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 and shutdown. The postmaster itself doesn't do those
  11  *        operations, mind you --- it just forks off a subprocess to do them
  12  *        at the right times.  It also takes care of resetting the system
  13  *        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) 1996-2017, PostgreSQL Global Development Group
  36  * Portions Copyright (c) 1994, Regents of the University of California
  37  *
  38  *
  39  * IDENTIFICATION
  40  *        src/backend/postmaster/postmaster.c
  41  *
  42  * NOTES
  43  *
  44  * Initialization:
  45  *              The Postmaster sets up shared memory data structures
  46  *              for the backends.
  47  *
  48  * Synchronization:
  49  *              The Postmaster shares memory with the backends but should avoid
  50  *              touching shared memory, so as not to become stuck if a crashing
  51  *              backend screws up locks or shared memory.  Likewise, the Postmaster
  52  *              should never block on messages from frontend clients.
  53  *
  54  * Garbage Collection:
  55  *              The Postmaster cleans up after backends if they have an emergency
  56  *              exit and/or core dump.
  57  *
  58  * Error Reporting:
  59  *              Use write_stderr() only for reporting "interactive" errors
  60  *              (essentially, bogus arguments on the command line).  Once the
  61  *              postmaster is launched, use ereport().
  62  *
  63  *-------------------------------------------------------------------------
  64  */
  65
  66 #include "postgres.h"
  67
  68 #include <unistd.h>
  69 #include <signal.h>
  70 #include <time.h>
  71 #include <sys/wait.h>
  72 #include <ctype.h>
  73 #include <sys/stat.h>
  74 #include <sys/socket.h>
  75 #include <fcntl.h>
  76 #include <sys/param.h>
  77 #include <netinet/in.h>
  78 #include <arpa/inet.h>
  79 #include <netdb.h>
  80 #include <limits.h>
  81
  82 #ifdef HAVE_SYS_SELECT_H
  83 #include <sys/select.h>
  84 #endif
  85
  86 #ifdef USE_BONJOUR
  87 #include <dns_sd.h>
  88 #endif
  89
  90 #ifdef USE_SYSTEMD
  91 #include <systemd/sd-daemon.h>
  92 #endif
  93
  94 #ifdef HAVE_PTHREAD_IS_THREADED_NP
  95 #include <pthread.h>
  96 #endif
  97
  98 #include "access/transam.h"
  99 #include "access/xlog.h"
 100 #include "bootstrap/bootstrap.h"
 101 #include "catalog/pg_control.h"
 102 #include "common/ip.h"
 103 #include "lib/ilist.h"
 104 #include "libpq/auth.h"
 105 #include "libpq/libpq.h"
 106 #include "libpq/pqsignal.h"
 107 #include "miscadmin.h"
 108 #include "pg_getopt.h"
 109 #include "pgstat.h"
 110 #include "postmaster/autovacuum.h"
 111 #include "postmaster/bgworker_internals.h"
 112 #include "postmaster/fork_process.h"
 113 #include "postmaster/pgarch.h"
 114 #include "postmaster/postmaster.h"
 115 #include "postmaster/syslogger.h"
 116 #include "replication/logicallauncher.h"
 117 #include "replication/walsender.h"
 118 #include "storage/fd.h"
 119 #include "storage/ipc.h"
 120 #include "storage/pg_shmem.h"
 121 #include "storage/pmsignal.h"
 122 #include "storage/proc.h"
 123 #include "tcop/tcopprot.h"
 124 #include "utils/builtins.h"
 125 #include "utils/datetime.h"
 126 #include "utils/dynamic_loader.h"
 127 #include "utils/memutils.h"
 128 #include "utils/pidfile.h"
 129 #include "utils/ps_status.h"
 130 #include "utils/timeout.h"
 131 #include "utils/varlena.h"
 132
 133 #ifdef EXEC_BACKEND
 134 #include "storage/spin.h"
 135 #endif
 136
 137
 138 /*
 139  * Possible types of a backend. Beyond being the possible bkend_type values in
 140  * struct bkend, these are OR-able request flag bits for SignalSomeChildren()
 141  * and CountChildren().
