src/backend/replication/walsender.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * walsender.c
   4  *
   5  * The WAL sender process (walsender) is new as of Postgres 9.0. It takes
   6  * care of sending XLOG from the primary server to a single recipient.
   7  * (Note that there can be more than one walsender process concurrently.)
   8  * It is started by the postmaster when the walreceiver of a standby server
   9  * connects to the primary server and requests XLOG streaming replication.
  10  * It attempts to keep reading XLOG records from the disk and sending them
  11  * to the standby server, as long as the connection is alive (i.e., like
  12  * any backend, there is a one-to-one relationship between a connection
  13  * and a walsender process).
  14  *
  15  * Normal termination is by SIGTERM, which instructs the walsender to
  16  * close the connection and exit(0) at next convenient moment. Emergency
  17  * termination is by SIGQUIT; like any backend, the walsender will simply
  18  * abort and exit on SIGQUIT. A close of the connection and a FATAL error
  19  * are treated as not a crash but approximately normal termination;
  20  * the walsender will exit quickly without sending any more XLOG records.
  21  *
  22  * If the server is shut down, postmaster sends us SIGUSR2 after all
  23  * regular backends have exited and the shutdown checkpoint has been written.
  24  * This instruct walsender to send any outstanding WAL, including the
  25  * shutdown checkpoint record, and then exit.
  26  *
  27  *
  28  * Portions Copyright (c) 2010-2010, PostgreSQL Global Development Group
  29  *
  30  * IDENTIFICATION
  31  *        $PostgreSQL: pgsql/src/backend/replication/walsender.c,v 1.32 2010/09/15 06:51:19 heikki Exp $
  32  *
  33  *-------------------------------------------------------------------------
  34  */
  35 #include "postgres.h"
  36
  37 #include <signal.h>
  38 #include <unistd.h>
  39
  40 #include "access/xlog_internal.h"
  41 #include "catalog/pg_type.h"
  42 #include "libpq/libpq.h"
  43 #include "libpq/pqformat.h"
  44 #include "libpq/pqsignal.h"
  45 #include "miscadmin.h"
  46 #include "replication/walprotocol.h"
  47 #include "replication/walsender.h"
  48 #include "storage/fd.h"
  49 #include "storage/ipc.h"
  50 #include "storage/pmsignal.h"
  51 #include "tcop/tcopprot.h"
  52 #include "utils/guc.h"
  53 #include "utils/memutils.h"
  54 #include "utils/ps_status.h"
  55
  56
  57 /* Array of WalSnds in shared memory */
  58 WalSndCtlData *WalSndCtl = NULL;
  59
  60 /* My slot in the shared memory array */
  61 static WalSnd *MyWalSnd = NULL;
  62
  63 /* Global state */
  64 bool            am_walsender = false;           /* Am I a walsender process ? */
  65
  66 /* User-settable parameters for walsender */
  67 int                     max_wal_senders = 0;    /* the maximum number of concurrent walsenders */
  68 int                     WalSndDelay = 200;      /* max sleep time between some actions */
  69
  70 /*
  71  * These variables are used similarly to openLogFile/Id/Seg/Off,
  72  * but for walsender to read the XLOG.
  73  */
  74 static int      sendFile = -1;
  75 static uint32 sendId = 0;
  76 static uint32 sendSeg = 0;
  77 static uint32 sendOff = 0;
  78
  79 /*
  80  * How far have we sent WAL already? This is also advertised in
  81  * MyWalSnd->sentPtr.  (Actually, this is the next WAL location to send.)
  82  */
  83 static XLogRecPtr sentPtr = {0, 0};
  84
  85 /*
  86  * How far have we completed replication already? This is also
  87  * advertised in MyWalSnd->ackdPtr. This is not used in asynchronous
  88  * replication case.
  89  */
  90 static XLogRecPtr ackdPtr = {0, 0};
  91
  92 /* Flags set by signal handlers for later service in main loop */
  93 static volatile sig_atomic_t got_SIGHUP = false;
  94 static volatile sig_atomic_t shutdown_requested = false;
  95 static volatile sig_atomic_t ready_to_stop = false;
  96
  97 /* Signal handlers */
  98 static void WalSndSigHupHandler(SIGNAL_ARGS);
  99 static void WalSndShutdownHandler(SIGNAL_ARGS);
 100 static void WalSndQuickDieHandler(SIGNAL_ARGS);
 101 static void WalSndXLogSendHandler(SIGNAL_ARGS);
 102 static void WalSndLastCycleHandler(SIGNAL_ARGS);
 103
 104 /* Prototypes for private functions */
 105 static int      WalSndLoop(void);
 106 static void InitWalSnd(void);
 107 static void WalSndHandshake(void);
 108 static void WalSndKill(int code, Datum arg);
 109 static void XLogRead(char *buf, XLogRecPtr recptr, Size nbytes);
 110 static bool XLogSend(char *msgbuf, bool *caughtup);
 111 static void ProcessStreamMsgs(StringInfo inMsg);
 112
 113
 114 /* Main entry point for walsender process */
 115 int
 116 WalSenderMain(void)
 117 {
 118         MemoryContext walsnd_context;
 119
 120         if (RecoveryInProgress())
 121                 ereport(FATAL,
 122                                 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 123                                  errmsg("recovery is still in progress, can't accept WAL streaming connections")));
 124
 125         /* Create a per-walsender data structure in shared memory */
 126         InitWalSnd();
 127
 128         /*
 129          * Create a memory context that we will do all our work in.  We do this so
 130          * that we can reset the context during error recovery and thereby avoid
 131          * possible memory leaks.  Formerly this code just ran in
 132          * TopMemoryContext, but resetting that would be a really bad idea.
