PostgreSQL Source Code  git master
clog.c
Go to the documentation of this file.
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  */
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);
101  XLogRecPtr lsn, int slotno);
102 static void set_status_by_pages(int nsubxids, TransactionId *subxids,
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
164  TransactionId *subxids, XidStatus status, XLogRecPtr lsn)
165 {
166  int pageno = TransactionIdToPage(xid); /* get page of parent */
167  int i;
168 
170  status == TRANSACTION_STATUS_ABORTED);
171 
172  /*
173  * See how many subxids, if any, are on the same page as the parent, if
174  * any.
175  */
176  for (i = 0; i < nsubxids; i++)
177  {
178  if (TransactionIdToPage(subxids[i]) != pageno)
179  break;
180  }
181 
182  /*
183  * Do all items fit on a single page?
184  */
185  if (i == nsubxids)
186  {
187  /*
188  * Set the parent and all subtransactions in a single call
189  */
190  TransactionIdSetPageStatus(xid, nsubxids, subxids, status, lsn,
191  pageno, true);
192  }
193  else
194  {
195  int nsubxids_on_first_page = i;
196 
197  /*
198  * If this is a commit then we care about doing this correctly (i.e.
199  * using the subcommitted intermediate status). By here, we know
200  * we're updating more than one page of clog, so we must mark entries
201  * that are *not* on the first page so that they show as subcommitted
202  * before we then return to update the status to fully committed.
203  *
204  * To avoid touching the first page twice, skip marking subcommitted
205  * for the subxids on that first page.
206  */
207  if (status == TRANSACTION_STATUS_COMMITTED)
208  set_status_by_pages(nsubxids - nsubxids_on_first_page,
209  subxids + nsubxids_on_first_page,
211 
212  /*
213  * Now set the parent and subtransactions on same page as the parent,
214  * if any
215  */
216  pageno = TransactionIdToPage(xid);
217  TransactionIdSetPageStatus(xid, nsubxids_on_first_page, subxids, status,
218  lsn, pageno, false);
219 
220  /*
221  * Now work through the rest of the subxids one clog page at a time,
222  * starting from the second page onwards, like we did above.
223  */
224  set_status_by_pages(nsubxids - nsubxids_on_first_page,
225  subxids + nsubxids_on_first_page,
226  status, lsn);
227  }
228 }
229 
230 /*
231  * Helper for TransactionIdSetTreeStatus: set the status for a bunch of
232  * transactions, chunking in the separate CLOG pages involved. We never
233  * pass the whole transaction tree to this function, only subtransactions
234  * that are on different pages to the top level transaction id.
235  */
236 static void
237 set_status_by_pages(int nsubxids, TransactionId *subxids,
239 {
240  int pageno = TransactionIdToPage(subxids[0]);
241  int offset = 0;
242  int i = 0;
243 
244  Assert(nsubxids > 0); /* else the pageno fetch above is unsafe */
245 
246  while (i < nsubxids)
247  {
248  int num_on_page = 0;
249  int nextpageno;
250 
251  do
252  {
253  nextpageno = TransactionIdToPage(subxids[i]);
254  if (nextpageno != pageno)
255  break;
256  num_on_page++;
257  i++;
258  } while (i < nsubxids);
259 
261  num_on_page, subxids + offset,
262  status, lsn, pageno, false);
263  offset = i;
264  pageno = nextpageno;
265  }
266 }
267 
268 /*
269  * Record the final state of transaction entries in the commit log for all
270  * entries on a single page. Atomic only on this page.
271  */
272 static void
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. */
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 and subxids in MyProc must be
289  * the same as the ones for which we're setting the status. Check that
290  * this is the case.
291  *
292  * For this optimization to be efficient, we shouldn't have too many
293  * sub-XIDs and all of the XIDs for which we're adjusting clog should be
294  * on the same page. Check those conditions, too.
