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) 1996-2017, 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 "pg_trace.h"
43
44 /*
45 * Defines for CLOG page sizes. A page is the same BLCKSZ as is used
46 * everywhere else in Postgres.
47 *
48 * Note: because TransactionIds are 32 bits and wrap around at 0xFFFFFFFF,
49 * CLOG page numbering also wraps around at 0xFFFFFFFF/CLOG_XACTS_PER_PAGE,
50 * and CLOG segment numbering at
51 * 0xFFFFFFFF/CLOG_XACTS_PER_PAGE/SLRU_PAGES_PER_SEGMENT. We need take no
52 * explicit notice of that fact in this module, except when comparing segment
53 * and page numbers in TruncateCLOG (see CLOGPagePrecedes).
54 */
55
56 /* We need two bits per xact, so four xacts fit in a byte */
57 #define CLOG_BITS_PER_XACT 2
58 #define CLOG_XACTS_PER_BYTE 4
59 #define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
60 #define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
61
62 #define TransactionIdToPage(xid) ((xid) / (TransactionId) CLOG_XACTS_PER_PAGE)
63 #define TransactionIdToPgIndex(xid) ((xid) % (TransactionId) CLOG_XACTS_PER_PAGE)
64 #define TransactionIdToByte(xid) (TransactionIdToPgIndex(xid) / CLOG_XACTS_PER_BYTE)
65 #define TransactionIdToBIndex(xid) ((xid) % (TransactionId) CLOG_XACTS_PER_BYTE)
66
67 /* We store the latest async LSN for each group of transactions */
68 #define CLOG_XACTS_PER_LSN_GROUP 32 /* keep this a power of 2 */
69 #define CLOG_LSNS_PER_PAGE (CLOG_XACTS_PER_PAGE / CLOG_XACTS_PER_LSN_GROUP)
70
71 #define GetLSNIndex(slotno, xid) ((slotno) * CLOG_LSNS_PER_PAGE + \
72 ((xid) % (TransactionId) CLOG_XACTS_PER_PAGE) / CLOG_XACTS_PER_LSN_GROUP)
73
74
75 /*
76 * Link to shared-memory data structures for CLOG control
77 */
78 static SlruCtlData ClogCtlData;
79
80 #define ClogCtl (&ClogCtlData)
81
82
83 static int ZeroCLOGPage(int pageno, bool writeXlog);
84 static bool CLOGPagePrecedes(int page1, int page2);
85 static void WriteZeroPageXlogRec(int pageno);
86 static void WriteTruncateXlogRec(int pageno, TransactionId oldestXact,
87 Oid oldestXidDb);
88 static void TransactionIdSetPageStatus(TransactionId xid, int nsubxids,
89 TransactionId *subxids, XidStatus status,
90 XLogRecPtr lsn, int pageno);
91 static void TransactionIdSetStatusBit(TransactionId xid, XidStatus status,
92 XLogRecPtr lsn, int slotno);
93 static void set_status_by_pages(int nsubxids, TransactionId *subxids,
94 XidStatus status, XLogRecPtr lsn);
95
96
97 /*
98 * TransactionIdSetTreeStatus
99 *
100 * Record the final state of transaction entries in the commit log for
101 * a transaction and its subtransaction tree. Take care to ensure this is
102 * efficient, and as atomic as possible.
103 *
104 * xid is a single xid to set status for. This will typically be
105 * the top level transactionid for a top level commit or abort. It can
106 * also be a subtransaction when we record transaction aborts.
107 *
108 * subxids is an array of xids of length nsubxids, representing subtransactions
109 * in the tree of xid. In various cases nsubxids may be zero.
110 *
111 * lsn must be the WAL location of the commit record when recording an async
112 * commit. For a synchronous commit it can be InvalidXLogRecPtr, since the
113 * caller guarantees the commit record is already flushed in that case. It
114 * should be InvalidXLogRecPtr for abort cases, too.
115 *
116 * In the commit case, atomicity is limited by whether all the subxids are in
117 * the same CLOG page as xid. If they all are, then the lock will be grabbed
118 * only once, and the status will be set to committed directly. Otherwise
119 * we must
120 * 1. set sub-committed all subxids that are not on the same page as the
121 * main xid
122 * 2. atomically set committed the main xid and the subxids on the same page
123 * 3. go over the first bunch again and set them committed
124 * Note that as far as concurrent checkers are concerned, main transaction
125 * commit as a whole is still atomic.
