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 1.9 2010/02/26 02:00:58 momjian 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/lock.h"
  51 #include "storage/pmsignal.h"
  52 #include "tcop/tcopprot.h"
  53 #include "utils/guc.h"
  54 #include "utils/memutils.h"
  55 #include "utils/ps_status.h"
  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                     MaxWalSenders = 0;      /* the maximum number of concurrent walsenders */
  68 int                     WalSndDelay = 200;      /* max sleep time between some actions */
  69
  70 #define NAPTIME_PER_CYCLE 100   /* 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 (!superuser())
 118                 ereport(FATAL,
 119                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
 120                                  errmsg("must be superuser to start walsender")));
 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                                                 /* Send a CopyOutResponse message, and start streaming */
 257                                                 pq_beginmessage(&buf, 'H');
 258                                                 pq_sendbyte(&buf, 0);
 259                                                 pq_sendint(&buf, 0, 2);
 260                                                 pq_endmessage(&buf);
 261
 262                                                 /*
 263                                                  * Initialize position to the received one, then the
 264                                                  * xlog records begin to be shipped from that position
 265                                                  */
 266                                                 sentPtr = recptr;
 267
 268                                                 /* break out of the loop */
 269                                                 replication_started = true;
 270                                         }
 271                                         else
 272                                         {
 273                                                 ereport(FATAL,
 274                                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 275                                                                  errmsg("invalid standby query string: %s", query_string)));
 276                                         }
 277                                         break;
 278                                 }
 279
 280                         case 'X':
 281                                 /* standby is closing the connection */
 282                                 proc_exit(0);
 283
 284                         case EOF:
 285                                 /* standby disconnected unexpectedly */
 286                                 ereport(COMMERROR,
 287                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 288                                                  errmsg("unexpected EOF on standby connection")));
 289                                 proc_exit(0);
 290
 291                         default:
 292                                 ereport(FATAL,
 293                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 294                                                  errmsg("invalid standby handshake message type %d", firstchar)));
 295                 }
 296         }
 297 }
 298
 299 /*
 300  * Check if the remote end has closed the connection.
 301  */
 302 static void
 303 CheckClosedConnection(void)
 304 {
 305         unsigned char firstchar;
 306         int                     r;
 307
 308         r = pq_getbyte_if_available(&firstchar);
 309         if (r < 0)
 310         {
 311                 /* unexpected error or EOF */
 312                 ereport(COMMERROR,
 313                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
 314                                  errmsg("unexpected EOF on standby connection")));
 315                 proc_exit(0);
 316         }
 317         if (r == 0)
 318         {
 319                 /* no data available without blocking */
 320                 return;
 321         }
 322
 323         /* Handle the very limited subset of commands expected in this phase */
 324         switch (firstchar)
 325         {
 326                         /*
 327                          * 'X' means that the standby is closing down the socket.
 328                          */
 329                 case 'X':
 330                         proc_exit(0);
 331
 332                 default:
 333                         ereport(FATAL,
 334                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
 335                                          errmsg("invalid standby closing message type %d",
 336                                                         firstchar)));
 337         }
 338 }
 339
 340 /* Main loop of walsender process */
 341 static int
 342 WalSndLoop(void)
 343 {
 344         StringInfoData output_message;
 345
 346         initStringInfo(&output_message);
 347
 348         /* Loop forever */
 349         for (;;)
 350         {
 351                 int                     remain;         /* remaining time (ms) */
 352
 353                 /*
 354                  * Emergency bailout if postmaster has died.  This is to avoid the
 355                  * necessity for manual cleanup of all postmaster children.
 356                  */
 357                 if (!PostmasterIsAlive(true))
 358                         exit(1);
 359                 /* Process any requests or signals received recently */
 360                 if (got_SIGHUP)
 361                 {
 362                         got_SIGHUP = false;
 363                         ProcessConfigFile(PGC_SIGHUP);
 364                 }
 365
 366                 /*
 367                  * When SIGUSR2 arrives, we send all outstanding logs up to the
 368                  * shutdown checkpoint record (i.e., the latest record) and exit.
