1 /*-------------------------------------------------------------------------
5 * The WAL sender process (walsender) is new as of Postgres 9.0. It takes
6 * charge of XLOG streaming sender in the primary server. At first, it is
7 * started by the postmaster when the walreceiver in the standby server
8 * connects to the primary server and requests XLOG streaming replication.
9 * It attempts to keep reading XLOG records from the disk and sending them
10 * to the standby server, as long as the connection is alive (i.e., like
11 * any backend, there is a one-to-one relationship between a connection
12 * and a walsender process).
14 * Normal termination is by SIGTERM, which instructs the walsender to
15 * close the connection and exit(0) at next convenient moment. Emergency
16 * termination is by SIGQUIT; like any backend, the walsender will simply
17 * abort and exit on SIGQUIT. A close of the connection and a FATAL error
18 * are treated as not a crash but approximately normal termination;
19 * the walsender will exit quickly without sending any more XLOG records.
21 * If the server is shut down, postmaster sends us SIGUSR2 after all
22 * regular backends have exited and the shutdown checkpoint has been written.
23 * This instruct walsender to send any outstanding WAL, including the
24 * shutdown checkpoint record, and then exit.
26 * Note that there can be more than one walsender process concurrently.
28 * Portions Copyright (c) 2010-2010, PostgreSQL Global Development Group
32 * $PostgreSQL: pgsql/src/backend/replication/walsender.c,v 1.25 2010/06/03 22:17:32 tgl Exp $
34 *-------------------------------------------------------------------------
40 #include "access/xlog_internal.h"
41 #include "catalog/pg_type.h"
42 #include "libpq/libpq.h"
43 #include "libpq/pqformat.h"
44 #include "libpq/pqsignal.h"
45 #include "miscadmin.h"
46 #include "replication/walprotocol.h"
47 #include "replication/walsender.h"
48 #include "storage/fd.h"
49 #include "storage/ipc.h"
50 #include "storage/pmsignal.h"
51 #include "tcop/tcopprot.h"
52 #include "utils/guc.h"
53 #include "utils/memutils.h"
54 #include "utils/ps_status.h"
57 /* Array of WalSnds in shared memory */
58 WalSndCtlData *WalSndCtl = NULL;
60 /* My slot in the shared memory array */
61 static WalSnd *MyWalSnd = NULL;
64 bool am_walsender = false; /* Am I a walsender process ? */
66 /* User-settable parameters for walsender */
67 int max_wal_senders = 0; /* the maximum number of concurrent walsenders */
68 int WalSndDelay = 200; /* max sleep time between some actions */
70 #define NAPTIME_PER_CYCLE 100000L /* max sleep time between cycles (100ms) */
73 * These variables are used similarly to openLogFile/Id/Seg/Off,
74 * but for walsender to read the XLOG.
76 static int sendFile = -1;
77 static uint32 sendId = 0;
78 static uint32 sendSeg = 0;
79 static uint32 sendOff = 0;
82 * How far have we sent WAL already? This is also advertised in
83 * MyWalSnd->sentPtr. (Actually, this is the next WAL location to send.)
85 static XLogRecPtr sentPtr = {0, 0};
87 /* Flags set by signal handlers for later service in main loop */
88 static volatile sig_atomic_t got_SIGHUP = false;
89 static volatile sig_atomic_t shutdown_requested = false;
90 static volatile sig_atomic_t ready_to_stop = false;
93 static void WalSndSigHupHandler(SIGNAL_ARGS);
94 static void WalSndShutdownHandler(SIGNAL_ARGS);
95 static void WalSndQuickDieHandler(SIGNAL_ARGS);
97 /* Prototypes for private functions */
98 static int WalSndLoop(void);
99 static void InitWalSnd(void);
100 static void WalSndHandshake(void);
101 static void WalSndKill(int code, Datum arg);
102 static void XLogRead(char *buf, XLogRecPtr recptr, Size nbytes);
103 static bool XLogSend(char *msgbuf, bool *caughtup);
104 static void CheckClosedConnection(void);
107 /* Main entry point for walsender process */
111 MemoryContext walsnd_context;
113 if (RecoveryInProgress())
115 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
116 errmsg("recovery is still in progress, can't accept WAL streaming connections")));
118 /* Create a per-walsender data structure in shared memory */
122 * Create a memory context that we will do all our work in. We do this so
123 * that we can reset the context during error recovery and thereby avoid
124 * possible memory leaks. Formerly this code just ran in
125 * TopMemoryContext, but resetting that would be a really bad idea.
127 * XXX: we don't actually attempt error recovery in walsender, we just
128 * close the connection and exit.