126 *
127 * Example:
128 * TransactionId t commits and has subxids t1, t2, t3, t4
129 * t is on page p1, t1 is also on p1, t2 and t3 are on p2, t4 is on p3
130 * 1. update pages2-3:
131 * page2: set t2,t3 as sub-committed
132 * page3: set t4 as sub-committed
133 * 2. update page1:
134 * set t1 as sub-committed,
135 * then set t as committed,
136 then set t1 as committed
137 * 3. update pages2-3:
138 * page2: set t2,t3 as committed
139 * page3: set t4 as committed
140 *
141 * NB: this is a low-level routine and is NOT the preferred entry point
142 * for most uses; functions in transam.c are the intended callers.
143 *
144 * XXX Think about issuing FADVISE_WILLNEED on pages that we will need,
145 * but aren't yet in cache, as well as hinting pages not to fall out of
146 * cache yet.
147 */
148 void
TransactionIdSetTreeStatus(TransactionId xid,int nsubxids,TransactionId * subxids,XidStatus status,XLogRecPtr lsn)149 TransactionIdSetTreeStatus(TransactionId xid, int nsubxids,
150 TransactionId *subxids, XidStatus status, XLogRecPtr lsn)
151 {
152 int pageno = TransactionIdToPage(xid); /* get page of parent */
153 int i;
154
155 Assert(status == TRANSACTION_STATUS_COMMITTED ||
156 status == TRANSACTION_STATUS_ABORTED);
157
158 /*
159 * See how many subxids, if any, are on the same page as the parent, if
160 * any.
161 */
162 for (i = 0; i < nsubxids; i++)
163 {
164 if (TransactionIdToPage(subxids[i]) != pageno)
165 break;
166 }
167
168 /*
169 * Do all items fit on a single page?
170 */
171 if (i == nsubxids)
172 {
173 /*
174 * Set the parent and all subtransactions in a single call
175 */
176 TransactionIdSetPageStatus(xid, nsubxids, subxids, status, lsn,
177 pageno);
178 }
179 else
180 {
181 int nsubxids_on_first_page = i;
182
183 /*
184 * If this is a commit then we care about doing this correctly (i.e.
185 * using the subcommitted intermediate status). By here, we know
186 * we're updating more than one page of clog, so we must mark entries
187 * that are *not* on the first page so that they show as subcommitted
188 * before we then return to update the status to fully committed.
189 *
190 * To avoid touching the first page twice, skip marking subcommitted
191 * for the subxids on that first page.
192 */
193 if (status == TRANSACTION_STATUS_COMMITTED)
194 set_status_by_pages(nsubxids - nsubxids_on_first_page,
195 subxids + nsubxids_on_first_page,
196 TRANSACTION_STATUS_SUB_COMMITTED, lsn);
197
198 /*
199 * Now set the parent and subtransactions on same page as the parent,
200 * if any
201 */
202 pageno = TransactionIdToPage(xid);
203 TransactionIdSetPageStatus(xid, nsubxids_on_first_page, subxids, status,
204 lsn, pageno);
205
206 /*
207 * Now work through the rest of the subxids one clog page at a time,
208 * starting from the second page onwards, like we did above.
209 */
210 set_status_by_pages(nsubxids - nsubxids_on_first_page,
211 subxids + nsubxids_on_first_page,
212 status, lsn);
213 }
214 }
215
216 /*
217 * Helper for TransactionIdSetTreeStatus: set the status for a bunch of
218 * transactions, chunking in the separate CLOG pages involved. We never
219 * pass the whole transaction tree to this function, only subtransactions
220 * that are on different pages to the top level transaction id.
