1 /*-------------------------------------------------------------------------
5 * The WAL sender process (walsender) is new as of Postgres 9.0. It takes
6 * care of sending XLOG from the primary server to a single recipient.
7 * (Note that there can be more than one walsender process concurrently.)
8 * It is started by the postmaster when the walreceiver of a standby server
9 * connects to the primary server and requests XLOG streaming replication.
10 * It attempts to keep reading XLOG records from the disk and sending them
11 * to the standby server, as long as the connection is alive (i.e., like
12 * any backend, there is a one-to-one relationship between a connection
13 * and a walsender process).
15 * Normal termination is by SIGTERM, which instructs the walsender to
16 * close the connection and exit(0) at next convenient moment. Emergency
17 * termination is by SIGQUIT; like any backend, the walsender will simply
18 * abort and exit on SIGQUIT. A close of the connection and a FATAL error
19 * are treated as not a crash but approximately normal termination;
20 * the walsender will exit quickly without sending any more XLOG records.
22 * If the server is shut down, postmaster sends us SIGUSR2 after all
23 * regular backends have exited and the shutdown checkpoint has been written.
24 * This instruct walsender to send any outstanding WAL, including the
25 * shutdown checkpoint record, and then exit.
28 * Portions Copyright (c) 2010-2012, PostgreSQL Global Development Group
31 * src/backend/replication/walsender.c
33 *-------------------------------------------------------------------------
40 #include "access/transam.h"
41 #include "access/xlog_internal.h"
42 #include "catalog/pg_type.h"
44 #include "libpq/libpq.h"
45 #include "libpq/pqformat.h"
46 #include "libpq/pqsignal.h"
47 #include "miscadmin.h"
48 #include "nodes/replnodes.h"
49 #include "replication/basebackup.h"
50 #include "replication/syncrep.h"
51 #include "replication/walprotocol.h"
52 #include "replication/walreceiver.h"
53 #include "replication/walsender.h"
54 #include "replication/walsender_private.h"
55 #include "storage/fd.h"
56 #include "storage/ipc.h"
57 #include "storage/pmsignal.h"
58 #include "storage/proc.h"
59 #include "storage/procarray.h"
60 #include "tcop/tcopprot.h"
61 #include "utils/builtins.h"
62 #include "utils/guc.h"
63 #include "utils/memutils.h"
64 #include "utils/ps_status.h"
65 #include "utils/resowner.h"
66 #include "utils/timestamp.h"
69 /* Array of WalSnds in shared memory */
70 WalSndCtlData *WalSndCtl = NULL;
72 /* My slot in the shared memory array */
73 WalSnd *MyWalSnd = NULL;
76 bool am_walsender = false; /* Am I a walsender process ? */
77 bool am_cascading_walsender = false; /* Am I cascading WAL to
78 * another standby ? */
80 /* User-settable parameters for walsender */
81 int max_wal_senders = 0; /* the maximum number of concurrent walsenders */
82 int replication_timeout = 60 * 1000; /* maximum time to send one
86 * These variables are used similarly to openLogFile/Id/Seg/Off,
87 * but for walsender to read the XLOG.
89 static int sendFile = -1;
90 static XLogSegNo sendSegNo = 0;
91 static uint32 sendOff = 0;
94 * How far have we sent WAL already? This is also advertised in
95 * MyWalSnd->sentPtr. (Actually, this is the next WAL location to send.)
97 static XLogRecPtr sentPtr = 0;
100 * Buffer for processing reply messages.
102 static StringInfoData reply_message;
105 * Timestamp of the last receipt of the reply from the standby.
107 static TimestampTz last_reply_timestamp;
109 /* Flags set by signal handlers for later service in main loop */
110 static volatile sig_atomic_t got_SIGHUP = false;
111 volatile sig_atomic_t walsender_shutdown_requested = false;
112 volatile sig_atomic_t walsender_ready_to_stop = false;
114 /* Signal handlers */
115 static void WalSndSigHupHandler(SIGNAL_ARGS);
116 static void WalSndShutdownHandler(SIGNAL_ARGS);
117 static void WalSndQuickDieHandler(SIGNAL_ARGS);
118 static void WalSndXLogSendHandler(SIGNAL_ARGS);
119 static void WalSndLastCycleHandler(SIGNAL_ARGS);
121 /* Prototypes for private functions */
122 static bool HandleReplicationCommand(const char *cmd_string);
123 static void WalSndLoop(void) __attribute__((noreturn));
124 static void InitWalSnd(void);
125 static void WalSndHandshake(void);
126 static void WalSndKill(int code, Datum arg);
127 static void XLogSend(char *msgbuf, bool *caughtup);
128 static void IdentifySystem(void);
129 static void StartReplication(StartReplicationCmd *cmd);
130 static void ProcessStandbyMessage(void);
131 static void ProcessStandbyReplyMessage(void);
132 static void ProcessStandbyHSFeedbackMessage(void);
133 static void ProcessRepliesIfAny(void);
134 static void WalSndKeepalive(char *msgbuf);
137 /* Main entry point for walsender process */
141 MemoryContext walsnd_context;
143 am_cascading_walsender = RecoveryInProgress();
145 /* Create a per-walsender data structure in shared memory */
149 * Create a memory context that we will do all our work in. We do this so
150 * that we can reset the context during error recovery and thereby avoid
151 * possible memory leaks. Formerly this code just ran in
152 * TopMemoryContext, but resetting that would be a really bad idea.
