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 uint32 sendId = 0;
91 static uint32 sendSeg = 0;
92 static uint32 sendOff = 0;
95 * How far have we sent WAL already? This is also advertised in
96 * MyWalSnd->sentPtr. (Actually, this is the next WAL location to send.)
98 static XLogRecPtr sentPtr = {0, 0};
101 * Buffer for processing reply messages.
103 static StringInfoData reply_message;
106 * Timestamp of the last receipt of the reply from the standby.
108 static TimestampTz last_reply_timestamp;
110 /* Flags set by signal handlers for later service in main loop */
111 static volatile sig_atomic_t got_SIGHUP = false;
112 volatile sig_atomic_t walsender_shutdown_requested = false;
113 volatile sig_atomic_t walsender_ready_to_stop = false;
115 /* Signal handlers */
116 static void WalSndSigHupHandler(SIGNAL_ARGS);
117 static void WalSndShutdownHandler(SIGNAL_ARGS);
118 static void WalSndQuickDieHandler(SIGNAL_ARGS);
119 static void WalSndXLogSendHandler(SIGNAL_ARGS);
120 static void WalSndLastCycleHandler(SIGNAL_ARGS);
122 /* Prototypes for private functions */
123 static bool HandleReplicationCommand(const char *cmd_string);
124 static int WalSndLoop(void);
125 static void InitWalSnd(void);
126 static void WalSndHandshake(void);
127 static void WalSndKill(int code, Datum arg);
128 static void XLogSend(char *msgbuf, bool *caughtup);
129 static void IdentifySystem(void);
130 static void StartReplication(StartReplicationCmd *cmd);
131 static void ProcessStandbyMessage(void);
132 static void ProcessStandbyReplyMessage(void);
133 static void ProcessStandbyHSFeedbackMessage(void);
134 static void ProcessRepliesIfAny(void);
135 static void WalSndKeepalive(char *msgbuf);
138 /* Main entry point for walsender process */
142 MemoryContext walsnd_context;
144 am_cascading_walsender = RecoveryInProgress();
146 /* Create a per-walsender data structure in shared memory */
150 * Create a memory context that we will do all our work in. We do this so
151 * that we can reset the context during error recovery and thereby avoid
152 * possible memory leaks. Formerly this code just ran in
153 * TopMemoryContext, but resetting that would be a really bad idea.
155 * XXX: we don't actually attempt error recovery in walsender, we just
156 * close the connection and exit.
158 walsnd_context = AllocSetContextCreate(TopMemoryContext,
160 ALLOCSET_DEFAULT_MINSIZE,
161 ALLOCSET_DEFAULT_INITSIZE,
162 ALLOCSET_DEFAULT_MAXSIZE);
163 MemoryContextSwitchTo(walsnd_context);
165 /* Set up resource owner */
166 CurrentResourceOwner = ResourceOwnerCreate(NULL, "walsender top-level resource owner");
168 /* Unblock signals (they were blocked when the postmaster forked us) */
169 PG_SETMASK(&UnBlockSig);
172 * Use the recovery target timeline ID during recovery
174 if (am_cascading_walsender)
175 ThisTimeLineID = GetRecoveryTargetTLI();
177 /* Tell the standby that walsender is ready for receiving commands */
178 ReadyForQuery(DestRemote);
180 /* Handle handshake messages before streaming */
183 /* Initialize shared memory status */
185 /* use volatile pointer to prevent code rearrangement */
186 volatile WalSnd *walsnd = MyWalSnd;
188 SpinLockAcquire(&walsnd->mutex);
189 walsnd->sentPtr = sentPtr;
190 SpinLockRelease(&walsnd->mutex);
195 /* Main loop of walsender */
200 * Execute commands from walreceiver, until we enter streaming mode.
203 WalSndHandshake(void)
205 StringInfoData input_message;
206 bool replication_started = false;
208 initStringInfo(&input_message);
210 while (!replication_started)
214 WalSndSetState(WALSNDSTATE_STARTUP);
215 set_ps_display("idle", false);
217 /* Wait for a command to arrive */
218 firstchar = pq_getbyte();
221 * Emergency bailout if postmaster has died. This is to avoid the
222 * necessity for manual cleanup of all postmaster children.
224 if (!PostmasterIsAlive())
228 * Check for any other interesting events that happened while we
234 ProcessConfigFile(PGC_SIGHUP);
237 if (firstchar != EOF)
240 * Read the message contents. This is expected to be done without
241 * blocking because we've been able to get message type code.
243 if (pq_getmessage(&input_message, 0))
244 firstchar = EOF; /* suitable message already logged */
247 /* Handle the very limited subset of commands expected in this phase */
250 case 'Q': /* Query message */
252 const char *query_string;
254 query_string = pq_getmsgstring(&input_message);
255 pq_getmsgend(&input_message);
257 if (HandleReplicationCommand(query_string))
258 replication_started = true;
263 /* standby is closing the connection */
267 /* standby disconnected unexpectedly */
269 (errcode(ERRCODE_PROTOCOL_VIOLATION),
270 errmsg("unexpected EOF on standby connection")));
275 (errcode(ERRCODE_PROTOCOL_VIOLATION),
276 errmsg("invalid standby handshake message type %d", firstchar)));
290 char xpos[MAXFNAMELEN];
294 * Reply with a result set with one row, three columns. First col is
295 * system ID, second is timeline ID, and third is current xlog location.
