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) 2010-2010, PostgreSQL Global Development Group
  29  *
  30  * IDENTIFICATION
  31  *        $PostgreSQL: pgsql/src/backend/replication/walsender.c,v 1.32 2010/09/15 06:51:19 heikki Exp $
  32  *
  33  *-------------------------------------------------------------------------
  34  */
  35 #include "postgres.h"
  36
  37 #include <signal.h>
  38 #include <unistd.h>
  39
  40 #include "access/xlog_internal.h"
  41 #include "catalog/pg_type.h"
  42 #include "libpq/libpq.h"
  43 #include "libpq/pqformat.h"
  44 #include "libpq/pqsignal.h"
  45 #include "miscadmin.h"
  46 #include "replication/walprotocol.h"
  47 #include "replication/walsender.h"
  48 #include "storage/fd.h"
  49 #include "storage/ipc.h"
  50 #include "storage/pmsignal.h"
  51 #include "storage/proc.h"
  52 #include "tcop/tcopprot.h"
  53 #include "utils/guc.h"
  54 #include "utils/memutils.h"
  55 #include "utils/ps_status.h"
  56
  57
  58 /* Array of WalSnds in shared memory */
  59 WalSndCtlData *WalSndCtl = NULL;
  60
  61 /* My slot in the shared memory array */
  62 static WalSnd *MyWalSnd = NULL;
  63
  64 /* Array of WalSndWaiter in shared memory */
  65 static WalSndWaiter  *WalSndWaiters;
  66
  67 /* Global state */
  68 bool            am_walsender = false;           /* Am I a walsender process ? */
  69
  70 /* User-settable parameters for walsender */
  71 int                     max_wal_senders = 0;    /* the maximum number of concurrent walsenders */
  72 int                     WalSndDelay = 200;      /* max sleep time between some actions */
  73 int                     replication_timeout = 0;        /* maximum time to wait for the Ack from the standby */
  74
  75 /*
  76  * Buffer for WAL sending
  77  *
  78  * WalSndOutBuffer is a work area in which the output message is constructed.
  79  * It's used in just so we can avoid re-palloc'ing the buffer on each cycle.
  80  * It must be of size 6 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
  81  */
  82 static char        *WalSndOutBuffer;
  83 static int              WalSndOutHead;          /* head of pending output */
  84 static int              WalSndOutTail;          /* tail of pending output */
  85
  86 /*
  87  * These variables are used similarly to openLogFile/Id/Seg/Off,
  88  * but for walsender to read the XLOG.
  89  */
  90 static int      sendFile = -1;
  91 static uint32 sendId = 0;
  92 static uint32 sendSeg = 0;
  93 static uint32 sendOff = 0;
  94
  95 /*
  96  * How far have we sent WAL already? This is also advertised in
  97  * MyWalSnd->sentPtr.  (Actually, this is the next WAL location to send.)
  98  */
  99 static XLogRecPtr sentPtr = {0, 0};
 100
 101 /*
 102  * How far have we completed replication already? This is also
 103  * advertised in MyWalSnd->ackdPtr. This is not used in asynchronous
 104  * replication case.
 105  */
 106 static XLogRecPtr ackdPtr = {0, 0};
 107
 108 /* Replication mode requested by connected standby */
 109 static int      rplMode = REPLICATION_MODE_ASYNC;
 110
 111 /* Flags set by signal handlers for later service in main loop */
 112 static volatile sig_atomic_t got_SIGHUP = false;
 113 static volatile sig_atomic_t shutdown_requested = false;
 114 static volatile sig_atomic_t ready_to_stop = false;
 115
 116 /* Flag set by signal handler of backends for replication */
 117 static volatile sig_atomic_t replication_done = false;
 118
 119 /* Signal handlers */
 120 static void WalSndSigHupHandler(SIGNAL_ARGS);
 121 static void WalSndShutdownHandler(SIGNAL_ARGS);
 122 static void WalSndQuickDieHandler(SIGNAL_ARGS);
 123 static void WalSndXLogSendHandler(SIGNAL_ARGS);
 124 static void WalSndLastCycleHandler(SIGNAL_ARGS);
 125
 126 /* Prototypes for private functions */
 127 static int      WalSndLoop(void);
 128 static void InitWalSnd(void);
 129 static void WalSndHandshake(void);
 130 static void WalSndKill(int code, Datum arg);
 131 static void XLogRead(char *buf, XLogRecPtr recptr, Size nbytes);
 132 static bool XLogSend(bool *caughtup, bool *pending);
 133 static void ProcessStreamMsgs(StringInfo inMsg);
 134
 135 static void RegisterWalSndWaiter(BackendId backendId, XLogRecPtr record,
 136                                                                  Latch *latch);
 137 static void WakeupWalSndWaiters(XLogRecPtr record);
 138 static XLogRecPtr GetOldestAckdPtr(void);
 139
 140
 141 /* Main entry point for walsender process */
 142 int
 143 WalSenderMain(void)
 144 {
 145         MemoryContext walsnd_context;
 146
 147         if (RecoveryInProgress())
 148                 ereport(FATAL,
 149                                 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 150                                  errmsg("recovery is still in progress, can't accept WAL streaming connections")));
 151
 152         /* Create a per-walsender data structure in shared memory */
 153         InitWalSnd();
 154
 155         /*
 156          * Create a memory context that we will do all our work in.  We do this so
 157          * that we can reset the context during error recovery and thereby avoid
 158          * possible memory leaks.  Formerly this code just ran in
 159          * TopMemoryContext, but resetting that would be a really bad idea.
 160          *
 161          * XXX: we don't actually attempt error recovery in walsender, we just
 162          * close the connection and exit.
