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 int                     replication_timeout = 60 * 1000;        /* maximum time to send one WAL data message */
  78
  79 /*
  80  * These variables are used similarly to openLogFile/Id/Seg/Off,
  81  * but for walsender to read the XLOG.
  82  */
  83 static int      sendFile = -1;
  84 static uint32 sendId = 0;
  85 static uint32 sendSeg = 0;
  86 static uint32 sendOff = 0;
  87
  88 /*
  89  * How far have we sent WAL already? This is also advertised in
  90  * MyWalSnd->sentPtr.  (Actually, this is the next WAL location to send.)
  91  */
  92 static XLogRecPtr sentPtr = {0, 0};
  93
  94 /*
  95  * Buffer for processing reply messages.
  96  */
  97 static StringInfoData reply_message;
  98
  99 /*
 100  * Timestamp of the last receipt of the reply from the standby.
 101  */
 102 static TimestampTz last_reply_timestamp;
 103
 104 /* Flags set by signal handlers for later service in main loop */
 105 static volatile sig_atomic_t got_SIGHUP = false;
 106 volatile sig_atomic_t walsender_shutdown_requested = false;
 107 volatile sig_atomic_t walsender_ready_to_stop = false;
 108
 109 /* Signal handlers */
 110 static void WalSndSigHupHandler(SIGNAL_ARGS);
 111 static void WalSndShutdownHandler(SIGNAL_ARGS);
 112 static void WalSndQuickDieHandler(SIGNAL_ARGS);
 113 static void WalSndXLogSendHandler(SIGNAL_ARGS);
 114 static void WalSndLastCycleHandler(SIGNAL_ARGS);
 115
 116 /* Prototypes for private functions */
 117 static bool HandleReplicationCommand(const char *cmd_string);
 118 static int      WalSndLoop(void);
 119 static void InitWalSnd(void);
 120 static void WalSndHandshake(void);
 121 static void WalSndKill(int code, Datum arg);
 122 static void XLogSend(char *msgbuf, bool *caughtup);
 123 static void IdentifySystem(void);
 124 static void StartReplication(StartReplicationCmd * cmd);
 125 static void ProcessStandbyMessage(void);
 126 static void ProcessStandbyReplyMessage(void);
 127 static void ProcessStandbyHSFeedbackMessage(void);
 128 static void ProcessRepliesIfAny(void);
 129
 130
 131 /* Main entry point for walsender process */
 132 int
 133 WalSenderMain(void)
 134 {
 135         MemoryContext walsnd_context;
 136
 137         if (RecoveryInProgress())
 138                 ereport(FATAL,
 139                                 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 140                                  errmsg("recovery is still in progress, can't accept WAL streaming connections")));
 141
 142         /* Create a per-walsender data structure in shared memory */
 143         InitWalSnd();
 144
 145         /*
 146          * Create a memory context that we will do all our work in.  We do this so
 147          * that we can reset the context during error recovery and thereby avoid
 148          * possible memory leaks.  Formerly this code just ran in
 149          * TopMemoryContext, but resetting that would be a really bad idea.
 150          *
 151          * XXX: we don't actually attempt error recovery in walsender, we just
 152          * close the connection and exit.
 153          */
 154         walsnd_context = AllocSetContextCreate(TopMemoryContext,
 155                                                                                    "Wal Sender",
 156                                                                                    ALLOCSET_DEFAULT_MINSIZE,
 157                                                                                    ALLOCSET_DEFAULT_INITSIZE,
 158                                                                                    ALLOCSET_DEFAULT_MAXSIZE);
 159         MemoryContextSwitchTo(walsnd_context);
 160
 161         /* Set up resource owner */
 162         CurrentResourceOwner = ResourceOwnerCreate(NULL, "walsender top-level resource owner");
 163
 164         /* Unblock signals (they were blocked when the postmaster forked us) */
 165         PG_SETMASK(&UnBlockSig);
 166
 167         /* Tell the standby that walsender is ready for receiving commands */
 168         ReadyForQuery(DestRemote);
 169
 170         /* Handle handshake messages before streaming */
 171         WalSndHandshake();
 172
 173         /* Initialize shared memory status */
 174         {
 175                 /* use volatile pointer to prevent code rearrangement */
 176                 volatile WalSnd *walsnd = MyWalSnd;
 177
 178                 SpinLockAcquire(&walsnd->mutex);
 179                 walsnd->sentPtr = sentPtr;
 180                 SpinLockRelease(&walsnd->mutex);
 181         }
 182
 183         SyncRepInitConfig();
 184
 185         /* Main loop of walsender */
 186         return WalSndLoop();
 187 }
 188
 189 /*
 190  * Execute commands from walreceiver, until we enter streaming mode.
 191  */
 192 static void
 193 WalSndHandshake(void)
 194 {
 195         StringInfoData input_message;
 196         bool            replication_started = false;
 197
 198         initStringInfo(&input_message);
 199
 200         while (!replication_started)
 201         {
 202                 int                     firstchar;
 203
 204                 WalSndSetState(WALSNDSTATE_STARTUP);
 205                 set_ps_display("idle", false);
 206
 207                 /* Wait for a command to arrive */
 208                 firstchar = pq_getbyte();
 209
 210                 /*
 211                  * Emergency bailout if postmaster has died.  This is to avoid the
 212                  * necessity for manual cleanup of all postmaster children.
 213                  */
 214                 if (!PostmasterIsAlive(true))
 215                         exit(1);
 216
 217                 /*
 218                  * Check for any other interesting events that happened while we
 219                  * slept.
 220                  */
 221                 if (got_SIGHUP)
 222                 {
 223                         got_SIGHUP = false;
 224                         ProcessConfigFile(PGC_SIGHUP);
 225                 }
 226
 227                 if (firstchar != EOF)
 228                 {
 229                         /*
 230                          * Read the message contents. This is expected to be done without
 231                          * blocking because we've been able to get message type code.
