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) 2010-2011, 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 "funcapi.h"
  41 #include "access/xlog_internal.h"
  42 #include "access/transam.h"
  43 #include "catalog/pg_type.h"
  44 #include "libpq/libpq.h"
  45 #include "libpq/pqformat.h"
  46 #include "libpq/pqsignal.h"
  47 #include "miscadmin.h"
  48 #include "replication/basebackup.h"
  49 #include "replication/replnodes.h"
  50 #include "replication/walprotocol.h"
  51 #include "replication/walsender.h"
  52 #include "storage/fd.h"
  53 #include "storage/ipc.h"
  54 #include "storage/pmsignal.h"
  55 #include "storage/proc.h"
  56 #include "storage/procarray.h"
  57 #include "tcop/tcopprot.h"
  58 #include "utils/builtins.h"
  59 #include "utils/guc.h"
  60 #include "utils/memutils.h"
  61 #include "utils/ps_status.h"
  62 #include "utils/resowner.h"
  63
  64
  65 /* Array of WalSnds in shared memory */
  66 WalSndCtlData *WalSndCtl = NULL;
  67
  68 /* My slot in the shared memory array */
  69 WalSnd *MyWalSnd = NULL;
  70
  71 /* Global state */
  72 bool            am_walsender = false;           /* Am I a walsender process ? */
  73
  74 /* User-settable parameters for walsender */
  75 int                     max_wal_senders = 0;    /* the maximum number of concurrent walsenders */
  76 int                     WalSndDelay = 1000;     /* max sleep time between some actions */
  77
  78 /*
  79  * These variables are used similarly to openLogFile/Id/Seg/Off,
  80  * but for walsender to read the XLOG.
  81  */
  82 static int      sendFile = -1;
  83 static uint32 sendId = 0;
  84 static uint32 sendSeg = 0;
  85 static uint32 sendOff = 0;
  86
  87 /*
  88  * How far have we sent WAL already? This is also advertised in
  89  * MyWalSnd->sentPtr.  (Actually, this is the next WAL location to send.)
  90  */
  91 static XLogRecPtr sentPtr = {0, 0};
  92
  93 /*
  94  * Buffer for processing reply messages.
  95  */
  96 static StringInfoData reply_message;
  97
  98 /* Flags set by signal handlers for later service in main loop */
  99 static volatile sig_atomic_t got_SIGHUP = false;
 100 volatile sig_atomic_t walsender_shutdown_requested = false;
 101 volatile sig_atomic_t walsender_ready_to_stop = false;
 102
 103 /* Signal handlers */
 104 static void WalSndSigHupHandler(SIGNAL_ARGS);
 105 static void WalSndShutdownHandler(SIGNAL_ARGS);
 106 static void WalSndQuickDieHandler(SIGNAL_ARGS);
 107 static void WalSndXLogSendHandler(SIGNAL_ARGS);
 108 static void WalSndLastCycleHandler(SIGNAL_ARGS);
 109
 110 /* Prototypes for private functions */
 111 static bool HandleReplicationCommand(const char *cmd_string);
 112 static int      WalSndLoop(void);
 113 static void InitWalSnd(void);
 114 static void WalSndHandshake(void);
 115 static void WalSndKill(int code, Datum arg);
 116 static bool XLogSend(char *msgbuf, bool *caughtup);
 117 static void IdentifySystem(void);
 118 static void StartReplication(StartReplicationCmd * cmd);
 119 static void ProcessStandbyMessage(void);
 120 static void ProcessStandbyReplyMessage(void);
 121 static void ProcessStandbyHSFeedbackMessage(void);
 122 static void ProcessRepliesIfAny(void);
 123
 124
 125 /* Main entry point for walsender process */
 126 int
 127 WalSenderMain(void)
 128 {
 129         MemoryContext walsnd_context;
 130
 131         if (RecoveryInProgress())
 132                 ereport(FATAL,
 133                                 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 134                                  errmsg("recovery is still in progress, can't accept WAL streaming connections")));
 135
 136         /* Create a per-walsender data structure in shared memory */
 137         InitWalSnd();
 138
 139         /*
 140          * Create a memory context that we will do all our work in.  We do this so
 141          * that we can reset the context during error recovery and thereby avoid
 142          * possible memory leaks.  Formerly this code just ran in
 143          * TopMemoryContext, but resetting that would be a really bad idea.
 144          *
 145          * XXX: we don't actually attempt error recovery in walsender, we just
 146          * close the connection and exit.
 147          */
 148         walsnd_context = AllocSetContextCreate(TopMemoryContext,
 149                                                                                    "Wal Sender",
 150                                                                                    ALLOCSET_DEFAULT_MINSIZE,
 151                                                                                    ALLOCSET_DEFAULT_INITSIZE,
 152                                                                                    ALLOCSET_DEFAULT_MAXSIZE);
 153         MemoryContextSwitchTo(walsnd_context);
 154
 155         /* Set up resource owner */
 156         CurrentResourceOwner = ResourceOwnerCreate(NULL, "walsender top-level resource owner");
 157
 158         /* Unblock signals (they were blocked when the postmaster forked us) */
 159         PG_SETMASK(&UnBlockSig);
 160
 161         /* Tell the standby that walsender is ready for receiving commands */
 162         ReadyForQuery(DestRemote);
 163
 164         /* Handle handshake messages before streaming */
 165         WalSndHandshake();
 166
 167         /* Initialize shared memory status */
 168         {
 169                 /* use volatile pointer to prevent code rearrangement */
 170                 volatile WalSnd *walsnd = MyWalSnd;
 171
 172                 SpinLockAcquire(&walsnd->mutex);
 173                 walsnd->sentPtr = sentPtr;
 174                 SpinLockRelease(&walsnd->mutex);
 175         }
 176
 177         SyncRepInitConfig();
 178
 179         /* Main loop of walsender */
 180         return WalSndLoop();
 181 }
 182
 183 /*
 184  * Execute commands from walreceiver, until we enter streaming mode.
 185  */
 186 static void
 187 WalSndHandshake(void)
 188 {
 189         StringInfoData input_message;
 190         bool            replication_started = false;
 191
 192         initStringInfo(&input_message);
 193
 194         while (!replication_started)
 195         {
 196                 int                     firstchar;
 197
 198                 WalSndSetState(WALSNDSTATE_STARTUP);
 199                 set_ps_display("idle", false);
 200
 201                 /* Wait for a command to arrive */
 202                 firstchar = pq_getbyte();
 203
 204                 /*
 205                  * Emergency bailout if postmaster has died.  This is to avoid the
 206                  * necessity for manual cleanup of all postmaster children.