221 */
222 static void
set_status_by_pages(int nsubxids,TransactionId * subxids,XidStatus status,XLogRecPtr lsn)223 set_status_by_pages(int nsubxids, TransactionId *subxids,
224 XidStatus status, XLogRecPtr lsn)
225 {
226 int pageno = TransactionIdToPage(subxids[0]);
227 int offset = 0;
228 int i = 0;
229
230 Assert(nsubxids > 0); /* else the pageno fetch above is unsafe */
231
232 while (i < nsubxids)
233 {
234 int num_on_page = 0;
235 int nextpageno;
236
237 do
238 {
239 nextpageno = TransactionIdToPage(subxids[i]);
240 if (nextpageno != pageno)
241 break;
242 num_on_page++;
243 i++;
244 } while (i < nsubxids);
245
246 TransactionIdSetPageStatus(InvalidTransactionId,
247 num_on_page, subxids + offset,
248 status, lsn, pageno);
249 offset = i;
250 pageno = nextpageno;
251 }
252 }
253
254 /*
255 * Record the final state of transaction entries in the commit log for
256 * all entries on a single page. Atomic only on this page.
257 *
258 * Otherwise API is same as TransactionIdSetTreeStatus()
259 */
260 static void
TransactionIdSetPageStatus(TransactionId xid,int nsubxids,TransactionId * subxids,XidStatus status,XLogRecPtr lsn,int pageno)261 TransactionIdSetPageStatus(TransactionId xid, int nsubxids,
262 TransactionId *subxids, XidStatus status,
263 XLogRecPtr lsn, int pageno)
264 {
265 int slotno;
266 int i;
267
268 Assert(status == TRANSACTION_STATUS_COMMITTED ||
269 status == TRANSACTION_STATUS_ABORTED ||
270 (status == TRANSACTION_STATUS_SUB_COMMITTED && !TransactionIdIsValid(xid)));
271
272 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
273
274 /*
275 * If we're doing an async commit (ie, lsn is valid), then we must wait
276 * for any active write on the page slot to complete. Otherwise our
277 * update could reach disk in that write, which will not do since we
278 * mustn't let it reach disk until we've done the appropriate WAL flush.
279 * But when lsn is invalid, it's OK to scribble on a page while it is
280 * write-busy, since we don't care if the update reaches disk sooner than
281 * we think.
282 */
283 slotno = SimpleLruReadPage(ClogCtl, pageno, XLogRecPtrIsInvalid(lsn), xid);
284
285 /*
286 * Set the main transaction id, if any.
287 *
288 * If we update more than one xid on this page while it is being written
289 * out, we might find that some of the bits go to disk and others don't.
290 * If we are updating commits on the page with the top-level xid that
291 * could break atomicity, so we subcommit the subxids first before we mark
292 * the top-level commit.
293 */
294 if (TransactionIdIsValid(xid))
295 {
296 /* Subtransactions first, if needed ... */
297 if (status == TRANSACTION_STATUS_COMMITTED)
298 {
299 for (i = 0; i < nsubxids; i++)
300 {
301 Assert(ClogCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
302 TransactionIdSetStatusBit(subxids[i],
303 TRANSACTION_STATUS_SUB_COMMITTED,
304 lsn, slotno);
305 }
306 }
307
308 /* ... then the main transaction */
309 TransactionIdSetStatusBit(xid, status, lsn, slotno);
310 }
311
312 /* Set the subtransactions */
313 for (i = 0; i < nsubxids; i++)
314 {
315 Assert(ClogCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
316 TransactionIdSetStatusBit(subxids[i], status, lsn, slotno);
317 }
318
319 ClogCtl->shared->page_dirty[slotno] = true;
320
321 LWLockRelease(CLogControlLock);
322 }
323
324 /*
325 * Sets the commit status of a single transaction.
326 *
327 * Must be called with CLogControlLock held
328 */
329 static void
TransactionIdSetStatusBit(TransactionId xid,XidStatus status,XLogRecPtr lsn,int slotno)330 TransactionIdSetStatusBit(TransactionId xid, XidStatus status, XLogRecPtr lsn, int slotno)
331 {
332 int byteno = TransactionIdToByte(xid);
333 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
334 char *byteptr;
335 char byteval;
336 char curval;
337
338 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
339 curval = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
340
341 /*
342 * When replaying transactions during recovery we still need to perform
343 * the two phases of subcommit and then commit. However, some transactions
344 * are already correctly marked, so we just treat those as a no-op which
345 * allows us to keep the following Assert as restrictive as possible.
