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) 1996-2020, 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 : #include "pgstat.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 XactCtlData;
87 :
88 : #define XactCtl (&XactCtlData)
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 oldestXactDb);
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 POSIX_FADV_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 278168 : TransactionIdSetTreeStatus(TransactionId xid, int nsubxids,
164 : TransactionId *subxids, XidStatus status, XLogRecPtr lsn)
165 : {
166 278168 : 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 284664 : for (i = 0; i < nsubxids; i++)
177 : {
178 6496 : if (TransactionIdToPage(subxids[i]) != pageno)
179 0 : break;
180 : }
181 :
182 : /*
183 : * Do all items fit on a single page?
184 : */
185 278168 : if (i == nsubxids)
186 : {
187 : /*
188 : * Set the parent and all subtransactions in a single call
189 : */
190 278168 : 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 278168 : }
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 : Assert(nsubxids > 0); /* else the pageno fetch above is unsafe */
245 :
246 0 : while (i < nsubxids)
247 : {
248 0 : int num_on_page = 0;
249 : int nextpageno;
250 :
251 : do
252 : {
253 0 : nextpageno = TransactionIdToPage(subxids[i]);
254 0 : if (nextpageno != pageno)
255 0 : break;
256 0 : num_on_page++;
257 0 : i++;
258 0 : } while (i < nsubxids);
259 :
260 0 : TransactionIdSetPageStatus(InvalidTransactionId,
261 0 : num_on_page, subxids + offset,
262 : status, lsn, pageno, false);
263 0 : offset = i;
264 0 : pageno = nextpageno;
265 : }
266 0 : }
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
273 278168 : 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 XactSLRULock, 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 XactSLRULock
286 : * 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 278168 : if (all_xact_same_page && xid == MyPgXact->xid &&
297 273312 : nsubxids <= THRESHOLD_SUBTRANS_CLOG_OPT &&
298 273312 : nsubxids == MyPgXact->nxids &&
299 273312 : memcmp(subxids, MyProc->subxids.xids,
300 : nsubxids * sizeof(TransactionId)) == 0)
301 : {
302 : /*
303 : * If we can immediately acquire XactSLRULock, we update the status of
304 : * 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 273312 : if (LWLockConditionalAcquire(XactSLRULock, LW_EXCLUSIVE))
309 : {
310 : /* Got the lock without waiting! Do the update. */
311 273274 : TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
312 : lsn, pageno);
313 273274 : LWLockRelease(XactSLRULock);
314 273274 : return;
315 : }
316 38 : else if (TransactionGroupUpdateXidStatus(xid, status, lsn, pageno))
317 : {
318 : /* Group update mechanism has done the work. */
319 38 : 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 4856 : LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
327 4856 : TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
328 : lsn, pageno);
329 4856 : LWLockRelease(XactSLRULock);
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
338 278168 : 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(XactSLRULock, 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 278168 : slotno = SimpleLruReadPage(XactCtl, 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 278168 : if (TransactionIdIsValid(xid))
371 : {
372 : /* Subtransactions first, if needed ... */
373 278168 : if (status == TRANSACTION_STATUS_COMMITTED)
374 : {
375 278276 : for (i = 0; i < nsubxids; i++)
376 : {
377 : Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
378 5908 : TransactionIdSetStatusBit(subxids[i],
