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  * 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 1.23 2010/05/31 10:44:37 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, bool *caughtup);
 104 static void CheckClosedConnection(void);
 105
 106 /*
 107  * How much WAL to send in one message? Must be >= XLOG_BLCKSZ.
 108  *
 109  * We don't have a good idea of what a good value would be; there's some
 110  * overhead per message in both walsender and walreceiver, but on the other
 111  * hand sending large batches makes walsender less responsive to signals
 112  * because signals are checked only between messages. 128kB (with
 113  * default 8k blocks) seems like a reasonable guess for now.
 114  */
 115 #define MAX_SEND_SIZE (XLOG_BLCKSZ * 16)
 116
 117 /* Main entry point for walsender process */
 118 int
 119 WalSenderMain(void)
 120 {
 121         MemoryContext walsnd_context;
 122
 123         if (RecoveryInProgress())
 124                 ereport(FATAL,
 125                                 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 126                                  errmsg("recovery is still in progress, can't accept WAL streaming connections")));
 127
 128         /* Create a per-walsender data structure in shared memory */
 129         InitWalSnd();
 130
 131         /*
 132          * Create a memory context that we will do all our work in.  We do this so
 133          * that we can reset the context during error recovery and thereby avoid
 134          * possible memory leaks.  Formerly this code just ran in
 135          * TopMemoryContext, but resetting that would be a really bad idea.
 136          *
 137          * XXX: we don't actually attempt error recovery in walsender, we just
 138          * close the connection and exit.
 139          */
 140         walsnd_context = AllocSetContextCreate(TopMemoryContext,
 141                                                                                    "Wal Sender",
 142                                                                                    ALLOCSET_DEFAULT_MINSIZE,
 143                                                                                    ALLOCSET_DEFAULT_INITSIZE,
 144                                                                                    ALLOCSET_DEFAULT_MAXSIZE);
 145         MemoryContextSwitchTo(walsnd_context);
 146
 147         /* Unblock signals (they were blocked when the postmaster forked us) */
 148         PG_SETMASK(&UnBlockSig);
 149
 150         /* Tell the standby that walsender is ready for receiving commands */
 151         ReadyForQuery(DestRemote);
 152
 153         /* Handle handshake messages before streaming */
 154         WalSndHandshake();
 155
 156         /* Main loop of walsender */
 157         return WalSndLoop();
 158 }
 159
 160 static void
 161 WalSndHandshake(void)
 162 {
 163         StringInfoData input_message;
 164         bool            replication_started = false;
 165
 166         initStringInfo(&input_message);
 167
 168         while (!replication_started)
 169         {
 170                 int                     firstchar;
 171
 172                 /* Wait for a command to arrive */
 173                 firstchar = pq_getbyte();
 174
 175                 /*
 176                  * Check for any other interesting events that happened while we
 177                  * slept.
 178                  */
 179                 if (got_SIGHUP)
 180                 {
 181                         got_SIGHUP = false;
 182                         ProcessConfigFile(PGC_SIGHUP);
 183                 }
 184
 185                 if (firstchar != EOF)
 186                 {
 187                         /*
 188                          * Read the message contents. This is expected to be done without
 189                          * blocking because we've been able to get message type code.
 190                          */
 191                         if (pq_getmessage(&input_message, 0))
 192                                 firstchar = EOF;        /* suitable message already logged */
 193                 }
 194
 195                 /* Handle the very limited subset of commands expected in this phase */
 196                 switch (firstchar)
 197                 {
 198                         case 'Q':                       /* Query message */
 199                                 {
 200                                         const char *query_string;
 201                                         XLogRecPtr      recptr;
 202
 203                                         query_string = pq_getmsgstring(&input_message);
 204                                         pq_getmsgend(&input_message);
 205
 206                                         if (strcmp(query_string, "IDENTIFY_SYSTEM") == 0)
 207                                         {
 208                                                 StringInfoData buf;
 209                                                 char            sysid[32];
 210                                                 char            tli[11];
 211
 212                                                 /*
 213                                                  * Reply with a result set with one row, two columns.