295  */
296  if (all_xact_same_page && xid == MyProc->xid &&
297  nsubxids <= THRESHOLD_SUBTRANS_CLOG_OPT &&
298  nsubxids == MyProc->subxidStatus.count &&
299  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  if (LWLockConditionalAcquire(XactSLRULock, LW_EXCLUSIVE))
309  {
310  /* Got the lock without waiting! Do the update. */
311  TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
312  lsn, pageno);
313  LWLockRelease(XactSLRULock);
314  return;
315  }
316  else if (TransactionGroupUpdateXidStatus(xid, status, lsn, pageno))
317  {
318  /* Group update mechanism has done the work. */
319  return;
320  }
321 
322  /* Fall through only if update isn't done yet. */
323  }
324 
325  /* Group update not applicable, or couldn't accept this page number. */
326  LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
327  TransactionIdSetPageStatusInternal(xid, nsubxids, subxids, status,
328  lsn, pageno);
329  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
339  TransactionId *subxids, XidStatus status,
340  XLogRecPtr lsn, int pageno)
341 {
342  int slotno;
343  int i;
344 
346  status == TRANSACTION_STATUS_ABORTED ||
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  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  if (TransactionIdIsValid(xid))
371  {
372  /* Subtransactions first, if needed ... */
373  if (status == TRANSACTION_STATUS_COMMITTED)
374  {
375  for (i = 0; i < nsubxids; i++)
376  {
377  Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
378  TransactionIdSetStatusBit(subxids[i],
380  lsn, slotno);
381  }
382  }
383 
384  /* ... then the main transaction */
385  TransactionIdSetStatusBit(xid, status, lsn, slotno);
386  }
387 
388  /* Set the subtransactions */
389  for (i = 0; i < nsubxids; i++)
390  {
391  Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
392  TransactionIdSetStatusBit(subxids[i], status, lsn, slotno);
393  }
394 
395  XactCtl->shared->page_dirty[slotno] = true;
396 }
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
414  XLogRecPtr lsn, int pageno)
415 {
416  volatile PROC_HDR *procglobal = ProcGlobal;
417  PGPROC *proc = MyProc;
418  uint32 nextidx;
419  uint32 wakeidx;
420 
421  /* We should definitely have an XID whose status needs to be updated. */
423 
424  /*
425  * Add ourselves to the list of processes needing a group XID status
426  * update.
427  */
428  proc->clogGroupMember = true;
429  proc->clogGroupMemberXid = xid;
431  proc->clogGroupMemberPage = pageno;
432  proc->clogGroupMemberLsn = lsn;
433 
434  nextidx = pg_atomic_read_u32(&procglobal->clogGroupFirst);
435 
436  while (true)
437  {
438  /*
439  * Add the proc to list, if the clog page where we need to update the
440  * current transaction status is same as group leader's clog page.
441  *
442  * There is a race condition here, which is that after doing the below
443  * check and before adding this proc's clog update to a group, the
444  * group leader might have already finished the group update for this
445  * page and becomes group leader of another group. This will lead to a
446  * situation where a single group can have different clog page
447  * updates. This isn't likely and will still work, just maybe a bit
448  * less efficiently.
449  */
450  if (nextidx != INVALID_PGPROCNO &&
452  {
453  proc->clogGroupMember = false;
454  return false;
455  }
456 
457  pg_atomic_write_u32(&proc->clogGroupNext, nextidx);
458 
460  &nextidx,
461  (uint32) proc->pgprocno))
462  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  if (nextidx != INVALID_PGPROCNO)
472  {
473  int extraWaits = 0;
474 
475  /* Sleep until the leader updates our XID status. */
477  for (;;)
478  {
479  /* acts as a read barrier */
480  PGSemaphoreLock(proc->sem);
481  if (!proc->clogGroupMember)
482  break;
483  extraWaits++;
484  }
486 
488 
489  /* Fix semaphore count for any absorbed wakeups */
490  while (extraWaits-- > 0)
491  PGSemaphoreUnlock(proc->sem);
492  return true;
493  }
494 
495  /* We are the leader. Acquire the lock on behalf of everyone. */
496  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  nextidx = pg_atomic_exchange_u32(&procglobal->clogGroupFirst,
505 
506  /* Remember head of list so we can perform wakeups after dropping lock. */
507  wakeidx = nextidx;
508 
509  /* Walk the list and update the status of all XIDs. */
510  while (nextidx != INVALID_PGPROCNO)
511  {
512  PGPROC *proc = &ProcGlobal->allProcs[nextidx];
513 
514  /*
515  * Transactions with more than THRESHOLD_SUBTRANS_CLOG_OPT sub-XIDs
516  * should not use group XID status update mechanism.
