Line data Source code
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 "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
163 10602 : TransactionIdSetTreeStatus(TransactionId xid, int nsubxids,
164 : TransactionId *subxids, XidStatus status, XLogRecPtr lsn)
165 : {
166 10602 : int pageno = TransactionIdToPage(xid); /* get page of parent */
167 : int i;
168 :
169 10602 : 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 10626 : for (i = 0; i < nsubxids; i++)
177 : {
178 24 : if (TransactionIdToPage(subxids[i]) != pageno)
179 0 : break;
180 : }
181 :
182 : /*
183 : * Do all items fit on a single page?
184 : */
185 10602 : if (i == nsubxids)
186 : {
187 : /*
188 : * Set the parent and all subtransactions in a single call
189 : */
190 10602 : TransactionIdSetPageStatus(xid, nsubxids, subxids, status, lsn,
191 : pageno, true);
192 : }
193 : else
194 : {
195 0 : 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 0 : if (status == TRANSACTION_STATUS_COMMITTED)
208 0 : set_status_by_pages(nsubxids - nsubxids_on_first_page,
209 0 : 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 0 : pageno = TransactionIdToPage(xid);
217 0 : 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 0 : set_status_by_pages(nsubxids - nsubxids_on_first_page,
225 0 : subxids + nsubxids_on_first_page,
226 : status, lsn);
227 : }
228 10602 : }
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
237 0 : set_status_by_pages(int nsubxids, TransactionId *subxids,
238 : XidStatus status, XLogRecPtr lsn)
239 : {
240 0 : int pageno = TransactionIdToPage(subxids[0]);
241 0 : int offset = 0;
242 0 : int i = 0;
243 :
244 0 : while (i < nsubxids)
245 : {
246 0 : int num_on_page = 0;
247 :
248 0 : while (TransactionIdToPage(subxids[i]) == pageno && i < nsubxids)
249 : {
250 0 : num_on_page++;
251 0 : i++;
252 : }
253 :
254 0 : TransactionIdSetPageStatus(InvalidTransactionId,
255 0 : num_on_page, subxids + offset,
256 : status, lsn, pageno, false);
257 0 : offset = i;
258 0 : pageno = TransactionIdToPage(subxids[offset]);
259 : }
260 0 : }
261 :
262 : /*
263 : * Record the final state of transaction entries in the commit log for all
264 : * entries on a single page. Atomic only on this page.
265 : */
266 : static void
267 10602 : TransactionIdSetPageStatus(TransactionId xid, int nsubxids,
268 : TransactionId *subxids, XidStatus status,
269 : XLogRecPtr lsn, int pageno,
270 : bool all_xact_same_page)
271 : {
272 : /* Can't use group update when PGPROC overflows. */
273 : StaticAssertStmt(THRESHOLD_SUBTRANS_CLOG_OPT <= PGPROC_MAX_CACHED_SUBXIDS,
274 : "group clog threshold less than PGPROC cached subxids");
275 :
276 : /*
277 : * When there is contention on CLogControlLock, we try to group multiple
278 : * updates; a single leader process will perform transaction status
279 : * updates for multiple backends so that the number of times
280 : * CLogControlLock needs to be acquired is reduced.
281 : *
282 : * For this optimization to be safe, the XID in MyPgXact and the subxids
283 : * in MyProc must be the same as the ones for which we're setting the
284 : * status. Check that this is the case.
285 : *
286 : * For this optimization to be efficient, we shouldn't have too many
287 : * sub-XIDs and all of the XIDs for which we're adjusting clog should be
288 : * on the same page. Check those conditions, too.
289 : */
290 10602 : if (all_xact_same_page && xid == MyPgXact->xid &&
291 10556 : nsubxids <= THRESHOLD_SUBTRANS_CLOG_OPT &&
292 21112 : nsubxids == MyPgXact->nxids &&
293 10556 : memcmp(subxids, MyProc->subxids.xids,
294 : nsubxids * sizeof(TransactionId)) == 0)
295 : {
296 : /*
297 : * We don't try to do group update optimization if a process has
298 : * overflowed the subxids array in its PGPROC, since in that case we
299 : * don't have a complete list of XIDs for it.
