granicus.if.org Git - postgresql/blob - src/backend/replication/walsender.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * walsender.c
   4  *
   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).
  14  *
  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.
  21  *
  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.
  26  *
  27  *
  28  * Portions Copyright (c) 2010-2012, PostgreSQL Global Development Group
  29  *
  30  * IDENTIFICATION
  31  *        src/backend/replication/walsender.c
  32  *
  33  *-------------------------------------------------------------------------
  34  */
  35 #include "postgres.h"
  36
  37 #include <signal.h>
  38 #include <unistd.h>
  39
  40 #include "access/transam.h"
  41 #include "access/xlog_internal.h"
  42 #include "catalog/pg_type.h"
  43 #include "funcapi.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"
  67
  68
  69 /* Array of WalSnds in shared memory */
  70 WalSndCtlData *WalSndCtl = NULL;
  71
  72 /* My slot in the shared memory array */
  73 WalSnd     *MyWalSnd = NULL;
  74
  75 /* Global state */
  76 bool            am_walsender = false;           /* Am I a walsender process ? */
  77 bool            am_cascading_walsender = false;         /* Am I cascading WAL to
  78                                                                                                  * another standby ? */
  79
  80 /* User-settable parameters for walsender */
  81 int                     max_wal_senders = 0;    /* the maximum number of concurrent walsenders */
  82 int                     replication_timeout = 60 * 1000;        /* maximum time to send one
  83                                                                                                  * WAL data message */
  84
  85 /*
  86  * These variables are used similarly to openLogFile/Id/Seg/Off,
  87  * but for walsender to read the XLOG.
  88  */
  89 static int      sendFile = -1;
  90 static XLogSegNo sendSegNo = 0;
  91 static uint32 sendOff = 0;
  92
  93 /*
  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.)
  96  */
  97 static XLogRecPtr sentPtr = 0;
  98
  99 /*
 100  * Buffer for processing reply messages.
 101  */
 102 static StringInfoData reply_message;
 103
 104 /*
 105  * Timestamp of the last receipt of the reply from the standby.
 106  */
 107 static TimestampTz last_reply_timestamp;
 108
 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;
 113
 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);
 120
 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);
 135
 136
 137 /* Main entry point for walsender process */
 138 int
 139 WalSenderMain(void)
 140 {
 141         MemoryContext walsnd_context;
 142
 143         am_cascading_walsender = RecoveryInProgress();
 144
 145         /* Create a per-walsender data structure in shared memory */
 146         InitWalSnd();
 147
 148         /*
 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.
 153          *
 154          * XXX: we don't actually attempt error recovery in walsender, we just
 155          * close the connection and exit.
 156          */
 157         walsnd_context = AllocSetContextCreate(TopMemoryContext,
 158                                                                                    "Wal Sender",
 159                                                                                    ALLOCSET_DEFAULT_MINSIZE,
 160                                                                                    ALLOCSET_DEFAULT_INITSIZE,
 161                                                                                    ALLOCSET_DEFAULT_MAXSIZE);
 162         MemoryContextSwitchTo(walsnd_context);
 163
 164         /* Set up resource owner */
 165         CurrentResourceOwner = ResourceOwnerCreate(NULL, "walsender top-level resource owner");
 166
 167         /* Unblock signals (they were blocked when the postmaster forked us) */
 168         PG_SETMASK(&UnBlockSig);
 169
 170         /*
 171          * Use the recovery target timeline ID during recovery
 172          */
 173         if (am_cascading_walsender)
 174                 ThisTimeLineID = GetRecoveryTargetTLI();
 175
 176         /* Tell the standby that walsender is ready for receiving commands */
 177         ReadyForQuery(DestRemote);
 178
 179         /* Handle handshake messages before streaming */
 180         WalSndHandshake();
 181
 182         /* Initialize shared memory status */
 183         {
 184                 /* use volatile pointer to prevent code rearrangement */
 185                 volatile WalSnd *walsnd = MyWalSnd;
 186
 187                 SpinLockAcquire(&walsnd->mutex);
 188                 walsnd->sentPtr = sentPtr;
 189                 SpinLockRelease(&walsnd->mutex);
 190         }
 191
 192         SyncRepInitConfig();
 193
 194         /* Main loop of walsender */
 195         return WalSndLoop();
 196 }
 197
 198 /*
 199  * Execute commands from walreceiver, until we enter streaming mode.
 200  */
 201 static void
 202 WalSndHandshake(void)
 203 {
 204         StringInfoData input_message;
 205         bool            replication_started = false;
 206
 207         initStringInfo(&input_message);
 208
 209         while (!replication_started)
 210         {
 211                 int                     firstchar;
 212
 213                 WalSndSetState(WALSNDSTATE_STARTUP);
 214                 set_ps_display("idle", false);
 215
 216                 /* Wait for a command to arrive */
 217                 firstchar = pq_getbyte();
 218
 219                 /*
 220                  * Emergency bailout if postmaster has died.  This is to avoid the
 221                  * necessity for manual cleanup of all postmaster children.
 222                  */
 223                 if (!PostmasterIsAlive())
 224                         exit(1);
 225
 226                 /*
 227                  * Check for any other interesting events that happened while we
 228                  * slept.
 229                  */
 230                 if (got_SIGHUP)
 231                 {
 232                         got_SIGHUP = false;
 233                         ProcessConfigFile(PGC_SIGHUP);
 234                 }
 235
 236                 if (firstchar != EOF)
 237                 {
 238                         /*
 239                          * Read the message contents. This is expected to be done without
 240                          * blocking because we've been able to get message type code.
 241                          */
 242                         if (pq_getmessage(&input_message, 0))
 243                                 firstchar = EOF;        /* suitable message already logged */
 244                 }
 245
 246                 /* Handle the very limited subset of commands expected in this phase */
 247                 switch (firstchar)
 248                 {
 249                         case 'Q':                       /* Query message */
 250                                 {
 251                                         const char *query_string;
 252
 253                                         query_string = pq_getmsgstring(&input_message);
 254                                         pq_getmsgend(&input_message);
 255
 256                                         if (HandleReplicationCommand(query_string))
 257                                                 replication_started = true;
 258                                 }
 259                                 break;
 260
 261                         case 'X':
 262                                 /* standby is closing the connection */
 263                                 proc_exit(0);
 264
 265                         case EOF:
 266                                 /* standby disconnected unexpectedly */
 267                                 ereport(COMMERROR,
 268                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 269                                                  errmsg("unexpected EOF on standby connection")));
 270                                 proc_exit(0);
 271
 272                         default:
 273                                 ereport(FATAL,
 274                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 275                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 276                 }
 277         }
 278 }
 279
 280 /*
 281  * IDENTIFY_SYSTEM
 282  */
 283 static void
 284 IdentifySystem(void)
 285 {
 286         StringInfoData buf;
 287         char            sysid[32];
 288         char            tli[11];
 289         char            xpos[MAXFNAMELEN];
 290         XLogRecPtr      logptr;
 291
 292         /*
 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.