517  */
519 
521  proc->subxidStatus.count,
522  proc->subxids.xids,
524  proc->clogGroupMemberLsn,
525  proc->clogGroupMemberPage);
526 
527  /* Move to next proc in list. */
528  nextidx = pg_atomic_read_u32(&proc->clogGroupNext);
529  }
530 
531  /* We're done with the lock now. */
532  LWLockRelease(XactSLRULock);
533 
534  /*
535  * Now that we've released the lock, go back and wake everybody up. We
536  * don't do this under the lock so as to keep lock hold times to a
537  * minimum.
538  */
539  while (wakeidx != INVALID_PGPROCNO)
540  {
541  PGPROC *proc = &ProcGlobal->allProcs[wakeidx];
542 
543  wakeidx = pg_atomic_read_u32(&proc->clogGroupNext);
545 
546  /* ensure all previous writes are visible before follower continues. */
548 
549  proc->clogGroupMember = false;
550 
551  if (proc != MyProc)
552  PGSemaphoreUnlock(proc->sem);
553  }
554 
555  return true;
556 }
557 
558 /*
559  * Sets the commit status of a single transaction.
560  *
561  * Must be called with XactSLRULock held
562  */
563 static void
565 {
566  int byteno = TransactionIdToByte(xid);
567  int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
568  char *byteptr;
569  char byteval;
570  char curval;
571 
572  byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
573  curval = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
574 
575  /*
576  * When replaying transactions during recovery we still need to perform
577  * the two phases of subcommit and then commit. However, some transactions
578  * are already correctly marked, so we just treat those as a no-op which
579  * allows us to keep the following Assert as restrictive as possible.
580  */
581  if (InRecovery && status == TRANSACTION_STATUS_SUB_COMMITTED &&
583  return;
584 
585  /*
586  * Current state change should be from 0 or subcommitted to target state
587  * or we should already be there when replaying changes during recovery.
588  */
589  Assert(curval == 0 ||
591  status != TRANSACTION_STATUS_IN_PROGRESS) ||
592  curval == status);
593 
594  /* note this assumes exclusive access to the clog page */
595  byteval = *byteptr;
596  byteval &= ~(((1 << CLOG_BITS_PER_XACT) - 1) << bshift);
597  byteval |= (status << bshift);
598  *byteptr = byteval;
599 
600  /*
601  * Update the group LSN if the transaction completion LSN is higher.
602  *
603  * Note: lsn will be invalid when supplied during InRecovery processing,
604  * so we don't need to do anything special to avoid LSN updates during
605  * recovery. After recovery completes the next clog change will set the
606  * LSN correctly.
607  */
608  if (!XLogRecPtrIsInvalid(lsn))
609  {
610  int lsnindex = GetLSNIndex(slotno, xid);
611 
612  if (XactCtl->shared->group_lsn[lsnindex] < lsn)
613  XactCtl->shared->group_lsn[lsnindex] = lsn;
614  }
615 }
616 
617 /*
618  * Interrogate the state of a transaction in the commit log.
619  *
620  * Aside from the actual commit status, this function returns (into *lsn)
621  * an LSN that is late enough to be able to guarantee that if we flush up to
622  * that LSN then we will have flushed the transaction's commit record to disk.
623  * The result is not necessarily the exact LSN of the transaction's commit
624  * record! For example, for long-past transactions (those whose clog pages
625  * already migrated to disk), we'll return InvalidXLogRecPtr. Also, because
626  * we group transactions on the same clog page to conserve storage, we might
627  * return the LSN of a later transaction that falls into the same group.
628  *
629  * NB: this is a low-level routine and is NOT the preferred entry point
630  * for most uses; TransactionLogFetch() in transam.c is the intended caller.
631  */
632 XidStatus
634 {
635  int pageno = TransactionIdToPage(xid);
636  int byteno = TransactionIdToByte(xid);
637  int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
638  int slotno;
639  int lsnindex;
640  char *byteptr;
642 
643  /* lock is acquired by SimpleLruReadPage_ReadOnly */
644 
645  slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
646  byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
647 
648  status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
649 
650  lsnindex = GetLSNIndex(slotno, xid);
651  *lsn = XactCtl->shared->group_lsn[lsnindex];
652 
653  LWLockRelease(XactSLRULock);
654 
655  return status;
656 }
657 
658 /*
659  * Number of shared CLOG buffers.