300 : */
301 : Assert(THRESHOLD_SUBTRANS_CLOG_OPT <= PGPROC_MAX_CACHED_SUBXIDS);
302 :
303 : /*
304 : * If we can immediately acquire CLogControlLock, we update the status
305 : * of our own XID and release the lock. If not, try use group XID
306 : * update. If that doesn't work out, fall back to waiting for the
307 : * lock to perform an update for this transaction only.
308 : */
309 10556 : if (LWLockConditionalAcquire(CLogControlLock, LW_EXCLUSIVE))
310 : {
311 : /* Got the lock without waiting! Do the update. */
312 10540 : TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
313 : lsn, pageno);
314 10540 : LWLockRelease(CLogControlLock);
315 10540 : return;
316 : }
317 16 : else if (TransactionGroupUpdateXidStatus(xid, status, lsn, pageno))
318 : {
319 : /* Group update mechanism has done the work. */
320 16 : return;
321 : }
322 :
323 : /* Fall through only if update isn't done yet. */
324 : }
325 :
326 : /* Group update not applicable, or couldn't accept this page number. */
327 46 : LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
328 46 : TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
329 : lsn, pageno);
330 46 : LWLockRelease(CLogControlLock);
331 : }
332 :
333 : /*
334 : * Record the final state of transaction entry in the commit log
335 : *
336 : * We don't do any locking here; caller must handle that.
337 : */
338 : static void
339 10602 : TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids,
340 : TransactionId *subxids, XidStatus status,
341 : XLogRecPtr lsn, int pageno)
342 : {
343 : int slotno;
344 : int i;
345 :
346 10602 : Assert(status == TRANSACTION_STATUS_COMMITTED ||
347 : status == TRANSACTION_STATUS_ABORTED ||
348 : (status == TRANSACTION_STATUS_SUB_COMMITTED && !TransactionIdIsValid(xid)));
349 10602 : Assert(LWLockHeldByMeInMode(CLogControlLock, LW_EXCLUSIVE));
350 :
351 : /*
352 : * If we're doing an async commit (ie, lsn is valid), then we must wait
353 : * for any active write on the page slot to complete. Otherwise our
354 : * update could reach disk in that write, which will not do since we
355 : * mustn't let it reach disk until we've done the appropriate WAL flush.
356 : * But when lsn is invalid, it's OK to scribble on a page while it is
357 : * write-busy, since we don't care if the update reaches disk sooner than
358 : * we think.
359 : */
360 10602 : slotno = SimpleLruReadPage(ClogCtl, pageno, XLogRecPtrIsInvalid(lsn), xid);
361 :
362 : /*
363 : * Set the main transaction id, if any.
364 : *
365 : * If we update more than one xid on this page while it is being written
366 : * out, we might find that some of the bits go to disk and others don't.
367 : * If we are updating commits on the page with the top-level xid that
368 : * could break atomicity, so we subcommit the subxids first before we mark
369 : * the top-level commit.
370 : */
371 10602 : if (TransactionIdIsValid(xid))
372 : {
373 : /* Subtransactions first, if needed ... */
374 10602 : if (status == TRANSACTION_STATUS_COMMITTED)
375 : {
376 9897 : for (i = 0; i < nsubxids; i++)
377 : {
378 19 : Assert(ClogCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
379 19 : TransactionIdSetStatusBit(subxids[i],
380 : TRANSACTION_STATUS_SUB_COMMITTED,
381 : lsn, slotno);
382 : }
383 : }
384 :
385 : /* ... then the main transaction */
386 10602 : TransactionIdSetStatusBit(xid, status, lsn, slotno);
387 : }
388 :
389 : /* Set the subtransactions */
390 10626 : for (i = 0; i < nsubxids; i++)
391 : {
392 24 : Assert(ClogCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
393 24 : TransactionIdSetStatusBit(subxids[i], status, lsn, slotno);
394 : }
395 :
396 10602 : ClogCtl->shared->page_dirty[slotno] = true;
397 10602 : }
398 :
399 : /*
400 : * When we cannot immediately acquire CLogControlLock in exclusive mode at
401 : * commit time, add ourselves to a list of processes that need their XIDs
402 : * status update. The first process to add itself to the list will acquire
403 : * CLogControlLock in exclusive mode and set transaction status as required
404 : * on behalf of all group members. This avoids a great deal of contention
405 : * around CLogControlLock when many processes are trying to commit at once,
406 : * since the lock need not be repeatedly handed off from one committing
407 : * process to the next.
