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 another standby ? */
79 /* User-settable parameters for walsender */
80 int max_wal_senders = 0; /* the maximum number of concurrent walsenders */
81 int replication_timeout = 60 * 1000; /* maximum time to send one
85 * These variables are used similarly to openLogFile/Id/Seg/Off,
86 * but for walsender to read the XLOG.
88 static int sendFile = -1;
89 static uint32 sendId = 0;
90 static uint32 sendSeg = 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, 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 int WalSndLoop(void);
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",
304 logptr.xlogid, logptr.xrecoff);
306 /* Send a RowDescription message */
307 pq_beginmessage(&buf, 'T');
308 pq_sendint(&buf, 3, 2); /* 3 fields */
311 pq_sendstring(&buf, "systemid"); /* col name */
312 pq_sendint(&buf, 0, 4); /* table oid */
313 pq_sendint(&buf, 0, 2); /* attnum */
314 pq_sendint(&buf, TEXTOID, 4); /* type oid */
315 pq_sendint(&buf, -1, 2); /* typlen */
316 pq_sendint(&buf, 0, 4); /* typmod */
317 pq_sendint(&buf, 0, 2); /* format code */
320 pq_sendstring(&buf, "timeline"); /* col name */
321 pq_sendint(&buf, 0, 4); /* table oid */
322 pq_sendint(&buf, 0, 2); /* attnum */
323 pq_sendint(&buf, INT4OID, 4); /* type oid */
324 pq_sendint(&buf, 4, 2); /* typlen */
325 pq_sendint(&buf, 0, 4); /* typmod */
326 pq_sendint(&buf, 0, 2); /* format code */
329 pq_sendstring(&buf, "xlogpos");
330 pq_sendint(&buf, 0, 4);
331 pq_sendint(&buf, 0, 2);
332 pq_sendint(&buf, TEXTOID, 4);
333 pq_sendint(&buf, -1, 2);
334 pq_sendint(&buf, 0, 4);
335 pq_sendint(&buf, 0, 2);
338 /* Send a DataRow message */
339 pq_beginmessage(&buf, 'D');
340 pq_sendint(&buf, 3, 2); /* # of columns */
341 pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
342 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
343 pq_sendint(&buf, strlen(tli), 4); /* col2 len */
344 pq_sendbytes(&buf, (char *) tli, strlen(tli));
345 pq_sendint(&buf, strlen(xpos), 4); /* col3 len */
346 pq_sendbytes(&buf, (char *) xpos, strlen(xpos));
350 /* Send CommandComplete and ReadyForQuery messages */
351 EndCommand("SELECT", DestRemote);
352 ReadyForQuery(DestRemote);
353 /* ReadyForQuery did pq_flush for us */
360 StartReplication(StartReplicationCmd *cmd)
365 * Let postmaster know that we're streaming. Once we've declared us as a
366 * WAL sender process, postmaster will let us outlive the bgwriter and
367 * kill us last in the shutdown sequence, so we get a chance to stream all
368 * remaining WAL at shutdown, including the shutdown checkpoint. Note that
369 * there's no going back, and we mustn't write any WAL records after this.
371 MarkPostmasterChildWalSender();
372 SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
375 * When promoting a cascading standby, postmaster sends SIGUSR2 to
376 * any cascading walsenders to kill them. But there is a corner-case where
377 * such walsender fails to receive SIGUSR2 and survives a standby promotion
378 * unexpectedly. This happens when postmaster sends SIGUSR2 before
379 * the walsender marks itself as a WAL sender, because postmaster sends
380 * SIGUSR2 to only the processes marked as a WAL sender.
382 * To avoid this corner-case, if recovery is NOT in progress even though
383 * the walsender is cascading one, we do the same thing as SIGUSR2 signal
384 * handler does, i.e., set walsender_ready_to_stop to true. Which causes
385 * the walsender to end later.
387 * When terminating cascading walsenders, usually postmaster writes
388 * the log message announcing the terminations. But there is a race condition
389 * here. If there is no walsender except this process before reaching here,
390 * postmaster thinks that there is no walsender and suppresses that
391 * log message. To handle this case, we always emit that log message here.
392 * This might cause duplicate log messages, but which is less likely to happen,
393 * so it's not worth writing some code to suppress them.
395 if (am_cascading_walsender && !RecoveryInProgress())
398 (errmsg("terminating walsender process to force cascaded standby "
399 "to update timeline and reconnect")));
400 walsender_ready_to_stop = true;
404 * We assume here that we're logging enough information in the WAL for
405 * log-shipping, since this is checked in PostmasterMain().