 232                          */
 233                         if (pq_getmessage(&input_message, 0))
 234                                 firstchar = EOF;        /* suitable message already logged */
 235                 }
 236
 237                 /* Handle the very limited subset of commands expected in this phase */
 238                 switch (firstchar)
 239                 {
 240                         case 'Q':                       /* Query message */
 241                                 {
 242                                         const char *query_string;
 243
 244                                         query_string = pq_getmsgstring(&input_message);
 245                                         pq_getmsgend(&input_message);
 246
 247                                         if (HandleReplicationCommand(query_string))
 248                                                 replication_started = true;
 249                                 }
 250                                 break;
 251
 252                         case 'X':
 253                                 /* standby is closing the connection */
 254                                 proc_exit(0);
 255
 256                         case EOF:
 257                                 /* standby disconnected unexpectedly */
 258                                 ereport(COMMERROR,
 259                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 260                                                  errmsg("unexpected EOF on standby connection")));
 261                                 proc_exit(0);
 262
 263                         default:
 264                                 ereport(FATAL,
 265                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 266                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 267                 }
 268         }
 269 }
 270
 271 /*
 272  * IDENTIFY_SYSTEM
 273  */
 274 static void
 275 IdentifySystem(void)
 276 {
 277         StringInfoData buf;
 278         char            sysid[32];
 279         char            tli[11];
 280         char            xpos[MAXFNAMELEN];
 281         XLogRecPtr      logptr;
 282
 283         /*
 284          * Reply with a result set with one row, three columns. First col is system
 285          * ID, second is timeline ID, and third is current xlog location.
 286          */
 287
 288         snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
 289                          GetSystemIdentifier());
 290         snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
 291
 292         logptr = GetInsertRecPtr();
 293
 294         snprintf(xpos, sizeof(xpos), "%X/%X",
 295                          logptr.xlogid, logptr.xrecoff);
 296
 297         /* Send a RowDescription message */
 298         pq_beginmessage(&buf, 'T');
 299         pq_sendint(&buf, 3, 2);         /* 3 fields */
 300
 301         /* first field */
 302         pq_sendstring(&buf, "systemid");        /* col name */
 303         pq_sendint(&buf, 0, 4);         /* table oid */
 304         pq_sendint(&buf, 0, 2);         /* attnum */
 305         pq_sendint(&buf, TEXTOID, 4);           /* type oid */
 306         pq_sendint(&buf, -1, 2);        /* typlen */
 307         pq_sendint(&buf, 0, 4);         /* typmod */
 308         pq_sendint(&buf, 0, 2);         /* format code */
 309
 310         /* second field */
 311         pq_sendstring(&buf, "timeline");        /* col name */
 312         pq_sendint(&buf, 0, 4);         /* table oid */
 313         pq_sendint(&buf, 0, 2);         /* attnum */
 314         pq_sendint(&buf, INT4OID, 4);           /* type oid */
 315         pq_sendint(&buf, 4, 2);         /* typlen */
 316         pq_sendint(&buf, 0, 4);         /* typmod */
 317         pq_sendint(&buf, 0, 2);         /* format code */
 318
 319         /* third field */
 320         pq_sendstring(&buf, "xlogpos");
 321         pq_sendint(&buf, 0, 4);
 322         pq_sendint(&buf, 0, 2);
 323         pq_sendint(&buf, TEXTOID, 4);
 324         pq_sendint(&buf, -1, 2);
 325         pq_sendint(&buf, 0, 4);
 326         pq_sendint(&buf, 0, 2);
 327         pq_endmessage(&buf);
 328
 329         /* Send a DataRow message */
 330         pq_beginmessage(&buf, 'D');
 331         pq_sendint(&buf, 3, 2);         /* # of columns */
 332         pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
 333         pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
 334         pq_sendint(&buf, strlen(tli), 4);       /* col2 len */
 335         pq_sendbytes(&buf, (char *) tli, strlen(tli));
 336         pq_sendint(&buf, strlen(xpos), 4);      /* col3 len */
 337         pq_sendbytes(&buf, (char *) xpos, strlen(xpos));
 338
 339         pq_endmessage(&buf);
 340
 341         /* Send CommandComplete and ReadyForQuery messages */
 342         EndCommand("SELECT", DestRemote);
 343         ReadyForQuery(DestRemote);
 344         /* ReadyForQuery did pq_flush for us */
 345 }
 346
 347 /*
 348  * START_REPLICATION
 349  */
 350 static void
 351 StartReplication(StartReplicationCmd * cmd)
 352 {
 353         StringInfoData buf;
 354
 355         /*
 356          * Let postmaster know that we're streaming. Once we've declared us as
 357          * a WAL sender process, postmaster will let us outlive the bgwriter and
 358          * kill us last in the shutdown sequence, so we get a chance to stream
 359          * all remaining WAL at shutdown, including the shutdown checkpoint.
 360          * Note that there's no going back, and we mustn't write any WAL records
 361          * after this.
 362          */
 363         MarkPostmasterChildWalSender();
 364
 365         /*
 366          * Check that we're logging enough information in the WAL for
 367          * log-shipping.
 368          *
 369          * NOTE: This only checks the current value of wal_level. Even if the
 370          * current setting is not 'minimal', there can be old WAL in the pg_xlog
 371          * directory that was created with 'minimal'. So this is not bulletproof,
 372          * the purpose is just to give a user-friendly error message that hints
 373          * how to configure the system correctly.
 374          */
 375         if (wal_level == WAL_LEVEL_MINIMAL)
 376                 ereport(FATAL,
 377                                 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 378                 errmsg("standby connections not allowed because wal_level=minimal")));
 379
 380         /*
 381          * When we first start replication the standby will be behind the primary.
 382          * For some applications, for example, synchronous replication, it is
 383          * important to have a clear state for this initial catchup mode, so we
 384          * can trigger actions when we change streaming state later. We may stay
 385          * in this state for a long time, which is exactly why we want to be
 386          * able to monitor whether or not we are still here.
 387          */
 388         WalSndSetState(WALSNDSTATE_CATCHUP);
 389
 390         /* Send a CopyBothResponse message, and start streaming */
 391         pq_beginmessage(&buf, 'W');
 392         pq_sendbyte(&buf, 0);
 393         pq_sendint(&buf, 0, 2);
 394         pq_endmessage(&buf);
 395         pq_flush();
 396
 397         /*
 398          * Initialize position to the received one, then the xlog records begin to
 399          * be shipped from that position
 400          */
 401         sentPtr = cmd->startpoint;
 402 }
 403
 404 /*
 405  * Execute an incoming replication command.