298 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
299 GetSystemIdentifier());
300 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
302 logptr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetInsertRecPtr();
304 snprintf(xpos, sizeof(xpos), "%X/%X",
305 logptr.xlogid, logptr.xrecoff);
307 /* Send a RowDescription message */
308 pq_beginmessage(&buf, 'T');
309 pq_sendint(&buf, 3, 2); /* 3 fields */
312 pq_sendstring(&buf, "systemid"); /* col name */
313 pq_sendint(&buf, 0, 4); /* table oid */
314 pq_sendint(&buf, 0, 2); /* attnum */
315 pq_sendint(&buf, TEXTOID, 4); /* type oid */
316 pq_sendint(&buf, -1, 2); /* typlen */
317 pq_sendint(&buf, 0, 4); /* typmod */
318 pq_sendint(&buf, 0, 2); /* format code */
321 pq_sendstring(&buf, "timeline"); /* col name */
322 pq_sendint(&buf, 0, 4); /* table oid */
323 pq_sendint(&buf, 0, 2); /* attnum */
324 pq_sendint(&buf, INT4OID, 4); /* type oid */
325 pq_sendint(&buf, 4, 2); /* typlen */
326 pq_sendint(&buf, 0, 4); /* typmod */
327 pq_sendint(&buf, 0, 2); /* format code */
330 pq_sendstring(&buf, "xlogpos");
331 pq_sendint(&buf, 0, 4);
332 pq_sendint(&buf, 0, 2);
333 pq_sendint(&buf, TEXTOID, 4);
334 pq_sendint(&buf, -1, 2);
335 pq_sendint(&buf, 0, 4);
336 pq_sendint(&buf, 0, 2);
339 /* Send a DataRow message */
340 pq_beginmessage(&buf, 'D');
341 pq_sendint(&buf, 3, 2); /* # of columns */
342 pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
343 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
344 pq_sendint(&buf, strlen(tli), 4); /* col2 len */
345 pq_sendbytes(&buf, (char *) tli, strlen(tli));
346 pq_sendint(&buf, strlen(xpos), 4); /* col3 len */
347 pq_sendbytes(&buf, (char *) xpos, strlen(xpos));
351 /* Send CommandComplete and ReadyForQuery messages */
352 EndCommand("SELECT", DestRemote);
353 ReadyForQuery(DestRemote);
354 /* ReadyForQuery did pq_flush for us */
361 StartReplication(StartReplicationCmd *cmd)
366 * Let postmaster know that we're streaming. Once we've declared us as a
367 * WAL sender process, postmaster will let us outlive the bgwriter and
368 * kill us last in the shutdown sequence, so we get a chance to stream all
369 * remaining WAL at shutdown, including the shutdown checkpoint. Note that
370 * there's no going back, and we mustn't write any WAL records after this.
372 MarkPostmasterChildWalSender();
373 SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
376 * When promoting a cascading standby, postmaster sends SIGUSR2 to any
377 * cascading walsenders to kill them. But there is a corner-case where
378 * such walsender fails to receive SIGUSR2 and survives a standby
379 * promotion unexpectedly. This happens when postmaster sends SIGUSR2
380 * before the walsender marks itself as a WAL sender, because postmaster
381 * sends SIGUSR2 to only the processes marked as a WAL sender.
383 * To avoid this corner-case, if recovery is NOT in progress even though
384 * the walsender is cascading one, we do the same thing as SIGUSR2 signal
385 * handler does, i.e., set walsender_ready_to_stop to true. Which causes
386 * the walsender to end later.
388 * When terminating cascading walsenders, usually postmaster writes the
389 * log message announcing the terminations. But there is a race condition
390 * here. If there is no walsender except this process before reaching
391 * here, postmaster thinks that there is no walsender and suppresses that
392 * log message. To handle this case, we always emit that log message here.
393 * This might cause duplicate log messages, but which is less likely to
394 * happen, so it's not worth writing some code to suppress them.
396 if (am_cascading_walsender && !RecoveryInProgress())
399 (errmsg("terminating walsender process to force cascaded standby "
400 "to update timeline and reconnect")));
401 walsender_ready_to_stop = true;
405 * We assume here that we're logging enough information in the WAL for
406 * log-shipping, since this is checked in PostmasterMain().
408 * NOTE: wal_level can only change at shutdown, so in most cases it is
409 * difficult for there to be WAL data that we can still see that was
410 * written at wal_level='minimal'.
414 * When we first start replication the standby will be behind the primary.
415 * For some applications, for example, synchronous replication, it is
416 * important to have a clear state for this initial catchup mode, so we
417 * can trigger actions when we change streaming state later. We may stay
418 * in this state for a long time, which is exactly why we want to be able
419 * to monitor whether or not we are still here.
