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 "access/transam.h"
  41 #include "access/xlog_internal.h"
  42 #include "catalog/pg_type.h"
  43 #include "funcapi.h"
  44 #include "libpq/libpq.h"
  45 #include "libpq/pqformat.h"
  46 #include "libpq/pqsignal.h"
  47 #include "miscadmin.h"
  48 #include "nodes/replnodes.h"
  49 #include "replication/basebackup.h"
  50 #include "replication/walprotocol.h"
  51 #include "replication/walreceiver.h"
  52 #include "replication/walsender.h"
  53 #include "storage/fd.h"
  54 #include "storage/ipc.h"
  55 #include "storage/pmsignal.h"
  56 #include "storage/proc.h"
  57 #include "storage/procarray.h"
  58 #include "tcop/tcopprot.h"
  59 #include "utils/builtins.h"
  60 #include "utils/guc.h"
  61 #include "utils/memutils.h"
  62 #include "utils/ps_status.h"
  63 #include "utils/resowner.h"
  64
  65
  66 /* Array of WalSnds in shared memory */
  67 WalSndCtlData *WalSndCtl = NULL;
  68
  69 /* My slot in the shared memory array */
  70 WalSnd     *MyWalSnd = NULL;
  71
  72 /* Global state */
  73 bool            am_walsender = false;           /* Am I a walsender process ? */
  74 bool            am_cascading_walsender = false; /* Am I cascading WAL to another standby ? */
  75
  76 /* User-settable parameters for walsender */
  77 int                     max_wal_senders = 0;    /* the maximum number of concurrent walsenders */
  78 int                     WalSndDelay = 1000; /* max sleep time between some actions */
  79 int                     replication_timeout = 60 * 1000;        /* maximum time to send one
  80                                                                                                  * WAL data message */
  81
  82 /*
  83  * These variables are used similarly to openLogFile/Id/Seg/Off,
  84  * but for walsender to read the XLOG.
  85  */
  86 static int      sendFile = -1;
  87 static uint32 sendId = 0;
  88 static uint32 sendSeg = 0;
  89 static uint32 sendOff = 0;
  90
  91 /*
  92  * How far have we sent WAL already? This is also advertised in
  93  * MyWalSnd->sentPtr.  (Actually, this is the next WAL location to send.)
  94  */
  95 static XLogRecPtr sentPtr = {0, 0};
  96
  97 /*
  98  * Buffer for processing reply messages.
  99  */
 100 static StringInfoData reply_message;
 101
 102 /*
 103  * Timestamp of the last receipt of the reply from the standby.
 104  */
 105 static TimestampTz last_reply_timestamp;
 106
 107 /* Flags set by signal handlers for later service in main loop */
 108 static volatile sig_atomic_t got_SIGHUP = false;
 109 volatile sig_atomic_t walsender_shutdown_requested = false;
 110 volatile sig_atomic_t walsender_ready_to_stop = false;
 111
 112 /* Signal handlers */
 113 static void WalSndSigHupHandler(SIGNAL_ARGS);
 114 static void WalSndShutdownHandler(SIGNAL_ARGS);
 115 static void WalSndQuickDieHandler(SIGNAL_ARGS);
 116 static void WalSndXLogSendHandler(SIGNAL_ARGS);
 117 static void WalSndLastCycleHandler(SIGNAL_ARGS);
 118
 119 /* Prototypes for private functions */
 120 static bool HandleReplicationCommand(const char *cmd_string);
 121 static int      WalSndLoop(void);
 122 static void InitWalSnd(void);
 123 static void WalSndHandshake(void);
 124 static void WalSndKill(int code, Datum arg);
 125 static void XLogSend(char *msgbuf, bool *caughtup);
 126 static void IdentifySystem(void);
 127 static void StartReplication(StartReplicationCmd *cmd);
 128 static void ProcessStandbyMessage(void);
 129 static void ProcessStandbyReplyMessage(void);
 130 static void ProcessStandbyHSFeedbackMessage(void);
 131 static void ProcessRepliesIfAny(void);
 132
 133
 134 /* Main entry point for walsender process */
 135 int
 136 WalSenderMain(void)
 137 {
 138         MemoryContext walsnd_context;
 139
 140         am_cascading_walsender = RecoveryInProgress();
 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         /*
 168          * Use the recovery target timeline ID during recovery
 169          */
 170         if (am_cascading_walsender)
 171                 ThisTimeLineID = GetRecoveryTargetTLI();
 172
 173         /* Tell the standby that walsender is ready for receiving commands */
 174         ReadyForQuery(DestRemote);
 175
 176         /* Handle handshake messages before streaming */
 177         WalSndHandshake();
 178
 179         /* Initialize shared memory status */
 180         {
 181                 /* use volatile pointer to prevent code rearrangement */
 182                 volatile WalSnd *walsnd = MyWalSnd;
 183
 184                 SpinLockAcquire(&walsnd->mutex);
 185                 walsnd->sentPtr = sentPtr;
 186                 SpinLockRelease(&walsnd->mutex);
 187         }
 188
 189         SyncRepInitConfig();
 190
 191         /* Main loop of walsender */
 192         return WalSndLoop();
 193 }
 194
 195 /*
 196  * Execute commands from walreceiver, until we enter streaming mode.
 197  */
 198 static void
 199 WalSndHandshake(void)
 200 {
 201         StringInfoData input_message;
 202         bool            replication_started = false;
 203
 204         initStringInfo(&input_message);
 205
 206         while (!replication_started)
 207         {
 208                 int                     firstchar;
 209
 210                 WalSndSetState(WALSNDSTATE_STARTUP);
 211                 set_ps_display("idle", false);
 212
 213                 /* Wait for a command to arrive */
 214                 firstchar = pq_getbyte();
 215
 216                 /*
 217                  * Emergency bailout if postmaster has died.  This is to avoid the
 218                  * necessity for manual cleanup of all postmaster children.