346 */
347 if (InRecovery && status == TRANSACTION_STATUS_SUB_COMMITTED &&
348 curval == TRANSACTION_STATUS_COMMITTED)
349 return;
350
351 /*
352 * Current state change should be from 0 or subcommitted to target state
353 * or we should already be there when replaying changes during recovery.
354 */
355 Assert(curval == 0 ||
356 (curval == TRANSACTION_STATUS_SUB_COMMITTED &&
357 status != TRANSACTION_STATUS_IN_PROGRESS) ||
358 curval == status);
359
360 /* note this assumes exclusive access to the clog page */
361 byteval = *byteptr;
362 byteval &= ~(((1 << CLOG_BITS_PER_XACT) - 1) << bshift);
363 byteval |= (status << bshift);
364 *byteptr = byteval;
365
366 /*
367 * Update the group LSN if the transaction completion LSN is higher.
368 *
369 * Note: lsn will be invalid when supplied during InRecovery processing,
370 * so we don't need to do anything special to avoid LSN updates during
371 * recovery. After recovery completes the next clog change will set the
372 * LSN correctly.
373 */
374 if (!XLogRecPtrIsInvalid(lsn))
375 {
376 int lsnindex = GetLSNIndex(slotno, xid);
377
378 if (ClogCtl->shared->group_lsn[lsnindex] < lsn)
379 ClogCtl->shared->group_lsn[lsnindex] = lsn;
380 }
381 }
382
383 /*
384 * Interrogate the state of a transaction in the commit log.
385 *
386 * Aside from the actual commit status, this function returns (into *lsn)
387 * an LSN that is late enough to be able to guarantee that if we flush up to
388 * that LSN then we will have flushed the transaction's commit record to disk.
389 * The result is not necessarily the exact LSN of the transaction's commit
390 * record! For example, for long-past transactions (those whose clog pages
391 * already migrated to disk), we'll return InvalidXLogRecPtr. Also, because
392 * we group transactions on the same clog page to conserve storage, we might
393 * return the LSN of a later transaction that falls into the same group.
394 *
395 * NB: this is a low-level routine and is NOT the preferred entry point
396 * for most uses; TransactionLogFetch() in transam.c is the intended caller.
397 */
398 XidStatus
TransactionIdGetStatus(TransactionId xid,XLogRecPtr * lsn)399 TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
400 {
401 int pageno = TransactionIdToPage(xid);
402 int byteno = TransactionIdToByte(xid);
403 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
404 int slotno;
405 int lsnindex;
406 char *byteptr;
407 XidStatus status;
408
409 /* lock is acquired by SimpleLruReadPage_ReadOnly */
410
411 slotno = SimpleLruReadPage_ReadOnly(ClogCtl, pageno, xid);
412 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
413
414 status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
415
416 lsnindex = GetLSNIndex(slotno, xid);
417 *lsn = ClogCtl->shared->group_lsn[lsnindex];
418
419 LWLockRelease(CLogControlLock);
420
421 return status;
422 }
423
424 /*
425 * Number of shared CLOG buffers.
426 *
427 * On larger multi-processor systems, it is possible to have many CLOG page
428 * requests in flight at one time which could lead to disk access for CLOG
429 * page if the required page is not found in memory. Testing revealed that we
430 * can get the best performance by having 128 CLOG buffers, more than that it
431 * doesn't improve performance.
432 *
433 * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
434 * a good idea, because it would increase the minimum amount of shared memory
435 * required to start, which could be a problem for people running very small
436 * configurations. The following formula seems to represent a reasonable
437 * compromise: people with very low values for shared_buffers will get fewer
438 * CLOG buffers as well, and everyone else will get 128.
439 */
440 Size
CLOGShmemBuffers(void)441 CLOGShmemBuffers(void)
442 {
443 return Min(128, Max(4, NBuffers / 512));
444 }
445
446 /*
447 * Initialization of shared memory for CLOG
448 */
449 Size
CLOGShmemSize(void)450 CLOGShmemSize(void)
451 {
452 return SimpleLruShmemSize(CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE);
453 }
454
455 void
CLOGShmemInit(void)456 CLOGShmemInit(void)
457 {
458 ClogCtl->PagePrecedes = CLOGPagePrecedes;
459 SimpleLruInit(ClogCtl, "clog", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
460 CLogControlLock, "pg_xact", LWTRANCHE_CLOG_BUFFERS);
461 SlruPagePrecedesUnitTests(ClogCtl, CLOG_XACTS_PER_PAGE);
462 }
463
464 /*
465 * This func must be called ONCE on system install. It creates
466 * the initial CLOG segment. (The CLOG directory is assumed to
467 * have been created by initdb, and CLOGShmemInit must have been
468 * called already.)