421 WalSndSetState(WALSNDSTATE_CATCHUP);
423 /* Send a CopyBothResponse message, and start streaming */
424 pq_beginmessage(&buf, 'W');
425 pq_sendbyte(&buf, 0);
426 pq_sendint(&buf, 0, 2);
431 * Initialize position to the received one, then the xlog records begin to
432 * be shipped from that position
434 sentPtr = cmd->startpoint;
438 * Execute an incoming replication command.
441 HandleReplicationCommand(const char *cmd_string)
443 bool replication_started = false;
446 MemoryContext cmd_context;
447 MemoryContext old_context;
449 elog(DEBUG1, "received replication command: %s", cmd_string);
451 cmd_context = AllocSetContextCreate(CurrentMemoryContext,
452 "Replication command context",
453 ALLOCSET_DEFAULT_MINSIZE,
454 ALLOCSET_DEFAULT_INITSIZE,
455 ALLOCSET_DEFAULT_MAXSIZE);
456 old_context = MemoryContextSwitchTo(cmd_context);
458 replication_scanner_init(cmd_string);
459 parse_rc = replication_yyparse();
462 (errcode(ERRCODE_SYNTAX_ERROR),
463 (errmsg_internal("replication command parser returned %d",
466 cmd_node = replication_parse_result;
468 switch (cmd_node->type)
470 case T_IdentifySystemCmd:
474 case T_StartReplicationCmd:
475 StartReplication((StartReplicationCmd *) cmd_node);
477 /* break out of the loop */
478 replication_started = true;
481 case T_BaseBackupCmd:
482 SendBaseBackup((BaseBackupCmd *) cmd_node);
484 /* Send CommandComplete and ReadyForQuery messages */
485 EndCommand("SELECT", DestRemote);
486 ReadyForQuery(DestRemote);
487 /* ReadyForQuery did pq_flush for us */
492 (errcode(ERRCODE_PROTOCOL_VIOLATION),
493 errmsg("invalid standby query string: %s", cmd_string)));
497 MemoryContextSwitchTo(old_context);
498 MemoryContextDelete(cmd_context);
500 return replication_started;
504 * Check if the remote end has closed the connection.
507 ProcessRepliesIfAny(void)
509 unsigned char firstchar;
511 bool received = false;
515 r = pq_getbyte_if_available(&firstchar);
518 /* unexpected error or EOF */
520 (errcode(ERRCODE_PROTOCOL_VIOLATION),
521 errmsg("unexpected EOF on standby connection")));
526 /* no data available without blocking */
530 /* Handle the very limited subset of commands expected in this phase */
534 * 'd' means a standby reply wrapped in a CopyData packet.
537 ProcessStandbyMessage();
542 * 'X' means that the standby is closing down the socket.
549 (errcode(ERRCODE_PROTOCOL_VIOLATION),
550 errmsg("invalid standby message type \"%c\"",
556 * Save the last reply timestamp if we've received at least one reply.
559 last_reply_timestamp = GetCurrentTimestamp();
563 * Process a status update message received from standby.
566 ProcessStandbyMessage(void)
570 resetStringInfo(&reply_message);
573 * Read the message contents.
575 if (pq_getmessage(&reply_message, 0))
578 (errcode(ERRCODE_PROTOCOL_VIOLATION),
579 errmsg("unexpected EOF on standby connection")));
584 * Check message type from the first byte.
586 msgtype = pq_getmsgbyte(&reply_message);
591 ProcessStandbyReplyMessage();
595 ProcessStandbyHSFeedbackMessage();
600 (errcode(ERRCODE_PROTOCOL_VIOLATION),
601 errmsg("unexpected message type \"%c\"", msgtype)));
607 * Regular reply from standby advising of WAL positions on standby server.
610 ProcessStandbyReplyMessage(void)
612 StandbyReplyMessage reply;
614 pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyReplyMessage));
616 elog(DEBUG2, "write %X/%X flush %X/%X apply %X/%X",
617 reply.write.xlogid, reply.write.xrecoff,
618 reply.flush.xlogid, reply.flush.xrecoff,
619 reply.apply.xlogid, reply.apply.xrecoff);
622 * Update shared state for this WalSender process based on reply data from
626 /* use volatile pointer to prevent code rearrangement */
627 volatile WalSnd *walsnd = MyWalSnd;
629 SpinLockAcquire(&walsnd->mutex);
630 walsnd->write = reply.write;
631 walsnd->flush = reply.flush;
632 walsnd->apply = reply.apply;
633 SpinLockRelease(&walsnd->mutex);
636 if (!am_cascading_walsender)
637 SyncRepReleaseWaiters();
641 * Hot Standby feedback
644 ProcessStandbyHSFeedbackMessage(void)
646 StandbyHSFeedbackMessage reply;
647 TransactionId nextXid;
650 /* Decipher the reply message */
651 pq_copymsgbytes(&reply_message, (char *) &reply,
652 sizeof(StandbyHSFeedbackMessage));
654 elog(DEBUG2, "hot standby feedback xmin %u epoch %u",
658 /* Ignore invalid xmin (can't actually happen with current walreceiver) */
659 if (!TransactionIdIsNormal(reply.xmin))
663 * Check that the provided xmin/epoch are sane, that is, not in the future
664 * and not so far back as to be already wrapped around. Ignore if not.