130 walsnd_context = AllocSetContextCreate(TopMemoryContext,
132 ALLOCSET_DEFAULT_MINSIZE,
133 ALLOCSET_DEFAULT_INITSIZE,
134 ALLOCSET_DEFAULT_MAXSIZE);
135 MemoryContextSwitchTo(walsnd_context);
137 /* Unblock signals (they were blocked when the postmaster forked us) */
138 PG_SETMASK(&UnBlockSig);
140 /* Tell the standby that walsender is ready for receiving commands */
141 ReadyForQuery(DestRemote);
143 /* Handle handshake messages before streaming */
146 /* Main loop of walsender */
151 * Execute commands from walreceiver, until we enter streaming mode.
154 WalSndHandshake(void)
156 StringInfoData input_message;
157 bool replication_started = false;
159 initStringInfo(&input_message);
161 while (!replication_started)
165 /* Wait for a command to arrive */
166 firstchar = pq_getbyte();
169 * Emergency bailout if postmaster has died. This is to avoid the
170 * necessity for manual cleanup of all postmaster children.
172 if (!PostmasterIsAlive(true))
176 * Check for any other interesting events that happened while we
182 ProcessConfigFile(PGC_SIGHUP);
185 if (firstchar != EOF)
188 * Read the message contents. This is expected to be done without
189 * blocking because we've been able to get message type code.
191 if (pq_getmessage(&input_message, 0))
192 firstchar = EOF; /* suitable message already logged */
195 /* Handle the very limited subset of commands expected in this phase */
198 case 'Q': /* Query message */
200 const char *query_string;
203 query_string = pq_getmsgstring(&input_message);
204 pq_getmsgend(&input_message);
206 if (strcmp(query_string, "IDENTIFY_SYSTEM") == 0)
213 * Reply with a result set with one row, two columns.
214 * First col is system ID, and second is timeline ID
217 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
218 GetSystemIdentifier());
219 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
221 /* Send a RowDescription message */
222 pq_beginmessage(&buf, 'T');
223 pq_sendint(&buf, 2, 2); /* 2 fields */
226 pq_sendstring(&buf, "systemid"); /* col name */
227 pq_sendint(&buf, 0, 4); /* table oid */
228 pq_sendint(&buf, 0, 2); /* attnum */
229 pq_sendint(&buf, TEXTOID, 4); /* type oid */
230 pq_sendint(&buf, -1, 2); /* typlen */
231 pq_sendint(&buf, 0, 4); /* typmod */
232 pq_sendint(&buf, 0, 2); /* format code */
235 pq_sendstring(&buf, "timeline"); /* col name */
236 pq_sendint(&buf, 0, 4); /* table oid */
237 pq_sendint(&buf, 0, 2); /* attnum */
238 pq_sendint(&buf, INT4OID, 4); /* type oid */
239 pq_sendint(&buf, 4, 2); /* typlen */
240 pq_sendint(&buf, 0, 4); /* typmod */
241 pq_sendint(&buf, 0, 2); /* format code */
244 /* Send a DataRow message */
245 pq_beginmessage(&buf, 'D');
246 pq_sendint(&buf, 2, 2); /* # of columns */
247 pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
248 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
249 pq_sendint(&buf, strlen(tli), 4); /* col2 len */
250 pq_sendbytes(&buf, (char *) tli, strlen(tli));
253 /* Send CommandComplete and ReadyForQuery messages */
254 EndCommand("SELECT", DestRemote);
255 ReadyForQuery(DestRemote);
256 /* ReadyForQuery did pq_flush for us */
258 else if (sscanf(query_string, "START_REPLICATION %X/%X",
259 &recptr.xlogid, &recptr.xrecoff) == 2)
264 * Check that we're logging enough information in the
265 * WAL for log-shipping.
267 * NOTE: This only checks the current value of
268 * wal_level. Even if the current setting is not
269 * 'minimal', there can be old WAL in the pg_xlog
270 * directory that was created with 'minimal'.
271 * So this is not bulletproof, the purpose is
272 * just to give a user-friendly error message that
273 * hints how to configure the system correctly.