 163          */
 164         walsnd_context = AllocSetContextCreate(TopMemoryContext,
 165                                                                                    "Wal Sender",
 166                                                                                    ALLOCSET_DEFAULT_MINSIZE,
 167                                                                                    ALLOCSET_DEFAULT_INITSIZE,
 168                                                                                    ALLOCSET_DEFAULT_MAXSIZE);
 169         MemoryContextSwitchTo(walsnd_context);
 170
 171         /* Unblock signals (they were blocked when the postmaster forked us) */
 172         PG_SETMASK(&UnBlockSig);
 173
 174         /* Tell the standby that walsender is ready for receiving commands */
 175         ReadyForQuery(DestRemote);
 176
 177         /* Handle handshake messages before streaming */
 178         WalSndHandshake();
 179
 180         /* Initialize shared memory status */
 181         {
 182                 /* use volatile pointer to prevent code rearrangement */
 183                 volatile WalSnd *walsnd = MyWalSnd;
 184
 185                 SpinLockAcquire(&walsnd->mutex);
 186                 walsnd->sentPtr = sentPtr;
 187                 SpinLockRelease(&walsnd->mutex);
 188         }
 189
 190         /* Main loop of walsender */
 191         return WalSndLoop();
 192 }
 193
 194 /*
 195  * Execute commands from walreceiver, until we enter streaming mode.
 196  */
 197 static void
 198 WalSndHandshake(void)
 199 {
 200         StringInfoData input_message;
 201         bool            replication_started = false;
 202
 203         initStringInfo(&input_message);
 204
 205         while (!replication_started)
 206         {
 207                 int                     firstchar;
 208
 209                 /* Wait for a command to arrive */
 210                 firstchar = pq_getbyte();
 211
 212                 /*
 213                  * Emergency bailout if postmaster has died.  This is to avoid the
 214                  * necessity for manual cleanup of all postmaster children.
 215                  */
 216                 if (!PostmasterIsAlive(true))
 217                         exit(1);
 218
 219                 /*
 220                  * Check for any other interesting events that happened while we
 221                  * slept.
 222                  */
 223                 if (got_SIGHUP)
 224                 {
 225                         got_SIGHUP = false;
 226                         ProcessConfigFile(PGC_SIGHUP);
 227                 }
 228
 229                 if (firstchar != EOF)
 230                 {
 231                         /*
 232                          * Read the message contents. This is expected to be done without
 233                          * blocking because we've been able to get message type code.
 234                          */
 235                         if (pq_getmessage(&input_message, 0))
 236                                 firstchar = EOF;        /* suitable message already logged */
 237                 }
 238
 239                 /* Handle the very limited subset of commands expected in this phase */
 240                 switch (firstchar)
 241                 {
 242                         char            rplModeStr[6];
 243
 244                         case 'Q':                       /* Query message */
 245                                 {
 246                                         const char *query_string;
 247                                         XLogRecPtr      recptr;
 248
 249                                         query_string = pq_getmsgstring(&input_message);
 250                                         pq_getmsgend(&input_message);
 251
 252                                         if (strcmp(query_string, "IDENTIFY_SYSTEM") == 0)
 253                                         {
 254                                                 StringInfoData buf;
 255                                                 char            sysid[32];
 256                                                 char            tli[11];
 257
 258                                                 /*
 259                                                  * Reply with a result set with one row, two columns.
 260                                                  * First col is system ID, and second is timeline ID
 261                                                  */
 262
 263                                                 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
 264                                                                  GetSystemIdentifier());
 265                                                 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
 266
 267                                                 /* Send a RowDescription message */
 268                                                 pq_beginmessage(&buf, 'T');
 269                                                 pq_sendint(&buf, 2, 2); /* 2 fields */
 270
 271                                                 /* first field */
 272                                                 pq_sendstring(&buf, "systemid");                /* col name */
 273                                                 pq_sendint(&buf, 0, 4); /* table oid */
 274                                                 pq_sendint(&buf, 0, 2); /* attnum */
 275                                                 pq_sendint(&buf, TEXTOID, 4);   /* type oid */
 276                                                 pq_sendint(&buf, -1, 2);                /* typlen */
 277                                                 pq_sendint(&buf, 0, 4); /* typmod */
 278                                                 pq_sendint(&buf, 0, 2); /* format code */
 279
 280                                                 /* second field */
 281                                                 pq_sendstring(&buf, "timeline");                /* col name */
 282                                                 pq_sendint(&buf, 0, 4); /* table oid */
 283                                                 pq_sendint(&buf, 0, 2); /* attnum */
 284                                                 pq_sendint(&buf, INT4OID, 4);   /* type oid */
 285                                                 pq_sendint(&buf, 4, 2); /* typlen */
 286                                                 pq_sendint(&buf, 0, 4); /* typmod */
 287                                                 pq_sendint(&buf, 0, 2); /* format code */
 288                                                 pq_endmessage(&buf);
 289
 290                                                 /* Send a DataRow message */
 291                                                 pq_beginmessage(&buf, 'D');
 292                                                 pq_sendint(&buf, 2, 2); /* # of columns */
 293                                                 pq_sendint(&buf, strlen(sysid), 4);             /* col1 len */
 294                                                 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
 295                                                 pq_sendint(&buf, strlen(tli), 4);               /* col2 len */
 296                                                 pq_sendbytes(&buf, (char *) tli, strlen(tli));
 297                                                 pq_endmessage(&buf);
 298
 299                                                 /* Send CommandComplete and ReadyForQuery messages */
 300                                                 EndCommand("SELECT", DestRemote);
 301                                                 ReadyForQuery(DestRemote);
 302                                                 /* ReadyForQuery did pq_flush for us */
 303                                         }
 304                                         else if (sscanf(query_string, "START_REPLICATION %X/%X MODE %5s",
 305                                                                         &recptr.xlogid, &recptr.xrecoff, rplModeStr) == 3)
 306                                         {
 307                                                 StringInfoData buf;
 308
 309                                                 /*
 310                                                  * Check that we're logging enough information in the
 311                                                  * WAL for log-shipping.
 312                                                  *
 313                                                  * NOTE: This only checks the current value of
 314                                                  * wal_level. Even if the current setting is not
 315                                                  * 'minimal', there can be old WAL in the pg_xlog
 316                                                  * directory that was created with 'minimal'. So this
 317                                                  * is not bulletproof, the purpose is just to give a
 318                                                  * user-friendly error message that hints how to
 319                                                  * configure the system correctly.
