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  * charge of XLOG streaming sender in the primary server. At first, it is
   7  * started by the postmaster when the walreceiver in the standby server
   8  * connects to the primary server and requests XLOG streaming replication,
   9  * i.e., unlike any auxiliary process, it is not an always-running process.
  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 an 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  * Note that there can be more than one walsender process concurrently.
  28  *
  29  * Portions Copyright (c) 2010-2010, PostgreSQL Global Development Group
  30  *
  31  *
  32  * IDENTIFICATION
  33  *        $PostgreSQL: pgsql/src/backend/replication/walsender.c,v 1.18 2010/04/28 16:10:42 heikki Exp $
  34  *
  35  *-------------------------------------------------------------------------
  36  */
  37 #include "postgres.h"
  38
  39 #include <unistd.h>
  40
  41 #include "access/xlog_internal.h"
  42 #include "catalog/pg_type.h"
  43 #include "libpq/libpq.h"
  44 #include "libpq/pqformat.h"
  45 #include "libpq/pqsignal.h"
  46 #include "miscadmin.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 #define NAPTIME_PER_CYCLE 100000L       /* max sleep time between cycles (100ms) */
  71
  72 /*
  73  * These variables are used similarly to openLogFile/Id/Seg/Off,
  74  * but for walsender to read the XLOG.
  75  */
  76 static int      sendFile = -1;
  77 static uint32 sendId = 0;
  78 static uint32 sendSeg = 0;
  79 static uint32 sendOff = 0;
  80
  81 /*
  82  * How far have we sent WAL already? This is also advertised in
  83  * MyWalSnd->sentPtr.
  84  */
  85 static XLogRecPtr sentPtr = {0, 0};
  86
  87 /* Flags set by signal handlers for later service in main loop */
  88 static volatile sig_atomic_t got_SIGHUP = false;
  89 static volatile sig_atomic_t shutdown_requested = false;
  90 static volatile sig_atomic_t ready_to_stop = false;
  91
  92 /* Signal handlers */
  93 static void WalSndSigHupHandler(SIGNAL_ARGS);
  94 static void WalSndShutdownHandler(SIGNAL_ARGS);
  95 static void WalSndQuickDieHandler(SIGNAL_ARGS);
  96
  97 /* Prototypes for private functions */
  98 static int      WalSndLoop(void);
  99 static void InitWalSnd(void);
 100 static void WalSndHandshake(void);
 101 static void WalSndKill(int code, Datum arg);
 102 static void XLogRead(char *buf, XLogRecPtr recptr, Size nbytes);
 103 static bool XLogSend(StringInfo outMsg);
 104 static void CheckClosedConnection(void);
 105
 106 /*
 107  * How much WAL to send in one message? Must be >= XLOG_BLCKSZ.
 108  */
 109 #define MAX_SEND_SIZE (XLOG_SEG_SIZE / 2)
 110
 111 /* Main entry point for walsender process */
 112 int
 113 WalSenderMain(void)
 114 {
 115         MemoryContext walsnd_context;
 116
 117         if (RecoveryInProgress())
 118                 ereport(FATAL,
 119                                 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 120                                  errmsg("recovery is still in progress, can't accept WAL streaming connections")));
 121
 122         /* Create a per-walsender data structure in shared memory */
 123         InitWalSnd();
 124
 125         /*
 126          * Create a memory context that we will do all our work in.  We do this so
 127          * that we can reset the context during error recovery and thereby avoid
 128          * possible memory leaks.  Formerly this code just ran in
 129          * TopMemoryContext, but resetting that would be a really bad idea.
 130          *
 131          * XXX: we don't actually attempt error recovery in walsender, we just
 132          * close the connection and exit.