 142  */
 143 #define BACKEND_TYPE_NORMAL             0x0001  /* normal backend */
 144 #define BACKEND_TYPE_AUTOVAC    0x0002  /* autovacuum worker process */
 145 #define BACKEND_TYPE_WALSND             0x0004  /* walsender process */
 146 #define BACKEND_TYPE_BGWORKER   0x0008  /* bgworker process */
 147 #define BACKEND_TYPE_ALL                0x000F  /* OR of all the above */
 148
 149 #define BACKEND_TYPE_WORKER             (BACKEND_TYPE_AUTOVAC | BACKEND_TYPE_BGWORKER)
 150
 151 /*
 152  * List of active backends (or child processes anyway; we don't actually
 153  * know whether a given child has become a backend or is still in the
 154  * authorization phase).  This is used mainly to keep track of how many
 155  * children we have and send them appropriate signals when necessary.
 156  *
 157  * "Special" children such as the startup, bgwriter and autovacuum launcher
 158  * tasks are not in this list.  Autovacuum worker and walsender are in it.
 159  * Also, "dead_end" children are in it: these are children launched just for
 160  * the purpose of sending a friendly rejection message to a would-be client.
 161  * We must track them because they are attached to shared memory, but we know
 162  * they will never become live backends.  dead_end children are not assigned a
 163  * PMChildSlot.
 164  *
 165  * Background workers are in this list, too.
 166  */
 167 typedef struct bkend
 168 {
 169         pid_t           pid;                    /* process id of backend */
 170         int32           cancel_key;             /* cancel key for cancels for this backend */
 171         int                     child_slot;             /* PMChildSlot for this backend, if any */
 172
 173         /*
 174          * Flavor of backend or auxiliary process.  Note that BACKEND_TYPE_WALSND
 175          * backends initially announce themselves as BACKEND_TYPE_NORMAL, so if
 176          * bkend_type is normal, you should check for a recent transition.
 177          */
 178         int                     bkend_type;
 179         bool            dead_end;               /* is it going to send an error and quit? */
 180         bool            bgworker_notify;        /* gets bgworker start/stop notifications */
 181         dlist_node      elem;                   /* list link in BackendList */
 182 } Backend;
 183
 184 static dlist_head BackendList = DLIST_STATIC_INIT(BackendList);
 185
 186 #ifdef EXEC_BACKEND
 187 static Backend *ShmemBackendArray;
 188 #endif
 189
 190 BackgroundWorker *MyBgworkerEntry = NULL;
 191
 192
 193
 194 /* The socket number we are listening for connections on */
 195 int                     PostPortNumber;
 196
 197 /* The directory names for Unix socket(s) */
 198 char       *Unix_socket_directories;
 199
 200 /* The TCP listen address(es) */
 201 char       *ListenAddresses;
 202
 203 /*
 204  * ReservedBackends is the number of backends reserved for superuser use.
 205  * This number is taken out of the pool size given by MaxBackends so
 206  * number of backend slots available to non-superusers is
 207  * (MaxBackends - ReservedBackends).  Note what this really means is
 208  * "if there are <= ReservedBackends connections available, only superusers
 209  * can make new connections" --- pre-existing superuser connections don't
 210  * count against the limit.
 211  */
 212 int                     ReservedBackends;
 213
 214 /* The socket(s) we're listening to. */
 215 #define MAXLISTEN       64
 216 static pgsocket ListenSocket[MAXLISTEN];
 217
 218 /*
 219  * Set by the -o option
 220  */
 221 static char ExtraOptions[MAXPGPATH];
 222
 223 /*
 224  * These globals control the behavior of the postmaster in case some
 225  * backend dumps core.  Normally, it kills all peers of the dead backend
 226  * and reinitializes shared memory.  By specifying -s or -n, we can have
 227  * the postmaster stop (rather than kill) peers and not reinitialize
 228  * shared data structures.  (Reinit is currently dead code, though.)