 219                  */
 220                 if (!PostmasterIsAlive())
 221                         exit(1);
 222
 223                 /*
 224                  * Check for any other interesting events that happened while we
 225                  * slept.
 226                  */
 227                 if (got_SIGHUP)
 228                 {
 229                         got_SIGHUP = false;
 230                         ProcessConfigFile(PGC_SIGHUP);
 231                 }
 232
 233                 if (firstchar != EOF)
 234                 {
 235                         /*
 236                          * Read the message contents. This is expected to be done without
 237                          * blocking because we've been able to get message type code.
 238                          */
 239                         if (pq_getmessage(&input_message, 0))
 240                                 firstchar = EOF;        /* suitable message already logged */
 241                 }
 242
 243                 /* Handle the very limited subset of commands expected in this phase */
 244                 switch (firstchar)
 245                 {
 246                         case 'Q':                       /* Query message */
 247                                 {
 248                                         const char *query_string;
 249
 250                                         query_string = pq_getmsgstring(&input_message);
 251                                         pq_getmsgend(&input_message);
 252
 253                                         if (HandleReplicationCommand(query_string))
 254                                                 replication_started = true;
 255                                 }
 256                                 break;
 257
 258                         case 'X':
 259                                 /* standby is closing the connection */
 260                                 proc_exit(0);
 261
 262                         case EOF:
 263                                 /* standby disconnected unexpectedly */
 264                                 ereport(COMMERROR,
 265                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 266                                                  errmsg("unexpected EOF on standby connection")));
 267                                 proc_exit(0);
 268
 269                         default:
 270                                 ereport(FATAL,
 271                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 272                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 273                 }
 274         }
 275 }
 276
 277 /*
 278  * IDENTIFY_SYSTEM
 279  */
 280 static void
 281 IdentifySystem(void)
 282 {
 283         StringInfoData buf;
 284         char            sysid[32];
 285         char            tli[11];
 286         char            xpos[MAXFNAMELEN];
 287         XLogRecPtr      logptr;
 288
 289         /*
 290          * Reply with a result set with one row, three columns. First col is
 291          * system ID, second is timeline ID, and third is current xlog location.
 292          */
 293
 294         snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
 295                          GetSystemIdentifier());
 296         snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
 297
 298         logptr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetInsertRecPtr();
 299
 300         snprintf(xpos, sizeof(xpos), "%X/%X",
 301                          logptr.xlogid, logptr.xrecoff);
 302
 303         /* Send a RowDescription message */
 304         pq_beginmessage(&buf, 'T');
 305         pq_sendint(&buf, 3, 2);         /* 3 fields */
 306
 307         /* first field */
 308         pq_sendstring(&buf, "systemid");        /* col name */
 309         pq_sendint(&buf, 0, 4);         /* table oid */
 310         pq_sendint(&buf, 0, 2);         /* attnum */
 311         pq_sendint(&buf, TEXTOID, 4);           /* type oid */
 312         pq_sendint(&buf, -1, 2);        /* typlen */
 313         pq_sendint(&buf, 0, 4);         /* typmod */
 314         pq_sendint(&buf, 0, 2);         /* format code */
 315
 316         /* second field */
 317         pq_sendstring(&buf, "timeline");        /* col name */
 318         pq_sendint(&buf, 0, 4);         /* table oid */
 319         pq_sendint(&buf, 0, 2);         /* attnum */
 320         pq_sendint(&buf, INT4OID, 4);           /* type oid */
 321         pq_sendint(&buf, 4, 2);         /* typlen */
 322         pq_sendint(&buf, 0, 4);         /* typmod */
 323         pq_sendint(&buf, 0, 2);         /* format code */
 324
 325         /* third field */
 326         pq_sendstring(&buf, "xlogpos");
 327         pq_sendint(&buf, 0, 4);
 328         pq_sendint(&buf, 0, 2);
 329         pq_sendint(&buf, TEXTOID, 4);
 330         pq_sendint(&buf, -1, 2);
 331         pq_sendint(&buf, 0, 4);
 332         pq_sendint(&buf, 0, 2);
 333         pq_endmessage(&buf);
 334
 335         /* Send a DataRow message */
 336         pq_beginmessage(&buf, 'D');
 337         pq_sendint(&buf, 3, 2);         /* # of columns */
 338         pq_sendint(&buf, strlen(sysid), 4); /* col1 len */
 339         pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
 340         pq_sendint(&buf, strlen(tli), 4);       /* col2 len */
 341         pq_sendbytes(&buf, (char *) tli, strlen(tli));
 342         pq_sendint(&buf, strlen(xpos), 4);      /* col3 len */
 343         pq_sendbytes(&buf, (char *) xpos, strlen(xpos));
 344
 345         pq_endmessage(&buf);
 346
 347         /* Send CommandComplete and ReadyForQuery messages */
 348         EndCommand("SELECT", DestRemote);
 349         ReadyForQuery(DestRemote);
 350         /* ReadyForQuery did pq_flush for us */
 351 }
 352
 353 /*
 354  * START_REPLICATION
 355  */
 356 static void
 357 StartReplication(StartReplicationCmd *cmd)
 358 {
 359         StringInfoData buf;
 360
 361         /*
 362          * Let postmaster know that we're streaming. Once we've declared us as a
 363          * WAL sender process, postmaster will let us outlive the bgwriter and
 364          * kill us last in the shutdown sequence, so we get a chance to stream all
 365          * remaining WAL at shutdown, including the shutdown checkpoint. Note that
 366          * there's no going back, and we mustn't write any WAL records after this.