 133          *
 134          * XXX: we don't actually attempt error recovery in walsender, we just
 135          * close the connection and exit.
 136          */
 137         walsnd_context = AllocSetContextCreate(TopMemoryContext,
 138                                                                                    "Wal Sender",
 139                                                                                    ALLOCSET_DEFAULT_MINSIZE,
 140                                                                                    ALLOCSET_DEFAULT_INITSIZE,
 141                                                                                    ALLOCSET_DEFAULT_MAXSIZE);
 142         MemoryContextSwitchTo(walsnd_context);
 143
 144         /* Unblock signals (they were blocked when the postmaster forked us) */
 145         PG_SETMASK(&UnBlockSig);
 146
 147         /* Tell the standby that walsender is ready for receiving commands */
 148         ReadyForQuery(DestRemote);
 149
 150         /* Handle handshake messages before streaming */
 151         WalSndHandshake();
 152
 153         /* Initialize shared memory status */
 154         {
 155                 /* use volatile pointer to prevent code rearrangement */
 156                 volatile WalSnd *walsnd = MyWalSnd;
 157
 158                 SpinLockAcquire(&walsnd->mutex);
 159                 walsnd->sentPtr = sentPtr;
 160                 SpinLockRelease(&walsnd->mutex);
 161         }
 162
 163         /* Main loop of walsender */
 164         return WalSndLoop();
 165 }
 166
 167 /*
 168  * Execute commands from walreceiver, until we enter streaming mode.
 169  */
 170 static void
 171 WalSndHandshake(void)
 172 {
 173         StringInfoData input_message;
 174         bool            replication_started = false;
 175
 176         initStringInfo(&input_message);
 177
 178         while (!replication_started)
 179         {
 180                 int                     firstchar;
 181
 182                 /* Wait for a command to arrive */
 183                 firstchar = pq_getbyte();
 184
 185                 /*
 186                  * Emergency bailout if postmaster has died.  This is to avoid the
 187                  * necessity for manual cleanup of all postmaster children.
 188                  */
 189                 if (!PostmasterIsAlive(true))
 190                         exit(1);
 191
 192                 /*
 193                  * Check for any other interesting events that happened while we
 194                  * slept.
 195                  */
 196                 if (got_SIGHUP)
 197                 {
 198                         got_SIGHUP = false;
 199                         ProcessConfigFile(PGC_SIGHUP);
 200                 }
 201
 202                 if (firstchar != EOF)
 203                 {
 204                         /*
 205                          * Read the message contents. This is expected to be done without
 206                          * blocking because we've been able to get message type code.
 207                          */
 208                         if (pq_getmessage(&input_message, 0))
 209                                 firstchar = EOF;        /* suitable message already logged */
 210                 }
 211
 212                 /* Handle the very limited subset of commands expected in this phase */
 213                 switch (firstchar)
 214                 {
 215                         case 'Q':                       /* Query message */
 216                                 {
 217                                         const char *query_string;
 218                                         XLogRecPtr      recptr;
 219
 220                                         query_string = pq_getmsgstring(&input_message);
 221                                         pq_getmsgend(&input_message);
 222
 223                                         if (strcmp(query_string, "IDENTIFY_SYSTEM") == 0)
 224                                         {
 225                                                 StringInfoData buf;
 226                                                 char            sysid[32];
 227                                                 char            tli[11];
 228
 229                                                 /*
 230                                                  * Reply with a result set with one row, two columns.