469 */
470 void
BootStrapCLOG(void)471 BootStrapCLOG(void)
472 {
473 int slotno;
474
475 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
476
477 /* Create and zero the first page of the commit log */
478 slotno = ZeroCLOGPage(0, false);
479
480 /* Make sure it's written out */
481 SimpleLruWritePage(ClogCtl, slotno);
482 Assert(!ClogCtl->shared->page_dirty[slotno]);
483
484 LWLockRelease(CLogControlLock);
485 }
486
487 /*
488 * Initialize (or reinitialize) a page of CLOG to zeroes.
489 * If writeXlog is TRUE, also emit an XLOG record saying we did this.
490 *
491 * The page is not actually written, just set up in shared memory.
492 * The slot number of the new page is returned.
493 *
494 * Control lock must be held at entry, and will be held at exit.
495 */
496 static int
ZeroCLOGPage(int pageno,bool writeXlog)497 ZeroCLOGPage(int pageno, bool writeXlog)
498 {
499 int slotno;
500
501 slotno = SimpleLruZeroPage(ClogCtl, pageno);
502
503 if (writeXlog)
504 WriteZeroPageXlogRec(pageno);
505
506 return slotno;
507 }
508
509 /*
510 * This must be called ONCE during postmaster or standalone-backend startup,
511 * after StartupXLOG has initialized ShmemVariableCache->nextXid.
512 */
513 void
StartupCLOG(void)514 StartupCLOG(void)
515 {
516 TransactionId xid = ShmemVariableCache->nextXid;
517 int pageno = TransactionIdToPage(xid);
518
519 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
520
521 /*
522 * Initialize our idea of the latest page number.
523 */
524 ClogCtl->shared->latest_page_number = pageno;
525
526 LWLockRelease(CLogControlLock);
527 }
528
529 /*
530 * This must be called ONCE at the end of startup/recovery.
531 */
532 void
TrimCLOG(void)533 TrimCLOG(void)
534 {
535 TransactionId xid = ShmemVariableCache->nextXid;
536 int pageno = TransactionIdToPage(xid);
537
538 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
539
540 /*
541 * Re-Initialize our idea of the latest page number.
542 */
543 ClogCtl->shared->latest_page_number = pageno;
544
545 /*
546 * Zero out the remainder of the current clog page. Under normal
547 * circumstances it should be zeroes already, but it seems at least
548 * theoretically possible that XLOG replay will have settled on a nextXID
549 * value that is less than the last XID actually used and marked by the
550 * previous database lifecycle (since subtransaction commit writes clog
551 * but makes no WAL entry). Let's just be safe. (We need not worry about
552 * pages beyond the current one, since those will be zeroed when first
553 * used. For the same reason, there is no need to do anything when
554 * nextXid is exactly at a page boundary; and it's likely that the
555 * "current" page doesn't exist yet in that case.)
556 */
557 if (TransactionIdToPgIndex(xid) != 0)
558 {
559 int byteno = TransactionIdToByte(xid);
560 int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
561 int slotno;
562 char *byteptr;
563
564 slotno = SimpleLruReadPage(ClogCtl, pageno, false, xid);
565 byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
566
567 /* Zero so-far-unused positions in the current byte */
568 *byteptr &= (1 << bshift) - 1;
569 /* Zero the rest of the page */
570 MemSet(byteptr + 1, 0, BLCKSZ - byteno - 1);
571
572 ClogCtl->shared->page_dirty[slotno] = true;
573 }
574
575 LWLockRelease(CLogControlLock);
576 }
577
578 /*
579 * This must be called ONCE during postmaster or standalone-backend shutdown
580 */
581 void
ShutdownCLOG(void)582 ShutdownCLOG(void)
583 {
584 /* Flush dirty CLOG pages to disk */
585 TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(false);
586 SimpleLruFlush(ClogCtl, false);
587
588 /*
589 * fsync pg_xact to ensure that any files flushed previously are durably
590 * on disk.