 133          */
 134         walsnd_context = AllocSetContextCreate(TopMemoryContext,
 135                                                                                    "Wal Sender",
 136                                                                                    ALLOCSET_DEFAULT_MINSIZE,
 137                                                                                    ALLOCSET_DEFAULT_INITSIZE,
 138                                                                                    ALLOCSET_DEFAULT_MAXSIZE);
 139         MemoryContextSwitchTo(walsnd_context);
 140
 141         /* Unblock signals (they were blocked when the postmaster forked us) */
 142         PG_SETMASK(&UnBlockSig);
 143
 144         /* Tell the standby that walsender is ready for receiving commands */
 145         ReadyForQuery(DestRemote);
 146
 147         /* Handle handshake messages before streaming */
 148         WalSndHandshake();
 149
 150         /* Main loop of walsender */
 151         return WalSndLoop();
 152 }
 153
 154 static void
 155 WalSndHandshake(void)
 156 {
 157         StringInfoData input_message;
 158         bool            replication_started = false;
 159
 160         initStringInfo(&input_message);
 161
 162         while (!replication_started)
 163         {
 164                 int                     firstchar;
 165
 166                 /* Wait for a command to arrive */
 167                 firstchar = pq_getbyte();
 168
 169                 /*
 170                  * Check for any other interesting events that happened while we
 171                  * slept.
 172                  */
 173                 if (got_SIGHUP)
 174                 {
 175                         got_SIGHUP = false;
 176                         ProcessConfigFile(PGC_SIGHUP);
 177                 }
 178
 179                 if (firstchar != EOF)
 180                 {
 181                         /*
 182                          * Read the message contents. This is expected to be done without
 183                          * blocking because we've been able to get message type code.
 184                          */
 185                         if (pq_getmessage(&input_message, 0))
 186                                 firstchar = EOF;        /* suitable message already logged */
 187                 }
 188
 189                 /* Handle the very limited subset of commands expected in this phase */
 190                 switch (firstchar)
 191                 {
 192                         case 'Q':                       /* Query message */
 193                                 {
 194                                         const char *query_string;
 195                                         XLogRecPtr      recptr;
 196
 197                                         query_string = pq_getmsgstring(&input_message);
 198                                         pq_getmsgend(&input_message);
 199
 200                                         if (strcmp(query_string, "IDENTIFY_SYSTEM") == 0)
 201                                         {
 202                                                 StringInfoData buf;
 203                                                 char            sysid[32];
 204                                                 char            tli[11];
 205
 206                                                 /*
 207                                                  * Reply with a result set with one row, two columns.
 208                                                  * First col is system ID, and second if timeline ID
 209                                                  */
 210
 211                                                 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
 212                                                                  GetSystemIdentifier());
 213                                                 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
 214
 215                                                 /* Send a RowDescription message */
 216                                                 pq_beginmessage(&buf, 'T');
 217                                                 pq_sendint(&buf, 2, 2); /* 2 fields */
 218
 219                                                 /* first field */
 220                                                 pq_sendstring(&buf, "systemid");                /* col name */
 221                                                 pq_sendint(&buf, 0, 4); /* table oid */
 222                                                 pq_sendint(&buf, 0, 2); /* attnum */
 223                                                 pq_sendint(&buf, TEXTOID, 4);   /* type oid */
 224                                                 pq_sendint(&buf, -1, 2);                /* typlen */
 225                                                 pq_sendint(&buf, 0, 4); /* typmod */
 226                                                 pq_sendint(&buf, 0, 2); /* format code */
 227
 228                                                 /* second field */
 229                                                 pq_sendstring(&buf, "timeline");                /* col name */
 230                                                 pq_sendint(&buf, 0, 4); /* table oid */
 231                                                 pq_sendint(&buf, 0, 2); /* attnum */
 232                                                 pq_sendint(&buf, INT4OID, 4);   /* type oid */
 233                                                 pq_sendint(&buf, 4, 2); /* typlen */
 234                                                 pq_sendint(&buf, 0, 4); /* typmod */
 235                                                 pq_sendint(&buf, 0, 2); /* format code */
 236                                                 pq_endmessage(&buf);
 237
 238                                                 /* Send a DataRow message */
 239                                                 pq_beginmessage(&buf, 'D');
 240                                                 pq_sendint(&buf, 2, 2); /* # of columns */
 241                                                 pq_sendint(&buf, strlen(sysid), 4);             /* col1 len */
 242                                                 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
 243                                                 pq_sendint(&buf, strlen(tli), 4);               /* col2 len */
 244                                                 pq_sendbytes(&buf, (char *) tli, strlen(tli));
 245                                                 pq_endmessage(&buf);
 246
 247                                                 /* Send CommandComplete and ReadyForQuery messages */
 248                                                 EndCommand("SELECT", DestRemote);
 249                                                 ReadyForQuery(DestRemote);
 250                                         }
 251                                         else if (sscanf(query_string, "START_REPLICATION %X/%X",
 252                                                                         &recptr.xlogid, &recptr.xrecoff) == 2)
 253                                         {
 254                                                 StringInfoData buf;
 255
 256                                                 /*
 257                                                  * Check that we're logging enough information in the
 258                                                  * WAL for log-shipping.