379 : TRANSACTION_STATUS_SUB_COMMITTED,
380 : lsn, slotno);
381 : }
382 : }
383 :
384 : /* ... then the main transaction */
385 278168 : TransactionIdSetStatusBit(xid, status, lsn, slotno);
386 : }
387 :
388 : /* Set the subtransactions */
389 284664 : for (i = 0; i < nsubxids; i++)
390 : {
391 : Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
392 6496 : TransactionIdSetStatusBit(subxids[i], status, lsn, slotno);
393 : }
394 :
395 278168 : XactCtl->shared->page_dirty[slotno] = true;
396 278168 : }
397 :
398 : /*
399 : * When we cannot immediately acquire XactSLRULock 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 : * XactSLRULock 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 XactSLRULock 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
413 38 : TransactionGroupUpdateXidStatus(TransactionId xid, XidStatus status,
414 : XLogRecPtr lsn, int pageno)
415 : {
416 38 : volatile PROC_HDR *procglobal = ProcGlobal;
417 38 : 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 38 : proc->clogGroupMember = true;
429 38 : proc->clogGroupMemberXid = xid;
430 38 : proc->clogGroupMemberXidStatus = status;
431 38 : proc->clogGroupMemberPage = pageno;
432 38 : proc->clogGroupMemberLsn = lsn;
433 :
434 38 : 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 38 : if (nextidx != INVALID_PGPROCNO &&
451 2 : ProcGlobal->allProcs[nextidx].clogGroupMemberPage != proc->clogGroupMemberPage)
452 : {
453 0 : proc->clogGroupMember = false;
454 0 : return false;
455 : }
456 :
457 38 : pg_atomic_write_u32(&proc->clogGroupNext, nextidx);
458 :
459 38 : if (pg_atomic_compare_exchange_u32(&procglobal->clogGroupFirst,
460 : &nextidx,
461 38 : (uint32) proc->pgprocno))
462 38 : break;
463 : }
464 :
465 : /*
466 : * If the list was not empty, the leader will update the status of our
467 : * XID. It is impossible to have followers without a leader because the
468 : * first process that has added itself to the list will always have
469 : * nextidx as INVALID_PGPROCNO.
470 : */
471 38 : if (nextidx != INVALID_PGPROCNO)
472 : {
473 2 : int extraWaits = 0;
474 :
475 : /* Sleep until the leader updates our XID status. */
476 2 : pgstat_report_wait_start(WAIT_EVENT_XACT_GROUP_UPDATE);
477 : for (;;)
478 : {
479 : /* acts as a read barrier */
480 2 : PGSemaphoreLock(proc->sem);
481 2 : if (!proc->clogGroupMember)
482 2 : break;
483 0 : extraWaits++;
484 : }
485 2 : pgstat_report_wait_end();
486 :
487 : Assert(pg_atomic_read_u32(&proc->clogGroupNext) == INVALID_PGPROCNO);
488 :
489 : /* Fix semaphore count for any absorbed wakeups */
490 2 : while (extraWaits-- > 0)
491 0 : PGSemaphoreUnlock(proc->sem);
492 2 : return true;
493 : }
494 :
495 : /* We are the leader. Acquire the lock on behalf of everyone. */
496 36 : LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
497 :
498 : /*
499 : * Now that we've got the lock, clear the list of processes waiting for
500 : * group XID status update, saving a pointer to the head of the list.
501 : * Trying to pop elements one at a time could lead to an ABA problem.
502 : */
503 36 : nextidx = pg_atomic_exchange_u32(&procglobal->clogGroupFirst,
504 : INVALID_PGPROCNO);
505 :
506 : /* Remember head of list so we can perform wakeups after dropping lock. */
507 36 : wakeidx = nextidx;
508 :
509 : /* Walk the list and update the status of all XIDs. */
510 74 : while (nextidx != INVALID_PGPROCNO)
511 : {
512 38 : PGPROC *proc = &ProcGlobal->allProcs[nextidx];
513 38 : PGXACT *pgxact = &ProcGlobal->allPgXact[nextidx];
514 :
515 : /*
516 : * Transactions with more than THRESHOLD_SUBTRANS_CLOG_OPT sub-XIDs
517 : * should not use group XID status update mechanism.
