CLOG data. We choose to emit a special TRUNCATE XLOG record too.
889 * Replaying the deletion from XLOG is not critical, since the files could
890 * just as well be removed later, but doing so prevents a long-running hot
891 * standby server from acquiring an unreasonably bloated CLOG directory.
892 *
893 * Since CLOG segments hold a large number of transactions, the opportunity to
894 * actually remove a segment is fairly rare, and so it seems best not to do
895 * the XLOG flush unless we have confirmed that there is a removable segment.
896 */
897 void
TruncateCLOG(TransactionId oldestXact,Oid oldestxid_datoid)898 TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
899 {
900 int cutoffPage;
901
902 /*
903 * The cutoff point is the start of the segment containing oldestXact. We
904 * pass the *page* containing oldestXact to SimpleLruTruncate.
905 */
906 cutoffPage = TransactionIdToPage(oldestXact);
907
908 /* Check to see if there's any files that could be removed */
909 if (!SlruScanDirectory(ClogCtl, SlruScanDirCbReportPresence, &cutoffPage))
910 return; /* nothing to remove */
911
912 /*
913 * Advance oldestClogXid before truncating clog, so concurrent xact status
914 * lookups can ensure they don't attempt to access truncated-away clog.
915 *
916 * It's only necessary to do this if we will actually truncate away clog
917 * pages.
918 */
919 AdvanceOldestClogXid(oldestXact);
920
921 /*
922 * Write XLOG record and flush XLOG to disk. We record the oldest xid
923 * we're keeping information about here so we can ensure that it's always
924 * ahead of clog truncation in case we crash, and so a standby finds out
925 * the new valid xid before the next checkpoint.
926 */
927 WriteTruncateXlogRec(cutoffPage, oldestXact, oldestxid_datoid);
928
929 /* Now we can remove the old CLOG segment(s) */
930 SimpleLruTruncate(ClogCtl, cutoffPage);
931 }
932
933
934 /*
935 * Decide whether a CLOG page number is "older" for truncation purposes.
936 *
937 * We need to use comparison of TransactionIds here in order to do the right
938 * thing with wraparound XID arithmetic. However, TransactionIdPrecedes()
939 * would get weird about permanent xact IDs. So, offset both such that xid1,
940 * xid2, and xid2 + CLOG_XACTS_PER_PAGE - 1 are all normal XIDs; this offset
941 * is relevant to page 0 and to the page preceding page 0.
942 *
943 * The page containing oldestXact-2^31 is the important edge case. The
944 * portion of that page equaling or following oldestXact-2^31 is expendable,
945 * but the portion preceding oldestXact-2^31 is not. When oldestXact-2^31 is
946 * the first XID of a page and segment, the entire page and segment is
947 * expendable, and we could truncate the segment. Recognizing that case would
948 * require making oldestXact, not just the page containing oldestXact,
949 * available to this callback. The benefit would be rare and small, so we
950 * don't optimize that edge case.
951 */
952 static bool
CLOGPagePrecedes(int page1,int page2)953 CLOGPagePrecedes(int page1, int page2)
954 {
955 TransactionId xid1;
956 TransactionId xid2;
957
958 xid1 = ((TransactionId) page1) * CLOG_XACTS_PER_PAGE;
959 xid1 += FirstNormalTransactionId + 1;
960 xid2 = ((TransactionId) page2) * CLOG_XACTS_PER_PAGE;
961 xid2 += FirstNormalTransactionId + 1;
962
963 return (TransactionIdPrecedes(xid1, xid2) &&
964 TransactionIdPrecedes(xid1, xid2 + CLOG_XACTS_PER_PAGE - 1));
965 }
966
967
968 /*
969 * Write a ZEROPAGE xlog record
970 */
971 static void
WriteZeroPageXlogRec(int pageno)972 WriteZeroPageXlogRec(int pageno)
973 {
974 XLogBeginInsert();
975 XLogRegisterData((char *) (&pageno), sizeof(int));
976 (void) XLogInsert(RM_CLOG_ID, CLOG_ZEROPAGE);
977 }
978
979 /*
980 * Write a TRUNCATE xlog record
981 *
982 * We must flush the xlog record to disk before returning --- see notes
983 * in TruncateCLOG().
984 */
985 static void
WriteTruncateXlogRec(int pageno,TransactionId oldestXact,Oid oldestXactDb)986 WriteTruncateXlogRec(int pageno, TransactionId oldestXact, Oid oldestXactDb)
987 {
988 XLogRecPtr recptr;
989 xl_clog_truncate xlrec;
990
991 xlrec.pageno = pageno;
992 xlrec.oldestXact = oldestXact;
993 xlrec.oldestXactDb = oldestXactDb;
994
995 XLogBeginInsert();
996 XLogRegisterData((char *) (&xlrec), sizeof(xl_clog_truncate));
997 recptr = XLogInsert(RM_CLOG_ID, CLOG_TRUNCATE);
998 XLogFlush(recptr);
999 }
1000
1001 /*
1002 * CLOG resource manager's routines
1003 */
1004 void
clog_redo(XLogReaderState * record)1005 clog_redo(XLogReaderState *record)
1006 {
1007 uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
1008
1009 /* Backup blocks are not used in clog records */
1010 Assert(!XLogRecHasAnyBlockRefs(record));
1011
1012 if (info == CLOG_ZEROPAGE)
1013 {
1014 int pageno;
1015 int slotno;
1016
1017 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
1018
1019 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
1020
1021 slotno = ZeroCLOGPage(pageno, false);
1022 SimpleLruWritePage(ClogCtl, slotno);
1023 Assert(!ClogCtl->shared->page_dirty[slotno]);
1024
1025 LWLockRelease(CLogControlLock);
1026 }
1027 else if (info == CLOG_TRUNCATE)
1028 {
1029 xl_clog_truncate xlrec;
1030
1031 memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_clog_truncate));
1032
1033 /*
1034 * During XLOG replay, latest_page_number isn't set up yet; insert a
1035 * suitable value to bypass the sanity test in SimpleLruTruncate.
1036 */
1037 ClogCtl->shared->latest_page_number = xlrec.pageno;
1038
1039 AdvanceOldestClogXid(xlrec.oldestXact);
1040
1041 SimpleLruTruncate(ClogCtl, xlrec.pageno);
1042 }
1043 else
1044 elog(PANIC, "clog_redo: unknown op code %u", info);
1045 }
1046