154 * XXX: we don't actually attempt error recovery in walsender, we just
155 * close the connection and exit.
157 walsnd_context = AllocSetContextCreate(TopMemoryContext,
159 ALLOCSET_DEFAULT_MINSIZE,
160 ALLOCSET_DEFAULT_INITSIZE,
161 ALLOCSET_DEFAULT_MAXSIZE);
162 MemoryContextSwitchTo(walsnd_context);
164 /* Set up resource owner */
165 CurrentResourceOwner = ResourceOwnerCreate(NULL, "walsender top-level resource owner");
167 /* Unblock signals (they were blocked when the postmaster forked us) */
168 PG_SETMASK(&UnBlockSig);
171 * Use the recovery target timeline ID during recovery
173 if (am_cascading_walsender)
174 ThisTimeLineID = GetRecoveryTargetTLI();
176 /* Tell the standby that walsender is ready for receiving commands */
177 ReadyForQuery(DestRemote);
179 /* Handle handshake messages before streaming */
182 /* Initialize shared memory status */
184 /* use volatile pointer to prevent code rearrangement */
185 volatile WalSnd *walsnd = MyWalSnd;
187 SpinLockAcquire(&walsnd->mutex);
188 walsnd->sentPtr = sentPtr;
189 SpinLockRelease(&walsnd->mutex);
194 /* Main loop of walsender */
199 * Execute commands from walreceiver, until we enter streaming mode.
202 WalSndHandshake(void)
204 StringInfoData input_message;
205 bool replication_started = false;
207 initStringInfo(&input_message);
209 while (!replication_started)
213 WalSndSetState(WALSNDSTATE_STARTUP);
214 set_ps_display("idle", false);
216 /* Wait for a command to arrive */
217 firstchar = pq_getbyte();
220 * Emergency bailout if postmaster has died. This is to avoid the
221 * necessity for manual cleanup of all postmaster children.
223 if (!PostmasterIsAlive())
227 * Check for any other interesting events that happened while we
233 ProcessConfigFile(PGC_SIGHUP);
236 if (firstchar != EOF)
239 * Read the message contents. This is expected to be done without
240 * blocking because we've been able to get message type code.
242 if (pq_getmessage(&input_message, 0))
243 firstchar = EOF; /* suitable message already logged */
246 /* Handle the very limited subset of commands expected in this phase */
249 case 'Q': /* Query message */
251 const char *query_string;
253 query_string = pq_getmsgstring(&input_message);
254 pq_getmsgend(&input_message);
256 if (HandleReplicationCommand(query_string))
257 replication_started = true;
262 /* standby is closing the connection */
266 /* standby disconnected unexpectedly */
268 (errcode(ERRCODE_PROTOCOL_VIOLATION),
269 errmsg("unexpected EOF on standby connection")));
274 (errcode(ERRCODE_PROTOCOL_VIOLATION),
275 errmsg("invalid standby handshake message type %d", firstchar)));
289 char xpos[MAXFNAMELEN];
293 * Reply with a result set with one row, three columns. First col is
294 * system ID, second is timeline ID, and third is current xlog location.
297 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
298 GetSystemIdentifier());
299 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
301 logptr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetInsertRecPtr();
303 snprintf(xpos, sizeof(xpos), "%X/%X", (uint32) (logptr >> 32), (uint32) logptr);
305 /* Send a RowDescription message */
306 pq_beginmessage(&buf, 'T');
307 pq_sendint(&buf, 3, 2); /* 3 fields */
310 pq_sendstring(&buf, "systemid"); /* col name */
311 pq_sendint(&buf, 0, 4); /* table oid */
312 pq_sendint(&buf, 0, 2); /* attnum */
313 pq_sendint(&buf, TEXTOID, 4); /* type oid */
314 pq_sendint(&buf, -1, 2); /* typlen */
315 pq_sendint(&buf, 0, 4); /* typmod */
316 pq_sendint(&buf, 0, 2); /* format code */
319 pq_sendstring(&buf, "timeline"); /* col name */
320 pq_sendint(&buf, 0, 4); /* table oid */
321 pq_sendint(&buf, 0, 2); /* attnum */
322 pq_sendint(&buf, INT4OID, 4); /* type oid */
323 pq_sendint(&buf, 4, 2); /* typlen */
324 pq_sendint(&buf, 0, 4); /* typmod */
325 pq_sendint(&buf, 0, 2); /* format code */
328 pq_sendstring(&buf, "xlogpos");
329 pq_sendint(&buf, 0, 4);
330 pq_sendint(&buf, 0, 2);
331 pq_sendint(&buf, TEXTOID, 4);
332 pq_sendint(&buf, -1, 2);
333 pq_sendint(&buf, 0, 4);
334 pq_sendint(&buf, 0, 2);
337 /* Send a DataRow message */
338 pq_beginmessage(&buf, 'D');
339 pq_sendint(&buf, 3, 2); /* # of columns */
340 pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
341 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
342 pq_sendint(&buf, strlen(tli), 4); /* col2 len */
343 pq_sendbytes(&buf, (char *) tli, strlen(tli));
344 pq_sendint(&buf, strlen(xpos), 4); /* col3 len */
345 pq_sendbytes(&buf, (char *) xpos, strlen(xpos));
349 /* Send CommandComplete and ReadyForQuery messages */
350 EndCommand("SELECT", DestRemote);
351 ReadyForQuery(DestRemote);
352 /* ReadyForQuery did pq_flush for us */
359 StartReplication(StartReplicationCmd *cmd)
364 * Let postmaster know that we're streaming. Once we've declared us as a
365 * WAL sender process, postmaster will let us outlive the bgwriter and
366 * kill us last in the shutdown sequence, so we get a chance to stream all
367 * remaining WAL at shutdown, including the shutdown checkpoint. Note that
368 * there's no going back, and we mustn't write any WAL records after this.