408 : *
409 : * Returns true when transaction status has been updated in clog; returns
410 : * false if we decided against applying the optimization because the page
411 : * number we need to update differs from those processes already waiting.
412 : */
413 : static bool
414 16 : TransactionGroupUpdateXidStatus(TransactionId xid, XidStatus status,
415 : XLogRecPtr lsn, int pageno)
416 : {
417 16 : volatile PROC_HDR *procglobal = ProcGlobal;
418 16 : PGPROC *proc = MyProc;
419 : uint32 nextidx;
420 : uint32 wakeidx;
421 :
422 : /* We should definitely have an XID whose status needs to be updated. */
423 16 : Assert(TransactionIdIsValid(xid));
424 :
425 : /*
426 : * Add ourselves to the list of processes needing a group XID status
427 : * update.
428 : */
429 16 : proc->clogGroupMember = true;
430 16 : proc->clogGroupMemberXid = xid;
431 16 : proc->clogGroupMemberXidStatus = status;
432 16 : proc->clogGroupMemberPage = pageno;
433 16 : proc->clogGroupMemberLsn = lsn;
434 :
435 16 : nextidx = pg_atomic_read_u32(&procglobal->clogGroupFirst);
436 :
437 : while (true)
438 : {
439 : /*
440 : * Add the proc to list, if the clog page where we need to update the
441 : * current transaction status is same as group leader's clog page.
442 : *
443 : * There is a race condition here, which is that after doing the below
444 : * check and before adding this proc's clog update to a group, the
445 : * group leader might have already finished the group update for this
446 : * page and becomes group leader of another group. This will lead to a
447 : * situation where a single group can have different clog page
448 : * updates. This isn't likely and will still work, just maybe a bit
449 : * less efficiently.
450 : */
451 18 : if (nextidx != INVALID_PGPROCNO &&
452 1 : ProcGlobal->allProcs[nextidx].clogGroupMemberPage != proc->clogGroupMemberPage)
453 : {
454 0 : proc->clogGroupMember = false;
455 0 : return false;
456 : }
457 :
458 17 : pg_atomic_write_u32(&proc->clogGroupNext, nextidx);
459 :
460 17 : if (pg_atomic_compare_exchange_u32(&procglobal->clogGroupFirst,
461 : &nextidx,
462 17 : (uint32) proc->pgprocno))
463 16 : break;
464 1 : }
465 :
466 : /*
467 : * If the list was not empty, the leader will update the status of our
468 : * XID. It is impossible to have followers without a leader because the
469 : * first process that has added itself to the list will always have
470 : * nextidx as INVALID_PGPROCNO.
471 : */
472 16 : if (nextidx != INVALID_PGPROCNO)
473 : {
474 1 : int extraWaits = 0;
475 :
476 : /* Sleep until the leader updates our XID status. */
477 1 : pgstat_report_wait_start(WAIT_EVENT_CLOG_GROUP_UPDATE);
478 : for (;;)
479 : {
480 : /* acts as a read barrier */
481 1 : PGSemaphoreLock(proc->sem);
482 1 : if (!proc->clogGroupMember)
483 1 : break;
484 0 : extraWaits++;
485 0 : }
486 1 : pgstat_report_wait_end();
487 :
488 1 : Assert(pg_atomic_read_u32(&proc->clogGroupNext) == INVALID_PGPROCNO);
489 :
490 : /* Fix semaphore count for any absorbed wakeups */
491 2 : while (extraWaits-- > 0)
492 0 : PGSemaphoreUnlock(proc->sem);
493 1 : return true;
494 : }
495 :
496 : /* We are the leader. Acquire the lock on behalf of everyone. */
497 15 : LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
498 :
499 : /*
500 : * Now that we've got the lock, clear the list of processes waiting for
501 : * group XID status update, saving a pointer to the head of the list.