 295          */
 296
 297         snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
 298                          GetSystemIdentifier());
 299         snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
 300
 301         logptr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetInsertRecPtr();
 302
 303         snprintf(xpos, sizeof(xpos), "%X/%X", (uint32) (logptr >> 32), (uint32) logptr);
 304
 305         /* Send a RowDescription message */
 306         pq_beginmessage(&buf, 'T');
 307         pq_sendint(&buf, 3, 2);         /* 3 fields */
 308
 309         /* first field */
 310         pq_sendstring(&buf, "systemid");        /* col name */
 311         pq_sendint(&buf, 0, 4);         /* table oid */
 312         pq_sendint(&buf, 0, 2);         /* attnum */
 313         pq_sendint(&buf, TEXTOID, 4);           /* type oid */
 314         pq_sendint(&buf, -1, 2);        /* typlen */
 315         pq_sendint(&buf, 0, 4);         /* typmod */
 316         pq_sendint(&buf, 0, 2);         /* format code */
 317
 318         /* second field */
 319         pq_sendstring(&buf, "timeline");        /* col name */
 320         pq_sendint(&buf, 0, 4);         /* table oid */
 321         pq_sendint(&buf, 0, 2);         /* attnum */
 322         pq_sendint(&buf, INT4OID, 4);           /* type oid */
 323         pq_sendint(&buf, 4, 2);         /* typlen */
 324         pq_sendint(&buf, 0, 4);         /* typmod */
 325         pq_sendint(&buf, 0, 2);         /* format code */
 326
 327         /* third field */
 328         pq_sendstring(&buf, "xlogpos");
 329         pq_sendint(&buf, 0, 4);
 330         pq_sendint(&buf, 0, 2);
 331         pq_sendint(&buf, TEXTOID, 4);
 332         pq_sendint(&buf, -1, 2);
 333         pq_sendint(&buf, 0, 4);
 334         pq_sendint(&buf, 0, 2);
 335         pq_endmessage(&buf);
 336
 337         /* Send a DataRow message */
 338         pq_beginmessage(&buf, 'D');
 339         pq_sendint(&buf, 3, 2);         /* # of columns */
 340         pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
 341         pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
 342         pq_sendint(&buf, strlen(tli), 4);       /* col2 len */
 343         pq_sendbytes(&buf, (char *) tli, strlen(tli));
 344         pq_sendint(&buf, strlen(xpos), 4);      /* col3 len */
 345         pq_sendbytes(&buf, (char *) xpos, strlen(xpos));
 346
 347         pq_endmessage(&buf);
 348
 349         /* Send CommandComplete and ReadyForQuery messages */
 350         EndCommand("SELECT", DestRemote);
 351         ReadyForQuery(DestRemote);
 352         /* ReadyForQuery did pq_flush for us */
 353 }
 354
 355 /*
 356  * START_REPLICATION
 357  */
 358 static void
 359 StartReplication(StartReplicationCmd *cmd)
 360 {
 361         StringInfoData buf;
 362
 363         /*
 364          * Let postmaster know that we're streaming. Once we've declared us as a
 365          * WAL sender process, postmaster will let us outlive the bgwriter and
 366          * kill us last in the shutdown sequence, so we get a chance to stream all
 367          * remaining WAL at shutdown, including the shutdown checkpoint. Note that
 368          * there's no going back, and we mustn't write any WAL records after this.
 369          */
 370         MarkPostmasterChildWalSender();
 371         SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
 372
 373         /*
 374          * When promoting a cascading standby, postmaster sends SIGUSR2 to any
 375          * cascading walsenders to kill them. But there is a corner-case where
 376          * such walsender fails to receive SIGUSR2 and survives a standby
 377          * promotion unexpectedly. This happens when postmaster sends SIGUSR2
 378          * before the walsender marks itself as a WAL sender, because postmaster
 379          * sends SIGUSR2 to only the processes marked as a WAL sender.
 380          *
 381          * To avoid this corner-case, if recovery is NOT in progress even though
 382          * the walsender is cascading one, we do the same thing as SIGUSR2 signal
 383          * handler does, i.e., set walsender_ready_to_stop to true. Which causes
 384          * the walsender to end later.
 385          *
 386          * When terminating cascading walsenders, usually postmaster writes the
 387          * log message announcing the terminations. But there is a race condition
 388          * here. If there is no walsender except this process before reaching
 389          * here, postmaster thinks that there is no walsender and suppresses that
 390          * log message. To handle this case, we always emit that log message here.
 391          * This might cause duplicate log messages, but which is less likely to
 392          * happen, so it's not worth writing some code to suppress them.
 393          */
 394         if (am_cascading_walsender && !RecoveryInProgress())
 395         {
 396                 ereport(LOG,
 397                    (errmsg("terminating walsender process to force cascaded standby "
 398                                    "to update timeline and reconnect")));
 399                 walsender_ready_to_stop = true;
 400         }
 401
 402         /*
 403          * We assume here that we're logging enough information in the WAL for
 404          * log-shipping, since this is checked in PostmasterMain().
 405          *
 406          * NOTE: wal_level can only change at shutdown, so in most cases it is
 407          * difficult for there to be WAL data that we can still see that was
 408          * written at wal_level='minimal'.
 409          */
 410
 411         /*
 412          * When we first start replication the standby will be behind the primary.
 413          * For some applications, for example, synchronous replication, it is
 414          * important to have a clear state for this initial catchup mode, so we
 415          * can trigger actions when we change streaming state later. We may stay
 416          * in this state for a long time, which is exactly why we want to be able
 417          * to monitor whether or not we are still here.
 418          */
 419         WalSndSetState(WALSNDSTATE_CATCHUP);
 420
 421         /* Send a CopyBothResponse message, and start streaming */
 422         pq_beginmessage(&buf, 'W');
 423         pq_sendbyte(&buf, 0);
 424         pq_sendint(&buf, 0, 2);
 425         pq_endmessage(&buf);
 426         pq_flush();
 427
 428         /*
 429          * Initialize position to the received one, then the xlog records begin to
 430          * be shipped from that position
 431          */
 432         sentPtr = cmd->startpoint;
 433 }
 434
 435 /*
 436  * Execute an incoming replication command.