370 MarkPostmasterChildWalSender();
371 SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
374 * When promoting a cascading standby, postmaster sends SIGUSR2 to any
375 * cascading walsenders to kill them. But there is a corner-case where
376 * such walsender fails to receive SIGUSR2 and survives a standby
377 * promotion unexpectedly. This happens when postmaster sends SIGUSR2
378 * before the walsender marks itself as a WAL sender, because postmaster
379 * sends SIGUSR2 to only the processes marked as a WAL sender.
381 * To avoid this corner-case, if recovery is NOT in progress even though
382 * the walsender is cascading one, we do the same thing as SIGUSR2 signal
383 * handler does, i.e., set walsender_ready_to_stop to true. Which causes
384 * the walsender to end later.
386 * When terminating cascading walsenders, usually postmaster writes the
387 * log message announcing the terminations. But there is a race condition
388 * here. If there is no walsender except this process before reaching
389 * here, postmaster thinks that there is no walsender and suppresses that
390 * log message. To handle this case, we always emit that log message here.
391 * This might cause duplicate log messages, but which is less likely to
392 * happen, so it's not worth writing some code to suppress them.
394 if (am_cascading_walsender && !RecoveryInProgress())
397 (errmsg("terminating walsender process to force cascaded standby "
398 "to update timeline and reconnect")));
399 walsender_ready_to_stop = true;
403 * We assume here that we're logging enough information in the WAL for
404 * log-shipping, since this is checked in PostmasterMain().
406 * NOTE: wal_level can only change at shutdown, so in most cases it is
407 * difficult for there to be WAL data that we can still see that was
408 * written at wal_level='minimal'.
412 * When we first start replication the standby will be behind the primary.
413 * For some applications, for example, synchronous replication, it is
414 * important to have a clear state for this initial catchup mode, so we
415 * can trigger actions when we change streaming state later. We may stay
416 * in this state for a long time, which is exactly why we want to be able
417 * to monitor whether or not we are still here.
419 WalSndSetState(WALSNDSTATE_CATCHUP);
421 /* Send a CopyBothResponse message, and start streaming */
422 pq_beginmessage(&buf, 'W');
423 pq_sendbyte(&buf, 0);
424 pq_sendint(&buf, 0, 2);
429 * Initialize position to the received one, then the xlog records begin to
430 * be shipped from that position
432 sentPtr = cmd->startpoint;
436 * Execute an incoming replication command.
439 HandleReplicationCommand(const char *cmd_string)
441 bool replication_started = false;
444 MemoryContext cmd_context;
445 MemoryContext old_context;
447 elog(DEBUG1, "received replication command: %s", cmd_string);
449 cmd_context = AllocSetContextCreate(CurrentMemoryContext,
450 "Replication command context",
451 ALLOCSET_DEFAULT_MINSIZE,
452 ALLOCSET_DEFAULT_INITSIZE,
453 ALLOCSET_DEFAULT_MAXSIZE);
454 old_context = MemoryContextSwitchTo(cmd_context);
456 replication_scanner_init(cmd_string);
457 parse_rc = replication_yyparse();
460 (errcode(ERRCODE_SYNTAX_ERROR),
461 (errmsg_internal("replication command parser returned %d",
464 cmd_node = replication_parse_result;
466 switch (cmd_node->type)
468 case T_IdentifySystemCmd:
472 case T_StartReplicationCmd:
473 StartReplication((StartReplicationCmd *) cmd_node);
475 /* break out of the loop */
476 replication_started = true;
479 case T_BaseBackupCmd:
480 SendBaseBackup((BaseBackupCmd *) cmd_node);
482 /* Send CommandComplete and ReadyForQuery messages */
483 EndCommand("SELECT", DestRemote);
484 ReadyForQuery(DestRemote);
485 /* ReadyForQuery did pq_flush for us */
490 (errcode(ERRCODE_PROTOCOL_VIOLATION),
491 errmsg("invalid standby query string: %s", cmd_string)));
495 MemoryContextSwitchTo(old_context);
496 MemoryContextDelete(cmd_context);
498 return replication_started;
502 * Check if the remote end has closed the connection.
505 ProcessRepliesIfAny(void)
507 unsigned char firstchar;
509 bool received = false;
513 r = pq_getbyte_if_available(&firstchar);
516 /* unexpected error or EOF */
518 (errcode(ERRCODE_PROTOCOL_VIOLATION),
519 errmsg("unexpected EOF on standby connection")));
524 /* no data available without blocking */
528 /* Handle the very limited subset of commands expected in this phase */
532 * 'd' means a standby reply wrapped in a CopyData packet.
535 ProcessStandbyMessage();
540 * 'X' means that the standby is closing down the socket.
547 (errcode(ERRCODE_PROTOCOL_VIOLATION),
548 errmsg("invalid standby message type \"%c\"",
554 * Save the last reply timestamp if we've received at least one reply.
557 last_reply_timestamp = GetCurrentTimestamp();
561 * Process a status update message received from standby.
564 ProcessStandbyMessage(void)
568 resetStringInfo(&reply_message);
571 * Read the message contents.