 229  */
 230 static bool Reinit = true;
 231 static int      SendStop = false;
 232
 233 /* still more option variables */
 234 bool            EnableSSL = false;
 235
 236 int                     PreAuthDelay = 0;
 237 int                     AuthenticationTimeout = 60;
 238
 239 bool            log_hostname;           /* for ps display and logging */
 240 bool            Log_connections = false;
 241 bool            Db_user_namespace = false;
 242
 243 bool            enable_bonjour = false;
 244 char       *bonjour_name;
 245 bool            restart_after_crash = true;
 246
 247 /* PIDs of special child processes; 0 when not running */
 248 static pid_t StartupPID = 0,
 249                         BgWriterPID = 0,
 250                         CheckpointerPID = 0,
 251                         WalWriterPID = 0,
 252                         WalReceiverPID = 0,
 253                         AutoVacPID = 0,
 254                         PgArchPID = 0,
 255                         PgStatPID = 0,
 256                         SysLoggerPID = 0;
 257
 258 /* Startup process's status */
 259 typedef enum
 260 {
 261         STARTUP_NOT_RUNNING,
 262         STARTUP_RUNNING,
 263         STARTUP_SIGNALED,                       /* we sent it a SIGQUIT or SIGKILL */
 264         STARTUP_CRASHED
 265 } StartupStatusEnum;
 266
 267 static StartupStatusEnum StartupStatus = STARTUP_NOT_RUNNING;
 268
 269 /* Startup/shutdown state */
 270 #define                 NoShutdown              0
 271 #define                 SmartShutdown   1
 272 #define                 FastShutdown    2
 273 #define                 ImmediateShutdown       3
 274
 275 static int      Shutdown = NoShutdown;
 276
 277 static bool FatalError = false; /* T if recovering from backend crash */
 278
 279 /*
 280  * We use a simple state machine to control startup, shutdown, and
 281  * crash recovery (which is rather like shutdown followed by startup).
 282  *
 283  * After doing all the postmaster initialization work, we enter PM_STARTUP
 284  * state and the startup process is launched. The startup process begins by
 285  * reading the control file and other preliminary initialization steps.
 286  * In a normal startup, or after crash recovery, the startup process exits
 287  * with exit code 0 and we switch to PM_RUN state.  However, archive recovery
 288  * is handled specially since it takes much longer and we would like to support
 289  * hot standby during archive recovery.
 290  *
 291  * When the startup process is ready to start archive recovery, it signals the
 292  * postmaster, and we switch to PM_RECOVERY state. The background writer and
 293  * checkpointer are launched, while the startup process continues applying WAL.
 294  * If Hot Standby is enabled, then, after reaching a consistent point in WAL
 295  * redo, startup process signals us again, and we switch to PM_HOT_STANDBY
 296  * state and begin accepting connections to perform read-only queries.  When
 297  * archive recovery is finished, the startup process exits with exit code 0
 298  * and we switch to PM_RUN state.
 299  *
 300  * Normal child backends can only be launched when we are in PM_RUN or
 301  * PM_HOT_STANDBY state.  (We also allow launch of normal
 302  * child backends in PM_WAIT_BACKUP state, but only for superusers.)
 303  * In other states we handle connection requests by launching "dead_end"
 304  * child processes, which will simply send the client an error message and
 305  * quit.  (We track these in the BackendList so that we can know when they
 306  * are all gone; this is important because they're still connected to shared
 307  * memory, and would interfere with an attempt to destroy the shmem segment,
 308  * possibly leading to SHMALL failure when we try to make a new one.)