660  *
661  * On larger multi-processor systems, it is possible to have many CLOG page
662  * requests in flight at one time which could lead to disk access for CLOG
663  * page if the required page is not found in memory. Testing revealed that we
664  * can get the best performance by having 128 CLOG buffers, more than that it
665  * doesn't improve performance.
666  *
667  * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
668  * a good idea, because it would increase the minimum amount of shared memory
669  * required to start, which could be a problem for people running very small
670  * configurations. The following formula seems to represent a reasonable
671  * compromise: people with very low values for shared_buffers will get fewer
672  * CLOG buffers as well, and everyone else will get 128.
673  */
674 Size
676 {
677  return Min(128, Max(4, NBuffers / 512));
678 }
679 
680 /*
681  * Initialization of shared memory for CLOG
682  */
683 Size
685 {
687 }
688 
689 void
691 {
692  XactCtl->PagePrecedes = CLOGPagePrecedes;
694  XactSLRULock, "pg_xact", LWTRANCHE_XACT_BUFFER);
695 }
696 
697 /*
698  * This func must be called ONCE on system install. It creates
699  * the initial CLOG segment. (The CLOG directory is assumed to
700  * have been created by initdb, and CLOGShmemInit must have been
701  * called already.)
702  */
703 void
705 {
706  int slotno;
707 
708  LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
709 
710  /* Create and zero the first page of the commit log */
711  slotno = ZeroCLOGPage(0, false);
712 
713  /* Make sure it's written out */
714  SimpleLruWritePage(XactCtl, slotno);
715  Assert(!XactCtl->shared->page_dirty[slotno]);
716 
717  LWLockRelease(XactSLRULock);
718 }
719 
720 /*
721  * Initialize (or reinitialize) a page of CLOG to zeroes.
722  * If writeXlog is true, also emit an XLOG record saying we did this.
723  *
724  * The page is not actually written, just set up in shared memory.
725  * The slot number of the new page is returned.
726  *
727  * Control lock must be held at entry, and will be held at exit.
728  */
729 static int
730 ZeroCLOGPage(int pageno, bool writeXlog)
731 {
732  int slotno;
733 
734  slotno = SimpleLruZeroPage(XactCtl, pageno);
735 
736  if (writeXlog)
737  WriteZeroPageXlogRec(pageno);
738 
739  return slotno;
740 }
741 
742 /*
743  * This must be called ONCE during postmaster or standalone-backend startup,
744  * after StartupXLOG has initialized ShmemVariableCache->nextXid.
745  */
746 void
748 {
750  int pageno = TransactionIdToPage(xid);
751 
752  LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
753 
754  /*
755  * Initialize our idea of the latest page number.
756  */
757  XactCtl->shared->latest_page_number = pageno;
758 
759  LWLockRelease(XactSLRULock);
760 }
761 
762 /*
763  * This must be called ONCE at the end of startup/recovery.
764  */
765 void
766 TrimCLOG(void)
767 {
769  int pageno = TransactionIdToPage(xid);
770 
771  LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
772 
773  /*
774  * Re-Initialize our idea of the latest page number.
775  */
776  XactCtl->shared->latest_page_number = pageno;
777 
778  /*
779  * Zero out the remainder of the current clog page. Under normal
780  * circumstances it should be zeroes already, but it seems at least
781  * theoretically possible that XLOG replay will have settled on a nextXID
782  * value that is less than the last XID actually used and marked by the
783  * previous database lifecycle (since subtransaction commit writes clog
784  * but makes no WAL entry). Let's just be safe. (We need not worry about
785  * pages beyond the current one, since those will be zeroed when first
786  * used. For the same reason, there is no need to do anything when
787  * nextXid is exactly at a page boundary; and it's likely that the
788  * "current" page doesn't exist yet in that case.)
789  */
790  if (TransactionIdToPgIndex(xid) != 0)
791  {
792  int byteno = TransactionIdToByte(xid);
793  int bshift = TransactionIdToBIndex(xid) * CLOG_BITS_PER_XACT;
794  int slotno;
795  char *byteptr;
796 
797  slotno = SimpleLruReadPage(XactCtl, pageno, false, xid);
798  byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
799 
800  /* Zero so-far-unused positions in the current byte */
801  *byteptr &= (1 << bshift) - 1;
802  /* Zero the rest of the page */
803  MemSet(byteptr + 1, 0, BLCKSZ - byteno - 1);
804 
805  XactCtl->shared->page_dirty[slotno] = true;
806  }
807 
808  LWLockRelease(XactSLRULock);
809 }
810 
811 /*
812  * This must be called ONCE during postmaster or standalone-backend shutdown
813  */
814 void
816 {
817  /* Flush dirty CLOG pages to disk */
818  TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(false);
819  SimpleLruFlush(XactCtl, false);
820 
821  /*
822  * fsync pg_xact to ensure that any files flushed previously are durably
823  * on disk.