502 : * Trying to pop elements one at a time could lead to an ABA problem.
503 : */
504 15 : nextidx = pg_atomic_exchange_u32(&procglobal->clogGroupFirst,
505 : INVALID_PGPROCNO);
506 :
507 : /* Remember head of list so we can perform wakeups after dropping lock. */
508 15 : wakeidx = nextidx;
509 :
510 : /* Walk the list and update the status of all XIDs. */
511 46 : while (nextidx != INVALID_PGPROCNO)
512 : {
513 16 : PGPROC *proc = &ProcGlobal->allProcs[nextidx];
514 16 : PGXACT *pgxact = &ProcGlobal->allPgXact[nextidx];
515 :
516 : /*
517 : * Overflowed transactions should not use group XID status update
518 : * mechanism.
519 : */
520 16 : Assert(!pgxact->overflowed);
521 :
522 32 : TransactionIdSetPageStatusInternal(proc->clogGroupMemberXid,
523 16 : pgxact->nxids,
524 16 : proc->subxids.xids,
525 : proc->clogGroupMemberXidStatus,
526 : proc->clogGroupMemberLsn,
527 : proc->clogGroupMemberPage);
528 :
529 : /* Move to next proc in list. */
530 16 : nextidx = pg_atomic_read_u32(&proc->clogGroupNext);
531 : }
532 :
533 : /* We're done with the lock now. */
534 15 : LWLockRelease(CLogControlLock);
535 :
536 : /*
537 : * Now that we've released the lock, go back and wake everybody up. We
538 : * don't do this under the lock so as to keep lock hold times to a
539 : * minimum.
540 : */
541 46 : while (wakeidx != INVALID_PGPROCNO)
542 : {
543 16 : PGPROC *proc = &ProcGlobal->allProcs[wakeidx];
544 :
545 16 : wakeidx = pg_atomic_read_u32(&proc->clogGroupNext);
546 16 : pg_atomic_write_u32(&proc->clogGroupNext, INVALID_PGPROCNO);
547 :
548 : /* ensure all previous writes are visible before follower continues. */
549 16 : pg_write_barrier();
550 :
551 16 : proc->clogGroupMember = false;
552 :
553 16 : if (proc != MyProc)
554 1 : PGSemaphoreUnlock(proc->sem);
555 : }
556 :
557 15 : return true;
558 : }
559 :
560 : /*
561 : * Sets the commit status of a single transaction.
562 : *
563 : * Must be called with CLogControlLock held
564 : */
565 : static void
566 10645 : TransactionIdSetStatusBit(TransactionId xid, XidStatus status, XLogRecPtr lsn, int slotno)
567 : {
568 10645 : int byteno = TransactionIdToByte(xid);
569 10645 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
570 : char *byteptr;
571 : char byteval;
572 : char curval;
573 :
574 10645 : byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
575 10645 : curval = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
576 :
577 : /*
578 : * When replaying transactions during recovery we still need to perform
579 : * the two phases of subcommit and then commit. However, some transactions
580 : * are already correctly marked, so we just treat those as a no-op which
581 : * allows us to keep the following Assert as restrictive as possible.
582 : */
583 10645 : if (InRecovery && status == TRANSACTION_STATUS_SUB_COMMITTED &&
584 : curval == TRANSACTION_STATUS_COMMITTED)
585 10645 : return;
586 :
587 : /*
588 : * Current state change should be from 0 or subcommitted to target state
589 : * or we should already be there when replaying changes during recovery.
590 : */
591 10645 : Assert(curval == 0 ||
592 : (curval == TRANSACTION_STATUS_SUB_COMMITTED &&
593 : status != TRANSACTION_STATUS_IN_PROGRESS) ||
594 : curval == status);
595 :
596 : /* note this assumes exclusive access to the clog page */
597 10645 : byteval = *byteptr;
598 10645 : byteval &= ~(((1 << CLOG_BITS_PER_XACT) - 1) << bshift);
599 10645 : byteval |= (status << bshift);
600 10645 : *byteptr = byteval;
601 :
602 : /*
603 : * Update the group LSN if the transaction completion LSN is higher.