 207                  */
 208                 if (!PostmasterIsAlive(true))
 209                         exit(1);
 210
 211                 /*
 212                  * Check for any other interesting events that happened while we
 213                  * slept.
 214                  */
 215                 if (got_SIGHUP)
 216                 {
 217                         got_SIGHUP = false;
 218                         ProcessConfigFile(PGC_SIGHUP);
 219                 }
 220
 221                 if (firstchar != EOF)
 222                 {
 223                         /*
 224                          * Read the message contents. This is expected to be done without
 225                          * blocking because we've been able to get message type code.
 226                          */
 227                         if (pq_getmessage(&input_message, 0))
 228                                 firstchar = EOF;        /* suitable message already logged */
 229                 }
 230
 231                 /* Handle the very limited subset of commands expected in this phase */
 232                 switch (firstchar)
 233                 {
 234                         case 'Q':                       /* Query message */
 235                                 {
 236                                         const char *query_string;
 237
 238                                         query_string = pq_getmsgstring(&input_message);
 239                                         pq_getmsgend(&input_message);
 240
 241                                         if (HandleReplicationCommand(query_string))
 242                                                 replication_started = true;
 243                                 }
 244                                 break;
 245
 246                         case 'X':
 247                                 /* standby is closing the connection */
 248                                 proc_exit(0);
 249
 250                         case EOF:
 251                                 /* standby disconnected unexpectedly */
 252                                 ereport(COMMERROR,
 253                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 254                                                  errmsg("unexpected EOF on standby connection")));
 255                                 proc_exit(0);
 256
 257                         default:
 258                                 ereport(FATAL,
 259                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 260                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 261                 }
 262         }
 263 }
 264
 265 /*
 266  * IDENTIFY_SYSTEM
 267  */
 268 static void
 269 IdentifySystem(void)
 270 {
 271         StringInfoData buf;
 272         char            sysid[32];
 273         char            tli[11];
 274         char            xpos[MAXFNAMELEN];
 275         XLogRecPtr      logptr;
 276
 277         /*
 278          * Reply with a result set with one row, three columns. First col is system
 279          * ID, second is timeline ID, and third is current xlog location.
 280          */
 281
 282         snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
 283                          GetSystemIdentifier());
 284         snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
 285
 286         logptr = GetInsertRecPtr();
 287
 288         snprintf(xpos, sizeof(xpos), "%X/%X",
 289                          logptr.xlogid, logptr.xrecoff);
 290
 291         /* Send a RowDescription message */
 292         pq_beginmessage(&buf, 'T');
 293         pq_sendint(&buf, 3, 2);         /* 3 fields */
 294
 295         /* first field */
 296         pq_sendstring(&buf, "systemid");        /* col name */
 297         pq_sendint(&buf, 0, 4);         /* table oid */
 298         pq_sendint(&buf, 0, 2);         /* attnum */
 299         pq_sendint(&buf, TEXTOID, 4);           /* type oid */
 300         pq_sendint(&buf, -1, 2);        /* typlen */
 301         pq_sendint(&buf, 0, 4);         /* typmod */
 302         pq_sendint(&buf, 0, 2);         /* format code */
 303
 304         /* second field */
 305         pq_sendstring(&buf, "timeline");        /* col name */
 306         pq_sendint(&buf, 0, 4);         /* table oid */
 307         pq_sendint(&buf, 0, 2);         /* attnum */
 308         pq_sendint(&buf, INT4OID, 4);           /* type oid */
 309         pq_sendint(&buf, 4, 2);         /* typlen */
 310         pq_sendint(&buf, 0, 4);         /* typmod */
 311         pq_sendint(&buf, 0, 2);         /* format code */
 312
 313         /* third field */
 314         pq_sendstring(&buf, "xlogpos");
 315         pq_sendint(&buf, 0, 4);
 316         pq_sendint(&buf, 0, 2);
 317         pq_sendint(&buf, TEXTOID, 4);
 318         pq_sendint(&buf, -1, 2);
 319         pq_sendint(&buf, 0, 4);
 320         pq_sendint(&buf, 0, 2);
 321         pq_endmessage(&buf);
 322
 323         /* Send a DataRow message */
 324         pq_beginmessage(&buf, 'D');
 325         pq_sendint(&buf, 3, 2);         /* # of columns */
 326         pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
 327         pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
 328         pq_sendint(&buf, strlen(tli), 4);       /* col2 len */
 329         pq_sendbytes(&buf, (char *) tli, strlen(tli));
 330         pq_sendint(&buf, strlen(xpos), 4);      /* col3 len */
 331         pq_sendbytes(&buf, (char *) xpos, strlen(xpos));
 332
 333         pq_endmessage(&buf);
 334
 335         /* Send CommandComplete and ReadyForQuery messages */
 336         EndCommand("SELECT", DestRemote);
 337         ReadyForQuery(DestRemote);
 338         /* ReadyForQuery did pq_flush for us */
 339 }
 340
 341 /*
 342  * START_REPLICATION
 343  */
 344 static void
 345 StartReplication(StartReplicationCmd * cmd)
 346 {
 347         StringInfoData buf;
 348
 349         /*
 350          * Let postmaster know that we're streaming. Once we've declared us as
 351          * a WAL sender process, postmaster will let us outlive the bgwriter and
 352          * kill us last in the shutdown sequence, so we get a chance to stream
 353          * all remaining WAL at shutdown, including the shutdown checkpoint.
 354          * Note that there's no going back, and we mustn't write any WAL records
 355          * after this.
 356          */
 357         MarkPostmasterChildWalSender();
 358
 359         /*
 360          * Check that we're logging enough information in the WAL for
 361          * log-shipping.
 362          *
 363          * NOTE: This only checks the current value of wal_level. Even if the
 364          * current setting is not 'minimal', there can be old WAL in the pg_xlog
 365          * directory that was created with 'minimal'. So this is not bulletproof,
 366          * the purpose is just to give a user-friendly error message that hints
 367          * how to configure the system correctly.
 368          */
 369         if (wal_level == WAL_LEVEL_MINIMAL)
 370                 ereport(FATAL,
 371                                 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 372                 errmsg("standby connections not allowed because wal_level=minimal")));
 373
 374         /*
 375          * When we first start replication the standby will be behind the primary.
