Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * clog.c
4 : * PostgreSQL transaction-commit-log manager
5 : *
6 : * This module stores two bits per transaction regarding its commit/abort
7 : * status; the status for four transactions fit in a byte.
8 : *
9 : * This would be a pretty simple abstraction on top of slru.c, except that
10 : * for performance reasons we allow multiple transactions that are
11 : * committing concurrently to form a queue, so that a single process can
12 : * update the status for all of them within a single lock acquisition run.
13 : *
14 : * XLOG interactions: this module generates an XLOG record whenever a new
15 : * CLOG page is initialized to zeroes. Other writes of CLOG come from
16 : * recording of transaction commit or abort in xact.c, which generates its
17 : * own XLOG records for these events and will re-perform the status update
18 : * on redo; so we need make no additional XLOG entry here. For synchronous
19 : * transaction commits, the XLOG is guaranteed flushed through the XLOG commit
20 : * record before we are called to log a commit, so the WAL rule "write xlog
21 : * before data" is satisfied automatically. However, for async commits we
22 : * must track the latest LSN affecting each CLOG page, so that we can flush
23 : * XLOG that far and satisfy the WAL rule. We don't have to worry about this
24 : * for aborts (whether sync or async), since the post-crash assumption would
25 : * be that such transactions failed anyway.
26 : *
27 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
28 : * Portions Copyright (c) 1994, Regents of the University of California
29 : *
30 : * src/backend/access/transam/clog.c
31 : *
32 : *-------------------------------------------------------------------------
33 : */
34 : #include "postgres.h"
35 :
36 : #include "access/clog.h"
37 : #include "access/slru.h"
38 : #include "access/transam.h"
39 : #include "access/xlog.h"
40 : #include "access/xloginsert.h"
41 : #include "access/xlogutils.h"
42 : #include "miscadmin.h"
43 : #include "pg_trace.h"
44 : #include "pgstat.h"
45 : #include "storage/proc.h"
46 : #include "storage/sync.h"
47 : #include "utils/guc_hooks.h"
48 : #include "utils/wait_event.h"
49 :
50 : /*
51 : * Defines for CLOG page sizes. A page is the same BLCKSZ as is used
52 : * everywhere else in Postgres.
53 : *
54 : * Note: because TransactionIds are 32 bits and wrap around at 0xFFFFFFFF,
55 : * CLOG page numbering also wraps around at 0xFFFFFFFF/CLOG_XACTS_PER_PAGE,
56 : * and CLOG segment numbering at
57 : * 0xFFFFFFFF/CLOG_XACTS_PER_PAGE/SLRU_PAGES_PER_SEGMENT. We need take no
58 : * explicit notice of that fact in this module, except when comparing segment
59 : * and page numbers in TruncateCLOG (see CLOGPagePrecedes).
60 : */
61 :
62 : /* We need two bits per xact, so four xacts fit in a byte */
63 : #define CLOG_BITS_PER_XACT 2
64 : #define CLOG_XACTS_PER_BYTE 4
65 : #define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
66 : #define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
67 :
68 : /*
69 : * Because space used in CLOG by each transaction is so small, we place a
70 : * smaller limit on the number of CLOG buffers than SLRU allows. No other
71 : * SLRU needs this.
72 : */
73 : #define CLOG_MAX_ALLOWED_BUFFERS \
74 : Min(SLRU_MAX_ALLOWED_BUFFERS, \
75 : (((MaxTransactionId / 2) + (CLOG_XACTS_PER_PAGE - 1)) / CLOG_XACTS_PER_PAGE))
76 :
77 :
78 : /*
79 : * Although we return an int64 the actual value can't currently exceed
80 : * 0xFFFFFFFF/CLOG_XACTS_PER_PAGE.
81 : */
82 : static inline int64
83 1367157 : TransactionIdToPage(TransactionId xid)
84 : {
85 1367157 : return xid / (int64) CLOG_XACTS_PER_PAGE;
86 : }
87 :
88 : #define TransactionIdToPgIndex(xid) ((xid) % (TransactionId) CLOG_XACTS_PER_PAGE)
89 : #define TransactionIdToByte(xid) (TransactionIdToPgIndex(xid) / CLOG_XACTS_PER_BYTE)
90 : #define TransactionIdToBIndex(xid) ((xid) % (TransactionId) CLOG_XACTS_PER_BYTE)
91 :
92 : /* We store the latest async LSN for each group of transactions */
93 : #define CLOG_XACTS_PER_LSN_GROUP 32 /* keep this a power of 2 */
94 : #define CLOG_LSNS_PER_PAGE (CLOG_XACTS_PER_PAGE / CLOG_XACTS_PER_LSN_GROUP)
95 :
96 : #define GetLSNIndex(slotno, xid) ((slotno) * CLOG_LSNS_PER_PAGE + \
97 : ((xid) % (TransactionId) CLOG_XACTS_PER_PAGE) / CLOG_XACTS_PER_LSN_GROUP)
98 :
99 : /*
100 : * The number of subtransactions below which we consider to apply clog group
101 : * update optimization. Testing reveals that the number higher than this can
102 : * hurt performance.
103 : */
104 : #define THRESHOLD_SUBTRANS_CLOG_OPT 5
105 :
106 : /*
107 : * Link to shared-memory data structures for CLOG control
108 : */
109 : static SlruCtlData XactCtlData;
110 :
111 : #define XactCtl (&XactCtlData)
112 :
113 :
114 : static bool CLOGPagePrecedes(int64 page1, int64 page2);
115 : static int clog_errdetail_for_io_error(const void *opaque_data);
116 : static void WriteTruncateXlogRec(int64 pageno, TransactionId oldestXact,
117 : Oid oldestXactDb);
118 : static void TransactionIdSetPageStatus(TransactionId xid, int nsubxids,
119 : TransactionId *subxids, XidStatus status,
120 : XLogRecPtr lsn, int64 pageno,
121 : bool all_xact_same_page);
122 : static void TransactionIdSetStatusBit(TransactionId xid, XidStatus status,
123 : XLogRecPtr lsn, int slotno);
124 : static void set_status_by_pages(int nsubxids, TransactionId *subxids,
125 : XidStatus status, XLogRecPtr lsn);
126 : static bool TransactionGroupUpdateXidStatus(TransactionId xid,
127 : XidStatus status, XLogRecPtr lsn, int64 pageno);
128 : static void TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids,
129 : TransactionId *subxids, XidStatus status,
130 : XLogRecPtr lsn, int64 pageno);
131 :
132 :
133 : /*
134 : * TransactionIdSetTreeStatus
135 : *
136 : * Record the final state of transaction entries in the commit log for
137 : * a transaction and its subtransaction tree. Take care to ensure this is
138 : * efficient, and as atomic as possible.