 320                                                  */
 321                                                 if (wal_level == WAL_LEVEL_MINIMAL)
 322                                                         ereport(FATAL,
 323                                                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 324                                                                          errmsg("standby connections not allowed because wal_level=minimal")));
 325
 326                                                 /* Verify that the specified replication mode is valid */
 327                                                 {
 328                                                         const struct config_enum_entry *entry;
 329
 330                                                         for (entry = replication_mode_options; entry && entry->name; entry++)
 331                                                         {
 332                                                                 if (strcmp(rplModeStr, entry->name) == 0)
 333                                                                 {
 334                                                                         rplMode = entry->val;
 335                                                                         break;
 336                                                                 }
 337                                                         }
 338                                                         if (entry == NULL || entry->name == NULL)
 339                                                                 ereport(FATAL,
 340                                                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 341                                                                                  errmsg("invalid replication mode: %s", rplModeStr)));
 342                                                 }
 343                                                 MyWalSnd->rplMode = rplMode;
 344
 345                                                 /* Send a CopyXLogResponse message, and start streaming */
 346                                                 pq_beginmessage(&buf, 'W');
 347                                                 pq_endmessage(&buf);
 348                                                 pq_flush();
 349
 350                                                 /*
 351                                                  * Initialize positions to the received one, then the
 352                                                  * xlog records begin to be shipped from that position
 353                                                  */
 354                                                 sentPtr = ackdPtr = recptr;
 355
 356                                                 /* break out of the loop */
 357                                                 replication_started = true;
 358                                         }
 359                                         else
 360                                         {
 361                                                 ereport(FATAL,
 362                                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 363                                                                  errmsg("invalid standby query string: %s", query_string)));
 364                                         }
 365                                         break;
 366                                 }
 367
 368                         case 'X':
 369                                 /* standby is closing the connection */
 370                                 proc_exit(0);
 371
 372                         case EOF:
 373                                 /* standby disconnected unexpectedly */
 374                                 ereport(COMMERROR,
 375                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 376                                                  errmsg("unexpected EOF on standby connection")));
 377                                 proc_exit(0);
 378
 379                         default:
 380                                 ereport(FATAL,
 381                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 382                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 383                 }
 384         }
 385 }
 386
 387 /*
 388  * Process messages received from the standby.
 389  *
 390  * ereports on error.
 391  */
 392 static void
 393 ProcessStreamMsgs(StringInfo inMsg)
 394 {
 395         bool    acked = false;
 396
 397         /* Loop to process successive complete messages available */
 398         for (;;)
 399         {
 400                 unsigned char firstchar;
 401                 int                     r;
 402
 403                 r = pq_getbyte_if_available(&firstchar);
 404                 if (r < 0)
 405                 {
 406                         /* unexpected error or EOF */
 407                         ereport(COMMERROR,
 408                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 409                                          errmsg("unexpected EOF on standby connection")));
 410                         proc_exit(0);
 411                 }
 412                 if (r == 0)
 413                 {
 414                         /* no data available without blocking */
 415                         break;
 416                 }
 417
 418                 /* Handle the very limited subset of commands expected in this phase */
 419                 switch (firstchar)
 420                 {
 421                         case 'd':       /* CopyData message */
 422                         {
 423                                 unsigned char   rpltype;
 424
 425                                 /*
 426                                  * Read the message contents. This is expected to be done without
 427                                  * blocking because we've been able to get message type code.
 428                                  */
 429                                 if (pq_getmessage(inMsg, 0))
 430                                         proc_exit(0);           /* suitable message already logged */
 431
 432                                 /* Read the replication message type from CopyData message */
 433                                 rpltype = pq_getmsgbyte(inMsg);
 434                                 switch (rpltype)
 435                                 {
 436                                         case 'l':
 437                                         {
 438                                                 WalAckMessageData  *msgdata;
 439
 440                                                 msgdata = (WalAckMessageData *) pq_getmsgbytes(inMsg, sizeof(WalAckMessageData));
 441
 442                                                 /*
 443                                                  * Update local status.
 444                                                  *
 445                                                  * The ackd ptr received from standby should not
 446                                                  * go backwards.
 447                                                  */
 448                                                 if (XLByteLE(ackdPtr, msgdata->ackEnd))
 449                                                         ackdPtr = msgdata->ackEnd;
 450                                                 else
 451                                                         ereport(FATAL,
 452                                                                         (errmsg("replication completion location went back from "
 453                                                                                         "%X/%X to %X/%X",
 454                                                                                         ackdPtr.xlogid, ackdPtr.xrecoff,
 455                                                                                         msgdata->ackEnd.xlogid, msgdata->ackEnd.xrecoff)));
 456
 457                                                 acked = true;   /* also need to update shared position */
 458                                                 break;
 459                                         }
 460                                         default:
 461                                                 ereport(FATAL,
 462                                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 463                                                                  errmsg("invalid replication message type %d",
 464                                                                                 rpltype)));
 465                                 }
 466                                 break;
 467                         }
 468
 469                         /*
 470                          * 'X' means that the standby is closing down the socket.
 471                          */
 472                         case 'X':
 473                                 proc_exit(0);
 474
 475                         default:
 476                                 ereport(FATAL,
 477                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 478                                                  errmsg("invalid standby closing message type %d",
 479                                                                 firstchar)));
 480                 }
 481         }
 482
 483         if (acked)
 484         {
 485                 /* use volatile pointer to prevent code rearrangement */
 486                 volatile WalSnd *walsnd = MyWalSnd;
 487
 488                 SpinLockAcquire(&walsnd->mutex);
 489                 walsnd->ackdPtr = ackdPtr;
 490                 SpinLockRelease(&walsnd->mutex);
 491         }
 492
 493         /* Wake up the backends that this walsender had been blocking */
 494         WakeupWalSndWaiters(ackdPtr);
 495 }
 496
 497 /* Main loop of walsender process */
 498 static int
 499 WalSndLoop(void)
 500 {
 501         StringInfoData  input_message;
 502         bool            caughtup = false;
 503         bool            pending = false;
 504
 505         initStringInfo(&input_message);
 506
 507         /*
 508          * Allocate buffer that will be used for each output message.  We do this
 509          * just once to reduce palloc overhead.  The buffer must be made large
 510          * enough for maximum-sized messages.
