granicus.if.org Git - postgresql/blob - src/backend/postmaster/postmaster.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * postmaster.c
   4  *        This program acts as a clearing house for requests to the
   5  *        POSTGRES system.      Frontend programs send a startup message
   6  *        to the Postmaster and the postmaster uses the info in the
   7  *        message to setup a backend process.
   8  *
   9  *        The postmaster also manages system-wide operations such as
  10  *        startup, shutdown, and periodic checkpoints.  The postmaster
  11  *        itself doesn't do those operations, mind you --- it just forks
  12  *        off a subprocess to do them at the right times.  It also takes
  13  *        care of resetting the system if a backend crashes.
  14  *
  15  *        The postmaster process creates the shared memory and semaphore
  16  *        pools during startup, but as a rule does not touch them itself.
  17  *        In particular, it is not a member of the PGPROC array of backends
  18  *        and so it cannot participate in lock-manager operations.      Keeping
  19  *        the postmaster away from shared memory operations makes it simpler
  20  *        and more reliable.  The postmaster is almost always able to recover
  21  *        from crashes of individual backends by resetting shared memory;
  22  *        if it did much with shared memory then it would be prone to crashing
  23  *        along with the backends.
  24  *
  25  *        When a request message is received, we now fork() immediately.
  26  *        The child process performs authentication of the request, and
  27  *        then becomes a backend if successful.  This allows the auth code
  28  *        to be written in a simple single-threaded style (as opposed to the
  29  *        crufty "poor man's multitasking" code that used to be needed).
  30  *        More importantly, it ensures that blockages in non-multithreaded
  31  *        libraries like SSL or PAM cannot cause denial of service to other
  32  *        clients.
  33  *
  34  *
  35  * Portions Copyright (c) 1996-2003, PostgreSQL Global Development Group
  36  * Portions Copyright (c) 1994, Regents of the University of California
  37  *
  38  *
  39  * IDENTIFICATION
  40  *        $PostgreSQL: pgsql/src/backend/postmaster/postmaster.c,v 1.371 2004/03/09 04:43:06 momjian Exp $
  41  *
  42  * NOTES
  43  *
  44  * Initialization:
  45  *              The Postmaster sets up a few shared memory data structures
  46  *              for the backends.  It should at the very least initialize the
  47  *              lock manager.
  48  *
  49  * Synchronization:
  50  *              The Postmaster shares memory with the backends but should avoid
  51  *              touching shared memory, so as not to become stuck if a crashing
  52  *              backend screws up locks or shared memory.  Likewise, the Postmaster
  53  *              should never block on messages from frontend clients.
  54  *
  55  * Garbage Collection:
  56  *              The Postmaster cleans up after backends if they have an emergency
  57  *              exit and/or core dump.
  58  *
  59  *-------------------------------------------------------------------------
  60  */
  61
  62 #include "postgres.h"
  63
  64 #include <unistd.h>
  65 #include <signal.h>
  66 #include <sys/wait.h>
  67 #include <ctype.h>
  68 #include <sys/stat.h>
  69 #include <sys/time.h>
  70 #include <sys/socket.h>
  71 #include <errno.h>
  72 #include <fcntl.h>
  73 #include <time.h>
  74 #include <sys/param.h>
  75 #include <netinet/in.h>
  76 #include <arpa/inet.h>
  77 #include <netdb.h>
  78 #include <limits.h>
  79
  80 #ifdef HAVE_SYS_SELECT_H
  81 #include <sys/select.h>
  82 #endif
  83
  84 #ifdef HAVE_GETOPT_H
  85 #include <getopt.h>
  86 #endif
  87
  88 #ifdef USE_RENDEZVOUS
  89 #include <DNSServiceDiscovery/DNSServiceDiscovery.h>
  90 #endif
  91
  92 #include "catalog/pg_database.h"
  93 #include "commands/async.h"
  94 #include "lib/dllist.h"
  95 #include "libpq/auth.h"
  96 #include "libpq/crypt.h"
  97 #include "libpq/libpq.h"
  98 #include "libpq/pqcomm.h"
  99 #include "libpq/pqsignal.h"
 100 #include "miscadmin.h"
 101 #include "nodes/nodes.h"
 102 #include "storage/fd.h"
 103 #include "storage/ipc.h"
 104 #include "storage/pg_shmem.h"
 105 #include "storage/pmsignal.h"
 106 #include "storage/proc.h"
 107 #include "storage/bufmgr.h"
 108 #include "access/xlog.h"
 109 #include "tcop/tcopprot.h"
 110 #include "utils/guc.h"
 111 #include "utils/memutils.h"
 112 #include "utils/ps_status.h"
 113 #include "bootstrap/bootstrap.h"
 114 #include "pgstat.h"
 115
 116
 117 #define INVALID_SOCK    (-1)
 118
 119 #ifdef HAVE_SIGPROCMASK
 120 sigset_t        UnBlockSig,
 121                         BlockSig,
 122                         AuthBlockSig;
 123
 124 #else
 125 int                     UnBlockSig,
 126                         BlockSig,
 127                         AuthBlockSig;
 128 #endif
 129
 130 /*
 131  * List of active backends (or child processes anyway; we don't actually
 132  * know whether a given child has become a backend or is still in the
 133  * authorization phase).  This is used mainly to keep track of how many
 134  * children we have and send them appropriate signals when necessary.
 135  */
 136 typedef struct bkend
 137 {
 138         pid_t           pid;                    /* process id of backend */
 139         long            cancel_key;             /* cancel key for cancels for this backend */
 140 } Backend;
 141
 142 static Dllist *BackendList;
 143
 144 #ifdef EXEC_BACKEND
 145 #define NUM_BACKENDARRAY_ELEMS (2*MaxBackends)
 146 static Backend *ShmemBackendArray;
 147 #endif
 148
 149 /* The socket number we are listening for connections on */
 150 int                     PostPortNumber;
 151 char       *UnixSocketDir;
 152 char       *VirtualHost;
 153
 154 /*
 155  * MaxBackends is the limit on the number of backends we can start.
 156  * Note that a larger MaxBackends value will increase the size of the
 157  * shared memory area as well as cause the postmaster to grab more
 158  * kernel semaphores, even if you never actually use that many
 159  * backends.
 160  */
 161 int                     MaxBackends;
 162
 163 /*
 164  * ReservedBackends is the number of backends reserved for superuser use.
 165  * This number is taken out of the pool size given by MaxBackends so
 166  * number of backend slots available to non-superusers is
 167  * (MaxBackends - ReservedBackends).  Note what this really means is
 168  * "if there are <= ReservedBackends connections available, only superusers
 169  * can make new connections" --- pre-existing superuser connections don't
 170  * count against the limit.
 171  */
 172 int                     ReservedBackends;
 173
 174
 175 static char *progname = NULL;
 176
 177 /* The socket(s) we're listening to. */
 178 #define MAXLISTEN       10
 179 static int      ListenSocket[MAXLISTEN];
 180
 181 /* Used to reduce macros tests */
 182 #ifdef EXEC_BACKEND
 183 const bool      ExecBackend = true;
 184
 185 #else
 186 const bool      ExecBackend = false;
 187 #endif
 188
 189 /*
 190  * Set by the -o option
 191  */
 192 static char ExtraOptions[MAXPGPATH];
 193
 194 /*
 195  * These globals control the behavior of the postmaster in case some
 196  * backend dumps core.  Normally, it kills all peers of the dead backend
 197  * and reinitializes shared memory.  By specifying -s or -n, we can have
 198  * the postmaster stop (rather than kill) peers and not reinitialize
 199  * shared data structures.
 200  */
 201 static bool Reinit = true;
 202 static int      SendStop = false;
 203
 204 /* still more option variables */
 205 bool            NetServer = false;      /* listen on TCP/IP */
 206 bool            EnableSSL = false;
 207 bool            SilentMode = false; /* silent mode (-S) */
 208
 209 int                     PreAuthDelay = 0;
 210 int                     AuthenticationTimeout = 60;
 211 int                     CheckPointTimeout = 300;
 212 int                     CheckPointWarning = 30;
 213 time_t          LastSignalledCheckpoint = 0;
 214
 215 bool            log_hostname;           /* for ps display */
 216 bool            LogSourcePort;
 217 bool            Log_connections = false;
 218 bool            Db_user_namespace = false;
 219
 220 char       *rendezvous_name;
 221
 222 /* For FNCTL_NONBLOCK */
 223 #if defined(WIN32) || defined(__BEOS__)
 224 long            ioctlsocket_ret=1;
 225 #endif
 226
 227 /* list of library:init-function to be preloaded */
 228 char       *preload_libraries_string = NULL;
 229
 230 /* Startup/shutdown state */
 231 static pid_t StartupPID = 0,
 232                         ShutdownPID = 0,
 233                         CheckPointPID = 0,
 234                         BgWriterPID = 0;
 235 static time_t checkpointed = 0;
 236
 237 #define                 NoShutdown              0
 238 #define                 SmartShutdown   1
 239 #define                 FastShutdown    2
 240
 241 static int      Shutdown = NoShutdown;
 242
 243 static bool FatalError = false; /* T if recovering from backend crash */
 244
 245 bool            ClientAuthInProgress = false;           /* T during new-client
 246                                                                                                  * authentication */
 247
 248 /*
 249  * State for assigning random salts and cancel keys.
 250  * Also, the global MyCancelKey passes the cancel key assigned to a given
 251  * backend from the postmaster to that backend (via fork).
 252  */
 253
 254 static unsigned int random_seed = 0;
 255
 256 static int      debug_flag = 0;
 257
 258 extern char *optarg;
 259 extern int      optind,
 260                         opterr;
 261
 262 #ifdef HAVE_INT_OPTRESET
 263 extern int      optreset;
 264 #endif
 265
 266 /*
 267  * postmaster.c - function prototypes
 268  */
 269 static void pmdaemonize(int argc, char *argv[]);
 270 static Port *ConnCreate(int serverFd);
 271 static void ConnFree(Port *port);
 272 static void reset_shared(unsigned short port);
 273 static void SIGHUP_handler(SIGNAL_ARGS);
 274 static void pmdie(SIGNAL_ARGS);
 275 static void reaper(SIGNAL_ARGS);
 276 static void sigusr1_handler(SIGNAL_ARGS);
 277 static void dummy_handler(SIGNAL_ARGS);
 278 static void CleanupProc(int pid, int exitstatus);
 279 static void LogChildExit(int lev, const char *procname,
 280                          int pid, int exitstatus);
 281 static void BackendInit(Port *port);
 282 static int  BackendRun(Port *port);
 283 static void ExitPostmaster(int status);
 284 static void usage(const char *);
 285 static int      ServerLoop(void);
 286 static int      BackendStartup(Port *port);
 287 static int      ProcessStartupPacket(Port *port, bool SSLdone);
 288 static void processCancelRequest(Port *port, void *pkt);
 289 static int      initMasks(fd_set *rmask);
 290 static void report_fork_failure_to_client(Port *port, int errnum);
 291 static enum CAC_state canAcceptConnections(void);
 292 static long PostmasterRandom(void);
 293 static void RandomSalt(char *cryptSalt, char *md5Salt);
 294 static void SignalChildren(int signal);
 295 static int      CountChildren(void);
 296 static bool CreateOptsFile(int argc, char *argv[]);
 297 NON_EXEC_STATIC void SSDataBaseInit(int xlop);
 298 static pid_t SSDataBase(int xlop);
 299 static void
 300 postmaster_error(const char *fmt,...)
 301 /* This lets gcc check the format string for consistency. */
 302 __attribute__((format(printf, 1, 2)));
 303
 304 #ifdef EXEC_BACKEND
 305 #ifdef WIN32
 306 pid_t win32_forkexec(const char* path, char *argv[]);
 307
 308 static void  win32_AddChild(pid_t pid, HANDLE handle);
 309 static void  win32_RemoveChild(pid_t pid);
 310 static pid_t win32_waitpid(int *exitstatus);
 311 static DWORD WINAPI win32_sigchld_waiter(LPVOID param);
 312
 313 static pid_t  *win32_childPIDArray;
 314 static HANDLE *win32_childHNDArray;
 315 static unsigned long win32_numChildren = 0;
 316 #endif
 317
 318 static pid_t Backend_forkexec(Port *port);
 319
 320 static unsigned long tmpBackendFileNum = 0;
 321 void read_backend_variables(unsigned long id, Port *port);
 322 static bool write_backend_variables(Port *port);
 323
 324 size_t          ShmemBackendArraySize(void);
 325 void            ShmemBackendArrayAllocation(void);
 326 static void     ShmemBackendArrayAdd(Backend *bn);
 327 static void ShmemBackendArrayRemove(pid_t pid);
 328 #endif
 329
 330 #define StartupDataBase()               SSDataBase(BS_XLOG_STARTUP)
 331 #define CheckPointDataBase()    SSDataBase(BS_XLOG_CHECKPOINT)
 332 #define StartBackgroundWriter() SSDataBase(BS_XLOG_BGWRITER)
 333 #define ShutdownDataBase()              SSDataBase(BS_XLOG_SHUTDOWN)
 334
 335
 336 static void
 337 checkDataDir(const char *checkdir)
 338 {
 339         char            path[MAXPGPATH];
 340         FILE       *fp;
 341         struct stat stat_buf;
 342
 343         if (checkdir == NULL)
 344         {
 345                 fprintf(stderr,
 346                                 gettext("%s does not know where to find the database system data.\n"
 347                                                 "You must specify the directory that contains the database system\n"
 348                                                 "either by specifying the -D invocation option or by setting the\n"
 349                                                 "PGDATA environment variable.\n"),
 350                                 progname);
 351                 ExitPostmaster(2);
 352         }
 353
 354         if (stat(checkdir, &stat_buf) == -1)
 355         {
 356                 if (errno == ENOENT)
 357                         ereport(FATAL,
 358                                         (errcode_for_file_access(),
 359                                          errmsg("data directory \"%s\" does not exist",
 360                                                         checkdir)));
 361                 else
 362                         ereport(FATAL,
 363                                         (errcode_for_file_access(),
 364                          errmsg("could not read permissions of directory \"%s\": %m",
 365                                         checkdir)));
 366         }
 367
 368         /*
 369          * Check if the directory has group or world access.  If so, reject.
 370          *
 371          * XXX temporarily suppress check when on Windows, because there may not
 372          * be proper support for Unix-y file permissions.  Need to think of a
 373          * reasonable check to apply on Windows.
 374          */
 375 #if !defined(__CYGWIN__) && !defined(WIN32)
 376         if (stat_buf.st_mode & (S_IRWXG | S_IRWXO))
 377                 ereport(FATAL,
 378                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 379                                  errmsg("data directory \"%s\" has group or world access",
 380                                                 checkdir),
 381                                  errdetail("Permissions should be u=rwx (0700).")));
 382 #endif
 383
 384         /* Look for PG_VERSION before looking for pg_control */
 385         ValidatePgVersion(checkdir);
 386
 387         snprintf(path, sizeof(path), "%s/global/pg_control", checkdir);
 388
 389         fp = AllocateFile(path, PG_BINARY_R);
 390         if (fp == NULL)
 391         {
 392                 fprintf(stderr,
 393                                 gettext("%s: could not find the database system\n"
 394                                                 "Expected to find it in the directory \"%s\",\n"
 395                                                 "but could not open file \"%s\": %s\n"),
 396                                 progname, checkdir, path, strerror(errno));
 397                 ExitPostmaster(2);
 398         }
 399         FreeFile(fp);
 400 }
 401
 402
 403 #ifdef USE_RENDEZVOUS
 404
 405 /* reg_reply -- empty callback function for DNSServiceRegistrationCreate() */
 406 static void
 407 reg_reply(DNSServiceRegistrationReplyErrorType errorCode, void *context)
 408 {
 409
 410 }
 411 #endif
 412
 413 int
 414 PostmasterMain(int argc, char *argv[])
 415 {
 416         int                     opt;
 417         int                     status;
 418         char            original_extraoptions[MAXPGPATH];
 419         char       *potential_DataDir = NULL;
 420         int                     i;
 421
 422         *original_extraoptions = '\0';
 423
 424         progname = argv[0];
 425
 426         IsPostmasterEnvironment = true;
 427
 428         /*
 429          * Catch standard options before doing much else.  This even works on
 430          * systems without getopt_long.