824  */
825  fsync_fname("pg_xact", true);
826 
827  TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(false);
828 }
829 
830 /*
831  * Perform a checkpoint --- either during shutdown, or on-the-fly
832  */
833 void
835 {
836  /* Flush dirty CLOG pages to disk */
837  TRACE_POSTGRESQL_CLOG_CHECKPOINT_START(true);
838  SimpleLruFlush(XactCtl, true);
839 
840  /*
841  * fsync pg_xact to ensure that any files flushed previously are durably
842  * on disk.
843  */
844  fsync_fname("pg_xact", true);
845 
846  TRACE_POSTGRESQL_CLOG_CHECKPOINT_DONE(true);
847 }
848 
849 
850 /*
851  * Make sure that CLOG has room for a newly-allocated XID.
852  *
853  * NB: this is called while holding XidGenLock. We want it to be very fast
854  * most of the time; even when it's not so fast, no actual I/O need happen
855  * unless we're forced to write out a dirty clog or xlog page to make room
856  * in shared memory.
857  */
858 void
860 {
861  int pageno;
862 
863  /*
864  * No work except at first XID of a page. But beware: just after
865  * wraparound, the first XID of page zero is FirstNormalTransactionId.
866  */
867  if (TransactionIdToPgIndex(newestXact) != 0 &&
869  return;
870 
871  pageno = TransactionIdToPage(newestXact);
872 
873  LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
874 
875  /* Zero the page and make an XLOG entry about it */
876  ZeroCLOGPage(pageno, true);
877 
878  LWLockRelease(XactSLRULock);
879 }
880 
881 
882 /*
883  * Remove all CLOG segments before the one holding the passed transaction ID
884  *
885  * Before removing any CLOG data, we must flush XLOG to disk, to ensure
886  * that any recently-emitted FREEZE_PAGE records have reached disk; otherwise
887  * a crash and restart might leave us with some unfrozen tuples referencing
888  * removed CLOG data. We choose to emit a special TRUNCATE XLOG record too.
889  * Replaying the deletion from XLOG is not critical, since the files could
890  * just as well be removed later, but doing so prevents a long-running hot
891  * standby server from acquiring an unreasonably bloated CLOG directory.
892  *
893  * Since CLOG segments hold a large number of transactions, the opportunity to
894  * actually remove a segment is fairly rare, and so it seems best not to do
895  * the XLOG flush unless we have confirmed that there is a removable segment.
896  */
897 void
898 TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
899 {
900  int cutoffPage;
901 
902  /*
903  * The cutoff point is the start of the segment containing oldestXact. We
904  * pass the *page* containing oldestXact to SimpleLruTruncate.
905  */
906  cutoffPage = TransactionIdToPage(oldestXact);
907 
908  /* Check to see if there's any files that could be removed */
910  return; /* nothing to remove */
911 
912  /*
913  * Advance oldestClogXid before truncating clog, so concurrent xact status
914  * lookups can ensure they don't attempt to access truncated-away clog.
915  *
916  * It's only necessary to do this if we will actually truncate away clog
917  * pages.
918  */
919  AdvanceOldestClogXid(oldestXact);
920 
921  /*
922  * Write XLOG record and flush XLOG to disk. We record the oldest xid
923  * we're keeping information about here so we can ensure that it's always
924  * ahead of clog truncation in case we crash, and so a standby finds out
925  * the new valid xid before the next checkpoint.
926  */
927  WriteTruncateXlogRec(cutoffPage, oldestXact, oldestxid_datoid);
928 
929  /* Now we can remove the old CLOG segment(s) */
930  SimpleLruTruncate(XactCtl, cutoffPage);
931 }
932 
933 
934 /*
935  * Decide which of two CLOG page numbers is "older" for truncation purposes.