 376          * For some applications, for example, synchronous replication, it is
 377          * important to have a clear state for this initial catchup mode, so we
 378          * can trigger actions when we change streaming state later. We may stay
 379          * in this state for a long time, which is exactly why we want to be
 380          * able to monitor whether or not we are still here.
 381          */
 382         WalSndSetState(WALSNDSTATE_CATCHUP);
 383
 384         /* Send a CopyBothResponse message, and start streaming */
 385         pq_beginmessage(&buf, 'W');
 386         pq_sendbyte(&buf, 0);
 387         pq_sendint(&buf, 0, 2);
 388         pq_endmessage(&buf);
 389         pq_flush();
 390
 391         /*
 392          * Initialize position to the received one, then the xlog records begin to
 393          * be shipped from that position
 394          */
 395         sentPtr = cmd->startpoint;
 396 }
 397
 398 /*
 399  * Execute an incoming replication command.
 400  */
 401 static bool
 402 HandleReplicationCommand(const char *cmd_string)
 403 {
 404         bool            replication_started = false;
 405         int                     parse_rc;
 406         Node       *cmd_node;
 407         MemoryContext cmd_context;
 408         MemoryContext old_context;
 409
 410         elog(DEBUG1, "received replication command: %s", cmd_string);
 411
 412         cmd_context = AllocSetContextCreate(CurrentMemoryContext,
 413                                                                                 "Replication command context",
 414                                                                                 ALLOCSET_DEFAULT_MINSIZE,
 415                                                                                 ALLOCSET_DEFAULT_INITSIZE,
 416                                                                                 ALLOCSET_DEFAULT_MAXSIZE);
 417         old_context = MemoryContextSwitchTo(cmd_context);
 418
 419         replication_scanner_init(cmd_string);
 420         parse_rc = replication_yyparse();
 421         if (parse_rc != 0)
 422                 ereport(ERROR,
 423                                 (errcode(ERRCODE_SYNTAX_ERROR),
 424                                  (errmsg_internal("replication command parser returned %d",
 425                                                                   parse_rc))));
 426
 427         cmd_node = replication_parse_result;
 428
 429         switch (cmd_node->type)
 430         {
 431                 case T_IdentifySystemCmd:
 432                         IdentifySystem();
 433                         break;
 434
 435                 case T_StartReplicationCmd:
 436                         StartReplication((StartReplicationCmd *) cmd_node);
 437
 438                         /* break out of the loop */
 439                         replication_started = true;
 440                         break;
 441
 442                 case T_BaseBackupCmd:
 443                         SendBaseBackup((BaseBackupCmd *) cmd_node);
 444
 445                         /* Send CommandComplete and ReadyForQuery messages */
 446                         EndCommand("SELECT", DestRemote);
 447                         ReadyForQuery(DestRemote);
 448                         /* ReadyForQuery did pq_flush for us */
 449                         break;
 450
 451                 default:
 452                         ereport(FATAL,
 453                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 454                                          errmsg("invalid standby query string: %s", cmd_string)));
 455         }
 456
 457         /* done */
 458         MemoryContextSwitchTo(old_context);
 459         MemoryContextDelete(cmd_context);
 460
 461         return replication_started;
 462 }
 463
 464 /*
 465  * Check if the remote end has closed the connection.
 466  */
 467 static void
 468 ProcessRepliesIfAny(void)
 469 {
 470         unsigned char firstchar;
 471         int                     r;
 472
 473         for (;;)
 474         {
 475                 r = pq_getbyte_if_available(&firstchar);
 476                 if (r < 0)
 477                 {
 478                         /* unexpected error or EOF */
 479                         ereport(COMMERROR,
 480                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 481                                          errmsg("unexpected EOF on standby connection")));
 482                         proc_exit(0);
 483                 }
 484                 if (r == 0)
 485                 {
 486                         /* no data available without blocking */
 487                         return;
 488                 }
 489
 490                 /* Handle the very limited subset of commands expected in this phase */
 491                 switch (firstchar)
 492                 {
 493                                 /*
 494                                  * 'd' means a standby reply wrapped in a CopyData packet.
 495                                  */
 496                         case 'd':
 497                                 ProcessStandbyMessage();
 498                                 break;
 499
 500                                 /*
 501                                  * 'X' means that the standby is closing down the socket.
 502                                  */
 503                         case 'X':
 504                                 proc_exit(0);
 505
 506                         default:
 507                                 ereport(FATAL,
 508                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 509                                                  errmsg("invalid standby closing message type %d",
 510                                                                 firstchar)));
 511                 }
 512         }
 513 }
 514
 515 /*
 516  * Process a status update message received from standby.
 517  */
 518 static void
 519 ProcessStandbyMessage(void)
 520 {
 521         char msgtype;
 522
 523         resetStringInfo(&reply_message);
 524
 525         /*
 526          * Read the message contents.
 527          */
 528         if (pq_getmessage(&reply_message, 0))
 529         {
 530                 ereport(COMMERROR,
 531                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 532                                  errmsg("unexpected EOF on standby connection")));
 533                 proc_exit(0);
 534         }
 535
 536         /*
 537          * Check message type from the first byte. At the moment, there is only
 538          * one type.
 539          */
 540         msgtype = pq_getmsgbyte(&reply_message);
 541
 542         switch (msgtype)
 543         {
 544                 case 'r':
 545                         ProcessStandbyReplyMessage();
 546                         break;
 547
 548                 case 'h':
 549                         ProcessStandbyHSFeedbackMessage();
 550                         break;
 551
 552                 default:
 553                         ereport(COMMERROR,
 554                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 555                                          errmsg("unexpected message type %c", msgtype)));
 556                         proc_exit(0);
 557         }
 558 }
 559
 560 /*
 561  * Regular reply from standby advising of WAL positions on standby server.
 562  */
 563 static void
 564 ProcessStandbyReplyMessage(void)
 565 {
 566         StandbyReplyMessage     reply;
 567
 568         pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyReplyMessage));
 569
 570         elog(DEBUG2, "write %X/%X flush %X/%X apply %X/%X",
 571                  reply.write.xlogid, reply.write.xrecoff,
 572                  reply.flush.xlogid, reply.flush.xrecoff,
 573                  reply.apply.xlogid, reply.apply.xrecoff);
 574
 575         /*
 576          * Update shared state for this WalSender process
 577          * based on reply data from standby.