 214                                                  * First col is system ID, and second if timeline ID
 215                                                  */
 216
 217                                                 snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
 218                                                                  GetSystemIdentifier());
 219                                                 snprintf(tli, sizeof(tli), "%u", ThisTimeLineID);
 220
 221                                                 /* Send a RowDescription message */
 222                                                 pq_beginmessage(&buf, 'T');
 223                                                 pq_sendint(&buf, 2, 2); /* 2 fields */
 224
 225                                                 /* first field */
 226                                                 pq_sendstring(&buf, "systemid");                /* col name */
 227                                                 pq_sendint(&buf, 0, 4); /* table oid */
 228                                                 pq_sendint(&buf, 0, 2); /* attnum */
 229                                                 pq_sendint(&buf, TEXTOID, 4);   /* type oid */
 230                                                 pq_sendint(&buf, -1, 2);                /* typlen */
 231                                                 pq_sendint(&buf, 0, 4); /* typmod */
 232                                                 pq_sendint(&buf, 0, 2); /* format code */
 233
 234                                                 /* second field */
 235                                                 pq_sendstring(&buf, "timeline");                /* col name */
 236                                                 pq_sendint(&buf, 0, 4); /* table oid */
 237                                                 pq_sendint(&buf, 0, 2); /* attnum */
 238                                                 pq_sendint(&buf, INT4OID, 4);   /* type oid */
 239                                                 pq_sendint(&buf, 4, 2); /* typlen */
 240                                                 pq_sendint(&buf, 0, 4); /* typmod */
 241                                                 pq_sendint(&buf, 0, 2); /* format code */
 242                                                 pq_endmessage(&buf);
 243
 244                                                 /* Send a DataRow message */
 245                                                 pq_beginmessage(&buf, 'D');
 246                                                 pq_sendint(&buf, 2, 2); /* # of columns */
 247                                                 pq_sendint(&buf, strlen(sysid), 4);             /* col1 len */
 248                                                 pq_sendbytes(&buf, (char *) &sysid, strlen(sysid));
 249                                                 pq_sendint(&buf, strlen(tli), 4);               /* col2 len */
 250                                                 pq_sendbytes(&buf, (char *) tli, strlen(tli));
 251                                                 pq_endmessage(&buf);
 252
 253                                                 /* Send CommandComplete and ReadyForQuery messages */
 254                                                 EndCommand("SELECT", DestRemote);
 255                                                 ReadyForQuery(DestRemote);
 256                                         }
 257                                         else if (sscanf(query_string, "START_REPLICATION %X/%X",
 258                                                                         &recptr.xlogid, &recptr.xrecoff) == 2)
 259                                         {
 260                                                 StringInfoData buf;
 261
 262                                                 /*
 263                                                  * Check that we're logging enough information in the
 264                                                  * WAL for log-shipping.
 265                                                  *
 266                                                  * NOTE: This only checks the current value of
 267                                                  * wal_level. Even if the current setting is not
 268                                                  * 'minimal', there can be old WAL in the pg_xlog
 269                                                  * directory that was created with 'minimal'.
 270                                                  * So this is not bulletproof, the purpose is
 271                                                  * just to give a user-friendly error message that
 272                                                  * hints how to configure the system correctly.
 273                                                  */
 274                                                 if (wal_level == WAL_LEVEL_MINIMAL)
 275                                                         ereport(FATAL,
 276                                                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
 277                                                                          errmsg("standby connections not allowed because wal_level=\"minimal\"")));
 278
 279                                                 /* Send a CopyOutResponse message, and start streaming */
 280                                                 pq_beginmessage(&buf, 'H');
 281                                                 pq_sendbyte(&buf, 0);
 282                                                 pq_sendint(&buf, 0, 2);
 283                                                 pq_endmessage(&buf);
 284                                                 pq_flush();
 285
 286                                                 /*
 287                                                  * Initialize position to the received one, then the
 288                                                  * xlog records begin to be shipped from that position
 289                                                  */
 290                                                 sentPtr = recptr;
 291
 292                                                 /* break out of the loop */
 293                                                 replication_started = true;
 294                                         }
 295                                         else
 296                                         {
 297                                                 ereport(FATAL,
 298                                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 299                                                                  errmsg("invalid standby query string: %s", query_string)));
 300                                         }
 301                                         break;
 302                                 }
 303
 304                         case 'X':
 305                                 /* standby is closing the connection */
 306                                 proc_exit(0);
 307
 308                         case EOF:
 309                                 /* standby disconnected unexpectedly */
 310                                 ereport(COMMERROR,
 311                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 312                                                  errmsg("unexpected EOF on standby connection")));
 313                                 proc_exit(0);
 314
 315                         default:
 316                                 ereport(FATAL,
 317                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 318                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 319                 }
 320         }
 321 }
 322
 323 /*
 324  * Check if the remote end has closed the connection.