 406  */
 407 static bool
 408 HandleReplicationCommand(const char *cmd_string)
 409 {
 410         bool            replication_started = false;
 411         int                     parse_rc;
 412         Node       *cmd_node;
 413         MemoryContext cmd_context;
 414         MemoryContext old_context;
 415
 416         elog(DEBUG1, "received replication command: %s", cmd_string);
 417
 418         cmd_context = AllocSetContextCreate(CurrentMemoryContext,
 419                                                                                 "Replication command context",
 420                                                                                 ALLOCSET_DEFAULT_MINSIZE,
 421                                                                                 ALLOCSET_DEFAULT_INITSIZE,
 422                                                                                 ALLOCSET_DEFAULT_MAXSIZE);
 423         old_context = MemoryContextSwitchTo(cmd_context);
 424
 425         replication_scanner_init(cmd_string);
 426         parse_rc = replication_yyparse();
 427         if (parse_rc != 0)
 428                 ereport(ERROR,
 429                                 (errcode(ERRCODE_SYNTAX_ERROR),
 430                                  (errmsg_internal("replication command parser returned %d",
 431                                                                   parse_rc))));
 432
 433         cmd_node = replication_parse_result;
 434
 435         switch (cmd_node->type)
 436         {
 437                 case T_IdentifySystemCmd:
 438                         IdentifySystem();
 439                         break;
 440
 441                 case T_StartReplicationCmd:
 442                         StartReplication((StartReplicationCmd *) cmd_node);
 443
 444                         /* break out of the loop */
 445                         replication_started = true;
 446                         break;
 447
 448                 case T_BaseBackupCmd:
 449                         SendBaseBackup((BaseBackupCmd *) cmd_node);
 450
 451                         /* Send CommandComplete and ReadyForQuery messages */
 452                         EndCommand("SELECT", DestRemote);
 453                         ReadyForQuery(DestRemote);
 454                         /* ReadyForQuery did pq_flush for us */
 455                         break;
 456
 457                 default:
 458                         ereport(FATAL,
 459                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 460                                          errmsg("invalid standby query string: %s", cmd_string)));
 461         }
 462
 463         /* done */
 464         MemoryContextSwitchTo(old_context);
 465         MemoryContextDelete(cmd_context);
 466
 467         return replication_started;
 468 }
 469
 470 /*
 471  * Check if the remote end has closed the connection.
 472  */
 473 static void
 474 ProcessRepliesIfAny(void)
 475 {
 476         unsigned char firstchar;
 477         int                     r;
 478         int             received = false;
 479
 480         for (;;)
 481         {
 482                 r = pq_getbyte_if_available(&firstchar);
 483                 if (r < 0)
 484                 {
 485                         /* unexpected error or EOF */
 486                         ereport(COMMERROR,
 487                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 488                                          errmsg("unexpected EOF on standby connection")));
 489                         proc_exit(0);
 490                 }
 491                 if (r == 0)
 492                 {
 493                         /* no data available without blocking */
 494                         break;
 495                 }
 496
 497                 /* Handle the very limited subset of commands expected in this phase */
 498                 switch (firstchar)
 499                 {
 500                                 /*
 501                                  * 'd' means a standby reply wrapped in a CopyData packet.
 502                                  */
 503                         case 'd':
 504                                 ProcessStandbyMessage();
 505                                 received = true;
 506                                 break;
 507
 508                                 /*
 509                                  * 'X' means that the standby is closing down the socket.
 510                                  */
 511                         case 'X':
 512                                 proc_exit(0);
 513
 514                         default:
 515                                 ereport(FATAL,
 516                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 517                                                  errmsg("invalid standby message type %d",
 518                                                                 firstchar)));
 519                 }
 520         }
 521         /*
 522          * Save the last reply timestamp if we've received at least
 523          * one reply.
 524          */
 525         if (received)
 526                 last_reply_timestamp = GetCurrentTimestamp();
 527 }
 528
 529 /*
 530  * Process a status update message received from standby.
 531  */
 532 static void
 533 ProcessStandbyMessage(void)
 534 {
 535         char msgtype;
 536
 537         resetStringInfo(&reply_message);
 538
 539         /*
 540          * Read the message contents.
 541          */
 542         if (pq_getmessage(&reply_message, 0))
 543         {
 544                 ereport(COMMERROR,
 545                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 546                                  errmsg("unexpected EOF on standby connection")));
 547                 proc_exit(0);
 548         }
 549
 550         /*
 551          * Check message type from the first byte.
 552          */
 553         msgtype = pq_getmsgbyte(&reply_message);
 554
 555         switch (msgtype)
 556         {
 557                 case 'r':
 558                         ProcessStandbyReplyMessage();
 559                         break;
 560
 561                 case 'h':
 562                         ProcessStandbyHSFeedbackMessage();
 563                         break;
 564
 565                 default:
 566                         ereport(COMMERROR,
 567                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 568                                          errmsg("unexpected message type %c", msgtype)));
 569                         proc_exit(0);
 570         }
 571 }
 572
 573 /*
 574  * Regular reply from standby advising of WAL positions on standby server.
 575  */
 576 static void
 577 ProcessStandbyReplyMessage(void)
 578 {
 579         StandbyReplyMessage     reply;
 580
 581         pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyReplyMessage));
 582
 583         elog(DEBUG2, "write %X/%X flush %X/%X apply %X/%X",
 584                  reply.write.xlogid, reply.write.xrecoff,
 585                  reply.flush.xlogid, reply.flush.xrecoff,
 586                  reply.apply.xlogid, reply.apply.xrecoff);
 587
 588         /*
 589          * Update shared state for this WalSender process
 590          * based on reply data from standby.
 591          */
 592         {
 593                 /* use volatile pointer to prevent code rearrangement */
 594                 volatile WalSnd *walsnd = MyWalSnd;
 595
 596                 SpinLockAcquire(&walsnd->mutex);
 597                 walsnd->write = reply.write;
 598                 walsnd->flush = reply.flush;
 599                 walsnd->apply = reply.apply;
 600                 SpinLockRelease(&walsnd->mutex);
 601         }
 602
 603         SyncRepReleaseWaiters();
 604 }
 605
 606 /*
 607  * Hot Standby feedback
 608  */
 609 static void
 610 ProcessStandbyHSFeedbackMessage(void)
 611 {
 612         StandbyHSFeedbackMessage        reply;
 613         TransactionId newxmin = InvalidTransactionId;
 614
 615         pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyHSFeedbackMessage));
 616
 617         elog(DEBUG2, "hot standby feedback xmin %u epoch %u",
 618                  reply.xmin,
 619                  reply.epoch);
 620
 621         /*
 622          * Update the WalSender's proc xmin to allow it to be visible
 623          * to snapshots. This will hold back the removal of dead rows
 624          * and thereby prevent the generation of cleanup conflicts
 625          * on the standby server.