518 : */
519 : Assert(pgxact->nxids <= THRESHOLD_SUBTRANS_CLOG_OPT);
520 :
521 76 : TransactionIdSetPageStatusInternal(proc->clogGroupMemberXid,
522 38 : pgxact->nxids,
523 38 : proc->subxids.xids,
524 : proc->clogGroupMemberXidStatus,
525 : proc->clogGroupMemberLsn,
526 : proc->clogGroupMemberPage);
527 :
528 : /* Move to next proc in list. */
529 38 : nextidx = pg_atomic_read_u32(&proc->clogGroupNext);
530 : }
531 :
532 : /* We're done with the lock now. */
533 36 : LWLockRelease(XactSLRULock);
534 :
535 : /*
536 : * Now that we've released the lock, go back and wake everybody up. We
537 : * don't do this under the lock so as to keep lock hold times to a
538 : * minimum.
539 : */
540 74 : while (wakeidx != INVALID_PGPROCNO)
541 : {
542 38 : PGPROC *proc = &ProcGlobal->allProcs[wakeidx];
543 :
544 38 : wakeidx = pg_atomic_read_u32(&proc->clogGroupNext);
545 38 : pg_atomic_write_u32(&proc->clogGroupNext, INVALID_PGPROCNO);
546 :
547 : /* ensure all previous writes are visible before follower continues. */
548 38 : pg_write_barrier();
549 :
550 38 : proc->clogGroupMember = false;
551 :
552 38 : if (proc != MyProc)
553 2 : PGSemaphoreUnlock(proc->sem);
554 : }
555 :
556 36 : return true;
557 : }
558 :
559 : /*
560 : * Sets the commit status of a single transaction.
561 : *
562 : * Must be called with XactSLRULock held
563 : */
564 : static void
565 290572 : TransactionIdSetStatusBit(TransactionId xid, XidStatus status, XLogRecPtr lsn, int slotno)
566 : {
567 290572 : int byteno = TransactionIdToByte(xid);
568 290572 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
569 : char *byteptr;
570 : char byteval;
571 : char curval;
572 :
573 290572 : byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
574 290572 : curval = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
575 :
576 : /*
577 : * When replaying transactions during recovery we still need to perform
578 : * the two phases of subcommit and then commit. However, some transactions
579 : * are already correctly marked, so we just treat those as a no-op which
580 : * allows us to keep the following Assert as restrictive as possible.
581 : */
582 290572 : if (InRecovery && status == TRANSACTION_STATUS_SUB_COMMITTED &&
583 : curval == TRANSACTION_STATUS_COMMITTED)
584 0 : return;
585 :
586 : /*
587 : * Current state change should be from 0 or subcommitted to target state
588 : * or we should already be there when replaying changes during recovery.
589 : */
590 : Assert(curval == 0 ||
591 : (curval == TRANSACTION_STATUS_SUB_COMMITTED &&
592 : status != TRANSACTION_STATUS_IN_PROGRESS) ||
593 : curval == status);
594 :
595 : /* note this assumes exclusive access to the clog page */
596 290572 : byteval = *byteptr;
597 290572 : byteval &= ~(((1 << CLOG_BITS_PER_XACT) - 1) << bshift);
598 290572 : byteval |= (status << bshift);
599 290572 : *byteptr = byteval;
600 :
601 : /*
602 : * Update the group LSN if the transaction completion LSN is higher.
603 : *
604 : * Note: lsn will be invalid when supplied during InRecovery processing,
605 : * so we don't need to do anything special to avoid LSN updates during
606 : * recovery. After recovery completes the next clog change will set the
607 : * LSN correctly.
608 : */
609 290572 : if (!XLogRecPtrIsInvalid(lsn))
610 : {
611 16522 : int lsnindex = GetLSNIndex(slotno, xid);
612 :
613 16522 : if (XactCtl->shared->group_lsn[lsnindex] < lsn)
614 11486 : XactCtl->shared->group_lsn[lsnindex] = lsn;
615 : }
616 : }
617 :
618 : /*
619 : * Interrogate the state of a transaction in the commit log.
620 : *
621 : * Aside from the actual commit status, this function returns (into *lsn)
622 : * an LSN that is late enough to be able to guarantee that if we flush up to
623 : * that LSN then we will have flushed the transaction's commit record to disk.