591 */
592 fsync_fname("pg_xact", true);
593
594 TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(false);
595 }
596
597 /*
598 * Perform a checkpoint --- either during shutdown, or on-the-fly
599 */
600 void
CheckPointCLOG(void)601 CheckPointCLOG(void)
602 {
603 /* Flush dirty CLOG pages to disk */
604 TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(true);
605 SimpleLruFlush(ClogCtl, true);
606 TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(true);
607 }
608
609
610 /*
611 * Make sure that CLOG has room for a newly-allocated XID.
612 *
613 * NB: this is called while holding XidGenLock. We want it to be very fast
614 * most of the time; even when it's not so fast, no actual I/O need happen
615 * unless we're forced to write out a dirty clog or xlog page to make room
616 * in shared memory.
617 */
618 void
ExtendCLOG(TransactionId newestXact)619 ExtendCLOG(TransactionId newestXact)
620 {
621 int pageno;
622
623 /*
624 * No work except at first XID of a page. But beware: just after
625 * wraparound, the first XID of page zero is FirstNormalTransactionId.
626 */
627 if (TransactionIdToPgIndex(newestXact) != 0 &&
628 !TransactionIdEquals(newestXact, FirstNormalTransactionId))
629 return;
630
631 pageno = TransactionIdToPage(newestXact);
632
633 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
634
635 /* Zero the page and make an XLOG entry about it */
636 ZeroCLOGPage(pageno, true);
637
638 LWLockRelease(CLogControlLock);
639 }
640
641
642 /*
643 * Remove all CLOG segments before the one holding the passed transaction ID
644 *
645 * Before removing any CLOG data, we must flush XLOG to disk, to ensure
646 * that any recently-emitted HEAP_FREEZE records have reached disk; otherwise
647 * a crash and restart might leave us with some unfrozen tuples referencing
648 * removed CLOG data. We choose to emit a special TRUNCATE XLOG record too.
649 * Replaying the deletion from XLOG is not critical, since the files could
650 * just as well be removed later, but doing so prevents a long-running hot
651 * standby server from acquiring an unreasonably bloated CLOG directory.
652 *
653 * Since CLOG segments hold a large number of transactions, the opportunity to
654 * actually remove a segment is fairly rare, and so it seems best not to do
655 * the XLOG flush unless we have confirmed that there is a removable segment.
656 */
657 void
TruncateCLOG(TransactionId oldestXact,Oid oldestxid_datoid)658 TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
659 {
660 int cutoffPage;
661
662 /*
663 * The cutoff point is the start of the segment containing oldestXact. We
664 * pass the *page* containing oldestXact to SimpleLruTruncate.
665 */
666 cutoffPage = TransactionIdToPage(oldestXact);
667
668 /* Check to see if there's any files that could be removed */
669 if (!SlruScanDirectory(ClogCtl, SlruScanDirCbReportPresence, &cutoffPage))
670 return; /* nothing to remove */
671
672 /*
673 * Advance oldestClogXid before truncating clog, so concurrent xact status
674 * lookups can ensure they don't attempt to access truncated-away clog.
675 *
676 * It's only necessary to do this if we will actually truncate away clog
677 * pages.
678 */
679 AdvanceOldestClogXid(oldestXact);
680
681 /*
682 * Write XLOG record and flush XLOG to disk. We record the oldest xid
683 * we're keeping information about here so we can ensure that it's always
684 * ahead of clog truncation in case we crash, and so a standby finds out
685 * the new valid xid before the next checkpoint.
686 */
687 WriteTruncateXlogRec(cutoffPage, oldestXact, oldestxid_datoid);
688
689 /* Now we can remove the old CLOG segment(s) */
690 SimpleLruTruncate(ClogCtl, cutoffPage);
691 }
692
693
694 /*
695 * Decide whether a CLOG page number is "older" for truncation purposes.