 259                                                  *
 260                                                  * NOTE: This only checks the current value of
 261                                                  * wal_level. Even if the current setting is not
 262                                                  * 'minimal', there can be old WAL in the pg_xlog
 263                                                  * directory that was created with 'minimal'.
 264                                                  * So this is not bulletproof, the purpose is
 265                                                  * just to give a user-friendly error message that
 266                                                  * hints how to configure the system correctly.
 267                                                  */
 268                                                 if (wal_level == WAL_LEVEL_MINIMAL)
 269                                                         ereport(FATAL,
 270                                                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 271                                                                          errmsg("standby connections not allowed because wal_level='minimal'")));
 272
 273
 274                                                 /* Send a CopyOutResponse message, and start streaming */
 275                                                 pq_beginmessage(&buf, 'H');
 276                                                 pq_sendbyte(&buf, 0);
 277                                                 pq_sendint(&buf, 0, 2);
 278                                                 pq_endmessage(&buf);
 279                                                 pq_flush();
 280
 281                                                 /*
 282                                                  * Initialize position to the received one, then the
 283                                                  * xlog records begin to be shipped from that position
 284                                                  */
 285                                                 sentPtr = recptr;
 286
 287                                                 /* break out of the loop */
 288                                                 replication_started = true;
 289                                         }
 290                                         else
 291                                         {
 292                                                 ereport(FATAL,
 293                                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 294                                                                  errmsg("invalid standby query string: %s", query_string)));
 295                                         }
 296                                         break;
 297                                 }
 298
 299                         case 'X':
 300                                 /* standby is closing the connection */
 301                                 proc_exit(0);
 302
 303                         case EOF:
 304                                 /* standby disconnected unexpectedly */
 305                                 ereport(COMMERROR,
 306                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 307                                                  errmsg("unexpected EOF on standby connection")));
 308                                 proc_exit(0);
 309
 310                         default:
 311                                 ereport(FATAL,
 312                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 313                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 314                 }
 315         }
 316 }
 317
 318 /*
 319  * Check if the remote end has closed the connection.
 320  */
 321 static void
 322 CheckClosedConnection(void)
 323 {
 324         unsigned char firstchar;
 325         int                     r;
 326
 327         r = pq_getbyte_if_available(&firstchar);
 328         if (r < 0)
 329         {
 330                 /* unexpected error or EOF */
 331                 ereport(COMMERROR,
 332                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 333                                  errmsg("unexpected EOF on standby connection")));
 334                 proc_exit(0);
 335         }
 336         if (r == 0)
 337         {
 338                 /* no data available without blocking */
 339                 return;
 340         }
 341
 342         /* Handle the very limited subset of commands expected in this phase */
 343         switch (firstchar)
 344         {
 345                         /*
 346                          * 'X' means that the standby is closing down the socket.
 347                          */
 348                 case 'X':
 349                         proc_exit(0);
 350
 351                 default:
 352                         ereport(FATAL,
 353                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 354                                          errmsg("invalid standby closing message type %d",
 355                                                         firstchar)));
 356         }
 357 }
 358
 359 /* Main loop of walsender process */
 360 static int
 361 WalSndLoop(void)
 362 {
 363         StringInfoData output_message;
 364
 365         initStringInfo(&output_message);
 366
 367         /* Loop forever */
 368         for (;;)
 369         {
 370                 long    remain;         /* remaining time (us) */
 371
 372                 /*
 373                  * Emergency bailout if postmaster has died.  This is to avoid the
 374                  * necessity for manual cleanup of all postmaster children.
 375                  */
 376                 if (!PostmasterIsAlive(true))
 377                         exit(1);
 378                 /* Process any requests or signals received recently */
 379                 if (got_SIGHUP)
 380                 {
 381                         got_SIGHUP = false;
 382                         ProcessConfigFile(PGC_SIGHUP);
 383                 }
 384
 385                 /*
 386                  * When SIGUSR2 arrives, we send all outstanding logs up to the
 387                  * shutdown checkpoint record (i.e., the latest record) and exit.