407 * NOTE: wal_level can only change at shutdown, so in most cases it is
408 * difficult for there to be WAL data that we can still see that was written
409 * at wal_level='minimal'.
413 * When we first start replication the standby will be behind the primary.
414 * For some applications, for example, synchronous replication, it is
415 * important to have a clear state for this initial catchup mode, so we
416 * can trigger actions when we change streaming state later. We may stay
417 * in this state for a long time, which is exactly why we want to be able
418 * to monitor whether or not we are still here.
420 WalSndSetState(WALSNDSTATE_CATCHUP);
422 /* Send a CopyBothResponse message, and start streaming */
423 pq_beginmessage(&buf, 'W');
424 pq_sendbyte(&buf, 0);
425 pq_sendint(&buf, 0, 2);
430 * Initialize position to the received one, then the xlog records begin to
431 * be shipped from that position
433 sentPtr = cmd->startpoint;
437 * Execute an incoming replication command.
440 HandleReplicationCommand(const char *cmd_string)
442 bool replication_started = false;
445 MemoryContext cmd_context;
446 MemoryContext old_context;
448 elog(DEBUG1, "received replication command: %s", cmd_string);
450 cmd_context = AllocSetContextCreate(CurrentMemoryContext,
451 "Replication command context",
452 ALLOCSET_DEFAULT_MINSIZE,
453 ALLOCSET_DEFAULT_INITSIZE,
454 ALLOCSET_DEFAULT_MAXSIZE);
455 old_context = MemoryContextSwitchTo(cmd_context);
457 replication_scanner_init(cmd_string);
458 parse_rc = replication_yyparse();
461 (errcode(ERRCODE_SYNTAX_ERROR),
462 (errmsg_internal("replication command parser returned %d",
465 cmd_node = replication_parse_result;
467 switch (cmd_node->type)
469 case T_IdentifySystemCmd:
473 case T_StartReplicationCmd:
474 StartReplication((StartReplicationCmd *) cmd_node);
476 /* break out of the loop */
477 replication_started = true;
480 case T_BaseBackupCmd:
481 SendBaseBackup((BaseBackupCmd *) cmd_node);
483 /* Send CommandComplete and ReadyForQuery messages */
484 EndCommand("SELECT", DestRemote);
485 ReadyForQuery(DestRemote);
486 /* ReadyForQuery did pq_flush for us */
491 (errcode(ERRCODE_PROTOCOL_VIOLATION),
492 errmsg("invalid standby query string: %s", cmd_string)));
496 MemoryContextSwitchTo(old_context);
497 MemoryContextDelete(cmd_context);
499 return replication_started;
503 * Check if the remote end has closed the connection.
506 ProcessRepliesIfAny(void)
508 unsigned char firstchar;
510 bool received = false;
514 r = pq_getbyte_if_available(&firstchar);
517 /* unexpected error or EOF */
519 (errcode(ERRCODE_PROTOCOL_VIOLATION),
520 errmsg("unexpected EOF on standby connection")));
525 /* no data available without blocking */
529 /* Handle the very limited subset of commands expected in this phase */
533 * 'd' means a standby reply wrapped in a CopyData packet.
536 ProcessStandbyMessage();
541 * 'X' means that the standby is closing down the socket.
548 (errcode(ERRCODE_PROTOCOL_VIOLATION),
549 errmsg("invalid standby message type \"%c\"",
555 * Save the last reply timestamp if we've received at least one reply.
558 last_reply_timestamp = GetCurrentTimestamp();
562 * Process a status update message received from standby.
565 ProcessStandbyMessage(void)
569 resetStringInfo(&reply_message);
572 * Read the message contents.
574 if (pq_getmessage(&reply_message, 0))
577 (errcode(ERRCODE_PROTOCOL_VIOLATION),
578 errmsg("unexpected EOF on standby connection")));
583 * Check message type from the first byte.
585 msgtype = pq_getmsgbyte(&reply_message);
590 ProcessStandbyReplyMessage();
594 ProcessStandbyHSFeedbackMessage();
599 (errcode(ERRCODE_PROTOCOL_VIOLATION),
600 errmsg("unexpected message type \"%c\"", msgtype)));
606 * Regular reply from standby advising of WAL positions on standby server.