 437  */
 438 static bool
 439 HandleReplicationCommand(const char *cmd_string)
 440 {
 441         bool            replication_started = false;
 442         int                     parse_rc;
 443         Node       *cmd_node;
 444         MemoryContext cmd_context;
 445         MemoryContext old_context;
 446
 447         elog(DEBUG1, "received replication command: %s", cmd_string);
 448
 449         cmd_context = AllocSetContextCreate(CurrentMemoryContext,
 450                                                                                 "Replication command context",
 451                                                                                 ALLOCSET_DEFAULT_MINSIZE,
 452                                                                                 ALLOCSET_DEFAULT_INITSIZE,
 453                                                                                 ALLOCSET_DEFAULT_MAXSIZE);
 454         old_context = MemoryContextSwitchTo(cmd_context);
 455
 456         replication_scanner_init(cmd_string);
 457         parse_rc = replication_yyparse();
 458         if (parse_rc != 0)
 459                 ereport(ERROR,
 460                                 (errcode(ERRCODE_SYNTAX_ERROR),
 461                                  (errmsg_internal("replication command parser returned %d",
 462                                                                   parse_rc))));
 463
 464         cmd_node = replication_parse_result;
 465
 466         switch (cmd_node->type)
 467         {
 468                 case T_IdentifySystemCmd:
 469                         IdentifySystem();
 470                         break;
 471
 472                 case T_StartReplicationCmd:
 473                         StartReplication((StartReplicationCmd *) cmd_node);
 474
 475                         /* break out of the loop */
 476                         replication_started = true;
 477                         break;
 478
 479                 case T_BaseBackupCmd:
 480                         SendBaseBackup((BaseBackupCmd *) cmd_node);
 481
 482                         /* Send CommandComplete and ReadyForQuery messages */
 483                         EndCommand("SELECT", DestRemote);
 484                         ReadyForQuery(DestRemote);
 485                         /* ReadyForQuery did pq_flush for us */
 486                         break;
 487
 488                 default:
 489                         ereport(FATAL,
 490                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 491                                          errmsg("invalid standby query string: %s", cmd_string)));
 492         }
 493
 494         /* done */
 495         MemoryContextSwitchTo(old_context);
 496         MemoryContextDelete(cmd_context);
 497
 498         return replication_started;
 499 }
 500
 501 /*
 502  * Check if the remote end has closed the connection.
 503  */
 504 static void
 505 ProcessRepliesIfAny(void)
 506 {
 507         unsigned char firstchar;
 508         int                     r;
 509         bool            received = false;
 510
 511         for (;;)
 512         {
 513                 r = pq_getbyte_if_available(&firstchar);
 514                 if (r < 0)
 515                 {
 516                         /* unexpected error or EOF */
 517                         ereport(COMMERROR,
 518                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 519                                          errmsg("unexpected EOF on standby connection")));
 520                         proc_exit(0);
 521                 }
 522                 if (r == 0)
 523                 {
 524                         /* no data available without blocking */
 525                         break;
 526                 }
 527
 528                 /* Handle the very limited subset of commands expected in this phase */
 529                 switch (firstchar)
 530                 {
 531                                 /*
 532                                  * 'd' means a standby reply wrapped in a CopyData packet.
 533                                  */
 534                         case 'd':
 535                                 ProcessStandbyMessage();
 536                                 received = true;
 537                                 break;
 538
 539                                 /*
 540                                  * 'X' means that the standby is closing down the socket.
 541                                  */
 542                         case 'X':
 543                                 proc_exit(0);
 544
 545                         default:
 546                                 ereport(FATAL,
 547                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 548                                                  errmsg("invalid standby message type \"%c\"",
 549                                                                 firstchar)));
 550                 }
 551         }
 552
 553         /*
 554          * Save the last reply timestamp if we've received at least one reply.
 555          */
 556         if (received)
 557                 last_reply_timestamp = GetCurrentTimestamp();
 558 }
 559
 560 /*
 561  * Process a status update message received from standby.
 562  */
 563 static void
 564 ProcessStandbyMessage(void)
 565 {
 566         char            msgtype;
 567
 568         resetStringInfo(&reply_message);
 569
 570         /*
 571          * Read the message contents.
 572          */
 573         if (pq_getmessage(&reply_message, 0))
 574         {
 575                 ereport(COMMERROR,
 576                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 577                                  errmsg("unexpected EOF on standby connection")));
 578                 proc_exit(0);
 579         }
 580
 581         /*
 582          * Check message type from the first byte.
 583          */
 584         msgtype = pq_getmsgbyte(&reply_message);
 585
 586         switch (msgtype)
 587         {
 588                 case 'r':
 589                         ProcessStandbyReplyMessage();
 590                         break;
 591
 592                 case 'h':
 593                         ProcessStandbyHSFeedbackMessage();
 594                         break;
 595
 596                 default:
 597                         ereport(COMMERROR,
 598                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 599                                          errmsg("unexpected message type \"%c\"", msgtype)));
 600                         proc_exit(0);
 601         }
 602 }
 603
 604 /*
 605  * Regular reply from standby advising of WAL positions on standby server.
 606  */
 607 static void
 608 ProcessStandbyReplyMessage(void)
 609 {
 610         StandbyReplyMessage reply;
 611
 612         pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyReplyMessage));
 613
 614         elog(DEBUG2, "write %X/%X flush %X/%X apply %X/%X",
 615                  (uint32) (reply.write << 32), (uint32) reply.write,
 616                  (uint32) (reply.flush << 32), (uint32) reply.flush,
 617                  (uint32) (reply.apply << 32), (uint32) reply.apply);
 618
 619         /*
 620          * Update shared state for this WalSender process based on reply data from
 621          * standby.
 622          */
 623         {
 624                 /* use volatile pointer to prevent code rearrangement */
 625                 volatile WalSnd *walsnd = MyWalSnd;
 626
 627                 SpinLockAcquire(&walsnd->mutex);
 628                 walsnd->write = reply.write;
 629                 walsnd->flush = reply.flush;
 630                 walsnd->apply = reply.apply;
 631                 SpinLockRelease(&walsnd->mutex);
 632         }
 633
 634         if (!am_cascading_walsender)
 635                 SyncRepReleaseWaiters();
 636 }
 637
 638 /*
 639  * Hot Standby feedback
 640  */
 641 static void
 642 ProcessStandbyHSFeedbackMessage(void)
 643 {
 644         StandbyHSFeedbackMessage reply;
 645         TransactionId nextXid;
 646         uint32          nextEpoch;
 647
 648         /* Decipher the reply message */
 649         pq_copymsgbytes(&reply_message, (char *) &reply,
 650                                         sizeof(StandbyHSFeedbackMessage));
 651
 652         elog(DEBUG2, "hot standby feedback xmin %u epoch %u",
 653                  reply.xmin,
 654                  reply.epoch);
 655
 656         /* Ignore invalid xmin (can't actually happen with current walreceiver) */
 657         if (!TransactionIdIsNormal(reply.xmin))
 658                 return;
 659
 660         /*
 661          * Check that the provided xmin/epoch are sane, that is, not in the future
 662          * and not so far back as to be already wrapped around.  Ignore if not.
 663          *
 664          * Epoch of nextXid should be same as standby, or if the counter has
 665          * wrapped, then one greater than standby.