624 : * The result is not necessarily the exact LSN of the transaction's commit
625 : * record! For example, for long-past transactions (those whose clog pages
626 : * already migrated to disk), we'll return InvalidXLogRecPtr. Also, because
627 : * we group transactions on the same clog page to conserve storage, we might
628 : * return the LSN of a later transaction that falls into the same group.
629 : *
630 : * NB: this is a low-level routine and is NOT the preferred entry point
631 : * for most uses; TransactionLogFetch() in transam.c is the intended caller.
632 : */
633 : XidStatus
634 1186080 : TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
635 : {
636 1186080 : int pageno = TransactionIdToPage(xid);
637 1186080 : int byteno = TransactionIdToByte(xid);
638 1186080 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
639 : int slotno;
640 : int lsnindex;
641 : char *byteptr;
642 : XidStatus status;
643 :
644 : /* lock is acquired by SimpleLruReadPage_ReadOnly */
645 :
646 1186080 : slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
647 1186080 : byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
648 :
649 1186080 : status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
650 :
651 1186080 : lsnindex = GetLSNIndex(slotno, xid);
652 1186080 : *lsn = XactCtl->shared->group_lsn[lsnindex];
653 :
654 1186080 : LWLockRelease(XactSLRULock);
655 :
656 1186080 : return status;
657 : }
658 :
659 : /*
660 : * Number of shared CLOG buffers.
661 : *
662 : * On larger multi-processor systems, it is possible to have many CLOG page
663 : * requests in flight at one time which could lead to disk access for CLOG
664 : * page if the required page is not found in memory. Testing revealed that we
665 : * can get the best performance by having 128 CLOG buffers, more than that it
666 : * doesn't improve performance.
667 : *
668 : * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
669 : * a good idea, because it would increase the minimum amount of shared memory
670 : * required to start, which could be a problem for people running very small
671 : * configurations. The following formula seems to represent a reasonable
672 : * compromise: people with very low values for shared_buffers will get fewer
673 : * CLOG buffers as well, and everyone else will get 128.
674 : */
675 : Size
676 4344 : CLOGShmemBuffers(void)
677 : {
678 4344 : return Min(128, Max(4, NBuffers / 512));
679 : }
680 :
681 : /*
682 : * Initialization of shared memory for CLOG
683 : */
684 : Size
685 2174 : CLOGShmemSize(void)
686 : {
687 2174 : return SimpleLruShmemSize(CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE);
688 : }
689 :
690 : void
691 2170 : CLOGShmemInit(void)
692 : {
693 2170 : XactCtl->PagePrecedes = CLOGPagePrecedes;
694 2170 : SimpleLruInit(XactCtl, "Xact", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
695 2170 : XactSLRULock, "pg_xact", LWTRANCHE_XACT_BUFFER);
696 2170 : }
697 :
698 : /*
699 : * This func must be called ONCE on system install. It creates
700 : * the initial CLOG segment. (The CLOG directory is assumed to
701 : * have been created by initdb, and CLOGShmemInit must have been
702 : * called already.)
703 : */
704 : void
705 358 : BootStrapCLOG(void)
706 : {
707 : int slotno;
708 :
709 358 : LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
710 :
711 : /* Create and zero the first page of the commit log */
712 358 : slotno = ZeroCLOGPage(0, false);
713 :
714 : /* Make sure it's written out */
715 358 : SimpleLruWritePage(XactCtl, slotno);
716 : Assert(!XactCtl->shared->page_dirty[slotno]);
717 :
718 358 : LWLockRelease(XactSLRULock);
719 358 : }
720 :
721 : /*
722 : * Initialize (or reinitialize) a page of CLOG to zeroes.
723 : * If writeXlog is true, also emit an XLOG record saying we did this.
724 : *
725 : * The page is not actually written, just set up in shared memory.
726 : * The slot number of the new page is returned.