 309  * In PM_WAIT_DEAD_END state we are waiting for all the dead_end children
 310  * to drain out of the system, and therefore stop accepting connection
 311  * requests at all until the last existing child has quit (which hopefully
 312  * will not be very long).
 313  *
 314  * Notice that this state variable does not distinguish *why* we entered
 315  * states later than PM_RUN --- Shutdown and FatalError must be consulted
 316  * to find that out.  FatalError is never true in PM_RECOVERY_* or PM_RUN
 317  * states, nor in PM_SHUTDOWN states (because we don't enter those states
 318  * when trying to recover from a crash).  It can be true in PM_STARTUP state,
 319  * because we don't clear it until we've successfully started WAL redo.
 320  */
 321 typedef enum
 322 {
 323         PM_INIT,                                        /* postmaster starting */
 324         PM_STARTUP,                                     /* waiting for startup subprocess */
 325         PM_RECOVERY,                            /* in archive recovery mode */
 326         PM_HOT_STANDBY,                         /* in hot standby mode */
 327         PM_RUN,                                         /* normal "database is alive" state */
 328         PM_WAIT_BACKUP,                         /* waiting for online backup mode to end */
 329         PM_WAIT_READONLY,                       /* waiting for read only backends to exit */
 330         PM_WAIT_BACKENDS,                       /* waiting for live backends to exit */
 331         PM_SHUTDOWN,                            /* waiting for checkpointer to do shutdown
 332                                                                  * ckpt */
 333         PM_SHUTDOWN_2,                          /* waiting for archiver and walsenders to
 334                                                                  * finish */
 335         PM_WAIT_DEAD_END,                       /* waiting for dead_end children to exit */
 336         PM_NO_CHILDREN                          /* all important children have exited */
 337 } PMState;
 338
 339 static PMState pmState = PM_INIT;
 340
 341 /* Start time of SIGKILL timeout during immediate shutdown or child crash */
 342 /* Zero means timeout is not running */
 343 static time_t AbortStartTime = 0;
 344
 345 /* Length of said timeout */
 346 #define SIGKILL_CHILDREN_AFTER_SECS             5
 347
 348 static bool ReachedNormalRunning = false;       /* T if we've reached PM_RUN */
 349
 350 bool            ClientAuthInProgress = false;   /* T during new-client
 351                                                                                          * authentication */
 352
 353 bool            redirection_done = false;       /* stderr redirected for syslogger? */
 354
 355 /* received START_AUTOVAC_LAUNCHER signal */
 356 static volatile sig_atomic_t start_autovac_launcher = false;
 357
 358 /* the launcher needs to be signalled to communicate some condition */
 359 static volatile bool avlauncher_needs_signal = false;
 360
 361 /* received START_WALRECEIVER signal */
 362 static volatile sig_atomic_t WalReceiverRequested = false;
 363
 364 /* set when there's a worker that needs to be started up */
 365 static volatile bool StartWorkerNeeded = true;
 366 static volatile bool HaveCrashedWorker = false;
 367
 368 #ifndef HAVE_STRONG_RANDOM
 369 /*
 370  * State for assigning cancel keys.
 371  * Also, the global MyCancelKey passes the cancel key assigned to a given
 372  * backend from the postmaster to that backend (via fork).