666 * Epoch of nextXid should be same as standby, or if the counter has
667 * wrapped, then one greater than standby.
669 GetNextXidAndEpoch(&nextXid, &nextEpoch);
671 if (reply.xmin <= nextXid)
673 if (reply.epoch != nextEpoch)
678 if (reply.epoch + 1 != nextEpoch)
682 if (!TransactionIdPrecedesOrEquals(reply.xmin, nextXid))
683 return; /* epoch OK, but it's wrapped around */
686 * Set the WalSender's xmin equal to the standby's requested xmin, so that
687 * the xmin will be taken into account by GetOldestXmin. This will hold
688 * back the removal of dead rows and thereby prevent the generation of
689 * cleanup conflicts on the standby server.
691 * There is a small window for a race condition here: although we just
692 * checked that reply.xmin precedes nextXid, the nextXid could have gotten
693 * advanced between our fetching it and applying the xmin below, perhaps
694 * far enough to make reply.xmin wrap around. In that case the xmin we
695 * set here would be "in the future" and have no effect. No point in
696 * worrying about this since it's too late to save the desired data
697 * anyway. Assuming that the standby sends us an increasing sequence of
698 * xmins, this could only happen during the first reply cycle, else our
699 * own xmin would prevent nextXid from advancing so far.
701 * We don't bother taking the ProcArrayLock here. Setting the xmin field
702 * is assumed atomic, and there's no real need to prevent a concurrent
703 * GetOldestXmin. (If we're moving our xmin forward, this is obviously
704 * safe, and if we're moving it backwards, well, the data is at risk
705 * already since a VACUUM could have just finished calling GetOldestXmin.)
707 MyPgXact->xmin = reply.xmin;
710 /* Main loop of walsender process */
714 char *output_message;
715 bool caughtup = false;
718 * Allocate buffer that will be used for each output message. We do this
719 * just once to reduce palloc overhead. The buffer must be made large
720 * enough for maximum-sized messages.
722 output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
725 * Allocate buffer that will be used for processing reply messages. As
726 * above, do this just once to reduce palloc overhead.
728 initStringInfo(&reply_message);
730 /* Initialize the last reply timestamp */
731 last_reply_timestamp = GetCurrentTimestamp();
733 /* Loop forever, unless we get an error */
736 /* Clear any already-pending wakeups */
737 ResetLatch(&MyWalSnd->latch);
740 * Emergency bailout if postmaster has died. This is to avoid the
741 * necessity for manual cleanup of all postmaster children.
743 if (!PostmasterIsAlive())
746 /* Process any requests or signals received recently */
750 ProcessConfigFile(PGC_SIGHUP);
754 /* Normal exit from the walsender is here */
755 if (walsender_shutdown_requested)
757 /* Inform the standby that XLOG streaming is done */
758 pq_puttextmessage('C', "COPY 0");
764 /* Check for input from the client */
765 ProcessRepliesIfAny();
768 * If we don't have any pending data in the output buffer, try to send
769 * some more. If there is some, we don't bother to call XLogSend
770 * again until we've flushed it ... but we'd better assume we are not
773 if (!pq_is_send_pending())
774 XLogSend(output_message, &caughtup);
778 /* Try to flush pending output to the client */
779 if (pq_flush_if_writable() != 0)
782 /* If nothing remains to be sent right now ... */
783 if (caughtup && !pq_is_send_pending())
786 * If we're in catchup state, move to streaming. This is an
787 * important state change for users to know about, since before
788 * this point data loss might occur if the primary dies and we
789 * need to failover to the standby. The state change is also
790 * important for synchronous replication, since commits that
791 * started to wait at that point might wait for some time.
793 if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
796 (errmsg("standby \"%s\" has now caught up with primary",
798 WalSndSetState(WALSNDSTATE_STREAMING);
802 * When SIGUSR2 arrives, we send any outstanding logs up to the
803 * shutdown checkpoint record (i.e., the latest record) and exit.
804 * This may be a normal termination at shutdown, or a promotion,
805 * the walsender is not sure which.
807 if (walsender_ready_to_stop)
809 /* ... let's just be real sure we're caught up ... */
810 XLogSend(output_message, &caughtup);
811 if (caughtup && !pq_is_send_pending())
813 walsender_shutdown_requested = true;
814 continue; /* don't want to wait more */
820 * We don't block if not caught up, unless there is unsent data
821 * pending in which case we'd better block until the socket is
822 * write-ready. This test is only needed for the case where XLogSend
823 * loaded a subset of the available data but then pq_flush_if_writable
824 * flushed it all --- we should immediately try to send more.
826 if (caughtup || pq_is_send_pending())
828 TimestampTz timeout = 0;
829 long sleeptime = 10000; /* 10 s */
832 wakeEvents = WL_LATCH_SET | WL_POSTMASTER_DEATH |
833 WL_SOCKET_READABLE | WL_TIMEOUT;
835 if (pq_is_send_pending())
836 wakeEvents |= WL_SOCKET_WRITEABLE;
839 WalSndKeepalive(output_message);
840 /* Try to flush pending output to the client */
841 if (pq_flush_if_writable() != 0)
845 /* Determine time until replication timeout */
846 if (replication_timeout > 0)
848 timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
849 replication_timeout);
850 sleeptime = 1 + (replication_timeout / 10);
853 /* Sleep until something happens or replication timeout */
854 WaitLatchOrSocket(&MyWalSnd->latch, wakeEvents,
855 MyProcPort->sock, sleeptime);
858 * Check for replication timeout. Note we ignore the corner case
859 * possibility that the client replied just as we reached the
860 * timeout ... he's supposed to reply *before* that.