727 : *
728 : * Control lock must be held at entry, and will be held at exit.
729 : */
730 : static int
731 716 : ZeroCLOGPage(int pageno, bool writeXlog)
732 : {
733 : int slotno;
734 :
735 716 : slotno = SimpleLruZeroPage(XactCtl, pageno);
736 :
737 716 : if (writeXlog)
738 358 : WriteZeroPageXlogRec(pageno);
739 :
740 716 : return slotno;
741 : }
742 :
743 : /*
744 : * This must be called ONCE during postmaster or standalone-backend startup,
745 : * after StartupXLOG has initialized ShmemVariableCache->nextFullXid.
746 : */
747 : void
748 1390 : StartupCLOG(void)
749 : {
750 1390 : TransactionId xid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
751 1390 : int pageno = TransactionIdToPage(xid);
752 :
753 1390 : LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
754 :
755 : /*
756 : * Initialize our idea of the latest page number.
757 : */
758 1390 : XactCtl->shared->latest_page_number = pageno;
759 :
760 1390 : LWLockRelease(XactSLRULock);
761 1390 : }
762 :
763 : /*
764 : * This must be called ONCE at the end of startup/recovery.
765 : */
766 : void
767 1356 : TrimCLOG(void)
768 : {
769 1356 : TransactionId xid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
770 1356 : int pageno = TransactionIdToPage(xid);
771 :
772 1356 : LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
773 :
774 : /*
775 : * Re-Initialize our idea of the latest page number.
776 : */
777 1356 : XactCtl->shared->latest_page_number = pageno;
778 :
779 : /*
780 : * Zero out the remainder of the current clog page. Under normal
781 : * circumstances it should be zeroes already, but it seems at least
782 : * theoretically possible that XLOG replay will have settled on a nextXID
783 : * value that is less than the last XID actually used and marked by the
784 : * previous database lifecycle (since subtransaction commit writes clog
785 : * but makes no WAL entry). Let's just be safe. (We need not worry about
786 : * pages beyond the current one, since those will be zeroed when first
787 : * used. For the same reason, there is no need to do anything when
788 : * nextFullXid is exactly at a page boundary; and it's likely that the
789 : * "current" page doesn't exist yet in that case.)
790 : */
791 1356 : if (TransactionIdToPgIndex(xid) != 0)
792 : {
793 1356 : int byteno = TransactionIdToByte(xid);
794 1356 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
795 : int slotno;
796 : char *byteptr;
797 :
798 1356 : slotno = SimpleLruReadPage(XactCtl, pageno, false, xid);
799 1356 : byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
800 :
801 : /* Zero so-far-unused positions in the current byte */
802 1356 : *byteptr &= (1 << bshift) - 1;
803 : /* Zero the rest of the page */
804 1356 : MemSet(byteptr + 1, 0, BLCKSZ - byteno - 1);
805 :
806 1356 : XactCtl->shared->page_dirty[slotno] = true;
807 : }
808 :
809 1356 : LWLockRelease(XactSLRULock);
810 1356 : }
811 :
812 : /*
813 : * This must be called ONCE during postmaster or standalone-backend shutdown
814 : */
815 : void
816 1126 : ShutdownCLOG(void)
817 : {
818 : /* Flush dirty CLOG pages to disk */
819 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(false);
820 1126 : SimpleLruFlush(XactCtl, false);
821 :
822 : /*
823 : * fsync pg_xact to ensure that any files flushed previously are durably
824 : * on disk.
825 : */
826 1126 : fsync_fname("pg_xact", true);
827 :
828 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(false);
829 1126 : }
830 :
831 : /*
832 : * Perform a checkpoint --- either during shutdown, or on-the-fly
833 : */
834 : void
835 3172 : CheckPointCLOG(void)
836 : {
837 : /* Flush dirty CLOG pages to disk */
838 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(true);
839 3172 : SimpleLruFlush(XactCtl, true);
840 :
841 : /*
842 : * fsync pg_xact to ensure that any files flushed previously are durably
843 : * on disk.