696 *
697 * We need to use comparison of TransactionIds here in order to do the right
698 * thing with wraparound XID arithmetic. However, TransactionIdPrecedes()
699 * would get weird about permanent xact IDs. So, offset both such that xid1,
700 * xid2, and xid2 + CLOG_XACTS_PER_PAGE - 1 are all normal XIDs; this offset
701 * is relevant to page 0 and to the page preceding page 0.
702 *
703 * The page containing oldestXact-2^31 is the important edge case. The
704 * portion of that page equaling or following oldestXact-2^31 is expendable,
705 * but the portion preceding oldestXact-2^31 is not. When oldestXact-2^31 is
706 * the first XID of a page and segment, the entire page and segment is
707 * expendable, and we could truncate the segment. Recognizing that case would
708 * require making oldestXact, not just the page containing oldestXact,
709 * available to this callback. The benefit would be rare and small, so we
710 * don't optimize that edge case.
711 */
712 static bool
CLOGPagePrecedes(int page1,int page2)713 CLOGPagePrecedes(int page1, int page2)
714 {
715 TransactionId xid1;
716 TransactionId xid2;
717
718 xid1 = ((TransactionId) page1) * CLOG_XACTS_PER_PAGE;
719 xid1 += FirstNormalTransactionId + 1;
720 xid2 = ((TransactionId) page2) * CLOG_XACTS_PER_PAGE;
721 xid2 += FirstNormalTransactionId + 1;
722
723 return (TransactionIdPrecedes(xid1, xid2) &&
724 TransactionIdPrecedes(xid1, xid2 + CLOG_XACTS_PER_PAGE - 1));
725 }
726
727
728 /*
729 * Write a ZEROPAGE xlog record
730 */
731 static void
WriteZeroPageXlogRec(int pageno)732 WriteZeroPageXlogRec(int pageno)
733 {
734 XLogBeginInsert();
735 XLogRegisterData((char *) (&pageno), sizeof(int));
736 (void) XLogInsert(RM_CLOG_ID, CLOG_ZEROPAGE);
737 }
738
739 /*
740 * Write a TRUNCATE xlog record
741 *
742 * We must flush the xlog record to disk before returning --- see notes
743 * in TruncateCLOG().
744 */
745 static void
WriteTruncateXlogRec(int pageno,TransactionId oldestXact,Oid oldestXactDb)746 WriteTruncateXlogRec(int pageno, TransactionId oldestXact, Oid oldestXactDb)
747 {
748 XLogRecPtr recptr;
749 xl_clog_truncate xlrec;
750
751 xlrec.pageno = pageno;
752 xlrec.oldestXact = oldestXact;
753 xlrec.oldestXactDb = oldestXactDb;
754
755 XLogBeginInsert();
756 XLogRegisterData((char *) (&xlrec), sizeof(xl_clog_truncate));
757 recptr = XLogInsert(RM_CLOG_ID, CLOG_TRUNCATE);
758 XLogFlush(recptr);
759 }
760
761 /*
762 * CLOG resource manager's routines
763 */
764 void
clog_redo(XLogReaderState * record)765 clog_redo(XLogReaderState *record)
766 {
767 uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
768
769 /* Backup blocks are not used in clog records */
770 Assert(!XLogRecHasAnyBlockRefs(record));
771
772 if (info == CLOG_ZEROPAGE)
773 {
774 int pageno;
775 int slotno;
776
777 memcpy(&pageno, XLogRecGetData(record), sizeof(int));
778
779 LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
780
781 slotno = ZeroCLOGPage(pageno, false);
782 SimpleLruWritePage(ClogCtl, slotno);
783 Assert(!ClogCtl->shared->page_dirty[slotno]);
784
785 LWLockRelease(CLogControlLock);
786 }
787 else if (info == CLOG_TRUNCATE)
788 {
789 xl_clog_truncate xlrec;
790
791 memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_clog_truncate));
792
793 /*
794 * During XLOG replay, latest_page_number isn't set up yet; insert a
795 * suitable value to bypass the sanity test in SimpleLruTruncate.
796 */
797 ClogCtl->shared->latest_page_number = xlrec.pageno;
798
799 AdvanceOldestClogXid(xlrec.oldestXact);
800
801 SimpleLruTruncate(ClogCtl, xlrec.pageno);
802 }
803 else
804 elog(PANIC, "clog_redo: unknown op code %u", info);
805 }
806