 373  */
 374 static unsigned int random_seed = 0;
 375 static struct timeval random_start_time;
 376 #endif
 377
 378 #ifdef USE_SSL
 379 /* Set when and if SSL has been initialized properly */
 380 static bool LoadedSSL = false;
 381 #endif
 382
 383 #ifdef USE_BONJOUR
 384 static DNSServiceRef bonjour_sdref = NULL;
 385 #endif
 386
 387 /*
 388  * postmaster.c - function prototypes
 389  */
 390 static void CloseServerPorts(int status, Datum arg);
 391 static void unlink_external_pid_file(int status, Datum arg);
 392 static void getInstallationPaths(const char *argv0);
 393 static void checkDataDir(void);
 394 static Port *ConnCreate(int serverFd);
 395 static void ConnFree(Port *port);
 396 static void reset_shared(int port);
 397 static void SIGHUP_handler(SIGNAL_ARGS);
 398 static void pmdie(SIGNAL_ARGS);
 399 static void reaper(SIGNAL_ARGS);
 400 static void sigusr1_handler(SIGNAL_ARGS);
 401 static void startup_die(SIGNAL_ARGS);
 402 static void dummy_handler(SIGNAL_ARGS);
 403 static void StartupPacketTimeoutHandler(void);
 404 static void CleanupBackend(int pid, int exitstatus);
 405 static bool CleanupBackgroundWorker(int pid, int exitstatus);
 406 static void HandleChildCrash(int pid, int exitstatus, const char *procname);
 407 static void LogChildExit(int lev, const char *procname,
 408                          int pid, int exitstatus);
 409 static void PostmasterStateMachine(void);
 410 static void BackendInitialize(Port *port);
 411 static void BackendRun(Port *port) pg_attribute_noreturn();
 412 static void ExitPostmaster(int status) pg_attribute_noreturn();
 413 static int      ServerLoop(void);
 414 static int      BackendStartup(Port *port);
 415 static int      ProcessStartupPacket(Port *port, bool SSLdone);
 416 static void processCancelRequest(Port *port, void *pkt);
 417 static int      initMasks(fd_set *rmask);
 418 static void report_fork_failure_to_client(Port *port, int errnum);
 419 static CAC_state canAcceptConnections(void);
 420 static bool RandomCancelKey(int32 *cancel_key);
 421 static void signal_child(pid_t pid, int signal);
 422 static bool SignalSomeChildren(int signal, int targets);
 423 static void TerminateChildren(int signal);
 424
 425 #define SignalChildren(sig)                        SignalSomeChildren(sig, BACKEND_TYPE_ALL)
 426
 427 static int      CountChildren(int target);
 428 static bool assign_backendlist_entry(RegisteredBgWorker *rw);
 429 static void maybe_start_bgworkers(void);
 430 static bool CreateOptsFile(int argc, char *argv[], char *fullprogname);
 431 static pid_t StartChildProcess(AuxProcType type);
 432 static void StartAutovacuumWorker(void);
 433 static void MaybeStartWalReceiver(void);
 434 static void InitPostmasterDeathWatchHandle(void);
 435
 436 /*
 437  * Archiver is allowed to start up at the current postmaster state?
 438  *
 439  * If WAL archiving is enabled always, we are allowed to start archiver
 440  * even during recovery.
 441  */
 442 #define PgArchStartupAllowed()  \
 443         ((XLogArchivingActive() && pmState == PM_RUN) ||        \
 444          (XLogArchivingAlways() &&      \
 445           (pmState == PM_RECOVERY || pmState == PM_HOT_STANDBY)))
 446
 447 #ifdef EXEC_BACKEND
 448
 449 #ifdef WIN32
 450 #define WNOHANG 0                               /* ignored, so any integer value will do */
 451
 452 static pid_t waitpid(pid_t pid, int *exitstatus, int options);
 453 static void WINAPI pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired);
 454
 455 static HANDLE win32ChildQueue;
 456
 457 typedef struct
 458 {
 459         HANDLE          waitHandle;
 460         HANDLE          procHandle;
 461         DWORD           procId;
 462 } win32_deadchild_waitinfo;
 463 #endif                                                  /* WIN32 */
 464
 465 static pid_t backend_forkexec(Port *port);
 466 static pid_t internal_forkexec(int argc, char *argv[], Port *port);