 367          */
 368         MarkPostmasterChildWalSender();
 369         SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
 370
 371         /*
 372          * We assume here that we're logging enough information in the WAL for
 373          * log-shipping, since this is checked in PostmasterMain().
 374          *
 375          * NOTE: wal_level can only change at shutdown, so in most cases it is
 376          * difficult for there to be WAL data that we can still see that was written
 377          * at wal_level='minimal'.
 378          */
 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 able
 386          * 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 \"%c\"",
 518                                                                 firstchar)));
 519                 }
 520         }
 521
 522         /*
 523          * Save the last reply timestamp if we've received at least 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 based on reply data from
 590          * 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         if (!am_cascading_walsender)
 604                 SyncRepReleaseWaiters();
 605 }
 606
 607 /*
 608  * Hot Standby feedback
 609  */
 610 static void
 611 ProcessStandbyHSFeedbackMessage(void)
 612 {
 613         StandbyHSFeedbackMessage reply;
 614         TransactionId newxmin = InvalidTransactionId;
 615
 616         pq_copymsgbytes(&reply_message, (char *) &reply, sizeof(StandbyHSFeedbackMessage));
 617
 618         elog(DEBUG2, "hot standby feedback xmin %u epoch %u",
 619                  reply.xmin,
 620                  reply.epoch);
 621
 622         /*
 623          * Update the WalSender's proc xmin to allow it to be visible to
 624          * snapshots. This will hold back the removal of dead rows and thereby
 625          * prevent the generation of cleanup conflicts 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 if the counter has
 637                  * 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
 660                                 if (TransactionIdPrecedes(oldestXmin, reply.xmin))
 661                                         newxmin = reply.xmin;
 662                                 else
 663                                         newxmin = oldestXmin;
 664                         }
 665                         else
 666                         {
 667                                 if (TransactionIdPrecedes(MyProc->xmin, reply.xmin))
 668                                         newxmin = reply.xmin;
 669                                 else
 670                                         newxmin = MyProc->xmin;         /* stay the same */
 671                         }
 672                 }
 673         }
 674
 675         /*
 676          * Grab the ProcArrayLock to set xmin, or invalidate for bad reply
 677          */
 678         if (MyProc->xmin != newxmin)
 679         {
 680                 LWLockAcquire(ProcArrayLock, LW_SHARED);
 681                 MyProc->xmin = newxmin;
 682                 LWLockRelease(ProcArrayLock);
 683         }
 684 }
 685
 686 /* Main loop of walsender process */
 687 static int
 688 WalSndLoop(void)
 689 {
 690         char       *output_message;
 691         bool            caughtup = false;
 692
 693         /*
 694          * Allocate buffer that will be used for each output message.  We do this
 695          * just once to reduce palloc overhead.  The buffer must be made large
 696          * enough for maximum-sized messages.
 697          */
 698         output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
 699
 700         /*
 701          * Allocate buffer that will be used for processing reply messages.  As
 702          * above, do this just once to reduce palloc overhead.
 703          */
 704         initStringInfo(&reply_message);
 705
 706         /* Initialize the last reply timestamp */
 707         last_reply_timestamp = GetCurrentTimestamp();
 708
 709         /* Loop forever, unless we get an error */
 710         for (;;)
 711         {
 712                 /*
 713                  * Emergency bailout if postmaster has died.  This is to avoid the
 714                  * necessity for manual cleanup of all postmaster children.
 715                  */
 716                 if (!PostmasterIsAlive())
 717                         exit(1);
 718
 719                 /* Process any requests or signals received recently */
 720                 if (got_SIGHUP)
 721                 {
 722                         got_SIGHUP = false;
 723                         ProcessConfigFile(PGC_SIGHUP);
 724                         SyncRepInitConfig();
 725                 }
 726
 727                 /* Normal exit from the walsender is here */
 728                 if (walsender_shutdown_requested)
 729                 {
 730                         /* Inform the standby that XLOG streaming was done */
 731                         pq_puttextmessage('C', "COPY 0");
 732                         pq_flush();
 733
 734                         proc_exit(0);
 735                 }
 736
 737                 /*
 738                  * If we don't have any pending data in the output buffer, try to send
 739                  * some more.
 740                  */
 741                 if (!pq_is_send_pending())
 742                 {
 743                         XLogSend(output_message, &caughtup);
 744
 745                         /*
 746                          * Even if we wrote all the WAL that was available when we started
 747                          * sending, more might have arrived while we were sending this
 748                          * batch. We had the latch set while sending, so we have not
 749                          * received any signals from that time. Let's arm the latch again,
 750                          * and after that check that we're still up-to-date.
 751                          */
 752                         if (caughtup && !pq_is_send_pending())
 753                         {
 754                                 ResetLatch(&MyWalSnd->latch);
 755
 756                                 XLogSend(output_message, &caughtup);
 757                         }
 758                 }
 759
 760                 /* Flush pending output to the client */
 761                 if (pq_flush_if_writable() != 0)
 762                         break;
 763
 764                 /*
 765                  * When SIGUSR2 arrives, we send any outstanding logs up to the
 766                  * shutdown checkpoint record (i.e., the latest record) and exit.
 767                  * This may be a normal termination at shutdown, or a promotion,
 768                  * the walsender is not sure which.