609 ProcessStandbyReplyMessage(void)
611 StandbyReplyMessage reply;
613 pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyReplyMessage));
615 elog(DEBUG2, "write %X/%X flush %X/%X apply %X/%X",
616 reply.write.xlogid, reply.write.xrecoff,
617 reply.flush.xlogid, reply.flush.xrecoff,
618 reply.apply.xlogid, reply.apply.xrecoff);
621 * Update shared state for this WalSender process based on reply data from
625 /* use volatile pointer to prevent code rearrangement */
626 volatile WalSnd *walsnd = MyWalSnd;
628 SpinLockAcquire(&walsnd->mutex);
629 walsnd->write = reply.write;
630 walsnd->flush = reply.flush;
631 walsnd->apply = reply.apply;
632 SpinLockRelease(&walsnd->mutex);
635 if (!am_cascading_walsender)
636 SyncRepReleaseWaiters();
640 * Hot Standby feedback
643 ProcessStandbyHSFeedbackMessage(void)
645 StandbyHSFeedbackMessage reply;
646 TransactionId nextXid;
649 /* Decipher the reply message */
650 pq_copymsgbytes(&reply_message, (char *) &reply,
651 sizeof(StandbyHSFeedbackMessage));
653 elog(DEBUG2, "hot standby feedback xmin %u epoch %u",
657 /* Ignore invalid xmin (can't actually happen with current walreceiver) */
658 if (!TransactionIdIsNormal(reply.xmin))
662 * Check that the provided xmin/epoch are sane, that is, not in the future
663 * and not so far back as to be already wrapped around. Ignore if not.
665 * Epoch of nextXid should be same as standby, or if the counter has
666 * wrapped, then one greater than standby.
668 GetNextXidAndEpoch(&nextXid, &nextEpoch);
670 if (reply.xmin <= nextXid)
672 if (reply.epoch != nextEpoch)
677 if (reply.epoch + 1 != nextEpoch)
681 if (!TransactionIdPrecedesOrEquals(reply.xmin, nextXid))
682 return; /* epoch OK, but it's wrapped around */
685 * Set the WalSender's xmin equal to the standby's requested xmin, so that
686 * the xmin will be taken into account by GetOldestXmin. This will hold
687 * back the removal of dead rows and thereby prevent the generation of
688 * cleanup conflicts on the standby server.
690 * There is a small window for a race condition here: although we just
691 * checked that reply.xmin precedes nextXid, the nextXid could have gotten
692 * advanced between our fetching it and applying the xmin below, perhaps
693 * far enough to make reply.xmin wrap around. In that case the xmin we
694 * set here would be "in the future" and have no effect. No point in
695 * worrying about this since it's too late to save the desired data
696 * anyway. Assuming that the standby sends us an increasing sequence of
697 * xmins, this could only happen during the first reply cycle, else our
698 * own xmin would prevent nextXid from advancing so far.
700 * We don't bother taking the ProcArrayLock here. Setting the xmin field
701 * is assumed atomic, and there's no real need to prevent a concurrent
702 * GetOldestXmin. (If we're moving our xmin forward, this is obviously
703 * safe, and if we're moving it backwards, well, the data is at risk
704 * already since a VACUUM could have just finished calling GetOldestXmin.)
706 MyPgXact->xmin = reply.xmin;
709 /* Main loop of walsender process */
713 char *output_message;
714 bool caughtup = false;
717 * Allocate buffer that will be used for each output message. We do this
718 * just once to reduce palloc overhead. The buffer must be made large
719 * enough for maximum-sized messages.
721 output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
724 * Allocate buffer that will be used for processing reply messages. As
725 * above, do this just once to reduce palloc overhead.
727 initStringInfo(&reply_message);
729 /* Initialize the last reply timestamp */
730 last_reply_timestamp = GetCurrentTimestamp();
732 /* Loop forever, unless we get an error */
735 /* Clear any already-pending wakeups */
736 ResetLatch(&MyWalSnd->latch);
739 * Emergency bailout if postmaster has died. This is to avoid the
740 * necessity for manual cleanup of all postmaster children.
742 if (!PostmasterIsAlive())
745 /* Process any requests or signals received recently */
749 ProcessConfigFile(PGC_SIGHUP);
753 /* Normal exit from the walsender is here */
754 if (walsender_shutdown_requested)
756 /* Inform the standby that XLOG streaming is done */
757 pq_puttextmessage('C', "COPY 0");
763 /* Check for input from the client */
764 ProcessRepliesIfAny();
767 * If we don't have any pending data in the output buffer, try to send
768 * some more. If there is some, we don't bother to call XLogSend
769 * again until we've flushed it ... but we'd better assume we are not
772 if (!pq_is_send_pending())
773 XLogSend(output_message, &caughtup);
777 /* Try to flush pending output to the client */
778 if (pq_flush_if_writable() != 0)
781 /* If nothing remains to be sent right now ... */
782 if (caughtup && !pq_is_send_pending())
785 * If we're in catchup state, move to streaming. This is an
786 * important state change for users to know about, since before
787 * this point data loss might occur if the primary dies and we
788 * need to failover to the standby. The state change is also
789 * important for synchronous replication, since commits that
790 * started to wait at that point might wait for some time.