862 if (replication_timeout > 0 &&
863 GetCurrentTimestamp() >= timeout)
866 * Since typically expiration of replication timeout means
867 * communication problem, we don't send the error message to
871 (errmsg("terminating walsender process due to replication timeout")));
878 * Get here on send failure. Clean up and exit.
880 * Reset whereToSendOutput to prevent ereport from attempting to send any
881 * more messages to the standby.
883 if (whereToSendOutput == DestRemote)
884 whereToSendOutput = DestNone;
887 return 1; /* keep the compiler quiet */
890 /* Initialize a per-walsender data structure for this walsender process */
897 * WalSndCtl should be set up already (we inherit this by fork() or
898 * EXEC_BACKEND mechanism from the postmaster).
900 Assert(WalSndCtl != NULL);
901 Assert(MyWalSnd == NULL);
904 * Find a free walsender slot and reserve it. If this fails, we must be
905 * out of WalSnd structures.
907 for (i = 0; i < max_wal_senders; i++)
909 /* use volatile pointer to prevent code rearrangement */
910 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
912 SpinLockAcquire(&walsnd->mutex);
914 if (walsnd->pid != 0)
916 SpinLockRelease(&walsnd->mutex);
922 * Found a free slot. Reserve it for us.
924 walsnd->pid = MyProcPid;
925 MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
926 walsnd->state = WALSNDSTATE_STARTUP;
927 SpinLockRelease(&walsnd->mutex);
928 /* don't need the lock anymore */
929 OwnLatch((Latch *) &walsnd->latch);
930 MyWalSnd = (WalSnd *) walsnd;
935 if (MyWalSnd == NULL)
937 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
938 errmsg("number of requested standby connections "
939 "exceeds max_wal_senders (currently %d)",
942 /* Arrange to clean up at walsender exit */
943 on_shmem_exit(WalSndKill, 0);
946 /* Destroy the per-walsender data structure for this walsender process */
948 WalSndKill(int code, Datum arg)
950 Assert(MyWalSnd != NULL);
953 * Mark WalSnd struct no longer in use. Assume that no lock is required
957 DisownLatch(&MyWalSnd->latch);
959 /* WalSnd struct isn't mine anymore */
964 * Read 'count' bytes from WAL into 'buf', starting at location 'startptr'
966 * XXX probably this should be improved to suck data directly from the
967 * WAL buffers when possible.
969 * Will open, and keep open, one WAL segment stored in the global file
970 * descriptor sendFile. This means if XLogRead is used once, there will
971 * always be one descriptor left open until the process ends, but never
975 XLogRead(char *buf, XLogRecPtr startptr, Size count)
980 uint32 lastRemovedLog;
981 uint32 lastRemovedSeg;
996 startoff = recptr.xrecoff % XLogSegSize;
998 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
1000 char path[MAXPGPATH];
1002 /* Switch to another logfile segment */
1006 XLByteToSeg(recptr, sendId, sendSeg);
1007 XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
1009 sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
1013 * If the file is not found, assume it's because the standby
1014 * asked for a too old WAL segment that has already been
1015 * removed or recycled.
1017 if (errno == ENOENT)
1019 char filename[MAXFNAMELEN];
1021 XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
1023 (errcode_for_file_access(),
1024 errmsg("requested WAL segment %s has already been removed",
1029 (errcode_for_file_access(),
1030 errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
1031 path, sendId, sendSeg)));
1036 /* Need to seek in the file? */
1037 if (sendOff != startoff)
1039 if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
1041 (errcode_for_file_access(),
1042 errmsg("could not seek in log file %u, segment %u to offset %u: %m",
1043 sendId, sendSeg, startoff)));
1047 /* How many bytes are within this segment? */
1048 if (nbytes > (XLogSegSize - startoff))
1049 segbytes = XLogSegSize - startoff;
1053 readbytes = read(sendFile, p, segbytes);
1056 (errcode_for_file_access(),
1057 errmsg("could not read from log file %u, segment %u, offset %u, "
1059 sendId, sendSeg, sendOff, (unsigned long) segbytes)));
1061 /* Update state for read */
1062 XLByteAdvance(recptr, readbytes);
1064 sendOff += readbytes;
1065 nbytes -= readbytes;
1070 * After reading into the buffer, check that what we read was valid. We do
1071 * this after reading, because even though the segment was present when we
1072 * opened it, it might get recycled or removed while we read it. The
1073 * read() succeeds in that case, but the data we tried to read might
1074 * already have been overwritten with new WAL records.
