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 * i.e., unlike any auxiliary process, it is not an always-running process.
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 an 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.
27 * Note that there can be more than one walsender process concurrently.
29 * Portions Copyright (c) 2010-2010, PostgreSQL Global Development Group
33 * $PostgreSQL: pgsql/src/backend/replication/walsender.c,v 1.13 2010/03/26 12:23:34 heikki Exp $
35 *-------------------------------------------------------------------------
41 #include "access/xlog_internal.h"
42 #include "catalog/pg_type.h"
43 #include "libpq/libpq.h"
44 #include "libpq/pqformat.h"
45 #include "libpq/pqsignal.h"
46 #include "miscadmin.h"
47 #include "replication/walsender.h"
48 #include "storage/fd.h"
49 #include "storage/ipc.h"
50 #include "storage/lock.h"
51 #include "storage/pmsignal.h"
52 #include "tcop/tcopprot.h"
53 #include "utils/guc.h"
54 #include "utils/memutils.h"
55 #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 MaxWalSenders = 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
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(StringInfo outMsg);
104 static void CheckClosedConnection(void);
107 * How much WAL to send in one message? Must be >= XLOG_BLCKSZ.
109 #define MAX_SEND_SIZE (XLOG_SEG_SIZE / 2)
111 /* Main entry point for walsender process */
115 MemoryContext walsnd_context;
119 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
120 errmsg("must be superuser to start walsender")));
122 if (RecoveryInProgress())
124 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
125 errmsg("recovery is still in progress, can't accept WAL streaming connections")));
127 /* Create a per-walsender data structure in shared memory */
131 * Create a memory context that we will do all our work in. We do this so
132 * that we can reset the context during error recovery and thereby avoid
133 * possible memory leaks. Formerly this code just ran in
134 * TopMemoryContext, but resetting that would be a really bad idea.
136 * XXX: we don't actually attempt error recovery in walsender, we just
137 * close the connection and exit.
139 walsnd_context = AllocSetContextCreate(TopMemoryContext,
141 ALLOCSET_DEFAULT_MINSIZE,
142 ALLOCSET_DEFAULT_INITSIZE,
143 ALLOCSET_DEFAULT_MAXSIZE);
144 MemoryContextSwitchTo(walsnd_context);
146 /* Unblock signals (they were blocked when the postmaster forked us) */
147 PG_SETMASK(&UnBlockSig);
149 /* Tell the standby that walsender is ready for receiving commands */
150 ReadyForQuery(DestRemote);
152 /* Handle handshake messages before streaming */
155 /* Main loop of walsender */
160 WalSndHandshake(void)
162 StringInfoData input_message;
163 bool replication_started = false;
165 initStringInfo(&input_message);
167 while (!replication_started)
171 /* Wait for a command to arrive */
172 firstchar = pq_getbyte();
175 * Check for any other interesting events that happened while we
181 ProcessConfigFile(PGC_SIGHUP);
184 if (firstchar != EOF)
187 * Read the message contents. This is expected to be done without
188 * blocking because we've been able to get message type code.
190 if (pq_getmessage(&input_message, 0))
191 firstchar = EOF; /* suitable message already logged */
194 /* Handle the very limited subset of commands expected in this phase */
197 case 'Q': /* Query message */
199 const char *query_string;
202 query_string = pq_getmsgstring(&input_message);
203 pq_getmsgend(&input_message);
205 if (strcmp(query_string, "IDENTIFY_SYSTEM") == 0)
212 * Reply with a result set with one row, two columns.