792 if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
795 (errmsg("standby \"%s\" has now caught up with primary",
797 WalSndSetState(WALSNDSTATE_STREAMING);
801 * When SIGUSR2 arrives, we send any outstanding logs up to the
802 * shutdown checkpoint record (i.e., the latest record) and exit.
803 * This may be a normal termination at shutdown, or a promotion,
804 * the walsender is not sure which.
806 if (walsender_ready_to_stop)
808 /* ... let's just be real sure we're caught up ... */
809 XLogSend(output_message, &caughtup);
810 if (caughtup && !pq_is_send_pending())
812 walsender_shutdown_requested = true;
813 continue; /* don't want to wait more */
819 * We don't block if not caught up, unless there is unsent data
820 * pending in which case we'd better block until the socket is
821 * write-ready. This test is only needed for the case where XLogSend
822 * loaded a subset of the available data but then pq_flush_if_writable
823 * flushed it all --- we should immediately try to send more.
825 if (caughtup || pq_is_send_pending())
827 TimestampTz timeout = 0;
828 long sleeptime = 10000; /* 10 s */
831 wakeEvents = WL_LATCH_SET | WL_POSTMASTER_DEATH |
832 WL_SOCKET_READABLE | WL_TIMEOUT;
834 if (pq_is_send_pending())
835 wakeEvents |= WL_SOCKET_WRITEABLE;
838 WalSndKeepalive(output_message);
839 /* Try to flush pending output to the client */
840 if (pq_flush_if_writable() != 0)
844 /* Determine time until replication timeout */
845 if (replication_timeout > 0)
847 timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
848 replication_timeout);
849 sleeptime = 1 + (replication_timeout / 10);
852 /* Sleep until something happens or replication timeout */
853 WaitLatchOrSocket(&MyWalSnd->latch, wakeEvents,
854 MyProcPort->sock, sleeptime);
857 * Check for replication timeout. Note we ignore the corner case
858 * possibility that the client replied just as we reached the
859 * timeout ... he's supposed to reply *before* that.
861 if (replication_timeout > 0 &&
862 GetCurrentTimestamp() >= timeout)
865 * Since typically expiration of replication timeout means
866 * communication problem, we don't send the error message to
870 (errmsg("terminating walsender process due to replication timeout")));
877 * Get here on send failure. Clean up and exit.
879 * Reset whereToSendOutput to prevent ereport from attempting to send any
880 * more messages to the standby.
882 if (whereToSendOutput == DestRemote)
883 whereToSendOutput = DestNone;
886 return 1; /* keep the compiler quiet */
889 /* Initialize a per-walsender data structure for this walsender process */
896 * WalSndCtl should be set up already (we inherit this by fork() or
897 * EXEC_BACKEND mechanism from the postmaster).
899 Assert(WalSndCtl != NULL);
900 Assert(MyWalSnd == NULL);
903 * Find a free walsender slot and reserve it. If this fails, we must be
904 * out of WalSnd structures.
906 for (i = 0; i < max_wal_senders; i++)
908 /* use volatile pointer to prevent code rearrangement */
909 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
911 SpinLockAcquire(&walsnd->mutex);
913 if (walsnd->pid != 0)
915 SpinLockRelease(&walsnd->mutex);
921 * Found a free slot. Reserve it for us.
923 walsnd->pid = MyProcPid;
924 MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
925 walsnd->state = WALSNDSTATE_STARTUP;
926 SpinLockRelease(&walsnd->mutex);
927 /* don't need the lock anymore */
928 OwnLatch((Latch *) &walsnd->latch);
929 MyWalSnd = (WalSnd *) walsnd;
934 if (MyWalSnd == NULL)
936 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
937 errmsg("number of requested standby connections "
938 "exceeds max_wal_senders (currently %d)",
941 /* Arrange to clean up at walsender exit */
942 on_shmem_exit(WalSndKill, 0);
945 /* Destroy the per-walsender data structure for this walsender process */
947 WalSndKill(int code, Datum arg)
949 Assert(MyWalSnd != NULL);
952 * Mark WalSnd struct no longer in use. Assume that no lock is required
956 DisownLatch(&MyWalSnd->latch);
958 /* WalSnd struct isn't mine anymore */
963 * Read 'count' bytes from WAL into 'buf', starting at location 'startptr'
965 * XXX probably this should be improved to suck data directly from the
966 * WAL buffers when possible.