 578          */
 579         {
 580                 /* use volatile pointer to prevent code rearrangement */
 581                 volatile WalSnd *walsnd = MyWalSnd;
 582
 583                 SpinLockAcquire(&walsnd->mutex);
 584                 walsnd->write = reply.write;
 585                 walsnd->flush = reply.flush;
 586                 walsnd->apply = reply.apply;
 587                 SpinLockRelease(&walsnd->mutex);
 588         }
 589
 590         SyncRepReleaseWaiters();
 591 }
 592
 593 /*
 594  * Hot Standby feedback
 595  */
 596 static void
 597 ProcessStandbyHSFeedbackMessage(void)
 598 {
 599         StandbyHSFeedbackMessage        reply;
 600         TransactionId newxmin = InvalidTransactionId;
 601
 602         pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyHSFeedbackMessage));
 603
 604         elog(DEBUG2, "hot standby feedback xmin %u epoch %u",
 605                  reply.xmin,
 606                  reply.epoch);
 607
 608         /*
 609          * Update the WalSender's proc xmin to allow it to be visible
 610          * to snapshots. This will hold back the removal of dead rows
 611          * and thereby prevent the generation of cleanup conflicts
 612          * on the standby server.
 613          */
 614         if (TransactionIdIsValid(reply.xmin))
 615         {
 616                 TransactionId   nextXid;
 617                 uint32                  nextEpoch;
 618                 bool                    epochOK = false;
 619
 620                 GetNextXidAndEpoch(&nextXid, &nextEpoch);
 621
 622                 /*
 623                  * Epoch of oldestXmin should be same as standby or
 624                  * if the counter has wrapped, then one less than reply.
 625                  */
 626                 if (reply.xmin <= nextXid)
 627                 {
 628                         if (reply.epoch == nextEpoch)
 629                                 epochOK = true;
 630                 }
 631                 else
 632                 {
 633                         if (nextEpoch > 0 && reply.epoch == nextEpoch - 1)
 634                                 epochOK = true;
 635                 }
 636
 637                 /*
 638                  * Feedback from standby must not go backwards, nor should it go
 639                  * forwards further than our most recent xid.
 640                  */
 641                 if (epochOK && TransactionIdPrecedesOrEquals(reply.xmin, nextXid))
 642                 {
 643                         if (!TransactionIdIsValid(MyProc->xmin))
 644                         {
 645                                 TransactionId oldestXmin = GetOldestXmin(true, true);
 646                                 if (TransactionIdPrecedes(oldestXmin, reply.xmin))
 647                                         newxmin = reply.xmin;
 648                                 else
 649                                         newxmin = oldestXmin;
 650                         }
 651                         else
 652                         {
 653                                 if (TransactionIdPrecedes(MyProc->xmin, reply.xmin))
 654                                         newxmin = reply.xmin;
 655                                 else
 656                                         newxmin = MyProc->xmin; /* stay the same */
 657                         }
 658                 }
 659         }
 660
 661         /*
 662          * Grab the ProcArrayLock to set xmin, or invalidate for bad reply
 663          */
 664         if (MyProc->xmin != newxmin)
 665         {
 666                 LWLockAcquire(ProcArrayLock, LW_SHARED);
 667                 MyProc->xmin = newxmin;
 668                 LWLockRelease(ProcArrayLock);
 669         }
 670 }
 671
 672 /* Main loop of walsender process */
 673 static int
 674 WalSndLoop(void)
 675 {
 676         char       *output_message;
 677         bool            caughtup = false;
 678
 679         /*
 680          * Allocate buffer that will be used for each output message.  We do this
 681          * just once to reduce palloc overhead.  The buffer must be made large
 682          * enough for maximum-sized messages.
 683          */
 684         output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
 685
 686         /*
 687          * Allocate buffer that will be used for processing reply messages.  As
 688          * above, do this just once to reduce palloc overhead.
 689          */
 690         initStringInfo(&reply_message);
 691
 692         /* Loop forever, unless we get an error */
 693         for (;;)
 694         {
 695                 /*
 696                  * Emergency bailout if postmaster has died.  This is to avoid the
 697                  * necessity for manual cleanup of all postmaster children.
 698                  */
 699                 if (!PostmasterIsAlive(true))
 700                         exit(1);
 701
 702                 /* Process any requests or signals received recently */
 703                 if (got_SIGHUP)
 704                 {
 705                         got_SIGHUP = false;
 706                         ProcessConfigFile(PGC_SIGHUP);
 707                         SyncRepInitConfig();
 708                 }
 709
 710                 /*
 711                  * When SIGUSR2 arrives, we send all outstanding logs up to the
 712                  * shutdown checkpoint record (i.e., the latest record) and exit.
 713                  */
 714                 if (walsender_ready_to_stop)
 715                 {
 716                         if (!XLogSend(output_message, &caughtup))
 717                                 break;
 718                         ProcessRepliesIfAny();
 719                         if (caughtup)
 720                                 walsender_shutdown_requested = true;
 721                 }
 722
 723                 /* Normal exit from the walsender is here */
 724                 if (walsender_shutdown_requested)
 725                 {
 726                         /* Inform the standby that XLOG streaming was done */
 727                         pq_puttextmessage('C', "COPY 0");
 728                         pq_flush();
 729
 730                         proc_exit(0);
 731                 }
 732
 733                 /*
 734                  * If we had sent all accumulated WAL in last round, nap for the
 735                  * configured time before retrying.
 736                  */
 737                 if (caughtup)
 738                 {
 739                         /*
 740                          * Even if we wrote all the WAL that was available when we started
 741                          * sending, more might have arrived while we were sending this
 742                          * batch. We had the latch set while sending, so we have not
 743                          * received any signals from that time. Let's arm the latch
 744                          * again, and after that check that we're still up-to-date.
 745                          */
 746                         ResetLatch(&MyWalSnd->latch);
 747
 748                         if (!XLogSend(output_message, &caughtup))
 749                                 break;
 750                         if (caughtup && !got_SIGHUP && !walsender_ready_to_stop && !walsender_shutdown_requested)
 751                         {
 752                                 /*
 753                                  * XXX: We don't really need the periodic wakeups anymore,
 754                                  * WaitLatchOrSocket should reliably wake up as soon as
 755                                  * something interesting happens.