139 : *
140 : * xid is a single xid to set status for. This will typically be
141 : * the top level transactionid for a top level commit or abort. It can
142 : * also be a subtransaction when we record transaction aborts.
143 : *
144 : * subxids is an array of xids of length nsubxids, representing subtransactions
145 : * in the tree of xid. In various cases nsubxids may be zero.
146 : *
147 : * lsn must be the WAL location of the commit record when recording an async
148 : * commit. For a synchronous commit it can be InvalidXLogRecPtr, since the
149 : * caller guarantees the commit record is already flushed in that case. It
150 : * should be InvalidXLogRecPtr for abort cases, too.
151 : *
152 : * In the commit case, atomicity is limited by whether all the subxids are in
153 : * the same CLOG page as xid. If they all are, then the lock will be grabbed
154 : * only once, and the status will be set to committed directly. Otherwise
155 : * we must
156 : * 1. set sub-committed all subxids that are not on the same page as the
157 : * main xid
158 : * 2. atomically set committed the main xid and the subxids on the same page
159 : * 3. go over the first bunch again and set them committed
160 : * Note that as far as concurrent checkers are concerned, main transaction
161 : * commit as a whole is still atomic.
162 : *
163 : * Example:
164 : * TransactionId t commits and has subxids t1, t2, t3, t4
165 : * t is on page p1, t1 is also on p1, t2 and t3 are on p2, t4 is on p3
166 : * 1. update pages2-3:
167 : * page2: set t2,t3 as sub-committed
168 : * page3: set t4 as sub-committed
169 : * 2. update page1:
170 : * page1: set t,t1 as committed
171 : * 3. update pages2-3:
172 : * page2: set t2,t3 as committed
173 : * page3: set t4 as committed
174 : *
175 : * NB: this is a low-level routine and is NOT the preferred entry point
176 : * for most uses; functions in transam.c are the intended callers.
177 : *
178 : * XXX Think about issuing POSIX_FADV_WILLNEED on pages that we will need,
179 : * but aren't yet in cache, as well as hinting pages not to fall out of
180 : * cache yet.
181 : */
182 : void
183 160617 : TransactionIdSetTreeStatus(TransactionId xid, int nsubxids,
184 : TransactionId *subxids, XidStatus status, XLogRecPtr lsn)
185 : {
186 160617 : int64 pageno = TransactionIdToPage(xid); /* get page of parent */
187 : int i;
188 :
189 : Assert(status == TRANSACTION_STATUS_COMMITTED ||
190 : status == TRANSACTION_STATUS_ABORTED);
191 :
192 : /*
193 : * See how many subxids, if any, are on the same page as the parent, if
194 : * any.
195 : */
196 165651 : for (i = 0; i < nsubxids; i++)
197 : {
198 5034 : if (TransactionIdToPage(subxids[i]) != pageno)
199 0 : break;
200 : }
201 :
202 : /*
203 : * Do all items fit on a single page?
204 : */
205 160617 : if (i == nsubxids)
206 : {
207 : /*
208 : * Set the parent and all subtransactions in a single call
209 : */
210 160617 : TransactionIdSetPageStatus(xid, nsubxids, subxids, status, lsn,
211 : pageno, true);
212 : }
213 : else
214 : {
215 0 : int nsubxids_on_first_page = i;
216 :
217 : /*
218 : * If this is a commit then we care about doing this correctly (i.e.
219 : * using the subcommitted intermediate status). By here, we know
220 : * we're updating more than one page of clog, so we must mark entries
221 : * that are *not* on the first page so that they show as subcommitted
222 : * before we then return to update the status to fully committed.
223 : *
224 : * To avoid touching the first page twice, skip marking subcommitted
225 : * for the subxids on that first page.
226 : */
227 0 : if (status == TRANSACTION_STATUS_COMMITTED)
228 0 : set_status_by_pages(nsubxids - nsubxids_on_first_page,
229 0 : subxids + nsubxids_on_first_page,
230 : TRANSACTION_STATUS_SUB_COMMITTED, lsn);
231 :
232 : /*
233 : * Now set the parent and subtransactions on same page as the parent,
234 : * if any
235 : */
236 0 : pageno = TransactionIdToPage(xid);
237 0 : TransactionIdSetPageStatus(xid, nsubxids_on_first_page, subxids, status,
238 : lsn, pageno, false);
239 :
240 : /*
241 : * Now work through the rest of the subxids one clog page at a time,
242 : * starting from the second page onwards, like we did above.
243 : */
244 0 : set_status_by_pages(nsubxids - nsubxids_on_first_page,
245 0 : subxids + nsubxids_on_first_page,
246 : status, lsn);
247 : }
248 160617 : }
249 :
250 : /*
251 : * Helper for TransactionIdSetTreeStatus: set the status for a bunch of
252 : * transactions, chunking in the separate CLOG pages involved. We never
253 : * pass the whole transaction tree to this function, only subtransactions
254 : * that are on different pages to the top level transaction id.