275 if (wal_level == WAL_LEVEL_MINIMAL)
277 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
278 errmsg("standby connections not allowed because wal_level=minimal")));
280 /* Send a CopyOutResponse message, and start streaming */
281 pq_beginmessage(&buf, 'H');
282 pq_sendbyte(&buf, 0);
283 pq_sendint(&buf, 0, 2);
288 * Initialize position to the received one, then the
289 * xlog records begin to be shipped from that position
293 /* break out of the loop */
294 replication_started = true;
299 (errcode(ERRCODE_PROTOCOL_VIOLATION),
300 errmsg("invalid standby query string: %s", query_string)));
306 /* standby is closing the connection */
310 /* standby disconnected unexpectedly */
312 (errcode(ERRCODE_PROTOCOL_VIOLATION),
313 errmsg("unexpected EOF on standby connection")));
318 (errcode(ERRCODE_PROTOCOL_VIOLATION),
319 errmsg("invalid standby handshake message type %d", firstchar)));
325 * Check if the remote end has closed the connection.
328 CheckClosedConnection(void)
330 unsigned char firstchar;
333 r = pq_getbyte_if_available(&firstchar);
336 /* unexpected error or EOF */
338 (errcode(ERRCODE_PROTOCOL_VIOLATION),
339 errmsg("unexpected EOF on standby connection")));
344 /* no data available without blocking */
348 /* Handle the very limited subset of commands expected in this phase */
352 * 'X' means that the standby is closing down the socket.
359 (errcode(ERRCODE_PROTOCOL_VIOLATION),
360 errmsg("invalid standby closing message type %d",
365 /* Main loop of walsender process */
369 char *output_message;
370 bool caughtup = false;
373 * Allocate buffer that will be used for each output message. We do this
374 * just once to reduce palloc overhead. The buffer must be made large
375 * enough for maximum-sized messages.
377 output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
379 /* Loop forever, unless we get an error */
382 long remain; /* remaining time (us) */
385 * Emergency bailout if postmaster has died. This is to avoid the
386 * necessity for manual cleanup of all postmaster children.
388 if (!PostmasterIsAlive(true))
391 /* Process any requests or signals received recently */
395 ProcessConfigFile(PGC_SIGHUP);
399 * When SIGUSR2 arrives, we send all outstanding logs up to the
400 * shutdown checkpoint record (i.e., the latest record) and exit.
404 if (!XLogSend(output_message, &caughtup))
407 shutdown_requested = true;
410 /* Normal exit from the walsender is here */
411 if (shutdown_requested)
413 /* Inform the standby that XLOG streaming was done */
414 pq_puttextmessage('C', "COPY 0");
421 * If we had sent all accumulated WAL in last round, nap for the
422 * configured time before retrying.
424 * On some platforms, signals won't interrupt the sleep. To ensure we
425 * respond reasonably promptly when someone signals us, break down the
426 * sleep into NAPTIME_PER_CYCLE increments, and check for
427 * interrupts after each nap.
431 remain = WalSndDelay * 1000L;
434 /* Check for interrupts */
435 if (got_SIGHUP || shutdown_requested || ready_to_stop)
438 /* Sleep and check that the connection is still alive */
439 pg_usleep(remain > NAPTIME_PER_CYCLE ? NAPTIME_PER_CYCLE : remain);
440 CheckClosedConnection();
442 remain -= NAPTIME_PER_CYCLE;
446 /* Attempt to send the log once every loop */
447 if (!XLogSend(output_message, &caughtup))
452 * Get here on send failure. Clean up and exit.
454 * Reset whereToSendOutput to prevent ereport from attempting to send any
455 * more messages to the standby.
457 if (whereToSendOutput == DestRemote)
458 whereToSendOutput = DestNone;
461 return 1; /* keep the compiler quiet */
464 /* Initialize a per-walsender data structure for this walsender process */
471 * WalSndCtl should be set up already (we inherit this by fork() or
472 * EXEC_BACKEND mechanism from the postmaster).
474 Assert(WalSndCtl != NULL);
475 Assert(MyWalSnd == NULL);
478 * Find a free walsender slot and reserve it. If this fails, we must be
479 * out of WalSnd structures.
481 for (i = 0; i < max_wal_senders; i++)
483 /* use volatile pointer to prevent code rearrangement */
484 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
486 SpinLockAcquire(&walsnd->mutex);
488 if (walsnd->pid != 0)
490 SpinLockRelease(&walsnd->mutex);
496 MyWalSnd = (WalSnd *) walsnd;
497 walsnd->pid = MyProcPid;
498 MemSet(&MyWalSnd->sentPtr, 0, sizeof(XLogRecPtr));
499 SpinLockRelease(&walsnd->mutex);
503 if (MyWalSnd == NULL)
505 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
506 errmsg("number of requested standby connections "
507 "exceeds max_wal_senders (currently %d)",
510 /* Arrange to clean up at walsender exit */
511 on_shmem_exit(WalSndKill, 0);
514 /* Destroy the per-walsender data structure for this walsender process */
516 WalSndKill(int code, Datum arg)
518 Assert(MyWalSnd != NULL);
521 * Mark WalSnd struct no longer in use. Assume that no lock is required
526 /* WalSnd struct isn't mine anymore */
531 * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
533 * XXX probably this should be improved to suck data directly from the
534 * WAL buffers when possible.