 666          */
 667         GetNextXidAndEpoch(&nextXid, &nextEpoch);
 668
 669         if (reply.xmin <= nextXid)
 670         {
 671                 if (reply.epoch != nextEpoch)
 672                         return;
 673         }
 674         else
 675         {
 676                 if (reply.epoch + 1 != nextEpoch)
 677                         return;
 678         }
 679
 680         if (!TransactionIdPrecedesOrEquals(reply.xmin, nextXid))
 681                 return;                                 /* epoch OK, but it's wrapped around */
 682
 683         /*
 684          * Set the WalSender's xmin equal to the standby's requested xmin, so that
 685          * the xmin will be taken into account by GetOldestXmin.  This will hold
 686          * back the removal of dead rows and thereby prevent the generation of
 687          * cleanup conflicts on the standby server.
 688          *
 689          * There is a small window for a race condition here: although we just
 690          * checked that reply.xmin precedes nextXid, the nextXid could have gotten
 691          * advanced between our fetching it and applying the xmin below, perhaps
 692          * far enough to make reply.xmin wrap around.  In that case the xmin we
 693          * set here would be "in the future" and have no effect.  No point in
 694          * worrying about this since it's too late to save the desired data
 695          * anyway.      Assuming that the standby sends us an increasing sequence of
 696          * xmins, this could only happen during the first reply cycle, else our
 697          * own xmin would prevent nextXid from advancing so far.
 698          *
 699          * We don't bother taking the ProcArrayLock here.  Setting the xmin field
 700          * is assumed atomic, and there's no real need to prevent a concurrent
 701          * GetOldestXmin.  (If we're moving our xmin forward, this is obviously
 702          * safe, and if we're moving it backwards, well, the data is at risk
 703          * already since a VACUUM could have just finished calling GetOldestXmin.)
 704          */
 705         MyPgXact->xmin = reply.xmin;
 706 }
 707
 708 /* Main loop of walsender process */
 709 static int
 710 WalSndLoop(void)
 711 {
 712         char       *output_message;
 713         bool            caughtup = false;
 714
 715         /*
 716          * Allocate buffer that will be used for each output message.  We do this
 717          * just once to reduce palloc overhead.  The buffer must be made large
 718          * enough for maximum-sized messages.
 719          */
 720         output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
 721
 722         /*
 723          * Allocate buffer that will be used for processing reply messages.  As
 724          * above, do this just once to reduce palloc overhead.
 725          */
 726         initStringInfo(&reply_message);
 727
 728         /* Initialize the last reply timestamp */
 729         last_reply_timestamp = GetCurrentTimestamp();
 730
 731         /* Loop forever, unless we get an error */
 732         for (;;)
 733         {
 734                 /* Clear any already-pending wakeups */
 735                 ResetLatch(&MyWalSnd->latch);
 736
 737                 /*
 738                  * Emergency bailout if postmaster has died.  This is to avoid the
 739                  * necessity for manual cleanup of all postmaster children.
 740                  */
 741                 if (!PostmasterIsAlive())
 742                         exit(1);
 743
 744                 /* Process any requests or signals received recently */
 745                 if (got_SIGHUP)
 746                 {
 747                         got_SIGHUP = false;
 748                         ProcessConfigFile(PGC_SIGHUP);
 749                         SyncRepInitConfig();
 750                 }
 751
 752                 /* Normal exit from the walsender is here */
 753                 if (walsender_shutdown_requested)
 754                 {
 755                         /* Inform the standby that XLOG streaming is done */
 756                         pq_puttextmessage('C', "COPY 0");
 757                         pq_flush();
 758
 759                         proc_exit(0);
 760                 }
 761
 762                 /* Check for input from the client */
 763                 ProcessRepliesIfAny();
 764
 765                 /*
 766                  * If we don't have any pending data in the output buffer, try to send
 767                  * some more.  If there is some, we don't bother to call XLogSend
 768                  * again until we've flushed it ... but we'd better assume we are not
 769                  * caught up.
 770                  */
 771                 if (!pq_is_send_pending())
 772                         XLogSend(output_message, &caughtup);
 773                 else
 774                         caughtup = false;
 775
 776                 /* Try to flush pending output to the client */
 777                 if (pq_flush_if_writable() != 0)
 778                         break;
 779
 780                 /* If nothing remains to be sent right now ... */
 781                 if (caughtup && !pq_is_send_pending())
 782                 {
 783                         /*
 784                          * If we're in catchup state, move to streaming.  This is an
 785                          * important state change for users to know about, since before
 786                          * this point data loss might occur if the primary dies and we
 787                          * need to failover to the standby. The state change is also
 788                          * important for synchronous replication, since commits that
 789                          * started to wait at that point might wait for some time.
 790                          */
 791                         if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
 792                         {
 793                                 ereport(DEBUG1,
 794                                          (errmsg("standby \"%s\" has now caught up with primary",
 795                                                          application_name)));
 796                                 WalSndSetState(WALSNDSTATE_STREAMING);
 797                         }
 798
 799                         /*
 800                          * When SIGUSR2 arrives, we send any outstanding logs up to the
 801                          * shutdown checkpoint record (i.e., the latest record) and exit.
 802                          * This may be a normal termination at shutdown, or a promotion,
 803                          * the walsender is not sure which.
 804                          */
 805                         if (walsender_ready_to_stop)
 806                         {
 807                                 /* ... let's just be real sure we're caught up ... */
 808                                 XLogSend(output_message, &caughtup);
 809                                 if (caughtup && !pq_is_send_pending())
 810                                 {
 811                                         walsender_shutdown_requested = true;
 812                                         continue;       /* don't want to wait more */
 813                                 }
 814                         }
 815                 }
 816
 817                 /*
 818                  * We don't block if not caught up, unless there is unsent data
 819                  * pending in which case we'd better block until the socket is
 820                  * write-ready.  This test is only needed for the case where XLogSend
 821                  * loaded a subset of the available data but then pq_flush_if_writable
 822                  * flushed it all --- we should immediately try to send more.
 823                  */
 824                 if (caughtup || pq_is_send_pending())
 825                 {
 826                         TimestampTz timeout = 0;
 827                         long            sleeptime = 10000;              /* 10 s */
 828                         int                     wakeEvents;
 829
 830                         wakeEvents = WL_LATCH_SET | WL_POSTMASTER_DEATH |
 831                                 WL_SOCKET_READABLE | WL_TIMEOUT;
 832
 833                         if (pq_is_send_pending())
 834                                 wakeEvents |= WL_SOCKET_WRITEABLE;
 835                         else
 836                         {
 837                                 WalSndKeepalive(output_message);
 838                                 /* Try to flush pending output to the client */
 839                                 if (pq_flush_if_writable() != 0)
 840                                         break;
 841                         }
 842
 843                         /* Determine time until replication timeout */
 844                         if (replication_timeout > 0)
 845                         {
 846                                 timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
 847                                                                                                           replication_timeout);
 848                                 sleeptime = 1 + (replication_timeout / 10);
 849                         }
 850
 851                         /* Sleep until something happens or replication timeout */
 852                         WaitLatchOrSocket(&MyWalSnd->latch, wakeEvents,
 853                                                           MyProcPort->sock, sleeptime);
 854
 855                         /*
 856                          * Check for replication timeout.  Note we ignore the corner case
 857                          * possibility that the client replied just as we reached the
 858                          * timeout ... he's supposed to reply *before* that.