 511          */
 512         WalSndOutBuffer = palloc(6 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
 513         WalSndOutHead = WalSndOutTail = 0;
 514
 515         /* Loop forever, unless we get an error */
 516         for (;;)
 517         {
 518                 /*
 519                  * Emergency bailout if postmaster has died.  This is to avoid the
 520                  * necessity for manual cleanup of all postmaster children.
 521                  */
 522                 if (!PostmasterIsAlive(true))
 523                         exit(1);
 524
 525                 /* Process any requests or signals received recently */
 526                 if (got_SIGHUP)
 527                 {
 528                         got_SIGHUP = false;
 529                         ProcessConfigFile(PGC_SIGHUP);
 530                 }
 531
 532                 /*
 533                  * When SIGUSR2 arrives, we send all outstanding logs up to the
 534                  * shutdown checkpoint record (i.e., the latest record) and exit.
 535                  */
 536                 if (ready_to_stop)
 537                 {
 538                         if (!XLogSend(&caughtup, &pending))
 539                                 break;
 540                         if (caughtup && !pending)
 541                                 shutdown_requested = true;
 542                 }
 543
 544                 /* Normal exit from the walsender is here */
 545                 if (shutdown_requested)
 546                 {
 547                         /* Inform the standby that XLOG streaming was done */
 548                         pq_puttextmessage('C', "COPY 0");
 549                         pq_flush();
 550
 551                         proc_exit(0);
 552                 }
 553
 554                 /*
 555                  * If we had sent all accumulated WAL in last round or could not
 556                  * flush pending WAL in output buffer because the socket was not
 557                  * writable, nap for the configured time before retrying.
 558                  */
 559                 if (caughtup || pending)
 560                 {
 561                         /*
 562                          * Even if we wrote all the WAL that was available when we started
 563                          * sending, more might have arrived while we were sending this
 564                          * batch. We had the latch set while sending, so we have not
 565                          * received any signals from that time. Let's arm the latch
 566                          * again, and after that check that we're still up-to-date.
 567                          */
 568                         ResetLatch(&MyWalSnd->latch);
 569
 570                         if (!XLogSend(&caughtup, &pending))
 571                                 break;
 572                         if ((caughtup || pending) && !got_SIGHUP && !ready_to_stop &&
 573                                 !shutdown_requested)
 574                         {
 575                                 bool            check_timeout;
 576                                 long            sleeptime;
 577                                 int                     res;
 578
 579                                 /*
 580                                  * XXX: We don't really need the periodic wakeups anymore,
 581                                  * WaitLatchOrSocket should reliably wake up as soon as
 582                                  * something interesting happens.
 583                                  */
 584
 585                                 /*
 586                                  * Check for replication timeout if it's enabled and we need to
 587                                  * wait for the socket to be writable to flush pending WAL in
 588                                  * output buffer.
 589                                  */
 590                                 check_timeout = replication_timeout > 0 && pending;
 591                                 if (check_timeout)
 592                                         sleeptime = replication_timeout;
 593                                 else
 594                                         sleeptime = WalSndDelay;
 595
 596                                 /* Sleep */
 597                                 res = WaitLatchOrSocket(&MyWalSnd->latch, MyProcPort->sock,
 598                                                                                 true, (WalSndOutTail > 0),
 599                                                                                 sleeptime * 1000L);
 600
 601                                 if (res == 0 && check_timeout)
 602                                 {
 603                                         /*
 604                                          * Since typically expiration of replication timeout means
 605                                          * communication problem, we don't send the error message
 606                                          * to the standby.
 607                                          */
 608                                         ereport(COMMERROR,
 609                                                         (errmsg("terminating walsender process due to replication timeout")));
 610                                         break;
 611                                 }
 612                         }
 613
 614                         /* Process messages received from the standby */
 615                         ProcessStreamMsgs(&input_message);
 616                 }
 617                 else
 618                 {
 619                         /* Attempt to send the log once every loop */
 620                         if (!XLogSend(&caughtup, &pending))
 621                                 break;
 622                 }
 623         }
 624
 625         /*
 626          * Get here on send failure.  Clean up and exit.
 627          *
 628          * Reset whereToSendOutput to prevent ereport from attempting to send any
 629          * more messages to the standby.
 630          */
 631         if (whereToSendOutput == DestRemote)
 632                 whereToSendOutput = DestNone;
 633
 634         proc_exit(0);
 635         return 1;                                       /* keep the compiler quiet */
 636 }
 637
 638 /* Initialize a per-walsender data structure for this walsender process */
 639 static void
 640 InitWalSnd(void)
 641 {
 642         int                     i;
 643
 644         /*
 645          * WalSndCtl should be set up already (we inherit this by fork() or
 646          * EXEC_BACKEND mechanism from the postmaster).
 647          */
 648         Assert(WalSndCtl != NULL);
 649         Assert(MyWalSnd == NULL);
 650
 651         /*
 652          * Find a free walsender slot and reserve it. If this fails, we must be
 653          * out of WalSnd structures.
 654          */
 655         for (i = 0; i < max_wal_senders; i++)
 656         {
 657                 /* use volatile pointer to prevent code rearrangement */
 658                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 659
 660                 SpinLockAcquire(&walsnd->mutex);
 661
 662                 if (walsnd->pid != 0)
 663                 {
 664                         SpinLockRelease(&walsnd->mutex);
 665                         continue;
 666                 }
 667                 else
 668                 {
 669                         /*
 670                          * Found a free slot. Reserve it for us.