537 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
539 XLogRecPtr startRecPtr = recptr;
540 char path[MAXPGPATH];
541 uint32 lastRemovedLog;
542 uint32 lastRemovedSeg;
552 startoff = recptr.xrecoff % XLogSegSize;
554 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
556 /* Switch to another logfile segment */
560 XLByteToSeg(recptr, sendId, sendSeg);
561 XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
563 sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
567 * If the file is not found, assume it's because the
568 * standby asked for a too old WAL segment that has already
569 * been removed or recycled.
573 char filename[MAXFNAMELEN];
574 XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
576 (errcode_for_file_access(),
577 errmsg("requested WAL segment %s has already been removed",
582 (errcode_for_file_access(),
583 errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
584 path, sendId, sendSeg)));
589 /* Need to seek in the file? */
590 if (sendOff != startoff)
592 if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
594 (errcode_for_file_access(),
595 errmsg("could not seek in log file %u, segment %u to offset %u: %m",
596 sendId, sendSeg, startoff)));
600 /* How many bytes are within this segment? */
601 if (nbytes > (XLogSegSize - startoff))
602 segbytes = XLogSegSize - startoff;
606 readbytes = read(sendFile, buf, segbytes);
609 (errcode_for_file_access(),
610 errmsg("could not read from log file %u, segment %u, offset %u, "
612 sendId, sendSeg, sendOff, (unsigned long) segbytes)));
614 /* Update state for read */
615 XLByteAdvance(recptr, readbytes);
617 sendOff += readbytes;
623 * After reading into the buffer, check that what we read was valid.
624 * We do this after reading, because even though the segment was present
625 * when we opened it, it might get recycled or removed while we read it.
626 * The read() succeeds in that case, but the data we tried to read might
627 * already have been overwritten with new WAL records.
629 XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
630 XLByteToSeg(startRecPtr, log, seg);
631 if (log < lastRemovedLog ||
632 (log == lastRemovedLog && seg <= lastRemovedSeg))
634 char filename[MAXFNAMELEN];
635 XLogFileName(filename, ThisTimeLineID, log, seg);
637 (errcode_for_file_access(),
638 errmsg("requested WAL segment %s has already been removed",
644 * Read up to MAX_SEND_SIZE bytes of WAL that's been written (and flushed),
645 * but not yet sent to the client, and send it.
647 * msgbuf is a work area in which the output message is constructed. It's
648 * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
649 * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
651 * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
652 * *caughtup is set to false.
654 * Returns true if OK, false if trouble.
657 XLogSend(char *msgbuf, bool *caughtup)
659 XLogRecPtr SendRqstPtr;
663 WalDataMessageHeader msghdr;
665 /* Attempt to send all records flushed to the disk already */
666 SendRqstPtr = GetWriteRecPtr();
668 /* Quick exit if nothing to do */
669 if (XLByteLE(SendRqstPtr, sentPtr))
676 * Figure out how much to send in one message. If there's no more than
677 * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
678 * MAX_SEND_SIZE bytes, but round to logfile or page boundary.
680 * The rounding is not only for performance reasons. Walreceiver
681 * relies on the fact that we never split a WAL record across two
682 * messages. Since a long WAL record is split at page boundary into
683 * continuation records, page boundary is always a safe cut-off point.
684 * We also assume that SendRqstPtr never points to the middle of a WAL
688 if (startptr.xrecoff >= XLogFileSize)
691 * crossing a logid boundary, skip the non-existent last log
692 * segment in previous logical log file.
694 startptr.xlogid += 1;
695 startptr.xrecoff = 0;
699 XLByteAdvance(endptr, MAX_SEND_SIZE);
700 if (endptr.xlogid != startptr.xlogid)
702 /* Don't cross a logfile boundary within one message */
703 Assert(endptr.xlogid == startptr.xlogid + 1);
704 endptr.xlogid = startptr.xlogid;
705 endptr.xrecoff = XLogFileSize;
708 /* if we went beyond SendRqstPtr, back off */
709 if (XLByteLE(SendRqstPtr, endptr))
711 endptr = SendRqstPtr;
716 /* round down to page boundary. */
717 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
721 nbytes = endptr.xrecoff - startptr.xrecoff;
722 Assert(nbytes <= MAX_SEND_SIZE);
725 * OK to read and send the slice.