 369                  */
 370                 if (ready_to_stop)
 371                 {
 372                         XLogSend(&output_message);
 373                         shutdown_requested = true;
 374                 }
 375
 376                 /* Normal exit from the walsender is here */
 377                 if (shutdown_requested)
 378                 {
 379                         /* Inform the standby that XLOG streaming was done */
 380                         pq_puttextmessage('C', "COPY 0");
 381                         pq_flush();
 382
 383                         proc_exit(0);
 384                 }
 385
 386                 /*
 387                  * Nap for the configured time or until a message arrives.
 388                  *
 389                  * On some platforms, signals won't interrupt the sleep.  To ensure we
 390                  * respond reasonably promptly when someone signals us, break down the
 391                  * sleep into NAPTIME_PER_CYCLE (ms) increments, and check for
 392                  * interrupts after each nap.
 393                  */
 394                 remain = WalSndDelay;
 395                 while (remain > 0)
 396                 {
 397                         if (got_SIGHUP || shutdown_requested || ready_to_stop)
 398                                 break;
 399
 400                         /*
 401                          * Check to see whether a message from the standby or an interrupt
 402                          * from other processes has arrived.
 403                          */
 404                         pg_usleep(remain > NAPTIME_PER_CYCLE ? NAPTIME_PER_CYCLE : remain);
 405                         CheckClosedConnection();
 406
 407                         remain -= NAPTIME_PER_CYCLE;
 408                 }
 409
 410                 /* Attempt to send the log once every loop */
 411                 if (!XLogSend(&output_message))
 412                         goto eof;
 413         }
 414
 415         /* can't get here because the above loop never exits */
 416         return 1;
 417
 418 eof:
 419
 420         /*
 421          * Reset whereToSendOutput to prevent ereport from attempting to send any
 422          * more messages to the standby.
 423          */
 424         if (whereToSendOutput == DestRemote)
 425                 whereToSendOutput = DestNone;
 426
 427         proc_exit(0);
 428         return 1;                                       /* keep the compiler quiet */
 429 }
 430
 431 /* Initialize a per-walsender data structure for this walsender process */
 432 static void
 433 InitWalSnd(void)
 434 {
 435         /* use volatile pointer to prevent code rearrangement */
 436         int                     i;
 437
 438         /*
 439          * WalSndCtl should be set up already (we inherit this by fork() or
 440          * EXEC_BACKEND mechanism from the postmaster).
 441          */
 442         Assert(WalSndCtl != NULL);
 443         Assert(MyWalSnd == NULL);
 444
 445         /*
 446          * Find a free walsender slot and reserve it. If this fails, we must be
 447          * out of WalSnd structures.
 448          */
 449         for (i = 0; i < MaxWalSenders; i++)
 450         {
 451                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 452
 453                 SpinLockAcquire(&walsnd->mutex);
 454
 455                 if (walsnd->pid != 0)
 456                 {
 457                         SpinLockRelease(&walsnd->mutex);
 458                         continue;
 459                 }
 460                 else
 461                 {
 462                         /* found */
 463                         MyWalSnd = (WalSnd *) walsnd;
 464                         walsnd->pid = MyProcPid;
 465                         MemSet(&MyWalSnd->sentPtr, 0, sizeof(XLogRecPtr));
 466                         SpinLockRelease(&walsnd->mutex);
 467                         break;
 468                 }
 469         }
 470         if (MyWalSnd == NULL)
 471                 ereport(FATAL,
 472                                 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
 473                                  errmsg("sorry, too many standbys already")));
 474
 475         /* Arrange to clean up at walsender exit */
 476         on_shmem_exit(WalSndKill, 0);
 477 }
 478
 479 /* Destroy the per-walsender data structure for this walsender process */
 480 static void
 481 WalSndKill(int code, Datum arg)
 482 {
 483         Assert(MyWalSnd != NULL);
 484
 485         /*
 486          * Mark WalSnd struct no longer in use. Assume that no lock is required
 487          * for this.