 756                                  */
 757
 758                                 /* Sleep */
 759                                 WaitLatchOrSocket(&MyWalSnd->latch, MyProcPort->sock,
 760                                                                   WalSndDelay * 1000L);
 761                         }
 762                 }
 763                 else
 764                 {
 765                         /* Attempt to send the log once every loop */
 766                         if (!XLogSend(output_message, &caughtup))
 767                                 break;
 768                 }
 769
 770                 /*
 771                  * If we're in catchup state, see if its time to move to streaming.
 772                  * This is an important state change for users, since before this
 773                  * point data loss might occur if the primary dies and we need to
 774                  * failover to the standby. The state change is also important for
 775                  * synchronous replication, since commits that started to wait at
 776                  * that point might wait for some time.
 777                  */
 778                 if (MyWalSnd->state == WALSNDSTATE_CATCHUP && caughtup)
 779                 {
 780                         ereport(DEBUG1,
 781                                         (errmsg("standby \"%s\" has now caught up with primary",
 782                                                                         application_name)));
 783                         WalSndSetState(WALSNDSTATE_STREAMING);
 784                 }
 785
 786                 ProcessRepliesIfAny();
 787         }
 788
 789         /*
 790          * Get here on send failure.  Clean up and exit.
 791          *
 792          * Reset whereToSendOutput to prevent ereport from attempting to send any
 793          * more messages to the standby.
 794          */
 795         if (whereToSendOutput == DestRemote)
 796                 whereToSendOutput = DestNone;
 797
 798         proc_exit(0);
 799         return 1;                                       /* keep the compiler quiet */
 800 }
 801
 802 /* Initialize a per-walsender data structure for this walsender process */
 803 static void
 804 InitWalSnd(void)
 805 {
 806         int                     i;
 807
 808         /*
 809          * WalSndCtl should be set up already (we inherit this by fork() or
 810          * EXEC_BACKEND mechanism from the postmaster).
 811          */
 812         Assert(WalSndCtl != NULL);
 813         Assert(MyWalSnd == NULL);
 814
 815         /*
 816          * Find a free walsender slot and reserve it. If this fails, we must be
 817          * out of WalSnd structures.
 818          */
 819         for (i = 0; i < max_wal_senders; i++)
 820         {
 821                 /* use volatile pointer to prevent code rearrangement */
 822                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 823
 824                 SpinLockAcquire(&walsnd->mutex);
 825
 826                 if (walsnd->pid != 0)
 827                 {
 828                         SpinLockRelease(&walsnd->mutex);
 829                         continue;
 830                 }
 831                 else
 832                 {
 833                         /*
 834                          * Found a free slot. Reserve it for us.
 835                          */
 836                         walsnd->pid = MyProcPid;
 837                         MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
 838                         walsnd->state = WALSNDSTATE_STARTUP;
 839                         SpinLockRelease(&walsnd->mutex);
 840                         /* don't need the lock anymore */
 841                         OwnLatch((Latch *) &walsnd->latch);
 842                         MyWalSnd = (WalSnd *) walsnd;
 843
 844                         break;
 845                 }
 846         }
 847         if (MyWalSnd == NULL)
 848                 ereport(FATAL,
 849                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 850                                  errmsg("number of requested standby connections "
 851                                                 "exceeds max_wal_senders (currently %d)",
 852                                                 max_wal_senders)));
 853
 854         /* Arrange to clean up at walsender exit */
 855         on_shmem_exit(WalSndKill, 0);
 856 }
 857
 858 /* Destroy the per-walsender data structure for this walsender process */
 859 static void
 860 WalSndKill(int code, Datum arg)
 861 {
 862         Assert(MyWalSnd != NULL);
 863
 864         /*
 865          * Mark WalSnd struct no longer in use. Assume that no lock is required
 866          * for this.
 867          */
 868         MyWalSnd->pid = 0;
 869         DisownLatch(&MyWalSnd->latch);
 870
 871         /* WalSnd struct isn't mine anymore */
 872         MyWalSnd = NULL;
 873 }
 874
 875 /*
 876  * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
 877  *
 878  * XXX probably this should be improved to suck data directly from the
 879  * WAL buffers when possible.
 880  *
 881  * Will open, and keep open, one WAL segment stored in the global file
 882  * descriptor sendFile. This means if XLogRead is used once, there will
 883  * always be one descriptor left open until the process ends, but never
 884  * more than one.
 885  */
 886 void
 887 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
 888 {
 889         XLogRecPtr      startRecPtr = recptr;
 890         char            path[MAXPGPATH];
 891         uint32          lastRemovedLog;
 892         uint32          lastRemovedSeg;
 893         uint32          log;
 894         uint32          seg;
 895
 896         while (nbytes > 0)
 897         {
 898                 uint32          startoff;
 899                 int                     segbytes;
 900                 int                     readbytes;
 901
 902                 startoff = recptr.xrecoff % XLogSegSize;
 903
 904                 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
 905                 {
 906                         /* Switch to another logfile segment */
 907                         if (sendFile >= 0)
 908                                 close(sendFile);
 909
 910                         XLByteToSeg(recptr, sendId, sendSeg);
 911                         XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
 912
 913                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
 914                         if (sendFile < 0)
 915                         {
 916                                 /*
 917                                  * If the file is not found, assume it's because the standby
 918                                  * asked for a too old WAL segment that has already been
 919                                  * removed or recycled.
