1 /*-------------------------------------------------------------------------
5 * The WAL sender process (walsender) is new as of Postgres 8.5. 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 1.4 2010/01/27 16:41:09 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 100 /* 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 /* Create a per-walsender data structure in shared memory */
126 * Create a memory context that we will do all our work in. We do this so
127 * that we can reset the context during error recovery and thereby avoid
128 * possible memory leaks. Formerly this code just ran in
129 * TopMemoryContext, but resetting that would be a really bad idea.
131 * XXX: we don't actually attempt error recovery in walsender, we just
132 * close the connection and exit.
134 walsnd_context = AllocSetContextCreate(TopMemoryContext,
136 ALLOCSET_DEFAULT_MINSIZE,
137 ALLOCSET_DEFAULT_INITSIZE,
138 ALLOCSET_DEFAULT_MAXSIZE);
139 MemoryContextSwitchTo(walsnd_context);
141 /* Unblock signals (they were blocked when the postmaster forked us) */
142 PG_SETMASK(&UnBlockSig);
144 /* Tell the standby that walsender is ready for receiving commands */
145 ReadyForQuery(DestRemote);
147 /* Handle handshake messages before streaming */
150 /* Main loop of walsender */
155 WalSndHandshake(void)
157 StringInfoData input_message;
158 bool replication_started = false;
160 initStringInfo(&input_message);
162 while (!replication_started)
166 /* Wait for a command to arrive */
167 firstchar = pq_getbyte();
170 * Check for any other interesting events that happened while we
176 ProcessConfigFile(PGC_SIGHUP);
179 if (firstchar == EOF)
181 /* standby disconnected */
183 (errcode(ERRCODE_PROTOCOL_VIOLATION),
184 errmsg("unexpected EOF on standby connection")));
189 * Read the message contents. This is expected to be done without
190 * blocking because we've been able to get message type code.
192 if (pq_getmessage(&input_message, 0))
193 firstchar = EOF; /* suitable message already logged */
197 /* Handle the very limited subset of commands expected in this phase */
201 case 'Q': /* Query message */
203 const char *query_string;
206 query_string = pq_getmsgstring(&input_message);
207 pq_getmsgend(&input_message);
209 if (strcmp(query_string, "IDENTIFY_SYSTEM") == 0)
216 * Reply with a result set with one row, two columns.
217 * First col is system ID, and second if timeline ID
220 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
221 GetSystemIdentifier());
222 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
224 /* Send a RowDescription message */
225 pq_beginmessage(&buf, 'T');
226 pq_sendint(&buf, 2, 2); /* 2 fields */
229 pq_sendstring(&buf, "systemid"); /* col name */
230 pq_sendint(&buf, 0, 4); /* table oid */
231 pq_sendint(&buf, 0, 2); /* attnum */
232 pq_sendint(&buf, TEXTOID, 4); /* type oid */
233 pq_sendint(&buf, -1, 2); /* typlen */
234 pq_sendint(&buf, 0, 4); /* typmod */
235 pq_sendint(&buf, 0, 2); /* format code */
238 pq_sendstring(&buf, "timeline"); /* col name */
239 pq_sendint(&buf, 0, 4); /* table oid */
240 pq_sendint(&buf, 0, 2); /* attnum */
241 pq_sendint(&buf, INT4OID, 4); /* type oid */
242 pq_sendint(&buf, 4, 2); /* typlen */
243 pq_sendint(&buf, 0, 4); /* typmod */
244 pq_sendint(&buf, 0, 2); /* format code */
247 /* Send a DataRow message */
248 pq_beginmessage(&buf, 'D');
249 pq_sendint(&buf, 2, 2); /* # of columns */
250 pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
251 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
252 pq_sendint(&buf, strlen(tli), 4); /* col2 len */
253 pq_sendbytes(&buf, (char *) tli, strlen(tli));
256 /* Send CommandComplete and ReadyForQuery messages */
257 EndCommand("SELECT", DestRemote);
258 ReadyForQuery(DestRemote);
260 else if (sscanf(query_string, "START_REPLICATION %X/%X",
261 &recptr.xlogid, &recptr.xrecoff) == 2)
265 /* Send a CopyOutResponse message, and start streaming */
266 pq_beginmessage(&buf, 'H');
267 pq_sendbyte(&buf, 0);
268 pq_sendint(&buf, 0, 2);
272 * Initialize position to the received one, then
273 * the xlog records begin to be shipped from that position
277 /* break out of the loop */
278 replication_started = true;
283 (errcode(ERRCODE_PROTOCOL_VIOLATION),
284 errmsg("invalid standby query string: %s", query_string)));
289 /* 'X' means that the standby is closing the connection */
295 (errcode(ERRCODE_PROTOCOL_VIOLATION),
296 errmsg("unexpected EOF on standby connection")));
300 (errcode(ERRCODE_PROTOCOL_VIOLATION),
301 errmsg("invalid standby handshake message type %d", firstchar)));
307 * Check if the remote end has closed the connection.