 859                          */
 860                         if (replication_timeout > 0 &&
 861                                 GetCurrentTimestamp() >= timeout)
 862                         {
 863                                 /*
 864                                  * Since typically expiration of replication timeout means
 865                                  * communication problem, we don't send the error message to
 866                                  * the standby.
 867                                  */
 868                                 ereport(COMMERROR,
 869                                                 (errmsg("terminating walsender process due to replication timeout")));
 870                                 break;
 871                         }
 872                 }
 873         }
 874
 875         /*
 876          * Get here on send failure.  Clean up and exit.
 877          *
 878          * Reset whereToSendOutput to prevent ereport from attempting to send any
 879          * more messages to the standby.
 880          */
 881         if (whereToSendOutput == DestRemote)
 882                 whereToSendOutput = DestNone;
 883
 884         proc_exit(0);
 885         return 1;                                       /* keep the compiler quiet */
 886 }
 887
 888 /* Initialize a per-walsender data structure for this walsender process */
 889 static void
 890 InitWalSnd(void)
 891 {
 892         int                     i;
 893
 894         /*
 895          * WalSndCtl should be set up already (we inherit this by fork() or
 896          * EXEC_BACKEND mechanism from the postmaster).
 897          */
 898         Assert(WalSndCtl != NULL);
 899         Assert(MyWalSnd == NULL);
 900
 901         /*
 902          * Find a free walsender slot and reserve it. If this fails, we must be
 903          * out of WalSnd structures.
 904          */
 905         for (i = 0; i < max_wal_senders; i++)
 906         {
 907                 /* use volatile pointer to prevent code rearrangement */
 908                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 909
 910                 SpinLockAcquire(&walsnd->mutex);
 911
 912                 if (walsnd->pid != 0)
 913                 {
 914                         SpinLockRelease(&walsnd->mutex);
 915                         continue;
 916                 }
 917                 else
 918                 {
 919                         /*
 920                          * Found a free slot. Reserve it for us.
 921                          */
 922                         walsnd->pid = MyProcPid;
 923                         MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
 924                         walsnd->state = WALSNDSTATE_STARTUP;
 925                         SpinLockRelease(&walsnd->mutex);
 926                         /* don't need the lock anymore */
 927                         OwnLatch((Latch *) &walsnd->latch);
 928                         MyWalSnd = (WalSnd *) walsnd;
 929
 930                         break;
 931                 }
 932         }
 933         if (MyWalSnd == NULL)
 934                 ereport(FATAL,
 935                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 936                                  errmsg("number of requested standby connections "
 937                                                 "exceeds max_wal_senders (currently %d)",
 938                                                 max_wal_senders)));
 939
 940         /* Arrange to clean up at walsender exit */
 941         on_shmem_exit(WalSndKill, 0);
 942 }
 943
 944 /* Destroy the per-walsender data structure for this walsender process */
 945 static void
 946 WalSndKill(int code, Datum arg)
 947 {
 948         Assert(MyWalSnd != NULL);
 949
 950         /*
 951          * Mark WalSnd struct no longer in use. Assume that no lock is required
 952          * for this.
 953          */
 954         MyWalSnd->pid = 0;
 955         DisownLatch(&MyWalSnd->latch);
 956
 957         /* WalSnd struct isn't mine anymore */
 958         MyWalSnd = NULL;
 959 }
 960
 961 /*
 962  * Read 'count' bytes from WAL into 'buf', starting at location 'startptr'
 963  *
 964  * XXX probably this should be improved to suck data directly from the
 965  * WAL buffers when possible.
 966  *
 967  * Will open, and keep open, one WAL segment stored in the global file
 968  * descriptor sendFile. This means if XLogRead is used once, there will
 969  * always be one descriptor left open until the process ends, but never
 970  * more than one.
 971  */
 972 void
 973 XLogRead(char *buf, XLogRecPtr startptr, Size count)
 974 {
 975         char       *p;
 976         XLogRecPtr      recptr;
 977         Size            nbytes;
 978         XLogSegNo       lastRemovedSegNo;
 979         XLogSegNo       segno;
 980
 981 retry:
 982         p = buf;
 983         recptr = startptr;
 984         nbytes = count;
 985
 986         while (nbytes > 0)
 987         {
 988                 uint32          startoff;
 989                 int                     segbytes;
 990                 int                     readbytes;
 991
 992                 startoff = recptr % XLogSegSize;
 993
 994                 if (sendFile < 0 || !XLByteInSeg(recptr, sendSegNo))
 995                 {
 996                         char            path[MAXPGPATH];
 997
 998                         /* Switch to another logfile segment */
 999                         if (sendFile >= 0)
1000                                 close(sendFile);
1001
1002                         XLByteToSeg(recptr, sendSegNo);
1003                         XLogFilePath(path, ThisTimeLineID, sendSegNo);
1004
1005                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
1006                         if (sendFile < 0)
1007                         {
1008                                 /*
1009                                  * If the file is not found, assume it's because the standby
1010                                  * asked for a too old WAL segment that has already been
1011                                  * removed or recycled.
1012                                  */
1013                                 if (errno == ENOENT)
1014                                         ereport(ERROR,
1015                                                         (errcode_for_file_access(),
1016                                                          errmsg("requested WAL segment %s has already been removed",
1017                                                                         XLogFileNameP(ThisTimeLineID, sendSegNo))));
1018                                 else
1019                                         ereport(ERROR,
1020                                                         (errcode_for_file_access(),
1021                                                          errmsg("could not open file \"%s\": %m",
1022                                                                         path)));
1023                         }
1024                         sendOff = 0;
1025                 }
1026
1027                 /* Need to seek in the file? */
1028                 if (sendOff != startoff)
1029                 {
1030                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
1031                                 ereport(ERROR,
1032                                                 (errcode_for_file_access(),
1033                                                  errmsg("could not seek in log segment %s to offset %u: %m",
1034                                                                 XLogFileNameP(ThisTimeLineID, sendSegNo),
1035                                                                 startoff)));
1036                         sendOff = startoff;
1037                 }
1038
1039                 /* How many bytes are within this segment? */
1040                 if (nbytes > (XLogSegSize - startoff))
1041                         segbytes = XLogSegSize - startoff;
1042                 else
1043                         segbytes = nbytes;
1044
1045                 readbytes = read(sendFile, p, segbytes);
1046                 if (readbytes <= 0)
1047                 {
1048                         ereport(ERROR,
1049                                         (errcode_for_file_access(),
1050                         errmsg("could not read from log segment %s, offset %u, length %lu: %m",
1051                                    XLogFileNameP(ThisTimeLineID, sendSegNo),
1052                                    sendOff, (unsigned long) segbytes)));
1053                 }
1054
1055                 /* Update state for read */
1056                 XLByteAdvance(recptr, readbytes);
1057
1058                 sendOff += readbytes;
1059                 nbytes -= readbytes;
1060                 p += readbytes;
1061         }
1062
1063         /*
1064          * After reading into the buffer, check that what we read was valid. We do
1065          * this after reading, because even though the segment was present when we
1066          * opened it, it might get recycled or removed while we read it. The
1067          * read() succeeds in that case, but the data we tried to read might
1068          * already have been overwritten with new WAL records.