255 : */
256 : static void
257 0 : set_status_by_pages(int nsubxids, TransactionId *subxids,
258 : XidStatus status, XLogRecPtr lsn)
259 : {
260 0 : int64 pageno = TransactionIdToPage(subxids[0]);
261 0 : int offset = 0;
262 0 : int i = 0;
263 :
264 : Assert(nsubxids > 0); /* else the pageno fetch above is unsafe */
265 :
266 0 : while (i < nsubxids)
267 : {
268 0 : int num_on_page = 0;
269 : int64 nextpageno;
270 :
271 : do
272 : {
273 0 : nextpageno = TransactionIdToPage(subxids[i]);
274 0 : if (nextpageno != pageno)
275 0 : break;
276 0 : num_on_page++;
277 0 : i++;
278 0 : } while (i < nsubxids);
279 :
280 0 : TransactionIdSetPageStatus(InvalidTransactionId,
281 0 : num_on_page, subxids + offset,
282 : status, lsn, pageno, false);
283 0 : offset = i;
284 0 : pageno = nextpageno;
285 : }
286 0 : }
287 :
288 : /*
289 : * Record the final state of transaction entries in the commit log for all
290 : * entries on a single page. Atomic only on this page.
291 : */
292 : static void
293 160617 : TransactionIdSetPageStatus(TransactionId xid, int nsubxids,
294 : TransactionId *subxids, XidStatus status,
295 : XLogRecPtr lsn, int64 pageno,
296 : bool all_xact_same_page)
297 : {
298 : LWLock *lock;
299 :
300 : /* Can't use group update when PGPROC overflows. */
301 : StaticAssertDecl(THRESHOLD_SUBTRANS_CLOG_OPT <= PGPROC_MAX_CACHED_SUBXIDS,
302 : "group clog threshold less than PGPROC cached subxids");
303 :
304 : /* Get the SLRU bank lock for the page we are going to access. */
305 160617 : lock = SimpleLruGetBankLock(XactCtl, pageno);
306 :
307 : /*
308 : * When there is contention on the SLRU bank lock we need, we try to group
309 : * multiple updates; a single leader process will perform transaction
310 : * status updates for multiple backends so that the number of times the
311 : * bank lock needs to be acquired is reduced.
312 : *
313 : * For this optimization to be safe, the XID and subxids in MyProc must be
314 : * the same as the ones for which we're setting the status. Check that
315 : * this is the case.
316 : *
317 : * For this optimization to be efficient, we shouldn't have too many
318 : * sub-XIDs and all of the XIDs for which we're adjusting clog should be
319 : * on the same page. Check those conditions, too.
320 : */
321 160617 : if (all_xact_same_page && xid == MyProc->xid &&
322 134710 : nsubxids <= THRESHOLD_SUBTRANS_CLOG_OPT &&
323 134710 : nsubxids == MyProc->subxidStatus.count &&
324 459 : (nsubxids == 0 ||
325 459 : memcmp(subxids, MyProc->subxids.xids,
326 : nsubxids * sizeof(TransactionId)) == 0))
327 : {
328 : /*
329 : * If we can immediately acquire the lock, we update the status of our
330 : * own XID and release the lock. If not, try use group XID update. If
331 : * that doesn't work out, fall back to waiting for the lock to perform
332 : * an update for this transaction only.
333 : */
334 134594 : if (LWLockConditionalAcquire(lock, LW_EXCLUSIVE))
335 : {
336 : /* Got the lock without waiting! Do the update. */
337 134512 : TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
338 : lsn, pageno);
339 134512 : LWLockRelease(lock);
340 134512 : return;
341 : }
342 82 : else if (TransactionGroupUpdateXidStatus(xid, status, lsn, pageno))
343 : {
344 : /* Group update mechanism has done the work. */
345 82 : return;
346 : }
347 :
348 : /* Fall through only if update isn't done yet. */
349 : }
350 :
351 : /* Group update not applicable, or couldn't accept this page number. */
352 26023 : LWLockAcquire(lock, LW_EXCLUSIVE);
353 26023 : TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
354 : lsn, pageno);
355 26023 : LWLockRelease(lock);
356 : }
357 :
358 : /*
359 : * Record the final state of transaction entry in the commit log
360 : *
361 : * We don't do any locking here; caller must handle that.
362 : */
363 : static void
364 160617 : TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids,
365 : TransactionId *subxids, XidStatus status,
366 : XLogRecPtr lsn, int64 pageno)
367 : {
368 : int slotno;
369 : int i;
370 :
371 : Assert(status == TRANSACTION_STATUS_COMMITTED ||
372 : status == TRANSACTION_STATUS_ABORTED ||
373 : (status == TRANSACTION_STATUS_SUB_COMMITTED && !TransactionIdIsValid(xid)));
374 : Assert(LWLockHeldByMeInMode(SimpleLruGetBankLock(XactCtl, pageno),
375 : LW_EXCLUSIVE));
376 :
377 : /*
378 : * If we're doing an async commit (ie, lsn is valid), then we must wait
379 : * for any active write on the page slot to complete. Otherwise our
380 : * update could reach disk in that write, which will not do since we
381 : * mustn't let it reach disk until we've done the appropriate WAL flush.
382 : * But when lsn is invalid, it's OK to scribble on a page while it is
383 : * write-busy, since we don't care if the update reaches disk sooner than
384 : * we think.
385 : */
386 160617 : slotno = SimpleLruReadPage(XactCtl, pageno, !XLogRecPtrIsValid(lsn), &xid);
387 :
388 : /*
389 : * Set the main transaction id, if any.
390 : *
391 : * If we update more than one xid on this page while it is being written
392 : * out, we might find that some of the bits go to disk and others don't.
393 : * If we are updating commits on the page with the top-level xid that
394 : * could break atomicity, so we subcommit the subxids first before we mark
395 : * the top-level commit.