213 * First col is system ID, and second if timeline ID
216 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
217 GetSystemIdentifier());
218 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
220 /* Send a RowDescription message */
221 pq_beginmessage(&buf, 'T');
222 pq_sendint(&buf, 2, 2); /* 2 fields */
225 pq_sendstring(&buf, "systemid"); /* col name */
226 pq_sendint(&buf, 0, 4); /* table oid */
227 pq_sendint(&buf, 0, 2); /* attnum */
228 pq_sendint(&buf, TEXTOID, 4); /* type oid */
229 pq_sendint(&buf, -1, 2); /* typlen */
230 pq_sendint(&buf, 0, 4); /* typmod */
231 pq_sendint(&buf, 0, 2); /* format code */
234 pq_sendstring(&buf, "timeline"); /* col name */
235 pq_sendint(&buf, 0, 4); /* table oid */
236 pq_sendint(&buf, 0, 2); /* attnum */
237 pq_sendint(&buf, INT4OID, 4); /* type oid */
238 pq_sendint(&buf, 4, 2); /* typlen */
239 pq_sendint(&buf, 0, 4); /* typmod */
240 pq_sendint(&buf, 0, 2); /* format code */
243 /* Send a DataRow message */
244 pq_beginmessage(&buf, 'D');
245 pq_sendint(&buf, 2, 2); /* # of columns */
246 pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
247 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
248 pq_sendint(&buf, strlen(tli), 4); /* col2 len */
249 pq_sendbytes(&buf, (char *) tli, strlen(tli));
252 /* Send CommandComplete and ReadyForQuery messages */
253 EndCommand("SELECT", DestRemote);
254 ReadyForQuery(DestRemote);
256 else if (sscanf(query_string, "START_REPLICATION %X/%X",
257 &recptr.xlogid, &recptr.xrecoff) == 2)
261 /* Send a CopyOutResponse message, and start streaming */
262 pq_beginmessage(&buf, 'H');
263 pq_sendbyte(&buf, 0);
264 pq_sendint(&buf, 0, 2);
269 * Initialize position to the received one, then the
270 * xlog records begin to be shipped from that position
274 /* break out of the loop */
275 replication_started = true;
280 (errcode(ERRCODE_PROTOCOL_VIOLATION),
281 errmsg("invalid standby query string: %s", query_string)));
287 /* standby is closing the connection */
291 /* standby disconnected unexpectedly */
293 (errcode(ERRCODE_PROTOCOL_VIOLATION),
294 errmsg("unexpected EOF on standby connection")));
299 (errcode(ERRCODE_PROTOCOL_VIOLATION),
300 errmsg("invalid standby handshake message type %d", firstchar)));
306 * Check if the remote end has closed the connection.
309 CheckClosedConnection(void)
311 unsigned char firstchar;
314 r = pq_getbyte_if_available(&firstchar);
317 /* unexpected error or EOF */
319 (errcode(ERRCODE_PROTOCOL_VIOLATION),
320 errmsg("unexpected EOF on standby connection")));
325 /* no data available without blocking */
329 /* Handle the very limited subset of commands expected in this phase */
333 * 'X' means that the standby is closing down the socket.
340 (errcode(ERRCODE_PROTOCOL_VIOLATION),
341 errmsg("invalid standby closing message type %d",
346 /* Main loop of walsender process */
350 StringInfoData output_message;
352 initStringInfo(&output_message);
357 long remain; /* remaining time (us) */
360 * Emergency bailout if postmaster has died. This is to avoid the
361 * necessity for manual cleanup of all postmaster children.
363 if (!PostmasterIsAlive(true))
365 /* Process any requests or signals received recently */
369 ProcessConfigFile(PGC_SIGHUP);
373 * When SIGUSR2 arrives, we send all outstanding logs up to the
374 * shutdown checkpoint record (i.e., the latest record) and exit.
378 XLogSend(&output_message);
379 shutdown_requested = true;
382 /* Normal exit from the walsender is here */
383 if (shutdown_requested)
385 /* Inform the standby that XLOG streaming was done */
386 pq_puttextmessage('C', "COPY 0");
393 * Nap for the configured time or until a message arrives.
395 * On some platforms, signals won't interrupt the sleep. To ensure we
396 * respond reasonably promptly when someone signals us, break down the
397 * sleep into NAPTIME_PER_CYCLE increments, and check for
398 * interrupts after each nap.
400 remain = WalSndDelay * 1000L;
403 if (got_SIGHUP || shutdown_requested || ready_to_stop)
407 * Check to see whether a message from the standby or an interrupt
408 * from other processes has arrived.
410 pg_usleep(remain > NAPTIME_PER_CYCLE ? NAPTIME_PER_CYCLE : remain);
411 CheckClosedConnection();
413 remain -= NAPTIME_PER_CYCLE;
416 /* Attempt to send the log once every loop */
417 if (!XLogSend(&output_message))
421 /* can't get here because the above loop never exits */
427 * Reset whereToSendOutput to prevent ereport from attempting to send any
428 * more messages to the standby.
430 if (whereToSendOutput == DestRemote)
431 whereToSendOutput = DestNone;
434 return 1; /* keep the compiler quiet */
437 /* Initialize a per-walsender data structure for this walsender process */
441 /* use volatile pointer to prevent code rearrangement */
445 * WalSndCtl should be set up already (we inherit this by fork() or
446 * EXEC_BACKEND mechanism from the postmaster).
448 Assert(WalSndCtl != NULL);
449 Assert(MyWalSnd == NULL);
452 * Find a free walsender slot and reserve it. If this fails, we must be
453 * out of WalSnd structures.