1076 XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
1077 XLByteToSeg(startptr, log, seg);
1078 if (log < lastRemovedLog ||
1079 (log == lastRemovedLog && seg <= lastRemovedSeg))
1081 char filename[MAXFNAMELEN];
1083 XLogFileName(filename, ThisTimeLineID, log, seg);
1085 (errcode_for_file_access(),
1086 errmsg("requested WAL segment %s has already been removed",
1091 * During recovery, the currently-open WAL file might be replaced with the
1092 * file of the same name retrieved from archive. So we always need to
1093 * check what we read was valid after reading into the buffer. If it's
1094 * invalid, we try to open and read the file again.
1096 if (am_cascading_walsender)
1098 /* use volatile pointer to prevent code rearrangement */
1099 volatile WalSnd *walsnd = MyWalSnd;
1102 SpinLockAcquire(&walsnd->mutex);
1103 reload = walsnd->needreload;
1104 walsnd->needreload = false;
1105 SpinLockRelease(&walsnd->mutex);
1107 if (reload && sendFile >= 0)
1118 * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
1119 * but not yet sent to the client, and buffer it in the libpq output
1122 * msgbuf is a work area in which the output message is constructed. It's
1123 * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
1124 * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
1126 * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
1127 * *caughtup is set to false.
1131 XLogSend(char *msgbuf, bool *caughtup)
1133 XLogRecPtr SendRqstPtr;
1134 XLogRecPtr startptr;
1137 WalDataMessageHeader msghdr;
1140 * Attempt to send all data that's already been written out and fsync'd to
1141 * disk. We cannot go further than what's been written out given the
1142 * current implementation of XLogRead(). And in any case it's unsafe to
1143 * send WAL that is not securely down to disk on the master: if the master
1144 * subsequently crashes and restarts, slaves must not have applied any WAL
1145 * that gets lost on the master.
1147 SendRqstPtr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetFlushRecPtr();
1149 /* Quick exit if nothing to do */
1150 if (XLByteLE(SendRqstPtr, sentPtr))
1157 * Figure out how much to send in one message. If there's no more than
1158 * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
1159 * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
1161 * The rounding is not only for performance reasons. Walreceiver relies on
1162 * the fact that we never split a WAL record across two messages. Since a
1163 * long WAL record is split at page boundary into continuation records,
1164 * page boundary is always a safe cut-off point. We also assume that
1165 * SendRqstPtr never points to the middle of a WAL record.
1168 if (startptr.xrecoff >= XLogFileSize)
1171 * crossing a logid boundary, skip the non-existent last log segment
1172 * in previous logical log file.
1174 startptr.xlogid += 1;
1175 startptr.xrecoff = 0;
1179 XLByteAdvance(endptr, MAX_SEND_SIZE);
1180 if (endptr.xlogid != startptr.xlogid)
1182 /* Don't cross a logfile boundary within one message */
1183 Assert(endptr.xlogid == startptr.xlogid + 1);
1184 endptr.xlogid = startptr.xlogid;
1185 endptr.xrecoff = XLogFileSize;
1188 /* if we went beyond SendRqstPtr, back off */
1189 if (XLByteLE(SendRqstPtr, endptr))
1191 endptr = SendRqstPtr;
1196 /* round down to page boundary. */
1197 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
1201 nbytes = endptr.xrecoff - startptr.xrecoff;
1202 Assert(nbytes <= MAX_SEND_SIZE);
1205 * OK to read and send the slice.
1210 * Read the log directly into the output buffer to avoid extra memcpy
1213 XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
1216 * We fill the message header last so that the send timestamp is taken as
1219 msghdr.dataStart = startptr;
1220 msghdr.walEnd = SendRqstPtr;
1221 msghdr.sendTime = GetCurrentTimestamp();
1223 memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
1225 pq_putmessage_noblock('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
1229 /* Update shared memory status */
1231 /* use volatile pointer to prevent code rearrangement */
1232 volatile WalSnd *walsnd = MyWalSnd;
1234 SpinLockAcquire(&walsnd->mutex);
1235 walsnd->sentPtr = sentPtr;
1236 SpinLockRelease(&walsnd->mutex);
1239 /* Report progress of XLOG streaming in PS display */
1240 if (update_process_title)
1242 char activitymsg[50];
1244 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
1245 sentPtr.xlogid, sentPtr.xrecoff);
1246 set_ps_display(activitymsg, false);
1253 * Request walsenders to reload the currently-open WAL file
1256 WalSndRqstFileReload(void)
1260 for (i = 0; i < max_wal_senders; i++)
1262 /* use volatile pointer to prevent code rearrangement */
1263 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1265 if (walsnd->pid == 0)
1268 SpinLockAcquire(&walsnd->mutex);
1269 walsnd->needreload = true;
1270 SpinLockRelease(&walsnd->mutex);
1274 /* SIGHUP: set flag to re-read config file at next convenient time */
1276 WalSndSigHupHandler(SIGNAL_ARGS)
1278 int save_errno = errno;
1282 SetLatch(&MyWalSnd->latch);
1287 /* SIGTERM: set flag to shut down */
1289 WalSndShutdownHandler(SIGNAL_ARGS)
1291 int save_errno = errno;
1293 walsender_shutdown_requested = true;
1295 SetLatch(&MyWalSnd->latch);
1298 * Set the standard (non-walsender) state as well, so that we can abort
1299 * things like do_pg_stop_backup().