 431          */
 432         if (argc > 1)
 433         {
 434                 if (strcmp(argv[1], "--help") == 0 || strcmp(argv[1], "-?") == 0)
 435                 {
 436                         usage(progname);
 437                         ExitPostmaster(0);
 438                 }
 439                 if (strcmp(argv[1], "--version") == 0 || strcmp(argv[1], "-V") == 0)
 440                 {
 441                         puts("postmaster (PostgreSQL) " PG_VERSION);
 442                         ExitPostmaster(0);
 443                 }
 444         }
 445
 446         /*
 447          * for security, no dir or file created can be group or other
 448          * accessible
 449          */
 450         umask((mode_t) 0077);
 451
 452         MyProcPid = PostmasterPid = getpid();
 453
 454         /*
 455          * Fire up essential subsystems: memory management
 456          */
 457         MemoryContextInit();
 458
 459         /*
 460          * By default, palloc() requests in the postmaster will be allocated
 461          * in the PostmasterContext, which is space that can be recycled by
 462          * backends.  Allocated data that needs to be available to backends
 463          * should be allocated in TopMemoryContext.
 464          */
 465         PostmasterContext = AllocSetContextCreate(TopMemoryContext,
 466                                                                                           "Postmaster",
 467                                                                                           ALLOCSET_DEFAULT_MINSIZE,
 468                                                                                           ALLOCSET_DEFAULT_INITSIZE,
 469                                                                                           ALLOCSET_DEFAULT_MAXSIZE);
 470         MemoryContextSwitchTo(PostmasterContext);
 471
 472         IgnoreSystemIndexes(false);
 473
 474         /*
 475          * Options setup
 476          */
 477         InitializeGUCOptions();
 478
 479         potential_DataDir = getenv("PGDATA");           /* default value */
 480
 481         opterr = 1;
 482
 483         while ((opt = getopt(argc, argv, "A:a:B:b:c:D:d:Fh:ik:lm:MN:no:p:Ss-:")) != -1)
 484         {
 485                 switch (opt)
 486                 {
 487                         case 'A':
 488 #ifdef USE_ASSERT_CHECKING
 489                                 SetConfigOption("debug_assertions", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 490 #else
 491                                 postmaster_error("assert checking is not compiled in");
 492 #endif
 493                                 break;
 494                         case 'a':
 495                                 /* Can no longer set authentication method. */
 496                                 break;
 497                         case 'B':
 498                                 SetConfigOption("shared_buffers", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 499                                 break;
 500                         case 'b':
 501                                 /* Can no longer set the backend executable file to use. */
 502                                 break;
 503                         case 'D':
 504                                 potential_DataDir = optarg;
 505                                 break;
 506                         case 'd':
 507                                 {
 508                                         /* Turn on debugging for the postmaster. */
 509                                         char       *debugstr = palloc(strlen("debug") + strlen(optarg) + 1);
 510
 511                                         sprintf(debugstr, "debug%s", optarg);
 512                                         SetConfigOption("log_min_messages", debugstr,
 513                                                                         PGC_POSTMASTER, PGC_S_ARGV);
 514                                         pfree(debugstr);
 515                                         debug_flag = atoi(optarg);
 516                                         break;
 517                                 }
 518                         case 'F':
 519                                 SetConfigOption("fsync", "false", PGC_POSTMASTER, PGC_S_ARGV);
 520                                 break;
 521                         case 'h':
 522                                 SetConfigOption("virtual_host", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 523                                 break;
 524                         case 'i':
 525                                 SetConfigOption("tcpip_socket", "true", PGC_POSTMASTER, PGC_S_ARGV);
 526                                 break;
 527                         case 'k':
 528                                 SetConfigOption("unix_socket_directory", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 529                                 break;
 530 #ifdef USE_SSL
 531                         case 'l':
 532                                 SetConfigOption("ssl", "true", PGC_POSTMASTER, PGC_S_ARGV);
 533                                 break;
 534 #endif
 535                         case 'm':
 536                                 /* Multiplexed backends no longer supported. */
 537                                 break;
 538                         case 'M':
 539
 540                                 /*
 541                                  * ignore this flag.  This may be passed in because the
 542                                  * program was run as 'postgres -M' instead of
 543                                  * 'postmaster'
 544                                  */
 545                                 break;
 546                         case 'N':
 547                                 /* The max number of backends to start. */
 548                                 SetConfigOption("max_connections", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 549                                 break;
 550                         case 'n':
 551                                 /* Don't reinit shared mem after abnormal exit */
 552                                 Reinit = false;
 553                                 break;
 554                         case 'o':
 555
 556                                 /*
 557                                  * Other options to pass to the backend on the command
 558                                  * line -- useful only for debugging.
 559                                  */
 560                                 strcat(ExtraOptions, " ");
 561                                 strcat(ExtraOptions, optarg);
 562                                 strcpy(original_extraoptions, optarg);
 563                                 break;
 564                         case 'p':
 565                                 SetConfigOption("port", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 566                                 break;
 567                         case 'S':
 568
 569                                 /*
 570                                  * Start in 'S'ilent mode (disassociate from controlling
 571                                  * tty). You may also think of this as 'S'ysV mode since
 572                                  * it's most badly needed on SysV-derived systems like
 573                                  * SVR4 and HP-UX.
 574                                  */
 575                                 SetConfigOption("silent_mode", "true", PGC_POSTMASTER, PGC_S_ARGV);
 576                                 break;
 577                         case 's':
 578
 579                                 /*
 580                                  * In the event that some backend dumps core, send
 581                                  * SIGSTOP, rather than SIGQUIT, to all its peers.      This
 582                                  * lets the wily post_hacker collect core dumps from
 583                                  * everyone.
 584                                  */
 585                                 SendStop = true;
 586                                 break;
 587                         case 'c':
 588                         case '-':
 589                                 {
 590                                         char       *name,
 591                                                            *value;
 592
 593                                         ParseLongOption(optarg, &name, &value);
 594                                         if (!value)
 595                                         {
 596                                                 if (opt == '-')
 597                                                         ereport(ERROR,
 598                                                                         (errcode(ERRCODE_SYNTAX_ERROR),
 599                                                                          errmsg("--%s requires a value",
 600                                                                                         optarg)));
 601                                                 else
 602                                                         ereport(ERROR,
 603                                                                         (errcode(ERRCODE_SYNTAX_ERROR),
 604                                                                          errmsg("-c %s requires a value",
 605                                                                                         optarg)));
 606                                         }
 607
 608                                         SetConfigOption(name, value, PGC_POSTMASTER, PGC_S_ARGV);
 609                                         free(name);
 610                                         if (value)
 611                                                 free(value);
 612                                         break;
 613                                 }
 614
 615                         default:
 616                                 fprintf(stderr,
 617                                           gettext("Try \"%s --help\" for more information.\n"),
 618                                                 progname);
 619                                 ExitPostmaster(1);
 620                 }
 621         }
 622
 623         /*
 624          * Postmaster accepts no non-option switch arguments.
 625          */
 626         if (optind < argc)
 627         {
 628                 postmaster_error("invalid argument: \"%s\"", argv[optind]);
 629                 fprintf(stderr,
 630                                 gettext("Try \"%s --help\" for more information.\n"),
 631                                 progname);
 632                 ExitPostmaster(1);
 633         }
 634
 635         /*
 636          * Now we can set the data directory, and then read postgresql.conf.
 637          */
 638         checkDataDir(potential_DataDir);        /* issues error messages */
 639         SetDataDir(potential_DataDir);
 640
 641         ProcessConfigFile(PGC_POSTMASTER);
 642 #ifdef EXEC_BACKEND
 643         write_nondefault_variables(PGC_POSTMASTER);
 644 #endif
 645
 646         /*
 647          * Check for invalid combinations of GUC settings.
 648          */
 649         if (NBuffers < 2 * MaxBackends || NBuffers < 16)
 650         {
 651                 /*
 652                  * Do not accept -B so small that backends are likely to starve
 653                  * for lack of buffers.  The specific choices here are somewhat
 654                  * arbitrary.
 655                  */
 656                 postmaster_error("the number of buffers (-B) must be at least twice the number of allowed connections (-N) and at least 16");
 657                 ExitPostmaster(1);
 658         }
 659
 660         if (ReservedBackends >= MaxBackends)
 661         {
 662                 postmaster_error("superuser_reserved_connections must be less than max_connections");
 663                 ExitPostmaster(1);
 664         }
 665
 666         /*
 667          * Other one-time internal sanity checks can go here.
 668          */
 669         if (!CheckDateTokenTables())
 670         {
 671                 postmaster_error("invalid datetoken tables, please fix");
 672                 ExitPostmaster(1);
 673         }
 674
 675         /*
 676          * Now that we are done processing the postmaster arguments, reset
 677          * getopt(3) library so that it will work correctly in subprocesses.
 678          */
 679         optind = 1;
 680 #ifdef HAVE_INT_OPTRESET
 681         optreset = 1;                           /* some systems need this too */
 682 #endif
 683
 684         /* For debugging: display postmaster environment */
 685         {
 686                 extern char **environ;
 687                 char      **p;
 688
 689                 ereport(DEBUG3,
 690                                 (errmsg_internal("%s: PostmasterMain: initial environ dump:",
 691                                                                  progname)));
 692                 ereport(DEBUG3,
 693                                 (errmsg_internal("-----------------------------------------")));
 694                 for (p = environ; *p; ++p)
 695                         ereport(DEBUG3,
 696                                         (errmsg_internal("\t%s", *p)));
 697                 ereport(DEBUG3,
 698                                 (errmsg_internal("-----------------------------------------")));
 699         }
 700
 701         /*
 702          * On some systems our dynloader code needs the executable's pathname.
 703          */
 704         if (FindExec(pg_pathname, progname, "postgres") < 0)
 705                 ereport(FATAL,
 706                                 (errmsg("%s: could not locate postgres executable",
 707                                                 progname)));
 708
 709         /*
 710          * Initialize SSL library, if specified.
 711          */
 712 #ifdef USE_SSL
 713         if (EnableSSL && !NetServer)
 714         {
 715                 postmaster_error("TCP/IP connections must be enabled for SSL");
 716                 ExitPostmaster(1);
 717         }
 718         if (EnableSSL)
 719                 secure_initialize();
 720 #endif
 721
 722         /*
 723          * process any libraries that should be preloaded and optionally
 724          * pre-initialized
 725          */
 726         if (preload_libraries_string)
 727                 process_preload_libraries(preload_libraries_string);
 728
 729         /*
 730          * Fork away from controlling terminal, if -S specified.
 731          *
 732          * Must do this before we grab any interlock files, else the interlocks
 733          * will show the wrong PID.
 734          */
 735         if (SilentMode)
 736                 pmdaemonize(argc, argv);
 737
 738         /*
 739          * Create lockfile for data directory.
 740          *
 741          * We want to do this before we try to grab the input sockets, because
 742          * the data directory interlock is more reliable than the socket-file
 743          * interlock (thanks to whoever decided to put socket files in /tmp
 744          * :-(). For the same reason, it's best to grab the TCP socket before
 745          * the Unix socket.
 746          */
 747         CreateDataDirLockFile(DataDir, true);
 748
 749         /*
 750          * Remove old temporary files.  At this point there can be no other
 751          * Postgres processes running in this directory, so this should be
 752          * safe.
 753          */
 754         RemovePgTempFiles();
 755
 756         /*
 757          * Establish input sockets.
 758          */
 759         for (i = 0; i < MAXLISTEN; i++)
 760                 ListenSocket[i] = -1;
 761
 762         if (NetServer)
 763         {
 764                 if (VirtualHost && VirtualHost[0])
 765                 {
 766                         char       *curhost,
 767                                            *endptr;
 768                         char            c = 0;
 769
 770                         curhost = VirtualHost;
 771                         for (;;)
 772                         {
 773                                 while (*curhost == ' ') /* skip any extra spaces */
 774                                         curhost++;
 775                                 if (*curhost == '\0')
 776                                         break;
 777                                 endptr = strchr(curhost, ' ');
 778                                 if (endptr)
 779                                 {
 780                                         c = *endptr;
 781                                         *endptr = '\0';
 782                                 }
 783                                 status = StreamServerPort(AF_UNSPEC, curhost,
 784                                                                                   (unsigned short) PostPortNumber,
 785                                                                                   UnixSocketDir,
 786                                                                                   ListenSocket, MAXLISTEN);
 787                                 if (status != STATUS_OK)
 788                                         ereport(FATAL,
 789                                          (errmsg("could not create listen socket for \"%s\"",
 790                                                          curhost)));
 791                                 if (endptr)
 792                                 {
 793                                         *endptr = c;
 794                                         curhost = endptr + 1;
 795                                 }
 796                                 else
 797                                         break;
 798                         }
 799                 }
 800                 else
 801                 {
 802                         status = StreamServerPort(AF_UNSPEC, NULL,
 803                                                                           (unsigned short) PostPortNumber,
 804                                                                           UnixSocketDir,
 805                                                                           ListenSocket, MAXLISTEN);
 806                         if (status != STATUS_OK)
 807                                 ereport(FATAL,
 808                                           (errmsg("could not create TCP/IP listen socket")));
 809                 }
 810
 811 #ifdef USE_RENDEZVOUS
 812                 if (rendezvous_name != NULL)
 813                 {
 814                         DNSServiceRegistrationCreate(rendezvous_name,
 815                                                                                  "_postgresql._tcp.",
 816                                                                                  "",
 817                                                                                  htonl(PostPortNumber),
 818                                                                                  "",
 819                                                                  (DNSServiceRegistrationReply) reg_reply,
 820                                                                                  NULL);
 821                 }
 822 #endif
 823         }
 824
 825 #ifdef HAVE_UNIX_SOCKETS
 826         status = StreamServerPort(AF_UNIX, NULL,
 827                                                           (unsigned short) PostPortNumber,
 828                                                           UnixSocketDir,
 829                                                           ListenSocket, MAXLISTEN);
 830         if (status != STATUS_OK)
 831                 ereport(FATAL,
 832                                 (errmsg("could not create Unix-domain socket")));
 833 #endif
 834
 835         XLOGPathInit();
 836
 837         /*
 838          * Set up shared memory and semaphores.
 839          */
 840         reset_shared(PostPortNumber);
 841
 842         /*
 843          * Estimate number of openable files.  This must happen after setting up
 844          * semaphores, because on some platforms semaphores count as open files.
 845          */
 846         set_max_safe_fds();
 847
 848         /*
 849          * Initialize the list of active backends.
 850          */
 851         BackendList = DLNewList();
 852
 853 #ifdef WIN32
 854         /*
 855          * Initialize the child pid/HANDLE arrays
 856          */
 857         win32_childPIDArray = (pid_t*)malloc(NUM_BACKENDARRAY_ELEMS*sizeof(pid_t));
 858         win32_childHNDArray = (HANDLE*)malloc(NUM_BACKENDARRAY_ELEMS*sizeof(HANDLE));
 859         if (!win32_childPIDArray || !win32_childHNDArray)
 860                 ereport(FATAL,
 861                                 (errcode(ERRCODE_OUT_OF_MEMORY),
 862                                  errmsg("out of memory")));
 863 #endif
 864
 865         /*
 866          * Record postmaster options.  We delay this till now to avoid
 867          * recording bogus options (eg, NBuffers too high for available
 868          * memory).
 869          */
 870         if (!CreateOptsFile(argc, argv))
 871                 ExitPostmaster(1);
 872
 873         /*
 874          * Set up signal handlers for the postmaster process.
 875          *
 876          * CAUTION: when changing this list, check for side-effects on the signal
 877          * handling setup of child processes.  See tcop/postgres.c,
 878          * bootstrap/bootstrap.c, and postmaster/pgstat.c.