 388                  */
 389                 if (ready_to_stop)
 390                 {
 391                         XLogSend(&output_message);
 392                         shutdown_requested = true;
 393                 }
 394
 395                 /* Normal exit from the walsender is here */
 396                 if (shutdown_requested)
 397                 {
 398                         /* Inform the standby that XLOG streaming was done */
 399                         pq_puttextmessage('C', "COPY 0");
 400                         pq_flush();
 401
 402                         proc_exit(0);
 403                 }
 404
 405                 /*
 406                  * Nap for the configured time or until a message arrives.
 407                  *
 408                  * On some platforms, signals won't interrupt the sleep.  To ensure we
 409                  * respond reasonably promptly when someone signals us, break down the
 410                  * sleep into NAPTIME_PER_CYCLE increments, and check for
 411                  * interrupts after each nap.
 412                  */
 413                 remain = WalSndDelay * 1000L;
 414                 while (remain > 0)
 415                 {
 416                         if (got_SIGHUP || shutdown_requested || ready_to_stop)
 417                                 break;
 418
 419                         /*
 420                          * Check to see whether a message from the standby or an interrupt
 421                          * from other processes has arrived.
 422                          */
 423                         pg_usleep(remain > NAPTIME_PER_CYCLE ? NAPTIME_PER_CYCLE : remain);
 424                         CheckClosedConnection();
 425
 426                         remain -= NAPTIME_PER_CYCLE;
 427                 }
 428
 429                 /* Attempt to send the log once every loop */
 430                 if (!XLogSend(&output_message))
 431                         goto eof;
 432         }
 433
 434         /* can't get here because the above loop never exits */
 435         return 1;
 436
 437 eof:
 438
 439         /*
 440          * Reset whereToSendOutput to prevent ereport from attempting to send any
 441          * more messages to the standby.
 442          */
 443         if (whereToSendOutput == DestRemote)
 444                 whereToSendOutput = DestNone;
 445
 446         proc_exit(0);
 447         return 1;                                       /* keep the compiler quiet */
 448 }
 449
 450 /* Initialize a per-walsender data structure for this walsender process */
 451 static void
 452 InitWalSnd(void)
 453 {
 454         /* use volatile pointer to prevent code rearrangement */
 455         int                     i;
 456
 457         /*
 458          * WalSndCtl should be set up already (we inherit this by fork() or
 459          * EXEC_BACKEND mechanism from the postmaster).
 460          */
 461         Assert(WalSndCtl != NULL);
 462         Assert(MyWalSnd == NULL);
 463
 464         /*
 465          * Find a free walsender slot and reserve it. If this fails, we must be
 466          * out of WalSnd structures.
 467          */
 468         for (i = 0; i < max_wal_senders; i++)
 469         {
 470                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 471
 472                 SpinLockAcquire(&walsnd->mutex);
 473
 474                 if (walsnd->pid != 0)
 475                 {
 476                         SpinLockRelease(&walsnd->mutex);
 477                         continue;
 478                 }
 479                 else
 480                 {
 481                         /* found */
 482                         MyWalSnd = (WalSnd *) walsnd;
 483                         walsnd->pid = MyProcPid;
 484                         MemSet(&MyWalSnd->sentPtr, 0, sizeof(XLogRecPtr));
 485                         SpinLockRelease(&walsnd->mutex);
 486                         break;
 487                 }
 488         }
 489         if (MyWalSnd == NULL)
 490                 ereport(FATAL,
 491                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 492                                  errmsg("number of requested standby connections "
 493                                         "exceeds max_wal_senders (currently %d)",
 494                                         max_wal_senders)));
 495
 496         /* Arrange to clean up at walsender exit */
 497         on_shmem_exit(WalSndKill, 0);
 498 }
 499
 500 /* Destroy the per-walsender data structure for this walsender process */
 501 static void
 502 WalSndKill(int code, Datum arg)
 503 {
 504         Assert(MyWalSnd != NULL);
 505
 506         /*
 507          * Mark WalSnd struct no longer in use. Assume that no lock is required
 508          * for this.