 920                                  */
 921                                 if (errno == ENOENT)
 922                                 {
 923                                         char            filename[MAXFNAMELEN];
 924
 925                                         XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
 926                                         ereport(ERROR,
 927                                                         (errcode_for_file_access(),
 928                                                          errmsg("requested WAL segment %s has already been removed",
 929                                                                         filename)));
 930                                 }
 931                                 else
 932                                         ereport(ERROR,
 933                                                         (errcode_for_file_access(),
 934                                                          errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
 935                                                                         path, sendId, sendSeg)));
 936                         }
 937                         sendOff = 0;
 938                 }
 939
 940                 /* Need to seek in the file? */
 941                 if (sendOff != startoff)
 942                 {
 943                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
 944                                 ereport(ERROR,
 945                                                 (errcode_for_file_access(),
 946                                                  errmsg("could not seek in log file %u, segment %u to offset %u: %m",
 947                                                                 sendId, sendSeg, startoff)));
 948                         sendOff = startoff;
 949                 }
 950
 951                 /* How many bytes are within this segment? */
 952                 if (nbytes > (XLogSegSize - startoff))
 953                         segbytes = XLogSegSize - startoff;
 954                 else
 955                         segbytes = nbytes;
 956
 957                 readbytes = read(sendFile, buf, segbytes);
 958                 if (readbytes <= 0)
 959                         ereport(ERROR,
 960                                         (errcode_for_file_access(),
 961                         errmsg("could not read from log file %u, segment %u, offset %u, "
 962                                    "length %lu: %m",
 963                                    sendId, sendSeg, sendOff, (unsigned long) segbytes)));
 964
 965                 /* Update state for read */
 966                 XLByteAdvance(recptr, readbytes);
 967
 968                 sendOff += readbytes;
 969                 nbytes -= readbytes;
 970                 buf += readbytes;
 971         }
 972
 973         /*
 974          * After reading into the buffer, check that what we read was valid. We do
 975          * this after reading, because even though the segment was present when we
 976          * opened it, it might get recycled or removed while we read it. The
 977          * read() succeeds in that case, but the data we tried to read might
 978          * already have been overwritten with new WAL records.
 979          */
 980         XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
 981         XLByteToSeg(startRecPtr, log, seg);
 982         if (log < lastRemovedLog ||
 983                 (log == lastRemovedLog && seg <= lastRemovedSeg))
 984         {
 985                 char            filename[MAXFNAMELEN];
 986
 987                 XLogFileName(filename, ThisTimeLineID, log, seg);
 988                 ereport(ERROR,
 989                                 (errcode_for_file_access(),
 990                                  errmsg("requested WAL segment %s has already been removed",
 991                                                 filename)));
 992         }
 993 }
 994
 995 /*
 996  * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
 997  * but not yet sent to the client, and send it.
 998  *
 999  * msgbuf is a work area in which the output message is constructed.  It's
1000  * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
1001  * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
1002  *
1003  * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
1004  * *caughtup is set to false.
1005  *
1006  * Returns true if OK, false if trouble.
1007  */
1008 static bool
1009 XLogSend(char *msgbuf, bool *caughtup)
1010 {
1011         XLogRecPtr      SendRqstPtr;
1012         XLogRecPtr      startptr;
1013         XLogRecPtr      endptr;
1014         Size            nbytes;
1015         WalDataMessageHeader msghdr;
1016
1017         /*
1018          * Attempt to send all data that's already been written out and fsync'd to
1019          * disk.  We cannot go further than what's been written out given the
1020          * current implementation of XLogRead().  And in any case it's unsafe to
1021          * send WAL that is not securely down to disk on the master: if the master
1022          * subsequently crashes and restarts, slaves must not have applied any WAL
1023          * that gets lost on the master.
1024          */
1025         SendRqstPtr = GetFlushRecPtr();
1026
1027         /* Quick exit if nothing to do */
1028         if (XLByteLE(SendRqstPtr, sentPtr))
1029         {
1030                 *caughtup = true;
1031                 return true;
1032         }
1033
1034         /*
1035          * Figure out how much to send in one message. If there's no more than
1036          * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
1037          * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
1038          *
1039          * The rounding is not only for performance reasons. Walreceiver relies on
1040          * the fact that we never split a WAL record across two messages. Since a
1041          * long WAL record is split at page boundary into continuation records,
1042          * page boundary is always a safe cut-off point. We also assume that
1043          * SendRqstPtr never points to the middle of a WAL record.
1044          */
1045         startptr = sentPtr;
1046         if (startptr.xrecoff >= XLogFileSize)
1047         {
1048                 /*
1049                  * crossing a logid boundary, skip the non-existent last log segment
1050                  * in previous logical log file.
1051                  */
1052                 startptr.xlogid += 1;
1053                 startptr.xrecoff = 0;
1054         }
1055
1056         endptr = startptr;
1057         XLByteAdvance(endptr, MAX_SEND_SIZE);
1058         if (endptr.xlogid != startptr.xlogid)
1059         {
1060                 /* Don't cross a logfile boundary within one message */
1061                 Assert(endptr.xlogid == startptr.xlogid + 1);
1062                 endptr.xlogid = startptr.xlogid;
1063                 endptr.xrecoff = XLogFileSize;
1064         }
1065
1066         /* if we went beyond SendRqstPtr, back off */
1067         if (XLByteLE(SendRqstPtr, endptr))
1068         {
1069                 endptr = SendRqstPtr;
1070                 *caughtup = true;
1071         }
1072         else
1073         {
1074                 /* round down to page boundary. */
1075                 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
1076                 *caughtup = false;
1077         }
1078
1079         nbytes = endptr.xrecoff - startptr.xrecoff;
1080         Assert(nbytes <= MAX_SEND_SIZE);
1081
1082         /*
1083          * OK to read and send the slice.
1084          */
1085         msgbuf[0] = 'w';
1086
1087         /*
1088          * Read the log directly into the output buffer to avoid extra memcpy
1089          * calls.
1090          */
1091         XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
1092
1093         /*
1094          * We fill the message header last so that the send timestamp is taken as
1095          * late as possible.
1096          */
1097         msghdr.dataStart = startptr;
1098         msghdr.walEnd = SendRqstPtr;
1099         msghdr.sendTime = GetCurrentTimestamp();
1100
1101         memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
1102
1103         pq_putmessage('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
1104
1105         /* Flush pending output to the client */
1106         if (pq_flush())
1107                 return false;
1108
1109         sentPtr = endptr;
1110
1111         /* Update shared memory status */
1112         {
1113                 /* use volatile pointer to prevent code rearrangement */
1114                 volatile WalSnd *walsnd = MyWalSnd;
1115
1116                 SpinLockAcquire(&walsnd->mutex);
1117                 walsnd->sentPtr = sentPtr;
1118                 SpinLockRelease(&walsnd->mutex);
1119         }
1120
1121         /* Report progress of XLOG streaming in PS display */
1122         if (update_process_title)
1123         {
1124                 char            activitymsg[50];
1125
1126                 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
1127                                  sentPtr.xlogid, sentPtr.xrecoff);
1128                 set_ps_display(activitymsg, false);
1129         }
1130
1131         return true;
1132 }
1133
1134 /* SIGHUP: set flag to re-read config file at next convenient time */
1135 static void
1136 WalSndSigHupHandler(SIGNAL_ARGS)
1137 {
1138         got_SIGHUP = true;
1139         if (MyWalSnd)
1140                 SetLatch(&MyWalSnd->latch);
1141 }
1142
1143 /* SIGTERM: set flag to shut down */
1144 static void
1145 WalSndShutdownHandler(SIGNAL_ARGS)
1146 {
1147         walsender_shutdown_requested = true;
1148         if (MyWalSnd)
1149                 SetLatch(&MyWalSnd->latch);
1150 }
1151
1152 /*
1153  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
1154  *
1155  * Some backend has bought the farm,
1156  * so we need to stop what we're doing and exit.