 769                  */
 770                 if (walsender_ready_to_stop && !pq_is_send_pending())
 771                 {
 772                         XLogSend(output_message, &caughtup);
 773                         ProcessRepliesIfAny();
 774                         if (caughtup && !pq_is_send_pending())
 775                                 walsender_shutdown_requested = true;
 776                 }
 777
 778                 if ((caughtup || pq_is_send_pending()) &&
 779                         !got_SIGHUP &&
 780                         !walsender_shutdown_requested)
 781                 {
 782                         TimestampTz finish_time = 0;
 783                         long            sleeptime;
 784                         int                     wakeEvents;
 785
 786                         /* Reschedule replication timeout */
 787                         if (replication_timeout > 0)
 788                         {
 789                                 long            secs;
 790                                 int                     usecs;
 791
 792                                 finish_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
 793                                                                                                                 replication_timeout);
 794                                 TimestampDifference(GetCurrentTimestamp(),
 795                                                                         finish_time, &secs, &usecs);
 796                                 sleeptime = secs * 1000 + usecs / 1000;
 797                                 if (WalSndDelay < sleeptime)
 798                                         sleeptime = WalSndDelay;
 799                         }
 800                         else
 801                         {
 802                                 /*
 803                                  * XXX: Without timeout, we don't really need the periodic
 804                                  * wakeups anymore, WaitLatchOrSocket should reliably wake up
 805                                  * as soon as something interesting happens.
 806                                  */
 807                                 sleeptime = WalSndDelay;
 808                         }
 809
 810                         /* Sleep */
 811                         wakeEvents  = WL_LATCH_SET | WL_SOCKET_READABLE | WL_TIMEOUT;
 812                         if (pq_is_send_pending())
 813                                 wakeEvents |= WL_SOCKET_WRITEABLE;
 814                         WaitLatchOrSocket(&MyWalSnd->latch, wakeEvents,
 815                                                           MyProcPort->sock, sleeptime);
 816
 817                         /* Check for replication timeout */
 818                         if (replication_timeout > 0 &&
 819                                 GetCurrentTimestamp() >= finish_time)
 820                         {
 821                                 /*
 822                                  * Since typically expiration of replication timeout means
 823                                  * communication problem, we don't send the error message to
 824                                  * the standby.
 825                                  */
 826                                 ereport(COMMERROR,
 827                                                 (errmsg("terminating walsender process due to replication timeout")));
 828                                 break;
 829                         }
 830                 }
 831
 832                 /*
 833                  * If we're in catchup state, see if its time to move to streaming.
 834                  * This is an important state change for users, since before this
 835                  * point data loss might occur if the primary dies and we need to
 836                  * failover to the standby. The state change is also important for
 837                  * synchronous replication, since commits that started to wait at that
 838                  * point might wait for some time.
 839                  */
 840                 if (MyWalSnd->state == WALSNDSTATE_CATCHUP && caughtup)
 841                 {
 842                         ereport(DEBUG1,
 843                                         (errmsg("standby \"%s\" has now caught up with primary",
 844                                                         application_name)));
 845                         WalSndSetState(WALSNDSTATE_STREAMING);
 846                 }
 847
 848                 ProcessRepliesIfAny();
 849         }
 850
 851         /*
 852          * Get here on send failure.  Clean up and exit.
 853          *
 854          * Reset whereToSendOutput to prevent ereport from attempting to send any
 855          * more messages to the standby.
 856          */
 857         if (whereToSendOutput == DestRemote)
 858                 whereToSendOutput = DestNone;
 859
 860         proc_exit(0);
 861         return 1;                                       /* keep the compiler quiet */
 862 }
 863
 864 /* Initialize a per-walsender data structure for this walsender process */
 865 static void
 866 InitWalSnd(void)
 867 {
 868         int                     i;
 869
 870         /*
 871          * WalSndCtl should be set up already (we inherit this by fork() or
 872          * EXEC_BACKEND mechanism from the postmaster).
 873          */
 874         Assert(WalSndCtl != NULL);
 875         Assert(MyWalSnd == NULL);
 876
 877         /*
 878          * Find a free walsender slot and reserve it. If this fails, we must be
 879          * out of WalSnd structures.
 880          */
 881         for (i = 0; i < max_wal_senders; i++)
 882         {
 883                 /* use volatile pointer to prevent code rearrangement */
 884                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 885
 886                 SpinLockAcquire(&walsnd->mutex);
 887
 888                 if (walsnd->pid != 0)
 889                 {
 890                         SpinLockRelease(&walsnd->mutex);
 891                         continue;
 892                 }
 893                 else
 894                 {
 895                         /*
 896                          * Found a free slot. Reserve it for us.
 897                          */
 898                         walsnd->pid = MyProcPid;
 899                         MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
 900                         walsnd->state = WALSNDSTATE_STARTUP;
 901                         SpinLockRelease(&walsnd->mutex);
 902                         /* don't need the lock anymore */
 903                         OwnLatch((Latch *) &walsnd->latch);
 904                         MyWalSnd = (WalSnd *) walsnd;
 905
 906                         break;
 907                 }
 908         }
 909         if (MyWalSnd == NULL)
 910                 ereport(FATAL,
 911                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 912                                  errmsg("number of requested standby connections "
 913                                                 "exceeds max_wal_senders (currently %d)",
 914                                                 max_wal_senders)));
 915
 916         /* Arrange to clean up at walsender exit */
 917         on_shmem_exit(WalSndKill, 0);
 918 }
 919
 920 /* Destroy the per-walsender data structure for this walsender process */
 921 static void
 922 WalSndKill(int code, Datum arg)
 923 {
 924         Assert(MyWalSnd != NULL);
 925
 926         /*
 927          * Mark WalSnd struct no longer in use. Assume that no lock is required
 928          * for this.
 929          */
 930         MyWalSnd->pid = 0;
 931         DisownLatch(&MyWalSnd->latch);
 932
 933         /* WalSnd struct isn't mine anymore */
 934         MyWalSnd = NULL;
 935 }
 936
 937 /*
 938  * Read 'count' bytes from WAL into 'buf', starting at location 'startptr'
 939  *
 940  * XXX probably this should be improved to suck data directly from the
 941  * WAL buffers when possible.