573 if (pq_getmessage(&reply_message, 0))
576 (errcode(ERRCODE_PROTOCOL_VIOLATION),
577 errmsg("unexpected EOF on standby connection")));
582 * Check message type from the first byte.
584 msgtype = pq_getmsgbyte(&reply_message);
589 ProcessStandbyReplyMessage();
593 ProcessStandbyHSFeedbackMessage();
598 (errcode(ERRCODE_PROTOCOL_VIOLATION),
599 errmsg("unexpected message type \"%c\"", msgtype)));
605 * Regular reply from standby advising of WAL positions on standby server.
608 ProcessStandbyReplyMessage(void)
610 StandbyReplyMessage reply;
612 pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyReplyMessage));
614 elog(DEBUG2, "write %X/%X flush %X/%X apply %X/%X",
615 (uint32) (reply.write >> 32), (uint32) reply.write,
616 (uint32) (reply.flush >> 32), (uint32) reply.flush,
617 (uint32) (reply.apply >> 32), (uint32) reply.apply);
620 * Update shared state for this WalSender process based on reply data from
624 /* use volatile pointer to prevent code rearrangement */
625 volatile WalSnd *walsnd = MyWalSnd;
627 SpinLockAcquire(&walsnd->mutex);
628 walsnd->write = reply.write;
629 walsnd->flush = reply.flush;
630 walsnd->apply = reply.apply;
631 SpinLockRelease(&walsnd->mutex);
634 if (!am_cascading_walsender)
635 SyncRepReleaseWaiters();
639 * Hot Standby feedback
642 ProcessStandbyHSFeedbackMessage(void)
644 StandbyHSFeedbackMessage reply;
645 TransactionId nextXid;
648 /* Decipher the reply message */
649 pq_copymsgbytes(&reply_message, (char *) &reply,
650 sizeof(StandbyHSFeedbackMessage));
652 elog(DEBUG2, "hot standby feedback xmin %u epoch %u",
656 /* Ignore invalid xmin (can't actually happen with current walreceiver) */
657 if (!TransactionIdIsNormal(reply.xmin))
661 * Check that the provided xmin/epoch are sane, that is, not in the future
662 * and not so far back as to be already wrapped around. Ignore if not.
664 * Epoch of nextXid should be same as standby, or if the counter has
665 * wrapped, then one greater than standby.
667 GetNextXidAndEpoch(&nextXid, &nextEpoch);
669 if (reply.xmin <= nextXid)
671 if (reply.epoch != nextEpoch)
676 if (reply.epoch + 1 != nextEpoch)
680 if (!TransactionIdPrecedesOrEquals(reply.xmin, nextXid))
681 return; /* epoch OK, but it's wrapped around */
684 * Set the WalSender's xmin equal to the standby's requested xmin, so that
685 * the xmin will be taken into account by GetOldestXmin. This will hold
686 * back the removal of dead rows and thereby prevent the generation of
687 * cleanup conflicts on the standby server.
689 * There is a small window for a race condition here: although we just
690 * checked that reply.xmin precedes nextXid, the nextXid could have gotten
691 * advanced between our fetching it and applying the xmin below, perhaps
692 * far enough to make reply.xmin wrap around. In that case the xmin we
693 * set here would be "in the future" and have no effect. No point in
694 * worrying about this since it's too late to save the desired data
695 * anyway. Assuming that the standby sends us an increasing sequence of
696 * xmins, this could only happen during the first reply cycle, else our
697 * own xmin would prevent nextXid from advancing so far.
699 * We don't bother taking the ProcArrayLock here. Setting the xmin field
700 * is assumed atomic, and there's no real need to prevent a concurrent
701 * GetOldestXmin. (If we're moving our xmin forward, this is obviously
702 * safe, and if we're moving it backwards, well, the data is at risk
703 * already since a VACUUM could have just finished calling GetOldestXmin.)
705 MyPgXact->xmin = reply.xmin;
708 /* Main loop of walsender process */
712 char *output_message;
713 bool caughtup = false;
716 * Allocate buffer that will be used for each output message. We do this
717 * just once to reduce palloc overhead. The buffer must be made large
718 * enough for maximum-sized messages.
720 output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
723 * Allocate buffer that will be used for processing reply messages. As
724 * above, do this just once to reduce palloc overhead.
726 initStringInfo(&reply_message);
728 /* Initialize the last reply timestamp */
729 last_reply_timestamp = GetCurrentTimestamp();
731 /* Loop forever, unless we get an error */
734 /* Clear any already-pending wakeups */
735 ResetLatch(&MyWalSnd->latch);
738 * Emergency bailout if postmaster has died. This is to avoid the
739 * necessity for manual cleanup of all postmaster children.
741 if (!PostmasterIsAlive())
744 /* Process any requests or signals received recently */
748 ProcessConfigFile(PGC_SIGHUP);
752 /* Normal exit from the walsender is here */
753 if (walsender_shutdown_requested)
755 /* Inform the standby that XLOG streaming is done */
756 pq_puttextmessage('C', "COPY 0");
762 /* Check for input from the client */
763 ProcessRepliesIfAny();
766 * If we don't have any pending data in the output buffer, try to send
767 * some more. If there is some, we don't bother to call XLogSend
768 * again until we've flushed it ... but we'd better assume we are not
771 if (!pq_is_send_pending())
772 XLogSend(output_message, &caughtup);
776 /* Try to flush pending output to the client */
777 if (pq_flush_if_writable() != 0)
780 /* If nothing remains to be sent right now ... */
781 if (caughtup && !pq_is_send_pending())
784 * If we're in catchup state, move to streaming. This is an
785 * important state change for users to know about, since before
786 * this point data loss might occur if the primary dies and we
787 * need to failover to the standby. The state change is also
788 * important for synchronous replication, since commits that
789 * started to wait at that point might wait for some time.