 231                                                  * First col is system ID, and second is timeline ID
 232                                                  */
 233
 234                                                 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
 235                                                                  GetSystemIdentifier());
 236                                                 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
 237
 238                                                 /* Send a RowDescription message */
 239                                                 pq_beginmessage(&buf, 'T');
 240                                                 pq_sendint(&buf, 2, 2); /* 2 fields */
 241
 242                                                 /* first field */
 243                                                 pq_sendstring(&buf, "systemid");                /* col name */
 244                                                 pq_sendint(&buf, 0, 4); /* table oid */
 245                                                 pq_sendint(&buf, 0, 2); /* attnum */
 246                                                 pq_sendint(&buf, TEXTOID, 4);   /* type oid */
 247                                                 pq_sendint(&buf, -1, 2);                /* typlen */
 248                                                 pq_sendint(&buf, 0, 4); /* typmod */
 249                                                 pq_sendint(&buf, 0, 2); /* format code */
 250
 251                                                 /* second field */
 252                                                 pq_sendstring(&buf, "timeline");                /* col name */
 253                                                 pq_sendint(&buf, 0, 4); /* table oid */
 254                                                 pq_sendint(&buf, 0, 2); /* attnum */
 255                                                 pq_sendint(&buf, INT4OID, 4);   /* type oid */
 256                                                 pq_sendint(&buf, 4, 2); /* typlen */
 257                                                 pq_sendint(&buf, 0, 4); /* typmod */
 258                                                 pq_sendint(&buf, 0, 2); /* format code */
 259                                                 pq_endmessage(&buf);
 260
 261                                                 /* Send a DataRow message */
 262                                                 pq_beginmessage(&buf, 'D');
 263                                                 pq_sendint(&buf, 2, 2); /* # of columns */
 264                                                 pq_sendint(&buf, strlen(sysid), 4);             /* col1 len */
 265                                                 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
 266                                                 pq_sendint(&buf, strlen(tli), 4);               /* col2 len */
 267                                                 pq_sendbytes(&buf, (char *) tli, strlen(tli));
 268                                                 pq_endmessage(&buf);
 269
 270                                                 /* Send CommandComplete and ReadyForQuery messages */
 271                                                 EndCommand("SELECT", DestRemote);
 272                                                 ReadyForQuery(DestRemote);
 273                                                 /* ReadyForQuery did pq_flush for us */
 274                                         }
 275                                         else if (sscanf(query_string, "START_REPLICATION %X/%X",
 276                                                                         &recptr.xlogid, &recptr.xrecoff) == 2)
 277                                         {
 278                                                 StringInfoData buf;
 279
 280                                                 /*
 281                                                  * Check that we're logging enough information in the
 282                                                  * WAL for log-shipping.
 283                                                  *
 284                                                  * NOTE: This only checks the current value of
 285                                                  * wal_level. Even if the current setting is not
 286                                                  * 'minimal', there can be old WAL in the pg_xlog
 287                                                  * directory that was created with 'minimal'. So this
 288                                                  * is not bulletproof, the purpose is just to give a
 289                                                  * user-friendly error message that hints how to
 290                                                  * configure the system correctly.
 291                                                  */
 292                                                 if (wal_level == WAL_LEVEL_MINIMAL)
 293                                                         ereport(FATAL,
 294                                                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 295                                                                          errmsg("standby connections not allowed because wal_level=minimal")));
 296
 297                                                 /* Send a CopyXLogResponse message, and start streaming */
 298                                                 pq_beginmessage(&buf, 'W');
 299                                                 pq_endmessage(&buf);
 300                                                 pq_flush();
 301
 302                                                 /*
 303                                                  * Initialize positions to the received one, then the
 304                                                  * xlog records begin to be shipped from that position
 305                                                  */
 306                                                 sentPtr = ackdPtr = recptr;
 307
 308                                                 /* break out of the loop */
 309                                                 replication_started = true;
 310                                         }
 311                                         else
 312                                         {
 313                                                 ereport(FATAL,
 314                                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 315                                                                  errmsg("invalid standby query string: %s", query_string)));
 316                                         }
 317                                         break;
 318                                 }
 319
 320                         case 'X':
 321                                 /* standby is closing the connection */
 322                                 proc_exit(0);
 323
 324                         case EOF:
 325                                 /* standby disconnected unexpectedly */
 326                                 ereport(COMMERROR,
 327                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 328                                                  errmsg("unexpected EOF on standby connection")));
 329                                 proc_exit(0);
 330
 331                         default:
 332                                 ereport(FATAL,
 333                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 334                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 335                 }
 336         }
 337 }
 338
 339 /*
 340  * Process messages received from the standby.
 341  *
 342  * ereports on error.