310 CheckClosedConnection(void)
312 unsigned char firstchar;
315 r = pq_getbyte_if_available(&firstchar);
318 /* no data available */
319 if (errno == EAGAIN || errno == EWOULDBLOCK)
323 * Ok if interrupted, though it shouldn't really happen with
324 * a non-blocking operation.
330 (errcode_for_socket_access(),
331 errmsg("could not receive data from client: %m")));
335 /* standby disconnected unexpectedly */
337 (errcode(ERRCODE_PROTOCOL_VIOLATION),
338 errmsg("unexpected EOF on standby connection")));
341 /* Handle the very limited subset of commands expected in this phase */
345 * 'X' means that the standby is closing down the socket. EOF means
346 * unexpected loss of standby connection. Either way, perform normal
354 (errcode(ERRCODE_PROTOCOL_VIOLATION),
355 errmsg("invalid standby closing message type %d",
360 /* Main loop of walsender process */
364 StringInfoData output_message;
366 initStringInfo(&output_message);
371 int remain; /* remaining time (ms) */
374 * Emergency bailout if postmaster has died. This is to avoid the
375 * necessity for manual cleanup of all postmaster children.
377 if (!PostmasterIsAlive(true))
379 /* Process any requests or signals received recently */
383 ProcessConfigFile(PGC_SIGHUP);
387 * When SIGUSR2 arrives, we send all outstanding logs up to the
388 * shutdown checkpoint record (i.e., the latest record) and exit.
392 XLogSend(&output_message);
393 shutdown_requested = true;
396 /* Normal exit from the walsender is here */
397 if (shutdown_requested)
399 /* Inform the standby that XLOG streaming was done */
400 pq_puttextmessage('C', "COPY 0");
407 * Nap for the configured time or until a message arrives.
409 * On some platforms, signals won't interrupt the sleep. To ensure we
410 * respond reasonably promptly when someone signals us, break down the
411 * sleep into NAPTIME_PER_CYCLE (ms) increments, and check for
412 * interrupts after each nap.
414 remain = WalSndDelay;
417 if (got_SIGHUP || shutdown_requested || ready_to_stop)
421 * Check to see whether a message from the standby or an interrupt
422 * from other processes has arrived.
424 pg_usleep(remain > NAPTIME_PER_CYCLE ? NAPTIME_PER_CYCLE : remain);
425 CheckClosedConnection();
427 remain -= NAPTIME_PER_CYCLE;
430 /* Attempt to send the log once every loop */
431 if (!XLogSend(&output_message))
435 /* can't get here because the above loop never exits */
440 * Reset whereToSendOutput to prevent ereport from attempting
441 * to send any more messages to the standby.
443 if (whereToSendOutput == DestRemote)
444 whereToSendOutput = DestNone;
447 return 1; /* keep the compiler quiet */
450 /* Initialize a per-walsender data structure for this walsender process */
454 /* use volatile pointer to prevent code rearrangement */
458 * WalSndCtl should be set up already (we inherit this by fork() or
459 * EXEC_BACKEND mechanism from the postmaster).
461 Assert(WalSndCtl != NULL);
462 Assert(MyWalSnd == NULL);
465 * Find a free walsender slot and reserve it. If this fails, we must be
466 * out of WalSnd structures.