 671                          */
 672                         walsnd->pid = MyProcPid;
 673                         MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
 674                         MemSet(&walsnd->ackdPtr, 0, sizeof(XLogRecPtr));
 675                         SpinLockRelease(&walsnd->mutex);
 676                         /* don't need the lock anymore */
 677                         OwnLatch((Latch *) &walsnd->latch);
 678                         MyWalSnd = (WalSnd *) walsnd;
 679
 680                         break;
 681                 }
 682         }
 683         if (MyWalSnd == NULL)
 684                 ereport(FATAL,
 685                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 686                                  errmsg("number of requested standby connections "
 687                                                 "exceeds max_wal_senders (currently %d)",
 688                                                 max_wal_senders)));
 689
 690         /* Arrange to clean up at walsender exit */
 691         on_shmem_exit(WalSndKill, 0);
 692 }
 693
 694 /* Destroy the per-walsender data structure for this walsender process */
 695 static void
 696 WalSndKill(int code, Datum arg)
 697 {
 698         Assert(MyWalSnd != NULL);
 699
 700         /* Wake up the backends that this walsender had been blocking */
 701         MyWalSnd->rplMode = REPLICATION_MODE_ASYNC;
 702         WakeupWalSndWaiters(GetOldestAckdPtr());
 703
 704         /*
 705          * Mark WalSnd struct no longer in use. Assume that no lock is required
 706          * for this.
 707          */
 708         MyWalSnd->pid = 0;
 709         DisownLatch(&MyWalSnd->latch);
 710
 711         /* WalSnd struct isn't mine anymore */
 712         MyWalSnd = NULL;
 713 }
 714
 715 /*
 716  * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
 717  *
 718  * XXX probably this should be improved to suck data directly from the
 719  * WAL buffers when possible.
 720  */
 721 static void
 722 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
 723 {
 724         XLogRecPtr      startRecPtr = recptr;
 725         char            path[MAXPGPATH];
 726         uint32          lastRemovedLog;
 727         uint32          lastRemovedSeg;
 728         uint32          log;
 729         uint32          seg;
 730
 731         while (nbytes > 0)
 732         {
 733                 uint32          startoff;
 734                 int                     segbytes;
 735                 int                     readbytes;
 736
 737                 startoff = recptr.xrecoff % XLogSegSize;
 738
 739                 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
 740                 {
 741                         /* Switch to another logfile segment */
 742                         if (sendFile >= 0)
 743                                 close(sendFile);
 744
 745                         XLByteToSeg(recptr, sendId, sendSeg);
 746                         XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
 747
 748                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
 749                         if (sendFile < 0)
 750                         {
 751                                 /*
 752                                  * If the file is not found, assume it's because the standby
 753                                  * asked for a too old WAL segment that has already been
 754                                  * removed or recycled.
 755                                  */
 756                                 if (errno == ENOENT)
 757                                 {
 758                                         char            filename[MAXFNAMELEN];
 759
 760                                         XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
 761                                         ereport(ERROR,
 762                                                         (errcode_for_file_access(),
 763                                                          errmsg("requested WAL segment %s has already been removed",
 764                                                                         filename)));
 765                                 }
 766                                 else
 767                                         ereport(ERROR,
 768                                                         (errcode_for_file_access(),
 769                                                          errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
 770                                                                         path, sendId, sendSeg)));
 771                         }
 772                         sendOff = 0;
 773                 }
 774
 775                 /* Need to seek in the file? */
 776                 if (sendOff != startoff)
 777                 {
 778                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
 779                                 ereport(ERROR,
 780                                                 (errcode_for_file_access(),
 781                                                  errmsg("could not seek in log file %u, segment %u to offset %u: %m",
 782                                                                 sendId, sendSeg, startoff)));
 783                         sendOff = startoff;
 784                 }
 785
 786                 /* How many bytes are within this segment? */
 787                 if (nbytes > (XLogSegSize - startoff))
 788                         segbytes = XLogSegSize - startoff;
 789                 else
 790                         segbytes = nbytes;
 791
 792                 readbytes = read(sendFile, buf, segbytes);
 793                 if (readbytes <= 0)
 794                         ereport(ERROR,
 795                                         (errcode_for_file_access(),
 796                         errmsg("could not read from log file %u, segment %u, offset %u, "
 797                                    "length %lu: %m",
 798                                    sendId, sendSeg, sendOff, (unsigned long) segbytes)));
 799
 800                 /* Update state for read */
 801                 XLByteAdvance(recptr, readbytes);
 802
 803                 sendOff += readbytes;
 804                 nbytes -= readbytes;
 805                 buf += readbytes;
 806         }
 807
 808         /*
 809          * After reading into the buffer, check that what we read was valid. We do
 810          * this after reading, because even though the segment was present when we
 811          * opened it, it might get recycled or removed while we read it. The
 812          * read() succeeds in that case, but the data we tried to read might
 813          * already have been overwritten with new WAL records.
 814          */
 815         XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
 816         XLByteToSeg(startRecPtr, log, seg);
 817         if (log < lastRemovedLog ||
 818                 (log == lastRemovedLog && seg <= lastRemovedSeg))
 819         {
 820                 char            filename[MAXFNAMELEN];
 821
 822                 XLogFileName(filename, ThisTimeLineID, log, seg);
 823                 ereport(ERROR,
 824                                 (errcode_for_file_access(),
 825                                  errmsg("requested WAL segment %s has already been removed",
 826                                                 filename)));
 827         }
 828 }
 829
 830 /*
 831  * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
 832  * but not yet sent to the client, and send it.
 833  *
 834  * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
 835  * *caughtup is set to false.
 836  *
 837  * If there is pending WAL in output buffer, *pending is set to true,
 838  * otherwise *pending is set to false.
 839  *
 840  * Returns true if OK, false if trouble.