 467
 468 /* Type for a socket that can be inherited to a client process */
 469 #ifdef WIN32
 470 typedef struct
 471 {
 472         SOCKET          origsocket;             /* Original socket value, or PGINVALID_SOCKET
 473                                                                  * if not a socket */
 474         WSAPROTOCOL_INFO wsainfo;
 475 } InheritableSocket;
 476 #else
 477 typedef int InheritableSocket;
 478 #endif
 479
 480 /*
 481  * Structure contains all variables passed to exec:ed backends
 482  */
 483 typedef struct
 484 {
 485         Port            port;
 486         InheritableSocket portsocket;
 487         char            DataDir[MAXPGPATH];
 488         pgsocket        ListenSocket[MAXLISTEN];
 489         int32           MyCancelKey;
 490         int                     MyPMChildSlot;
 491 #ifndef WIN32
 492         unsigned long UsedShmemSegID;
 493 #else
 494         HANDLE          UsedShmemSegID;
 495 #endif
 496         void       *UsedShmemSegAddr;
 497         slock_t    *ShmemLock;
 498         VariableCache ShmemVariableCache;
 499         Backend    *ShmemBackendArray;
 500 #ifndef HAVE_SPINLOCKS
 501         PGSemaphore *SpinlockSemaArray;
 502 #endif
 503         int                     NamedLWLockTrancheRequests;
 504         NamedLWLockTranche *NamedLWLockTrancheArray;
 505         LWLockPadded *MainLWLockArray;
 506         slock_t    *ProcStructLock;
 507         PROC_HDR   *ProcGlobal;
 508         PGPROC     *AuxiliaryProcs;
 509         PGPROC     *PreparedXactProcs;
 510         PMSignalData *PMSignalState;
 511         InheritableSocket pgStatSock;
 512         pid_t           PostmasterPid;
 513         TimestampTz PgStartTime;
 514         TimestampTz PgReloadTime;
 515         pg_time_t       first_syslogger_file_time;
 516         bool            redirection_done;
 517         bool            IsBinaryUpgrade;
 518         int                     max_safe_fds;
 519         int                     MaxBackends;
 520 #ifdef WIN32
 521         HANDLE          PostmasterHandle;
 522         HANDLE          initial_signal_pipe;
 523         HANDLE          syslogPipe[2];
 524 #else
 525         int                     postmaster_alive_fds[2];
 526         int                     syslogPipe[2];
 527 #endif
 528         char            my_exec_path[MAXPGPATH];
 529         char            pkglib_path[MAXPGPATH];
 530         char            ExtraOptions[MAXPGPATH];
 531 } BackendParameters;
 532
 533 static void read_backend_variables(char *id, Port *port);
 534 static void restore_backend_variables(BackendParameters *param, Port *port);
 535
 536 #ifndef WIN32
 537 static bool save_backend_variables(BackendParameters *param, Port *port);
 538 #else
 539 static bool save_backend_variables(BackendParameters *param, Port *port,
 540                                            HANDLE childProcess, pid_t childPid);
 541 #endif
 542
 543 static void ShmemBackendArrayAdd(Backend *bn);
 544 static void ShmemBackendArrayRemove(Backend *bn);
 545 #endif                                                  /* EXEC_BACKEND */
 546
 547 #define StartupDataBase()               StartChildProcess(StartupProcess)
 548 #define StartBackgroundWriter() StartChildProcess(BgWriterProcess)
 549 #define StartCheckpointer()             StartChildProcess(CheckpointerProcess)
 550 #define StartWalWriter()                StartChildProcess(WalWriterProcess)
 551 #define StartWalReceiver()              StartChildProcess(WalReceiverProcess)
 552
 553 /* Macros to check exit status of a child process */
 554 #define EXIT_STATUS_0(st)  ((st) == 0)
 555 #define EXIT_STATUS_1(st)  (WIFEXITED(st) && WEXITSTATUS(st) == 1)
 556 #define EXIT_STATUS_3(st)  (WIFEXITED(st) && WEXITSTATUS(st) == 3)
 557
 558 #ifndef WIN32
 559 /*
 560  * File descriptors for pipe used to monitor if postmaster is alive.