396 : */
397 160617 : if (TransactionIdIsValid(xid))
398 : {
399 : /* Subtransactions first, if needed ... */
400 160617 : if (status == TRANSACTION_STATUS_COMMITTED)
401 : {
402 156334 : for (i = 0; i < nsubxids; i++)
403 : {
404 : Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
405 4714 : TransactionIdSetStatusBit(subxids[i],
406 : TRANSACTION_STATUS_SUB_COMMITTED,
407 : lsn, slotno);
408 : }
409 : }
410 :
411 : /* ... then the main transaction */
412 160617 : TransactionIdSetStatusBit(xid, status, lsn, slotno);
413 : }
414 :
415 : /* Set the subtransactions */
416 165651 : for (i = 0; i < nsubxids; i++)
417 : {
418 : Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
419 5034 : TransactionIdSetStatusBit(subxids[i], status, lsn, slotno);
420 : }
421 :
422 160617 : XactCtl->shared->page_dirty[slotno] = true;
423 160617 : }
424 :
425 : /*
426 : * Subroutine for TransactionIdSetPageStatus, q.v.
427 : *
428 : * When we cannot immediately acquire the SLRU bank lock in exclusive mode at
429 : * commit time, add ourselves to a list of processes that need their XIDs
430 : * status update. The first process to add itself to the list will acquire
431 : * the lock in exclusive mode and set transaction status as required on behalf
432 : * of all group members. This avoids a great deal of contention when many
433 : * processes are trying to commit at once, since the lock need not be
434 : * repeatedly handed off from one committing process to the next.
435 : *
436 : * Returns true when transaction status has been updated in clog; returns
437 : * false if we decided against applying the optimization because the page
438 : * number we need to update differs from those processes already waiting.
439 : */
440 : static bool
441 82 : TransactionGroupUpdateXidStatus(TransactionId xid, XidStatus status,
442 : XLogRecPtr lsn, int64 pageno)
443 : {
444 82 : volatile PROC_HDR *procglobal = ProcGlobal;
445 82 : PGPROC *proc = MyProc;
446 : uint32 nextidx;
447 : uint32 wakeidx;
448 : int64 prevpageno;
449 82 : LWLock *prevlock = NULL;
450 :
451 : /* We should definitely have an XID whose status needs to be updated. */
452 : Assert(TransactionIdIsValid(xid));
453 :
454 : /*
455 : * Prepare to add ourselves to the list of processes needing a group XID
456 : * status update.
457 : */
458 82 : proc->clogGroupMember = true;
459 82 : proc->clogGroupMemberXid = xid;
460 82 : proc->clogGroupMemberXidStatus = status;
461 82 : proc->clogGroupMemberPage = pageno;
462 82 : proc->clogGroupMemberLsn = lsn;
463 :
464 : /*
465 : * We put ourselves in the queue by writing MyProcNumber to
466 : * ProcGlobal->clogGroupFirst. However, if there's already a process
467 : * listed there, we compare our pageno with that of that process; if it
468 : * differs, we cannot participate in the group, so we return for caller to
469 : * update pg_xact in the normal way.
470 : *
471 : * If we're not the first process in the list, we must follow the leader.
472 : * We do this by storing the data we want updated in our PGPROC entry
473 : * where the leader can find it, then going to sleep.
474 : *
475 : * If no process is already in the list, we're the leader; our first step
476 : * is to lock the SLRU bank to which our page belongs, then we close out
477 : * the group by resetting the list pointer from ProcGlobal->clogGroupFirst
478 : * (this lets other processes set up other groups later); finally we do
479 : * the SLRU updates, release the SLRU bank lock, and wake up the sleeping
480 : * processes.
481 : *
482 : * If another group starts to update a page in a different SLRU bank, they
483 : * can proceed concurrently, since the bank lock they're going to use is
484 : * different from ours. If another group starts to update a page in the
485 : * same bank as ours, they wait until we release the lock.
486 : */
487 82 : nextidx = pg_atomic_read_u32(&procglobal->clogGroupFirst);
488 :
489 : while (true)
490 : {
491 : /*
492 : * Add the proc to list, if the clog page where we need to update the
493 : * current transaction status is same as group leader's clog page.
494 : *
495 : * There is a race condition here, which is that after doing the below
496 : * check and before adding this proc's clog update to a group, the
497 : * group leader might have already finished the group update for this
498 : * page and becomes group leader of another group, updating a
499 : * different page. This will lead to a situation where a single group
500 : * can have different clog page updates. This isn't likely and will
501 : * still work, just less efficiently -- we handle this case by
502 : * switching to a different bank lock in the loop below.
503 : */
504 82 : if (nextidx != INVALID_PROC_NUMBER &&
505 7 : GetPGProcByNumber(nextidx)->clogGroupMemberPage != proc->clogGroupMemberPage)
506 : {
507 : /*
508 : * Ensure that this proc is not a member of any clog group that
509 : * needs an XID status update.
510 : */
511 0 : proc->clogGroupMember = false;
512 0 : pg_atomic_write_u32(&proc->clogGroupNext, INVALID_PROC_NUMBER);
513 0 : return false;
514 : }
515 :
516 82 : pg_atomic_write_u32(&proc->clogGroupNext, nextidx);
517 :
518 82 : if (pg_atomic_compare_exchange_u32(&procglobal->clogGroupFirst,
519 : &nextidx,
520 : (uint32) MyProcNumber))
521 82 : break;
522 : }
523 :
524 : /*
525 : * If the list was not empty, the leader will update the status of our
526 : * XID. It is impossible to have followers without a leader because the
527 : * first process that has added itself to the list will always have
528 : * nextidx as INVALID_PROC_NUMBER.