 942  *
 943  * Will open, and keep open, one WAL segment stored in the global file
 944  * descriptor sendFile. This means if XLogRead is used once, there will
 945  * always be one descriptor left open until the process ends, but never
 946  * more than one.
 947  */
 948 void
 949 XLogRead(char *buf, XLogRecPtr startptr, Size count)
 950 {
 951         char               *p;
 952         XLogRecPtr      recptr;
 953         Size                    nbytes;
 954         uint32          lastRemovedLog;
 955         uint32          lastRemovedSeg;
 956         uint32          log;
 957         uint32          seg;
 958
 959 retry:
 960         p = buf;
 961         recptr = startptr;
 962         nbytes = count;
 963
 964         while (nbytes > 0)
 965         {
 966                 uint32          startoff;
 967                 int                     segbytes;
 968                 int                     readbytes;
 969
 970                 startoff = recptr.xrecoff % XLogSegSize;
 971
 972                 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
 973                 {
 974                         char            path[MAXPGPATH];
 975
 976                         /* Switch to another logfile segment */
 977                         if (sendFile >= 0)
 978                                 close(sendFile);
 979
 980                         XLByteToSeg(recptr, sendId, sendSeg);
 981                         XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
 982
 983                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
 984                         if (sendFile < 0)
 985                         {
 986                                 /*
 987                                  * If the file is not found, assume it's because the standby
 988                                  * asked for a too old WAL segment that has already been
 989                                  * removed or recycled.
 990                                  */
 991                                 if (errno == ENOENT)
 992                                 {
 993                                         char            filename[MAXFNAMELEN];
 994
 995                                         XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
 996                                         ereport(ERROR,
 997                                                         (errcode_for_file_access(),
 998                                                          errmsg("requested WAL segment %s has already been removed",
 999                                                                         filename)));
1000                                 }
1001                                 else
1002                                         ereport(ERROR,
1003                                                         (errcode_for_file_access(),
1004                                                          errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
1005                                                                         path, sendId, sendSeg)));
1006                         }
1007                         sendOff = 0;
1008                 }
1009
1010                 /* Need to seek in the file? */
1011                 if (sendOff != startoff)
1012                 {
1013                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
1014                                 ereport(ERROR,
1015                                                 (errcode_for_file_access(),
1016                                                  errmsg("could not seek in log file %u, segment %u to offset %u: %m",
1017                                                                 sendId, sendSeg, startoff)));
1018                         sendOff = startoff;
1019                 }
1020
1021                 /* How many bytes are within this segment? */
1022                 if (nbytes > (XLogSegSize - startoff))
1023                         segbytes = XLogSegSize - startoff;
1024                 else
1025                         segbytes = nbytes;
1026
1027                 readbytes = read(sendFile, p, segbytes);
1028                 if (readbytes <= 0)
1029                         ereport(ERROR,
1030                                         (errcode_for_file_access(),
1031                         errmsg("could not read from log file %u, segment %u, offset %u, "
1032                                    "length %lu: %m",
1033                                    sendId, sendSeg, sendOff, (unsigned long) segbytes)));
1034
1035                 /* Update state for read */
1036                 XLByteAdvance(recptr, readbytes);
1037
1038                 sendOff += readbytes;
1039                 nbytes -= readbytes;
1040                 p += readbytes;
1041         }
1042
1043         /*
1044          * After reading into the buffer, check that what we read was valid. We do
1045          * this after reading, because even though the segment was present when we
1046          * opened it, it might get recycled or removed while we read it. The
1047          * read() succeeds in that case, but the data we tried to read might
1048          * already have been overwritten with new WAL records.
1049          */
1050         XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
1051         XLByteToSeg(startptr, log, seg);
1052         if (log < lastRemovedLog ||
1053                 (log == lastRemovedLog && seg <= lastRemovedSeg))
1054         {
1055                 char            filename[MAXFNAMELEN];
1056
1057                 XLogFileName(filename, ThisTimeLineID, log, seg);
1058                 ereport(ERROR,
1059                                 (errcode_for_file_access(),
1060                                  errmsg("requested WAL segment %s has already been removed",
1061                                                 filename)));
1062         }
1063
1064         /*
1065          * During recovery, the currently-open WAL file might be replaced with
1066          * the file of the same name retrieved from archive. So we always need
1067          * to check what we read was valid after reading into the buffer. If it's
1068          * invalid, we try to open and read the file again.
1069          */
1070         if (am_cascading_walsender)
1071         {
1072                 /* use volatile pointer to prevent code rearrangement */
1073                 volatile WalSnd *walsnd = MyWalSnd;
1074                 bool            reload;
1075
1076                 SpinLockAcquire(&walsnd->mutex);
1077                 reload = walsnd->needreload;
1078                 walsnd->needreload = false;
1079                 SpinLockRelease(&walsnd->mutex);
1080
1081                 if (reload && sendFile >= 0)
1082                 {
1083                         close(sendFile);
1084                         sendFile = -1;
1085
1086                         goto retry;
1087                 }
1088         }
1089 }
1090
1091 /*
1092  * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
1093  * but not yet sent to the client, and buffer it in the libpq output
1094  * buffer.
1095  *
1096  * msgbuf is a work area in which the output message is constructed.  It's
1097  * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
1098  * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
1099  *
1100  * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
1101  * *caughtup is set to false.
1102
1103  */
1104 static void
1105 XLogSend(char *msgbuf, bool *caughtup)
1106 {
1107         XLogRecPtr      SendRqstPtr;
1108         XLogRecPtr      startptr;
1109         XLogRecPtr      endptr;
1110         Size            nbytes;
1111         WalDataMessageHeader msghdr;
1112
1113         /*
1114          * Attempt to send all data that's already been written out and fsync'd to
1115          * disk.  We cannot go further than what's been written out given the
1116          * current implementation of XLogRead().  And in any case it's unsafe to
1117          * send WAL that is not securely down to disk on the master: if the master
1118          * subsequently crashes and restarts, slaves must not have applied any WAL
1119          * that gets lost on the master.