 488          */
 489         MyWalSnd->pid = 0;
 490
 491         /* WalSnd struct isn't mine anymore */
 492         MyWalSnd = NULL;
 493 }
 494
 495 /*
 496  * Read 'nbytes' bytes from WAL into 'buf', starting at location 'recptr'
 497  */
 498 void
 499 XLogRead(char *buf, XLogRecPtr recptr, Size nbytes)
 500 {
 501         char            path[MAXPGPATH];
 502         uint32          startoff;
 503
 504         while (nbytes > 0)
 505         {
 506                 int                     segbytes;
 507                 int                     readbytes;
 508
 509                 startoff = recptr.xrecoff % XLogSegSize;
 510
 511                 if (sendFile < 0 || !XLByteInSeg(recptr, sendId, sendSeg))
 512                 {
 513                         /* Switch to another logfile segment */
 514                         if (sendFile >= 0)
 515                                 close(sendFile);
 516
 517                         XLByteToSeg(recptr, sendId, sendSeg);
 518                         XLogFilePath(path, ThisTimeLineID, sendId, sendSeg);
 519
 520                         sendFile = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
 521                         if (sendFile < 0)
 522                                 ereport(FATAL,  /* XXX: Why FATAL? */
 523                                                 (errcode_for_file_access(),
 524                                                  errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
 525                                                                 path, sendId, sendSeg)));
 526                         sendOff = 0;
 527                 }
 528
 529                 /* Need to seek in the file? */
 530                 if (sendOff != startoff)
 531                 {
 532                         if (lseek(sendFile, (off_t) startoff, SEEK_SET) < 0)
 533                                 ereport(FATAL,
 534                                                 (errcode_for_file_access(),
 535                                                  errmsg("could not seek in log file %u, segment %u to offset %u: %m",
 536                                                                 sendId, sendSeg, startoff)));
 537                         sendOff = startoff;
 538                 }
 539
 540                 /* How many bytes are within this segment? */
 541                 if (nbytes > (XLogSegSize - startoff))
 542                         segbytes = XLogSegSize - startoff;
 543                 else
 544                         segbytes = nbytes;
 545
 546                 readbytes = read(sendFile, buf, segbytes);
 547                 if (readbytes <= 0)
 548                         ereport(FATAL,
 549                                         (errcode_for_file_access(),
 550                         errmsg("could not read from log file %u, segment %u, offset %u, "
 551                                    "length %lu: %m",
 552                                    sendId, sendSeg, sendOff, (unsigned long) segbytes)));
 553
 554                 /* Update state for read */
 555                 XLByteAdvance(recptr, readbytes);
 556
 557                 sendOff += readbytes;
 558                 nbytes -= readbytes;
 559                 buf += readbytes;
 560         }
 561 }
 562
 563 /*
 564  * Read all WAL that's been written (and flushed) since last cycle, and send
 565  * it to client.
 566  *
 567  * Returns true if OK, false if trouble.
 568  */
 569 static bool
 570 XLogSend(StringInfo outMsg)
 571 {
 572         XLogRecPtr      SendRqstPtr;
 573         char            activitymsg[50];
 574
 575         /* use volatile pointer to prevent code rearrangement */
 576         volatile WalSnd *walsnd = MyWalSnd;
 577
 578         /* Attempt to send all records flushed to the disk already */
 579         SendRqstPtr = GetWriteRecPtr();
 580
 581         /* Quick exit if nothing to do */
 582         if (!XLByteLT(sentPtr, SendRqstPtr))
 583                 return true;
 584
 585         /*
 586          * We gather multiple records together by issuing just one XLogRead() of a
 587          * suitable size, and send them as one CopyData message. Repeat until
 588          * we've sent everything we can.
 589          */
 590         while (XLByteLT(sentPtr, SendRqstPtr))
 591         {
 592                 XLogRecPtr      startptr;
 593                 XLogRecPtr      endptr;
 594                 Size            nbytes;
 595
 596                 /*
 597                  * Figure out how much to send in one message. If there's less than
 598                  * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
 599                  * MAX_SEND_SIZE bytes, but round to page boundary.
 600                  *
 601                  * The rounding is not only for performance reasons. Walreceiver
 602                  * relies on the fact that we never split a WAL record across two
 603                  * messages. Since a long WAL record is split at page boundary into
 604                  * continuation records, page boundary is always a safe cut-off point.
 605                  * We also assume that SendRqstPtr never points in the middle of a WAL
 606                  * record.