 841  */
 842 static bool
 843 XLogSend(bool *caughtup, bool *pending)
 844 {
 845         XLogRecPtr      SendRqstPtr;
 846         XLogRecPtr      startptr;
 847         static XLogRecPtr       endptr;
 848         Size            nbytes;
 849         uint32          n32;
 850         int                     res;
 851         WalDataMessageHeader msghdr;
 852
 853         /* Attempt to flush pending WAL in output buffer */
 854         if (*pending)
 855         {
 856                 if (WalSndOutHead != WalSndOutTail)
 857                 {
 858                         res = pq_putbytes_if_writable(WalSndOutBuffer + WalSndOutHead,
 859                                                                                   WalSndOutTail - WalSndOutHead);
 860                         if (res == EOF)
 861                                 return false;
 862                         WalSndOutHead += res;
 863                         if (WalSndOutHead != WalSndOutTail)
 864                                 return true;
 865                 }
 866
 867                 res = pq_flush_if_writable();
 868                 if (res == EOF)
 869                         return false;
 870                 if (res == 0)
 871                         return true;
 872
 873                 goto updt;
 874         }
 875
 876         /*
 877          * Attempt to send all data that's already been written out and fsync'd to
 878          * disk.  We cannot go further than what's been written out given the
 879          * current implementation of XLogRead().  And in any case it's unsafe to
 880          * send WAL that is not securely down to disk on the master: if the master
 881          * subsequently crashes and restarts, slaves must not have applied any WAL
 882          * that gets lost on the master.
 883          */
 884         SendRqstPtr = GetFlushRecPtr();
 885
 886         /* Quick exit if nothing to do */
 887         if (XLByteLE(SendRqstPtr, sentPtr))
 888         {
 889                 *caughtup = true;
 890                 return true;
 891         }
 892
 893         /*
 894          * Figure out how much to send in one message. If there's no more than
 895          * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
 896          * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
 897          *
 898          * The rounding is not only for performance reasons. Walreceiver relies on
 899          * the fact that we never split a WAL record across two messages. Since a
 900          * long WAL record is split at page boundary into continuation records,
 901          * page boundary is always a safe cut-off point. We also assume that
 902          * SendRqstPtr never points to the middle of a WAL record.
 903          */
 904         startptr = sentPtr;
 905         if (startptr.xrecoff >= XLogFileSize)
 906         {
 907                 /*
 908                  * crossing a logid boundary, skip the non-existent last log segment
 909                  * in previous logical log file.
 910                  */
 911                 startptr.xlogid += 1;
 912                 startptr.xrecoff = 0;
 913         }
 914
 915         endptr = startptr;
 916         XLByteAdvance(endptr, MAX_SEND_SIZE);
 917         if (endptr.xlogid != startptr.xlogid)
 918         {
 919                 /* Don't cross a logfile boundary within one message */
 920                 Assert(endptr.xlogid == startptr.xlogid + 1);
 921                 endptr.xlogid = startptr.xlogid;
 922                 endptr.xrecoff = XLogFileSize;
 923         }
 924
 925         /* if we went beyond SendRqstPtr, back off */
 926         if (XLByteLE(SendRqstPtr, endptr))
 927         {
 928                 endptr = SendRqstPtr;
 929                 *caughtup = true;
 930         }
 931         else
 932         {
 933                 /* round down to page boundary. */
 934                 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
 935                 *caughtup = false;
 936         }
 937
 938         nbytes = endptr.xrecoff - startptr.xrecoff;
 939         Assert(nbytes <= MAX_SEND_SIZE);
 940
 941         /*
 942          * OK to read and send the slice.
 943          */
 944         WalSndOutBuffer[0] = 'd';
 945         WalSndOutBuffer[5] = 'w';
 946         WalSndOutHead = 0;
 947         WalSndOutTail = 6 + sizeof(WalDataMessageHeader) + nbytes;
 948
 949         n32 = htonl((uint32) WalSndOutTail - 1);
 950         memcpy(WalSndOutBuffer + 1, &n32, 4);
 951
 952         /*
 953          * Read the log directly into the output buffer to avoid extra memcpy
 954          * calls.
 955          */
 956         XLogRead(WalSndOutBuffer + 6 + sizeof(WalDataMessageHeader), startptr, nbytes);
 957
 958         /*
 959          * We fill the message header last so that the send timestamp is taken as
 960          * late as possible.
 961          */
 962         msghdr.dataStart = startptr;
 963         msghdr.walEnd = SendRqstPtr;
 964         msghdr.sendTime = GetCurrentTimestamp();
 965
 966         memcpy(WalSndOutBuffer + 6, &msghdr, sizeof(WalDataMessageHeader));
 967
 968         res = pq_putbytes_if_writable(WalSndOutBuffer, WalSndOutTail);
 969         if (res == EOF)
 970                 return false;
 971
 972         WalSndOutHead = res;
 973         if (WalSndOutHead != WalSndOutTail)
 974         {
 975                 *caughtup = false;
 976                 *pending = true;
 977                 return true;
 978         }
 979
 980         /* Flush pending output to the client */
 981         res = pq_flush_if_writable();
 982         if (res == EOF)
 983                 return false;
 984         if (res == 0)
 985         {
 986                 *caughtup = false;
 987                 *pending = true;
 988                 return true;
 989         }
 990
 991 updt:
 992         WalSndOutHead = WalSndOutTail = 0;
 993         *pending = false;
 994
 995         sentPtr = endptr;
 996
 997         /* Update shared memory status */
 998         {
 999                 /* use volatile pointer to prevent code rearrangement */
1000                 volatile WalSnd *walsnd = MyWalSnd;
1001
1002                 SpinLockAcquire(&walsnd->mutex);
1003                 walsnd->sentPtr = sentPtr;
1004                 SpinLockRelease(&walsnd->mutex);
1005         }
1006
1007         /* Report progress of XLOG streaming in PS display */
1008         if (update_process_title)
1009         {
1010                 char            activitymsg[50];
1011
1012                 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
1013                                  sentPtr.xlogid, sentPtr.xrecoff);
1014                 set_ps_display(activitymsg, false);
1015         }
1016
1017         return true;
1018 }
1019
1020 /* SIGHUP: set flag to re-read config file at next convenient time */
1021 static void
1022 WalSndSigHupHandler(SIGNAL_ARGS)
1023 {
1024         got_SIGHUP = true;
1025         if (MyWalSnd)
1026                 SetLatch(&MyWalSnd->latch);
1027 }
1028
1029 /* SIGTERM: set flag to shut down */
1030 static void
1031 WalSndShutdownHandler(SIGNAL_ARGS)
1032 {
1033         shutdown_requested = true;
1034         if (MyWalSnd)
1035                 SetLatch(&MyWalSnd->latch);
1036 }
1037
1038 /*
1039  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
1040  *
1041  * Some backend has bought the farm,
1042  * so we need to stop what we're doing and exit.