 626          */
 627         if (TransactionIdIsValid(reply.xmin))
 628         {
 629                 TransactionId   nextXid;
 630                 uint32                  nextEpoch;
 631                 bool                    epochOK = false;
 632
 633                 GetNextXidAndEpoch(&nextXid, &nextEpoch);
 634
 635                 /*
 636                  * Epoch of oldestXmin should be same as standby or
 637                  * if the counter has wrapped, then one less than reply.
 638                  */
 639                 if (reply.xmin <= nextXid)
 640                 {
 641                         if (reply.epoch == nextEpoch)
 642                                 epochOK = true;
 643                 }
 644                 else
 645                 {
 646                         if (nextEpoch > 0 && reply.epoch == nextEpoch - 1)
 647                                 epochOK = true;
 648                 }
 649
 650                 /*
 651                  * Feedback from standby must not go backwards, nor should it go
 652                  * forwards further than our most recent xid.
 653                  */
 654                 if (epochOK && TransactionIdPrecedesOrEquals(reply.xmin, nextXid))
 655                 {
 656                         if (!TransactionIdIsValid(MyProc->xmin))
 657                         {
 658                                 TransactionId oldestXmin = GetOldestXmin(true, true);
 659                                 if (TransactionIdPrecedes(oldestXmin, reply.xmin))
 660                                         newxmin = reply.xmin;
 661                                 else
 662                                         newxmin = oldestXmin;
 663                         }
 664                         else
 665                         {
 666                                 if (TransactionIdPrecedes(MyProc->xmin, reply.xmin))
 667                                         newxmin = reply.xmin;
 668                                 else
 669                                         newxmin = MyProc->xmin; /* stay the same */
 670                         }
 671                 }
 672         }
 673
 674         /*
 675          * Grab the ProcArrayLock to set xmin, or invalidate for bad reply
 676          */
 677         if (MyProc->xmin != newxmin)
 678         {
 679                 LWLockAcquire(ProcArrayLock, LW_SHARED);
 680                 MyProc->xmin = newxmin;
 681                 LWLockRelease(ProcArrayLock);
 682         }
 683 }
 684
 685 /* Main loop of walsender process */
 686 static int
 687 WalSndLoop(void)
 688 {
 689         char       *output_message;
 690         bool            caughtup = false;
 691
 692         /*
 693          * Allocate buffer that will be used for each output message.  We do this
 694          * just once to reduce palloc overhead.  The buffer must be made large
 695          * enough for maximum-sized messages.
 696          */
 697         output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
 698
 699         /*
 700          * Allocate buffer that will be used for processing reply messages.  As
 701          * above, do this just once to reduce palloc overhead.
 702          */
 703         initStringInfo(&reply_message);
 704
 705         /* Initialize the last reply timestamp */
 706         last_reply_timestamp = GetCurrentTimestamp();
 707
 708         /* Loop forever, unless we get an error */
 709         for (;;)
 710         {
 711                 /*
 712                  * Emergency bailout if postmaster has died.  This is to avoid the
 713                  * necessity for manual cleanup of all postmaster children.
 714                  */
 715                 if (!PostmasterIsAlive(true))
 716                         exit(1);
 717
 718                 /* Process any requests or signals received recently */
 719                 if (got_SIGHUP)
 720                 {
 721                         got_SIGHUP = false;
 722                         ProcessConfigFile(PGC_SIGHUP);
 723                         SyncRepInitConfig();
 724                 }
 725
 726                 /* Normal exit from the walsender is here */
 727                 if (walsender_shutdown_requested)
 728                 {
 729                         /* Inform the standby that XLOG streaming was done */
 730                         pq_puttextmessage('C', "COPY 0");
 731                         pq_flush();
 732
 733                         proc_exit(0);
 734                 }
 735
 736                 /*
 737                  * If we don't have any pending data in the output buffer, try to
 738                  * send some more.
 739                  */
 740                 if (!pq_is_send_pending())
 741                 {
 742                         XLogSend(output_message, &caughtup);
 743
 744                         /*
 745                          * Even if we wrote all the WAL that was available when we started
 746                          * sending, more might have arrived while we were sending this
 747                          * batch. We had the latch set while sending, so we have not
 748                          * received any signals from that time. Let's arm the latch
 749                          * again, and after that check that we're still up-to-date.
 750                          */
 751                         if (caughtup && !pq_is_send_pending())
 752                         {
 753                                 ResetLatch(&MyWalSnd->latch);
 754
 755                                 XLogSend(output_message, &caughtup);
 756                         }
 757                 }
 758
 759                 /* Flush pending output to the client */
 760                 if (pq_flush_if_writable() != 0)
 761                         break;
 762
 763                 /*
 764                  * When SIGUSR2 arrives, we send any outstanding logs up to the
 765                  * shutdown checkpoint record (i.e., the latest record) and exit.
 766                  */
 767                 if (walsender_ready_to_stop && !pq_is_send_pending())
 768                 {
 769                         XLogSend(output_message, &caughtup);
 770                         ProcessRepliesIfAny();
 771                         if (caughtup && !pq_is_send_pending())
 772                                 walsender_shutdown_requested = true;
 773                 }
 774
 775                 if ((caughtup || pq_is_send_pending()) &&
 776                         !got_SIGHUP &&
 777                         !walsender_shutdown_requested)
 778                 {
 779                         TimestampTz     finish_time = 0;
 780                         long            sleeptime;
 781
 782                         /* Reschedule replication timeout */
 783                         if (replication_timeout > 0)
 784                         {
 785                                 long            secs;
 786                                 int             usecs;
 787
 788                                 finish_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
 789                                                                                                                   replication_timeout);
 790                                 TimestampDifference(GetCurrentTimestamp(),
 791                                                                         finish_time, &secs, &usecs);
 792                                 sleeptime = secs * 1000 + usecs / 1000;
 793                                 if (WalSndDelay < sleeptime)
 794                                         sleeptime = WalSndDelay;
 795                         }
 796                         else
 797                         {
 798                                 /*
 799                                  * XXX: Without timeout, we don't really need the periodic
 800                                  * wakeups anymore, WaitLatchOrSocket should reliably wake up
 801                                  * as soon as something interesting happens.