936  *
937  * We need to use comparison of TransactionIds here in order to do the right
938  * thing with wraparound XID arithmetic. However, if we are asked about
939  * page number zero, we don't want to hand InvalidTransactionId to
940  * TransactionIdPrecedes: it'll get weird about permanent xact IDs. So,
941  * offset both xids by FirstNormalTransactionId to avoid that.
942  */
943 static bool
944 CLOGPagePrecedes(int page1, int page2)
945 {
946  TransactionId xid1;
947  TransactionId xid2;
948 
949  xid1 = ((TransactionId) page1) * CLOG_XACTS_PER_PAGE;
950  xid1 += FirstNormalTransactionId;
951  xid2 = ((TransactionId) page2) * CLOG_XACTS_PER_PAGE;
952  xid2 += FirstNormalTransactionId;
953 
954  return TransactionIdPrecedes(xid1, xid2);
955 }
956 
957 
958 /*
959  * Write a ZEROPAGE xlog record
960  */
961 static void
963 {
964  XLogBeginInsert();
965  XLogRegisterData((char *) (&pageno), sizeof(int));
966  (void) XLogInsert(RM_CLOG_ID, CLOG_ZEROPAGE);
967 }
968 
969 /*
970  * Write a TRUNCATE xlog record
971  *
972  * We must flush the xlog record to disk before returning --- see notes
973  * in TruncateCLOG().
974  */
975 static void
976 WriteTruncateXlogRec(int pageno, TransactionId oldestXact, Oid oldestXactDb)
977 {
978  XLogRecPtr recptr;
979  xl_clog_truncate xlrec;
980 
981  xlrec.pageno = pageno;
982  xlrec.oldestXact = oldestXact;
983  xlrec.oldestXactDb = oldestXactDb;
984 
985  XLogBeginInsert();
986  XLogRegisterData((char *) (&xlrec), sizeof(xl_clog_truncate));
987  recptr = XLogInsert(RM_CLOG_ID, CLOG_TRUNCATE);
988  XLogFlush(recptr);
989 }
990 
991 /*
992  * CLOG resource manager's routines
993  */
994 void
996 {
997  uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
998 
999  /* Backup blocks are not used in clog records */
1000  Assert(!XLogRecHasAnyBlockRefs(record));
1001 
1002  if (info == CLOG_ZEROPAGE)
1003  {
1004  int pageno;
1005  int slotno;
1006 
1007  memcpy(&pageno, XLogRecGetData(record), sizeof(int));
1008 
1009  LWLockAcquire(XactSLRULock, LW_EXCLUSIVE);
1010 
1011  slotno = ZeroCLOGPage(pageno, false);
1012  SimpleLruWritePage(XactCtl, slotno);
1013  Assert(!XactCtl->shared->page_dirty[slotno]);
1014 
1015  LWLockRelease(XactSLRULock);
1016  }
1017  else if (info == CLOG_TRUNCATE)
1018  {
1019  xl_clog_truncate xlrec;
1020 
1021  memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_clog_truncate));
1022 
1023  /*
1024  * During XLOG replay, latest_page_number isn't set up yet; insert a
1025  * suitable value to bypass the sanity test in SimpleLruTruncate.