455 for (i = 0; i < MaxWalSenders; i++)
457 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
459 SpinLockAcquire(&walsnd->mutex);
461 if (walsnd->pid != 0)
463 SpinLockRelease(&walsnd->mutex);
469 MyWalSnd = (WalSnd *) walsnd;
470 walsnd->pid = MyProcPid;
471 MemSet(&MyWalSnd->sentPtr, 0, sizeof(XLogRecPtr));
472 SpinLockRelease(&walsnd->mutex);
476 if (MyWalSnd == NULL)
478 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
479 errmsg("sorry, too many standbys already")));
481 /* Arrange to clean up at walsender exit */
482 on_shmem_exit(WalSndKill, 0);
485 /* Destroy the per-walsender data structure for this walsender process */
487 WalSndKill(int code, Datum arg)
489 Assert(MyWalSnd != NULL);
492 * Mark WalSnd struct no longer in use. Assume that no lock is required
497 /* WalSnd struct isn't mine anymore */
502 * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
505 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
507 char path[MAXPGPATH];
515 startoff = recptr.xrecoff % XLogSegSize;
517 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
519 /* Switch to another logfile segment */
523 XLByteToSeg(recptr, sendId, sendSeg);
524 XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
526 sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
528 ereport(FATAL, /* XXX: Why FATAL? */
529 (errcode_for_file_access(),
530 errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
531 path, sendId, sendSeg)));
535 /* Need to seek in the file? */
536 if (sendOff != startoff)
538 if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
540 (errcode_for_file_access(),
541 errmsg("could not seek in log file %u, segment %u to offset %u: %m",
542 sendId, sendSeg, startoff)));
546 /* How many bytes are within this segment? */
547 if (nbytes > (XLogSegSize - startoff))
548 segbytes = XLogSegSize - startoff;
552 readbytes = read(sendFile, buf, segbytes);
555 (errcode_for_file_access(),
556 errmsg("could not read from log file %u, segment %u, offset %u, "
558 sendId, sendSeg, sendOff, (unsigned long) segbytes)));
560 /* Update state for read */
561 XLByteAdvance(recptr, readbytes);
563 sendOff += readbytes;
570 * Read all WAL that's been written (and flushed) since last cycle, and send
573 * Returns true if OK, false if trouble.
576 XLogSend(StringInfo outMsg)
578 XLogRecPtr SendRqstPtr;
579 char activitymsg[50];
581 /* use volatile pointer to prevent code rearrangement */
582 volatile WalSnd *walsnd = MyWalSnd;
584 /* Attempt to send all records flushed to the disk already */
585 SendRqstPtr = GetWriteRecPtr();
587 /* Quick exit if nothing to do */
588 if (!XLByteLT(sentPtr, SendRqstPtr))
592 * We gather multiple records together by issuing just one XLogRead() of a
593 * suitable size, and send them as one CopyData message. Repeat until
594 * we've sent everything we can.
596 while (XLByteLT(sentPtr, SendRqstPtr))
603 * Figure out how much to send in one message. If there's less than
604 * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
605 * MAX_SEND_SIZE bytes, but round to page boundary.
607 * The rounding is not only for performance reasons. Walreceiver
608 * relies on the fact that we never split a WAL record across two
609 * messages. Since a long WAL record is split at page boundary into
610 * continuation records, page boundary is always a safe cut-off point.
611 * We also assume that SendRqstPtr never points in the middle of a WAL
615 if (startptr.xrecoff >= XLogFileSize)
618 * crossing a logid boundary, skip the non-existent last log
619 * segment in previous logical log file.
621 startptr.xlogid += 1;
622 startptr.xrecoff = 0;
626 XLByteAdvance(endptr, MAX_SEND_SIZE);
627 /* round down to page boundary. */
628 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
629 /* if we went beyond SendRqstPtr, back off */
630 if (XLByteLT(SendRqstPtr, endptr))
631 endptr = SendRqstPtr;
634 * OK to read and send the slice.