 879          */
 880         pqinitmask();
 881         PG_SETMASK(&BlockSig);
 882
 883         pqsignal(SIGHUP, SIGHUP_handler);       /* reread config file and have
 884                                                                                  * children do same */
 885         pqsignal(SIGINT, pmdie);        /* send SIGTERM and ShutdownDataBase */
 886         pqsignal(SIGQUIT, pmdie);       /* send SIGQUIT and die */
 887         pqsignal(SIGTERM, pmdie);       /* wait for children and ShutdownDataBase */
 888         pqsignal(SIGALRM, SIG_IGN); /* ignored */
 889         pqsignal(SIGPIPE, SIG_IGN); /* ignored */
 890         pqsignal(SIGUSR1, sigusr1_handler); /* message from child process */
 891         pqsignal(SIGUSR2, dummy_handler);       /* unused, reserve for children */
 892         pqsignal(SIGCHLD, reaper);      /* handle child termination */
 893         pqsignal(SIGTTIN, SIG_IGN); /* ignored */
 894         pqsignal(SIGTTOU, SIG_IGN); /* ignored */
 895         /* ignore SIGXFSZ, so that ulimit violations work like disk full */
 896 #ifdef SIGXFSZ
 897         pqsignal(SIGXFSZ, SIG_IGN); /* ignored */
 898 #endif
 899
 900         /*
 901          * Reset whereToSendOutput from Debug (its starting state) to None.
 902          * This prevents ereport from sending log messages to stderr unless
 903          * the syslog/stderr switch permits.  We don't do this until the
 904          * postmaster is fully launched, since startup failures may as well be
 905          * reported to stderr.
 906          */
 907         whereToSendOutput = None;
 908
 909         /*
 910          * On many platforms, the first call of localtime() incurs significant
 911          * overhead to load timezone info from the system configuration files.
 912          * By doing it once in the postmaster, we avoid having to do it in
 913          * every started child process.  The savings are not huge, but they
 914          * add up...
 915          */
 916         {
 917                 time_t          now = time(NULL);
 918
 919                 (void) localtime(&now);
 920         }
 921
 922         /*
 923          * Initialize and try to startup the statistics collector process
 924          */
 925         pgstat_init();
 926         pgstat_start();
 927
 928         /*
 929          * Load cached files for client authentication.
 930          */
 931         load_hba();
 932         load_ident();
 933         load_user();
 934         load_group();
 935
 936         /*
 937          * We're ready to rock and roll...
 938          */
 939         StartupPID = StartupDataBase();
 940
 941         status = ServerLoop();
 942
 943         /*
 944          * ServerLoop probably shouldn't ever return, but if it does, close
 945          * down.
 946          */
 947         ExitPostmaster(status != STATUS_OK);
 948
 949         return 0;                                       /* not reached */
 950 }
 951
 952 static void
 953 pmdaemonize(int argc, char *argv[])
 954 {
 955         int                     i;
 956         pid_t           pid;
 957
 958 #ifdef LINUX_PROFILE
 959         struct itimerval prof_itimer;
 960 #endif
 961
 962 #ifdef LINUX_PROFILE
 963         /* see comments in BackendRun */
 964         getitimer(ITIMER_PROF, &prof_itimer);
 965 #endif
 966
 967 #ifdef WIN32
 968         /* FIXME: [fork/exec] to be implemented? */
 969         abort();
 970 #else
 971         pid = fork();
 972 #endif
 973         if (pid == (pid_t) -1)
 974         {
 975                 postmaster_error("could not fork background process: %s",
 976                                                  strerror(errno));
 977                 ExitPostmaster(1);
 978         }
 979         else if (pid)
 980         {                                                       /* parent */
 981                 /* Parent should just exit, without doing any atexit cleanup */
 982                 _exit(0);
 983         }
 984
 985 #ifdef LINUX_PROFILE
 986         setitimer(ITIMER_PROF, &prof_itimer, NULL);
 987 #endif
 988
 989         MyProcPid = getpid();           /* reset MyProcPid to child */
 990
 991 /* GH: If there's no setsid(), we hopefully don't need silent mode.
 992  * Until there's a better solution.
 993  */
 994 #ifdef HAVE_SETSID
 995         if (setsid() < 0)
 996         {
 997                 postmaster_error("could not dissociate from controlling TTY: %s",
 998                                                  strerror(errno));
 999                 ExitPostmaster(1);
1000         }
1001 #endif
1002         i = open(NULL_DEV, O_RDWR | PG_BINARY);
1003         dup2(i, 0);
1004         dup2(i, 1);
1005         dup2(i, 2);
1006         close(i);
1007 }
1008
1009
1010
1011 /*
1012  * Print out help message
1013  */
1014 static void
1015 usage(const char *progname)
1016 {
1017         printf(gettext("%s is the PostgreSQL server.\n\n"), progname);
1018         printf(gettext("Usage:\n  %s [OPTION]...\n\n"), progname);
1019         printf(gettext("Options:\n"));
1020 #ifdef USE_ASSERT_CHECKING
1021         printf(gettext("  -A 1|0          enable/disable run-time assert checking\n"));
1022 #endif
1023         printf(gettext("  -B NBUFFERS     number of shared buffers\n"));
1024         printf(gettext("  -c NAME=VALUE   set run-time parameter\n"));
1025         printf(gettext("  -d 1-5          debugging level\n"));
1026         printf(gettext("  -D DATADIR      database directory\n"));
1027         printf(gettext("  -F              turn fsync off\n"));
1028         printf(gettext("  -h HOSTNAME     host name or IP address to listen on\n"));
1029         printf(gettext("  -i              enable TCP/IP connections\n"));
1030         printf(gettext("  -k DIRECTORY    Unix-domain socket location\n"));
1031 #ifdef USE_SSL
1032         printf(gettext("  -l              enable SSL connections\n"));
1033 #endif
1034         printf(gettext("  -N MAX-CONNECT  maximum number of allowed connections\n"));
1035         printf(gettext("  -o OPTIONS      pass \"OPTIONS\" to each server process\n"));
1036         printf(gettext("  -p PORT         port number to listen on\n"));
1037         printf(gettext("  -S              silent mode (start in background without logging output)\n"));
1038         printf(gettext("  --help          show this help, then exit\n"));
1039         printf(gettext("  --version       output version information, then exit\n"));
1040
1041         printf(gettext("\nDeveloper options:\n"));
1042         printf(gettext("  -n              do not reinitialize shared memory after abnormal exit\n"));
1043         printf(gettext("  -s              send SIGSTOP to all backend servers if one dies\n"));
1044
1045         printf(gettext("\nPlease read the documentation for the complete list of run-time\n"
1046                                    "configuration settings and how to set them on the command line or in\n"
1047                                    "the configuration file.\n\n"
1048                                    "Report bugs to <pgsql-bugs@postgresql.org>.\n"));
1049 }
1050
1051 static int
1052 ServerLoop(void)
1053 {
1054         fd_set          readmask;
1055         int                     nSockets;
1056         struct timeval now,
1057                                 later;
1058         struct timezone tz;
1059         int                     i;
1060
1061         gettimeofday(&now, &tz);
1062
1063         nSockets = initMasks(&readmask);
1064
1065         for (;;)
1066         {
1067                 Port       *port;
1068                 fd_set          rmask;
1069                 struct timeval timeout;
1070
1071                 /*
1072                  * The timeout for the select() below is normally set on the basis
1073                  * of the time to the next checkpoint.  However, if for some
1074                  * reason we don't have a next-checkpoint time, time out after 60
1075                  * seconds. This keeps checkpoint scheduling from locking up when
1076                  * we get new connection requests infrequently (since we are
1077                  * likely to detect checkpoint completion just after enabling
1078                  * signals below, after we've already made the decision about how
1079                  * long to wait this time).
1080                  */
1081                 timeout.tv_sec = 60;
1082                 timeout.tv_usec = 0;
1083
1084                 if (CheckPointPID == 0 && checkpointed &&
1085                         StartupPID == 0 && Shutdown == NoShutdown &&
1086                         !FatalError && random_seed != 0)
1087                 {
1088                         time_t          now = time(NULL);
1089
1090                         if (CheckPointTimeout + checkpointed > now)
1091                         {
1092                                 /*
1093                                  * Not time for checkpoint yet, so set select timeout
1094                                  */
1095                                 timeout.tv_sec = CheckPointTimeout + checkpointed - now;
1096                         }
1097                         else
1098                         {
1099                                 /* Time to make the checkpoint... */
1100                                 CheckPointPID = CheckPointDataBase();
1101
1102                                 /*
1103                                  * if fork failed, schedule another try at 0.1 normal
1104                                  * delay
1105                                  */
1106                                 if (CheckPointPID == 0)
1107                                 {
1108                                         timeout.tv_sec = CheckPointTimeout / 10;
1109                                         checkpointed = now + timeout.tv_sec - CheckPointTimeout;
1110                                 }
1111                         }
1112                 }
1113
1114                 /*
1115                  * If no background writer process is running and we should
1116                  * do background writing, start one. It doesn't matter if
1117                  * this fails, we'll just try again later.
1118                  */
1119                 if (BgWriterPID == 0 && BgWriterPercent > 0 &&
1120                         StartupPID == 0 && Shutdown == NoShutdown &&
1121                         !FatalError && random_seed != 0)
1122                 {
1123                         BgWriterPID = StartBackgroundWriter();
1124                 }
1125
1126                 /*
1127                  * Wait for something to happen.
1128                  */
1129                 memcpy((char *) &rmask, (char *) &readmask, sizeof(fd_set));
1130
1131                 PG_SETMASK(&UnBlockSig);
1132
1133                 if (select(nSockets, &rmask, NULL, NULL, &timeout) < 0)
1134                 {
1135                         PG_SETMASK(&BlockSig);
1136                         if (errno == EINTR || errno == EWOULDBLOCK)
1137                                 continue;
1138                         ereport(LOG,
1139                                         (errcode_for_socket_access(),
1140                                          errmsg("select() failed in postmaster: %m")));
1141                         return STATUS_ERROR;
1142                 }
1143
1144                 /*
1145                  * Block all signals until we wait again.  (This makes it safe for
1146                  * our signal handlers to do nontrivial work.)
1147                  */
1148                 PG_SETMASK(&BlockSig);
1149
1150                 /*
1151                  * Select a random seed at the time of first receiving a request.
1152                  */
1153                 while (random_seed == 0)
1154                 {
1155                         gettimeofday(&later, &tz);
1156
1157                         /*
1158                          * We are not sure how much precision is in tv_usec, so we
1159                          * swap the nibbles of 'later' and XOR them with 'now'. On the
1160                          * off chance that the result is 0, we loop until it isn't.
1161                          */
1162                         random_seed = now.tv_usec ^
1163                                 ((later.tv_usec << 16) |
1164                                  ((later.tv_usec >> 16) & 0xffff));
1165                 }
1166
1167                 /*
1168                  * New connection pending on any of our sockets? If so, fork a
1169                  * child process to deal with it.
1170                  */
1171                 for (i = 0; i < MAXLISTEN; i++)
1172                 {
1173                         if (ListenSocket[i] == -1)
1174                                 break;
1175                         if (FD_ISSET(ListenSocket[i], &rmask))
1176                         {
1177                                 port = ConnCreate(ListenSocket[i]);
1178                                 if (port)
1179                                 {
1180                                         BackendStartup(port);
1181
1182                                         /*
1183                                          * We no longer need the open socket or port structure
1184                                          * in this process
1185                                          */
1186                                         StreamClose(port->sock);
1187                                         ConnFree(port);
1188                                 }
1189                         }
1190                 }
1191
1192                 /* If we have lost the stats collector, try to start a new one */
1193                 if (!pgstat_is_running)
1194                         pgstat_start();
1195         }
1196 }
1197
1198
1199 /*
1200  * Initialise the masks for select() for the ports
1201  * we are listening on.  Return the number of sockets to listen on.
1202  */
1203
1204 static int
1205 initMasks(fd_set *rmask)
1206 {
1207         int                     nsocks = -1;
1208         int                     i;
1209
1210         FD_ZERO(rmask);
1211
1212         for (i = 0; i < MAXLISTEN; i++)
1213         {
1214                 int                     fd = ListenSocket[i];
1215
1216                 if (fd == -1)
1217                         break;
1218                 FD_SET(fd, rmask);
1219                 if (fd > nsocks)
1220                         nsocks = fd;
1221         }
1222
1223         return nsocks + 1;
1224 }
1225
1226
1227 /*
1228  * Read the startup packet and do something according to it.
1229  *
1230  * Returns STATUS_OK or STATUS_ERROR, or might call ereport(FATAL) and
1231  * not return at all.
1232  *
1233  * (Note that ereport(FATAL) stuff is sent to the client, so only use it
1234  * if that's what you want.  Return STATUS_ERROR if you don't want to
1235  * send anything to the client, which would typically be appropriate
1236  * if we detect a communications failure.)
1237  */
1238 static int
1239 ProcessStartupPacket(Port *port, bool SSLdone)
1240 {
1241         int32           len;
1242         void       *buf;
1243         ProtocolVersion proto;
1244         MemoryContext oldcontext;
1245
1246         if (pq_getbytes((char *) &len, 4) == EOF)
1247         {
1248                 /*
1249                  * EOF after SSLdone probably means the client didn't like our
1250                  * response to NEGOTIATE_SSL_CODE.      That's not an error condition,
1251                  * so don't clutter the log with a complaint.
1252                  */
1253                 if (!SSLdone)
1254                         ereport(COMMERROR,
1255                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
1256                                          errmsg("incomplete startup packet")));
1257                 return STATUS_ERROR;
1258         }
1259
1260         len = ntohl(len);
1261         len -= 4;
1262
1263         if (len < (int32) sizeof(ProtocolVersion) ||
1264                 len > MAX_STARTUP_PACKET_LENGTH)
1265         {
1266                 ereport(COMMERROR,
1267                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1268                                  errmsg("invalid length of startup packet")));
1269                 return STATUS_ERROR;
1270         }
1271
1272         /*
1273          * Allocate at least the size of an old-style startup packet, plus one
1274          * extra byte, and make sure all are zeroes.  This ensures we will
1275          * have null termination of all strings, in both fixed- and
1276          * variable-length packet layouts.
1277          */
1278         if (len <= (int32) sizeof(StartupPacket))
1279                 buf = palloc0(sizeof(StartupPacket) + 1);
1280         else
1281                 buf = palloc0(len + 1);
1282
1283         if (pq_getbytes(buf, len) == EOF)
1284         {
1285                 ereport(COMMERROR,
1286                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1287                                  errmsg("incomplete startup packet")));
1288                 return STATUS_ERROR;
1289         }
1290
1291         /*
1292          * The first field is either a protocol version number or a special
1293          * request code.
1294          */
1295         port->proto = proto = ntohl(*((ProtocolVersion *) buf));
1296
1297         if (proto == CANCEL_REQUEST_CODE)
1298         {
1299                 processCancelRequest(port, buf);
1300                 return 127;                             /* XXX */
1301         }
1302
1303         if (proto == NEGOTIATE_SSL_CODE && !SSLdone)
1304         {
1305                 char            SSLok;
1306
1307 #ifdef USE_SSL
1308                 /* No SSL when disabled or on Unix sockets */
1309                 if (!EnableSSL || IS_AF_UNIX(port->laddr.addr.ss_family))
1310                         SSLok = 'N';
1311                 else
1312                         SSLok = 'S';            /* Support for SSL */
1313 #else
1314                 SSLok = 'N';                    /* No support for SSL */
1315 #endif
1316                 if (send(port->sock, &SSLok, 1, 0) != 1)
1317                 {
1318                         ereport(COMMERROR,
1319                                         (errcode_for_socket_access(),
1320                                  errmsg("failed to send SSL negotiation response: %m")));
1321                         return STATUS_ERROR;    /* close the connection */
1322                 }
1323
1324 #ifdef USE_SSL
1325                 if (SSLok == 'S' && secure_open_server(port) == -1)
1326                         return STATUS_ERROR;
1327 #endif
1328                 /* regular startup packet, cancel, etc packet should follow... */
1329                 /* but not another SSL negotiation request */
1330                 return ProcessStartupPacket(port, true);
1331         }
1332
1333         /* Could add additional special packet types here */
1334
1335         /*
1336          * Set FrontendProtocol now so that ereport() knows what format to
1337          * send if we fail during startup.