529 : */
530 82 : if (nextidx != INVALID_PROC_NUMBER)
531 : {
532 7 : int extraWaits = 0;
533 :
534 : /* Sleep until the leader updates our XID status. */
535 7 : pgstat_report_wait_start(WAIT_EVENT_XACT_GROUP_UPDATE);
536 : for (;;)
537 : {
538 : /* acts as a read barrier */
539 7 : PGSemaphoreLock(proc->sem);
540 7 : if (!proc->clogGroupMember)
541 7 : break;
542 0 : extraWaits++;
543 : }
544 7 : pgstat_report_wait_end();
545 :
546 : Assert(pg_atomic_read_u32(&proc->clogGroupNext) == INVALID_PROC_NUMBER);
547 :
548 : /* Fix semaphore count for any absorbed wakeups */
549 7 : while (extraWaits-- > 0)
550 0 : PGSemaphoreUnlock(proc->sem);
551 7 : return true;
552 : }
553 :
554 : /*
555 : * By here, we know we're the leader process. Acquire the SLRU bank lock
556 : * that corresponds to the page we originally wanted to modify.
557 : */
558 75 : prevpageno = proc->clogGroupMemberPage;
559 75 : prevlock = SimpleLruGetBankLock(XactCtl, prevpageno);
560 75 : LWLockAcquire(prevlock, LW_EXCLUSIVE);
561 :
562 : /*
563 : * Now that we've got the lock, clear the list of processes waiting for
564 : * group XID status update, saving a pointer to the head of the list.
565 : * (Trying to pop elements one at a time could lead to an ABA problem.)
566 : *
567 : * At this point, any processes trying to do this would create a separate
568 : * group.
569 : */
570 75 : nextidx = pg_atomic_exchange_u32(&procglobal->clogGroupFirst,
571 : INVALID_PROC_NUMBER);
572 :
573 : /* Remember head of list so we can perform wakeups after dropping lock. */
574 75 : wakeidx = nextidx;
575 :
576 : /* Walk the list and update the status of all XIDs. */
577 157 : while (nextidx != INVALID_PROC_NUMBER)
578 : {
579 82 : PGPROC *nextproc = GetPGProcByNumber(nextidx);
580 82 : int64 thispageno = nextproc->clogGroupMemberPage;
581 :
582 : /*
583 : * If the page to update belongs to a different bank than the previous
584 : * one, exchange bank lock to the new one. This should be quite rare,
585 : * as described above.
586 : *
587 : * (We could try to optimize this by waking up the processes for which
588 : * we have already updated the status while we exchange the lock, but
589 : * the code doesn't do that at present. I think it'd require
590 : * additional bookkeeping, making the common path slower in order to
591 : * improve an infrequent case.)
592 : */
593 82 : if (thispageno != prevpageno)
594 : {
595 0 : LWLock *lock = SimpleLruGetBankLock(XactCtl, thispageno);
596 :
597 0 : if (prevlock != lock)
598 : {
599 0 : LWLockRelease(prevlock);
600 0 : LWLockAcquire(lock, LW_EXCLUSIVE);
601 : }
602 0 : prevlock = lock;
603 0 : prevpageno = thispageno;
604 : }
605 :
606 : /*
607 : * Transactions with more than THRESHOLD_SUBTRANS_CLOG_OPT sub-XIDs
608 : * should not use group XID status update mechanism.
609 : */
610 : Assert(nextproc->subxidStatus.count <= THRESHOLD_SUBTRANS_CLOG_OPT);
611 :
612 82 : TransactionIdSetPageStatusInternal(nextproc->clogGroupMemberXid,
613 82 : nextproc->subxidStatus.count,
614 82 : nextproc->subxids.xids,
615 : nextproc->clogGroupMemberXidStatus,
616 : nextproc->clogGroupMemberLsn,
617 : nextproc->clogGroupMemberPage);
618 :
619 : /* Move to next proc in list. */
620 82 : nextidx = pg_atomic_read_u32(&nextproc->clogGroupNext);
621 : }
622 :
623 : /* We're done with the lock now. */
624 75 : if (prevlock != NULL)
625 75 : LWLockRelease(prevlock);
626 :
627 : /*
628 : * Now that we've released the lock, go back and wake everybody up. We
629 : * don't do this under the lock so as to keep lock hold times to a
630 : * minimum.
631 : *
632 : * (Perhaps we could do this in two passes, the first setting
633 : * clogGroupNext to invalid while saving the semaphores to an array, then
634 : * a single write barrier, then another pass unlocking the semaphores.)
635 : */
636 157 : while (wakeidx != INVALID_PROC_NUMBER)
637 : {
638 82 : PGPROC *wakeproc = GetPGProcByNumber(wakeidx);
639 :
640 82 : wakeidx = pg_atomic_read_u32(&wakeproc->clogGroupNext);
641 82 : pg_atomic_write_u32(&wakeproc->clogGroupNext, INVALID_PROC_NUMBER);
642 :
643 : /* ensure all previous writes are visible before follower continues. */
644 82 : pg_write_barrier();
645 :
646 82 : wakeproc->clogGroupMember = false;
647 :
648 82 : if (wakeproc != MyProc)
649 7 : PGSemaphoreUnlock(wakeproc->sem);
650 : }
651 :
652 75 : return true;
653 : }
654 :
655 : /*
656 : * Sets the commit status of a single transaction.
657 : *
658 : * Caller must hold the corresponding SLRU bank lock, will be held at exit.
659 : */
660 : static void
661 170365 : TransactionIdSetStatusBit(TransactionId xid, XidStatus status, XLogRecPtr lsn, int slotno)
662 : {
663 170365 : int byteno = TransactionIdToByte(xid);
664 170365 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
665 : char *byteptr;
666 : char byteval;
667 : char curval;
668 :
669 : Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(xid));
670 : Assert(LWLockHeldByMeInMode(SimpleLruGetBankLock(XactCtl,
671 : XactCtl->shared->page_number[slotno]),
672 : LW_EXCLUSIVE));
673 :
674 170365 : byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
675 170365 : curval = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
676 :
677 : /*
678 : * When replaying transactions during recovery we still need to perform
679 : * the two phases of subcommit and then commit. However, some transactions
680 : * are already correctly marked, so we just treat those as a no-op which
681 : * allows us to keep the following Assert as restrictive as possible.