968 * Will open, and keep open, one WAL segment stored in the global file
969 * descriptor sendFile. This means if XLogRead is used once, there will
970 * always be one descriptor left open until the process ends, but never
974 XLogRead(char *buf, XLogRecPtr startptr, Size count)
979 uint32 lastRemovedLog;
980 uint32 lastRemovedSeg;
995 startoff = recptr.xrecoff % XLogSegSize;
997 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
999 char path[MAXPGPATH];
1001 /* Switch to another logfile segment */
1005 XLByteToSeg(recptr, sendId, sendSeg);
1006 XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
1008 sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
1012 * If the file is not found, assume it's because the standby
1013 * asked for a too old WAL segment that has already been
1014 * removed or recycled.
1016 if (errno == ENOENT)
1018 char filename[MAXFNAMELEN];
1020 XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
1022 (errcode_for_file_access(),
1023 errmsg("requested WAL segment %s has already been removed",
1028 (errcode_for_file_access(),
1029 errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
1030 path, sendId, sendSeg)));
1035 /* Need to seek in the file? */
1036 if (sendOff != startoff)
1038 if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
1040 (errcode_for_file_access(),
1041 errmsg("could not seek in log file %u, segment %u to offset %u: %m",
1042 sendId, sendSeg, startoff)));
1046 /* How many bytes are within this segment? */
1047 if (nbytes > (XLogSegSize - startoff))
1048 segbytes = XLogSegSize - startoff;
1052 readbytes = read(sendFile, p, segbytes);
1055 (errcode_for_file_access(),
1056 errmsg("could not read from log file %u, segment %u, offset %u, "
1058 sendId, sendSeg, sendOff, (unsigned long) segbytes)));
1060 /* Update state for read */
1061 XLByteAdvance(recptr, readbytes);
1063 sendOff += readbytes;
1064 nbytes -= readbytes;
1069 * After reading into the buffer, check that what we read was valid. We do
1070 * this after reading, because even though the segment was present when we
1071 * opened it, it might get recycled or removed while we read it. The
1072 * read() succeeds in that case, but the data we tried to read might
1073 * already have been overwritten with new WAL records.
1075 XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
1076 XLByteToSeg(startptr, log, seg);
1077 if (log < lastRemovedLog ||
1078 (log == lastRemovedLog && seg <= lastRemovedSeg))
1080 char filename[MAXFNAMELEN];
1082 XLogFileName(filename, ThisTimeLineID, log, seg);
1084 (errcode_for_file_access(),
1085 errmsg("requested WAL segment %s has already been removed",
1090 * During recovery, the currently-open WAL file might be replaced with
1091 * the file of the same name retrieved from archive. So we always need
1092 * to check what we read was valid after reading into the buffer. If it's
1093 * invalid, we try to open and read the file again.
1095 if (am_cascading_walsender)
1097 /* use volatile pointer to prevent code rearrangement */
1098 volatile WalSnd *walsnd = MyWalSnd;
1101 SpinLockAcquire(&walsnd->mutex);
1102 reload = walsnd->needreload;
1103 walsnd->needreload = false;
1104 SpinLockRelease(&walsnd->mutex);
1106 if (reload && sendFile >= 0)
1117 * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
1118 * but not yet sent to the client, and buffer it in the libpq output
1121 * msgbuf is a work area in which the output message is constructed. It's
1122 * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
1123 * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
1125 * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
1126 * *caughtup is set to false.
1130 XLogSend(char *msgbuf, bool *caughtup)
1132 XLogRecPtr SendRqstPtr;
1133 XLogRecPtr startptr;
1136 WalDataMessageHeader msghdr;
1139 * Attempt to send all data that's already been written out and fsync'd to
1140 * disk. We cannot go further than what's been written out given the
1141 * current implementation of XLogRead(). And in any case it's unsafe to
1142 * send WAL that is not securely down to disk on the master: if the master
1143 * subsequently crashes and restarts, slaves must not have applied any WAL
1144 * that gets lost on the master.