 509          */
 510         MyWalSnd->pid = 0;
 511
 512         /* WalSnd struct isn't mine anymore */
 513         MyWalSnd = NULL;
 514 }
 515
 516 /*
 517  * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
 518  */
 519 void
 520 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
 521 {
 522         XLogRecPtr      startRecPtr = recptr;
 523         char            path[MAXPGPATH];
 524         uint32          lastRemovedLog;
 525         uint32          lastRemovedSeg;
 526         uint32          log;
 527         uint32          seg;
 528
 529         while (nbytes > 0)
 530         {
 531                 uint32          startoff;
 532                 int                     segbytes;
 533                 int                     readbytes;
 534
 535                 startoff = recptr.xrecoff % XLogSegSize;
 536
 537                 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
 538                 {
 539                         /* Switch to another logfile segment */
 540                         if (sendFile >= 0)
 541                                 close(sendFile);
 542
 543                         XLByteToSeg(recptr, sendId, sendSeg);
 544                         XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
 545
 546                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
 547                         if (sendFile < 0)
 548                         {
 549                                 /*
 550                                  * If the file is not found, assume it's because the
 551                                  * standby asked for a too old WAL segment that has already
 552                                  * been removed or recycled.
 553                                  */
 554                                 if (errno == ENOENT)
 555                                 {
 556                                         char filename[MAXFNAMELEN];
 557                                         XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
 558                                         ereport(ERROR,
 559                                                         (errcode_for_file_access(),
 560                                                          errmsg("requested WAL segment %s has already been removed",
 561                                                                         filename)));
 562                                 }
 563                                 else
 564                                         ereport(ERROR,
 565                                                         (errcode_for_file_access(),
 566                                                          errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
 567                                                                         path, sendId, sendSeg)));
 568                         }
 569                         sendOff = 0;
 570                 }
 571
 572                 /* Need to seek in the file? */
 573                 if (sendOff != startoff)
 574                 {
 575                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
 576                                 ereport(ERROR,
 577                                                 (errcode_for_file_access(),
 578                                                  errmsg("could not seek in log file %u, segment %u to offset %u: %m",
 579                                                                 sendId, sendSeg, startoff)));
 580                         sendOff = startoff;
 581                 }
 582
 583                 /* How many bytes are within this segment? */
 584                 if (nbytes > (XLogSegSize - startoff))
 585                         segbytes = XLogSegSize - startoff;
 586                 else
 587                         segbytes = nbytes;
 588
 589                 readbytes = read(sendFile, buf, segbytes);
 590                 if (readbytes <= 0)
 591                         ereport(ERROR,
 592                                         (errcode_for_file_access(),
 593                         errmsg("could not read from log file %u, segment %u, offset %u, "
 594                                    "length %lu: %m",
 595                                    sendId, sendSeg, sendOff, (unsigned long) segbytes)));
 596
 597                 /* Update state for read */
 598                 XLByteAdvance(recptr, readbytes);
 599
 600                 sendOff += readbytes;
 601                 nbytes -= readbytes;
 602                 buf += readbytes;
 603         }
 604
 605         /*
 606          * After reading into the buffer, check that what we read was valid.
 607          * We do this after reading, because even though the segment was present
 608          * when we opened it, it might get recycled or removed while we read it.
 609          * The read() succeeds in that case, but the data we tried to read might
 610          * already have been overwritten with new WAL records.
 611          */
 612         XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
 613         XLByteToSeg(startRecPtr, log, seg);
 614         if (log < lastRemovedLog ||
 615                 (log == lastRemovedLog && seg <= lastRemovedSeg))
 616         {
 617                 char filename[MAXFNAMELEN];
 618                 XLogFileName(filename, ThisTimeLineID, log, seg);
 619                 ereport(ERROR,
 620                                 (errcode_for_file_access(),
 621                                  errmsg("requested WAL segment %s has already been removed",
 622                                                 filename)));
 623         }
 624 }
 625
 626 /*
 627  * Read all WAL that's been written (and flushed) since last cycle, and send
 628  * it to client.
 629  *
 630  * Returns true if OK, false if trouble.
 631  */
 632 static bool
 633 XLogSend(StringInfo outMsg)
 634 {
 635         XLogRecPtr      SendRqstPtr;
 636         char            activitymsg[50];
 637
 638         /* use volatile pointer to prevent code rearrangement */
 639         volatile WalSnd *walsnd = MyWalSnd;
 640
 641         /* Attempt to send all records flushed to the disk already */
 642         SendRqstPtr = GetWriteRecPtr();
 643
 644         /* Quick exit if nothing to do */
 645         if (!XLByteLT(sentPtr, SendRqstPtr))
 646                 return true;
 647
 648         /*
 649          * We gather multiple records together by issuing just one XLogRead() of a
 650          * suitable size, and send them as one CopyData message. Repeat until
 651          * we've sent everything we can.