468 for (i = 0; i < MaxWalSenders; i++)
470 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
472 SpinLockAcquire(&walsnd->mutex);
474 if (walsnd->pid != 0)
476 SpinLockRelease(&walsnd->mutex);
482 MyWalSnd = (WalSnd *) walsnd;
483 walsnd->pid = MyProcPid;
484 MemSet(&MyWalSnd->sentPtr, 0, sizeof(XLogRecPtr));
485 SpinLockRelease(&walsnd->mutex);
489 if (MyWalSnd == NULL)
491 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
492 errmsg("sorry, too many standbys already")));
494 /* Arrange to clean up at walsender exit */
495 on_shmem_exit(WalSndKill, 0);
498 /* Destroy the per-walsender data structure for this walsender process */
500 WalSndKill(int code, Datum arg)
502 Assert(MyWalSnd != NULL);
505 * Mark WalSnd struct no longer in use. Assume that no lock is required
510 /* WalSnd struct isn't mine anymore */
515 * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
518 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
520 char path[MAXPGPATH];
528 startoff = recptr.xrecoff % XLogSegSize;
530 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
532 /* Switch to another logfile segment */
536 XLByteToSeg(recptr, sendId, sendSeg);
537 XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
539 sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
541 ereport(FATAL, /* XXX: Why FATAL? */
542 (errcode_for_file_access(),
543 errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
544 path, sendId, sendSeg)));
548 /* Need to seek in the file? */
549 if (sendOff != startoff)
551 if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
553 (errcode_for_file_access(),
554 errmsg("could not seek in log file %u, segment %u to offset %u: %m",
555 sendId, sendSeg, startoff)));
559 /* How many bytes are within this segment? */
560 if (nbytes > (XLogSegSize - startoff))
561 segbytes = XLogSegSize - startoff;
565 readbytes = read(sendFile, buf, segbytes);
568 (errcode_for_file_access(),
569 errmsg("could not read from log file %u, segment %u, offset %u, "
571 sendId, sendSeg, sendOff, (unsigned long) segbytes)));
573 /* Update state for read */
574 XLByteAdvance(recptr, readbytes);
576 sendOff += readbytes;
583 * Read all WAL that's been written (and flushed) since last cycle, and send
586 * Returns true if OK, false if trouble.
589 XLogSend(StringInfo outMsg)
591 XLogRecPtr SendRqstPtr;
592 char activitymsg[50];
593 /* use volatile pointer to prevent code rearrangement */
594 volatile WalSnd *walsnd = MyWalSnd;
597 * Invalid position means that we have not yet received the initial
598 * CopyData message from the slave that indicates where to start the
601 if (sentPtr.xlogid == 0 &&
602 sentPtr.xrecoff == 0)
605 /* Attempt to send all records flushed to the disk already */
606 SendRqstPtr = GetWriteRecPtr();
608 /* Quick exit if nothing to do */
609 if (!XLByteLT(sentPtr, SendRqstPtr))
613 * We gather multiple records together by issuing just one XLogRead()
614 * of a suitable size, and send them as one CopyData message. Repeat
615 * until we've sent everything we can.
617 while (XLByteLT(sentPtr, SendRqstPtr))
624 * Figure out how much to send in one message. If there's less than
625 * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
626 * MAX_SEND_SIZE bytes, but round to page boundary.
628 * The rounding is not only for performance reasons. Walreceiver
629 * relies on the fact that we never split a WAL record across two
630 * messages. Since a long WAL record is split at page boundary into
631 * continuation records, page boundary is always a safe cut-off point.
632 * We also assume that SendRqstPtr never points in the middle of a
636 if (startptr.xrecoff >= XLogFileSize)
639 * crossing a logid boundary, skip the non-existent last log
640 * segment in previous logical log file.
642 startptr.xlogid += 1;
643 startptr.xrecoff = 0;
647 XLByteAdvance(endptr, MAX_SEND_SIZE);
648 /* round down to page boundary. */
649 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
650 /* if we went beyond SendRqstPtr, back off */
651 if (XLByteLT(SendRqstPtr, endptr))
652 endptr = SendRqstPtr;
655 * OK to read and send the slice.
657 * We don't need to convert the xlogid/xrecoff from host byte order
658 * to network byte order because the both server can be expected to
659 * have the same byte order. If they have different byte order, we
662 pq_sendbytes(outMsg, (char *) &startptr, sizeof(startptr));
664 if (endptr.xlogid != startptr.xlogid)
666 Assert(endptr.xlogid == startptr.xlogid + 1);
667 nbytes = endptr.xrecoff + XLogFileSize - startptr.xrecoff;
670 nbytes = endptr.xrecoff - startptr.xrecoff;
675 * Read the log directly into the output buffer to prevent
676 * extra memcpy calls.