604 : *
605 : * Note: lsn will be invalid when supplied during InRecovery processing,
606 : * so we don't need to do anything special to avoid LSN updates during
607 : * recovery. After recovery completes the next clog change will set the
608 : * LSN correctly.
609 : */
610 10645 : if (!XLogRecPtrIsInvalid(lsn))
611 : {
612 406 : int lsnindex = GetLSNIndex(slotno, xid);
613 :
614 406 : if (ClogCtl->shared->group_lsn[lsnindex] < lsn)
615 404 : ClogCtl->shared->group_lsn[lsnindex] = lsn;
616 : }
617 : }
618 :
619 : /*
620 : * Interrogate the state of a transaction in the commit log.
621 : *
622 : * Aside from the actual commit status, this function returns (into *lsn)
623 : * an LSN that is late enough to be able to guarantee that if we flush up to
624 : * that LSN then we will have flushed the transaction's commit record to disk.
625 : * The result is not necessarily the exact LSN of the transaction's commit
626 : * record! For example, for long-past transactions (those whose clog pages
627 : * already migrated to disk), we'll return InvalidXLogRecPtr. Also, because
628 : * we group transactions on the same clog page to conserve storage, we might
629 : * return the LSN of a later transaction that falls into the same group.
630 : *
631 : * NB: this is a low-level routine and is NOT the preferred entry point
632 : * for most uses; TransactionLogFetch() in transam.c is the intended caller.
633 : */
634 : XidStatus
635 30583 : TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
636 : {
637 30583 : int pageno = TransactionIdToPage(xid);
638 30583 : int byteno = TransactionIdToByte(xid);
639 30583 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
640 : int slotno;
641 : int lsnindex;
642 : char *byteptr;
643 : XidStatus status;
644 :
645 : /* lock is acquired by SimpleLruReadPage_ReadOnly */
646 :
647 30583 : slotno = SimpleLruReadPage_ReadOnly(ClogCtl, pageno, xid);
648 30583 : byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
649 :
650 30583 : status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
651 :
652 30583 : lsnindex = GetLSNIndex(slotno, xid);
653 30583 : *lsn = ClogCtl->shared->group_lsn[lsnindex];
654 :
655 30583 : LWLockRelease(CLogControlLock);
656 :
657 30583 : return status;
658 : }
659 :
660 : /*
661 : * Number of shared CLOG buffers.
662 : *
663 : * On larger multi-processor systems, it is possible to have many CLOG page
664 : * requests in flight at one time which could lead to disk access for CLOG
665 : * page if the required page is not found in memory. Testing revealed that we
666 : * can get the best performance by having 128 CLOG buffers, more than that it
667 : * doesn't improve performance.
668 : *
669 : * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
670 : * a good idea, because it would increase the minimum amount of shared memory
671 : * required to start, which could be a problem for people running very small
672 : * configurations. The following formula seems to represent a reasonable
673 : * compromise: people with very low values for shared_buffers will get fewer
674 : * CLOG buffers as well, and everyone else will get 128.
675 : */
676 : Size
677 10 : CLOGShmemBuffers(void)
678 : {
679 10 : return Min(128, Max(4, NBuffers / 512));
680 : }
681 :
682 : /*
683 : * Initialization of shared memory for CLOG
684 : */
685 : Size
686 5 : CLOGShmemSize(void)
687 : {
688 5 : return SimpleLruShmemSize(CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE);
689 : }
690 :
691 : void
692 5 : CLOGShmemInit(void)
693 : {
694 5 : ClogCtl->PagePrecedes = CLOGPagePrecedes;
695 5 : SimpleLruInit(ClogCtl, "clog", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
696 5 : CLogControlLock, "pg_xact", LWTRANCHE_CLOG_BUFFERS);
697 5 : }
698 :
699 : /*
700 : * This func must be called ONCE on system install. It creates
701 : * the initial CLOG segment. (The CLOG directory is assumed to
702 : * have been created by initdb, and CLOGShmemInit must have been
703 : * called already.)