1043  */
1044 static void
1045 WalSndQuickDieHandler(SIGNAL_ARGS)
1046 {
1047         PG_SETMASK(&BlockSig);
1048
1049         /*
1050          * We DO NOT want to run proc_exit() callbacks -- we're here because
1051          * shared memory may be corrupted, so we don't want to try to clean up our
1052          * transaction.  Just nail the windows shut and get out of town.  Now that
1053          * there's an atexit callback to prevent third-party code from breaking
1054          * things by calling exit() directly, we have to reset the callbacks
1055          * explicitly to make this work as intended.
1056          */
1057         on_exit_reset();
1058
1059         /*
1060          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
1061          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
1062          * backend.  This is necessary precisely because we don't clean up our
1063          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
1064          * should ensure the postmaster sees this as a crash, too, but no harm in
1065          * being doubly sure.)
1066          */
1067         exit(2);
1068 }
1069
1070 /* SIGUSR1: set flag to send WAL records */
1071 static void
1072 WalSndXLogSendHandler(SIGNAL_ARGS)
1073 {
1074         latch_sigusr1_handler();
1075 }
1076
1077 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
1078 static void
1079 WalSndLastCycleHandler(SIGNAL_ARGS)
1080 {
1081         ready_to_stop = true;
1082         if (MyWalSnd)
1083                 SetLatch(&MyWalSnd->latch);
1084 }
1085
1086 /* Set up signal handlers */
1087 void
1088 WalSndSignals(void)
1089 {
1090         /* Set up signal handlers */
1091         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
1092                                                                                                  * file */
1093         pqsignal(SIGINT, SIG_IGN);      /* not used */
1094         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
1095         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
1096         pqsignal(SIGALRM, SIG_IGN);
1097         pqsignal(SIGPIPE, SIG_IGN);
1098         pqsignal(SIGUSR1, WalSndXLogSendHandler);       /* request WAL sending */
1099         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
1100                                                                                                  * shutdown */
1101
1102         /* Reset some signals that are accepted by postmaster but not here */
1103         pqsignal(SIGCHLD, SIG_DFL);
1104         pqsignal(SIGTTIN, SIG_DFL);
1105         pqsignal(SIGTTOU, SIG_DFL);
1106         pqsignal(SIGCONT, SIG_DFL);
1107         pqsignal(SIGWINCH, SIG_DFL);
1108 }
1109
1110 /* Report shared-memory space needed by WalSndShmemInit */
1111 Size
1112 WalSndShmemSize(void)
1113 {
1114         Size            size = 0;
1115
1116         size = offsetof(WalSndCtlData, walsnds);
1117         size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
1118
1119         /*
1120          * If replication is enabled, we have a data structure called
1121          * WalSndWaiters, created in shared memory.
1122          */
1123         if (max_wal_senders > 0)
1124                 size = add_size(size, mul_size(MaxBackends, sizeof(WalSndWaiter)));
1125
1126         return size;
1127 }
1128
1129 /* Allocate and initialize walsender-related shared memory */
1130 void
1131 WalSndShmemInit(void)
1132 {
1133         bool            found;
1134         int                     i;
1135         Size            size = add_size(offsetof(WalSndCtlData, walsnds),
1136                                                                 mul_size(max_wal_senders, sizeof(WalSnd)));
1137
1138         WalSndCtl = (WalSndCtlData *)
1139                 ShmemInitStruct("Wal Sender Ctl", size, &found);
1140
1141         if (!found)
1142         {
1143                 /* First time through, so initialize */
1144                 MemSet(WalSndCtl, 0, size);
1145
1146                 for (i = 0; i < max_wal_senders; i++)
1147                 {
1148                         WalSnd     *walsnd = &WalSndCtl->walsnds[i];
1149
1150                         SpinLockInit(&walsnd->mutex);
1151                         InitSharedLatch(&walsnd->latch);
1152                 }
1153         }
1154
1155         /* Create or attach to the WalSndWaiters array too, if needed */
1156         if (max_wal_senders > 0)
1157         {
1158                 WalSndWaiters = (WalSndWaiter *)
1159                         ShmemInitStruct("WalSndWaiters",
1160                                                         mul_size(MaxBackends, sizeof(WalSndWaiter)),
1161                                                         &found);
1162                 WalSndCtl->maxWaiters = MaxBackends;
1163         }
1164 }
1165
1166 /* Wake up all walsenders */
1167 void
1168 WalSndWakeup(void)
1169 {
1170         int             i;
1171
1172         for (i = 0; i < max_wal_senders; i++)
1173                 SetLatch(&WalSndCtl->walsnds[i].latch);
1174 }
1175
1176 /*
1177  * Ensure that replication has been completed up to the given position.
1178  */
1179 void
1180 WaitXLogSend(XLogRecPtr record)
1181 {
1182         int             i;
1183         bool    mustwait = false;
1184
1185         Assert(max_wal_senders > 0);
1186
1187         for (i = 0; i < max_wal_senders; i++)
1188         {
1189                 /* use volatile pointer to prevent code rearrangement */
1190                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1191                 XLogRecPtr              recptr;
1192
1193                 /* Don't need to wait for asynchronous walsender */
1194                 if (walsnd->pid == 0 ||
1195                         walsnd->rplMode <= REPLICATION_MODE_ASYNC)
1196                         continue;
1197
1198                 SpinLockAcquire(&walsnd->mutex);
1199                 recptr = walsnd->ackdPtr;
1200                 SpinLockRelease(&walsnd->mutex);
1201
1202                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1203                         continue;
1204
1205                 /* Quick exit if already known replicated */
1206                 if (XLByteLE(record, recptr))
1207                         return;
1208
1209                 mustwait = true;
1210         }
1211
1212         /*
1213          * Don't need to wait for replication if there is no synchronous
1214          * standby
1215          */
1216         if (!mustwait)
1217                 return;
1218
1219         /*
1220          * Register myself into the wait list and sleep until replication
1221          * has been completed up to the given position and the walsender
1222          * signals me.