1338          */
1339         FrontendProtocol = proto;
1340
1341         /* Check we can handle the protocol the frontend is using. */
1342
1343         if (PG_PROTOCOL_MAJOR(proto) < PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST) ||
1344           PG_PROTOCOL_MAJOR(proto) > PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST) ||
1345         (PG_PROTOCOL_MAJOR(proto) == PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST) &&
1346          PG_PROTOCOL_MINOR(proto) > PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST)))
1347                 ereport(FATAL,
1348                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1349                                  errmsg("unsupported frontend protocol %u.%u: server supports %u.0 to %u.%u",
1350                                           PG_PROTOCOL_MAJOR(proto), PG_PROTOCOL_MINOR(proto),
1351                                                 PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST),
1352                                                 PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST),
1353                                                 PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST))));
1354
1355         /*
1356          * Now fetch parameters out of startup packet and save them into the
1357          * Port structure.      All data structures attached to the Port struct
1358          * must be allocated in TopMemoryContext so that they won't disappear
1359          * when we pass them to PostgresMain (see BackendRun).  We need not
1360          * worry about leaking this storage on failure, since we aren't in the
1361          * postmaster process anymore.
1362          */
1363         oldcontext = MemoryContextSwitchTo(TopMemoryContext);
1364
1365         if (PG_PROTOCOL_MAJOR(proto) >= 3)
1366         {
1367                 int32           offset = sizeof(ProtocolVersion);
1368
1369                 /*
1370                  * Scan packet body for name/option pairs.      We can assume any
1371                  * string beginning within the packet body is null-terminated,
1372                  * thanks to zeroing extra byte above.
1373                  */
1374                 port->guc_options = NIL;
1375
1376                 while (offset < len)
1377                 {
1378                         char       *nameptr = ((char *) buf) + offset;
1379                         int32           valoffset;
1380                         char       *valptr;
1381
1382                         if (*nameptr == '\0')
1383                                 break;                  /* found packet terminator */
1384                         valoffset = offset + strlen(nameptr) + 1;
1385                         if (valoffset >= len)
1386                                 break;                  /* missing value, will complain below */
1387                         valptr = ((char *) buf) + valoffset;
1388
1389                         if (strcmp(nameptr, "database") == 0)
1390                                 port->database_name = pstrdup(valptr);
1391                         else if (strcmp(nameptr, "user") == 0)
1392                                 port->user_name = pstrdup(valptr);
1393                         else if (strcmp(nameptr, "options") == 0)
1394                                 port->cmdline_options = pstrdup(valptr);
1395                         else
1396                         {
1397                                 /* Assume it's a generic GUC option */
1398                                 port->guc_options = lappend(port->guc_options,
1399                                                                                         pstrdup(nameptr));
1400                                 port->guc_options = lappend(port->guc_options,
1401                                                                                         pstrdup(valptr));
1402                         }
1403                         offset = valoffset + strlen(valptr) + 1;
1404                 }
1405
1406                 /*
1407                  * If we didn't find a packet terminator exactly at the end of the
1408                  * given packet length, complain.
1409                  */
1410                 if (offset != len - 1)
1411                         ereport(FATAL,
1412                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
1413                                          errmsg("invalid startup packet layout: expected terminator as last byte")));
1414         }
1415         else
1416         {
1417                 /*
1418                  * Get the parameters from the old-style, fixed-width-fields
1419                  * startup packet as C strings.  The packet destination was
1420                  * cleared first so a short packet has zeros silently added.  We
1421                  * have to be prepared to truncate the pstrdup result for oversize
1422                  * fields, though.
1423                  */
1424                 StartupPacket *packet = (StartupPacket *) buf;
1425
1426                 port->database_name = pstrdup(packet->database);
1427                 if (strlen(port->database_name) > sizeof(packet->database))
1428                         port->database_name[sizeof(packet->database)] = '\0';
1429                 port->user_name = pstrdup(packet->user);
1430                 if (strlen(port->user_name) > sizeof(packet->user))
1431                         port->user_name[sizeof(packet->user)] = '\0';
1432                 port->cmdline_options = pstrdup(packet->options);
1433                 if (strlen(port->cmdline_options) > sizeof(packet->options))
1434                         port->cmdline_options[sizeof(packet->options)] = '\0';
1435                 port->guc_options = NIL;
1436         }
1437
1438         /* Check a user name was given. */
1439         if (port->user_name == NULL || port->user_name[0] == '\0')
1440                 ereport(FATAL,
1441                                 (errcode(ERRCODE_INVALID_AUTHORIZATION_SPECIFICATION),
1442                  errmsg("no PostgreSQL user name specified in startup packet")));
1443
1444         /* The database defaults to the user name. */
1445         if (port->database_name == NULL || port->database_name[0] == '\0')
1446                 port->database_name = pstrdup(port->user_name);
1447
1448         if (Db_user_namespace)
1449         {
1450                 /*
1451                  * If user@, it is a global user, remove '@'. We only want to do
1452                  * this if there is an '@' at the end and no earlier in the user
1453                  * string or they may fake as a local user of another database
1454                  * attaching to this database.
1455                  */
1456                 if (strchr(port->user_name, '@') ==
1457                         port->user_name + strlen(port->user_name) - 1)
1458                         *strchr(port->user_name, '@') = '\0';
1459                 else
1460                 {
1461                         /* Append '@' and dbname */
1462                         char       *db_user;
1463
1464                         db_user = palloc(strlen(port->user_name) +
1465                                                          strlen(port->database_name) + 2);
1466                         sprintf(db_user, "%s@%s", port->user_name, port->database_name);
1467                         port->user_name = db_user;
1468                 }
1469         }
1470
1471         /*
1472          * Truncate given database and user names to length of a Postgres
1473          * name.  This avoids lookup failures when overlength names are given.
1474          */
1475         if (strlen(port->database_name) >= NAMEDATALEN)
1476                 port->database_name[NAMEDATALEN - 1] = '\0';
1477         if (strlen(port->user_name) >= NAMEDATALEN)
1478                 port->user_name[NAMEDATALEN - 1] = '\0';
1479
1480         /*
1481          * Done putting stuff in TopMemoryContext.
1482          */
1483         MemoryContextSwitchTo(oldcontext);
1484
1485         /*
1486          * If we're going to reject the connection due to database state, say
1487          * so now instead of wasting cycles on an authentication exchange.
1488          * (This also allows a pg_ping utility to be written.)
1489          */
1490         switch (port->canAcceptConnections)
1491         {
1492                 case CAC_STARTUP:
1493                         ereport(FATAL,
1494                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
1495                                          errmsg("the database system is starting up")));
1496                         break;
1497                 case CAC_SHUTDOWN:
1498                         ereport(FATAL,
1499                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
1500                                          errmsg("the database system is shutting down")));
1501                         break;
1502                 case CAC_RECOVERY:
1503                         ereport(FATAL,
1504                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
1505                                          errmsg("the database system is in recovery mode")));
1506                         break;
1507                 case CAC_TOOMANY:
1508                         ereport(FATAL,
1509                                         (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
1510                                          errmsg("sorry, too many clients already")));
1511                         break;
1512                 case CAC_OK:
1513                 default:
1514                         break;
1515         }
1516
1517         return STATUS_OK;
1518 }
1519
1520
1521 /*
1522  * The client has sent a cancel request packet, not a normal
1523  * start-a-new-connection packet.  Perform the necessary processing.
1524  * Nothing is sent back to the client.
1525  */
1526 static void
1527 processCancelRequest(Port *port, void *pkt)
1528 {
1529         CancelRequestPacket *canc = (CancelRequestPacket *) pkt;
1530         int                     backendPID;
1531         long            cancelAuthCode;
1532         Backend    *bp;
1533 #ifndef EXEC_BACKEND
1534         Dlelem     *curr;
1535 #else
1536         int i;
1537 #endif
1538
1539         backendPID = (int) ntohl(canc->backendPID);
1540         cancelAuthCode = (long) ntohl(canc->cancelAuthCode);
1541
1542         if (backendPID == CheckPointPID)
1543         {
1544                 ereport(DEBUG2,
1545                                 (errmsg_internal("ignoring cancel request for checkpoint process %d",
1546                                                                  backendPID)));
1547                 return;
1548         }
1549         else if (backendPID == BgWriterPID)
1550         {
1551                 ereport(DEBUG2,
1552                                 (errmsg_internal("ignoring cancel request for bgwriter process %d",
1553                                                                  backendPID)));
1554                 return;
1555         }
1556
1557         /* See if we have a matching backend */
1558 #ifndef EXEC_BACKEND
1559         for (curr = DLGetHead(BackendList); curr; curr = DLGetSucc(curr))
1560         {
1561                 bp = (Backend *) DLE_VAL(curr);
1562 #else
1563         for (i = 0; i < NUM_BACKENDARRAY_ELEMS; i++)
1564         {
1565                 bp = (Backend*) &ShmemBackendArray[i];
1566 #endif
1567                 if (bp->pid == backendPID)
1568                 {
1569                         if (bp->cancel_key == cancelAuthCode)
1570                         {
1571                                 /* Found a match; signal that backend to cancel current op */
1572                                 ereport(DEBUG2,
1573                                                 (errmsg_internal("processing cancel request: sending SIGINT to process %d",
1574                                                                                  backendPID)));
1575                                 kill(bp->pid, SIGINT);
1576                         }
1577                         else
1578                                 /* Right PID, wrong key: no way, Jose */
1579                                 ereport(DEBUG2,
1580                                                 (errmsg_internal("bad key in cancel request for process %d",
1581                                                                                  backendPID)));
1582                         return;
1583                 }
1584         }
1585
1586         /* No matching backend */
1587         ereport(DEBUG2,
1588                         (errmsg_internal("bad pid in cancel request for process %d",
1589                                                          backendPID)));
1590 }
1591
1592 /*
1593  * canAcceptConnections --- check to see if database state allows connections.
1594  */
1595 static enum CAC_state
1596 canAcceptConnections(void)
1597 {
1598         /* Can't start backends when in startup/shutdown/recovery state. */
1599         if (Shutdown > NoShutdown)
1600                 return CAC_SHUTDOWN;
1601         if (StartupPID)
1602                 return CAC_STARTUP;
1603         if (FatalError)
1604                 return CAC_RECOVERY;
1605
1606         /*
1607          * Don't start too many children.
1608          *
1609          * We allow more connections than we can have backends here because some
1610          * might still be authenticating; they might fail auth, or some
1611          * existing backend might exit before the auth cycle is completed. The
1612          * exact MaxBackends limit is enforced when a new backend tries to
1613          * join the shared-inval backend array.
1614          */
1615         if (CountChildren() >= 2 * MaxBackends)
1616                 return CAC_TOOMANY;
1617
1618         return CAC_OK;
1619 }
1620
1621
1622 /*
1623  * ConnCreate -- create a local connection data structure
1624  */
1625 static Port *
1626 ConnCreate(int serverFd)
1627 {
1628         Port       *port;
1629
1630         if (!(port = (Port *) calloc(1, sizeof(Port))))
1631         {
1632                 ereport(LOG,
1633                                 (errcode(ERRCODE_OUT_OF_MEMORY),
1634                                  errmsg("out of memory")));
1635                 ExitPostmaster(1);
1636         }
1637
1638         if (StreamConnection(serverFd, port) != STATUS_OK)
1639         {
1640                 StreamClose(port->sock);
1641                 ConnFree(port);
1642                 port = NULL;
1643         }
1644         else
1645         {
1646                 /*
1647                  * Precompute password salt values to use for this connection.
1648                  * It's slightly annoying to do this long in advance of knowing
1649                  * whether we'll need 'em or not, but we must do the random()
1650                  * calls before we fork, not after.  Else the postmaster's random
1651                  * sequence won't get advanced, and all backends would end up
1652                  * using the same salt...
1653                  */
1654                 RandomSalt(port->cryptSalt, port->md5Salt);
1655         }
1656
1657         return port;
1658 }
1659
1660
1661 /*
1662  * ConnFree -- free a local connection data structure
1663  */
1664 static void
1665 ConnFree(Port *conn)
1666 {
1667 #ifdef USE_SSL
1668         secure_close(conn);
1669 #endif
1670         free(conn);
1671 }
1672
1673
1674 /*
1675  * ClosePostmasterPorts -- close all the postmaster's open sockets
1676  *
1677  * This is called during child process startup to release file descriptors
1678  * that are not needed by that child process.  The postmaster still has
1679  * them open, of course.
1680  */
1681 void
1682 ClosePostmasterPorts(bool pgstat_too)
1683 {
1684         int                     i;
1685
1686         /* Close the listen sockets */
1687         for (i = 0; i < MAXLISTEN; i++)
1688         {
1689                 if (ListenSocket[i] != -1)
1690                 {
1691                         StreamClose(ListenSocket[i]);
1692                         ListenSocket[i] = -1;
1693                 }
1694         }
1695
1696         /* Close pgstat control sockets, unless we're starting pgstat itself */
1697         if (pgstat_too)
1698                 pgstat_close_sockets();
1699 }
1700
1701
1702 /*
1703  * reset_shared -- reset shared memory and semaphores
1704  */
1705 static void
1706 reset_shared(unsigned short port)
1707 {
1708         /*
1709          * Create or re-create shared memory and semaphores.
1710          *
1711          * Note: in each "cycle of life" we will normally assign the same IPC
1712          * keys (if using SysV shmem and/or semas), since the port number is
1713          * used to determine IPC keys.  This helps ensure that we will clean
1714          * up dead IPC objects if the postmaster crashes and is restarted.
1715          */
1716         CreateSharedMemoryAndSemaphores(false, MaxBackends, port);
1717 }
1718
1719
1720 /*
1721  * SIGHUP -- reread config files, and tell children to do same
1722  */
1723 static void
1724 SIGHUP_handler(SIGNAL_ARGS)
1725 {
1726         int                     save_errno = errno;
1727
1728         PG_SETMASK(&BlockSig);
1729
1730         if (Shutdown <= SmartShutdown)
1731         {
1732                 ereport(LOG,
1733                          (errmsg("received SIGHUP, reloading configuration files")));
1734                 ProcessConfigFile(PGC_SIGHUP);
1735 #ifdef EXEC_BACKEND
1736                 write_nondefault_variables(PGC_SIGHUP);
1737 #endif
1738                 SignalChildren(SIGHUP);
1739                 load_hba();
1740                 load_ident();
1741
1742                 /*
1743                  * Tell the background writer to terminate so that we
1744                  * will start a new one with a possibly changed config
1745                  */
1746                 if (BgWriterPID != 0)
1747                         kill(BgWriterPID, SIGTERM);
1748         }
1749
1750         PG_SETMASK(&UnBlockSig);
1751
1752         errno = save_errno;
1753 }
1754
1755
1756
1757 /*
1758  * pmdie -- signal handler for processing various postmaster signals.
1759  */
1760 static void
1761 pmdie(SIGNAL_ARGS)
1762 {
1763         int                     save_errno = errno;
1764
1765         PG_SETMASK(&BlockSig);
1766
1767         ereport(DEBUG2,
1768                         (errmsg_internal("postmaster received signal %d",
1769                                                          postgres_signal_arg)));
1770
1771         switch (postgres_signal_arg)
1772         {
1773                 case SIGTERM:
1774
1775                         /*
1776                          * Smart Shutdown:
1777                          *
1778                          * Wait for children to end their work and ShutdownDataBase.