1026  */
1027  XactCtl->shared->latest_page_number = xlrec.pageno;
1028 
1030 
1032  }
1033  else
1034  elog(PANIC, "clog_redo: unknown op code %u", info);
1035 }
#define TransactionIdToBIndex(xid)
Definition: clog.c:67
#define CLOG_XACT_BITMASK
Definition: clog.c:62
#define CLOG_XACTS_PER_PAGE
Definition: clog.c:61
#define TransactionIdToPage(xid)
Definition: clog.c:64
XidStatus clogGroupMemberXidStatus
Definition: proc.h:220
bool LWLockHeldByMeInMode(LWLock *l, LWLockMode mode)
Definition: lwlock.c:1946
static void WriteZeroPageXlogRec(int pageno)
Definition: clog.c:962
void TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
Definition: clog.c:898
#define TransactionIdEquals(id1, id2)
Definition: transam.h:43
#define TRANSACTION_STATUS_COMMITTED
Definition: clog.h:27
uint32 TransactionId
Definition: c.h:520
#define GetLSNIndex(slotno, xid)
Definition: clog.c:73
void PGSemaphoreUnlock(PGSemaphore sema)
Definition: posix_sema.c:340
void clog_redo(XLogReaderState *record)
Definition: clog.c:995
static bool TransactionGroupUpdateXidStatus(TransactionId xid, XidStatus status, XLogRecPtr lsn, int pageno)
Definition: clog.c:413
XLogRecPtr clogGroupMemberLsn
Definition: proc.h:224
static bool pg_atomic_compare_exchange_u32(volatile pg_atomic_uint32 *ptr, uint32 *expected, uint32 newval)
Definition: atomics.h:311
PGPROC * MyProc
Definition: proc.c:67
static void TransactionIdSetStatusBit(TransactionId xid, XidStatus status, XLogRecPtr lsn, int slotno)
Definition: clog.c:564
void SimpleLruTruncate(SlruCtl ctl, int cutoffPage)
Definition: slru.c:1196
void ExtendCLOG(TransactionId newestXact)
Definition: clog.c:859
bool InRecovery
Definition: xlog.c:204
#define Min(x, y)
Definition: c.h:927
#define TransactionIdToPgIndex(xid)
Definition: clog.c:65
void ShutdownCLOG(void)
Definition: clog.c:815
void fsync_fname(const char *fname, bool isdir)
Definition: fd.c:632
#define XactCtl
Definition: clog.c:88
unsigned char uint8
Definition: c.h:372
int XidStatus
Definition: clog.h:24
PROC_HDR * ProcGlobal
Definition: proc.c:79
#define MemSet(start, val, len)
Definition: c.h:949
pg_atomic_uint32 clogGroupNext
Definition: proc.h:218
unsigned int Oid
Definition: postgres_ext.h:31
FullTransactionId nextXid
Definition: transam.h:213
#define PANIC
Definition: elog.h:53
Size SimpleLruShmemSize(int nslots, int nlsns)
Definition: slru.c:144
void SimpleLruFlush(SlruCtl ctl, bool allow_redirtied)
Definition: slru.c:1128
void XLogFlush(XLogRecPtr record)
Definition: xlog.c:2844
XidCacheStatus subxidStatus
Definition: proc.h:198
#define XidFromFullTransactionId(x)
Definition: transam.h:48
void LWLockRelease(LWLock *lock)
Definition: lwlock.c:1812
static uint32 pg_atomic_exchange_u32(volatile pg_atomic_uint32 *ptr, uint32 newval)
Definition: atomics.h:292
#define CLOG_LSNS_PER_PAGE
Definition: clog.c:71
#define StaticAssertStmt(condition, errmessage)
Definition: c.h:859
#define CLOG_ZEROPAGE
Definition: clog.h:54
void AdvanceOldestClogXid(TransactionId oldest_datfrozenxid)
Definition: varsup.c:328
void TransactionIdSetTreeStatus(TransactionId xid, int nsubxids, TransactionId *subxids, XidStatus status, XLogRecPtr lsn)
Definition: clog.c:163
#define XLogRecGetData(decoder)
Definition: xlogreader.h:310
static void WriteTruncateXlogRec(int pageno, TransactionId oldestXact, Oid oldestXactDb)
Definition: clog.c:976
#define FirstNormalTransactionId
Definition: transam.h:34
int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok, TransactionId xid)
Definition: slru.c:382
#define TRANSACTION_STATUS_ABORTED
Definition: clog.h:28
#define THRESHOLD_SUBTRANS_CLOG_OPT
Definition: clog.c:81
static bool CLOGPagePrecedes(int page1, int page2)
Definition: clog.c:944
bool LWLockConditionalAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1380
int clogGroupMemberPage
Definition: proc.h:222
VariableCache ShmemVariableCache
Definition: varsup.c:34
Size CLOGShmemBuffers(void)
Definition: clog.c:675
#define TRANSACTION_STATUS_SUB_COMMITTED
Definition: clog.h:29
#define InvalidTransactionId