730 * Read the log directly into the output buffer to avoid extra memcpy
733 XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
736 * We fill the message header last so that the send timestamp is taken
737 * as late as possible.
739 msghdr.dataStart = startptr;
740 msghdr.walEnd = SendRqstPtr;
741 msghdr.sendTime = GetCurrentTimestamp();
743 memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
745 pq_putmessage('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
747 /* Flush pending output */
753 /* Update shared memory status */
755 /* use volatile pointer to prevent code rearrangement */
756 volatile WalSnd *walsnd = MyWalSnd;
758 SpinLockAcquire(&walsnd->mutex);
759 walsnd->sentPtr = sentPtr;
760 SpinLockRelease(&walsnd->mutex);
763 /* Report progress of XLOG streaming in PS display */
764 if (update_process_title)
766 char activitymsg[50];
768 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
769 sentPtr.xlogid, sentPtr.xrecoff);
770 set_ps_display(activitymsg, false);
776 /* SIGHUP: set flag to re-read config file at next convenient time */
778 WalSndSigHupHandler(SIGNAL_ARGS)
783 /* SIGTERM: set flag to shut down */
785 WalSndShutdownHandler(SIGNAL_ARGS)
787 shutdown_requested = true;
791 * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
793 * Some backend has bought the farm,
794 * so we need to stop what we're doing and exit.
797 WalSndQuickDieHandler(SIGNAL_ARGS)
799 PG_SETMASK(&BlockSig);
802 * We DO NOT want to run proc_exit() callbacks -- we're here because
803 * shared memory may be corrupted, so we don't want to try to clean up our
804 * transaction. Just nail the windows shut and get out of town. Now that
805 * there's an atexit callback to prevent third-party code from breaking
806 * things by calling exit() directly, we have to reset the callbacks
807 * explicitly to make this work as intended.
812 * Note we do exit(2) not exit(0). This is to force the postmaster into a
813 * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
814 * backend. This is necessary precisely because we don't clean up our
815 * shared memory state. (The "dead man switch" mechanism in pmsignal.c
816 * should ensure the postmaster sees this as a crash, too, but no harm in
817 * being doubly sure.)
822 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
824 WalSndLastCycleHandler(SIGNAL_ARGS)
826 ready_to_stop = true;
829 /* Set up signal handlers */
833 /* Set up signal handlers */
834 pqsignal(SIGHUP, WalSndSigHupHandler); /* set flag to read config
836 pqsignal(SIGINT, SIG_IGN); /* not used */
837 pqsignal(SIGTERM, WalSndShutdownHandler); /* request shutdown */
838 pqsignal(SIGQUIT, WalSndQuickDieHandler); /* hard crash time */
839 pqsignal(SIGALRM, SIG_IGN);
840 pqsignal(SIGPIPE, SIG_IGN);
841 pqsignal(SIGUSR1, SIG_IGN); /* not used */
842 pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
845 /* Reset some signals that are accepted by postmaster but not here */
846 pqsignal(SIGCHLD, SIG_DFL);
847 pqsignal(SIGTTIN, SIG_DFL);
848 pqsignal(SIGTTOU, SIG_DFL);
849 pqsignal(SIGCONT, SIG_DFL);
850 pqsignal(SIGWINCH, SIG_DFL);
853 /* Report shared-memory space needed by WalSndShmemInit */
855 WalSndShmemSize(void)
859 size = offsetof(WalSndCtlData, walsnds);
860 size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
865 /* Allocate and initialize walsender-related shared memory */
867 WalSndShmemInit(void)
872 WalSndCtl = (WalSndCtlData *)
873 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
877 /* First time through, so initialize */
878 MemSet(WalSndCtl, 0, WalSndShmemSize());
880 for (i = 0; i < max_wal_senders; i++)
882 WalSnd *walsnd = &WalSndCtl->walsnds[i];
884 SpinLockInit(&walsnd->mutex);
890 * This isn't currently used for anything. Monitoring tools might be
891 * interested in the future, and we'll need something like this in the
892 * future for synchronous replication.
896 * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
900 GetOldestWALSendPointer(void)
902 XLogRecPtr oldest = {0, 0};
906 for (i = 0; i < max_wal_senders; i++)
908 /* use volatile pointer to prevent code rearrangement */
909 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
912 if (walsnd->pid == 0)
915 SpinLockAcquire(&walsnd->mutex);
916 recptr = walsnd->sentPtr;
917 SpinLockRelease(&walsnd->mutex);
919 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
922 if (!found || XLByteLT(recptr, oldest))