682 : */
683 170365 : if (InRecovery && status == TRANSACTION_STATUS_SUB_COMMITTED &&
684 : curval == TRANSACTION_STATUS_COMMITTED)
685 0 : return;
686 :
687 : /*
688 : * Current state change should be from 0 or subcommitted to target state
689 : * or we should already be there when replaying changes during recovery.
690 : */
691 : Assert(curval == 0 ||
692 : (curval == TRANSACTION_STATUS_SUB_COMMITTED &&
693 : status != TRANSACTION_STATUS_IN_PROGRESS) ||
694 : curval == status);
695 :
696 : /* note this assumes exclusive access to the clog page */
697 170365 : byteval = *byteptr;
698 170365 : byteval &= ~(((1 << CLOG_BITS_PER_XACT) - 1) << bshift);
699 170365 : byteval |= (status << bshift);
700 170365 : *byteptr = byteval;
701 :
702 : /*
703 : * Update the group LSN if the transaction completion LSN is higher.
704 : *
705 : * Note: lsn will be invalid when supplied during InRecovery processing,
706 : * so we don't need to do anything special to avoid LSN updates during
707 : * recovery. After recovery completes the next clog change will set the
708 : * LSN correctly.
709 : */
710 170365 : if (XLogRecPtrIsValid(lsn))
711 : {
712 29779 : int lsnindex = GetLSNIndex(slotno, xid);
713 :
714 29779 : if (XactCtl->shared->group_lsn[lsnindex] < lsn)
715 27166 : XactCtl->shared->group_lsn[lsnindex] = lsn;
716 : }
717 : }
718 :
719 : /*
720 : * Interrogate the state of a transaction in the commit log.
721 : *
722 : * Aside from the actual commit status, this function returns (into *lsn)
723 : * an LSN that is late enough to be able to guarantee that if we flush up to
724 : * that LSN then we will have flushed the transaction's commit record to disk.
725 : * The result is not necessarily the exact LSN of the transaction's commit
726 : * record! For example, for long-past transactions (those whose clog pages
727 : * already migrated to disk), we'll return InvalidXLogRecPtr. Also, because
728 : * we group transactions on the same clog page to conserve storage, we might
729 : * return the LSN of a later transaction that falls into the same group.
730 : *
731 : * NB: this is a low-level routine and is NOT the preferred entry point
732 : * for most uses; TransactionLogFetch() in transam.c is the intended caller.
733 : */
734 : XidStatus
735 766329 : TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
736 : {
737 766329 : int64 pageno = TransactionIdToPage(xid);
738 766329 : int byteno = TransactionIdToByte(xid);
739 766329 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
740 : int slotno;
741 : int lsnindex;
742 : char *byteptr;
743 : XidStatus status;
744 :
745 : /* lock is acquired by SimpleLruReadPage_ReadOnly */
746 :
747 766329 : slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, &xid);
748 766329 : byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
749 :
750 766329 : status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
751 :
752 766329 : lsnindex = GetLSNIndex(slotno, xid);
753 766329 : *lsn = XactCtl->shared->group_lsn[lsnindex];
754 :
755 766329 : LWLockRelease(SimpleLruGetBankLock(XactCtl, pageno));
756 :
757 766329 : return status;
758 : }
759 :
760 : /*
761 : * Number of shared CLOG buffers.
762 : *
763 : * If asked to autotune, use 2MB for every 1GB of shared buffers, up to 8MB.
764 : * Otherwise just cap the configured amount to be between 16 and the maximum
765 : * allowed.
766 : */
767 : static int
768 4478 : CLOGShmemBuffers(void)
769 : {
770 : /* auto-tune based on shared buffers */
771 4478 : if (transaction_buffers == 0)
772 3314 : return SimpleLruAutotuneBuffers(512, 1024);
773 :
774 1164 : return Min(Max(16, transaction_buffers), CLOG_MAX_ALLOWED_BUFFERS);
775 : }
776 :
777 : /*
778 : * Initialization of shared memory for CLOG
779 : */
780 : Size
781 2165 : CLOGShmemSize(void)
782 : {
783 2165 : return SimpleLruShmemSize(CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE);
784 : }
785 :
786 : void
787 1159 : CLOGShmemInit(void)
788 : {
789 : /* If auto-tuning is requested, now is the time to do it */
790 1159 : if (transaction_buffers == 0)
791 : {
792 : char buf[32];
793 :
794 1154 : snprintf(buf, sizeof(buf), "%d", CLOGShmemBuffers());
795 1154 : SetConfigOption("transaction_buffers", buf, PGC_POSTMASTER,
796 : PGC_S_DYNAMIC_DEFAULT);
797 :
798 : /*
799 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
800 : * However, if the DBA explicitly set transaction_buffers = 0 in the
801 : * config file, then PGC_S_DYNAMIC_DEFAULT will fail to override that
802 : * and we must force the matter with PGC_S_OVERRIDE.
803 : */
804 1154 : if (transaction_buffers == 0) /* failed to apply it? */
805 0 : SetConfigOption("transaction_buffers", buf, PGC_POSTMASTER,
806 : PGC_S_OVERRIDE);
807 : }
808 : Assert(transaction_buffers != 0);
809 :
810 1159 : XactCtl->PagePrecedes = CLOGPagePrecedes;
811 1159 : XactCtl->errdetail_for_io_error = clog_errdetail_for_io_error;
812 1159 : SimpleLruInit(XactCtl, "transaction", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
813 : "pg_xact", LWTRANCHE_XACT_BUFFER,
814 : LWTRANCHE_XACT_SLRU, SYNC_HANDLER_CLOG, false);
815 : SlruPagePrecedesUnitTests(XactCtl, CLOG_XACTS_PER_PAGE);
816 1159 : }
817 :
818 : /*
819 : * GUC check_hook for transaction_buffers
820 : */
821 : bool
822 2347 : check_transaction_buffers(int *newval, void **extra, GucSource source)
823 : {
824 2347 : return check_slru_buffers("transaction_buffers", newval);
825 : }
826 :
827 : /*
828 : * This func must be called ONCE on system install. It creates
829 : * the initial CLOG segment. (The CLOG directory is assumed to
830 : * have been created by initdb, and CLOGShmemInit must have been
831 : * called already.)