 343  */
 344 static void
 345 ProcessStreamMsgs(StringInfo inMsg)
 346 {
 347         bool    acked = false;
 348
 349         /* Loop to process successive complete messages available */
 350         for (;;)
 351         {
 352                 unsigned char firstchar;
 353                 int                     r;
 354
 355                 r = pq_getbyte_if_available(&firstchar);
 356                 if (r < 0)
 357                 {
 358                         /* unexpected error or EOF */
 359                         ereport(COMMERROR,
 360                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 361                                          errmsg("unexpected EOF on standby connection")));
 362                         proc_exit(0);
 363                 }
 364                 if (r == 0)
 365                 {
 366                         /* no data available without blocking */
 367                         break;
 368                 }
 369
 370                 /* Handle the very limited subset of commands expected in this phase */
 371                 switch (firstchar)
 372                 {
 373                         case 'd':       /* CopyData message */
 374                         {
 375                                 unsigned char   rpltype;
 376
 377                                 /*
 378                                  * Read the message contents. This is expected to be done without
 379                                  * blocking because we've been able to get message type code.
 380                                  */
 381                                 if (pq_getmessage(inMsg, 0))
 382                                         proc_exit(0);           /* suitable message already logged */
 383
 384                                 /* Read the replication message type from CopyData message */
 385                                 rpltype = pq_getmsgbyte(inMsg);
 386                                 switch (rpltype)
 387                                 {
 388                                         case 'l':
 389                                         {
 390                                                 WalAckMessageData  *msgdata;
 391
 392                                                 msgdata = (WalAckMessageData *) pq_getmsgbytes(inMsg, sizeof(WalAckMessageData));
 393
 394                                                 /*
 395                                                  * Update local status.
 396                                                  *
 397                                                  * The ackd ptr received from standby should not
 398                                                  * go backwards.
 399                                                  */
 400                                                 if (XLByteLE(ackdPtr, msgdata->ackEnd))
 401                                                         ackdPtr = msgdata->ackEnd;
 402                                                 else
 403                                                         ereport(FATAL,
 404                                                                         (errmsg("replication completion location went back from "
 405                                                                                         "%X/%X to %X/%X",
 406                                                                                         ackdPtr.xlogid, ackdPtr.xrecoff,
 407                                                                                         msgdata->ackEnd.xlogid, msgdata->ackEnd.xrecoff)));
 408
 409                                                 acked = true;   /* also need to update shared position */
 410                                                 break;
 411                                         }
 412                                         default:
 413                                                 ereport(FATAL,
 414                                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 415                                                                  errmsg("invalid replication message type %d",
 416                                                                                 rpltype)));
 417                                 }
 418                                 break;
 419                         }
 420
 421                         /*
 422                          * 'X' means that the standby is closing down the socket.
 423                          */
 424                         case 'X':
 425                                 proc_exit(0);
 426
 427                         default:
 428                                 ereport(FATAL,
 429                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 430                                                  errmsg("invalid standby closing message type %d",
 431                                                                 firstchar)));
 432                 }
 433         }
 434
 435         if (acked)
 436         {
 437                 /* use volatile pointer to prevent code rearrangement */
 438                 volatile WalSnd *walsnd = MyWalSnd;
 439
 440                 SpinLockAcquire(&walsnd->mutex);
 441                 walsnd->ackdPtr = ackdPtr;
 442                 SpinLockRelease(&walsnd->mutex);
 443         }
 444 }
 445
 446 /* Main loop of walsender process */
 447 static int
 448 WalSndLoop(void)
 449 {
 450         StringInfoData  input_message;
 451         char       *output_message;
 452         bool            caughtup = false;
 453
 454         initStringInfo(&input_message);
 455
 456         /*
 457          * Allocate buffer that will be used for each output message.  We do this
 458          * just once to reduce palloc overhead.  The buffer must be made large
 459          * enough for maximum-sized messages.
 460          */
 461         output_message = palloc(1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE);
 462
 463         /* Loop forever, unless we get an error */
 464         for (;;)
 465         {
 466                 /*
 467                  * Emergency bailout if postmaster has died.  This is to avoid the
 468                  * necessity for manual cleanup of all postmaster children.
 469                  */
 470                 if (!PostmasterIsAlive(true))
 471                         exit(1);
 472
 473                 /* Process any requests or signals received recently */
 474                 if (got_SIGHUP)
 475                 {
 476                         got_SIGHUP = false;
 477                         ProcessConfigFile(PGC_SIGHUP);
 478                 }
 479
 480                 /*
 481                  * When SIGUSR2 arrives, we send all outstanding logs up to the
 482                  * shutdown checkpoint record (i.e., the latest record) and exit.
 483                  */
 484                 if (ready_to_stop)
 485                 {
 486                         if (!XLogSend(output_message, &caughtup))
 487                                 break;
 488                         if (caughtup)
 489                                 shutdown_requested = true;
 490                 }
 491
 492                 /* Normal exit from the walsender is here */
 493                 if (shutdown_requested)
 494                 {
 495                         /* Inform the standby that XLOG streaming was done */
 496                         pq_puttextmessage('C', "COPY 0");
 497                         pq_flush();
 498
 499                         proc_exit(0);
 500                 }
 501
 502                 /*
 503                  * If we had sent all accumulated WAL in last round, nap for the
 504                  * configured time before retrying.