704 : */
705 : void
706 1 : BootStrapCLOG(void)
707 : {
708 : int slotno;
709 :
710 1 : LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
711 :
712 : /* Create and zero the first page of the commit log */
713 1 : slotno = ZeroCLOGPage(0, false);
714 :
715 : /* Make sure it's written out */
716 1 : SimpleLruWritePage(ClogCtl, slotno);
717 1 : Assert(!ClogCtl->shared->page_dirty[slotno]);
718 :
719 1 : LWLockRelease(CLogControlLock);
720 1 : }
721 :
722 : /*
723 : * Initialize (or reinitialize) a page of CLOG to zeroes.
724 : * If writeXlog is TRUE, also emit an XLOG record saying we did this.
725 : *
726 : * The page is not actually written, just set up in shared memory.
727 : * The slot number of the new page is returned.
728 : *
729 : * Control lock must be held at entry, and will be held at exit.
730 : */
731 : static int
732 2 : ZeroCLOGPage(int pageno, bool writeXlog)
733 : {
734 : int slotno;
735 :
736 2 : slotno = SimpleLruZeroPage(ClogCtl, pageno);
737 :
738 2 : if (writeXlog)
739 1 : WriteZeroPageXlogRec(pageno);
740 :
741 2 : return slotno;
742 : }
743 :
744 : /*
745 : * This must be called ONCE during postmaster or standalone-backend startup,
746 : * after StartupXLOG has initialized ShmemVariableCache->nextXid.
747 : */
748 : void
749 3 : StartupCLOG(void)
750 : {
751 3 : TransactionId xid = ShmemVariableCache->nextXid;
752 3 : int pageno = TransactionIdToPage(xid);
753 :
754 3 : LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
755 :
756 : /*
757 : * Initialize our idea of the latest page number.
758 : */
759 3 : ClogCtl->shared->latest_page_number = pageno;
760 :
761 3 : LWLockRelease(CLogControlLock);
762 3 : }
763 :
764 : /*
765 : * This must be called ONCE at the end of startup/recovery.
766 : */
767 : void
768 3 : TrimCLOG(void)
769 : {
770 3 : TransactionId xid = ShmemVariableCache->nextXid;
771 3 : int pageno = TransactionIdToPage(xid);
772 :
773 3 : LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
774 :
775 : /*
776 : * Re-Initialize our idea of the latest page number.
777 : */
778 3 : ClogCtl->shared->latest_page_number = pageno;
779 :
780 : /*
781 : * Zero out the remainder of the current clog page. Under normal
782 : * circumstances it should be zeroes already, but it seems at least
783 : * theoretically possible that XLOG replay will have settled on a nextXID
784 : * value that is less than the last XID actually used and marked by the
785 : * previous database lifecycle (since subtransaction commit writes clog
786 : * but makes no WAL entry). Let's just be safe. (We need not worry about
787 : * pages beyond the current one, since those will be zeroed when first
788 : * used. For the same reason, there is no need to do anything when
789 : * nextXid is exactly at a page boundary; and it's likely that the
790 : * "current" page doesn't exist yet in that case.)
791 : */
792 3 : if (TransactionIdToPgIndex(xid) != 0)
793 : {
794 3 : int byteno = TransactionIdToByte(xid);
795 3 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
796 : int slotno;
797 : char *byteptr;
798 :
799 3 : slotno = SimpleLruReadPage(ClogCtl, pageno, false, xid);
800 3 : byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
801 :
802 : /* Zero so-far-unused positions in the current byte */
803 3 : *byteptr &= (1 << bshift) - 1;
804 : /* Zero the rest of the page */
805 3 : MemSet(byteptr + 1, 0, BLCKSZ - byteno - 1);
806 :
807 3 : ClogCtl->shared->page_dirty[slotno] = true;
808 : }
809 :
810 3 : LWLockRelease(CLogControlLock);
811 3 : }
812 :
813 : /*
814 : * This must be called ONCE during postmaster or standalone-backend shutdown
815 : */
816 : void
817 3 : ShutdownCLOG(void)
818 : {
819 : /* Flush dirty CLOG pages to disk */
820 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(false);
821 3 : SimpleLruFlush(ClogCtl, false);
822 :
823 : /*
824 : * fsync pg_xact to ensure that any files flushed previously are durably
825 : * on disk.