 325  */
 326 static void
 327 CheckClosedConnection(void)
 328 {
 329         unsigned char firstchar;
 330         int                     r;
 331
 332         r = pq_getbyte_if_available(&firstchar);
 333         if (r < 0)
 334         {
 335                 /* unexpected error or EOF */
 336                 ereport(COMMERROR,
 337                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 338                                  errmsg("unexpected EOF on standby connection")));
 339                 proc_exit(0);
 340         }
 341         if (r == 0)
 342         {
 343                 /* no data available without blocking */
 344                 return;
 345         }
 346
 347         /* Handle the very limited subset of commands expected in this phase */
 348         switch (firstchar)
 349         {
 350                         /*
 351                          * 'X' means that the standby is closing down the socket.
 352                          */
 353                 case 'X':
 354                         proc_exit(0);
 355
 356                 default:
 357                         ereport(FATAL,
 358                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 359                                          errmsg("invalid standby closing message type %d",
 360                                                         firstchar)));
 361         }
 362 }
 363
 364 /* Main loop of walsender process */
 365 static int
 366 WalSndLoop(void)
 367 {
 368         StringInfoData output_message;
 369         bool            caughtup = false;
 370
 371         initStringInfo(&output_message);
 372
 373         /* Loop forever */
 374         for (;;)
 375         {
 376                 long    remain;         /* remaining time (us) */
 377
 378                 /*
 379                  * Emergency bailout if postmaster has died.  This is to avoid the
 380                  * necessity for manual cleanup of all postmaster children.
 381                  */
 382                 if (!PostmasterIsAlive(true))
 383                         exit(1);
 384                 /* Process any requests or signals received recently */
 385                 if (got_SIGHUP)
 386                 {
 387                         got_SIGHUP = false;
 388                         ProcessConfigFile(PGC_SIGHUP);
 389                 }
 390
 391                 /*
 392                  * When SIGUSR2 arrives, we send all outstanding logs up to the
 393                  * shutdown checkpoint record (i.e., the latest record) and exit.
 394                  */
 395                 if (ready_to_stop)
 396                 {
 397                         if (!XLogSend(&output_message, &caughtup))
 398                                 goto eof;
 399                         if (caughtup)
 400                                 shutdown_requested = true;
 401                 }
 402
 403                 /* Normal exit from the walsender is here */
 404                 if (shutdown_requested)
 405                 {
 406                         /* Inform the standby that XLOG streaming was done */
 407                         pq_puttextmessage('C', "COPY 0");
 408                         pq_flush();
 409
 410                         proc_exit(0);
 411                 }
 412
 413                 /*
 414                  * If we had sent all accumulated WAL in last round, nap for the
 415                  * configured time before retrying.
 416                  *
 417                  * On some platforms, signals won't interrupt the sleep.  To ensure we
 418                  * respond reasonably promptly when someone signals us, break down the
 419                  * sleep into NAPTIME_PER_CYCLE increments, and check for
 420                  * interrupts after each nap.