 505                  */
 506                 if (caughtup)
 507                 {
 508                         /*
 509                          * Even if we wrote all the WAL that was available when we started
 510                          * sending, more might have arrived while we were sending this
 511                          * batch. We had the latch set while sending, so we have not
 512                          * received any signals from that time. Let's arm the latch
 513                          * again, and after that check that we're still up-to-date.
 514                          */
 515                         ResetLatch(&MyWalSnd->latch);
 516
 517                         if (!XLogSend(output_message, &caughtup))
 518                                 break;
 519                         if (caughtup && !got_SIGHUP && !ready_to_stop && !shutdown_requested)
 520                         {
 521                                 /*
 522                                  * XXX: We don't really need the periodic wakeups anymore,
 523                                  * WaitLatchOrSocket should reliably wake up as soon as
 524                                  * something interesting happens.
 525                                  */
 526
 527                                 /* Sleep */
 528                                 WaitLatchOrSocket(&MyWalSnd->latch, MyProcPort->sock,
 529                                                                   WalSndDelay * 1000L);
 530                         }
 531
 532                         /* Process messages received from the standby */
 533                         ProcessStreamMsgs(&input_message);
 534                 }
 535                 else
 536                 {
 537                         /* Attempt to send the log once every loop */
 538                         if (!XLogSend(output_message, &caughtup))
 539                                 break;
 540                 }
 541         }
 542
 543         /*
 544          * Get here on send failure.  Clean up and exit.
 545          *
 546          * Reset whereToSendOutput to prevent ereport from attempting to send any
 547          * more messages to the standby.
 548          */
 549         if (whereToSendOutput == DestRemote)
 550                 whereToSendOutput = DestNone;
 551
 552         proc_exit(0);
 553         return 1;                                       /* keep the compiler quiet */
 554 }
 555
 556 /* Initialize a per-walsender data structure for this walsender process */
 557 static void
 558 InitWalSnd(void)
 559 {
 560         int                     i;
 561
 562         /*
 563          * WalSndCtl should be set up already (we inherit this by fork() or
 564          * EXEC_BACKEND mechanism from the postmaster).
 565          */
 566         Assert(WalSndCtl != NULL);
 567         Assert(MyWalSnd == NULL);
 568
 569         /*
 570          * Find a free walsender slot and reserve it. If this fails, we must be
 571          * out of WalSnd structures.
 572          */
 573         for (i = 0; i < max_wal_senders; i++)
 574         {
 575                 /* use volatile pointer to prevent code rearrangement */
 576                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 577
 578                 SpinLockAcquire(&walsnd->mutex);
 579
 580                 if (walsnd->pid != 0)
 581                 {
 582                         SpinLockRelease(&walsnd->mutex);
 583                         continue;
 584                 }
 585                 else
 586                 {
 587                         /*
 588                          * Found a free slot. Reserve it for us.
 589                          */
 590                         walsnd->pid = MyProcPid;
 591                         MemSet(&walsnd->sentPtr, 0, sizeof(XLogRecPtr));
 592                         MemSet(&walsnd->ackdPtr, 0, sizeof(XLogRecPtr));
 593                         SpinLockRelease(&walsnd->mutex);
 594                         /* don't need the lock anymore */
 595                         OwnLatch((Latch *) &walsnd->latch);
 596                         MyWalSnd = (WalSnd *) walsnd;
 597
 598                         break;
 599                 }
 600         }
 601         if (MyWalSnd == NULL)
 602                 ereport(FATAL,
 603                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 604                                  errmsg("number of requested standby connections "
 605                                                 "exceeds max_wal_senders (currently %d)",
 606                                                 max_wal_senders)));
 607
 608         /* Arrange to clean up at walsender exit */
 609         on_shmem_exit(WalSndKill, 0);
 610 }
 611
 612 /* Destroy the per-walsender data structure for this walsender process */
 613 static void
 614 WalSndKill(int code, Datum arg)
 615 {
 616         Assert(MyWalSnd != NULL);
 617
 618         /*
 619          * Mark WalSnd struct no longer in use. Assume that no lock is required
 620          * for this.
 621          */
 622         MyWalSnd->pid = 0;
 623         DisownLatch(&MyWalSnd->latch);
 624
 625         /* WalSnd struct isn't mine anymore */
 626         MyWalSnd = NULL;
 627 }
 628
 629 /*
 630  * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
 631  *
 632  * XXX probably this should be improved to suck data directly from the
 633  * WAL buffers when possible.