1223          *
1224          * If replication has been completed up to the latest position
1225          * before the registration, walsender might be unable to send the
1226          * signal immediately. We must wake up the walsender after the
1227          * registration.
1228          */
1229         ResetLatch(&MyProc->latch);
1230         RegisterWalSndWaiter(MyBackendId, record, &MyProc->latch);
1231         WalSndWakeup();
1232
1233         for (;;)
1234         {
1235                 WaitLatch(&MyProc->latch, 1000000L);
1236
1237                 /* If done already, we finish waiting */
1238                 if (replication_done)
1239                 {
1240                         replication_done = false;
1241                         return;
1242                 }
1243         }
1244 }
1245
1246 /*
1247  * Register the given backend into the wait list.
1248  */
1249 static void
1250 RegisterWalSndWaiter(BackendId backendId, XLogRecPtr record, Latch *latch)
1251 {
1252         /* use volatile pointer to prevent code rearrangement */
1253         volatile WalSndCtlData  *walsndctl = WalSndCtl;
1254         int             i;
1255         int             count = 0;
1256
1257         LWLockAcquire(WalSndWaiterLock, LW_EXCLUSIVE);
1258
1259         /* Out of slots. This should not happen. */
1260         if (walsndctl->numWaiters + 1 > walsndctl->maxWaiters)
1261                 elog(PANIC, "out of replication waiters slots");
1262
1263         /*
1264          * The given position is expected to be relatively new in the
1265          * wait list. Since the entries in the list are sorted in an
1266          * increasing order of XLogRecPtr, we can shorten the time it
1267          * takes to find an insert slot by scanning the list backwards.
1268          */
1269         for (i = walsndctl->numWaiters; i > 0; i--)
1270         {
1271                 if (XLByteLE(WalSndWaiters[i - 1].record, record))
1272                         break;
1273                 count++;
1274         }
1275
1276         /* Shuffle the list if needed */
1277         if (count > 0)
1278                 memmove(&WalSndWaiters[i + 1], &WalSndWaiters[i],
1279                                 count * sizeof(WalSndWaiter));
1280
1281         WalSndWaiters[i].backendId = backendId;
1282         WalSndWaiters[i].record = record;
1283         WalSndWaiters[i].latch = latch;
1284         walsndctl->numWaiters++;
1285
1286         LWLockRelease(WalSndWaiterLock);
1287 }
1288
1289 /*
1290  * Wake up the backends waiting until replication has been completed
1291  * up to the position older than or equal to the given one.
1292  *
1293  * Wake up all waiters if InvalidXLogRecPtr is given.
1294    */
1295 static void
1296 WakeupWalSndWaiters(XLogRecPtr record)
1297 {
1298         /* use volatile pointer to prevent code rearrangement */
1299         volatile WalSndCtlData  *walsndctl = WalSndCtl;
1300         int             i;
1301         int             count = 0;
1302         bool    all_wakeup = (record.xlogid == 0 && record.xrecoff == 0);
1303
1304         LWLockAcquire(WalSndWaiterLock, LW_EXCLUSIVE);
1305
1306         for (i = 0; i < walsndctl->numWaiters; i++)
1307         {
1308                 /* use volatile pointer to prevent code rearrangement */
1309                 volatile WalSndWaiter  *waiter = &WalSndWaiters[i];
1310
1311                 if (all_wakeup || XLByteLE(waiter->record, record))
1312                 {
1313                         SetProcLatch(waiter->latch, PROCSIG_REPLICATION_INTERRUPT,
1314                                                  waiter->backendId);
1315                         count++;
1316                 }
1317                 else
1318                 {
1319                         /*
1320                          * If the backend waiting for the Ack position newer than
1321                          * the given one is found, we don't need to search the wait
1322                          * list any more. This is because the waiters in the list
1323                          * are guaranteed to be sorted in an increasing order of
1324                          * XLogRecPtr.
1325                          */
1326                         break;
1327                 }
1328         }
1329
1330         /* If there are still some waiters, left-justify them in the list */
1331         walsndctl->numWaiters -= count;
1332         if (walsndctl->numWaiters > 0 && count > 0)
1333                 memmove(&WalSndWaiters[0], &WalSndWaiters[i],
1334                                 walsndctl->numWaiters * sizeof(WalSndWaiter));
1335
1336         LWLockRelease(WalSndWaiterLock);
1337 }
1338
1339 /*
1340  * Returns the oldest Ack position in synchronous walsenders. Or
1341  * InvalidXLogRecPtr if none.
1342  */
1343 static XLogRecPtr
1344 GetOldestAckdPtr(void)
1345 {
1346         XLogRecPtr      oldest = {0, 0};
1347         int             i;
1348         bool    found = false;
1349
1350         for (i = 0; i < max_wal_senders; i++)
1351         {
1352                 /* use volatile pointer to prevent code rearrangement */
1353                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1354                 XLogRecPtr              recptr;
1355
1356                 /*
1357                  * Ignore the Ack position that asynchronous walsender has
1358                  * since it has never received any Ack.
1359                  */
1360                 if (walsnd->pid == 0 ||
1361                         walsnd->rplMode <= REPLICATION_MODE_ASYNC)
1362                         continue;
1363
1364                 SpinLockAcquire(&walsnd->mutex);
1365                 recptr = walsnd->ackdPtr;
1366                 SpinLockRelease(&walsnd->mutex);
1367
1368                 /*
1369                  * Ignore the Ack position that the walsender which has not
1370                  * received any Ack yet has.
1371                  */
1372                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1373                         continue;
1374
1375                 if (!found || XLByteLT(recptr, oldest))
1376                         oldest = recptr;
1377                 found = true;
1378         }
1379         return oldest;
1380 }
1381
1382 /*
1383  * This is called when PROCSIG_REPLICATION_INTERRUPT is received.
1384  */
1385 void
1386 HandleReplicationInterrupt(void)
1387 {
1388         replication_done = true;
1389 }