 421                  */
 422                 if (caughtup)
 423                 {
 424                         remain = WalSndDelay * 1000L;
 425                         while (remain > 0)
 426                         {
 427                                 /* Check for interrupts */
 428                                 if (got_SIGHUP || shutdown_requested || ready_to_stop)
 429                                         break;
 430
 431                                 /* Sleep and check that the connection is still alive */
 432                                 pg_usleep(remain > NAPTIME_PER_CYCLE ? NAPTIME_PER_CYCLE : remain);
 433                                 CheckClosedConnection();
 434
 435                                 remain -= NAPTIME_PER_CYCLE;
 436                         }
 437                 }
 438                 /* Attempt to send the log once every loop */
 439                 if (!XLogSend(&output_message, &caughtup))
 440                         goto eof;
 441         }
 442
 443         /* can't get here because the above loop never exits */
 444         return 1;
 445
 446 eof:
 447
 448         /*
 449          * Reset whereToSendOutput to prevent ereport from attempting to send any
 450          * more messages to the standby.
 451          */
 452         if (whereToSendOutput == DestRemote)
 453                 whereToSendOutput = DestNone;
 454
 455         proc_exit(0);
 456         return 1;                                       /* keep the compiler quiet */
 457 }
 458
 459 /* Initialize a per-walsender data structure for this walsender process */
 460 static void
 461 InitWalSnd(void)
 462 {
 463         int                     i;
 464
 465         /*
 466          * WalSndCtl should be set up already (we inherit this by fork() or
 467          * EXEC_BACKEND mechanism from the postmaster).
 468          */
 469         Assert(WalSndCtl != NULL);
 470         Assert(MyWalSnd == NULL);
 471
 472         /*
 473          * Find a free walsender slot and reserve it. If this fails, we must be
 474          * out of WalSnd structures.
 475          */
 476         for (i = 0; i < max_wal_senders; i++)
 477         {
 478                 /* use volatile pointer to prevent code rearrangement */
 479                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 480
 481                 SpinLockAcquire(&walsnd->mutex);
 482
 483                 if (walsnd->pid != 0)
 484                 {
 485                         SpinLockRelease(&walsnd->mutex);
 486                         continue;
 487                 }
 488                 else
 489                 {
 490                         /* found */
 491                         MyWalSnd = (WalSnd *) walsnd;
 492                         walsnd->pid = MyProcPid;
 493                         MemSet(&MyWalSnd->sentPtr, 0, sizeof(XLogRecPtr));
 494                         SpinLockRelease(&walsnd->mutex);
 495                         break;
 496                 }
 497         }
 498         if (MyWalSnd == NULL)
 499                 ereport(FATAL,
 500                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 501                                  errmsg("number of requested standby connections "
 502                                         "exceeds max_wal_senders (currently %d)",
 503                                         max_wal_senders)));
 504
 505         /* Arrange to clean up at walsender exit */
 506         on_shmem_exit(WalSndKill, 0);
 507 }
 508
 509 /* Destroy the per-walsender data structure for this walsender process */
 510 static void
 511 WalSndKill(int code, Datum arg)
 512 {
 513         Assert(MyWalSnd != NULL);
 514
 515         /*
 516          * Mark WalSnd struct no longer in use. Assume that no lock is required
 517          * for this.
 518          */
 519         MyWalSnd->pid = 0;
 520
 521         /* WalSnd struct isn't mine anymore */
 522         MyWalSnd = NULL;
 523 }
 524
 525 /*
 526  * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
 527  */
 528 static void
 529 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
 530 {
 531         XLogRecPtr      startRecPtr = recptr;
 532         char            path[MAXPGPATH];
 533         uint32          lastRemovedLog;
 534         uint32          lastRemovedSeg;
 535         uint32          log;
 536         uint32          seg;
 537
 538         while (nbytes > 0)
 539         {
 540                 uint32          startoff;
 541                 int                     segbytes;
 542                 int                     readbytes;
 543
 544                 startoff = recptr.xrecoff % XLogSegSize;
 545
 546                 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
 547                 {
 548                         /* Switch to another logfile segment */
 549                         if (sendFile >= 0)
 550                                 close(sendFile);
 551
 552                         XLByteToSeg(recptr, sendId, sendSeg);
 553                         XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
 554
 555                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
 556                         if (sendFile < 0)
 557                         {
 558                                 /*
 559                                  * If the file is not found, assume it's because the
 560                                  * standby asked for a too old WAL segment that has already
 561                                  * been removed or recycled.