1146 SendRqstPtr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetFlushRecPtr();
1148 /* Quick exit if nothing to do */
1149 if (XLByteLE(SendRqstPtr, sentPtr))
1156 * Figure out how much to send in one message. If there's no more than
1157 * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
1158 * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
1160 * The rounding is not only for performance reasons. Walreceiver relies on
1161 * the fact that we never split a WAL record across two messages. Since a
1162 * long WAL record is split at page boundary into continuation records,
1163 * page boundary is always a safe cut-off point. We also assume that
1164 * SendRqstPtr never points to the middle of a WAL record.
1167 if (startptr.xrecoff >= XLogFileSize)
1170 * crossing a logid boundary, skip the non-existent last log segment
1171 * in previous logical log file.
1173 startptr.xlogid += 1;
1174 startptr.xrecoff = 0;
1178 XLByteAdvance(endptr, MAX_SEND_SIZE);
1179 if (endptr.xlogid != startptr.xlogid)
1181 /* Don't cross a logfile boundary within one message */
1182 Assert(endptr.xlogid == startptr.xlogid + 1);
1183 endptr.xlogid = startptr.xlogid;
1184 endptr.xrecoff = XLogFileSize;
1187 /* if we went beyond SendRqstPtr, back off */
1188 if (XLByteLE(SendRqstPtr, endptr))
1190 endptr = SendRqstPtr;
1195 /* round down to page boundary. */
1196 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
1200 nbytes = endptr.xrecoff - startptr.xrecoff;
1201 Assert(nbytes <= MAX_SEND_SIZE);
1204 * OK to read and send the slice.
1209 * Read the log directly into the output buffer to avoid extra memcpy
1212 XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
1215 * We fill the message header last so that the send timestamp is taken as
1218 msghdr.dataStart = startptr;
1219 msghdr.walEnd = SendRqstPtr;
1220 msghdr.sendTime = GetCurrentTimestamp();
1222 memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
1224 pq_putmessage_noblock('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
1228 /* Update shared memory status */
1230 /* use volatile pointer to prevent code rearrangement */
1231 volatile WalSnd *walsnd = MyWalSnd;
1233 SpinLockAcquire(&walsnd->mutex);
1234 walsnd->sentPtr = sentPtr;
1235 SpinLockRelease(&walsnd->mutex);
1238 /* Report progress of XLOG streaming in PS display */
1239 if (update_process_title)
1241 char activitymsg[50];
1243 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
1244 sentPtr.xlogid, sentPtr.xrecoff);
1245 set_ps_display(activitymsg, false);
1252 * Request walsenders to reload the currently-open WAL file
1255 WalSndRqstFileReload(void)
1259 for (i = 0; i < max_wal_senders; i++)
1261 /* use volatile pointer to prevent code rearrangement */
1262 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1264 if (walsnd->pid == 0)
1267 SpinLockAcquire(&walsnd->mutex);
1268 walsnd->needreload = true;
1269 SpinLockRelease(&walsnd->mutex);
1273 /* SIGHUP: set flag to re-read config file at next convenient time */
1275 WalSndSigHupHandler(SIGNAL_ARGS)
1277 int save_errno = errno;
1281 SetLatch(&MyWalSnd->latch);
1286 /* SIGTERM: set flag to shut down */
1288 WalSndShutdownHandler(SIGNAL_ARGS)
1290 int save_errno = errno;
1292 walsender_shutdown_requested = true;
1294 SetLatch(&MyWalSnd->latch);
1297 * Set the standard (non-walsender) state as well, so that we can
1298 * abort things like do_pg_stop_backup().
1300 InterruptPending = true;
1301 ProcDiePending = true;
1307 * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
1309 * Some backend has bought the farm,
1310 * so we need to stop what we're doing and exit.
1313 WalSndQuickDieHandler(SIGNAL_ARGS)
1315 PG_SETMASK(&BlockSig);
1318 * We DO NOT want to run proc_exit() callbacks -- we're here because
1319 * shared memory may be corrupted, so we don't want to try to clean up our
1320 * transaction. Just nail the windows shut and get out of town. Now that
1321 * there's an atexit callback to prevent third-party code from breaking
1322 * things by calling exit() directly, we have to reset the callbacks
1323 * explicitly to make this work as intended.
1328 * Note we do exit(2) not exit(0). This is to force the postmaster into a
1329 * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
1330 * backend. This is necessary precisely because we don't clean up our
1331 * shared memory state. (The "dead man switch" mechanism in pmsignal.c
1332 * should ensure the postmaster sees this as a crash, too, but no harm in
1333 * being doubly sure.)