791 if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
794 (errmsg("standby \"%s\" has now caught up with primary",
796 WalSndSetState(WALSNDSTATE_STREAMING);
800 * When SIGUSR2 arrives, we send any outstanding logs up to the
801 * shutdown checkpoint record (i.e., the latest record) and exit.
802 * This may be a normal termination at shutdown, or a promotion,
803 * the walsender is not sure which.
805 if (walsender_ready_to_stop)
807 /* ... let's just be real sure we're caught up ... */
808 XLogSend(output_message, &caughtup);
809 if (caughtup && !pq_is_send_pending())
811 walsender_shutdown_requested = true;
812 continue; /* don't want to wait more */
818 * We don't block if not caught up, unless there is unsent data
819 * pending in which case we'd better block until the socket is
820 * write-ready. This test is only needed for the case where XLogSend
821 * loaded a subset of the available data but then pq_flush_if_writable
822 * flushed it all --- we should immediately try to send more.
824 if (caughtup || pq_is_send_pending())
826 TimestampTz timeout = 0;
827 long sleeptime = 10000; /* 10 s */
830 wakeEvents = WL_LATCH_SET | WL_POSTMASTER_DEATH |
831 WL_SOCKET_READABLE | WL_TIMEOUT;
833 if (pq_is_send_pending())
834 wakeEvents |= WL_SOCKET_WRITEABLE;
837 WalSndKeepalive(output_message);
838 /* Try to flush pending output to the client */
839 if (pq_flush_if_writable() != 0)
843 /* Determine time until replication timeout */
844 if (replication_timeout > 0)
846 timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
847 replication_timeout);
848 sleeptime = 1 + (replication_timeout / 10);
851 /* Sleep until something happens or replication timeout */
852 WaitLatchOrSocket(&MyWalSnd->latch, wakeEvents,
853 MyProcPort->sock, sleeptime);
856 * Check for replication timeout. Note we ignore the corner case
857 * possibility that the client replied just as we reached the
858 * timeout ... he's supposed to reply *before* that.
860 if (replication_timeout > 0 &&
861 GetCurrentTimestamp() >= timeout)
864 * Since typically expiration of replication timeout means
865 * communication problem, we don't send the error message to
869 (errmsg("terminating walsender process due to replication timeout")));
876 * Get here on send failure. Clean up and exit.
878 * Reset whereToSendOutput to prevent ereport from attempting to send any
879 * more messages to the standby.
881 if (whereToSendOutput == DestRemote)
882 whereToSendOutput = DestNone;
885 abort(); /* keep the compiler quiet */
888 /* Initialize a per-walsender data structure for this walsender process */
895 * WalSndCtl should be set up already (we inherit this by fork() or
896 * EXEC_BACKEND mechanism from the postmaster).
898 Assert(WalSndCtl != NULL);
899 Assert(MyWalSnd == NULL);
902 * Find a free walsender slot and reserve it. If this fails, we must be
903 * out of WalSnd structures.
905 for (i = 0; i < max_wal_senders; i++)
907 /* use volatile pointer to prevent code rearrangement */
908 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
910 SpinLockAcquire(&walsnd->mutex);
912 if (walsnd->pid != 0)
914 SpinLockRelease(&walsnd->mutex);
920 * Found a free slot. Reserve it for us.
922 walsnd->pid = MyProcPid;
923 MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
924 walsnd->state = WALSNDSTATE_STARTUP;
925 SpinLockRelease(&walsnd->mutex);
926 /* don't need the lock anymore */
927 OwnLatch((Latch *) &walsnd->latch);
928 MyWalSnd = (WalSnd *) walsnd;
933 if (MyWalSnd == NULL)
935 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
936 errmsg("number of requested standby connections "
937 "exceeds max_wal_senders (currently %d)",
940 /* Arrange to clean up at walsender exit */
941 on_shmem_exit(WalSndKill, 0);
944 /* Destroy the per-walsender data structure for this walsender process */
946 WalSndKill(int code, Datum arg)
948 Assert(MyWalSnd != NULL);
951 * Mark WalSnd struct no longer in use. Assume that no lock is required
955 DisownLatch(&MyWalSnd->latch);
957 /* WalSnd struct isn't mine anymore */
962 * Read 'count' bytes from WAL into 'buf', starting at location 'startptr'
964 * XXX probably this should be improved to suck data directly from the
965 * WAL buffers when possible.
967 * Will open, and keep open, one WAL segment stored in the global file
968 * descriptor sendFile. This means if XLogRead is used once, there will
969 * always be one descriptor left open until the process ends, but never
973 XLogRead(char *buf, XLogRecPtr startptr, Size count)
978 XLogSegNo lastRemovedSegNo;
992 startoff = recptr % XLogSegSize;
994 if (sendFile < 0 || !XLByteInSeg(recptr, sendSegNo))
996 char path[MAXPGPATH];
998 /* Switch to another logfile segment */
1002 XLByteToSeg(recptr, sendSegNo);
1003 XLogFilePath(path, ThisTimeLineID, sendSegNo);
1005 sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
1009 * If the file is not found, assume it's because the standby
1010 * asked for a too old WAL segment that has already been
1011 * removed or recycled.