636 * We don't need to convert the xlogid/xrecoff from host byte order to
637 * network byte order because the both server can be expected to have
638 * the same byte order. If they have different byte order, we don't
641 pq_sendbyte(outMsg, 'w');
642 pq_sendbytes(outMsg, (char *) &startptr, sizeof(startptr));
644 if (endptr.xlogid != startptr.xlogid)
646 Assert(endptr.xlogid == startptr.xlogid + 1);
647 nbytes = endptr.xrecoff + XLogFileSize - startptr.xrecoff;
650 nbytes = endptr.xrecoff - startptr.xrecoff;
655 * Read the log directly into the output buffer to prevent extra
658 enlargeStringInfo(outMsg, nbytes);
660 XLogRead(&outMsg->data[outMsg->len], startptr, nbytes);
661 outMsg->len += nbytes;
662 outMsg->data[outMsg->len] = '\0';
664 pq_putmessage('d', outMsg->data, outMsg->len);
665 resetStringInfo(outMsg);
668 /* Update shared memory status */
669 SpinLockAcquire(&walsnd->mutex);
670 walsnd->sentPtr = sentPtr;
671 SpinLockRelease(&walsnd->mutex);
673 /* Flush pending output */
677 /* Report progress of XLOG streaming in PS display */
678 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
679 sentPtr.xlogid, sentPtr.xrecoff);
680 set_ps_display(activitymsg, false);
685 /* SIGHUP: set flag to re-read config file at next convenient time */
687 WalSndSigHupHandler(SIGNAL_ARGS)
692 /* SIGTERM: set flag to shut down */
694 WalSndShutdownHandler(SIGNAL_ARGS)
696 shutdown_requested = true;
700 * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
702 * Some backend has bought the farm,
703 * so we need to stop what we're doing and exit.
706 WalSndQuickDieHandler(SIGNAL_ARGS)
708 PG_SETMASK(&BlockSig);
711 * We DO NOT want to run proc_exit() callbacks -- we're here because
712 * shared memory may be corrupted, so we don't want to try to clean up our
713 * transaction. Just nail the windows shut and get out of town. Now that
714 * there's an atexit callback to prevent third-party code from breaking
715 * things by calling exit() directly, we have to reset the callbacks
716 * explicitly to make this work as intended.
721 * Note we do exit(2) not exit(0). This is to force the postmaster into a
722 * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
723 * backend. This is necessary precisely because we don't clean up our
724 * shared memory state. (The "dead man switch" mechanism in pmsignal.c
725 * should ensure the postmaster sees this as a crash, too, but no harm in
726 * being doubly sure.)
731 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
733 WalSndLastCycleHandler(SIGNAL_ARGS)
735 ready_to_stop = true;
738 /* Set up signal handlers */
742 /* Set up signal handlers */
743 pqsignal(SIGHUP, WalSndSigHupHandler); /* set flag to read config
745 pqsignal(SIGINT, SIG_IGN); /* not used */
746 pqsignal(SIGTERM, WalSndShutdownHandler); /* request shutdown */
747 pqsignal(SIGQUIT, WalSndQuickDieHandler); /* hard crash time */
748 pqsignal(SIGALRM, SIG_IGN);
749 pqsignal(SIGPIPE, SIG_IGN);
750 pqsignal(SIGUSR1, SIG_IGN); /* not used */
751 pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
754 /* Reset some signals that are accepted by postmaster but not here */
755 pqsignal(SIGCHLD, SIG_DFL);
756 pqsignal(SIGTTIN, SIG_DFL);
757 pqsignal(SIGTTOU, SIG_DFL);
758 pqsignal(SIGCONT, SIG_DFL);
759 pqsignal(SIGWINCH, SIG_DFL);
762 /* Report shared-memory space needed by WalSndShmemInit */
764 WalSndShmemSize(void)
768 size = offsetof(WalSndCtlData, walsnds);
769 size = add_size(size, mul_size(MaxWalSenders, sizeof(WalSnd)));
774 /* Allocate and initialize walsender-related shared memory */
776 WalSndShmemInit(void)
781 WalSndCtl = (WalSndCtlData *)
782 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
784 if (WalSndCtl == NULL)
786 (errcode(ERRCODE_OUT_OF_MEMORY),
787 errmsg("not enough shared memory for walsender")));
789 return; /* already initialized */
791 /* Initialize the data structures */
792 MemSet(WalSndCtl, 0, WalSndShmemSize());
794 for (i = 0; i < MaxWalSenders; i++)
796 WalSnd *walsnd = &WalSndCtl->walsnds[i];
798 SpinLockInit(&walsnd->mutex);
803 * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
807 GetOldestWALSendPointer(void)
809 XLogRecPtr oldest = {0, 0};
813 for (i = 0; i < MaxWalSenders; i++)
815 /* use volatile pointer to prevent code rearrangement */
816 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
819 if (walsnd->pid == 0)
822 SpinLockAcquire(&walsnd->mutex);
823 recptr = walsnd->sentPtr;
824 SpinLockRelease(&walsnd->mutex);
826 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
829 if (!found || XLByteLT(recptr, oldest))