Definition: transam.h:31
TransactionId clogGroupMemberXid
Definition: proc.h:219
unsigned int uint32
Definition: c.h:374
#define CLOG_BITS_PER_XACT
Definition: clog.c:59
static SlruCtlData XactCtlData
Definition: clog.c:86
void SimpleLruWritePage(SlruCtl ctl, int slotno)
Definition: slru.c:597
int pageno
Definition: clog.h:33
static void pgstat_report_wait_end(void)
Definition: pgstat.h:1381
struct XidCache subxids
Definition: proc.h:200
void CheckPointCLOG(void)
Definition: clog.c:834
#define CLOG_TRUNCATE
Definition: clog.h:55
#define XLogRecGetInfo(decoder)
Definition: xlogreader.h:305
#define INVALID_PGPROCNO
Definition: proc.h:76
bool TransactionIdPrecedes(TransactionId id1, TransactionId id2)
Definition: transam.c:300
Definition: proc.h:303
#define XLogRecPtrIsInvalid(r)
Definition: xlogdefs.h:29
void XLogRegisterData(char *data, int len)
Definition: xloginsert.c:330
TransactionId oldestXact
Definition: clog.h:34
XLogRecPtr XLogInsert(RmgrId rmid, uint8 info)
Definition: xloginsert.c:422
bool SlruScanDirCbReportPresence(SlruCtl ctl, char *filename, int segpage, void *data)
Definition: slru.c:1352
int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
Definition: slru.c:482
Size CLOGShmemSize(void)
Definition: clog.c:684
static void set_status_by_pages(int nsubxids, TransactionId *subxids, XidStatus status, XLogRecPtr lsn)
Definition: clog.c:237
#define PGPROC_MAX_CACHED_SUBXIDS
Definition: proc.h:36
void BootStrapCLOG(void)
Definition: clog.c:704
#define Max(x, y)
Definition: c.h:921
TransactionId xids[PGPROC_MAX_CACHED_SUBXIDS]
Definition: proc.h:48
uint64 XLogRecPtr
Definition: xlogdefs.h:21
#define Assert(condition)
Definition: c.h:745
#define XLR_INFO_MASK
Definition: xlogrecord.h:62
bool SlruScanDirectory(SlruCtl ctl, SlruScanCallback callback, void *data)
Definition: slru.c:1407
size_t Size
Definition: c.h:473
static void pgstat_report_wait_start(uint32 wait_event_info)
Definition: pgstat.h:1357
static void TransactionIdSetPageStatus(TransactionId xid, int nsubxids, TransactionId *subxids, XidStatus status, XLogRecPtr lsn, int pageno, bool all_xact_same_page)
Definition: clog.c:273
bool LWLockAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1208
void StartupCLOG(void)
Definition: clog.c:747
void PGSemaphoreLock(PGSemaphore sema)
Definition: posix_sema.c:320
XidStatus TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
Definition: clog.c:633
uint8 count
Definition: proc.h:41
TransactionId xid
Definition: proc.h:124
pg_atomic_uint32 clogGroupFirst
Definition: proc.h:336
#define TransactionIdToByte(xid)
Definition: clog.c:66
int pgprocno
Definition: proc.h:141
static int ZeroCLOGPage(int pageno, bool writeXlog)
Definition: clog.c:730
bool clogGroupMember
Definition: proc.h:217
#define elog(elevel,...)
Definition: elog.h:214
int i
void CLOGShmemInit(void)
Definition: clog.c:690
#define pg_write_barrier()
Definition: atomics.h:159
int NBuffers
Definition: globals.c:132
Oid oldestXactDb
Definition: clog.h:35
static void TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids, TransactionId *subxids, XidStatus status, XLogRecPtr lsn, int pageno)
Definition: clog.c:338
#define XLogRecHasAnyBlockRefs(decoder)
Definition: xlogreader.h:312
PGPROC * allProcs
Definition: proc.h:306
#define TransactionIdIsValid(xid)
Definition: transam.h:41
static void pg_atomic_write_u32(volatile pg_atomic_uint32 *ptr, uint32 val)
Definition: atomics.h:258
PGSemaphore sem
Definition: proc.h:118
static void static void status(const char *fmt,...) pg_attribute_printf(1
Definition: pg_regress.c:225
void XLogBeginInsert(void)
Definition: xloginsert.c:123
Definition: proc.h:112
int SimpleLruZeroPage(SlruCtl ctl, int pageno)
Definition: slru.c:267
void TrimCLOG(void)
Definition: clog.c:766
#define TRANSACTION_STATUS_IN_PROGRESS
Definition: clog.h:26
static uint32 pg_atomic_read_u32(volatile pg_atomic_uint32 *ptr)
Definition: atomics.h:241
void SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns, LWLock *ctllock, const char *subdir, int tranche_id)
Definition: slru.c:175