 652          */
 653         while (XLByteLT(sentPtr, SendRqstPtr))
 654         {
 655                 XLogRecPtr      startptr;
 656                 XLogRecPtr      endptr;
 657                 Size            nbytes;
 658
 659                 /*
 660                  * Figure out how much to send in one message. If there's less than
 661                  * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
 662                  * MAX_SEND_SIZE bytes, but round to page boundary.
 663                  *
 664                  * The rounding is not only for performance reasons. Walreceiver
 665                  * relies on the fact that we never split a WAL record across two
 666                  * messages. Since a long WAL record is split at page boundary into
 667                  * continuation records, page boundary is always a safe cut-off point.
 668                  * We also assume that SendRqstPtr never points in the middle of a WAL
 669                  * record.
 670                  */
 671                 startptr = sentPtr;
 672                 if (startptr.xrecoff >= XLogFileSize)
 673                 {
 674                         /*
 675                          * crossing a logid boundary, skip the non-existent last log
 676                          * segment in previous logical log file.
 677                          */
 678                         startptr.xlogid += 1;
 679                         startptr.xrecoff = 0;
 680                 }
 681
 682                 endptr = startptr;
 683                 XLByteAdvance(endptr, MAX_SEND_SIZE);
 684                 /* round down to page boundary. */
 685                 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
 686                 /* if we went beyond SendRqstPtr, back off */
 687                 if (XLByteLT(SendRqstPtr, endptr))
 688                         endptr = SendRqstPtr;
 689
 690                 /*
 691                  * OK to read and send the slice.
 692                  *
 693                  * We don't need to convert the xlogid/xrecoff from host byte order to
 694                  * network byte order because the both server can be expected to have
 695                  * the same byte order. If they have different byte order, we don't
 696                  * reach here.
 697                  */
 698                 pq_sendbyte(outMsg, 'w');
 699                 pq_sendbytes(outMsg, (char *) &startptr, sizeof(startptr));
 700
 701                 if (endptr.xlogid != startptr.xlogid)
 702                 {
 703                         Assert(endptr.xlogid == startptr.xlogid + 1);
 704                         nbytes = endptr.xrecoff + XLogFileSize - startptr.xrecoff;
 705                 }
 706                 else
 707                         nbytes = endptr.xrecoff - startptr.xrecoff;
 708
 709                 sentPtr = endptr;
 710
 711                 /*
 712                  * Read the log directly into the output buffer to prevent extra
 713                  * memcpy calls.
 714                  */
 715                 enlargeStringInfo(outMsg, nbytes);
 716
 717                 XLogRead(&outMsg->data[outMsg->len], startptr, nbytes);
 718                 outMsg->len += nbytes;
 719                 outMsg->data[outMsg->len] = '\0';
 720
 721                 pq_putmessage('d', outMsg->data, outMsg->len);
 722                 resetStringInfo(outMsg);
 723         }
 724
 725         /* Update shared memory status */
 726         SpinLockAcquire(&walsnd->mutex);
 727         walsnd->sentPtr = sentPtr;
 728         SpinLockRelease(&walsnd->mutex);
 729
 730         /* Flush pending output */
 731         if (pq_flush())
 732                 return false;
 733
 734         /* Report progress of XLOG streaming in PS display */
 735         snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
 736                          sentPtr.xlogid, sentPtr.xrecoff);
 737         set_ps_display(activitymsg, false);
 738
 739         return true;
 740 }
 741
 742 /* SIGHUP: set flag to re-read config file at next convenient time */
 743 static void
 744 WalSndSigHupHandler(SIGNAL_ARGS)
 745 {
 746         got_SIGHUP = true;
 747 }
 748
 749 /* SIGTERM: set flag to shut down */
 750 static void
 751 WalSndShutdownHandler(SIGNAL_ARGS)
 752 {
 753         shutdown_requested = true;
 754 }
 755
 756 /*
 757  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
 758  *
 759  * Some backend has bought the farm,
 760  * so we need to stop what we're doing and exit.