1779                          */
1780                         if (Shutdown >= SmartShutdown)
1781                                 break;
1782                         Shutdown = SmartShutdown;
1783                         ereport(LOG,
1784                                         (errmsg("received smart shutdown request")));
1785
1786                         /* Must tell bgwriter to quit, or it never will... */
1787                         if (BgWriterPID != 0)
1788                                 kill(BgWriterPID, SIGTERM);
1789
1790                         if (DLGetHead(BackendList)) /* let reaper() handle this */
1791                                 break;
1792
1793                         /*
1794                          * No children left. Shutdown data base system.
1795                          */
1796                         if (StartupPID > 0 || FatalError)       /* let reaper() handle
1797                                                                                                  * this */
1798                                 break;
1799                         if (ShutdownPID > 0)
1800                         {
1801                                 elog(PANIC, "shutdown process %d already running",
1802                                          (int) ShutdownPID);
1803                                 abort();
1804                         }
1805
1806                         ShutdownPID = ShutdownDataBase();
1807                         break;
1808
1809                 case SIGINT:
1810
1811                         /*
1812                          * Fast Shutdown:
1813                          *
1814                          * Abort all children with SIGTERM (rollback active transactions
1815                          * and exit) and ShutdownDataBase when they are gone.
1816                          */
1817                         if (Shutdown >= FastShutdown)
1818                                 break;
1819                         Shutdown = FastShutdown;
1820                         ereport(LOG,
1821                                         (errmsg("received fast shutdown request")));
1822
1823                         if (DLGetHead(BackendList))
1824                         {
1825                                 if (!FatalError)
1826                                 {
1827                                         ereport(LOG,
1828                                                         (errmsg("aborting any active transactions")));
1829                                         SignalChildren(SIGTERM);
1830                                         /* reaper() does the rest */
1831                                 }
1832                                 break;
1833                         }
1834
1835                         /*
1836                          * No children left. Shutdown data base system.
1837                          *
1838                          * Unlike the previous case, it is not an error for the shutdown
1839                          * process to be running already (we could get SIGTERM followed
1840                          * shortly later by SIGINT).
1841                          */
1842                         if (StartupPID > 0 || FatalError)       /* let reaper() handle
1843                                                                                                  * this */
1844                                 break;
1845                         if (ShutdownPID == 0)
1846                                 ShutdownPID = ShutdownDataBase();
1847                         break;
1848
1849                 case SIGQUIT:
1850
1851                         /*
1852                          * Immediate Shutdown:
1853                          *
1854                          * abort all children with SIGQUIT and exit without attempt to
1855                          * properly shutdown data base system.
1856                          */
1857                         ereport(LOG,
1858                                         (errmsg("received immediate shutdown request")));
1859                         if (ShutdownPID > 0)
1860                                 kill(ShutdownPID, SIGQUIT);
1861                         if (StartupPID > 0)
1862                                 kill(StartupPID, SIGQUIT);
1863                         if (DLGetHead(BackendList))
1864                                 SignalChildren(SIGQUIT);
1865                         ExitPostmaster(0);
1866                         break;
1867         }
1868
1869         PG_SETMASK(&UnBlockSig);
1870
1871         errno = save_errno;
1872 }
1873
1874 /*
1875  * Reaper -- signal handler to cleanup after a backend (child) dies.
1876  */
1877 static void
1878 reaper(SIGNAL_ARGS)
1879 {
1880         int                     save_errno = errno;
1881
1882 #ifdef HAVE_WAITPID
1883         int                     status;                 /* backend exit status */
1884 #else
1885 #ifndef WIN32
1886         union wait      status;                 /* backend exit status */
1887 #endif
1888 #endif
1889         int                     exitstatus;
1890         int                     pid;                    /* process id of dead backend */
1891
1892         PG_SETMASK(&BlockSig);
1893
1894         ereport(DEBUG4,
1895                         (errmsg_internal("reaping dead processes")));
1896 #ifdef HAVE_WAITPID
1897         while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
1898         {
1899                 exitstatus = status;
1900 #else
1901 #ifndef WIN32
1902         while ((pid = wait3(&status, WNOHANG, NULL)) > 0)
1903         {
1904                 exitstatus = status.w_status;
1905 #else
1906         while ((pid = win32_waitpid(&exitstatus)) > 0)
1907         {
1908                 /*
1909                  * We need to do this here, and not in CleanupProc, since this
1910                  * is to be called on all children when we are done with them.
1911                  * Could move to LogChildExit, but that seems like asking for
1912                  * future trouble...
1913                  */
1914                 win32_RemoveChild(pid);
1915 #endif
1916 #endif
1917
1918                 /*
1919                  * Check if this child was the statistics collector. If so, try to
1920                  * start a new one.  (If fail, we'll try again in future cycles of
1921                  * the main loop.)
1922                  */
1923                 if (pgstat_ispgstat(pid))
1924                 {
1925                         LogChildExit(LOG, gettext("statistics collector process"),
1926                                                  pid, exitstatus);
1927                         pgstat_start();
1928                         continue;
1929                 }
1930
1931                 /*
1932                  * Check if this child was a shutdown or startup process.
1933                  */
1934                 if (ShutdownPID > 0 && pid == ShutdownPID)
1935                 {
1936                         if (exitstatus != 0)
1937                         {
1938                                 LogChildExit(LOG, gettext("shutdown process"),
1939                                                          pid, exitstatus);
1940                                 ExitPostmaster(1);
1941                         }
1942                         /* Normal postmaster exit is here */
1943                         ExitPostmaster(0);
1944                 }
1945
1946                 if (StartupPID > 0 && pid == StartupPID)
1947                 {
1948                         if (exitstatus != 0)
1949                         {
1950                                 LogChildExit(LOG, gettext("startup process"),
1951                                                          pid, exitstatus);
1952                                 ereport(LOG,
1953                                                 (errmsg("aborting startup due to startup process failure")));
1954                                 ExitPostmaster(1);
1955                         }
1956                         StartupPID = 0;
1957
1958                         /*
1959                          * Startup succeeded - remember its ID and RedoRecPtr.
1960                          *
1961                          * NB: this MUST happen before we fork a checkpoint or shutdown
1962                          * subprocess, else they will have wrong local ThisStartUpId.
1963                          */
1964                         SetThisStartUpID();
1965
1966                         FatalError = false; /* done with recovery */
1967
1968                         /*
1969                          * Arrange for first checkpoint to occur after standard delay.
1970                          */
1971                         CheckPointPID = 0;
1972                         checkpointed = time(NULL);
1973
1974                         /*
1975                          * Go to shutdown mode if a shutdown request was pending.
1976                          */
1977                         if (Shutdown > NoShutdown)
1978                         {
1979                                 if (ShutdownPID > 0)
1980                                 {
1981                                         elog(PANIC, "startup process %d died while shutdown process %d already running",
1982                                                  pid, (int) ShutdownPID);
1983                                         abort();
1984                                 }
1985                                 ShutdownPID = ShutdownDataBase();
1986                         }
1987
1988                         goto reaper_done;
1989                 }
1990
1991                 /*
1992                  * Else do standard child cleanup.
1993                  */
1994                 CleanupProc(pid, exitstatus);
1995
1996         }                                                       /* loop over pending child-death reports */
1997
1998         if (FatalError)
1999         {
2000                 /*
2001                  * Wait for all children exit, then reset shmem and
2002                  * StartupDataBase.
2003                  */
2004                 if (DLGetHead(BackendList) || StartupPID > 0 || ShutdownPID > 0)
2005                         goto reaper_done;
2006                 ereport(LOG,
2007                         (errmsg("all server processes terminated; reinitializing")));
2008
2009                 shmem_exit(0);
2010                 reset_shared(PostPortNumber);
2011
2012                 StartupPID = StartupDataBase();
2013
2014                 goto reaper_done;
2015         }
2016
2017         if (Shutdown > NoShutdown)
2018         {
2019                 if (DLGetHead(BackendList))
2020                         goto reaper_done;
2021                 if (StartupPID > 0 || ShutdownPID > 0)
2022                         goto reaper_done;
2023                 ShutdownPID = ShutdownDataBase();
2024         }
2025
2026 reaper_done:
2027         PG_SETMASK(&UnBlockSig);
2028
2029         errno = save_errno;
2030 }
2031
2032 /*
2033  * CleanupProc -- cleanup after terminated backend.
2034  *
2035  * Remove all local state associated with backend.
2036  */
2037 static void
2038 CleanupProc(int pid,
2039                         int exitstatus)         /* child's exit status. */
2040 {
2041         Dlelem     *curr,
2042                            *next;
2043         Backend    *bp;
2044
2045         LogChildExit(DEBUG2, gettext("child process"), pid, exitstatus);
2046
2047         /*
2048          * If a backend dies in an ugly way (i.e. exit status not 0) then we
2049          * must signal all other backends to quickdie.  If exit status is zero
2050          * we assume everything is hunky dory and simply remove the backend
2051          * from the active backend list.
2052          */
2053         if (exitstatus == 0)
2054         {
2055                 curr = DLGetHead(BackendList);
2056                 while (curr)
2057                 {
2058                         bp = (Backend *) DLE_VAL(curr);
2059                         if (bp->pid == pid)
2060                         {
2061 #ifdef EXEC_BACKEND
2062                                 ShmemBackendArrayRemove(bp->pid);
2063 #endif
2064                                 DLRemove(curr);
2065                                 free(bp);
2066                                 DLFreeElem(curr);
2067                                 break;
2068                         }
2069                         curr = DLGetSucc(curr);
2070                 }
2071
2072                 if (pid == CheckPointPID)
2073                 {
2074                         CheckPointPID = 0;
2075                         if (!FatalError)
2076                         {
2077                                 checkpointed = time(NULL);
2078                                 /* Update RedoRecPtr for future child backends */
2079                                 GetSavedRedoRecPtr();
2080                         }
2081                 }
2082                 else if (pid == BgWriterPID)
2083                         BgWriterPID = 0;
2084                 else
2085                         pgstat_beterm(pid);
2086
2087                 return;
2088         }
2089
2090         /* below here we're dealing with a non-normal exit */
2091
2092         /* Make log entry unless we did so already */
2093         if (!FatalError)
2094         {
2095                 LogChildExit(LOG,
2096                                          (pid == CheckPointPID) ? gettext("checkpoint process") :
2097                                          (pid == BgWriterPID) ? gettext("bgwriter process") :
2098                                          gettext("server process"),
2099                                          pid, exitstatus);
2100                 ereport(LOG,
2101                           (errmsg("terminating any other active server processes")));
2102         }
2103
2104         curr = DLGetHead(BackendList);
2105         while (curr)
2106         {
2107                 next = DLGetSucc(curr);
2108                 bp = (Backend *) DLE_VAL(curr);
2109                 if (bp->pid != pid)
2110                 {
2111                         /*
2112                          * This backend is still alive.  Unless we did so already,
2113                          * tell it to commit hara-kiri.
2114                          *
2115                          * SIGQUIT is the special signal that says exit without proc_exit
2116                          * and let the user know what's going on. But if SendStop is
2117                          * set (-s on command line), then we send SIGSTOP instead, so
2118                          * that we can get core dumps from all backends by hand.
2119                          */
2120                         if (!FatalError)
2121                         {
2122                                 ereport(DEBUG2,
2123                                                 (errmsg_internal("sending %s to process %d",
2124                                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
2125                                                                                  (int) bp->pid)));
2126                                 kill(bp->pid, (SendStop ? SIGSTOP : SIGQUIT));
2127                         }
2128                 }
2129                 else
2130                 {
2131                         /*
2132                          * Found entry for freshly-dead backend, so remove it.
2133                          */
2134 #ifdef EXEC_BACKEND
2135                         ShmemBackendArrayRemove(bp->pid);
2136 #endif
2137                         DLRemove(curr);
2138                         free(bp);
2139                         DLFreeElem(curr);
2140                 }
2141                 curr = next;
2142         }
2143
2144         if (pid == CheckPointPID)
2145         {
2146                 CheckPointPID = 0;
2147                 checkpointed = 0;
2148         }
2149         else if (pid == BgWriterPID)
2150         {
2151                 BgWriterPID = 0;
2152         }
2153         else
2154         {
2155                 /*
2156                  * Tell the collector about backend termination
2157                  */
2158                 pgstat_beterm(pid);
2159         }
2160
2161         FatalError = true;
2162 }
2163
2164 /*
2165  * Log the death of a child process.
2166  */
2167 static void
2168 LogChildExit(int lev, const char *procname, int pid, int exitstatus)
2169 {
2170         if (WIFEXITED(exitstatus))
2171                 ereport(lev,
2172
2173                 /*
2174                  * translator: %s is a noun phrase describing a child process,
2175                  * such as "server process"
2176                  */
2177                                 (errmsg("%s (PID %d) exited with exit code %d",
2178                                                 procname, pid, WEXITSTATUS(exitstatus))));
2179         else if (WIFSIGNALED(exitstatus))
2180                 ereport(lev,
2181
2182                 /*
2183                  * translator: %s is a noun phrase describing a child process,
2184                  * such as "server process"
2185                  */
2186                                 (errmsg("%s (PID %d) was terminated by signal %d",
2187                                                 procname, pid, WTERMSIG(exitstatus))));
2188         else
2189                 ereport(lev,
2190
2191                 /*
2192                  * translator: %s is a noun phrase describing a child process,
2193                  * such as "server process"
2194                  */
2195                                 (errmsg("%s (PID %d) exited with unexpected status %d",
2196                                                 procname, pid, exitstatus)));
2197 }
2198
2199 /*
2200  * Send a signal to all backend children.
2201  */
2202 static void
2203 SignalChildren(int signal)
2204 {
2205         Dlelem     *curr,
2206                            *next;
2207         Backend    *bp;
2208
2209         curr = DLGetHead(BackendList);
2210         while (curr)
2211         {
2212                 next = DLGetSucc(curr);
2213                 bp = (Backend *) DLE_VAL(curr);
2214
2215                 if (bp->pid != MyProcPid)
2216                 {
2217                         ereport(DEBUG2,
2218                                         (errmsg_internal("sending signal %d to process %d",
2219                                                                          signal,
2220                                                                          (int) bp->pid)));
2221                         kill(bp->pid, signal);
2222                 }
2223
2224                 curr = next;
2225         }
2226 }
2227
2228 /*
2229  * BackendStartup -- start backend process
2230  *
2231  * returns: STATUS_ERROR if the fork failed, STATUS_OK otherwise.
2232  */
2233 static int
2234 BackendStartup(Port *port)
2235 {
2236         Backend    *bn;                         /* for backend cleanup */
2237         pid_t           pid;
2238
2239 #ifdef LINUX_PROFILE
2240         struct itimerval prof_itimer;
2241 #endif
2242
2243         /*
2244          * Compute the cancel key that will be assigned to this backend. The
2245          * backend will have its own copy in the forked-off process' value of
2246          * MyCancelKey, so that it can transmit the key to the frontend.
2247          */
2248         MyCancelKey = PostmasterRandom();
2249
2250         /*
2251          * Make room for backend data structure.  Better before the fork() so
2252          * we can handle failure cleanly.
2253          */
2254         bn = (Backend *) malloc(sizeof(Backend));
2255         if (!bn)
2256         {
2257                 ereport(LOG,
2258                                 (errcode(ERRCODE_OUT_OF_MEMORY),
2259                                  errmsg("out of memory")));
2260                 return STATUS_ERROR;
2261         }
2262
2263         /*
2264          * Flush stdio channels just before fork, to avoid double-output
2265          * problems. Ideally we'd use fflush(NULL) here, but there are still a
2266          * few non-ANSI stdio libraries out there (like SunOS 4.1.x) that
2267          * coredump if we do. Presently stdout and stderr are the only stdio
2268          * output channels used by the postmaster, so fflush'ing them should
2269          * be sufficient.
2270          */
2271         fflush(stdout);
2272         fflush(stderr);
2273
2274 #ifdef LINUX_PROFILE
2275
2276         /*
2277          * Linux's fork() resets the profiling timer in the child process. If
2278          * we want to profile child processes then we need to save and restore
2279          * the timer setting.  This is a waste of time if not profiling,
2280          * however, so only do it if commanded by specific -DLINUX_PROFILE
2281          * switch.