1338 /* SIGUSR1: set flag to send WAL records */
1340 WalSndXLogSendHandler(SIGNAL_ARGS)
1342 int save_errno = errno;
1344 latch_sigusr1_handler();
1349 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
1351 WalSndLastCycleHandler(SIGNAL_ARGS)
1353 int save_errno = errno;
1355 walsender_ready_to_stop = true;
1357 SetLatch(&MyWalSnd->latch);
1362 /* Set up signal handlers */
1366 /* Set up signal handlers */
1367 pqsignal(SIGHUP, WalSndSigHupHandler); /* set flag to read config
1369 pqsignal(SIGINT, SIG_IGN); /* not used */
1370 pqsignal(SIGTERM, WalSndShutdownHandler); /* request shutdown */
1371 pqsignal(SIGQUIT, WalSndQuickDieHandler); /* hard crash time */
1372 pqsignal(SIGALRM, SIG_IGN);
1373 pqsignal(SIGPIPE, SIG_IGN);
1374 pqsignal(SIGUSR1, WalSndXLogSendHandler); /* request WAL sending */
1375 pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
1378 /* Reset some signals that are accepted by postmaster but not here */
1379 pqsignal(SIGCHLD, SIG_DFL);
1380 pqsignal(SIGTTIN, SIG_DFL);
1381 pqsignal(SIGTTOU, SIG_DFL);
1382 pqsignal(SIGCONT, SIG_DFL);
1383 pqsignal(SIGWINCH, SIG_DFL);
1386 /* Report shared-memory space needed by WalSndShmemInit */
1388 WalSndShmemSize(void)
1392 size = offsetof(WalSndCtlData, walsnds);
1393 size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
1398 /* Allocate and initialize walsender-related shared memory */
1400 WalSndShmemInit(void)
1405 WalSndCtl = (WalSndCtlData *)
1406 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
1410 /* First time through, so initialize */
1411 MemSet(WalSndCtl, 0, WalSndShmemSize());
1413 for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
1414 SHMQueueInit(&(WalSndCtl->SyncRepQueue[i]));
1416 for (i = 0; i < max_wal_senders; i++)
1418 WalSnd *walsnd = &WalSndCtl->walsnds[i];
1420 SpinLockInit(&walsnd->mutex);
1421 InitSharedLatch(&walsnd->latch);
1426 /* Wake up all walsenders */
1432 for (i = 0; i < max_wal_senders; i++)
1433 SetLatch(&WalSndCtl->walsnds[i].latch);
1436 /* Set state for current walsender (only called in walsender) */
1438 WalSndSetState(WalSndState state)
1440 /* use volatile pointer to prevent code rearrangement */
1441 volatile WalSnd *walsnd = MyWalSnd;
1443 Assert(am_walsender);
1445 if (walsnd->state == state)
1448 SpinLockAcquire(&walsnd->mutex);
1449 walsnd->state = state;
1450 SpinLockRelease(&walsnd->mutex);
1454 * Return a string constant representing the state. This is used
1455 * in system views, and should *not* be translated.
1458 WalSndGetStateString(WalSndState state)
1462 case WALSNDSTATE_STARTUP:
1464 case WALSNDSTATE_BACKUP:
1466 case WALSNDSTATE_CATCHUP:
1468 case WALSNDSTATE_STREAMING:
1476 * Returns activity of walsenders, including pids and xlog locations sent to
1480 pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
1482 #define PG_STAT_GET_WAL_SENDERS_COLS 8
1483 ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
1485 Tuplestorestate *tupstore;
1486 MemoryContext per_query_ctx;
1487 MemoryContext oldcontext;
1490 int sync_standby = -1;
1493 /* check to see if caller supports us returning a tuplestore */
1494 if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
1496 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1497 errmsg("set-valued function called in context that cannot accept a set")));
1498 if (!(rsinfo->allowedModes & SFRM_Materialize))
1500 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1501 errmsg("materialize mode required, but it is not " \
1502 "allowed in this context")));
1504 /* Build a tuple descriptor for our result type */
1505 if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
1506 elog(ERROR, "return type must be a row type");
1508 per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
1509 oldcontext = MemoryContextSwitchTo(per_query_ctx);
1511 tupstore = tuplestore_begin_heap(true, false, work_mem);
1512 rsinfo->returnMode = SFRM_Materialize;
1513 rsinfo->setResult = tupstore;
1514 rsinfo->setDesc = tupdesc;
1516 MemoryContextSwitchTo(oldcontext);
1519 * Get the priorities of sync standbys all in one go, to minimise lock
1520 * acquisitions and to allow us to evaluate who is the current sync
1521 * standby. This code must match the code in SyncRepReleaseWaiters().