 802                                  */
 803                                 sleeptime = WalSndDelay;
 804                         }
 805
 806                         /* Sleep */
 807                         WaitLatchOrSocket(&MyWalSnd->latch, MyProcPort->sock,
 808                                                           true, pq_is_send_pending(),
 809                                                           sleeptime * 1000L);
 810
 811                         /* Check for replication timeout */
 812                         if (replication_timeout > 0 &&
 813                                 GetCurrentTimestamp() >= finish_time)
 814                         {
 815                                 /*
 816                                  * Since typically expiration of replication timeout means
 817                                  * communication problem, we don't send the error message
 818                                  * to the standby.
 819                                  */
 820                                 ereport(COMMERROR,
 821                                                 (errmsg("terminating walsender process due to replication timeout")));
 822                                 break;
 823                         }
 824                 }
 825
 826                 /*
 827                  * If we're in catchup state, see if its time to move to streaming.
 828                  * This is an important state change for users, since before this
 829                  * point data loss might occur if the primary dies and we need to
 830                  * failover to the standby. The state change is also important for
 831                  * synchronous replication, since commits that started to wait at
 832                  * that point might wait for some time.
 833                  */
 834                 if (MyWalSnd->state == WALSNDSTATE_CATCHUP && caughtup)
 835                 {
 836                         ereport(DEBUG1,
 837                                         (errmsg("standby \"%s\" has now caught up with primary",
 838                                                                         application_name)));
 839                         WalSndSetState(WALSNDSTATE_STREAMING);
 840                 }
 841
 842                 ProcessRepliesIfAny();
 843         }
 844
 845         /*
 846          * Get here on send failure.  Clean up and exit.
 847          *
 848          * Reset whereToSendOutput to prevent ereport from attempting to send any
 849          * more messages to the standby.
 850          */
 851         if (whereToSendOutput == DestRemote)
 852                 whereToSendOutput = DestNone;
 853
 854         proc_exit(0);
 855         return 1;                                       /* keep the compiler quiet */
 856 }
 857
 858 /* Initialize a per-walsender data structure for this walsender process */
 859 static void
 860 InitWalSnd(void)
 861 {
 862         int                     i;
 863
 864         /*
 865          * WalSndCtl should be set up already (we inherit this by fork() or
 866          * EXEC_BACKEND mechanism from the postmaster).
 867          */
 868         Assert(WalSndCtl != NULL);
 869         Assert(MyWalSnd == NULL);
 870
 871         /*
 872          * Find a free walsender slot and reserve it. If this fails, we must be
 873          * out of WalSnd structures.
 874          */
 875         for (i = 0; i < max_wal_senders; i++)
 876         {
 877                 /* use volatile pointer to prevent code rearrangement */
 878                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 879
 880                 SpinLockAcquire(&walsnd->mutex);
 881
 882                 if (walsnd->pid != 0)
 883                 {
 884                         SpinLockRelease(&walsnd->mutex);
 885                         continue;
 886                 }
 887                 else
 888                 {
 889                         /*
 890                          * Found a free slot. Reserve it for us.
 891                          */
 892                         walsnd->pid = MyProcPid;
 893                         MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
 894                         walsnd->state = WALSNDSTATE_STARTUP;
 895                         SpinLockRelease(&walsnd->mutex);
 896                         /* don't need the lock anymore */
 897                         OwnLatch((Latch *) &walsnd->latch);
 898                         MyWalSnd = (WalSnd *) walsnd;
 899
 900                         break;
 901                 }
 902         }
 903         if (MyWalSnd == NULL)
 904                 ereport(FATAL,
 905                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 906                                  errmsg("number of requested standby connections "
 907                                                 "exceeds max_wal_senders (currently %d)",
 908                                                 max_wal_senders)));
 909
 910         /* Arrange to clean up at walsender exit */
 911         on_shmem_exit(WalSndKill, 0);
 912 }
 913
 914 /* Destroy the per-walsender data structure for this walsender process */
 915 static void
 916 WalSndKill(int code, Datum arg)
 917 {
 918         Assert(MyWalSnd != NULL);
 919
 920         /*
 921          * Mark WalSnd struct no longer in use. Assume that no lock is required
 922          * for this.
 923          */
 924         MyWalSnd->pid = 0;
 925         DisownLatch(&MyWalSnd->latch);
 926
 927         /* WalSnd struct isn't mine anymore */
 928         MyWalSnd = NULL;
 929 }
 930
 931 /*
 932  * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
 933  *
 934  * XXX probably this should be improved to suck data directly from the
 935  * WAL buffers when possible.
 936  *
 937  * Will open, and keep open, one WAL segment stored in the global file
 938  * descriptor sendFile. This means if XLogRead is used once, there will
 939  * always be one descriptor left open until the process ends, but never
 940  * more than one.
 941  */
 942 void
 943 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
 944 {
 945         XLogRecPtr      startRecPtr = recptr;
 946         char            path[MAXPGPATH];
 947         uint32          lastRemovedLog;
 948         uint32          lastRemovedSeg;
 949         uint32          log;
 950         uint32          seg;
 951
 952         while (nbytes > 0)
 953         {
 954                 uint32          startoff;
 955                 int                     segbytes;
 956                 int                     readbytes;
 957
 958                 startoff = recptr.xrecoff % XLogSegSize;
 959
 960                 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
 961                 {
 962                         /* Switch to another logfile segment */
 963                         if (sendFile >= 0)
 964                                 close(sendFile);
 965
 966                         XLByteToSeg(recptr, sendId, sendSeg);
 967                         XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
 968
 969                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
 970                         if (sendFile < 0)
 971                         {
 972                                 /*
 973                                  * If the file is not found, assume it's because the standby
 974                                  * asked for a too old WAL segment that has already been
 975                                  * removed or recycled.