1301 InterruptPending = true;
1302 ProcDiePending = true;
1308 * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
1310 * Some backend has bought the farm,
1311 * so we need to stop what we're doing and exit.
1314 WalSndQuickDieHandler(SIGNAL_ARGS)
1316 PG_SETMASK(&BlockSig);
1319 * We DO NOT want to run proc_exit() callbacks -- we're here because
1320 * shared memory may be corrupted, so we don't want to try to clean up our
1321 * transaction. Just nail the windows shut and get out of town. Now that
1322 * there's an atexit callback to prevent third-party code from breaking
1323 * things by calling exit() directly, we have to reset the callbacks
1324 * explicitly to make this work as intended.
1329 * Note we do exit(2) not exit(0). This is to force the postmaster into a
1330 * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
1331 * backend. This is necessary precisely because we don't clean up our
1332 * shared memory state. (The "dead man switch" mechanism in pmsignal.c
1333 * should ensure the postmaster sees this as a crash, too, but no harm in
1334 * being doubly sure.)
1339 /* SIGUSR1: set flag to send WAL records */
1341 WalSndXLogSendHandler(SIGNAL_ARGS)
1343 int save_errno = errno;
1345 latch_sigusr1_handler();
1350 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
1352 WalSndLastCycleHandler(SIGNAL_ARGS)
1354 int save_errno = errno;
1356 walsender_ready_to_stop = true;
1358 SetLatch(&MyWalSnd->latch);
1363 /* Set up signal handlers */
1367 /* Set up signal handlers */
1368 pqsignal(SIGHUP, WalSndSigHupHandler); /* set flag to read config
1370 pqsignal(SIGINT, SIG_IGN); /* not used */
1371 pqsignal(SIGTERM, WalSndShutdownHandler); /* request shutdown */
1372 pqsignal(SIGQUIT, WalSndQuickDieHandler); /* hard crash time */
1373 pqsignal(SIGALRM, SIG_IGN);
1374 pqsignal(SIGPIPE, SIG_IGN);
1375 pqsignal(SIGUSR1, WalSndXLogSendHandler); /* request WAL sending */
1376 pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
1379 /* Reset some signals that are accepted by postmaster but not here */
1380 pqsignal(SIGCHLD, SIG_DFL);
1381 pqsignal(SIGTTIN, SIG_DFL);
1382 pqsignal(SIGTTOU, SIG_DFL);
1383 pqsignal(SIGCONT, SIG_DFL);
1384 pqsignal(SIGWINCH, SIG_DFL);
1387 /* Report shared-memory space needed by WalSndShmemInit */
1389 WalSndShmemSize(void)
1393 size = offsetof(WalSndCtlData, walsnds);
1394 size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
1399 /* Allocate and initialize walsender-related shared memory */
1401 WalSndShmemInit(void)
1406 WalSndCtl = (WalSndCtlData *)
1407 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
1411 /* First time through, so initialize */
1412 MemSet(WalSndCtl, 0, WalSndShmemSize());
1414 for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
1415 SHMQueueInit(&(WalSndCtl->SyncRepQueue[i]));
1417 for (i = 0; i < max_wal_senders; i++)
1419 WalSnd *walsnd = &WalSndCtl->walsnds[i];
1421 SpinLockInit(&walsnd->mutex);
1422 InitSharedLatch(&walsnd->latch);
1427 /* Wake up all walsenders */
1433 for (i = 0; i < max_wal_senders; i++)
1434 SetLatch(&WalSndCtl->walsnds[i].latch);
1437 /* Set state for current walsender (only called in walsender) */
1439 WalSndSetState(WalSndState state)
1441 /* use volatile pointer to prevent code rearrangement */
1442 volatile WalSnd *walsnd = MyWalSnd;
1444 Assert(am_walsender);
1446 if (walsnd->state == state)
1449 SpinLockAcquire(&walsnd->mutex);
1450 walsnd->state = state;
1451 SpinLockRelease(&walsnd->mutex);
1455 * Return a string constant representing the state. This is used
1456 * in system views, and should *not* be translated.
1459 WalSndGetStateString(WalSndState state)
1463 case WALSNDSTATE_STARTUP:
1465 case WALSNDSTATE_BACKUP:
1467 case WALSNDSTATE_CATCHUP:
1469 case WALSNDSTATE_STREAMING:
1477 * Returns activity of walsenders, including pids and xlog locations sent to
1481 pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
1483 #define PG_STAT_GET_WAL_SENDERS_COLS 8
1484 ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
1486 Tuplestorestate *tupstore;
1487 MemoryContext per_query_ctx;
1488 MemoryContext oldcontext;
1491 int sync_standby = -1;
1494 /* check to see if caller supports us returning a tuplestore */
1495 if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
1497 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1498 errmsg("set-valued function called in context that cannot accept a set")));
1499 if (!(rsinfo->allowedModes & SFRM_Materialize))
1501 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1502 errmsg("materialize mode required, but it is not " \
1503 "allowed in this context")));
1505 /* Build a tuple descriptor for our result type */
1506 if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
1507 elog(ERROR, "return type must be a row type");
1509 per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
1510 oldcontext = MemoryContextSwitchTo(per_query_ctx);
1512 tupstore = tuplestore_begin_heap(true, false, work_mem);
1513 rsinfo->returnMode = SFRM_Materialize;
1514 rsinfo->setResult = tupstore;
1515 rsinfo->setDesc = tupdesc;
1517 MemoryContextSwitchTo(oldcontext);
1520 * Get the priorities of sync standbys all in one go, to minimise lock
1521 * acquisitions and to allow us to evaluate who is the current sync
1522 * standby. This code must match the code in SyncRepReleaseWaiters().