1013 if (errno == ENOENT)
1015 (errcode_for_file_access(),
1016 errmsg("requested WAL segment %s has already been removed",
1017 XLogFileNameP(ThisTimeLineID, sendSegNo))));
1020 (errcode_for_file_access(),
1021 errmsg("could not open file \"%s\": %m",
1027 /* Need to seek in the file? */
1028 if (sendOff != startoff)
1030 if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
1032 (errcode_for_file_access(),
1033 errmsg("could not seek in log segment %s to offset %u: %m",
1034 XLogFileNameP(ThisTimeLineID, sendSegNo),
1039 /* How many bytes are within this segment? */
1040 if (nbytes > (XLogSegSize - startoff))
1041 segbytes = XLogSegSize - startoff;
1045 readbytes = read(sendFile, p, segbytes);
1049 (errcode_for_file_access(),
1050 errmsg("could not read from log segment %s, offset %u, length %lu: %m",
1051 XLogFileNameP(ThisTimeLineID, sendSegNo),
1052 sendOff, (unsigned long) segbytes)));
1055 /* Update state for read */
1056 XLByteAdvance(recptr, readbytes);
1058 sendOff += readbytes;
1059 nbytes -= readbytes;
1064 * After reading into the buffer, check that what we read was valid. We do
1065 * this after reading, because even though the segment was present when we
1066 * opened it, it might get recycled or removed while we read it. The
1067 * read() succeeds in that case, but the data we tried to read might
1068 * already have been overwritten with new WAL records.
1070 XLogGetLastRemoved(&lastRemovedSegNo);
1071 XLByteToSeg(startptr, segno);
1072 if (segno <= lastRemovedSegNo)
1074 (errcode_for_file_access(),
1075 errmsg("requested WAL segment %s has already been removed",
1076 XLogFileNameP(ThisTimeLineID, segno))));
1079 * During recovery, the currently-open WAL file might be replaced with the
1080 * file of the same name retrieved from archive. So we always need to
1081 * check what we read was valid after reading into the buffer. If it's
1082 * invalid, we try to open and read the file again.
1084 if (am_cascading_walsender)
1086 /* use volatile pointer to prevent code rearrangement */
1087 volatile WalSnd *walsnd = MyWalSnd;
1090 SpinLockAcquire(&walsnd->mutex);
1091 reload = walsnd->needreload;
1092 walsnd->needreload = false;
1093 SpinLockRelease(&walsnd->mutex);
1095 if (reload && sendFile >= 0)
1106 * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
1107 * but not yet sent to the client, and buffer it in the libpq output
1110 * msgbuf is a work area in which the output message is constructed. It's
1111 * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
1112 * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
1114 * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
1115 * *caughtup is set to false.
1119 XLogSend(char *msgbuf, bool *caughtup)
1121 XLogRecPtr SendRqstPtr;
1122 XLogRecPtr startptr;
1125 WalDataMessageHeader msghdr;
1128 * Attempt to send all data that's already been written out and fsync'd to
1129 * disk. We cannot go further than what's been written out given the
1130 * current implementation of XLogRead(). And in any case it's unsafe to
1131 * send WAL that is not securely down to disk on the master: if the master
1132 * subsequently crashes and restarts, slaves must not have applied any WAL
1133 * that gets lost on the master.
1135 SendRqstPtr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetFlushRecPtr();
1137 /* Quick exit if nothing to do */
1138 if (XLByteLE(SendRqstPtr, sentPtr))
1145 * Figure out how much to send in one message. If there's no more than
1146 * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
1147 * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
1149 * The rounding is not only for performance reasons. Walreceiver relies on
1150 * the fact that we never split a WAL record across two messages. Since a
1151 * long WAL record is split at page boundary into continuation records,
1152 * page boundary is always a safe cut-off point. We also assume that
1153 * SendRqstPtr never points to the middle of a WAL record.
1157 XLByteAdvance(endptr, MAX_SEND_SIZE);
1159 /* if we went beyond SendRqstPtr, back off */
1160 if (XLByteLE(SendRqstPtr, endptr))
1162 endptr = SendRqstPtr;
1167 /* round down to page boundary. */
1168 endptr -= (endptr % XLOG_BLCKSZ);
1172 nbytes = endptr - startptr;
1173 Assert(nbytes <= MAX_SEND_SIZE);
1176 * OK to read and send the slice.
1181 * Read the log directly into the output buffer to avoid extra memcpy
1184 XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
1187 * We fill the message header last so that the send timestamp is taken as
1190 msghdr.dataStart = startptr;
1191 msghdr.walEnd = SendRqstPtr;
1192 msghdr.sendTime = GetCurrentTimestamp();
1194 memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
1196 pq_putmessage_noblock('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
1200 /* Update shared memory status */
1202 /* use volatile pointer to prevent code rearrangement */
1203 volatile WalSnd *walsnd = MyWalSnd;
1205 SpinLockAcquire(&walsnd->mutex);
1206 walsnd->sentPtr = sentPtr;
1207 SpinLockRelease(&walsnd->mutex);
1210 /* Report progress of XLOG streaming in PS display */
1211 if (update_process_title)
1213 char activitymsg[50];
1215 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
1216 (uint32) (sentPtr >> 32), (uint32) sentPtr);
1217 set_ps_display(activitymsg, false);
1224 * Request walsenders to reload the currently-open WAL file
1227 WalSndRqstFileReload(void)
1231 for (i = 0; i < max_wal_senders; i++)
1233 /* use volatile pointer to prevent code rearrangement */
1234 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1236 if (walsnd->pid == 0)
1239 SpinLockAcquire(&walsnd->mutex);
1240 walsnd->needreload = true;
1241 SpinLockRelease(&walsnd->mutex);
1245 /* SIGHUP: set flag to re-read config file at next convenient time */
1247 WalSndSigHupHandler(SIGNAL_ARGS)
1249 int save_errno = errno;
1253 SetLatch(&MyWalSnd->latch);
1258 /* SIGTERM: set flag to shut down */
1260 WalSndShutdownHandler(SIGNAL_ARGS)
1262 int save_errno = errno;
1264 walsender_shutdown_requested = true;
1266 SetLatch(&MyWalSnd->latch);
1269 * Set the standard (non-walsender) state as well, so that we can abort
1270 * things like do_pg_stop_backup().