 976                                  */
 977                                 if (errno == ENOENT)
 978                                 {
 979                                         char            filename[MAXFNAMELEN];
 980
 981                                         XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
 982                                         ereport(ERROR,
 983                                                         (errcode_for_file_access(),
 984                                                          errmsg("requested WAL segment %s has already been removed",
 985                                                                         filename)));
 986                                 }
 987                                 else
 988                                         ereport(ERROR,
 989                                                         (errcode_for_file_access(),
 990                                                          errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
 991                                                                         path, sendId, sendSeg)));
 992                         }
 993                         sendOff = 0;
 994                 }
 995
 996                 /* Need to seek in the file? */
 997                 if (sendOff != startoff)
 998                 {
 999                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
1000                                 ereport(ERROR,
1001                                                 (errcode_for_file_access(),
1002                                                  errmsg("could not seek in log file %u, segment %u to offset %u: %m",
1003                                                                 sendId, sendSeg, startoff)));
1004                         sendOff = startoff;
1005                 }
1006
1007                 /* How many bytes are within this segment? */
1008                 if (nbytes > (XLogSegSize - startoff))
1009                         segbytes = XLogSegSize - startoff;
1010                 else
1011                         segbytes = nbytes;
1012
1013                 readbytes = read(sendFile, buf, segbytes);
1014                 if (readbytes <= 0)
1015                         ereport(ERROR,
1016                                         (errcode_for_file_access(),
1017                         errmsg("could not read from log file %u, segment %u, offset %u, "
1018                                    "length %lu: %m",
1019                                    sendId, sendSeg, sendOff, (unsigned long) segbytes)));
1020
1021                 /* Update state for read */
1022                 XLByteAdvance(recptr, readbytes);
1023
1024                 sendOff += readbytes;
1025                 nbytes -= readbytes;
1026                 buf += readbytes;
1027         }
1028
1029         /*
1030          * After reading into the buffer, check that what we read was valid. We do
1031          * this after reading, because even though the segment was present when we
1032          * opened it, it might get recycled or removed while we read it. The
1033          * read() succeeds in that case, but the data we tried to read might
1034          * already have been overwritten with new WAL records.
1035          */
1036         XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
1037         XLByteToSeg(startRecPtr, log, seg);
1038         if (log < lastRemovedLog ||
1039                 (log == lastRemovedLog && seg <= lastRemovedSeg))
1040         {
1041                 char            filename[MAXFNAMELEN];
1042
1043                 XLogFileName(filename, ThisTimeLineID, log, seg);
1044                 ereport(ERROR,
1045                                 (errcode_for_file_access(),
1046                                  errmsg("requested WAL segment %s has already been removed",
1047                                                 filename)));
1048         }
1049 }
1050
1051 /*
1052  * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
1053  * but not yet sent to the client, and buffer it in the libpq output
1054  * buffer.
1055  *
1056  * msgbuf is a work area in which the output message is constructed.  It's
1057  * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
1058  * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
1059  *
1060  * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
1061  * *caughtup is set to false.
1062
1063  */
1064 static void
1065 XLogSend(char *msgbuf, bool *caughtup)
1066 {
1067         XLogRecPtr      SendRqstPtr;
1068         XLogRecPtr      startptr;
1069         XLogRecPtr      endptr;
1070         Size            nbytes;
1071         WalDataMessageHeader msghdr;
1072
1073         /*
1074          * Attempt to send all data that's already been written out and fsync'd to
1075          * disk.  We cannot go further than what's been written out given the
1076          * current implementation of XLogRead().  And in any case it's unsafe to
1077          * send WAL that is not securely down to disk on the master: if the master
1078          * subsequently crashes and restarts, slaves must not have applied any WAL
1079          * that gets lost on the master.
1080          */
1081         SendRqstPtr = GetFlushRecPtr();
1082
1083         /* Quick exit if nothing to do */
1084         if (XLByteLE(SendRqstPtr, sentPtr))
1085         {
1086                 *caughtup = true;
1087                 return;
1088         }
1089
1090         /*
1091          * Figure out how much to send in one message. If there's no more than
1092          * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
1093          * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
1094          *
1095          * The rounding is not only for performance reasons. Walreceiver relies on
1096          * the fact that we never split a WAL record across two messages. Since a
1097          * long WAL record is split at page boundary into continuation records,
1098          * page boundary is always a safe cut-off point. We also assume that
1099          * SendRqstPtr never points to the middle of a WAL record.
1100          */
1101         startptr = sentPtr;
1102         if (startptr.xrecoff >= XLogFileSize)
1103         {
1104                 /*
1105                  * crossing a logid boundary, skip the non-existent last log segment
1106                  * in previous logical log file.
1107                  */
1108                 startptr.xlogid += 1;
1109                 startptr.xrecoff = 0;
1110         }
1111
1112         endptr = startptr;
1113         XLByteAdvance(endptr, MAX_SEND_SIZE);
1114         if (endptr.xlogid != startptr.xlogid)
1115         {
1116                 /* Don't cross a logfile boundary within one message */
1117                 Assert(endptr.xlogid == startptr.xlogid + 1);
1118                 endptr.xlogid = startptr.xlogid;
1119                 endptr.xrecoff = XLogFileSize;
1120         }
1121
1122         /* if we went beyond SendRqstPtr, back off */
1123         if (XLByteLE(SendRqstPtr, endptr))
1124         {
1125                 endptr = SendRqstPtr;
1126                 *caughtup = true;
1127         }
1128         else
1129         {
1130                 /* round down to page boundary. */
1131                 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
1132                 *caughtup = false;
1133         }
1134
1135         nbytes = endptr.xrecoff - startptr.xrecoff;
1136         Assert(nbytes <= MAX_SEND_SIZE);
1137
1138         /*
1139          * OK to read and send the slice.
1140          */
1141         msgbuf[0] = 'w';
1142
1143         /*
1144          * Read the log directly into the output buffer to avoid extra memcpy
1145          * calls.
1146          */
1147         XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
1148
1149         /*
1150          * We fill the message header last so that the send timestamp is taken as
1151          * late as possible.
1152          */
1153         msghdr.dataStart = startptr;
1154         msghdr.walEnd = SendRqstPtr;
1155         msghdr.sendTime = GetCurrentTimestamp();
1156
1157         memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
1158
1159         pq_putmessage_noblock('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
1160
1161         sentPtr = endptr;
1162
1163         /* Update shared memory status */
1164         {
1165                 /* use volatile pointer to prevent code rearrangement */
1166                 volatile WalSnd *walsnd = MyWalSnd;
1167
1168                 SpinLockAcquire(&walsnd->mutex);
1169                 walsnd->sentPtr = sentPtr;
1170                 SpinLockRelease(&walsnd->mutex);
1171         }
1172
1173         /* Report progress of XLOG streaming in PS display */
1174         if (update_process_title)
1175         {
1176                 char            activitymsg[50];
1177
1178                 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
1179                                  sentPtr.xlogid, sentPtr.xrecoff);
1180                 set_ps_display(activitymsg, false);
1181         }
1182
1183         return;
1184 }
1185
1186 /* SIGHUP: set flag to re-read config file at next convenient time */
1187 static void
1188 WalSndSigHupHandler(SIGNAL_ARGS)
1189 {
1190         got_SIGHUP = true;
1191         if (MyWalSnd)
1192                 SetLatch(&MyWalSnd->latch);
1193 }
1194
1195 /* SIGTERM: set flag to shut down */
1196 static void
1197 WalSndShutdownHandler(SIGNAL_ARGS)
1198 {
1199         walsender_shutdown_requested = true;
1200         if (MyWalSnd)
1201                 SetLatch(&MyWalSnd->latch);
1202 }
1203
1204 /*
1205  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
1206  *
1207  * Some backend has bought the farm,
1208  * so we need to stop what we're doing and exit.