 562                                  */
 563                                 if (errno == ENOENT)
 564                                 {
 565                                         char filename[MAXFNAMELEN];
 566                                         XLogFileName(filename, ThisTimeLineID, sendId, sendSeg);
 567                                         ereport(ERROR,
 568                                                         (errcode_for_file_access(),
 569                                                          errmsg("requested WAL segment %s has already been removed",
 570                                                                         filename)));
 571                                 }
 572                                 else
 573                                         ereport(ERROR,
 574                                                         (errcode_for_file_access(),
 575                                                          errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
 576                                                                         path, sendId, sendSeg)));
 577                         }
 578                         sendOff = 0;
 579                 }
 580
 581                 /* Need to seek in the file? */
 582                 if (sendOff != startoff)
 583                 {
 584                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
 585                                 ereport(ERROR,
 586                                                 (errcode_for_file_access(),
 587                                                  errmsg("could not seek in log file %u, segment %u to offset %u: %m",
 588                                                                 sendId, sendSeg, startoff)));
 589                         sendOff = startoff;
 590                 }
 591
 592                 /* How many bytes are within this segment? */
 593                 if (nbytes > (XLogSegSize - startoff))
 594                         segbytes = XLogSegSize - startoff;
 595                 else
 596                         segbytes = nbytes;
 597
 598                 readbytes = read(sendFile, buf, segbytes);
 599                 if (readbytes <= 0)
 600                         ereport(ERROR,
 601                                         (errcode_for_file_access(),
 602                         errmsg("could not read from log file %u, segment %u, offset %u, "
 603                                    "length %lu: %m",
 604                                    sendId, sendSeg, sendOff, (unsigned long) segbytes)));
 605
 606                 /* Update state for read */
 607                 XLByteAdvance(recptr, readbytes);
 608
 609                 sendOff += readbytes;
 610                 nbytes -= readbytes;
 611                 buf += readbytes;
 612         }
 613
 614         /*
 615          * After reading into the buffer, check that what we read was valid.
 616          * We do this after reading, because even though the segment was present
 617          * when we opened it, it might get recycled or removed while we read it.
 618          * The read() succeeds in that case, but the data we tried to read might
 619          * already have been overwritten with new WAL records.
 620          */
 621         XLogGetLastRemoved(&lastRemovedLog, &lastRemovedSeg);
 622         XLByteToSeg(startRecPtr, log, seg);
 623         if (log < lastRemovedLog ||
 624                 (log == lastRemovedLog && seg <= lastRemovedSeg))
 625         {
 626                 char filename[MAXFNAMELEN];
 627                 XLogFileName(filename, ThisTimeLineID, log, seg);
 628                 ereport(ERROR,
 629                                 (errcode_for_file_access(),
 630                                  errmsg("requested WAL segment %s has already been removed",
 631                                                 filename)));
 632         }
 633 }
 634
 635 /*
 636  * Read up to MAX_SEND_SIZE bytes of WAL that's been written (and flushed),
 637  * but not yet sent to the client, and send it. If there is no unsent WAL,
 638  * *caughtup is set to true and nothing is sent, otherwise *caughtup is set
 639  * to false.
 640  *
 641  * Returns true if OK, false if trouble.
 642  */
 643 static bool
 644 XLogSend(StringInfo outMsg, bool *caughtup)
 645 {
 646         XLogRecPtr      SendRqstPtr;
 647         XLogRecPtr      startptr;
 648         XLogRecPtr      endptr;
 649         Size            nbytes;
 650         char            activitymsg[50];
 651
 652         /* use volatile pointer to prevent code rearrangement */
 653         volatile WalSnd *walsnd = MyWalSnd;
 654
 655         /* Attempt to send all records flushed to the disk already */
 656         SendRqstPtr = GetWriteRecPtr();
 657
 658         /* Quick exit if nothing to do */
 659         if (!XLByteLT(sentPtr, SendRqstPtr))
 660         {
 661                 *caughtup = true;
 662                 return true;
 663         }
 664         /*
 665          * Otherwise let the caller know that we're not fully caught up. Unless
 666          * there's a huge backlog, we'll be caught up to the current WriteRecPtr
 667          * after we've sent everything below, but more WAL could accumulate while
 668          * we're busy sending.