1069          */
1070         XLogGetLastRemoved(&lastRemovedSegNo);
1071         XLByteToSeg(startptr, segno);
1072         if (segno <= lastRemovedSegNo)
1073                 ereport(ERROR,
1074                                 (errcode_for_file_access(),
1075                                  errmsg("requested WAL segment %s has already been removed",
1076                                                 XLogFileNameP(ThisTimeLineID, segno))));
1077
1078         /*
1079          * During recovery, the currently-open WAL file might be replaced with the
1080          * file of the same name retrieved from archive. So we always need to
1081          * check what we read was valid after reading into the buffer. If it's
1082          * invalid, we try to open and read the file again.
1083          */
1084         if (am_cascading_walsender)
1085         {
1086                 /* use volatile pointer to prevent code rearrangement */
1087                 volatile WalSnd *walsnd = MyWalSnd;
1088                 bool            reload;
1089
1090                 SpinLockAcquire(&walsnd->mutex);
1091                 reload = walsnd->needreload;
1092                 walsnd->needreload = false;
1093                 SpinLockRelease(&walsnd->mutex);
1094
1095                 if (reload && sendFile >= 0)
1096                 {
1097                         close(sendFile);
1098                         sendFile = -1;
1099
1100                         goto retry;
1101                 }
1102         }
1103 }
1104
1105 /*
1106  * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
1107  * but not yet sent to the client, and buffer it in the libpq output
1108  * buffer.
1109  *
1110  * msgbuf is a work area in which the output message is constructed.  It's
1111  * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
1112  * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
1113  *
1114  * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
1115  * *caughtup is set to false.
1116
1117  */
1118 static void
1119 XLogSend(char *msgbuf, bool *caughtup)
1120 {
1121         XLogRecPtr      SendRqstPtr;
1122         XLogRecPtr      startptr;
1123         XLogRecPtr      endptr;
1124         Size            nbytes;
1125         WalDataMessageHeader msghdr;
1126
1127         /*
1128          * Attempt to send all data that's already been written out and fsync'd to
1129          * disk.  We cannot go further than what's been written out given the
1130          * current implementation of XLogRead().  And in any case it's unsafe to
1131          * send WAL that is not securely down to disk on the master: if the master
1132          * subsequently crashes and restarts, slaves must not have applied any WAL
1133          * that gets lost on the master.
1134          */
1135         SendRqstPtr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetFlushRecPtr();
1136
1137         /* Quick exit if nothing to do */
1138         if (XLByteLE(SendRqstPtr, sentPtr))
1139         {
1140                 *caughtup = true;
1141                 return;
1142         }
1143
1144         /*
1145          * Figure out how much to send in one message. If there's no more than
1146          * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
1147          * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
1148          *
1149          * The rounding is not only for performance reasons. Walreceiver relies on
1150          * the fact that we never split a WAL record across two messages. Since a
1151          * long WAL record is split at page boundary into continuation records,
1152          * page boundary is always a safe cut-off point. We also assume that
1153          * SendRqstPtr never points to the middle of a WAL record.
1154          */
1155         startptr = sentPtr;
1156         endptr = startptr;
1157         XLByteAdvance(endptr, MAX_SEND_SIZE);
1158
1159         /* if we went beyond SendRqstPtr, back off */
1160         if (XLByteLE(SendRqstPtr, endptr))
1161         {
1162                 endptr = SendRqstPtr;
1163                 *caughtup = true;
1164         }
1165         else
1166         {
1167                 /* round down to page boundary. */
1168                 endptr -= (endptr % XLOG_BLCKSZ);
1169                 *caughtup = false;
1170         }
1171
1172         nbytes = endptr - startptr;
1173         Assert(nbytes <= MAX_SEND_SIZE);
1174
1175         /*
1176          * OK to read and send the slice.
1177          */
1178         msgbuf[0] = 'w';
1179
1180         /*
1181          * Read the log directly into the output buffer to avoid extra memcpy
1182          * calls.
1183          */
1184         XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
1185
1186         /*
1187          * We fill the message header last so that the send timestamp is taken as
1188          * late as possible.
1189          */
1190         msghdr.dataStart = startptr;
1191         msghdr.walEnd = SendRqstPtr;
1192         msghdr.sendTime = GetCurrentTimestamp();
1193
1194         memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
1195
1196         pq_putmessage_noblock('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
1197
1198         sentPtr = endptr;
1199
1200         /* Update shared memory status */
1201         {
1202                 /* use volatile pointer to prevent code rearrangement */
1203                 volatile WalSnd *walsnd = MyWalSnd;
1204
1205                 SpinLockAcquire(&walsnd->mutex);
1206                 walsnd->sentPtr = sentPtr;
1207                 SpinLockRelease(&walsnd->mutex);
1208         }
1209
1210         /* Report progress of XLOG streaming in PS display */
1211         if (update_process_title)
1212         {
1213                 char            activitymsg[50];
1214
1215                 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
1216                                  (uint32) (sentPtr >> 32), (uint32) sentPtr);
1217                 set_ps_display(activitymsg, false);
1218         }
1219
1220         return;
1221 }
1222
1223 /*
1224  * Request walsenders to reload the currently-open WAL file
1225  */
1226 void
1227 WalSndRqstFileReload(void)
1228 {
1229         int                     i;
1230
1231         for (i = 0; i < max_wal_senders; i++)
1232         {
1233                 /* use volatile pointer to prevent code rearrangement */
1234                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1235
1236                 if (walsnd->pid == 0)
1237                         continue;
1238
1239                 SpinLockAcquire(&walsnd->mutex);
1240                 walsnd->needreload = true;
1241                 SpinLockRelease(&walsnd->mutex);
1242         }
1243 }
1244
1245 /* SIGHUP: set flag to re-read config file at next convenient time */
1246 static void
1247 WalSndSigHupHandler(SIGNAL_ARGS)
1248 {
1249         int                     save_errno = errno;
1250
1251         got_SIGHUP = true;
1252         if (MyWalSnd)
1253                 SetLatch(&MyWalSnd->latch);
1254
1255         errno = save_errno;
1256 }
1257
1258 /* SIGTERM: set flag to shut down */
1259 static void
1260 WalSndShutdownHandler(SIGNAL_ARGS)
1261 {
1262         int                     save_errno = errno;
1263
1264         walsender_shutdown_requested = true;
1265         if (MyWalSnd)
1266                 SetLatch(&MyWalSnd->latch);
1267
1268         /*
1269          * Set the standard (non-walsender) state as well, so that we can abort
1270          * things like do_pg_stop_backup().