1120          */
1121         SendRqstPtr = am_cascading_walsender ? GetStandbyFlushRecPtr() : GetFlushRecPtr();
1122
1123         /* Quick exit if nothing to do */
1124         if (XLByteLE(SendRqstPtr, sentPtr))
1125         {
1126                 *caughtup = true;
1127                 return;
1128         }
1129
1130         /*
1131          * Figure out how much to send in one message. If there's no more than
1132          * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
1133          * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
1134          *
1135          * The rounding is not only for performance reasons. Walreceiver relies on
1136          * the fact that we never split a WAL record across two messages. Since a
1137          * long WAL record is split at page boundary into continuation records,
1138          * page boundary is always a safe cut-off point. We also assume that
1139          * SendRqstPtr never points to the middle of a WAL record.
1140          */
1141         startptr = sentPtr;
1142         if (startptr.xrecoff >= XLogFileSize)
1143         {
1144                 /*
1145                  * crossing a logid boundary, skip the non-existent last log segment
1146                  * in previous logical log file.
1147                  */
1148                 startptr.xlogid += 1;
1149                 startptr.xrecoff = 0;
1150         }
1151
1152         endptr = startptr;
1153         XLByteAdvance(endptr, MAX_SEND_SIZE);
1154         if (endptr.xlogid != startptr.xlogid)
1155         {
1156                 /* Don't cross a logfile boundary within one message */
1157                 Assert(endptr.xlogid == startptr.xlogid + 1);
1158                 endptr.xlogid = startptr.xlogid;
1159                 endptr.xrecoff = XLogFileSize;
1160         }
1161
1162         /* if we went beyond SendRqstPtr, back off */
1163         if (XLByteLE(SendRqstPtr, endptr))
1164         {
1165                 endptr = SendRqstPtr;
1166                 *caughtup = true;
1167         }
1168         else
1169         {
1170                 /* round down to page boundary. */
1171                 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
1172                 *caughtup = false;
1173         }
1174
1175         nbytes = endptr.xrecoff - startptr.xrecoff;
1176         Assert(nbytes <= MAX_SEND_SIZE);
1177
1178         /*
1179          * OK to read and send the slice.
1180          */
1181         msgbuf[0] = 'w';
1182
1183         /*
1184          * Read the log directly into the output buffer to avoid extra memcpy
1185          * calls.
1186          */
1187         XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
1188
1189         /*
1190          * We fill the message header last so that the send timestamp is taken as
1191          * late as possible.
1192          */
1193         msghdr.dataStart = startptr;
1194         msghdr.walEnd = SendRqstPtr;
1195         msghdr.sendTime = GetCurrentTimestamp();
1196
1197         memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
1198
1199         pq_putmessage_noblock('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
1200
1201         sentPtr = endptr;
1202
1203         /* Update shared memory status */
1204         {
1205                 /* use volatile pointer to prevent code rearrangement */
1206                 volatile WalSnd *walsnd = MyWalSnd;
1207
1208                 SpinLockAcquire(&walsnd->mutex);
1209                 walsnd->sentPtr = sentPtr;
1210                 SpinLockRelease(&walsnd->mutex);
1211         }
1212
1213         /* Report progress of XLOG streaming in PS display */
1214         if (update_process_title)
1215         {
1216                 char            activitymsg[50];
1217
1218                 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
1219                                  sentPtr.xlogid, sentPtr.xrecoff);
1220                 set_ps_display(activitymsg, false);
1221         }
1222
1223         return;
1224 }
1225
1226 /*
1227  * Request walsenders to reload the currently-open WAL file
1228  */
1229 void
1230 WalSndRqstFileReload(void)
1231 {
1232         int                     i;
1233
1234         for (i = 0; i < max_wal_senders; i++)
1235         {
1236                 /* use volatile pointer to prevent code rearrangement */
1237                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1238
1239                 if (walsnd->pid == 0)
1240                         continue;
1241
1242                 SpinLockAcquire(&walsnd->mutex);
1243                 walsnd->needreload = true;
1244                 SpinLockRelease(&walsnd->mutex);
1245         }
1246 }
1247
1248 /* SIGHUP: set flag to re-read config file at next convenient time */
1249 static void
1250 WalSndSigHupHandler(SIGNAL_ARGS)
1251 {
1252         got_SIGHUP = true;
1253         if (MyWalSnd)
1254                 SetLatch(&MyWalSnd->latch);
1255 }
1256
1257 /* SIGTERM: set flag to shut down */
1258 static void
1259 WalSndShutdownHandler(SIGNAL_ARGS)
1260 {
1261         walsender_shutdown_requested = true;
1262         if (MyWalSnd)
1263                 SetLatch(&MyWalSnd->latch);
1264 }
1265
1266 /*
1267  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
1268  *
1269  * Some backend has bought the farm,
1270  * so we need to stop what we're doing and exit.
1271  */
1272 static void
1273 WalSndQuickDieHandler(SIGNAL_ARGS)
1274 {
1275         PG_SETMASK(&BlockSig);
1276
1277         /*
1278          * We DO NOT want to run proc_exit() callbacks -- we're here because
1279          * shared memory may be corrupted, so we don't want to try to clean up our
1280          * transaction.  Just nail the windows shut and get out of town.  Now that
1281          * there's an atexit callback to prevent third-party code from breaking
1282          * things by calling exit() directly, we have to reset the callbacks
1283          * explicitly to make this work as intended.