2282          */
2283         getitimer(ITIMER_PROF, &prof_itimer);
2284 #endif
2285
2286 #ifdef __BEOS__
2287         /* Specific beos actions before backend startup */
2288         beos_before_backend_startup();
2289 #endif
2290
2291         port->canAcceptConnections = canAcceptConnections();
2292 #ifdef EXEC_BACKEND
2293         pid = Backend_forkexec(port);
2294 #else
2295         pid = fork();
2296
2297         if (pid == 0)                           /* child */
2298         {
2299 #ifdef LINUX_PROFILE
2300                 setitimer(ITIMER_PROF, &prof_itimer, NULL);
2301 #endif
2302
2303 #ifdef __BEOS__
2304                 /* Specific beos backend startup actions */
2305                 beos_backend_startup();
2306 #endif
2307                 free(bn);
2308
2309                 proc_exit(BackendRun(port));
2310         }
2311 #endif
2312
2313         /* in parent, error */
2314         if (pid < 0)
2315         {
2316                 int                     save_errno = errno;
2317
2318 #ifdef __BEOS__
2319                 /* Specific beos backend startup actions */
2320                 beos_backend_startup_failed();
2321 #endif
2322                 free(bn);
2323                 errno = save_errno;
2324                 ereport(LOG,
2325                           (errmsg("could not fork new process for connection: %m")));
2326                 report_fork_failure_to_client(port, save_errno);
2327                 return STATUS_ERROR;
2328         }
2329
2330         /* in parent, normal */
2331         ereport(DEBUG2,
2332                         (errmsg_internal("forked new backend, pid=%d socket=%d",
2333                                                          (int) pid, port->sock)));
2334
2335         /*
2336          * Everything's been successful, it's safe to add this backend to our
2337          * list of backends.
2338          */
2339         bn->pid = pid;
2340         bn->cancel_key = MyCancelKey;
2341 #ifdef EXEC_BACKEND
2342         ShmemBackendArrayAdd(bn);
2343 #endif
2344         DLAddHead(BackendList, DLNewElem(bn));
2345
2346         return STATUS_OK;
2347 }
2348
2349 /*
2350  * Try to report backend fork() failure to client before we close the
2351  * connection.  Since we do not care to risk blocking the postmaster on
2352  * this connection, we set the connection to non-blocking and try only once.
2353  *
2354  * This is grungy special-purpose code; we cannot use backend libpq since
2355  * it's not up and running.
2356  */
2357 static void
2358 report_fork_failure_to_client(Port *port, int errnum)
2359 {
2360         char            buffer[1000];
2361
2362         /* Format the error message packet (always V2 protocol) */
2363         snprintf(buffer, sizeof(buffer), "E%s%s\n",
2364                          gettext("could not fork new process for connection: "),
2365                          strerror(errnum));
2366
2367         /* Set port to non-blocking.  Don't do send() if this fails */
2368         if (FCNTL_NONBLOCK(port->sock) < 0)
2369                 return;
2370
2371         send(port->sock, buffer, strlen(buffer) + 1, 0);
2372 }
2373
2374
2375 /*
2376  * split_opts -- split a string of options and append it to an argv array
2377  *
2378  * NB: the string is destructively modified!
2379  *
2380  * Since no current POSTGRES arguments require any quoting characters,
2381  * we can use the simple-minded tactic of assuming each set of space-
2382  * delimited characters is a separate argv element.
2383  *
2384  * If you don't like that, well, we *used* to pass the whole option string
2385  * as ONE argument to execl(), which was even less intelligent...
2386  */
2387 static void
2388 split_opts(char **argv, int *argcp, char *s)
2389 {
2390         while (s && *s)
2391         {
2392                 while (isspace((unsigned char) *s))
2393                         ++s;
2394                 if (*s == '\0')
2395                         break;
2396                 argv[(*argcp)++] = s;
2397                 while (*s && !isspace((unsigned char) *s))
2398                         ++s;
2399                 if (*s)
2400                         *s++ = '\0';
2401         }
2402 }
2403
2404
2405 /*
2406  * BackendInit/Run -- perform authentication [BackendInit], and if successful,
2407  *              set up the backend's argument list [BackendRun] and invoke
2408  *              backend main()
2409  *
2410  * returns:
2411  *              Shouldn't return at all.
2412  *              If PostgresMain() fails, return status.
2413  */
2414 static void
2415 BackendInit(Port *port)
2416 {
2417         int                     status;
2418         struct timeval now;
2419         struct timezone tz;
2420         char            remote_host[NI_MAXHOST];
2421         char            remote_port[NI_MAXSERV];
2422
2423         IsUnderPostmaster = true;       /* we are a postmaster subprocess now */
2424
2425         ClientAuthInProgress = true;    /* limit visibility of log messages */
2426
2427         /* We don't want the postmaster's proc_exit() handlers */
2428         on_exit_reset();
2429
2430         /*
2431          * Signal handlers setting is moved to tcop/postgres...
2432          */
2433
2434         /* save start time for end of session reporting */
2435         gettimeofday(&(port->session_start),NULL);
2436
2437         /* set these to empty in case they are needed before we set them up */
2438         port->remote_host = "";
2439         port->remote_port = "";
2440         port->commandTag = "";
2441
2442         /* Save port etc. for ps status */
2443         MyProcPort = port;
2444
2445         /* Reset MyProcPid to new backend's pid */
2446         MyProcPid = getpid();
2447
2448         /*
2449          * Initialize libpq and enable reporting of ereport errors to the
2450          * client. Must do this now because authentication uses libpq to send
2451          * messages.
2452          */
2453         pq_init();                                      /* initialize libpq to talk to client */
2454         whereToSendOutput = Remote; /* now safe to ereport to client */
2455
2456         /*
2457          * We arrange for a simple exit(0) if we receive SIGTERM or SIGQUIT
2458          * during any client authentication related communication. Otherwise
2459          * the postmaster cannot shutdown the database FAST or IMMED cleanly
2460          * if a buggy client blocks a backend during authentication.
2461          */
2462         pqsignal(SIGTERM, authdie);
2463         pqsignal(SIGQUIT, authdie);
2464         pqsignal(SIGALRM, authdie);
2465         PG_SETMASK(&AuthBlockSig);
2466
2467         /*
2468          * Get the remote host name and port for logging and status display.
2469          */
2470         remote_host[0] = '\0';
2471         remote_port[0] = '\0';
2472         if (getnameinfo_all(&port->raddr.addr, port->raddr.salen,
2473                                                 remote_host, sizeof(remote_host),
2474                                                 remote_port, sizeof(remote_port),
2475                                    (log_hostname ? 0 : NI_NUMERICHOST) | NI_NUMERICSERV))
2476         {
2477                 getnameinfo_all(&port->raddr.addr, port->raddr.salen,
2478                                                 remote_host, sizeof(remote_host),
2479                                                 remote_port, sizeof(remote_port),
2480                                                 NI_NUMERICHOST | NI_NUMERICSERV);
2481         }
2482
2483         if (Log_connections)
2484                 ereport(LOG,
2485                                 (errmsg("connection received: host=%s port=%s",
2486                                                 remote_host, remote_port)));
2487
2488         if (LogSourcePort)
2489         {
2490                 /* modify remote_host for use in ps status */
2491                 char            tmphost[NI_MAXHOST];
2492
2493                 snprintf(tmphost, sizeof(tmphost), "%s(%s)", remote_host, remote_port);
2494                 StrNCpy(remote_host, tmphost, sizeof(remote_host));
2495         }
2496
2497         /*
2498          * save remote_host and remote_port in port stucture
2499          */
2500         port->remote_host = strdup(remote_host);
2501         port->remote_port = strdup(remote_port);
2502
2503         /*
2504          * Ready to begin client interaction.  We will give up and exit(0)
2505          * after a time delay, so that a broken client can't hog a connection
2506          * indefinitely.  PreAuthDelay doesn't count against the time limit.
2507          */
2508         if (!enable_sig_alarm(AuthenticationTimeout * 1000, false))
2509                 elog(FATAL, "could not set timer for authorization timeout");
2510
2511         /*
2512          * Receive the startup packet (which might turn out to be a cancel
2513          * request packet).
2514          */
2515         status = ProcessStartupPacket(port, false);
2516
2517         if (status != STATUS_OK)
2518                 proc_exit(0);
2519
2520         /*
2521          * Now that we have the user and database name, we can set the process
2522          * title for ps.  It's good to do this as early as possible in
2523          * startup.
2524          */
2525         init_ps_display(port->user_name, port->database_name, remote_host);
2526         set_ps_display("authentication");
2527
2528         /*
2529          * Now perform authentication exchange.
2530          */
2531         ClientAuthentication(port); /* might not return, if failure */
2532
2533         /*
2534          * Done with authentication.  Disable timeout, and prevent
2535          * SIGTERM/SIGQUIT again until backend startup is complete.
2536          */
2537         if (!disable_sig_alarm(false))
2538                 elog(FATAL, "could not disable timer for authorization timeout");
2539         PG_SETMASK(&BlockSig);
2540
2541         if (Log_connections)
2542                 ereport(LOG,
2543                                 (errmsg("connection authorized: user=%s database=%s",
2544                                                 port->user_name, port->database_name)));
2545
2546         /*
2547          * Don't want backend to be able to see the postmaster random number
2548          * generator state.  We have to clobber the static random_seed *and*
2549          * start a new random sequence in the random() library function.
2550          */
2551         random_seed = 0;
2552         gettimeofday(&now, &tz);
2553         srandom((unsigned int) now.tv_usec);
2554 }
2555
2556
2557 static int
2558 BackendRun(Port *port)
2559 {
2560         char      **av;
2561         int                     maxac;
2562         int                     ac;
2563         char            debugbuf[32];
2564         char            protobuf[32];
2565         int                     i;
2566
2567         /*
2568          * Let's clean up ourselves as the postmaster child, and
2569          * close the postmaster's other sockets
2570          */
2571         ClosePostmasterPorts(true);
2572
2573         /*
2574          * PreAuthDelay is a debugging aid for investigating problems in the
2575          * authentication cycle: it can be set in postgresql.conf to allow
2576          * time to attach to the newly-forked backend with a debugger. (See
2577          * also the -W backend switch, which we allow clients to pass through
2578          * PGOPTIONS, but it is not honored until after authentication.)
2579          */
2580         if (PreAuthDelay > 0)
2581                 sleep(PreAuthDelay);
2582
2583         /* Will exit on failure */
2584         BackendInit(port);
2585
2586
2587         /* ----------------
2588          * Now, build the argv vector that will be given to PostgresMain.
2589          *
2590          * The layout of the command line is
2591          *              postgres [secure switches] -p databasename [insecure switches]
2592          * where the switches after -p come from the client request.
2593          *
2594          * The maximum possible number of commandline arguments that could come
2595          * from ExtraOptions or port->cmdline_options is (strlen + 1) / 2; see
2596          * split_opts().
2597          * ----------------
2598          */
2599         maxac = 10;                                     /* for fixed args supplied below */
2600         maxac += (strlen(ExtraOptions) + 1) / 2;
2601         if (port->cmdline_options)
2602                 maxac += (strlen(port->cmdline_options) + 1) / 2;
2603
2604         av = (char **) MemoryContextAlloc(TopMemoryContext,
2605                                                                           maxac * sizeof(char *));
2606         ac = 0;
2607
2608         av[ac++] = "postgres";
2609
2610         /*
2611          * Pass the requested debugging level along to the backend.
2612          */
2613         if (debug_flag > 0)
2614         {
2615                 snprintf(debugbuf, sizeof(debugbuf), "-d%d", debug_flag);
2616                 av[ac++] = debugbuf;
2617         }
2618
2619         /*
2620          * Pass any backend switches specified with -o in the postmaster's own
2621          * command line.  We assume these are secure.
2622          */
2623         split_opts(av, &ac, ExtraOptions);
2624
2625         /* Tell the backend what protocol the frontend is using. */
2626         snprintf(protobuf, sizeof(protobuf), "-v%u", port->proto);
2627         av[ac++] = protobuf;
2628
2629 #ifdef EXEC_BACKEND
2630         /* pass data dir before end of secure switches (-p) */
2631         av[ac++] = "-D";
2632         av[ac++] = DataDir;
2633 #endif
2634
2635         /*
2636          * Tell the backend it is being called from the postmaster, and which
2637          * database to use.  -p marks the end of secure switches.
2638          */
2639         av[ac++] = "-p";
2640         av[ac++] = port->database_name;
2641
2642         /*
2643          * Pass the (insecure) option switches from the connection request.
2644          * (It's OK to mangle port->cmdline_options now.)
2645          */
2646         if (port->cmdline_options)
2647                 split_opts(av, &ac, port->cmdline_options);
2648
2649         av[ac] = NULL;
2650
2651         Assert(ac < maxac);
2652
2653         /*
2654          * Release postmaster's working memory context so that backend can
2655          * recycle the space.  Note this does not trash *MyProcPort, because
2656          * ConnCreate() allocated that space with malloc() ... else we'd need
2657          * to copy the Port data here.  Also, subsidiary data such as the
2658          * username isn't lost either; see ProcessStartupPacket().
2659          */
2660         MemoryContextSwitchTo(TopMemoryContext);
2661 #ifndef EXEC_BACKEND
2662         MemoryContextDelete(PostmasterContext);
2663 #endif
2664         PostmasterContext = NULL;
2665
2666         /*
2667          * Debug: print arguments being passed to backend
2668          */
2669         ereport(DEBUG3,
2670                         (errmsg_internal("%s child[%d]: starting with (",
2671                                                          progname, getpid())));
2672         for (i = 0; i < ac; ++i)
2673                 ereport(DEBUG3,
2674                                 (errmsg_internal("\t%s", av[i])));
2675         ereport(DEBUG3,
2676                         (errmsg_internal(")")));
2677
2678         ClientAuthInProgress = false;           /* client_min_messages is active
2679                                                                                  * now */
2680
2681         return (PostgresMain(ac, av, port->user_name));
2682 }
2683
2684
2685 #ifdef EXEC_BACKEND
2686
2687
2688 /*
2689  * SubPostmasterMain -- prepare the fork/exec'd process to be in an equivalent
2690  *                      state (for calling BackendRun) as a forked process.
2691  *
2692  * returns:
2693  *              Shouldn't return at all.
2694  */
2695 void
2696 SubPostmasterMain(int argc, char* argv[])
2697 {
2698         unsigned long   backendID;
2699         Port                    port;
2700
2701         memset((void*)&port, 0, sizeof(Port));
2702         Assert(argc == 2);
2703
2704         /* Do this sooner rather than later... */
2705         IsUnderPostmaster = true;       /* we are a postmaster subprocess now */
2706
2707         /* In EXEC case we will not have inherited these settings */
2708         IsPostmasterEnvironment = true;
2709         whereToSendOutput = None;
2710
2711         /* Setup global context */
2712         MemoryContextInit();
2713         InitializeGUCOptions();
2714
2715         /* Parse passed-in context */
2716         argc = 0;
2717         backendID               = (unsigned long)atol(argv[argc++]);
2718         DataDir                 = strdup(argv[argc++]);
2719
2720         /* Read in file-based context */
2721         read_nondefault_variables();
2722         read_backend_variables(backendID,&port);
2723
2724         /* FIXME: [fork/exec] Ugh */
2725         load_hba();
2726         load_ident();
2727         load_user();
2728         load_group();
2729
2730         /* Attach process to shared segments */
2731         CreateSharedMemoryAndSemaphores(false, MaxBackends, 0);
2732
2733         /* Run backend */
2734         proc_exit(BackendRun(&port));
2735 }
2736
2737
2738 /*
2739  * Backend_forkexec -- fork/exec off a backend process
2740  *
2741  * returns:
2742  *              the pid of the fork/exec'd process
2743  */
2744 static pid_t
2745 Backend_forkexec(Port *port)
2746 {
2747         pid_t pid;
2748         char *av[5];
2749         int ac = 0, bufc = 0, i;
2750         char buf[2][MAXPGPATH];
2751
2752         if (!write_backend_variables(port))
2753                 return -1; /* log made by write_backend_variables */
2754
2755         av[ac++] = "postgres";
2756         av[ac++] = "-forkexec";
2757
2758         /* Format up context to pass to exec'd process */
2759         snprintf(buf[bufc++],MAXPGPATH,"%lu",tmpBackendFileNum);
2760         /* FIXME: [fork/exec] whitespaces in directories? */
2761         snprintf(buf[bufc++],MAXPGPATH,"%s",DataDir);
2762
2763         /* Add to the arg list */
2764         Assert(bufc <= lengthof(buf));
2765         for (i = 0; i < bufc; i++)
2766                 av[ac++] = buf[i];
2767
2768         /* FIXME: [fork/exec] ExtraOptions? */
2769
2770         av[ac++] = NULL;
2771         Assert(ac <= lengthof(av));
2772
2773 #ifdef WIN32
2774         pid = win32_forkexec(pg_pathname,av); /* logs on error */
2775 #else
2776         /* Fire off execv in child */
2777         if ((pid = fork()) == 0 && (execv(pg_pathname,av) == -1))
2778                 /*
2779                  * FIXME: [fork/exec] suggestions for what to do here?