1271          */
1272         InterruptPending = true;
1273         ProcDiePending = true;
1274
1275         errno = save_errno;
1276 }
1277
1278 /*
1279  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
1280  *
1281  * Some backend has bought the farm,
1282  * so we need to stop what we're doing and exit.
1283  */
1284 static void
1285 WalSndQuickDieHandler(SIGNAL_ARGS)
1286 {
1287         PG_SETMASK(&BlockSig);
1288
1289         /*
1290          * We DO NOT want to run proc_exit() callbacks -- we're here because
1291          * shared memory may be corrupted, so we don't want to try to clean up our
1292          * transaction.  Just nail the windows shut and get out of town.  Now that
1293          * there's an atexit callback to prevent third-party code from breaking
1294          * things by calling exit() directly, we have to reset the callbacks
1295          * explicitly to make this work as intended.
1296          */
1297         on_exit_reset();
1298
1299         /*
1300          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
1301          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
1302          * backend.  This is necessary precisely because we don't clean up our
1303          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
1304          * should ensure the postmaster sees this as a crash, too, but no harm in
1305          * being doubly sure.)
1306          */
1307         exit(2);
1308 }
1309
1310 /* SIGUSR1: set flag to send WAL records */
1311 static void
1312 WalSndXLogSendHandler(SIGNAL_ARGS)
1313 {
1314         int                     save_errno = errno;
1315
1316         latch_sigusr1_handler();
1317
1318         errno = save_errno;
1319 }
1320
1321 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
1322 static void
1323 WalSndLastCycleHandler(SIGNAL_ARGS)
1324 {
1325         int                     save_errno = errno;
1326
1327         walsender_ready_to_stop = true;
1328         if (MyWalSnd)
1329                 SetLatch(&MyWalSnd->latch);
1330
1331         errno = save_errno;
1332 }
1333
1334 /* Set up signal handlers */
1335 void
1336 WalSndSignals(void)
1337 {
1338         /* Set up signal handlers */
1339         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
1340                                                                                                  * file */
1341         pqsignal(SIGINT, SIG_IGN);      /* not used */
1342         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
1343         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
1344         pqsignal(SIGALRM, SIG_IGN);
1345         pqsignal(SIGPIPE, SIG_IGN);
1346         pqsignal(SIGUSR1, WalSndXLogSendHandler);       /* request WAL sending */
1347         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
1348                                                                                                  * shutdown */
1349
1350         /* Reset some signals that are accepted by postmaster but not here */
1351         pqsignal(SIGCHLD, SIG_DFL);
1352         pqsignal(SIGTTIN, SIG_DFL);
1353         pqsignal(SIGTTOU, SIG_DFL);
1354         pqsignal(SIGCONT, SIG_DFL);
1355         pqsignal(SIGWINCH, SIG_DFL);
1356 }
1357
1358 /* Report shared-memory space needed by WalSndShmemInit */
1359 Size
1360 WalSndShmemSize(void)
1361 {
1362         Size            size = 0;
1363
1364         size = offsetof(WalSndCtlData, walsnds);
1365         size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
1366
1367         return size;
1368 }
1369
1370 /* Allocate and initialize walsender-related shared memory */
1371 void
1372 WalSndShmemInit(void)
1373 {
1374         bool            found;
1375         int                     i;
1376
1377         WalSndCtl = (WalSndCtlData *)
1378                 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
1379
1380         if (!found)
1381         {
1382                 /* First time through, so initialize */
1383                 MemSet(WalSndCtl, 0, WalSndShmemSize());
1384
1385                 for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
1386                         SHMQueueInit(&(WalSndCtl->SyncRepQueue[i]));
1387
1388                 for (i = 0; i < max_wal_senders; i++)
1389                 {
1390                         WalSnd     *walsnd = &WalSndCtl->walsnds[i];
1391
1392                         SpinLockInit(&walsnd->mutex);
1393                         InitSharedLatch(&walsnd->latch);
1394                 }
1395         }
1396 }
1397
1398 /* Wake up all walsenders */
1399 void
1400 WalSndWakeup(void)
1401 {
1402         int                     i;
1403
1404         for (i = 0; i < max_wal_senders; i++)
1405                 SetLatch(&WalSndCtl->walsnds[i].latch);
1406 }
1407
1408 /* Set state for current walsender (only called in walsender) */
1409 void
1410 WalSndSetState(WalSndState state)
1411 {
1412         /* use volatile pointer to prevent code rearrangement */
1413         volatile WalSnd *walsnd = MyWalSnd;
1414
1415         Assert(am_walsender);
1416
1417         if (walsnd->state == state)
1418                 return;
1419
1420         SpinLockAcquire(&walsnd->mutex);
1421         walsnd->state = state;
1422         SpinLockRelease(&walsnd->mutex);
1423 }
1424
1425 /*
1426  * Return a string constant representing the state. This is used
1427  * in system views, and should *not* be translated.
1428  */
1429 static const char *
1430 WalSndGetStateString(WalSndState state)
1431 {
1432         switch (state)
1433         {
1434                 case WALSNDSTATE_STARTUP:
1435                         return "startup";
1436                 case WALSNDSTATE_BACKUP:
1437                         return "backup";
1438                 case WALSNDSTATE_CATCHUP:
1439                         return "catchup";
1440                 case WALSNDSTATE_STREAMING:
1441                         return "streaming";
1442         }
1443         return "UNKNOWN";
1444 }
1445
1446
1447 /*
1448  * Returns activity of walsenders, including pids and xlog locations sent to
1449  * standby servers.