832 : */
833 : void
834 51 : BootStrapCLOG(void)
835 : {
836 : /* Zero the initial page and flush it to disk */
837 51 : SimpleLruZeroAndWritePage(XactCtl, 0);
838 51 : }
839 :
840 : /*
841 : * This must be called ONCE during postmaster or standalone-backend startup,
842 : * after StartupXLOG has initialized TransamVariables->nextXid.
843 : */
844 : void
845 1010 : StartupCLOG(void)
846 : {
847 1010 : TransactionId xid = XidFromFullTransactionId(TransamVariables->nextXid);
848 1010 : int64 pageno = TransactionIdToPage(xid);
849 :
850 : /*
851 : * Initialize our idea of the latest page number.
852 : */
853 1010 : pg_atomic_write_u64(&XactCtl->shared->latest_page_number, pageno);
854 1010 : }
855 :
856 : /*
857 : * This must be called ONCE at the end of startup/recovery.
858 : */
859 : void
860 949 : TrimCLOG(void)
861 : {
862 949 : TransactionId xid = XidFromFullTransactionId(TransamVariables->nextXid);
863 949 : int64 pageno = TransactionIdToPage(xid);
864 949 : LWLock *lock = SimpleLruGetBankLock(XactCtl, pageno);
865 :
866 949 : LWLockAcquire(lock, LW_EXCLUSIVE);
867 :
868 : /*
869 : * Zero out the remainder of the current clog page. Under normal
870 : * circumstances it should be zeroes already, but it seems at least
871 : * theoretically possible that XLOG replay will have settled on a nextXID
872 : * value that is less than the last XID actually used and marked by the
873 : * previous database lifecycle (since subtransaction commit writes clog
874 : * but makes no WAL entry). Let's just be safe. (We need not worry about
875 : * pages beyond the current one, since those will be zeroed when first
876 : * used. For the same reason, there is no need to do anything when
877 : * nextXid is exactly at a page boundary; and it's likely that the
878 : * "current" page doesn't exist yet in that case.)
879 : */
880 949 : if (TransactionIdToPgIndex(xid) != 0)
881 : {
882 948 : int byteno = TransactionIdToByte(xid);
883 948 : int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
884 : int slotno;
885 : char *byteptr;
886 :
887 948 : slotno = SimpleLruReadPage(XactCtl, pageno, false, &xid);
888 948 : byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
889 :
890 : /* Zero so-far-unused positions in the current byte */
891 948 : *byteptr &= (1 << bshift) - 1;
892 : /* Zero the rest of the page */
893 948 : MemSet(byteptr + 1, 0, BLCKSZ - byteno - 1);
894 :
895 948 : XactCtl->shared->page_dirty[slotno] = true;
896 : }
897 :
898 949 : LWLockRelease(lock);
899 949 : }
900 :
901 : /*
902 : * Perform a checkpoint --- either during shutdown, or on-the-fly
903 : */
904 : void
905 1807 : CheckPointCLOG(void)
906 : {
907 : /*
908 : * Write dirty CLOG pages to disk. This may result in sync requests
909 : * queued for later handling by ProcessSyncRequests(), as part of the
910 : * checkpoint.
911 : */
912 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(true);
913 1807 : SimpleLruWriteAll(XactCtl, true);
914 : TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(true);
915 1807 : }
916 :
917 :
918 : /*
919 : * Make sure that CLOG has room for a newly-allocated XID.
920 : *
921 : * NB: this is called while holding XidGenLock. We want it to be very fast
922 : * most of the time; even when it's not so fast, no actual I/O need happen
923 : * unless we're forced to write out a dirty clog or xlog page to make room
924 : * in shared memory.
925 : */
926 : void
927 24524718 : ExtendCLOG(TransactionId newestXact)
928 : {
929 : int64 pageno;
930 : LWLock *lock;
931 :
932 : /*
933 : * No work except at first XID of a page. But beware: just after
934 : * wraparound, the first XID of page zero is FirstNormalTransactionId.
935 : */
936 24524718 : if (TransactionIdToPgIndex(newestXact) != 0 &&
937 : !TransactionIdEquals(newestXact, FirstNormalTransactionId))
938 24092710 : return;
939 :
940 432008 : pageno = TransactionIdToPage(newestXact);
941 432008 : lock = SimpleLruGetBankLock(XactCtl, pageno);
942 :
943 432008 : LWLockAcquire(lock, LW_EXCLUSIVE);
944 :
945 : /* Zero the page and make a WAL entry about it */
946 432008 : SimpleLruZeroPage(XactCtl, pageno);
947 432008 : XLogSimpleInsertInt64(RM_CLOG_ID, CLOG_ZEROPAGE, pageno);
948 :
949 432008 : LWLockRelease(lock);
950 : }
951 :
952 :
953 : /*
954 : * Remove all CLOG segments before the one holding the passed transaction ID
955 : *
956 : * Before removing any CLOG data, we must flush XLOG to disk, to ensure that
957 : * any recently-emitted records with freeze plans have reached disk; otherwise
958 : * a crash and restart might leave us with some unfrozen tuples referencing
959 : * removed CLOG data. We choose to emit a special TRUNCATE XLOG record too.
960 : * Replaying the deletion from XLOG is not critical, since the files could
961 : * just as well be removed later, but doing so prevents a long-running hot
962 : * standby server from acquiring an unreasonably bloated CLOG directory.
963 : *
964 : * Since CLOG segments hold a large number of transactions, the opportunity to
965 : * actually remove a segment is fairly rare, and so it seems best not to do
966 : * the XLOG flush unless we have confirmed that there is a removable segment.
967 : */
968 : void
969 1210 : TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
970 : {
971 : int64 cutoffPage;
972 :
973 : /*
974 : * The cutoff point is the start of the segment containing oldestXact. We
975 : * pass the *page* containing oldestXact to SimpleLruTruncate.