 634  */
 635 static void
 636 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
 637 {
 638         XLogRecPtr      startRecPtr = recptr;
 639         char            path[MAXPGPATH];
 640         uint32          lastRemovedLog;
 641         uint32          lastRemovedSeg;
 642         uint32          log;
 643         uint32          seg;
 644
 645         while (nbytes > 0)
 646         {
 647                 uint32          startoff;
 648                 int                     segbytes;
 649                 int                     readbytes;
 650
 651                 startoff = recptr.xrecoff % XLogSegSize;
 652
 653                 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
 654                 {
 655                         /* Switch to another logfile segment */
 656                         if (sendFile >= 0)
 657                                 close(sendFile);
 658
 659                         XLByteToSeg(recptr, sendId, sendSeg);
 660                         XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
 661
 662                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
 663                         if (sendFile < 0)
 664                         {
 665                                 /*
 666                                  * If the file is not found, assume it's because the standby
 667                                  * asked for a too old WAL segment that has already been
 668                                  * removed or recycled.
 669                                  */
 670                                 if (errno == ENOENT)
 671                                 {
 672                                         char            filename[MAXFNAMELEN];
 673
 674                                         XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
 675                                         ereport(ERROR,
 676                                                         (errcode_for_file_access(),
 677                                                          errmsg("requested WAL segment %s has already been removed",
 678                                                                         filename)));
 679                                 }
 680                                 else
 681                                         ereport(ERROR,
 682                                                         (errcode_for_file_access(),
 683                                                          errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
 684                                                                         path, sendId, sendSeg)));
 685                         }
 686                         sendOff = 0;
 687                 }
 688
 689                 /* Need to seek in the file? */
 690                 if (sendOff != startoff)
 691                 {
 692                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
 693                                 ereport(ERROR,
 694                                                 (errcode_for_file_access(),
 695                                                  errmsg("could not seek in log file %u, segment %u to offset %u: %m",
 696                                                                 sendId, sendSeg, startoff)));
 697                         sendOff = startoff;
 698                 }
 699
 700                 /* How many bytes are within this segment? */
 701                 if (nbytes > (XLogSegSize - startoff))
 702                         segbytes = XLogSegSize - startoff;
 703                 else
 704                         segbytes = nbytes;
 705
 706                 readbytes = read(sendFile, buf, segbytes);
 707                 if (readbytes <= 0)
 708                         ereport(ERROR,
 709                                         (errcode_for_file_access(),
 710                         errmsg("could not read from log file %u, segment %u, offset %u, "
 711                                    "length %lu: %m",
 712                                    sendId, sendSeg, sendOff, (unsigned long) segbytes)));
 713
 714                 /* Update state for read */
 715                 XLByteAdvance(recptr, readbytes);
 716
 717                 sendOff += readbytes;
 718                 nbytes -= readbytes;
 719                 buf += readbytes;
 720         }
 721
 722         /*
 723          * After reading into the buffer, check that what we read was valid. We do
 724          * this after reading, because even though the segment was present when we
 725          * opened it, it might get recycled or removed while we read it. The
 726          * read() succeeds in that case, but the data we tried to read might
 727          * already have been overwritten with new WAL records.
 728          */
 729         XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
 730         XLByteToSeg(startRecPtr, log, seg);
 731         if (log < lastRemovedLog ||
 732                 (log == lastRemovedLog && seg <= lastRemovedSeg))
 733         {
 734                 char            filename[MAXFNAMELEN];
 735
 736                 XLogFileName(filename, ThisTimeLineID, log, seg);
 737                 ereport(ERROR,
 738                                 (errcode_for_file_access(),
 739                                  errmsg("requested WAL segment %s has already been removed",
 740                                                 filename)));
 741         }
 742 }
 743
 744 /*
 745  * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
 746  * but not yet sent to the client, and send it.
 747  *
 748  * msgbuf is a work area in which the output message is constructed.  It's
 749  * passed in just so we can avoid re-palloc'ing the buffer on each cycle.
 750  * It must be of size 1 + sizeof(WalDataMessageHeader) + MAX_SEND_SIZE.
 751  *
 752  * If there is no unsent WAL remaining, *caughtup is set to true, otherwise
 753  * *caughtup is set to false.
 754  *
 755  * Returns true if OK, false if trouble.
 756  */
 757 static bool
 758 XLogSend(char *msgbuf, bool *caughtup)
 759 {
 760         XLogRecPtr      SendRqstPtr;
 761         XLogRecPtr      startptr;
 762         XLogRecPtr      endptr;
 763         Size            nbytes;
 764         WalDataMessageHeader msghdr;
 765
 766         /*
 767          * Attempt to send all data that's already been written out and fsync'd to
 768          * disk.  We cannot go further than what's been written out given the
 769          * current implementation of XLogRead().  And in any case it's unsafe to
 770          * send WAL that is not securely down to disk on the master: if the master
 771          * subsequently crashes and restarts, slaves must not have applied any WAL
 772          * that gets lost on the master.