1284          */
1285         on_exit_reset();
1286
1287         /*
1288          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
1289          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
1290          * backend.  This is necessary precisely because we don't clean up our
1291          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
1292          * should ensure the postmaster sees this as a crash, too, but no harm in
1293          * being doubly sure.)
1294          */
1295         exit(2);
1296 }
1297
1298 /* SIGUSR1: set flag to send WAL records */
1299 static void
1300 WalSndXLogSendHandler(SIGNAL_ARGS)
1301 {
1302         latch_sigusr1_handler();
1303 }
1304
1305 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
1306 static void
1307 WalSndLastCycleHandler(SIGNAL_ARGS)
1308 {
1309         walsender_ready_to_stop = true;
1310         if (MyWalSnd)
1311                 SetLatch(&MyWalSnd->latch);
1312 }
1313
1314 /* Set up signal handlers */
1315 void
1316 WalSndSignals(void)
1317 {
1318         /* Set up signal handlers */
1319         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
1320                                                                                                  * file */
1321         pqsignal(SIGINT, SIG_IGN);      /* not used */
1322         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
1323         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
1324         pqsignal(SIGALRM, SIG_IGN);
1325         pqsignal(SIGPIPE, SIG_IGN);
1326         pqsignal(SIGUSR1, WalSndXLogSendHandler);       /* request WAL sending */
1327         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
1328                                                                                                  * shutdown */
1329
1330         /* Reset some signals that are accepted by postmaster but not here */
1331         pqsignal(SIGCHLD, SIG_DFL);
1332         pqsignal(SIGTTIN, SIG_DFL);
1333         pqsignal(SIGTTOU, SIG_DFL);
1334         pqsignal(SIGCONT, SIG_DFL);
1335         pqsignal(SIGWINCH, SIG_DFL);
1336 }
1337
1338 /* Report shared-memory space needed by WalSndShmemInit */
1339 Size
1340 WalSndShmemSize(void)
1341 {
1342         Size            size = 0;
1343
1344         size = offsetof(WalSndCtlData, walsnds);
1345         size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
1346
1347         return size;
1348 }
1349
1350 /* Allocate and initialize walsender-related shared memory */
1351 void
1352 WalSndShmemInit(void)
1353 {
1354         bool            found;
1355         int                     i;
1356
1357         WalSndCtl = (WalSndCtlData *)
1358                 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
1359
1360         if (!found)
1361         {
1362                 /* First time through, so initialize */
1363                 MemSet(WalSndCtl, 0, WalSndShmemSize());
1364
1365                 SHMQueueInit(&(WalSndCtl->SyncRepQueue));
1366
1367                 for (i = 0; i < max_wal_senders; i++)
1368                 {
1369                         WalSnd     *walsnd = &WalSndCtl->walsnds[i];
1370
1371                         SpinLockInit(&walsnd->mutex);
1372                         InitSharedLatch(&walsnd->latch);
1373                 }
1374         }
1375 }
1376
1377 /* Wake up all walsenders */
1378 void
1379 WalSndWakeup(void)
1380 {
1381         int                     i;
1382
1383         for (i = 0; i < max_wal_senders; i++)
1384                 SetLatch(&WalSndCtl->walsnds[i].latch);
1385 }
1386
1387 /* Set state for current walsender (only called in walsender) */
1388 void
1389 WalSndSetState(WalSndState state)
1390 {
1391         /* use volatile pointer to prevent code rearrangement */
1392         volatile WalSnd *walsnd = MyWalSnd;
1393
1394         Assert(am_walsender);
1395
1396         if (walsnd->state == state)
1397                 return;
1398
1399         SpinLockAcquire(&walsnd->mutex);
1400         walsnd->state = state;
1401         SpinLockRelease(&walsnd->mutex);
1402 }
1403
1404 /*
1405  * Return a string constant representing the state. This is used
1406  * in system views, and should *not* be translated.
1407  */
1408 static const char *
1409 WalSndGetStateString(WalSndState state)
1410 {
1411         switch (state)
1412         {
1413                 case WALSNDSTATE_STARTUP:
1414                         return "startup";
1415                 case WALSNDSTATE_BACKUP:
1416                         return "backup";
1417                 case WALSNDSTATE_CATCHUP:
1418                         return "catchup";
1419                 case WALSNDSTATE_STREAMING:
1420                         return "streaming";
1421         }
1422         return "UNKNOWN";
1423 }
1424
1425
1426 /*
1427  * Returns activity of walsenders, including pids and xlog locations sent to
1428  * standby servers.
1429  */
1430 Datum
1431 pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
1432 {
1433 #define PG_STAT_GET_WAL_SENDERS_COLS    8
1434         ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
1435         TupleDesc       tupdesc;
1436         Tuplestorestate *tupstore;
1437         MemoryContext per_query_ctx;
1438         MemoryContext oldcontext;
1439         int                *sync_priority;
1440         int                     priority = 0;
1441         int                     sync_standby = -1;
1442         int                     i;
1443
1444         /* check to see if caller supports us returning a tuplestore */
1445         if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
1446                 ereport(ERROR,
1447                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1448                                  errmsg("set-valued function called in context that cannot accept a set")));
1449         if (!(rsinfo->allowedModes & SFRM_Materialize))
1450                 ereport(ERROR,
1451                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1452                                  errmsg("materialize mode required, but it is not " \
1453                                                 "allowed in this context")));
1454
1455         /* Build a tuple descriptor for our result type */
1456         if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
1457                 elog(ERROR, "return type must be a row type");
1458
1459         per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
1460         oldcontext = MemoryContextSwitchTo(per_query_ctx);
1461
1462         tupstore = tuplestore_begin_heap(true, false, work_mem);
1463         rsinfo->returnMode = SFRM_Materialize;
1464         rsinfo->setResult = tupstore;
1465         rsinfo->setDesc = tupdesc;
1466
1467         MemoryContextSwitchTo(oldcontext);
1468
1469         /*
1470          * Get the priorities of sync standbys all in one go, to minimise lock
1471          * acquisitions and to allow us to evaluate who is the current sync
1472          * standby. This code must match the code in SyncRepReleaseWaiters().