 607                  */
 608                 startptr = sentPtr;
 609                 if (startptr.xrecoff >= XLogFileSize)
 610                 {
 611                         /*
 612                          * crossing a logid boundary, skip the non-existent last log
 613                          * segment in previous logical log file.
 614                          */
 615                         startptr.xlogid += 1;
 616                         startptr.xrecoff = 0;
 617                 }
 618
 619                 endptr = startptr;
 620                 XLByteAdvance(endptr, MAX_SEND_SIZE);
 621                 /* round down to page boundary. */
 622                 endptr.xrecoff -= (endptr.xrecoff % XLOG_BLCKSZ);
 623                 /* if we went beyond SendRqstPtr, back off */
 624                 if (XLByteLT(SendRqstPtr, endptr))
 625                         endptr = SendRqstPtr;
 626
 627                 /*
 628                  * OK to read and send the slice.
 629                  *
 630                  * We don't need to convert the xlogid/xrecoff from host byte order to
 631                  * network byte order because the both server can be expected to have
 632                  * the same byte order. If they have different byte order, we don't
 633                  * reach here.
 634                  */
 635                 pq_sendbyte(outMsg, 'w');
 636                 pq_sendbytes(outMsg, (char *) &startptr, sizeof(startptr));
 637
 638                 if (endptr.xlogid != startptr.xlogid)
 639                 {
 640                         Assert(endptr.xlogid == startptr.xlogid + 1);
 641                         nbytes = endptr.xrecoff + XLogFileSize - startptr.xrecoff;
 642                 }
 643                 else
 644                         nbytes = endptr.xrecoff - startptr.xrecoff;
 645
 646                 sentPtr = endptr;
 647
 648                 /*
 649                  * Read the log directly into the output buffer to prevent extra
 650                  * memcpy calls.
 651                  */
 652                 enlargeStringInfo(outMsg, nbytes);
 653
 654                 XLogRead(&outMsg->data[outMsg->len], startptr, nbytes);
 655                 outMsg->len += nbytes;
 656                 outMsg->data[outMsg->len] = '\0';
 657
 658                 pq_putmessage('d', outMsg->data, outMsg->len);
 659                 resetStringInfo(outMsg);
 660         }
 661
 662         /* Update shared memory status */
 663         SpinLockAcquire(&walsnd->mutex);
 664         walsnd->sentPtr = sentPtr;
 665         SpinLockRelease(&walsnd->mutex);
 666
 667         /* Flush pending output */
 668         if (pq_flush())
 669                 return false;
 670
 671         /* Report progress of XLOG streaming in PS display */
 672         snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
 673                          sentPtr.xlogid, sentPtr.xrecoff);
 674         set_ps_display(activitymsg, false);
 675
 676         return true;
 677 }
 678
 679 /* SIGHUP: set flag to re-read config file at next convenient time */
 680 static void
 681 WalSndSigHupHandler(SIGNAL_ARGS)
 682 {
 683         got_SIGHUP = true;
 684 }
 685
 686 /* SIGTERM: set flag to shut down */
 687 static void
 688 WalSndShutdownHandler(SIGNAL_ARGS)
 689 {
 690         shutdown_requested = true;
 691 }
 692
 693 /*
 694  * WalSndQuickDieHandler() occurs when signalled SIGQUIT by the postmaster.
 695  *
 696  * Some backend has bought the farm,
 697  * so we need to stop what we're doing and exit.
 698  */
 699 static void
 700 WalSndQuickDieHandler(SIGNAL_ARGS)
 701 {
 702         PG_SETMASK(&BlockSig);
 703
 704         /*
 705          * We DO NOT want to run proc_exit() callbacks -- we're here because
 706          * shared memory may be corrupted, so we don't want to try to clean up our
 707          * transaction.  Just nail the windows shut and get out of town.  Now that
 708          * there's an atexit callback to prevent third-party code from breaking
 709          * things by calling exit() directly, we have to reset the callbacks
 710          * explicitly to make this work as intended.
 711          */
 712         on_exit_reset();
 713
 714         /*
 715          * Note we do exit(2) not exit(0).      This is to force the postmaster into a
 716          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
 717          * backend.  This is necessary precisely because we don't clean up our
 718          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
 719          * should ensure the postmaster sees this as a crash, too, but no harm in
 720          * being doubly sure.)