1157  */
1158 static void
1159 WalSndQuickDieHandler(SIGNAL_ARGS)
1160 {
1161         PG_SETMASK(&BlockSig);
1162
1163         /*
1164          * We DO NOT want to run proc_exit() callbacks -- we're here because
1165          * shared memory may be corrupted, so we don't want to try to clean up our
1166          * transaction.  Just nail the windows shut and get out of town.  Now that
1167          * there's an atexit callback to prevent third-party code from breaking
1168          * things by calling exit() directly, we have to reset the callbacks
1169          * explicitly to make this work as intended.
1170          */
1171         on_exit_reset();
1172
1173         /*
1174          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
1175          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
1176          * backend.  This is necessary precisely because we don't clean up our
1177          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
1178          * should ensure the postmaster sees this as a crash, too, but no harm in
1179          * being doubly sure.)
1180          */
1181         exit(2);
1182 }
1183
1184 /* SIGUSR1: set flag to send WAL records */
1185 static void
1186 WalSndXLogSendHandler(SIGNAL_ARGS)
1187 {
1188         latch_sigusr1_handler();
1189 }
1190
1191 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
1192 static void
1193 WalSndLastCycleHandler(SIGNAL_ARGS)
1194 {
1195         walsender_ready_to_stop = true;
1196         if (MyWalSnd)
1197                 SetLatch(&MyWalSnd->latch);
1198 }
1199
1200 /* Set up signal handlers */
1201 void
1202 WalSndSignals(void)
1203 {
1204         /* Set up signal handlers */
1205         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
1206                                                                                                  * file */
1207         pqsignal(SIGINT, SIG_IGN);      /* not used */
1208         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
1209         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
1210         pqsignal(SIGALRM, SIG_IGN);
1211         pqsignal(SIGPIPE, SIG_IGN);
1212         pqsignal(SIGUSR1, WalSndXLogSendHandler);       /* request WAL sending */
1213         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
1214                                                                                                  * shutdown */
1215
1216         /* Reset some signals that are accepted by postmaster but not here */
1217         pqsignal(SIGCHLD, SIG_DFL);
1218         pqsignal(SIGTTIN, SIG_DFL);
1219         pqsignal(SIGTTOU, SIG_DFL);
1220         pqsignal(SIGCONT, SIG_DFL);
1221         pqsignal(SIGWINCH, SIG_DFL);
1222 }
1223
1224 /* Report shared-memory space needed by WalSndShmemInit */
1225 Size
1226 WalSndShmemSize(void)
1227 {
1228         Size            size = 0;
1229
1230         size = offsetof(WalSndCtlData, walsnds);
1231         size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
1232
1233         return size;
1234 }
1235
1236 /* Allocate and initialize walsender-related shared memory */
1237 void
1238 WalSndShmemInit(void)
1239 {
1240         bool            found;
1241         int                     i;
1242
1243         WalSndCtl = (WalSndCtlData *)
1244                 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
1245
1246         if (!found)
1247         {
1248                 /* First time through, so initialize */
1249                 MemSet(WalSndCtl, 0, WalSndShmemSize());
1250
1251                 SHMQueueInit(&(WalSndCtl->SyncRepQueue));
1252
1253                 for (i = 0; i < max_wal_senders; i++)
1254                 {
1255                         WalSnd     *walsnd = &WalSndCtl->walsnds[i];
1256
1257                         SpinLockInit(&walsnd->mutex);
1258                         InitSharedLatch(&walsnd->latch);
1259                 }
1260         }
1261 }
1262
1263 /* Wake up all walsenders */
1264 void
1265 WalSndWakeup(void)
1266 {
1267         int             i;
1268
1269         for (i = 0; i < max_wal_senders; i++)
1270                 SetLatch(&WalSndCtl->walsnds[i].latch);
1271 }
1272
1273 /* Set state for current walsender (only called in walsender) */
1274 void
1275 WalSndSetState(WalSndState state)
1276 {
1277         /* use volatile pointer to prevent code rearrangement */
1278         volatile WalSnd *walsnd = MyWalSnd;
1279
1280         Assert(am_walsender);
1281
1282         if (walsnd->state == state)
1283                 return;
1284
1285         SpinLockAcquire(&walsnd->mutex);
1286         walsnd->state = state;
1287         SpinLockRelease(&walsnd->mutex);
1288 }
1289
1290 /*
1291  * Return a string constant representing the state. This is used
1292  * in system views, and should *not* be translated.
1293  */
1294 static const char *
1295 WalSndGetStateString(WalSndState state)
1296 {
1297         switch (state)
1298         {
1299                 case WALSNDSTATE_STARTUP:
1300                         return "STARTUP";
1301                 case WALSNDSTATE_BACKUP:
1302                         return "BACKUP";
1303                 case WALSNDSTATE_CATCHUP:
1304                         return "CATCHUP";
1305                 case WALSNDSTATE_STREAMING:
1306                         return "STREAMING";
1307         }
1308         return "UNKNOWN";
1309 }
1310
1311
1312 /*
1313  * Returns activity of walsenders, including pids and xlog locations sent to
1314  * standby servers.