844 : */
845 3172 : fsync_fname("pg_xact", true);
846 :
847 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(true);
848 3172 : }
849 :
850 :
851 : /*
852 : * Make sure that CLOG has room for a newly-allocated XID.
853 : *
854 : * NB: this is called while holding XidGenLock. We want it to be very fast
855 : * most of the time; even when it's not so fast, no actual I/O need happen
856 : * unless we're forced to write out a dirty clog or xlog page to make room
857 : * in shared memory.
858 : */
859 : void
860 277310 : ExtendCLOG(TransactionId newestXact)
861 : {
862 : int pageno;
863 :
864 : /*
865 : * No work except at first XID of a page. But beware: just after
866 : * wraparound, the first XID of page zero is FirstNormalTransactionId.
867 : */
868 277310 : if (TransactionIdToPgIndex(newestXact) != 0 &&
869 : !TransactionIdEquals(newestXact, FirstNormalTransactionId))
870 276952 : return;
871 :
872 358 : pageno = TransactionIdToPage(newestXact);
873 :
874 358 : LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
875 :
876 : /* Zero the page and make an XLOG entry about it */
877 358 : ZeroCLOGPage(pageno, true);
878 :
879 358 : LWLockRelease(XactSLRULock);
880 : }
881 :
882 :
883 : /*
884 : * Remove all CLOG segments before the one holding the passed transaction ID
885 : *
886 : * Before removing any CLOG data, we must flush XLOG to disk, to ensure
887 : * that any recently-emitted FREEZE_PAGE records have reached disk; otherwise
888 : * a crash and restart might leave us with some unfrozen tuples referencing
889 : * removed CLOG data. We choose to emit a special TRUNCATE XLOG record too.
890 : * Replaying the deletion from XLOG is not critical, since the files could
891 : * just as well be removed later, but doing so prevents a long-running hot
892 : * standby server from acquiring an unreasonably bloated CLOG directory.
893 : *
894 : * Since CLOG segments hold a large number of transactions, the opportunity to
895 : * actually remove a segment is fairly rare, and so it seems best not to do
896 : * the XLOG flush unless we have confirmed that there is a removable segment.
897 : */
898 : void
899 782 : TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
900 : {
901 : int cutoffPage;
902 :
903 : /*
904 : * The cutoff point is the start of the segment containing oldestXact. We
905 : * pass the *page* containing oldestXact to SimpleLruTruncate.
906 : */
907 782 : cutoffPage = TransactionIdToPage(oldestXact);
908 :
909 : /* Check to see if there's any files that could be removed */
910 782 : if (!SlruScanDirectory(XactCtl, SlruScanDirCbReportPresence, &cutoffPage))
911 782 : return; /* nothing to remove */
912 :
913 : /*
914 : * Advance oldestClogXid before truncating clog, so concurrent xact status
915 : * lookups can ensure they don't attempt to access truncated-away clog.
916 : *
917 : * It's only necessary to do this if we will actually truncate away clog
918 : * pages.
919 : */
920 0 : AdvanceOldestClogXid(oldestXact);
921 :
922 : /*
923 : * Write XLOG record and flush XLOG to disk. We record the oldest xid
924 : * we're keeping information about here so we can ensure that it's always
925 : * ahead of clog truncation in case we crash, and so a standby finds out
926 : * the new valid xid before the next checkpoint.
927 : */
928 0 : WriteTruncateXlogRec(cutoffPage, oldestXact, oldestxid_datoid);
929 :
930 : /* Now we can remove the old CLOG segment(s) */
931 0 : SimpleLruTruncate(XactCtl, cutoffPage);
932 : }
933 :
934 :
935 : /*
936 : * Decide which of two CLOG page numbers is "older" for truncation purposes.