826 : */
827 3 : fsync_fname("pg_xact", true);
828 :
829 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(false);
830 3 : }
831 :
832 : /*
833 : * Perform a checkpoint --- either during shutdown, or on-the-fly
834 : */
835 : void
836 11 : CheckPointCLOG(void)
837 : {
838 : /* Flush dirty CLOG pages to disk */
839 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(true);
840 11 : SimpleLruFlush(ClogCtl, true);
841 :
842 : /*
843 : * fsync pg_xact to ensure that any files flushed previously are durably
844 : * on disk.
845 : */
846 11 : fsync_fname("pg_xact", true);
847 :
848 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(true);
849 11 : }
850 :
851 :
852 : /*
853 : * Make sure that CLOG has room for a newly-allocated XID.
854 : *
855 : * NB: this is called while holding XidGenLock. We want it to be very fast
856 : * most of the time; even when it's not so fast, no actual I/O need happen
857 : * unless we're forced to write out a dirty clog or xlog page to make room
858 : * in shared memory.
859 : */
860 : void
861 10625 : ExtendCLOG(TransactionId newestXact)
862 : {
863 : int pageno;
864 :
865 : /*
866 : * No work except at first XID of a page. But beware: just after
867 : * wraparound, the first XID of page zero is FirstNormalTransactionId.
868 : */
869 10625 : if (TransactionIdToPgIndex(newestXact) != 0 &&
870 : !TransactionIdEquals(newestXact, FirstNormalTransactionId))
871 21249 : return;
872 :
873 1 : pageno = TransactionIdToPage(newestXact);
874 :
875 1 : LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
876 :
877 : /* Zero the page and make an XLOG entry about it */
878 1 : ZeroCLOGPage(pageno, true);
879 :
880 1 : LWLockRelease(CLogControlLock);
881 : }
882 :
883 :
884 : /*
885 : * Remove all CLOG segments before the one holding the passed transaction ID
886 : *
887 : * Before removing any CLOG data, we must flush XLOG to disk, to ensure
888 : * that any recently-emitted HEAP_FREEZE records have reached disk; otherwise
889 : * a crash and restart might leave us with some unfrozen tuples referencing
890 : * removed CLOG data. We choose to emit a special TRUNCATE XLOG record too.
891 : * Replaying the deletion from XLOG is not critical, since the files could
892 : * just as well be removed later, but doing so prevents a long-running hot
893 : * standby server from acquiring an unreasonably bloated CLOG directory.
894 : *
895 : * Since CLOG segments hold a large number of transactions, the opportunity to
896 : * actually remove a segment is fairly rare, and so it seems best not to do
897 : * the XLOG flush unless we have confirmed that there is a removable segment.
898 : */
899 : void
900 2 : TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
901 : {
902 : int cutoffPage;
903 :
904 : /*
905 : * The cutoff point is the start of the segment containing oldestXact. We
906 : * pass the *page* containing oldestXact to SimpleLruTruncate.
907 : */
908 2 : cutoffPage = TransactionIdToPage(oldestXact);
909 :
910 : /* Check to see if there's any files that could be removed */
911 2 : if (!SlruScanDirectory(ClogCtl, SlruScanDirCbReportPresence, &cutoffPage))
912 4 : return; /* nothing to remove */
913 :
914 : /*
915 : * Advance oldestClogXid before truncating clog, so concurrent xact status
916 : * lookups can ensure they don't attempt to access truncated-away clog.
917 : *
918 : * It's only necessary to do this if we will actually truncate away clog
919 : * pages.
920 : */
921 0 : AdvanceOldestClogXid(oldestXact);
922 :
923 : /*
924 : * Write XLOG record and flush XLOG to disk. We record the oldest xid
925 : * we're keeping information about here so we can ensure that it's always
926 : * ahead of clog truncation in case we crash, and so a standby finds out
927 : * the new valid xid before the next checkpoint.
928 : */
929 0 : WriteTruncateXlogRec(cutoffPage, oldestXact, oldestxid_datoid);
930 :
931 : /* Now we can remove the old CLOG segment(s) */
932 0 : SimpleLruTruncate(ClogCtl, cutoffPage);
933 : }
934 :
935 :
936 : /*
937 : * Decide which of two CLOG page numbers is "older" for truncation purposes.