1315  */
1316 Datum
1317 pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
1318 {
1319 #define PG_STAT_GET_WAL_SENDERS_COLS    8
1320         ReturnSetInfo      *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
1321         TupleDesc                       tupdesc;
1322         Tuplestorestate    *tupstore;
1323         MemoryContext           per_query_ctx;
1324         MemoryContext           oldcontext;
1325         int                                     sync_priority[max_wal_senders];
1326         int                                     priority = 0;
1327         int                                     sync_standby = -1;
1328         int                                     i;
1329
1330         /* check to see if caller supports us returning a tuplestore */
1331         if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
1332                 ereport(ERROR,
1333                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1334                                  errmsg("set-valued function called in context that cannot accept a set")));
1335         if (!(rsinfo->allowedModes & SFRM_Materialize))
1336                 ereport(ERROR,
1337                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1338                                  errmsg("materialize mode required, but it is not " \
1339                                                 "allowed in this context")));
1340
1341         /* Build a tuple descriptor for our result type */
1342         if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
1343                 elog(ERROR, "return type must be a row type");
1344
1345         per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
1346         oldcontext = MemoryContextSwitchTo(per_query_ctx);
1347
1348         tupstore = tuplestore_begin_heap(true, false, work_mem);
1349         rsinfo->returnMode = SFRM_Materialize;
1350         rsinfo->setResult = tupstore;
1351         rsinfo->setDesc = tupdesc;
1352
1353         MemoryContextSwitchTo(oldcontext);
1354
1355         /*
1356          * Get the priorities of sync standbys all in one go, to minimise
1357          * lock acquisitions and to allow us to evaluate who is the current
1358          * sync standby. This code must match the code in SyncRepReleaseWaiters().
1359          */
1360         LWLockAcquire(SyncRepLock, LW_SHARED);
1361         for (i = 0; i < max_wal_senders; i++)
1362         {
1363                 /* use volatile pointer to prevent code rearrangement */
1364                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1365
1366                 if (walsnd->pid != 0)
1367                 {
1368                         sync_priority[i] = walsnd->sync_standby_priority;
1369
1370                         if (walsnd->state == WALSNDSTATE_STREAMING &&
1371                                 walsnd->sync_standby_priority > 0 &&
1372                                 (priority == 0 ||
1373                                  priority > walsnd->sync_standby_priority))
1374                         {
1375                                 priority = walsnd->sync_standby_priority;
1376                                 sync_standby = i;
1377                         }
1378                 }
1379         }
1380         LWLockRelease(SyncRepLock);
1381
1382         for (i = 0; i < max_wal_senders; i++)
1383         {
1384                 /* use volatile pointer to prevent code rearrangement */
1385                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1386                 char            location[MAXFNAMELEN];
1387                 XLogRecPtr      sentPtr;
1388                 XLogRecPtr      write;
1389                 XLogRecPtr      flush;
1390                 XLogRecPtr      apply;
1391                 WalSndState     state;
1392                 Datum           values[PG_STAT_GET_WAL_SENDERS_COLS];
1393                 bool            nulls[PG_STAT_GET_WAL_SENDERS_COLS];
1394
1395                 if (walsnd->pid == 0)
1396                         continue;
1397
1398                 SpinLockAcquire(&walsnd->mutex);
1399                 sentPtr = walsnd->sentPtr;
1400                 state = walsnd->state;
1401                 write = walsnd->write;
1402                 flush = walsnd->flush;
1403                 apply = walsnd->apply;
1404                 SpinLockRelease(&walsnd->mutex);
1405
1406                 memset(nulls, 0, sizeof(nulls));
1407                 values[0] = Int32GetDatum(walsnd->pid);
1408
1409                 if (!superuser())
1410                 {
1411                         /*
1412                          * Only superusers can see details. Other users only get
1413                          * the pid value to know it's a walsender, but no details.
1414                          */
1415                         MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
1416                 }
1417                 else
1418                 {
1419                         values[1] = CStringGetTextDatum(WalSndGetStateString(state));
1420
1421                         snprintf(location, sizeof(location), "%X/%X",
1422                                          sentPtr.xlogid, sentPtr.xrecoff);
1423                         values[2] = CStringGetTextDatum(location);
1424
1425                         if (write.xlogid == 0 && write.xrecoff == 0)
1426                                 nulls[3] = true;
1427                         snprintf(location, sizeof(location), "%X/%X",
1428                                          write.xlogid, write.xrecoff);
1429                         values[3] = CStringGetTextDatum(location);
1430
1431                         if (flush.xlogid == 0 && flush.xrecoff == 0)
1432                                 nulls[4] = true;
1433                         snprintf(location, sizeof(location), "%X/%X",
1434                                         flush.xlogid, flush.xrecoff);
1435                         values[4] = CStringGetTextDatum(location);
1436
1437                         if (apply.xlogid == 0 && apply.xrecoff == 0)
1438                                 nulls[5] = true;
1439                         snprintf(location, sizeof(location), "%X/%X",
1440                                          apply.xlogid, apply.xrecoff);
1441                         values[5] = CStringGetTextDatum(location);
1442
1443                         values[6] = Int32GetDatum(sync_priority[i]);
1444
1445                         /*
1446                          * More easily understood version of standby state.
1447                          * This is purely informational, not different from priority.
1448                          */
1449                         if (sync_priority[i] == 0)
1450                                 values[7] = CStringGetTextDatum("ASYNC");
1451                         else if (i == sync_standby)
1452                                 values[7] = CStringGetTextDatum("SYNC");
1453                         else
1454                                 values[7] = CStringGetTextDatum("POTENTIAL");
1455                 }
1456
1457                 tuplestore_putvalues(tupstore, tupdesc, values, nulls);
1458         }
1459
1460         /* clean up and return the tuplestore */
1461         tuplestore_donestoring(tupstore);
1462
1463         return (Datum) 0;
1464 }
1465
1466 /*
1467  * This isn't currently used for anything. Monitoring tools might be
1468  * interested in the future, and we'll need something like this in the
1469  * future for synchronous replication.
1470  */
1471 #ifdef NOT_USED
1472 /*
1473  * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
1474  * if none.
1475  */
1476 XLogRecPtr
1477 GetOldestWALSendPointer(void)
1478 {
1479         XLogRecPtr      oldest = {0, 0};
1480         int                     i;
1481         bool            found = false;
1482
1483         for (i = 0; i < max_wal_senders; i++)
1484         {
1485                 /* use volatile pointer to prevent code rearrangement */
1486                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1487                 XLogRecPtr      recptr;
1488
1489                 if (walsnd->pid == 0)
1490                         continue;
1491
1492                 SpinLockAcquire(&walsnd->mutex);
1493                 recptr = walsnd->sentPtr;
1494                 SpinLockRelease(&walsnd->mutex);
1495
1496                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1497                         continue;
1498
1499                 if (!found || XLByteLT(recptr, oldest))
1500                         oldest = recptr;
1501                 found = true;
1502         }
1503         return oldest;
1504 }
1505
1506 #endif