2780                  *  Probably OK to issue error (unlike pgstat case)
2781                  */
2782                 abort();
2783 #endif
2784         return pid; /* Parent returns pid */
2785 }
2786
2787 #endif
2788
2789
2790 /*
2791  * ExitPostmaster -- cleanup
2792  *
2793  * Do NOT call exit() directly --- always go through here!
2794  */
2795 static void
2796 ExitPostmaster(int status)
2797 {
2798         /* should cleanup shared memory and kill all backends */
2799
2800         /*
2801          * Not sure of the semantics here.      When the Postmaster dies, should
2802          * the backends all be killed? probably not.
2803          *
2804          * MUST         -- vadim 05-10-1999
2805          */
2806         /* Should I use true instead? */
2807         ClosePostmasterPorts(false);
2808
2809         proc_exit(status);
2810 }
2811
2812 /*
2813  * sigusr1_handler - handle signal conditions from child processes
2814  */
2815 static void
2816 sigusr1_handler(SIGNAL_ARGS)
2817 {
2818         int                     save_errno = errno;
2819
2820         PG_SETMASK(&BlockSig);
2821
2822         if (CheckPostmasterSignal(PMSIGNAL_DO_CHECKPOINT))
2823         {
2824                 if (CheckPointWarning != 0)
2825                 {
2826                         /*
2827                          * This only times checkpoints forced by running out of
2828                          * segment files.  Other checkpoints could reduce the
2829                          * frequency of forced checkpoints.
2830                          */
2831                         time_t          now = time(NULL);
2832
2833                         if (LastSignalledCheckpoint != 0)
2834                         {
2835                                 int                     elapsed_secs = now - LastSignalledCheckpoint;
2836
2837                                 if (elapsed_secs < CheckPointWarning)
2838                                         ereport(LOG,
2839                                                         (errmsg("checkpoints are occurring too frequently (%d seconds apart)",
2840                                                                         elapsed_secs),
2841                                         errhint("Consider increasing the configuration parameter \"checkpoint_segments\".")));
2842                         }
2843                         LastSignalledCheckpoint = now;
2844                 }
2845
2846                 /*
2847                  * Request to schedule a checkpoint
2848                  *
2849                  * Ignore request if checkpoint is already running or checkpointing
2850                  * is currently disabled
2851                  */
2852                 if (CheckPointPID == 0 && checkpointed &&
2853                         StartupPID == 0 && Shutdown == NoShutdown &&
2854                         !FatalError && random_seed != 0)
2855                 {
2856                         CheckPointPID = CheckPointDataBase();
2857                         /* note: if fork fails, CheckPointPID stays 0; nothing happens */
2858                 }
2859         }
2860
2861         if (CheckPostmasterSignal(PMSIGNAL_PASSWORD_CHANGE))
2862         {
2863                 /*
2864                  * Password or group file has changed.
2865                  */
2866                 load_user();
2867                 load_group();
2868         }
2869
2870         if (CheckPostmasterSignal(PMSIGNAL_WAKEN_CHILDREN))
2871         {
2872                 /*
2873                  * Send SIGUSR2 to all children (triggers AsyncNotifyHandler). See
2874                  * storage/ipc/sinvaladt.c for the use of this.
2875                  */
2876                 if (Shutdown == NoShutdown)
2877                         SignalChildren(SIGUSR2);
2878         }
2879
2880         PG_SETMASK(&UnBlockSig);
2881
2882         errno = save_errno;
2883 }
2884
2885
2886 /*
2887  * Dummy signal handler
2888  *
2889  * We use this for signals that we don't actually use in the postmaster,
2890  * but we do use in backends.  If we SIG_IGN such signals in the postmaster,
2891  * then a newly started backend might drop a signal that arrives before it's
2892  * able to reconfigure its signal processing.  (See notes in postgres.c.)
2893  */
2894 static void
2895 dummy_handler(SIGNAL_ARGS)
2896 {
2897 }
2898
2899
2900 /*
2901  * CharRemap: given an int in range 0..61, produce textual encoding of it
2902  * per crypt(3) conventions.
2903  */
2904 static char
2905 CharRemap(long ch)
2906 {
2907         if (ch < 0)
2908                 ch = -ch;
2909         ch = ch % 62;
2910
2911         if (ch < 26)
2912                 return 'A' + ch;
2913
2914         ch -= 26;
2915         if (ch < 26)
2916                 return 'a' + ch;
2917
2918         ch -= 26;
2919         return '0' + ch;
2920 }
2921
2922 /*
2923  * RandomSalt
2924  */
2925 static void
2926 RandomSalt(char *cryptSalt, char *md5Salt)
2927 {
2928         long            rand = PostmasterRandom();
2929
2930         cryptSalt[0] = CharRemap(rand % 62);
2931         cryptSalt[1] = CharRemap(rand / 62);
2932
2933         /*
2934          * It's okay to reuse the first random value for one of the MD5 salt
2935          * bytes, since only one of the two salts will be sent to the client.
2936          * After that we need to compute more random bits.
2937          *
2938          * We use % 255, sacrificing one possible byte value, so as to ensure
2939          * that all bits of the random() value participate in the result.
2940          * While at it, add one to avoid generating any null bytes.
2941          */
2942         md5Salt[0] = (rand % 255) + 1;
2943         rand = PostmasterRandom();
2944         md5Salt[1] = (rand % 255) + 1;
2945         rand = PostmasterRandom();
2946         md5Salt[2] = (rand % 255) + 1;
2947         rand = PostmasterRandom();
2948         md5Salt[3] = (rand % 255) + 1;
2949 }
2950
2951 /*
2952  * PostmasterRandom
2953  */
2954 static long
2955 PostmasterRandom(void)
2956 {
2957         static bool initialized = false;
2958
2959         if (!initialized)
2960         {
2961                 Assert(random_seed != 0);
2962                 srandom(random_seed);
2963                 initialized = true;
2964         }
2965
2966         return random();
2967 }
2968
2969 /*
2970  * Count up number of child processes.
2971  */
2972 static int
2973 CountChildren(void)
2974 {
2975         Dlelem     *curr;
2976         Backend    *bp;
2977         int                     cnt = 0;
2978
2979         for (curr = DLGetHead(BackendList); curr; curr = DLGetSucc(curr))
2980         {
2981                 bp = (Backend *) DLE_VAL(curr);
2982                 if (bp->pid != MyProcPid)
2983                         cnt++;
2984         }
2985         /* Checkpoint and bgwriter will be in the list, discount them */
2986         if (CheckPointPID != 0)
2987                 cnt--;
2988         if (BgWriterPID != 0)
2989                 cnt--;
2990         return cnt;
2991 }
2992
2993 /*
2994  * Fire off a subprocess for startup/shutdown/checkpoint/bgwriter.
2995  *
2996  * Return value of SSDataBase is subprocess' PID, or 0 if failed to start subprocess
2997  * (0 is returned only for checkpoint/bgwriter cases).
2998  *
2999  * note: in the EXEC_BACKEND case, we delay the fork until argument list has been
3000  *      established
3001  */
3002 NON_EXEC_STATIC void
3003 SSDataBaseInit(int xlop)
3004 {
3005         const char *statmsg;
3006
3007         IsUnderPostmaster = true;               /* we are a postmaster subprocess
3008                                                                          * now */
3009
3010 #ifdef EXEC_BACKEND
3011         /* In EXEC case we will not have inherited these settings */
3012         IsPostmasterEnvironment = true;
3013         whereToSendOutput = None;
3014 #endif
3015
3016         MyProcPid = getpid();           /* reset MyProcPid */
3017
3018         /* Lose the postmaster's on-exit routines and port connections */
3019         on_exit_reset();
3020
3021         /*
3022          * Identify myself via ps
3023          */
3024         switch (xlop)
3025         {
3026                 case BS_XLOG_STARTUP:
3027                         statmsg = "startup subprocess";
3028                         break;
3029                 case BS_XLOG_CHECKPOINT:
3030                         statmsg = "checkpoint subprocess";
3031                         break;
3032                 case BS_XLOG_BGWRITER:
3033                         statmsg = "bgwriter subprocess";
3034                         break;
3035                 case BS_XLOG_SHUTDOWN:
3036                         statmsg = "shutdown subprocess";
3037                         break;
3038                 default:
3039                         statmsg = "??? subprocess";
3040                         break;
3041         }
3042         init_ps_display(statmsg, "", "");
3043         set_ps_display("");
3044 }
3045
3046
3047 static pid_t
3048 SSDataBase(int xlop)
3049 {
3050         pid_t           pid;
3051         Backend    *bn;
3052 #ifndef EXEC_BACKEND
3053 #ifdef LINUX_PROFILE
3054         struct itimerval prof_itimer;
3055 #endif
3056 #else
3057         char            idbuf[32];
3058 #endif
3059
3060         fflush(stdout);
3061         fflush(stderr);
3062
3063 #ifndef EXEC_BACKEND
3064 #ifdef LINUX_PROFILE
3065         /* see comments in BackendRun */
3066         getitimer(ITIMER_PROF, &prof_itimer);
3067 #endif
3068
3069 #ifdef __BEOS__
3070         /* Specific beos actions before backend startup */
3071         beos_before_backend_startup();
3072 #endif
3073
3074         /* Non EXEC_BACKEND case; fork here */
3075         if ((pid = fork()) == 0)        /* child */
3076 #endif
3077         {
3078                 char       *av[10];
3079                 int                     ac = 0;
3080                 char            nbbuf[32];
3081                 char            xlbuf[32];
3082
3083 #ifndef EXEC_BACKEND
3084 #ifdef LINUX_PROFILE
3085                 setitimer(ITIMER_PROF, &prof_itimer, NULL);
3086 #endif
3087
3088 #ifdef __BEOS__
3089                 /* Specific beos actions after backend startup */
3090                 beos_backend_startup();
3091 #endif
3092
3093                 /* Close the postmaster's sockets */
3094                 ClosePostmasterPorts(true);
3095
3096                 SSDataBaseInit(xlop);
3097 #else
3098                 if (!write_backend_variables(NULL))
3099                         return -1; /* log issued by write_backend_variables */
3100 #endif
3101
3102                 /* Set up command-line arguments for subprocess */
3103                 av[ac++] = "postgres";
3104
3105 #ifdef EXEC_BACKEND
3106                 av[ac++] = "-boot";
3107 #endif
3108                 snprintf(nbbuf, sizeof(nbbuf), "-B%d", NBuffers);
3109                 av[ac++] = nbbuf;
3110
3111                 snprintf(xlbuf, sizeof(xlbuf), "-x%d", xlop);
3112                 av[ac++] = xlbuf;
3113
3114 #ifdef EXEC_BACKEND
3115                 /* pass data dir before end of secure switches (-p) */
3116                 av[ac++] = "-D";
3117                 av[ac++] = DataDir;
3118
3119                 /* and the backend identifier + dbname */
3120                 snprintf(idbuf, sizeof(idbuf), "-p%lu,template1", tmpBackendFileNum);
3121                 av[ac++] = idbuf;
3122 #else
3123                 av[ac++] = "-p";
3124                 av[ac++] = "template1";
3125 #endif
3126
3127                 av[ac] = NULL;
3128
3129                 Assert(ac < lengthof(av));
3130
3131 #ifdef EXEC_BACKEND
3132                 /* EXEC_BACKEND case; fork/exec here */
3133 #ifdef WIN32
3134                 pid = win32_forkexec(pg_pathname,av); /* logs on error */
3135 #else
3136                 if ((pid = fork()) == 0 && (execv(pg_pathname,av) == -1))
3137                 {
3138                         /* in child */
3139                         elog(ERROR,"unable to execv in SSDataBase: %m");
3140                         exit(0);
3141                 }
3142 #endif
3143 #else
3144                 BootstrapMain(ac, av);
3145                 ExitPostmaster(0);
3146 #endif
3147         }
3148
3149         /* in parent */
3150         if (pid < 0)
3151         {
3152 #ifndef EXEC_BACKEND
3153 #ifdef __BEOS__
3154                 /* Specific beos actions before backend startup */
3155                 beos_backend_startup_failed();
3156 #endif
3157 #endif
3158                 switch (xlop)
3159                 {
3160                         case BS_XLOG_STARTUP:
3161                                 ereport(LOG,
3162                                                 (errmsg("could not fork startup process: %m")));
3163                                 break;
3164                         case BS_XLOG_CHECKPOINT:
3165                                 ereport(LOG,
3166                                           (errmsg("could not fork checkpoint process: %m")));
3167                                 break;
3168                         case BS_XLOG_BGWRITER:
3169                                 ereport(LOG,
3170                                           (errmsg("could not fork bgwriter process: %m")));
3171                                 break;
3172                         case BS_XLOG_SHUTDOWN:
3173                                 ereport(LOG,
3174                                                 (errmsg("could not fork shutdown process: %m")));
3175                                 break;
3176                         default:
3177                                 ereport(LOG,
3178                                                 (errmsg("could not fork process: %m")));
3179                                 break;
3180                 }
3181
3182                 /*
3183                  * fork failure is fatal during startup/shutdown, but there's no
3184                  * need to choke if a routine checkpoint or starting a background
3185                  * writer fails.
3186                  */
3187                 if (xlop == BS_XLOG_CHECKPOINT)
3188                         return 0;
3189                 if (xlop == BS_XLOG_BGWRITER)
3190                         return 0;
3191                 ExitPostmaster(1);
3192         }
3193
3194         /*
3195          * The startup and shutdown processes are not considered normal
3196          * backends, but the checkpoint and bgwriter processes are.
3197          * They must be added to the list of backends.
3198          */
3199         if (xlop == BS_XLOG_CHECKPOINT || xlop == BS_XLOG_BGWRITER)
3200         {
3201                 if (!(bn = (Backend *) malloc(sizeof(Backend))))
3202                 {
3203                         ereport(LOG,
3204                                         (errcode(ERRCODE_OUT_OF_MEMORY),
3205                                          errmsg("out of memory")));
3206                         ExitPostmaster(1);
3207                 }
3208
3209                 bn->pid = pid;
3210                 bn->cancel_key = PostmasterRandom();
3211 #ifdef EXEC_BACKEND
3212                 ShmemBackendArrayAdd(bn);
3213 #endif
3214                 DLAddHead(BackendList, DLNewElem(bn));
3215
3216                 /*
3217                  * Since this code is executed periodically, it's a fine place to
3218                  * do other actions that should happen every now and then on no
3219                  * particular schedule.  Such as...