 773          */
 774         SendRqstPtr = GetFlushRecPtr();
 775
 776         /* Quick exit if nothing to do */
 777         if (XLByteLE(SendRqstPtr, sentPtr))
 778         {
 779                 *caughtup = true;
 780                 return true;
 781         }
 782
 783         /*
 784          * Figure out how much to send in one message. If there's no more than
 785          * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
 786          * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
 787          *
 788          * The rounding is not only for performance reasons. Walreceiver relies on
 789          * the fact that we never split a WAL record across two messages. Since a
 790          * long WAL record is split at page boundary into continuation records,
 791          * page boundary is always a safe cut-off point. We also assume that
 792          * SendRqstPtr never points to the middle of a WAL record.
 793          */
 794         startptr = sentPtr;
 795         if (startptr.xrecoff >= XLogFileSize)
 796         {
 797                 /*
 798                  * crossing a logid boundary, skip the non-existent last log segment
 799                  * in previous logical log file.
 800                  */
 801                 startptr.xlogid += 1;
 802                 startptr.xrecoff = 0;
 803         }
 804
 805         endptr = startptr;
 806         XLByteAdvance(endptr, MAX_SEND_SIZE);
 807         if (endptr.xlogid != startptr.xlogid)
 808         {
 809                 /* Don't cross a logfile boundary within one message */
 810                 Assert(endptr.xlogid == startptr.xlogid + 1);
 811                 endptr.xlogid = startptr.xlogid;
 812                 endptr.xrecoff = XLogFileSize;
 813         }
 814
 815         /* if we went beyond SendRqstPtr, back off */
 816         if (XLByteLE(SendRqstPtr, endptr))
 817         {
 818                 endptr = SendRqstPtr;
 819                 *caughtup = true;
 820         }
 821         else
 822         {
 823                 /* round down to page boundary. */
 824                 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
 825                 *caughtup = false;
 826         }
 827
 828         nbytes = endptr.xrecoff - startptr.xrecoff;
 829         Assert(nbytes <= MAX_SEND_SIZE);
 830
 831         /*
 832          * OK to read and send the slice.
 833          */
 834         msgbuf[0] = 'w';
 835
 836         /*
 837          * Read the log directly into the output buffer to avoid extra memcpy
 838          * calls.
 839          */
 840         XLogRead(msgbuf + 1 + sizeof(WalDataMessageHeader), startptr, nbytes);
 841
 842         /*
 843          * We fill the message header last so that the send timestamp is taken as
 844          * late as possible.
 845          */
 846         msghdr.dataStart = startptr;
 847         msghdr.walEnd = SendRqstPtr;
 848         msghdr.sendTime = GetCurrentTimestamp();
 849
 850         memcpy(msgbuf + 1, &msghdr, sizeof(WalDataMessageHeader));
 851
 852         pq_putmessage('d', msgbuf, 1 + sizeof(WalDataMessageHeader) + nbytes);
 853
 854         /* Flush pending output to the client */
 855         if (pq_flush())
 856                 return false;
 857
 858         sentPtr = endptr;
 859
 860         /* Update shared memory status */
 861         {
 862                 /* use volatile pointer to prevent code rearrangement */
 863                 volatile WalSnd *walsnd = MyWalSnd;
 864
 865                 SpinLockAcquire(&walsnd->mutex);
 866                 walsnd->sentPtr = sentPtr;
 867                 SpinLockRelease(&walsnd->mutex);
 868         }
 869
 870         /* Report progress of XLOG streaming in PS display */
 871         if (update_process_title)
 872         {
 873                 char            activitymsg[50];
 874
 875                 snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
 876                                  sentPtr.xlogid, sentPtr.xrecoff);
 877                 set_ps_display(activitymsg, false);
 878         }
 879
 880         return true;
 881 }
 882
 883 /* SIGHUP: set flag to re-read config file at next convenient time */
 884 static void
 885 WalSndSigHupHandler(SIGNAL_ARGS)
 886 {
 887         got_SIGHUP = true;
 888         if (MyWalSnd)
 889                 SetLatch(&MyWalSnd->latch);
 890 }
 891
 892 /* SIGTERM: set flag to shut down */
 893 static void
 894 WalSndShutdownHandler(SIGNAL_ARGS)
 895 {
 896         shutdown_requested = true;
 897         if (MyWalSnd)
 898                 SetLatch(&MyWalSnd->latch);
 899 }
 900
 901 /*
 902  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
 903  *
 904  * Some backend has bought the farm,
 905  * so we need to stop what we're doing and exit.