1209  */
1210 static void
1211 WalSndQuickDieHandler(SIGNAL_ARGS)
1212 {
1213         PG_SETMASK(&BlockSig);
1214
1215         /*
1216          * We DO NOT want to run proc_exit() callbacks -- we're here because
1217          * shared memory may be corrupted, so we don't want to try to clean up our
1218          * transaction.  Just nail the windows shut and get out of town.  Now that
1219          * there's an atexit callback to prevent third-party code from breaking
1220          * things by calling exit() directly, we have to reset the callbacks
1221          * explicitly to make this work as intended.
1222          */
1223         on_exit_reset();
1224
1225         /*
1226          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
1227          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
1228          * backend.  This is necessary precisely because we don't clean up our
1229          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
1230          * should ensure the postmaster sees this as a crash, too, but no harm in
1231          * being doubly sure.)
1232          */
1233         exit(2);
1234 }
1235
1236 /* SIGUSR1: set flag to send WAL records */
1237 static void
1238 WalSndXLogSendHandler(SIGNAL_ARGS)
1239 {
1240         latch_sigusr1_handler();
1241 }
1242
1243 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
1244 static void
1245 WalSndLastCycleHandler(SIGNAL_ARGS)
1246 {
1247         walsender_ready_to_stop = true;
1248         if (MyWalSnd)
1249                 SetLatch(&MyWalSnd->latch);
1250 }
1251
1252 /* Set up signal handlers */
1253 void
1254 WalSndSignals(void)
1255 {
1256         /* Set up signal handlers */
1257         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
1258                                                                                                  * file */
1259         pqsignal(SIGINT, SIG_IGN);      /* not used */
1260         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
1261         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
1262         pqsignal(SIGALRM, SIG_IGN);
1263         pqsignal(SIGPIPE, SIG_IGN);
1264         pqsignal(SIGUSR1, WalSndXLogSendHandler);       /* request WAL sending */
1265         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
1266                                                                                                  * shutdown */
1267
1268         /* Reset some signals that are accepted by postmaster but not here */
1269         pqsignal(SIGCHLD, SIG_DFL);
1270         pqsignal(SIGTTIN, SIG_DFL);
1271         pqsignal(SIGTTOU, SIG_DFL);
1272         pqsignal(SIGCONT, SIG_DFL);
1273         pqsignal(SIGWINCH, SIG_DFL);
1274 }
1275
1276 /* Report shared-memory space needed by WalSndShmemInit */
1277 Size
1278 WalSndShmemSize(void)
1279 {
1280         Size            size = 0;
1281
1282         size = offsetof(WalSndCtlData, walsnds);
1283         size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
1284
1285         return size;
1286 }
1287
1288 /* Allocate and initialize walsender-related shared memory */
1289 void
1290 WalSndShmemInit(void)
1291 {
1292         bool            found;
1293         int                     i;
1294
1295         WalSndCtl = (WalSndCtlData *)
1296                 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
1297
1298         if (!found)
1299         {
1300                 /* First time through, so initialize */
1301                 MemSet(WalSndCtl, 0, WalSndShmemSize());
1302
1303                 SHMQueueInit(&(WalSndCtl->SyncRepQueue));
1304
1305                 for (i = 0; i < max_wal_senders; i++)
1306                 {
1307                         WalSnd     *walsnd = &WalSndCtl->walsnds[i];
1308
1309                         SpinLockInit(&walsnd->mutex);
1310                         InitSharedLatch(&walsnd->latch);
1311                 }
1312         }
1313 }
1314
1315 /* Wake up all walsenders */
1316 void
1317 WalSndWakeup(void)
1318 {
1319         int             i;
1320
1321         for (i = 0; i < max_wal_senders; i++)
1322                 SetLatch(&WalSndCtl->walsnds[i].latch);
1323 }
1324
1325 /* Set state for current walsender (only called in walsender) */
1326 void
1327 WalSndSetState(WalSndState state)
1328 {
1329         /* use volatile pointer to prevent code rearrangement */
1330         volatile WalSnd *walsnd = MyWalSnd;
1331
1332         Assert(am_walsender);
1333
1334         if (walsnd->state == state)
1335                 return;
1336
1337         SpinLockAcquire(&walsnd->mutex);
1338         walsnd->state = state;
1339         SpinLockRelease(&walsnd->mutex);
1340 }
1341
1342 /*
1343  * Return a string constant representing the state. This is used
1344  * in system views, and should *not* be translated.
1345  */
1346 static const char *
1347 WalSndGetStateString(WalSndState state)
1348 {
1349         switch (state)
1350         {
1351                 case WALSNDSTATE_STARTUP:
1352                         return "STARTUP";
1353                 case WALSNDSTATE_BACKUP:
1354                         return "BACKUP";
1355                 case WALSNDSTATE_CATCHUP:
1356                         return "CATCHUP";
1357                 case WALSNDSTATE_STREAMING:
1358                         return "STREAMING";
1359         }
1360         return "UNKNOWN";
1361 }
1362
1363
1364 /*
1365  * Returns activity of walsenders, including pids and xlog locations sent to
1366  * standby servers.