976 : */
977 1210 : cutoffPage = TransactionIdToPage(oldestXact);
978 :
979 : /* Check to see if there's any files that could be removed */
980 1210 : if (!SlruScanDirectory(XactCtl, SlruScanDirCbReportPresence, &cutoffPage))
981 1107 : return; /* nothing to remove */
982 :
983 : /*
984 : * Advance oldestClogXid before truncating clog, so concurrent xact status
985 : * lookups can ensure they don't attempt to access truncated-away clog.
986 : *
987 : * It's only necessary to do this if we will actually truncate away clog
988 : * pages.
989 : */
990 103 : AdvanceOldestClogXid(oldestXact);
991 :
992 : /*
993 : * Write XLOG record and flush XLOG to disk. We record the oldest xid
994 : * we're keeping information about here so we can ensure that it's always
995 : * ahead of clog truncation in case we crash, and so a standby finds out
996 : * the new valid xid before the next checkpoint.
997 : */
998 103 : WriteTruncateXlogRec(cutoffPage, oldestXact, oldestxid_datoid);
999 :
1000 : /* Now we can remove the old CLOG segment(s) */
1001 103 : SimpleLruTruncate(XactCtl, cutoffPage);
1002 : }
1003 :
1004 :
1005 : /*
1006 : * Decide whether a CLOG page number is "older" for truncation purposes.
1007 : *
1008 : * We need to use comparison of TransactionIds here in order to do the right
1009 : * thing with wraparound XID arithmetic. However, TransactionIdPrecedes()
1010 : * would get weird about permanent xact IDs. So, offset both such that xid1,
1011 : * xid2, and xid2 + CLOG_XACTS_PER_PAGE - 1 are all normal XIDs; this offset
1012 : * is relevant to page 0 and to the page preceding page 0.
1013 : *
1014 : * The page containing oldestXact-2^31 is the important edge case. The
1015 : * portion of that page equaling or following oldestXact-2^31 is expendable,
1016 : * but the portion preceding oldestXact-2^31 is not. When oldestXact-2^31 is
1017 : * the first XID of a page and segment, the entire page and segment is
1018 : * expendable, and we could truncate the segment. Recognizing that case would
1019 : * require making oldestXact, not just the page containing oldestXact,
1020 : * available to this callback. The benefit would be rare and small, so we
1021 : * don't optimize that edge case.
1022 : */
1023 : static bool
1024 911126 : CLOGPagePrecedes(int64 page1, int64 page2)
1025 : {
1026 : TransactionId xid1;
1027 : TransactionId xid2;
1028 :
1029 911126 : xid1 = ((TransactionId) page1) * CLOG_XACTS_PER_PAGE;
1030 911126 : xid1 += FirstNormalTransactionId + 1;
1031 911126 : xid2 = ((TransactionId) page2) * CLOG_XACTS_PER_PAGE;
1032 911126 : xid2 += FirstNormalTransactionId + 1;
1033 :
1034 938262 : return (TransactionIdPrecedes(xid1, xid2) &&
1035 27136 : TransactionIdPrecedes(xid1, xid2 + CLOG_XACTS_PER_PAGE - 1));
1036 : }
1037 :
1038 : static int
1039 0 : clog_errdetail_for_io_error(const void *opaque_data)
1040 : {
1041 0 : TransactionId xid = *(const TransactionId *) opaque_data;
1042 :
1043 0 : return errdetail("Could not access commit status of transaction %u.", xid);
1044 : }
1045 :
1046 :
1047 : /*
1048 : * Write a TRUNCATE xlog record
1049 : *
1050 : * We must flush the xlog record to disk before returning --- see notes
1051 : * in TruncateCLOG().
1052 : */
1053 : static void
1054 103 : WriteTruncateXlogRec(int64 pageno, TransactionId oldestXact, Oid oldestXactDb)
1055 : {
1056 : XLogRecPtr recptr;
1057 : xl_clog_truncate xlrec;
1058 :
1059 103 : xlrec.pageno = pageno;
1060 103 : xlrec.oldestXact = oldestXact;
1061 103 : xlrec.oldestXactDb = oldestXactDb;
1062 :
1063 103 : XLogBeginInsert();
1064 103 : XLogRegisterData(&xlrec, sizeof(xl_clog_truncate));
1065 103 : recptr = XLogInsert(RM_CLOG_ID, CLOG_TRUNCATE);
1066 103 : XLogFlush(recptr);
1067 103 : }
1068 :
1069 : /*
1070 : * CLOG resource manager's routines
1071 : */
1072 : void
1073 0 : clog_redo(XLogReaderState *record)
1074 : {
1075 0 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
1076 :
1077 : /* Backup blocks are not used in clog records */
1078 : Assert(!XLogRecHasAnyBlockRefs(record));
1079 :
1080 0 : if (info == CLOG_ZEROPAGE)
1081 : {
1082 : int64 pageno;
1083 :
1084 0 : memcpy(&pageno, XLogRecGetData(record), sizeof(pageno));
1085 0 : SimpleLruZeroAndWritePage(XactCtl, pageno);
1086 : }
1087 0 : else if (info == CLOG_TRUNCATE)
1088 : {
1089 : xl_clog_truncate xlrec;
1090 :
1091 0 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_clog_truncate));
1092 :
1093 0 : AdvanceOldestClogXid(xlrec.oldestXact);
1094 :
1095 0 : SimpleLruTruncate(XactCtl, xlrec.pageno);
1096 : }
1097 : else
1098 0 : elog(PANIC, "clog_redo: unknown op code %u", info);
1099 0 : }
1100 :
1101 : /*
1102 : * Entrypoint for sync.c to sync clog files.
1103 : */
1104 : int
1105 0 : clogsyncfiletag(const FileTag *ftag, char *path)
1106 : {
1107 0 : return SlruSyncFileTag(XactCtl, ftag, path);
1108 : }
|