678 enlargeStringInfo(outMsg, nbytes);
680 XLogRead(&outMsg->data[outMsg->len], startptr, nbytes);
681 outMsg->len += nbytes;
682 outMsg->data[outMsg->len] = '\0';
684 pq_putmessage('d', outMsg->data, outMsg->len);
685 resetStringInfo(outMsg);
688 /* Update shared memory status */
689 SpinLockAcquire(&walsnd->mutex);
690 walsnd->sentPtr = sentPtr;
691 SpinLockRelease(&walsnd->mutex);
693 /* Flush pending output */
697 /* Report progress of XLOG streaming in PS display */
698 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
699 sentPtr.xlogid, sentPtr.xrecoff);
700 set_ps_display(activitymsg, false);
705 /* SIGHUP: set flag to re-read config file at next convenient time */
707 WalSndSigHupHandler(SIGNAL_ARGS)
712 /* SIGTERM: set flag to shut down */
714 WalSndShutdownHandler(SIGNAL_ARGS)
716 shutdown_requested = true;
720 * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
722 * Some backend has bought the farm,
723 * so we need to stop what we're doing and exit.
726 WalSndQuickDieHandler(SIGNAL_ARGS)
728 PG_SETMASK(&BlockSig);
731 * We DO NOT want to run proc_exit() callbacks -- we're here because
732 * shared memory may be corrupted, so we don't want to try to clean up our
733 * transaction. Just nail the windows shut and get out of town. Now that
734 * there's an atexit callback to prevent third-party code from breaking
735 * things by calling exit() directly, we have to reset the callbacks
736 * explicitly to make this work as intended.
741 * Note we do exit(2) not exit(0). This is to force the postmaster into a
742 * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
743 * backend. This is necessary precisely because we don't clean up our
744 * shared memory state. (The "dead man switch" mechanism in pmsignal.c
745 * should ensure the postmaster sees this as a crash, too, but no harm
746 * in being doubly sure.)
751 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
753 WalSndLastCycleHandler(SIGNAL_ARGS)
755 ready_to_stop = true;
758 /* Set up signal handlers */
762 /* Set up signal handlers */
763 pqsignal(SIGHUP, WalSndSigHupHandler); /* set flag to read config file */
764 pqsignal(SIGINT, SIG_IGN); /* not used */
765 pqsignal(SIGTERM, WalSndShutdownHandler); /* request shutdown */
766 pqsignal(SIGQUIT, WalSndQuickDieHandler); /* hard crash time */
767 pqsignal(SIGALRM, SIG_IGN);
768 pqsignal(SIGPIPE, SIG_IGN);
769 pqsignal(SIGUSR1, SIG_IGN); /* not used */
770 pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and shutdown */
772 /* Reset some signals that are accepted by postmaster but not here */
773 pqsignal(SIGCHLD, SIG_DFL);
774 pqsignal(SIGTTIN, SIG_DFL);
775 pqsignal(SIGTTOU, SIG_DFL);
776 pqsignal(SIGCONT, SIG_DFL);
777 pqsignal(SIGWINCH, SIG_DFL);
780 /* Report shared-memory space needed by WalSndShmemInit */
782 WalSndShmemSize(void)
786 size = offsetof(WalSndCtlData, walsnds);
787 size = add_size(size, mul_size(MaxWalSenders, sizeof(WalSnd)));
792 /* Allocate and initialize walsender-related shared memory */
794 WalSndShmemInit(void)
799 WalSndCtl = (WalSndCtlData *)
800 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
802 if (WalSndCtl == NULL)
804 (errcode(ERRCODE_OUT_OF_MEMORY),
805 errmsg("not enough shared memory for walsender")));
807 return; /* already initialized */
809 /* Initialize the data structures */
810 MemSet(WalSndCtl, 0, WalSndShmemSize());
812 for (i = 0; i < MaxWalSenders; i++)
814 WalSnd *walsnd = &WalSndCtl->walsnds[i];
815 SpinLockInit(&walsnd->mutex);
820 * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
824 GetOldestWALSendPointer(void)
826 XLogRecPtr oldest = {0, 0};
830 for (i = 0; i < MaxWalSenders; i++)
832 /* use volatile pointer to prevent code rearrangement */
833 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
836 if (walsnd->pid == 0)
839 SpinLockAcquire(&walsnd->mutex);
840 recptr = walsnd->sentPtr;
841 SpinLockRelease(&walsnd->mutex);
843 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
846 if (!found || XLByteLT(recptr, oldest))