937 : *
938 : * We need to use comparison of TransactionIds here in order to do the right
939 : * thing with wraparound XID arithmetic. However, if we are asked about
940 : * page number zero, we don't want to hand InvalidTransactionId to
941 : * TransactionIdPrecedes: it'll get weird about permanent xact IDs. So,
942 : * offset both xids by FirstNormalTransactionId to avoid that.
943 : */
944 : static bool
945 782 : CLOGPagePrecedes(int page1, int page2)
946 : {
947 : TransactionId xid1;
948 : TransactionId xid2;
949 :
950 782 : xid1 = ((TransactionId) page1) * CLOG_XACTS_PER_PAGE;
951 782 : xid1 += FirstNormalTransactionId;
952 782 : xid2 = ((TransactionId) page2) * CLOG_XACTS_PER_PAGE;
953 782 : xid2 += FirstNormalTransactionId;
954 :
955 782 : return TransactionIdPrecedes(xid1, xid2);
956 : }
957 :
958 :
959 : /*
960 : * Write a ZEROPAGE xlog record
961 : */
962 : static void
963 358 : WriteZeroPageXlogRec(int pageno)
964 : {
965 358 : XLogBeginInsert();
966 358 : XLogRegisterData((char *) (&pageno), sizeof(int));
967 358 : (void) XLogInsert(RM_CLOG_ID, CLOG_ZEROPAGE);
968 358 : }
969 :
970 : /*
971 : * Write a TRUNCATE xlog record
972 : *
973 : * We must flush the xlog record to disk before returning --- see notes
974 : * in TruncateCLOG().
975 : */
976 : static void
977 0 : WriteTruncateXlogRec(int pageno, TransactionId oldestXact, Oid oldestXactDb)
978 : {
979 : XLogRecPtr recptr;
980 : xl_clog_truncate xlrec;
981 :
982 0 : xlrec.pageno = pageno;
983 0 : xlrec.oldestXact = oldestXact;
984 0 : xlrec.oldestXactDb = oldestXactDb;
985 :
986 0 : XLogBeginInsert();
987 0 : XLogRegisterData((char *) (&xlrec), sizeof(xl_clog_truncate));
988 0 : recptr = XLogInsert(RM_CLOG_ID, CLOG_TRUNCATE);
989 0 : XLogFlush(recptr);
990 0 : }
991 :
992 : /*
993 : * CLOG resource manager's routines
994 : */
995 : void
996 0 : clog_redo(XLogReaderState *record)
997 : {
998 0 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
999 :
1000 : /* Backup blocks are not used in clog records */
1001 : Assert(!XLogRecHasAnyBlockRefs(record));
1002 :
1003 0 : if (info == CLOG_ZEROPAGE)
1004 : {
1005 : int pageno;
1006 : int slotno;
1007 :
1008 0 : memcpy(&pageno, XLogRecGetData(record), sizeof(int));
1009 :
1010 0 : LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
1011 :
1012 0 : slotno = ZeroCLOGPage(pageno, false);
1013 0 : SimpleLruWritePage(XactCtl, slotno);
1014 : Assert(!XactCtl->shared->page_dirty[slotno]);
1015 :
1016 0 : LWLockRelease(XactSLRULock);
1017 : }
1018 0 : else if (info == CLOG_TRUNCATE)
1019 : {
1020 : xl_clog_truncate xlrec;
1021 :
1022 0 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_clog_truncate));
1023 :
1024 : /*
1025 : * During XLOG replay, latest_page_number isn't set up yet; insert a
1026 : * suitable value to bypass the sanity test in SimpleLruTruncate.
1027 : */
1028 0 : XactCtl->shared->latest_page_number = xlrec.pageno;
1029 :
1030 0 : AdvanceOldestClogXid(xlrec.oldestXact);
1031 :
1032 0 : SimpleLruTruncate(XactCtl, xlrec.pageno);
1033 : }
1034 : else
1035 0 : elog(PANIC, "clog_redo: unknown op code %u", info);
1036 0 : }
|