1473          */
1474         sync_priority = palloc(sizeof(int) * max_wal_senders);
1475         LWLockAcquire(SyncRepLock, LW_SHARED);
1476         for (i = 0; i < max_wal_senders; i++)
1477         {
1478                 /* use volatile pointer to prevent code rearrangement */
1479                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1480
1481                 if (walsnd->pid != 0)
1482                 {
1483                         sync_priority[i] = walsnd->sync_standby_priority;
1484
1485                         if (walsnd->state == WALSNDSTATE_STREAMING &&
1486                                 walsnd->sync_standby_priority > 0 &&
1487                                 (priority == 0 ||
1488                                  priority > walsnd->sync_standby_priority))
1489                         {
1490                                 priority = walsnd->sync_standby_priority;
1491                                 sync_standby = i;
1492                         }
1493                 }
1494         }
1495         LWLockRelease(SyncRepLock);
1496
1497         for (i = 0; i < max_wal_senders; i++)
1498         {
1499                 /* use volatile pointer to prevent code rearrangement */
1500                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1501                 char            location[MAXFNAMELEN];
1502                 XLogRecPtr      sentPtr;
1503                 XLogRecPtr      write;
1504                 XLogRecPtr      flush;
1505                 XLogRecPtr      apply;
1506                 WalSndState state;
1507                 Datum           values[PG_STAT_GET_WAL_SENDERS_COLS];
1508                 bool            nulls[PG_STAT_GET_WAL_SENDERS_COLS];
1509
1510                 if (walsnd->pid == 0)
1511                         continue;
1512
1513                 SpinLockAcquire(&walsnd->mutex);
1514                 sentPtr = walsnd->sentPtr;
1515                 state = walsnd->state;
1516                 write = walsnd->write;
1517                 flush = walsnd->flush;
1518                 apply = walsnd->apply;
1519                 SpinLockRelease(&walsnd->mutex);
1520
1521                 memset(nulls, 0, sizeof(nulls));
1522                 values[0] = Int32GetDatum(walsnd->pid);
1523
1524                 if (!superuser())
1525                 {
1526                         /*
1527                          * Only superusers can see details. Other users only get the pid
1528                          * value to know it's a walsender, but no details.
1529                          */
1530                         MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
1531                 }
1532                 else
1533                 {
1534                         values[1] = CStringGetTextDatum(WalSndGetStateString(state));
1535
1536                         snprintf(location, sizeof(location), "%X/%X",
1537                                          sentPtr.xlogid, sentPtr.xrecoff);
1538                         values[2] = CStringGetTextDatum(location);
1539
1540                         if (write.xlogid == 0 && write.xrecoff == 0)
1541                                 nulls[3] = true;
1542                         snprintf(location, sizeof(location), "%X/%X",
1543                                          write.xlogid, write.xrecoff);
1544                         values[3] = CStringGetTextDatum(location);
1545
1546                         if (flush.xlogid == 0 && flush.xrecoff == 0)
1547                                 nulls[4] = true;
1548                         snprintf(location, sizeof(location), "%X/%X",
1549                                          flush.xlogid, flush.xrecoff);
1550                         values[4] = CStringGetTextDatum(location);
1551
1552                         if (apply.xlogid == 0 && apply.xrecoff == 0)
1553                                 nulls[5] = true;
1554                         snprintf(location, sizeof(location), "%X/%X",
1555                                          apply.xlogid, apply.xrecoff);
1556                         values[5] = CStringGetTextDatum(location);
1557
1558                         values[6] = Int32GetDatum(sync_priority[i]);
1559
1560                         /*
1561                          * More easily understood version of standby state. This is purely
1562                          * informational, not different from priority.
1563                          */
1564                         if (sync_priority[i] == 0)
1565                                 values[7] = CStringGetTextDatum("async");
1566                         else if (i == sync_standby)
1567                                 values[7] = CStringGetTextDatum("sync");
1568                         else
1569                                 values[7] = CStringGetTextDatum("potential");
1570                 }
1571
1572                 tuplestore_putvalues(tupstore, tupdesc, values, nulls);
1573         }
1574         pfree(sync_priority);
1575
1576         /* clean up and return the tuplestore */
1577         tuplestore_donestoring(tupstore);
1578
1579         return (Datum) 0;
1580 }
1581
1582 /*
1583  * This isn't currently used for anything. Monitoring tools might be
1584  * interested in the future, and we'll need something like this in the
1585  * future for synchronous replication.
1586  */
1587 #ifdef NOT_USED
1588 /*
1589  * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
1590  * if none.
1591  */
1592 XLogRecPtr
1593 GetOldestWALSendPointer(void)
1594 {
1595         XLogRecPtr      oldest = {0, 0};
1596         int                     i;
1597         bool            found = false;
1598
1599         for (i = 0; i < max_wal_senders; i++)
1600         {
1601                 /* use volatile pointer to prevent code rearrangement */
1602                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1603                 XLogRecPtr      recptr;
1604
1605                 if (walsnd->pid == 0)
1606                         continue;
1607
1608                 SpinLockAcquire(&walsnd->mutex);
1609                 recptr = walsnd->sentPtr;
1610                 SpinLockRelease(&walsnd->mutex);
1611
1612                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1613                         continue;
1614
1615                 if (!found || XLByteLT(recptr, oldest))
1616                         oldest = recptr;
1617                 found = true;
1618         }
1619         return oldest;
1620 }
1621
1622 #endif