1272 InterruptPending = true;
1273 ProcDiePending = true;
1279 * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
1281 * Some backend has bought the farm,
1282 * so we need to stop what we're doing and exit.
1285 WalSndQuickDieHandler(SIGNAL_ARGS)
1287 PG_SETMASK(&BlockSig);
1290 * We DO NOT want to run proc_exit() callbacks -- we're here because
1291 * shared memory may be corrupted, so we don't want to try to clean up our
1292 * transaction. Just nail the windows shut and get out of town. Now that
1293 * there's an atexit callback to prevent third-party code from breaking
1294 * things by calling exit() directly, we have to reset the callbacks
1295 * explicitly to make this work as intended.
1300 * Note we do exit(2) not exit(0). This is to force the postmaster into a
1301 * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
1302 * backend. This is necessary precisely because we don't clean up our
1303 * shared memory state. (The "dead man switch" mechanism in pmsignal.c
1304 * should ensure the postmaster sees this as a crash, too, but no harm in
1305 * being doubly sure.)
1310 /* SIGUSR1: set flag to send WAL records */
1312 WalSndXLogSendHandler(SIGNAL_ARGS)
1314 int save_errno = errno;
1316 latch_sigusr1_handler();
1321 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
1323 WalSndLastCycleHandler(SIGNAL_ARGS)
1325 int save_errno = errno;
1327 walsender_ready_to_stop = true;
1329 SetLatch(&MyWalSnd->latch);
1334 /* Set up signal handlers */
1338 /* Set up signal handlers */
1339 pqsignal(SIGHUP, WalSndSigHupHandler); /* set flag to read config
1341 pqsignal(SIGINT, SIG_IGN); /* not used */
1342 pqsignal(SIGTERM, WalSndShutdownHandler); /* request shutdown */
1343 pqsignal(SIGQUIT, WalSndQuickDieHandler); /* hard crash time */
1344 pqsignal(SIGALRM, SIG_IGN);
1345 pqsignal(SIGPIPE, SIG_IGN);
1346 pqsignal(SIGUSR1, WalSndXLogSendHandler); /* request WAL sending */
1347 pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
1350 /* Reset some signals that are accepted by postmaster but not here */
1351 pqsignal(SIGCHLD, SIG_DFL);
1352 pqsignal(SIGTTIN, SIG_DFL);
1353 pqsignal(SIGTTOU, SIG_DFL);
1354 pqsignal(SIGCONT, SIG_DFL);
1355 pqsignal(SIGWINCH, SIG_DFL);
1358 /* Report shared-memory space needed by WalSndShmemInit */
1360 WalSndShmemSize(void)
1364 size = offsetof(WalSndCtlData, walsnds);
1365 size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
1370 /* Allocate and initialize walsender-related shared memory */
1372 WalSndShmemInit(void)
1377 WalSndCtl = (WalSndCtlData *)
1378 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
1382 /* First time through, so initialize */
1383 MemSet(WalSndCtl, 0, WalSndShmemSize());
1385 for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
1386 SHMQueueInit(&(WalSndCtl->SyncRepQueue[i]));
1388 for (i = 0; i < max_wal_senders; i++)
1390 WalSnd *walsnd = &WalSndCtl->walsnds[i];
1392 SpinLockInit(&walsnd->mutex);
1393 InitSharedLatch(&walsnd->latch);
1398 /* Wake up all walsenders */
1404 for (i = 0; i < max_wal_senders; i++)
1405 SetLatch(&WalSndCtl->walsnds[i].latch);
1408 /* Set state for current walsender (only called in walsender) */
1410 WalSndSetState(WalSndState state)
1412 /* use volatile pointer to prevent code rearrangement */
1413 volatile WalSnd *walsnd = MyWalSnd;
1415 Assert(am_walsender);
1417 if (walsnd->state == state)
1420 SpinLockAcquire(&walsnd->mutex);
1421 walsnd->state = state;
1422 SpinLockRelease(&walsnd->mutex);
1426 * Return a string constant representing the state. This is used
1427 * in system views, and should *not* be translated.