 669          */
 670         *caughtup = false;
 671
 672         /*
 673          * Figure out how much to send in one message. If there's less than
 674          * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
 675          * MAX_SEND_SIZE bytes, but round to page boundary.
 676          *
 677          * The rounding is not only for performance reasons. Walreceiver
 678          * relies on the fact that we never split a WAL record across two
 679          * messages. Since a long WAL record is split at page boundary into
 680          * continuation records, page boundary is always a safe cut-off point.
 681          * We also assume that SendRqstPtr never points in the middle of a WAL
 682          * record.
 683          */
 684         startptr = sentPtr;
 685         if (startptr.xrecoff >= XLogFileSize)
 686         {
 687                 /*
 688                  * crossing a logid boundary, skip the non-existent last log
 689                  * segment in previous logical log file.
 690                  */
 691                 startptr.xlogid += 1;
 692                 startptr.xrecoff = 0;
 693         }
 694
 695         endptr = startptr;
 696         XLByteAdvance(endptr, MAX_SEND_SIZE);
 697         /* round down to page boundary. */
 698         endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
 699         /* if we went beyond SendRqstPtr, back off */
 700         if (XLByteLT(SendRqstPtr, endptr))
 701                 endptr = SendRqstPtr;
 702
 703         /*
 704          * OK to read and send the slice.
 705          *
 706          * We don't need to convert the xlogid/xrecoff from host byte order to
 707          * network byte order because the both server can be expected to have
 708          * the same byte order. If they have different byte order, we don't
 709          * reach here.
 710          */
 711         pq_sendbyte(outMsg, 'w');
 712         pq_sendbytes(outMsg, (char *) &startptr, sizeof(startptr));
 713
 714         if (endptr.xlogid != startptr.xlogid)
 715         {
 716                 Assert(endptr.xlogid == startptr.xlogid + 1);
 717                 nbytes = endptr.xrecoff + XLogFileSize - startptr.xrecoff;
 718         }
 719         else
 720                 nbytes = endptr.xrecoff - startptr.xrecoff;
 721
 722         sentPtr = endptr;
 723
 724         /*
 725          * Read the log directly into the output buffer to prevent extra
 726          * memcpy calls.
 727          */
 728         enlargeStringInfo(outMsg, nbytes);
 729
 730         XLogRead(&outMsg->data[outMsg->len], startptr, nbytes);
 731         outMsg->len += nbytes;
 732         outMsg->data[outMsg->len] = '\0';
 733
 734         pq_putmessage('d', outMsg->data, outMsg->len);
 735         resetStringInfo(outMsg);
 736
 737         /* Update shared memory status */
 738         SpinLockAcquire(&walsnd->mutex);
 739         walsnd->sentPtr = sentPtr;
 740         SpinLockRelease(&walsnd->mutex);
 741
 742         /* Flush pending output */
 743         if (pq_flush())
 744                 return false;
 745
 746         /* Report progress of XLOG streaming in PS display */
 747         snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
 748                          sentPtr.xlogid, sentPtr.xrecoff);
 749         set_ps_display(activitymsg, false);
 750
 751         return true;
 752 }
 753
 754 /* SIGHUP: set flag to re-read config file at next convenient time */
 755 static void
 756 WalSndSigHupHandler(SIGNAL_ARGS)
 757 {
 758         got_SIGHUP = true;
 759 }
 760
 761 /* SIGTERM: set flag to shut down */
 762 static void
 763 WalSndShutdownHandler(SIGNAL_ARGS)
 764 {
 765         shutdown_requested = true;
 766 }
 767
 768 /*
 769  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
 770  *
 771  * Some backend has bought the farm,
 772  * so we need to stop what we're doing and exit.