 561  * First is POSTMASTER_FD_WATCH, second is POSTMASTER_FD_OWN.
 562  */
 563 int                     postmaster_alive_fds[2] = {-1, -1};
 564 #else
 565 /* Process handle of postmaster used for the same purpose on Windows */
 566 HANDLE          PostmasterHandle;
 567 #endif
 568
 569 /*
 570  * Postmaster main entry point
 571  */
 572 void
 573 PostmasterMain(int argc, char *argv[])
 574 {
 575         int                     opt;
 576         int                     status;
 577         char       *userDoption = NULL;
 578         bool            listen_addr_saved = false;
 579         int                     i;
 580         char       *output_config_variable = NULL;
 581
 582         MyProcPid = PostmasterPid = getpid();
 583
 584         MyStartTime = time(NULL);
 585
 586         IsPostmasterEnvironment = true;
 587
 588         /*
 589          * for security, no dir or file created can be group or other accessible
 590          */
 591         umask(S_IRWXG | S_IRWXO);
 592
 593         /*
 594          * Initialize random(3) so we don't get the same values in every run.
 595          *
 596          * Note: the seed is pretty predictable from externally-visible facts such
 597          * as postmaster start time, so avoid using random() for security-critical
 598          * random values during postmaster startup.  At the time of first
 599          * connection, PostmasterRandom will select a hopefully-more-random seed.
 600          */
 601         srandom((unsigned int) (MyProcPid ^ MyStartTime));
 602
 603         /*
 604          * By default, palloc() requests in the postmaster will be allocated in
 605          * the PostmasterContext, which is space that can be recycled by backends.
 606          * Allocated data that needs to be available to backends should be
 607          * allocated in TopMemoryContext.
 608          */
 609         PostmasterContext = AllocSetContextCreate(TopMemoryContext,
 610                                                                                           "Postmaster",
 611                                                                                           ALLOCSET_DEFAULT_SIZES);
 612         MemoryContextSwitchTo(PostmasterContext);
 613
 614         /* Initialize paths to installation files */
 615         getInstallationPaths(argv[0]);
 616
 617         /*
 618          * Set up signal handlers for the postmaster process.
 619          *
 620          * In the postmaster, we want to install non-ignored handlers *without*
 621          * SA_RESTART.  This is because they'll be blocked at all times except
 622          * when ServerLoop is waiting for something to happen, and during that
 623          * window, we want signals to exit the select(2) wait so that ServerLoop
 624          * can respond if anything interesting happened.  On some platforms,
 625          * signals marked SA_RESTART would not cause the select() wait to end.
 626          * Child processes will generally want SA_RESTART, but we expect them to
 627          * set up their own handlers before unblocking signals.
 628          *
 629          * CAUTION: when changing this list, check for side-effects on the signal
 630          * handling setup of child processes.  See tcop/postgres.c,
 631          * bootstrap/bootstrap.c, postmaster/bgwriter.c, postmaster/walwriter.c,
 632          * postmaster/autovacuum.c, postmaster/pgarch.c, postmaster/pgstat.c,
 633          * postmaster/syslogger.c, postmaster/bgworker.c and
 634          * postmaster/checkpointer.c.