 906  */
 907 static void
 908 WalSndQuickDieHandler(SIGNAL_ARGS)
 909 {
 910         PG_SETMASK(&BlockSig);
 911
 912         /*
 913          * We DO NOT want to run proc_exit() callbacks -- we're here because
 914          * shared memory may be corrupted, so we don't want to try to clean up our
 915          * transaction.  Just nail the windows shut and get out of town.  Now that
 916          * there's an atexit callback to prevent third-party code from breaking
 917          * things by calling exit() directly, we have to reset the callbacks
 918          * explicitly to make this work as intended.
 919          */
 920         on_exit_reset();
 921
 922         /*
 923          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
 924          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
 925          * backend.  This is necessary precisely because we don't clean up our
 926          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
 927          * should ensure the postmaster sees this as a crash, too, but no harm in
 928          * being doubly sure.)
 929          */
 930         exit(2);
 931 }
 932
 933 /* SIGUSR1: set flag to send WAL records */
 934 static void
 935 WalSndXLogSendHandler(SIGNAL_ARGS)
 936 {
 937         latch_sigusr1_handler();
 938 }
 939
 940 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
 941 static void
 942 WalSndLastCycleHandler(SIGNAL_ARGS)
 943 {
 944         ready_to_stop = true;
 945         if (MyWalSnd)
 946                 SetLatch(&MyWalSnd->latch);
 947 }
 948
 949 /* Set up signal handlers */
 950 void
 951 WalSndSignals(void)
 952 {
 953         /* Set up signal handlers */
 954         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
 955                                                                                                  * file */
 956         pqsignal(SIGINT, SIG_IGN);      /* not used */
 957         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
 958         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
 959         pqsignal(SIGALRM, SIG_IGN);
 960         pqsignal(SIGPIPE, SIG_IGN);
 961         pqsignal(SIGUSR1, WalSndXLogSendHandler);       /* request WAL sending */
 962         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
 963                                                                                                  * shutdown */
 964
 965         /* Reset some signals that are accepted by postmaster but not here */
 966         pqsignal(SIGCHLD, SIG_DFL);
 967         pqsignal(SIGTTIN, SIG_DFL);
 968         pqsignal(SIGTTOU, SIG_DFL);
 969         pqsignal(SIGCONT, SIG_DFL);
 970         pqsignal(SIGWINCH, SIG_DFL);
 971 }
 972
 973 /* Report shared-memory space needed by WalSndShmemInit */
 974 Size
 975 WalSndShmemSize(void)
 976 {
 977         Size            size = 0;
 978
 979         size = offsetof(WalSndCtlData, walsnds);
 980         size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
 981
 982         return size;
 983 }
 984
 985 /* Allocate and initialize walsender-related shared memory */
 986 void
 987 WalSndShmemInit(void)
 988 {
 989         bool            found;
 990         int                     i;
 991
 992         WalSndCtl = (WalSndCtlData *)
 993                 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
 994
 995         if (!found)
 996         {
 997                 /* First time through, so initialize */
 998                 MemSet(WalSndCtl, 0, WalSndShmemSize());
 999
1000                 for (i = 0; i < max_wal_senders; i++)
1001                 {
1002                         WalSnd     *walsnd = &WalSndCtl->walsnds[i];
1003
1004                         SpinLockInit(&walsnd->mutex);
1005                         InitSharedLatch(&walsnd->latch);
1006                 }
1007         }
1008 }
1009
1010 /* Wake up all walsenders */
1011 void
1012 WalSndWakeup(void)
1013 {
1014         int             i;
1015
1016         for (i = 0; i < max_wal_senders; i++)
1017                 SetLatch(&WalSndCtl->walsnds[i].latch);
1018 }
1019
1020 /*
1021  * This isn't currently used for anything. Monitoring tools might be
1022  * interested in the future, and we'll need something like this in the
1023  * future for synchronous replication.
1024  */
1025 #ifdef NOT_USED
1026 /*
1027  * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
1028  * if none.
1029  */
1030 XLogRecPtr
1031 GetOldestWALSendPointer(void)
1032 {
1033         XLogRecPtr      oldest = {0, 0};
1034         int                     i;
1035         bool            found = false;
1036
1037         for (i = 0; i < max_wal_senders; i++)
1038         {
1039                 /* use volatile pointer to prevent code rearrangement */
1040                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
1041                 XLogRecPtr      recptr;
1042
1043                 if (walsnd->pid == 0)
1044                         continue;
1045
1046                 SpinLockAcquire(&walsnd->mutex);
1047                 recptr = walsnd->sentPtr;
1048                 SpinLockRelease(&walsnd->mutex);
1049
1050                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
1051                         continue;
1052
1053                 if (!found || XLByteLT(recptr, oldest))
1054                         oldest = recptr;
1055                 found = true;
1056         }
1057         return oldest;
1058 }
1059
1060 #endif