1367  */
1368 Datum
1369 pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
1370 {
1371 #define PG_STAT_GET_WAL_SENDERS_COLS    8
1372         ReturnSetInfo      *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
1373         TupleDesc                       tupdesc;
1374         Tuplestorestate    *tupstore;
1375         MemoryContext           per_query_ctx;
1376         MemoryContext           oldcontext;
1377         int                                     *sync_priority;
1378         int                                     priority = 0;
1379         int                                     sync_standby = -1;
1380         int                                     i;
1381
1382         /* check to see if caller supports us returning a tuplestore */
1383         if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
1384                 ereport(ERROR,
1385                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1386                                  errmsg("set-valued function called in context that cannot accept a set")));
1387         if (!(rsinfo->allowedModes & SFRM_Materialize))
1388                 ereport(ERROR,
1389                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1390                                  errmsg("materialize mode required, but it is not " \
1391                                                 "allowed in this context")));
1392
1393         /* Build a tuple descriptor for our result type */
1394         if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
1395                 elog(ERROR, "return type must be a row type");
1396
1397         per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
1398         oldcontext = MemoryContextSwitchTo(per_query_ctx);
1399
1400         tupstore = tuplestore_begin_heap(true, false, work_mem);
1401         rsinfo->returnMode = SFRM_Materialize;
1402         rsinfo->setResult = tupstore;
1403         rsinfo->setDesc = tupdesc;
1404
1405         MemoryContextSwitchTo(oldcontext);
1406
1407         /*
1408          * Get the priorities of sync standbys all in one go, to minimise
1409          * lock acquisitions and to allow us to evaluate who is the current
1410          * sync standby. This code must match the code in SyncRepReleaseWaiters().
1411          */
1412         sync_priority = palloc(sizeof(int) * max_wal_senders);
1413         LWLockAcquire(SyncRepLock, LW_SHARED);
1414         for (i = 0; i < max_wal_senders; i++)
1415         {
1416                 /* use volatile pointer to prevent code rearrangement */
1417                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1418
1419                 if (walsnd->pid != 0)
1420                 {
1421                         sync_priority[i] = walsnd->sync_standby_priority;
1422
1423                         if (walsnd->state == WALSNDSTATE_STREAMING &&
1424                                 walsnd->sync_standby_priority > 0 &&
1425                                 (priority == 0 ||
1426                                  priority > walsnd->sync_standby_priority))
1427                         {
1428                                 priority = walsnd->sync_standby_priority;
1429                                 sync_standby = i;
1430                         }
1431                 }
1432         }
1433         LWLockRelease(SyncRepLock);
1434
1435         for (i = 0; i < max_wal_senders; i++)
1436         {
1437                 /* use volatile pointer to prevent code rearrangement */
1438                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1439                 char            location[MAXFNAMELEN];
1440                 XLogRecPtr      sentPtr;
1441                 XLogRecPtr      write;
1442                 XLogRecPtr      flush;
1443                 XLogRecPtr      apply;
1444                 WalSndState     state;
1445                 Datum           values[PG_STAT_GET_WAL_SENDERS_COLS];
1446                 bool            nulls[PG_STAT_GET_WAL_SENDERS_COLS];
1447
1448                 if (walsnd->pid == 0)
1449                         continue;
1450
1451                 SpinLockAcquire(&walsnd->mutex);
1452                 sentPtr = walsnd->sentPtr;
1453                 state = walsnd->state;
1454                 write = walsnd->write;
1455                 flush = walsnd->flush;
1456                 apply = walsnd->apply;
1457                 SpinLockRelease(&walsnd->mutex);
1458
1459                 memset(nulls, 0, sizeof(nulls));
1460                 values[0] = Int32GetDatum(walsnd->pid);
1461
1462                 if (!superuser())
1463                 {
1464                         /*
1465                          * Only superusers can see details. Other users only get
1466                          * the pid value to know it's a walsender, but no details.
1467                          */
1468                         MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
1469                 }
1470                 else
1471                 {
1472                         values[1] = CStringGetTextDatum(WalSndGetStateString(state));
1473
1474                         snprintf(location, sizeof(location), "%X/%X",
1475                                          sentPtr.xlogid, sentPtr.xrecoff);
1476                         values[2] = CStringGetTextDatum(location);
1477
1478                         if (write.xlogid == 0 && write.xrecoff == 0)
1479                                 nulls[3] = true;
1480                         snprintf(location, sizeof(location), "%X/%X",
1481                                          write.xlogid, write.xrecoff);
1482                         values[3] = CStringGetTextDatum(location);
1483
1484                         if (flush.xlogid == 0 && flush.xrecoff == 0)
1485                                 nulls[4] = true;
1486                         snprintf(location, sizeof(location), "%X/%X",
1487                                         flush.xlogid, flush.xrecoff);
1488                         values[4] = CStringGetTextDatum(location);
1489
1490                         if (apply.xlogid == 0 && apply.xrecoff == 0)
1491                                 nulls[5] = true;
1492                         snprintf(location, sizeof(location), "%X/%X",
1493                                          apply.xlogid, apply.xrecoff);
1494                         values[5] = CStringGetTextDatum(location);
1495
1496                         values[6] = Int32GetDatum(sync_priority[i]);
1497
1498                         /*
1499                          * More easily understood version of standby state.
1500                          * This is purely informational, not different from priority.
1501                          */
1502                         if (sync_priority[i] == 0)
1503                                 values[7] = CStringGetTextDatum("ASYNC");
1504                         else if (i == sync_standby)
1505                                 values[7] = CStringGetTextDatum("SYNC");
1506                         else
1507                                 values[7] = CStringGetTextDatum("POTENTIAL");
1508                 }
1509
1510                 tuplestore_putvalues(tupstore, tupdesc, values, nulls);
1511         }
1512         pfree(sync_priority);
1513
1514         /* clean up and return the tuplestore */
1515         tuplestore_donestoring(tupstore);
1516
1517         return (Datum) 0;
1518 }
1519
1520 /*
1521  * This isn't currently used for anything. Monitoring tools might be
1522  * interested in the future, and we'll need something like this in the
1523  * future for synchronous replication.
1524  */
1525 #ifdef NOT_USED
1526 /*
1527  * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
1528  * if none.
1529  */
1530 XLogRecPtr
1531 GetOldestWALSendPointer(void)
1532 {
1533         XLogRecPtr      oldest = {0, 0};
1534         int                     i;
1535         bool            found = false;
1536
1537         for (i = 0; i < max_wal_senders; i++)
1538         {
1539                 /* use volatile pointer to prevent code rearrangement */
1540                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1541                 XLogRecPtr      recptr;
1542
1543                 if (walsnd->pid == 0)
1544                         continue;
1545
1546                 SpinLockAcquire(&walsnd->mutex);
1547                 recptr = walsnd->sentPtr;
1548                 SpinLockRelease(&walsnd->mutex);
1549
1550                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1551                         continue;
1552
1553                 if (!found || XLByteLT(recptr, oldest))
1554                         oldest = recptr;
1555                 found = true;
1556         }
1557         return oldest;
1558 }
1559
1560 #endif