1430 WalSndGetStateString(WalSndState state)
1434 case WALSNDSTATE_STARTUP:
1436 case WALSNDSTATE_BACKUP:
1438 case WALSNDSTATE_CATCHUP:
1440 case WALSNDSTATE_STREAMING:
1448 * Returns activity of walsenders, including pids and xlog locations sent to
1452 pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
1454 #define PG_STAT_GET_WAL_SENDERS_COLS 8
1455 ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
1457 Tuplestorestate *tupstore;
1458 MemoryContext per_query_ctx;
1459 MemoryContext oldcontext;
1462 int sync_standby = -1;
1465 /* check to see if caller supports us returning a tuplestore */
1466 if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
1468 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1469 errmsg("set-valued function called in context that cannot accept a set")));
1470 if (!(rsinfo->allowedModes & SFRM_Materialize))
1472 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1473 errmsg("materialize mode required, but it is not " \
1474 "allowed in this context")));
1476 /* Build a tuple descriptor for our result type */
1477 if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
1478 elog(ERROR, "return type must be a row type");
1480 per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
1481 oldcontext = MemoryContextSwitchTo(per_query_ctx);
1483 tupstore = tuplestore_begin_heap(true, false, work_mem);
1484 rsinfo->returnMode = SFRM_Materialize;
1485 rsinfo->setResult = tupstore;
1486 rsinfo->setDesc = tupdesc;
1488 MemoryContextSwitchTo(oldcontext);
1491 * Get the priorities of sync standbys all in one go, to minimise lock
1492 * acquisitions and to allow us to evaluate who is the current sync
1493 * standby. This code must match the code in SyncRepReleaseWaiters().
1495 sync_priority = palloc(sizeof(int) * max_wal_senders);
1496 LWLockAcquire(SyncRepLock, LW_SHARED);
1497 for (i = 0; i < max_wal_senders; i++)
1499 /* use volatile pointer to prevent code rearrangement */
1500 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1502 if (walsnd->pid != 0)
1504 sync_priority[i] = walsnd->sync_standby_priority;
1506 if (walsnd->state == WALSNDSTATE_STREAMING &&
1507 walsnd->sync_standby_priority > 0 &&
1509 priority > walsnd->sync_standby_priority))
1511 priority = walsnd->sync_standby_priority;
1516 LWLockRelease(SyncRepLock);
1518 for (i = 0; i < max_wal_senders; i++)
1520 /* use volatile pointer to prevent code rearrangement */
1521 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1522 char location[MAXFNAMELEN];
1528 Datum values[PG_STAT_GET_WAL_SENDERS_COLS];
1529 bool nulls[PG_STAT_GET_WAL_SENDERS_COLS];
1531 if (walsnd->pid == 0)
1534 SpinLockAcquire(&walsnd->mutex);
1535 sentPtr = walsnd->sentPtr;
1536 state = walsnd->state;
1537 write = walsnd->write;
1538 flush = walsnd->flush;
1539 apply = walsnd->apply;
1540 SpinLockRelease(&walsnd->mutex);
1542 memset(nulls, 0, sizeof(nulls));
1543 values[0] = Int32GetDatum(walsnd->pid);
1548 * Only superusers can see details. Other users only get the pid
1549 * value to know it's a walsender, but no details.
1551 MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
1555 values[1] = CStringGetTextDatum(WalSndGetStateString(state));
1557 snprintf(location, sizeof(location), "%X/%X",
1558 (uint32) (sentPtr >> 32), (uint32) sentPtr);
1559 values[2] = CStringGetTextDatum(location);
1563 snprintf(location, sizeof(location), "%X/%X",
1564 (uint32) (write >> 32), (uint32) write);
1565 values[3] = CStringGetTextDatum(location);
1569 snprintf(location, sizeof(location), "%X/%X",
1570 (uint32) (flush >> 32), (uint32) flush);
1571 values[4] = CStringGetTextDatum(location);
1575 snprintf(location, sizeof(location), "%X/%X",
1576 (uint32) (apply >> 32), (uint32) apply);
1577 values[5] = CStringGetTextDatum(location);
1579 values[6] = Int32GetDatum(sync_priority[i]);
1582 * More easily understood version of standby state. This is purely
1583 * informational, not different from priority.
1585 if (sync_priority[i] == 0)
1586 values[7] = CStringGetTextDatum("async");
1587 else if (i == sync_standby)
1588 values[7] = CStringGetTextDatum("sync");
1590 values[7] = CStringGetTextDatum("potential");
1593 tuplestore_putvalues(tupstore, tupdesc, values, nulls);
1595 pfree(sync_priority);
1597 /* clean up and return the tuplestore */
1598 tuplestore_donestoring(tupstore);
1604 WalSndKeepalive(char *msgbuf)
1606 PrimaryKeepaliveMessage keepalive_message;
1608 /* Construct a new message */
1609 keepalive_message.walEnd = sentPtr;
1610 keepalive_message.sendTime = GetCurrentTimestamp();
1612 elog(DEBUG2, "sending replication keepalive");
1614 /* Prepend with the message type and send it. */
1616 memcpy(msgbuf + 1, &keepalive_message, sizeof(PrimaryKeepaliveMessage));
1617 pq_putmessage_noblock('d', msgbuf, sizeof(PrimaryKeepaliveMessage) + 1);
1621 * This isn't currently used for anything. Monitoring tools might be
1622 * interested in the future, and we'll need something like this in the
1623 * future for synchronous replication.
1627 * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
1631 GetOldestWALSendPointer(void)
1633 XLogRecPtr oldest = {0, 0};
1637 for (i = 0; i < max_wal_senders; i++)
1639 /* use volatile pointer to prevent code rearrangement */
1640 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1643 if (walsnd->pid == 0)
1646 SpinLockAcquire(&walsnd->mutex);
1647 recptr = walsnd->sentPtr;
1648 SpinLockRelease(&walsnd->mutex);
1650 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1653 if (!found || XLByteLT(recptr, oldest))