1523 sync_priority = palloc(sizeof(int) * max_wal_senders);
1524 LWLockAcquire(SyncRepLock, LW_SHARED);
1525 for (i = 0; i < max_wal_senders; i++)
1527 /* use volatile pointer to prevent code rearrangement */
1528 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1530 if (walsnd->pid != 0)
1532 sync_priority[i] = walsnd->sync_standby_priority;
1534 if (walsnd->state == WALSNDSTATE_STREAMING &&
1535 walsnd->sync_standby_priority > 0 &&
1537 priority > walsnd->sync_standby_priority))
1539 priority = walsnd->sync_standby_priority;
1544 LWLockRelease(SyncRepLock);
1546 for (i = 0; i < max_wal_senders; i++)
1548 /* use volatile pointer to prevent code rearrangement */
1549 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1550 char location[MAXFNAMELEN];
1556 Datum values[PG_STAT_GET_WAL_SENDERS_COLS];
1557 bool nulls[PG_STAT_GET_WAL_SENDERS_COLS];
1559 if (walsnd->pid == 0)
1562 SpinLockAcquire(&walsnd->mutex);
1563 sentPtr = walsnd->sentPtr;
1564 state = walsnd->state;
1565 write = walsnd->write;
1566 flush = walsnd->flush;
1567 apply = walsnd->apply;
1568 SpinLockRelease(&walsnd->mutex);
1570 memset(nulls, 0, sizeof(nulls));
1571 values[0] = Int32GetDatum(walsnd->pid);
1576 * Only superusers can see details. Other users only get the pid
1577 * value to know it's a walsender, but no details.
1579 MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
1583 values[1] = CStringGetTextDatum(WalSndGetStateString(state));
1585 snprintf(location, sizeof(location), "%X/%X",
1586 sentPtr.xlogid, sentPtr.xrecoff);
1587 values[2] = CStringGetTextDatum(location);
1589 if (write.xlogid == 0 && write.xrecoff == 0)
1591 snprintf(location, sizeof(location), "%X/%X",
1592 write.xlogid, write.xrecoff);
1593 values[3] = CStringGetTextDatum(location);
1595 if (flush.xlogid == 0 && flush.xrecoff == 0)
1597 snprintf(location, sizeof(location), "%X/%X",
1598 flush.xlogid, flush.xrecoff);
1599 values[4] = CStringGetTextDatum(location);
1601 if (apply.xlogid == 0 && apply.xrecoff == 0)
1603 snprintf(location, sizeof(location), "%X/%X",
1604 apply.xlogid, apply.xrecoff);
1605 values[5] = CStringGetTextDatum(location);
1607 values[6] = Int32GetDatum(sync_priority[i]);
1610 * More easily understood version of standby state. This is purely
1611 * informational, not different from priority.
1613 if (sync_priority[i] == 0)
1614 values[7] = CStringGetTextDatum("async");
1615 else if (i == sync_standby)
1616 values[7] = CStringGetTextDatum("sync");
1618 values[7] = CStringGetTextDatum("potential");
1621 tuplestore_putvalues(tupstore, tupdesc, values, nulls);
1623 pfree(sync_priority);
1625 /* clean up and return the tuplestore */
1626 tuplestore_donestoring(tupstore);
1632 WalSndKeepalive(char *msgbuf)
1634 PrimaryKeepaliveMessage keepalive_message;
1636 /* Construct a new message */
1637 keepalive_message.walEnd = sentPtr;
1638 keepalive_message.sendTime = GetCurrentTimestamp();
1640 elog(DEBUG2, "sending replication keepalive");
1642 /* Prepend with the message type and send it. */
1644 memcpy(msgbuf + 1, &keepalive_message, sizeof(PrimaryKeepaliveMessage));
1645 pq_putmessage_noblock('d', msgbuf, sizeof(PrimaryKeepaliveMessage) + 1);
1649 * This isn't currently used for anything. Monitoring tools might be
1650 * interested in the future, and we'll need something like this in the
1651 * future for synchronous replication.
1655 * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
1659 GetOldestWALSendPointer(void)
1661 XLogRecPtr oldest = {0, 0};
1665 for (i = 0; i < max_wal_senders; i++)
1667 /* use volatile pointer to prevent code rearrangement */
1668 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1671 if (walsnd->pid == 0)
1674 SpinLockAcquire(&walsnd->mutex);
1675 recptr = walsnd->sentPtr;
1676 SpinLockRelease(&walsnd->mutex);
1678 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1681 if (!found || XLByteLT(recptr, oldest))