1524 sync_priority = palloc(sizeof(int) * max_wal_senders);
1525 LWLockAcquire(SyncRepLock, LW_SHARED);
1526 for (i = 0; i < max_wal_senders; i++)
1528 /* use volatile pointer to prevent code rearrangement */
1529 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1531 if (walsnd->pid != 0)
1533 sync_priority[i] = walsnd->sync_standby_priority;
1535 if (walsnd->state == WALSNDSTATE_STREAMING &&
1536 walsnd->sync_standby_priority > 0 &&
1538 priority > walsnd->sync_standby_priority))
1540 priority = walsnd->sync_standby_priority;
1545 LWLockRelease(SyncRepLock);
1547 for (i = 0; i < max_wal_senders; i++)
1549 /* use volatile pointer to prevent code rearrangement */
1550 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1551 char location[MAXFNAMELEN];
1557 Datum values[PG_STAT_GET_WAL_SENDERS_COLS];
1558 bool nulls[PG_STAT_GET_WAL_SENDERS_COLS];
1560 if (walsnd->pid == 0)
1563 SpinLockAcquire(&walsnd->mutex);
1564 sentPtr = walsnd->sentPtr;
1565 state = walsnd->state;
1566 write = walsnd->write;
1567 flush = walsnd->flush;
1568 apply = walsnd->apply;
1569 SpinLockRelease(&walsnd->mutex);
1571 memset(nulls, 0, sizeof(nulls));
1572 values[0] = Int32GetDatum(walsnd->pid);
1577 * Only superusers can see details. Other users only get the pid
1578 * value to know it's a walsender, but no details.
1580 MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
1584 values[1] = CStringGetTextDatum(WalSndGetStateString(state));
1586 snprintf(location, sizeof(location), "%X/%X",
1587 sentPtr.xlogid, sentPtr.xrecoff);
1588 values[2] = CStringGetTextDatum(location);
1590 if (write.xlogid == 0 && write.xrecoff == 0)
1592 snprintf(location, sizeof(location), "%X/%X",
1593 write.xlogid, write.xrecoff);
1594 values[3] = CStringGetTextDatum(location);
1596 if (flush.xlogid == 0 && flush.xrecoff == 0)
1598 snprintf(location, sizeof(location), "%X/%X",
1599 flush.xlogid, flush.xrecoff);
1600 values[4] = CStringGetTextDatum(location);
1602 if (apply.xlogid == 0 && apply.xrecoff == 0)
1604 snprintf(location, sizeof(location), "%X/%X",
1605 apply.xlogid, apply.xrecoff);
1606 values[5] = CStringGetTextDatum(location);
1608 values[6] = Int32GetDatum(sync_priority[i]);
1611 * More easily understood version of standby state. This is purely
1612 * informational, not different from priority.
1614 if (sync_priority[i] == 0)
1615 values[7] = CStringGetTextDatum("async");
1616 else if (i == sync_standby)
1617 values[7] = CStringGetTextDatum("sync");
1619 values[7] = CStringGetTextDatum("potential");
1622 tuplestore_putvalues(tupstore, tupdesc, values, nulls);
1624 pfree(sync_priority);
1626 /* clean up and return the tuplestore */
1627 tuplestore_donestoring(tupstore);
1633 WalSndKeepalive(char *msgbuf)
1635 PrimaryKeepaliveMessage keepalive_message;
1637 /* Construct a new message */
1638 keepalive_message.walEnd = sentPtr;
1639 keepalive_message.sendTime = GetCurrentTimestamp();
1641 elog(DEBUG2, "sending replication keepalive");
1643 /* Prepend with the message type and send it. */
1645 memcpy(msgbuf + 1, &keepalive_message, sizeof(PrimaryKeepaliveMessage));
1646 pq_putmessage_noblock('d', msgbuf, sizeof(PrimaryKeepaliveMessage) + 1);
1650 * This isn't currently used for anything. Monitoring tools might be
1651 * interested in the future, and we'll need something like this in the
1652 * future for synchronous replication.
1656 * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
1660 GetOldestWALSendPointer(void)
1662 XLogRecPtr oldest = {0, 0};
1666 for (i = 0; i < max_wal_senders; i++)
1668 /* use volatile pointer to prevent code rearrangement */
1669 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1672 if (walsnd->pid == 0)
1675 SpinLockAcquire(&walsnd->mutex);
1676 recptr = walsnd->sentPtr;
1677 SpinLockRelease(&walsnd->mutex);
1679 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1682 if (!found || XLByteLT(recptr, oldest))