3220                  */
3221                 TouchSocketFile();
3222                 TouchSocketLockFile();
3223         }
3224
3225         return pid;
3226 }
3227
3228
3229 /*
3230  * Create the opts file
3231  */
3232 static bool
3233 CreateOptsFile(int argc, char *argv[])
3234 {
3235         char            fullprogname[MAXPGPATH];
3236         char            filename[MAXPGPATH];
3237         FILE       *fp;
3238         int                     i;
3239
3240         if (FindExec(fullprogname, argv[0], "postmaster") < 0)
3241                 return false;
3242
3243         snprintf(filename, sizeof(filename), "%s/postmaster.opts", DataDir);
3244
3245         if ((fp = fopen(filename, "w")) == NULL)
3246         {
3247                 elog(LOG, "could not create file \"%s\": %m", filename);
3248                 return false;
3249         }
3250
3251         fprintf(fp, "%s", fullprogname);
3252         for (i = 1; i < argc; i++)
3253                 fprintf(fp, " '%s'", argv[i]);
3254         fputs("\n", fp);
3255
3256         if (fclose(fp))
3257         {
3258                 elog(LOG, "could not write file \"%s\": %m", filename);
3259                 return false;
3260         }
3261
3262         return true;
3263 }
3264
3265 /*
3266  * This should be used only for reporting "interactive" errors (essentially,
3267  * bogus arguments on the command line).  Once the postmaster is launched,
3268  * use ereport.  In particular, don't use this for anything that occurs
3269  * after pmdaemonize.
3270  */
3271 static void
3272 postmaster_error(const char *fmt,...)
3273 {
3274         va_list         ap;
3275
3276         fprintf(stderr, "%s: ", progname);
3277         va_start(ap, fmt);
3278         vfprintf(stderr, gettext(fmt), ap);
3279         va_end(ap);
3280         fprintf(stderr, "\n");
3281 }
3282
3283
3284 #ifdef EXEC_BACKEND
3285
3286 /*
3287  * The following need to be available to the read/write_backend_variables
3288  * functions
3289  */
3290 #include "storage/spin.h"
3291 extern XLogRecPtr RedoRecPtr;
3292 extern XLogwrtResult LogwrtResult;
3293 extern slock_t *ShmemLock;
3294 extern slock_t *ShmemIndexLock;
3295 extern void *ShmemIndexAlloc;
3296 typedef struct LWLock LWLock;
3297 extern LWLock *LWLockArray;
3298 extern slock_t  *ProcStructLock;
3299 extern int      pgStatSock;
3300
3301 #define write_var(var,fp) fwrite((void*)&(var),sizeof(var),1,fp)
3302 #define read_var(var,fp)  fread((void*)&(var),sizeof(var),1,fp)
3303 #define get_tmp_backend_file_name(buf,id)       \
3304                 do {                                                            \
3305                         Assert(DataDir);                                \
3306                         sprintf((buf),                                  \
3307                                 "%s/%s/%s.backend_var.%lu",     \
3308                                 DataDir,                                        \
3309                                 PG_TEMP_FILES_DIR,                      \
3310                                 PG_TEMP_FILE_PREFIX,            \
3311                                 (id));                                          \
3312                 } while (0)
3313
3314 static bool
3315 write_backend_variables(Port *port)
3316 {
3317         char    filename[MAXPGPATH];
3318         FILE    *fp;
3319         get_tmp_backend_file_name(filename,++tmpBackendFileNum);
3320
3321         /* Open file */
3322         fp = AllocateFile(filename, PG_BINARY_W);
3323         if (!fp)
3324         {
3325                 /* As per OpenTemporaryFile... */
3326                 char dirname[MAXPGPATH];
3327                 sprintf(dirname,"%s/%s",DataDir,PG_TEMP_FILES_DIR);
3328                 mkdir(dirname, S_IRWXU);
3329
3330                 fp = AllocateFile(filename, PG_BINARY_W);
3331                 if (!fp)
3332                 {
3333                         ereport(ERROR,
3334                                 (errcode_for_file_access(),
3335                                 errmsg("could not write to file \"%s\": %m", filename)));
3336                         return false;
3337                 }
3338         }
3339
3340         /* Write vars */
3341         if (port)
3342         {
3343                 write_var(port->sock,fp);
3344                 write_var(port->proto,fp);
3345                 write_var(port->laddr,fp);
3346                 write_var(port->raddr,fp);
3347                 write_var(port->canAcceptConnections,fp);
3348                 write_var(port->cryptSalt,fp);
3349                 write_var(port->md5Salt,fp);
3350         }
3351         write_var(MyCancelKey,fp);
3352
3353         write_var(RedoRecPtr,fp);
3354         write_var(LogwrtResult,fp);
3355
3356         write_var(UsedShmemSegID,fp);
3357         write_var(UsedShmemSegAddr,fp);
3358
3359         write_var(ShmemLock,fp);
3360         write_var(ShmemIndexLock,fp);
3361         write_var(ShmemVariableCache,fp);
3362         write_var(ShmemIndexAlloc,fp);
3363         write_var(ShmemBackendArray,fp);
3364
3365         write_var(LWLockArray,fp);
3366         write_var(ProcStructLock,fp);
3367         write_var(pgStatSock,fp);
3368
3369         write_var(PreAuthDelay,fp);
3370         write_var(debug_flag,fp);
3371         write_var(PostmasterPid,fp);
3372
3373         /* Release file */
3374         if (FreeFile(fp))
3375         {
3376                 ereport(ERROR,
3377                                 (errcode_for_file_access(),
3378                                  errmsg("could not write to file \"%s\": %m", filename)));
3379                 return false;
3380         }
3381
3382         return true;
3383 }
3384
3385 void
3386 read_backend_variables(unsigned long id, Port *port)
3387 {
3388         char    filename[MAXPGPATH];
3389         FILE    *fp;
3390         get_tmp_backend_file_name(filename,id);
3391
3392         /* Open file */
3393         fp = AllocateFile(filename, PG_BINARY_R);
3394         if (!fp)
3395         {
3396                 ereport(ERROR,
3397                         (errcode_for_file_access(),
3398                         errmsg("could not read from backend_variables file \"%s\": %m", filename)));
3399                 return;
3400         }
3401
3402         /* Read vars */
3403         if (port)
3404         {
3405                 read_var(port->sock,fp);
3406                 read_var(port->proto,fp);
3407                 read_var(port->laddr,fp);
3408                 read_var(port->raddr,fp);
3409                 read_var(port->canAcceptConnections,fp);
3410                 read_var(port->cryptSalt,fp);
3411                 read_var(port->md5Salt,fp);
3412         }
3413         read_var(MyCancelKey,fp);
3414
3415         read_var(RedoRecPtr,fp);
3416         read_var(LogwrtResult,fp);
3417
3418         read_var(UsedShmemSegID,fp);
3419         read_var(UsedShmemSegAddr,fp);
3420
3421         read_var(ShmemLock,fp);
3422         read_var(ShmemIndexLock,fp);
3423         read_var(ShmemVariableCache,fp);
3424         read_var(ShmemIndexAlloc,fp);
3425         read_var(ShmemBackendArray,fp);
3426
3427         read_var(LWLockArray,fp);
3428         read_var(ProcStructLock,fp);
3429         read_var(pgStatSock,fp);
3430
3431         read_var(PreAuthDelay,fp);
3432         read_var(debug_flag,fp);
3433         read_var(PostmasterPid,fp);
3434
3435         /* Release file */
3436         FreeFile(fp);
3437         if (unlink(filename) != 0)
3438                 ereport(WARNING,
3439                                 (errcode_for_file_access(),
3440                                  errmsg("could not remove file \"%s\": %m", filename)));
3441 }
3442
3443
3444 size_t ShmemBackendArraySize(void)
3445 {
3446         return (NUM_BACKENDARRAY_ELEMS*sizeof(Backend));
3447 }
3448
3449 void ShmemBackendArrayAllocation(void)
3450 {
3451         size_t size = ShmemBackendArraySize();
3452         ShmemBackendArray = (Backend*)ShmemAlloc(size);
3453         memset(ShmemBackendArray, 0, size);
3454 }
3455
3456 static void ShmemBackendArrayAdd(Backend *bn)
3457 {
3458         int i;
3459         for (i = 0; i < NUM_BACKENDARRAY_ELEMS; i++)
3460         {
3461                 /* Find an empty slot */
3462                 if (ShmemBackendArray[i].pid == 0)
3463                 {
3464                         ShmemBackendArray[i] = *bn;
3465                         return;
3466                 }
3467         }
3468
3469         /* FIXME: [fork/exec] some sort of error */
3470         abort();
3471 }
3472
3473 static void ShmemBackendArrayRemove(pid_t pid)
3474 {
3475         int i;
3476         for (i = 0; i < NUM_BACKENDARRAY_ELEMS; i++)
3477         {
3478                 if (ShmemBackendArray[i].pid == pid)
3479                 {
3480                         /* Mark the slot as empty */
3481                         ShmemBackendArray[i].pid = 0;
3482                         return;
3483                 }
3484         }
3485
3486         /* Something stronger than WARNING here? */
3487         ereport(WARNING,
3488                         (errmsg_internal("unable to find backend entry with pid %d",
3489                                                          pid)));
3490 }
3491
3492 #endif
3493
3494 #ifdef WIN32
3495
3496 pid_t win32_forkexec(const char* path, char *argv[])
3497 {
3498         STARTUPINFO si;
3499         PROCESS_INFORMATION pi;
3500         char *p;
3501         int i;
3502         char cmdLine[MAXPGPATH];
3503         HANDLE childHandleCopy;
3504         HANDLE waiterThread;
3505
3506         /* Format the cmd line */
3507         snprintf(cmdLine,sizeof(cmdLine),"%s",path);
3508         i = 0;
3509         while (argv[++i] != NULL)
3510         {
3511                 /* FIXME: [fork/exec] some strlen checks might be prudent here */
3512                 strcat(cmdLine," ");
3513                 strcat(cmdLine,argv[i]);
3514         }
3515
3516         /*
3517          * The following snippet can disappear when we consistently
3518          * use forward slashes.
3519          */
3520         p = cmdLine;
3521         while (*(p++) != '\0')
3522                 if (*p == '/') *p = '\\';
3523
3524         memset(&pi,0,sizeof(pi));
3525         memset(&si,0,sizeof(si));
3526         si.cb = sizeof(si);
3527         if (!CreateProcess(NULL,cmdLine,NULL,NULL,TRUE,0,NULL,NULL,&si,&pi))
3528         {
3529                 elog(ERROR,"CreateProcess call failed (%i): %m",(int)GetLastError());
3530                 return -1;
3531         }
3532
3533         if (!IsUnderPostmaster)
3534                 /* We are the Postmaster creating a child... */
3535                 win32_AddChild(pi.dwProcessId,pi.hProcess);
3536
3537         if (!DuplicateHandle(GetCurrentProcess(),
3538                                                  pi.hProcess,
3539                                                  GetCurrentProcess(),
3540                                                  &childHandleCopy,
3541                                                  0,
3542                                                  FALSE,
3543                                                  DUPLICATE_SAME_ACCESS))
3544                 ereport(FATAL,
3545                                 (errmsg_internal("failed to duplicate child handle: %i",(int)GetLastError())));
3546         waiterThread = CreateThread(NULL, 64*1024, win32_sigchld_waiter, (LPVOID)childHandleCopy, 0, NULL);
3547         if (!waiterThread)
3548                 ereport(FATAL,
3549                                 (errmsg_internal("failed to create sigchld waiter thread: %i",(int)GetLastError())));
3550         CloseHandle(waiterThread);
3551
3552         if (IsUnderPostmaster)
3553                 CloseHandle(pi.hProcess);
3554         CloseHandle(pi.hThread);
3555
3556         return pi.dwProcessId;
3557 }
3558
3559 /*
3560  * Note: The following three functions must not be interrupted (eg. by signals).
3561  *  As the Postgres Win32 signalling architecture (currently) requires polling,
3562  *  or APC checking functions which aren't used here, this is not an issue.
3563  *
3564  *  We keep two separate arrays, instead of a single array of pid/HANDLE structs,
3565  *  to avoid having to re-create a handle array for WaitForMultipleObjects on
3566  *  each call to win32_waitpid.
3567  */
3568
3569 static void win32_AddChild(pid_t pid, HANDLE handle)
3570 {
3571         Assert(win32_childPIDArray && win32_childHNDArray);
3572         if (win32_numChildren < NUM_BACKENDARRAY_ELEMS)
3573         {
3574                 win32_childPIDArray[win32_numChildren] = pid;
3575                 win32_childHNDArray[win32_numChildren] = handle;
3576                 ++win32_numChildren;
3577         }
3578         else
3579                 /* FIXME: [fork/exec] some sort of error */
3580                 abort();
3581 }
3582
3583 static void win32_RemoveChild(pid_t pid)
3584 {
3585         int i;
3586         Assert(win32_childPIDArray && win32_childHNDArray);
3587
3588         for (i = 0; i < win32_numChildren; i++)
3589         {
3590                 if (win32_childPIDArray[i] == pid)
3591                 {
3592                         CloseHandle(win32_childHNDArray[i]);
3593
3594                         /* Swap last entry into the "removed" one */
3595                         --win32_numChildren;
3596                         win32_childPIDArray[i] = win32_childPIDArray[win32_numChildren];
3597                         win32_childHNDArray[i] = win32_childHNDArray[win32_numChildren];
3598                         return;
3599                 }
3600         }
3601
3602         /* Something stronger than WARNING here? */
3603         ereport(WARNING,
3604                         (errmsg_internal("unable to find child entry with pid %lu",
3605                                                          pid)));
3606 }
3607
3608 static pid_t win32_waitpid(int *exitstatus)
3609 {
3610         Assert(win32_childPIDArray && win32_childHNDArray);
3611         elog(DEBUG3,"waiting on %lu children",win32_numChildren);
3612
3613         if (win32_numChildren > 0)
3614         {
3615                 /*
3616                  * Note: Do NOT use WaitForMultipleObjectsEx, as we don't
3617                  * want to run queued APCs here.
3618                  */
3619                 int index;
3620                 DWORD exitCode;
3621                 DWORD ret = WaitForMultipleObjects(win32_numChildren,win32_childHNDArray,FALSE,0);
3622
3623                 switch (ret)
3624                 {
3625                         case WAIT_FAILED:
3626                                 ereport(ERROR,
3627                                                 (errmsg_internal("failed to wait on %lu children: %i",
3628                                                                                  win32_numChildren,(int)GetLastError())));
3629                                 /* Fall through to WAIT_TIMEOUTs return */
3630
3631                         case WAIT_TIMEOUT:
3632                                 /* No children have finished */
3633                                 return -1;
3634
3635                         default:
3636                                 /* Get the exit code, and return the PID of, the respective process */
3637                                 index = ret-WAIT_OBJECT_0;
3638                                 Assert(index >= 0 && index < win32_numChildren);
3639                                 if (!GetExitCodeProcess(win32_childHNDArray[index],&exitCode))
3640                                         /*
3641                                          * If we get this far, this should never happen, but, then again...
3642                                          * No choice other than to assume a catastrophic failure.
3643                                          */
3644                                         ereport(FATAL,
3645                                                         (errmsg_internal("failed to get exit code for child %lu",
3646                                                                                          win32_childPIDArray[index])));
3647                                 *exitstatus = (int)exitCode;
3648                                 return win32_childPIDArray[index];
3649                 }
3650         }
3651
3652         /* No children */
3653         return -1;
3654 }
3655
3656 /* Note! Code belows executes on separate threads, one for
3657    each child process created */
3658 static DWORD WINAPI win32_sigchld_waiter(LPVOID param) {
3659         HANDLE procHandle = (HANDLE)param;
3660
3661         DWORD r = WaitForSingleObject(procHandle, INFINITE);
3662         if (r == WAIT_OBJECT_0)
3663                 pg_queue_signal(SIGCHLD);
3664         else
3665                 fprintf(stderr,"ERROR: Failed to wait on child process handle: %i\n",(int)GetLastError());
3666         CloseHandle(procHandle);
3667         return 0;
3668 }
3669
3670 #endif