938 : *
939 : * We need to use comparison of TransactionIds here in order to do the right
940 : * thing with wraparound XID arithmetic. However, if we are asked about
941 : * page number zero, we don't want to hand InvalidTransactionId to
942 : * TransactionIdPrecedes: it'll get weird about permanent xact IDs. So,
943 : * offset both xids by FirstNormalTransactionId to avoid that.
944 : */
945 : static bool
946 2 : CLOGPagePrecedes(int page1, int page2)
947 : {
948 : TransactionId xid1;
949 : TransactionId xid2;
950 :
951 2 : xid1 = ((TransactionId) page1) * CLOG_XACTS_PER_PAGE;
952 2 : xid1 += FirstNormalTransactionId;
953 2 : xid2 = ((TransactionId) page2) * CLOG_XACTS_PER_PAGE;
954 2 : xid2 += FirstNormalTransactionId;
955 :
956 2 : return TransactionIdPrecedes(xid1, xid2);
957 : }
958 :
959 :
960 : /*
961 : * Write a ZEROPAGE xlog record
962 : */
963 : static void
964 1 : WriteZeroPageXlogRec(int pageno)
965 : {
966 1 : XLogBeginInsert();
967 1 : XLogRegisterData((char *) (&pageno), sizeof(int));
968 1 : (void) XLogInsert(RM_CLOG_ID, CLOG_ZEROPAGE);
969 1 : }
970 :
971 : /*
972 : * Write a TRUNCATE xlog record
973 : *
974 : * We must flush the xlog record to disk before returning --- see notes
975 : * in TruncateCLOG().
976 : */
977 : static void
978 0 : WriteTruncateXlogRec(int pageno, TransactionId oldestXact, Oid oldestXactDb)
979 : {
980 : XLogRecPtr recptr;
981 : xl_clog_truncate xlrec;
982 :
983 0 : xlrec.pageno = pageno;
984 0 : xlrec.oldestXact = oldestXact;
985 0 : xlrec.oldestXactDb = oldestXactDb;
986 :
987 0 : XLogBeginInsert();
988 0 : XLogRegisterData((char *) (&xlrec), sizeof(xl_clog_truncate));
989 0 : recptr = XLogInsert(RM_CLOG_ID, CLOG_TRUNCATE);
990 0 : XLogFlush(recptr);
991 0 : }
992 :
993 : /*
994 : * CLOG resource manager's routines
995 : */
996 : void
997 0 : clog_redo(XLogReaderState *record)
998 : {
999 0 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
1000 :
1001 : /* Backup blocks are not used in clog records */
1002 0 : Assert(!XLogRecHasAnyBlockRefs(record));
1003 :
1004 0 : if (info == CLOG_ZEROPAGE)
1005 : {
1006 : int pageno;
1007 : int slotno;
1008 :
1009 0 : memcpy(&pageno, XLogRecGetData(record), sizeof(int));
1010 :
1011 0 : LWLockAcquire(CLogControlLock, LW_EXCLUSIVE);
1012 :
1013 0 : slotno = ZeroCLOGPage(pageno, false);
1014 0 : SimpleLruWritePage(ClogCtl, slotno);
1015 0 : Assert(!ClogCtl->shared->page_dirty[slotno]);
1016 :
1017 0 : LWLockRelease(CLogControlLock);
1018 : }
1019 0 : else if (info == CLOG_TRUNCATE)
1020 : {
1021 : xl_clog_truncate xlrec;
1022 :
1023 0 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_clog_truncate));
1024 :
1025 : /*
1026 : * During XLOG replay, latest_page_number isn't set up yet; insert a
1027 : * suitable value to bypass the sanity test in SimpleLruTruncate.
1028 : */
1029 0 : ClogCtl->shared->latest_page_number = xlrec.pageno;
1030 :
1031 0 : AdvanceOldestClogXid(xlrec.oldestXact);
1032 :
1033 0 : SimpleLruTruncate(ClogCtl, xlrec.pageno);
1034 : }
1035 : else
1036 0 : elog(PANIC, "clog_redo: unknown op code %u", info);
1037 0 : }
|