 721          */
 722         exit(2);
 723 }
 724
 725 /* SIGUSR2: set flag to do a last cycle and shut down afterwards */
 726 static void
 727 WalSndLastCycleHandler(SIGNAL_ARGS)
 728 {
 729         ready_to_stop = true;
 730 }
 731
 732 /* Set up signal handlers */
 733 void
 734 WalSndSignals(void)
 735 {
 736         /* Set up signal handlers */
 737         pqsignal(SIGHUP, WalSndSigHupHandler);          /* set flag to read config
 738                                                                                                  * file */
 739         pqsignal(SIGINT, SIG_IGN);      /* not used */
 740         pqsignal(SIGTERM, WalSndShutdownHandler);       /* request shutdown */
 741         pqsignal(SIGQUIT, WalSndQuickDieHandler);       /* hard crash time */
 742         pqsignal(SIGALRM, SIG_IGN);
 743         pqsignal(SIGPIPE, SIG_IGN);
 744         pqsignal(SIGUSR1, SIG_IGN); /* not used */
 745         pqsignal(SIGUSR2, WalSndLastCycleHandler);      /* request a last cycle and
 746                                                                                                  * shutdown */
 747
 748         /* Reset some signals that are accepted by postmaster but not here */
 749         pqsignal(SIGCHLD, SIG_DFL);
 750         pqsignal(SIGTTIN, SIG_DFL);
 751         pqsignal(SIGTTOU, SIG_DFL);
 752         pqsignal(SIGCONT, SIG_DFL);
 753         pqsignal(SIGWINCH, SIG_DFL);
 754 }
 755
 756 /* Report shared-memory space needed by WalSndShmemInit */
 757 Size
 758 WalSndShmemSize(void)
 759 {
 760         Size            size = 0;
 761
 762         size = offsetof(WalSndCtlData, walsnds);
 763         size = add_size(size, mul_size(MaxWalSenders, sizeof(WalSnd)));
 764
 765         return size;
 766 }
 767
 768 /* Allocate and initialize walsender-related shared memory */
 769 void
 770 WalSndShmemInit(void)
 771 {
 772         bool            found;
 773         int                     i;
 774
 775         WalSndCtl = (WalSndCtlData *)
 776                 ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
 777
 778         if (WalSndCtl == NULL)
 779                 ereport(FATAL,
 780                                 (errcode(ERRCODE_OUT_OF_MEMORY),
 781                                  errmsg("not enough shared memory for walsender")));
 782         if (found)
 783                 return;                                 /* already initialized */
 784
 785         /* Initialize the data structures */
 786         MemSet(WalSndCtl, 0, WalSndShmemSize());
 787
 788         for (i = 0; i < MaxWalSenders; i++)
 789         {
 790                 WalSnd     *walsnd = &WalSndCtl->walsnds[i];
 791
 792                 SpinLockInit(&walsnd->mutex);
 793         }
 794 }
 795
 796 /*
 797  * Returns the oldest Send position among walsenders. Or InvalidXLogRecPtr
 798  * if none.
 799  */
 800 XLogRecPtr
 801 GetOldestWALSendPointer(void)
 802 {
 803         XLogRecPtr      oldest = {0, 0};
 804         int                     i;
 805         bool            found = false;
 806
 807         for (i = 0; i < MaxWalSenders; i++)
 808         {
 809                 /* use volatile pointer to prevent code rearrangement */
 810                 volatile WalSnd *walsnd = &WalSndCtl->walsnds[i];
 811                 XLogRecPtr      recptr;
 812
 813                 if (walsnd->pid == 0)
 814                         continue;
 815
 816                 SpinLockAcquire(&walsnd->mutex);
 817                 recptr = walsnd->sentPtr;
 818                 SpinLockRelease(&walsnd->mutex);
 819
 820                 if (recptr.xlogid == 0 && recptr.xrecoff == 0)
 821                         continue;
 822
 823                 if (!found || XLByteLT(recptr, oldest))
 824                         oldest = recptr;
 825                 found = true;
 826         }
 827         return oldest;
 828 }