 773  */
 774 static void
 775 WalSndQuickDieHandler(SIGNAL_ARGS)
 776 {
 777         PG_SETMASK(&BlockSig);
 778
 779         /*
 780          * We DO NOT want to run proc_exit() callbacks -- we're here because
 781          * shared memory may be corrupted, so we don't want to try to clean up our
 782          * transaction.  Just nail the windows shut and get out of town.  Now that
 783          * there's an atexit callback to prevent third-party code from breaking
 784          * things by calling exit() directly, we have to reset the callbacks
 785          * explicitly to make this work as intended.
 786          */
 787         on_exit_reset();
 788
 789         /*
 790          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
 791          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
 792          * backend.  This is necessary precisely because we don't clean up our
 793          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
 794          * should ensure the postmaster sees this as a crash, too, but no harm in
 795          * being doubly sure.)
 796          */
 797         exit(2);
 798 }
 799
 800 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
 801 static void
 802 WalSndLastCycleHandler(SIGNAL_ARGS)
 803 {
 804         ready_to_stop = true;
 805 }
 806
 807 /* Set up signal handlers */
 808 void
 809 WalSndSignals(void)
 810 {
 811         /* Set up signal handlers */
 812         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
 813                                                                                                  * file */
 814         pqsignal(SIGINT, SIG_IGN);      /* not used */
 815         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
 816         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
 817         pqsignal(SIGALRM, SIG_IGN);
 818         pqsignal(SIGPIPE, SIG_IGN);
 819         pqsignal(SIGUSR1, SIG_IGN); /* not used */
 820         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
 821                                                                                                  * shutdown */
 822
 823         /* Reset some signals that are accepted by postmaster but not here */
 824         pqsignal(SIGCHLD, SIG_DFL);
 825         pqsignal(SIGTTIN, SIG_DFL);
 826         pqsignal(SIGTTOU, SIG_DFL);
 827         pqsignal(SIGCONT, SIG_DFL);
 828         pqsignal(SIGWINCH, SIG_DFL);
 829 }
 830
 831 /* Report shared-memory space needed by WalSndShmemInit */
 832 Size
 833 WalSndShmemSize(void)
 834 {
 835         Size            size = 0;
 836
 837         size = offsetof(WalSndCtlData, walsnds);
 838         size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
 839
 840         return size;
 841 }
 842
 843 /* Allocate and initialize walsender-related shared memory */
 844 void
 845 WalSndShmemInit(void)
 846 {
 847         bool            found;
 848         int                     i;
 849
 850         WalSndCtl = (WalSndCtlData *)
 851                 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
 852
 853         if (!found)
 854         {
 855                 /* First time through, so initialize */
 856                 MemSet(WalSndCtl, 0, WalSndShmemSize());
 857
 858                 for (i = 0; i < max_wal_senders; i++)
 859                 {
 860                         WalSnd     *walsnd = &WalSndCtl->walsnds[i];
 861
 862                         SpinLockInit(&walsnd->mutex);
 863                 }
 864         }
 865 }
 866
 867 /*
 868  * This isn't currently used for anything. Monitoring tools might be
 869  * interested in the future, and we'll need something like this in the
 870  * future for synchronous replication.
 871  */
 872 #ifdef NOT_USED
 873 /*
 874  * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
 875  * if none.
 876  */
 877 XLogRecPtr
 878 GetOldestWALSendPointer(void)
 879 {
 880         XLogRecPtr      oldest = {0, 0};
 881         int                     i;
 882         bool            found = false;
 883
 884         for (i = 0; i < max_wal_senders; i++)
 885         {
 886                 /* use volatile pointer to prevent code rearrangement */
 887                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 888                 XLogRecPtr      recptr;
 889
 890                 if (walsnd->pid == 0)
 891                         continue;
 892
 893                 SpinLockAcquire(&walsnd->mutex);
 894                 recptr = walsnd->sentPtr;
 895                 SpinLockRelease(&walsnd->mutex);
 896
 897                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
 898                         continue;
 899
 900                 if (!found || XLByteLT(recptr, oldest))
 901                         oldest = recptr;
 902                 found = true;
 903         }
 904         return oldest;
 905 }
 906 #endif