1450  */
1451 Datum
1452 pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
1453 {
1454 #define PG_STAT_GET_WAL_SENDERS_COLS    8
1455         ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
1456         TupleDesc       tupdesc;
1457         Tuplestorestate *tupstore;
1458         MemoryContext per_query_ctx;
1459         MemoryContext oldcontext;
1460         int                *sync_priority;
1461         int                     priority = 0;
1462         int                     sync_standby = -1;
1463         int                     i;
1464
1465         /* check to see if caller supports us returning a tuplestore */
1466         if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
1467                 ereport(ERROR,
1468                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1469                                  errmsg("set-valued function called in context that cannot accept a set")));
1470         if (!(rsinfo->allowedModes & SFRM_Materialize))
1471                 ereport(ERROR,
1472                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1473                                  errmsg("materialize mode required, but it is not " \
1474                                                 "allowed in this context")));
1475
1476         /* Build a tuple descriptor for our result type */
1477         if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
1478                 elog(ERROR, "return type must be a row type");
1479
1480         per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
1481         oldcontext = MemoryContextSwitchTo(per_query_ctx);
1482
1483         tupstore = tuplestore_begin_heap(true, false, work_mem);
1484         rsinfo->returnMode = SFRM_Materialize;
1485         rsinfo->setResult = tupstore;
1486         rsinfo->setDesc = tupdesc;
1487
1488         MemoryContextSwitchTo(oldcontext);
1489
1490         /*
1491          * Get the priorities of sync standbys all in one go, to minimise lock
1492          * acquisitions and to allow us to evaluate who is the current sync
1493          * standby. This code must match the code in SyncRepReleaseWaiters().
1494          */
1495         sync_priority = palloc(sizeof(int) * max_wal_senders);
1496         LWLockAcquire(SyncRepLock, LW_SHARED);
1497         for (i = 0; i < max_wal_senders; i++)
1498         {
1499                 /* use volatile pointer to prevent code rearrangement */
1500                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1501
1502                 if (walsnd->pid != 0)
1503                 {
1504                         sync_priority[i] = walsnd->sync_standby_priority;
1505
1506                         if (walsnd->state == WALSNDSTATE_STREAMING &&
1507                                 walsnd->sync_standby_priority > 0 &&
1508                                 (priority == 0 ||
1509                                  priority > walsnd->sync_standby_priority))
1510                         {
1511                                 priority = walsnd->sync_standby_priority;
1512                                 sync_standby = i;
1513                         }
1514                 }
1515         }
1516         LWLockRelease(SyncRepLock);
1517
1518         for (i = 0; i < max_wal_senders; i++)
1519         {
1520                 /* use volatile pointer to prevent code rearrangement */
1521                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1522                 char            location[MAXFNAMELEN];
1523                 XLogRecPtr      sentPtr;
1524                 XLogRecPtr      write;
1525                 XLogRecPtr      flush;
1526                 XLogRecPtr      apply;
1527                 WalSndState state;
1528                 Datum           values[PG_STAT_GET_WAL_SENDERS_COLS];
1529                 bool            nulls[PG_STAT_GET_WAL_SENDERS_COLS];
1530
1531                 if (walsnd->pid == 0)
1532                         continue;
1533
1534                 SpinLockAcquire(&walsnd->mutex);
1535                 sentPtr = walsnd->sentPtr;
1536                 state = walsnd->state;
1537                 write = walsnd->write;
1538                 flush = walsnd->flush;
1539                 apply = walsnd->apply;
1540                 SpinLockRelease(&walsnd->mutex);
1541
1542                 memset(nulls, 0, sizeof(nulls));
1543                 values[0] = Int32GetDatum(walsnd->pid);
1544
1545                 if (!superuser())
1546                 {
1547                         /*
1548                          * Only superusers can see details. Other users only get the pid
1549                          * value to know it's a walsender, but no details.
1550                          */
1551                         MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
1552                 }
1553                 else
1554                 {
1555                         values[1] = CStringGetTextDatum(WalSndGetStateString(state));
1556
1557                         snprintf(location, sizeof(location), "%X/%X",
1558                                          (uint32) (sentPtr >> 32), (uint32) sentPtr);
1559                         values[2] = CStringGetTextDatum(location);
1560
1561                         if (write == 0)
1562                                 nulls[3] = true;
1563                         snprintf(location, sizeof(location), "%X/%X",
1564                                          (uint32) (write >> 32), (uint32) write);
1565                         values[3] = CStringGetTextDatum(location);
1566
1567                         if (flush == 0)
1568                                 nulls[4] = true;
1569                         snprintf(location, sizeof(location), "%X/%X",
1570                                          (uint32) (flush >> 32), (uint32) flush);
1571                         values[4] = CStringGetTextDatum(location);
1572
1573                         if (apply == 0)
1574                                 nulls[5] = true;
1575                         snprintf(location, sizeof(location), "%X/%X",
1576                                          (uint32) (apply >> 32), (uint32) apply);
1577                         values[5] = CStringGetTextDatum(location);
1578
1579                         values[6] = Int32GetDatum(sync_priority[i]);
1580
1581                         /*
1582                          * More easily understood version of standby state. This is purely
1583                          * informational, not different from priority.
1584                          */
1585                         if (sync_priority[i] == 0)
1586                                 values[7] = CStringGetTextDatum("async");
1587                         else if (i == sync_standby)
1588                                 values[7] = CStringGetTextDatum("sync");
1589                         else
1590                                 values[7] = CStringGetTextDatum("potential");
1591                 }
1592
1593                 tuplestore_putvalues(tupstore, tupdesc, values, nulls);
1594         }
1595         pfree(sync_priority);
1596
1597         /* clean up and return the tuplestore */
1598         tuplestore_donestoring(tupstore);
1599
1600         return (Datum) 0;
1601 }
1602
1603 static void
1604 WalSndKeepalive(char *msgbuf)
1605 {
1606         PrimaryKeepaliveMessage keepalive_message;
1607
1608         /* Construct a new message */
1609         keepalive_message.walEnd = sentPtr;
1610         keepalive_message.sendTime = GetCurrentTimestamp();
1611
1612         elog(DEBUG2, "sending replication keepalive");
1613
1614         /* Prepend with the message type and send it. */
1615         msgbuf[0] = 'k';
1616         memcpy(msgbuf + 1, &keepalive_message, sizeof(PrimaryKeepaliveMessage));
1617         pq_putmessage_noblock('d', msgbuf, sizeof(PrimaryKeepaliveMessage) + 1);
1618 }
1619
1620 /*
1621  * This isn't currently used for anything. Monitoring tools might be
1622  * interested in the future, and we'll need something like this in the
1623  * future for synchronous replication.
1624  */
1625 #ifdef NOT_USED
1626 /*
1627  * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
1628  * if none.
1629  */
1630 XLogRecPtr
1631 GetOldestWALSendPointer(void)
1632 {
1633         XLogRecPtr      oldest = {0, 0};
1634         int                     i;
1635         bool            found = false;
1636
1637         for (i = 0; i < max_wal_senders; i++)
1638         {
1639                 /* use volatile pointer to prevent code rearrangement */
1640                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1641                 XLogRecPtr      recptr;
1642
1643                 if (walsnd->pid == 0)
1644                         continue;
1645
1646                 SpinLockAcquire(&walsnd->mutex);
1647                 recptr = walsnd->sentPtr;
1648                 SpinLockRelease(&walsnd->mutex);
1649
1650                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1651                         continue;
1652
1653                 if (!found || XLByteLT(recptr, oldest))
1654                         oldest = recptr;
1655                 found = true;
1656         }
1657         return oldest;
1658 }
1659
1660 #endif