 761  */
 762 static void
 763 WalSndQuickDieHandler(SIGNAL_ARGS)
 764 {
 765         PG_SETMASK(&BlockSig);
 766
 767         /*
 768          * We DO NOT want to run proc_exit() callbacks -- we're here because
 769          * shared memory may be corrupted, so we don't want to try to clean up our
 770          * transaction.  Just nail the windows shut and get out of town.  Now that
 771          * there's an atexit callback to prevent third-party code from breaking
 772          * things by calling exit() directly, we have to reset the callbacks
 773          * explicitly to make this work as intended.
 774          */
 775         on_exit_reset();
 776
 777         /*
 778          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
 779          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
 780          * backend.  This is necessary precisely because we don't clean up our
 781          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
 782          * should ensure the postmaster sees this as a crash, too, but no harm in
 783          * being doubly sure.)
 784          */
 785         exit(2);
 786 }
 787
 788 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
 789 static void
 790 WalSndLastCycleHandler(SIGNAL_ARGS)
 791 {
 792         ready_to_stop = true;
 793 }
 794
 795 /* Set up signal handlers */
 796 void
 797 WalSndSignals(void)
 798 {
 799         /* Set up signal handlers */
 800         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
 801                                                                                                  * file */
 802         pqsignal(SIGINT, SIG_IGN);      /* not used */
 803         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
 804         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
 805         pqsignal(SIGALRM, SIG_IGN);
 806         pqsignal(SIGPIPE, SIG_IGN);
 807         pqsignal(SIGUSR1, SIG_IGN); /* not used */
 808         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
 809                                                                                                  * shutdown */
 810
 811         /* Reset some signals that are accepted by postmaster but not here */
 812         pqsignal(SIGCHLD, SIG_DFL);
 813         pqsignal(SIGTTIN, SIG_DFL);
 814         pqsignal(SIGTTOU, SIG_DFL);
 815         pqsignal(SIGCONT, SIG_DFL);
 816         pqsignal(SIGWINCH, SIG_DFL);
 817 }
 818
 819 /* Report shared-memory space needed by WalSndShmemInit */
 820 Size
 821 WalSndShmemSize(void)
 822 {
 823         Size            size = 0;
 824
 825         size = offsetof(WalSndCtlData, walsnds);
 826         size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
 827
 828         return size;
 829 }
 830
 831 /* Allocate and initialize walsender-related shared memory */
 832 void
 833 WalSndShmemInit(void)
 834 {
 835         bool            found;
 836         int                     i;
 837
 838         WalSndCtl = (WalSndCtlData *)
 839                 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
 840
 841         if (WalSndCtl == NULL)
 842                 ereport(FATAL,
 843                                 (errcode(ERRCODE_OUT_OF_MEMORY),
 844                                  errmsg("not enough shared memory for walsender")));
 845         if (found)
 846                 return;                                 /* already initialized */
 847
 848         /* Initialize the data structures */
 849         MemSet(WalSndCtl, 0, WalSndShmemSize());
 850
 851         for (i = 0; i < max_wal_senders; i++)
 852         {
 853                 WalSnd     *walsnd = &WalSndCtl->walsnds[i];
 854
 855                 SpinLockInit(&walsnd->mutex);
 856         }
 857 }
 858
 859 /*
 860  * This isn't currently used for anything. Monitoring tools might be
 861  * interested in the future, and we'll need something like this in the
 862  * future for synchronous replication.
 863  */
 864 #ifdef NOT_USED
 865 /*
 866  * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
 867  * if none.
 868  */
 869 XLogRecPtr
 870 GetOldestWALSendPointer(void)
 871 {
 872         XLogRecPtr      oldest = {0, 0};
 873         int                     i;
 874         bool            found = false;
 875
 876         for (i = 0; i < max_wal_senders; i++)
 877         {
 878                 /* use volatile pointer to prevent code rearrangement */
 879                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 880                 XLogRecPtr      recptr;
 881
 882                 if (walsnd->pid == 0)
 883                         continue;
 884
 885                 SpinLockAcquire(&walsnd->mutex);
 886                 recptr = walsnd->sentPtr;
 887                 SpinLockRelease(&walsnd->mutex);
 888
 889                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
 890                         continue;
 891
 892                 if (!found || XLByteLT(recptr, oldest))
 893                         oldest = recptr;
 894                 found = true;
 895         }
 896         return oldest;
 897 }
 898 #endif