 635          */
 636         pqinitmask();
 637         PG_SETMASK(&BlockSig);
 638
 639         pqsignal_no_restart(SIGHUP, SIGHUP_handler);    /* reread config file and
 640                                                                                                          * have children do same */
 641         pqsignal_no_restart(SIGINT, pmdie); /* send SIGTERM and shut down */
 642         pqsignal_no_restart(SIGQUIT, pmdie);    /* send SIGQUIT and die */
 643         pqsignal_no_restart(SIGTERM, pmdie);    /* wait for children and shut down */
 644         pqsignal(SIGALRM, SIG_IGN); /* ignored */
 645         pqsignal(SIGPIPE, SIG_IGN); /* ignored */
 646         pqsignal_no_restart(SIGUSR1, sigusr1_handler);  /* message from child
 647                                                                                                          * process */
 648         pqsignal_no_restart(SIGUSR2, dummy_handler);    /* unused, reserve for
 649                                                                                                          * children */
 650         pqsignal_no_restart(SIGCHLD, reaper);   /* handle child termination */
 651         pqsignal(SIGTTIN, SIG_IGN); /* ignored */
 652         pqsignal(SIGTTOU, SIG_IGN); /* ignored */
 653         /* ignore SIGXFSZ, so that ulimit violations work like disk full */
 654 #ifdef SIGXFSZ
 655         pqsignal(SIGXFSZ, SIG_IGN); /* ignored */
 656 #endif
 657
 658         /*
 659          * Options setup
 660          */
 661         InitializeGUCOptions();
 662
 663         opterr = 1;
 664
 665         /*
 666          * Parse command-line options.  CAUTION: keep this in sync with
 667          * tcop/postgres.c (the option sets should not conflict) and with the
 668          * common help() function in main/main.c.
 669          */
 670         while ((opt = getopt(argc, argv, "B:bc:C:D:d:EeFf:h:ijk:lN:nOo:Pp:r:S:sTt:W:-:")) != -1)
 671         {
 672                 switch (opt)
 673                 {
 674                         case 'B':
 675                                 SetConfigOption("shared_buffers", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 676                                 break;
 677
 678                         case 'b':
 679                                 /* Undocumented flag used for binary upgrades */
 680                                 IsBinaryUpgrade = true;
 681                                 break;
 682
 683                         case 'C':
 684                                 output_config_variable = strdup(optarg);
 685                                 break;
 686
 687                         case 'D':
 688                                 userDoption = strdup(optarg);
 689                                 break;
 690
 691                         case 'd':
 692                                 set_debug_options(atoi(optarg), PGC_POSTMASTER, PGC_S_ARGV);
 693                                 break;
 694
 695                         case 'E':
 696                                 SetConfigOption("log_statement", "all", PGC_POSTMASTER, PGC_S_ARGV);
 697                                 break;
 698
 699                         case 'e':
 700                                 SetConfigOption("datestyle", "euro", PGC_POSTMASTER, PGC_S_ARGV);
 701                                 break;
 702
 703                         case 'F':
 704                                 SetConfigOption("fsync", "false", PGC_POSTMASTER, PGC_S_ARGV);
 705                                 break;
 706
 707                         case 'f':
 708                                 if (!set_plan_disabling_options(optarg, PGC_POSTMASTER, PGC_S_ARGV))
 709                                 {
 710                                         write_stderr("%s: invalid argument for option -f: \"%s\"\n",
 711                                                                  progname, optarg);
 712                                         ExitPostmaster(1);
 713                                 }
 714                                 break;
 715
 716                         case 'h':
 717                                 SetConfigOption("listen_addresses", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 718                                 break;
 719
 720                         case 'i':
 721                                 SetConfigOption("listen_addresses", "*", PGC_POSTMASTER, PGC_S_ARGV);
 722                                 break;
 723
 724                         case 'j':
 725                                 /* only used by interactive backend */
 726                                 break;
 727
 728                         case 'k':
 729                                 SetConfigOption("unix_socket_directories", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 730                                 break;
 731
 732                         case 'l':
 733                                 SetConfigOption("ssl", "true", PGC_POSTMASTER, PGC_S_ARGV);
 734                                 break;
 735
 736                         case 'N':
 737                                 SetConfigOption("max_connections", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 738                                 break;
 739
 740                         case 'n':
 741                                 /* Don't reinit shared mem after abnormal exit */
 742                                 Reinit = false;
 743                                 break;
 744
 745                         case 'O':
 746                                 SetConfigOption("allow_system_table_mods", "true", PGC_POSTMASTER, PGC_S_ARGV);
 747                                 break;
 748
 749                         case 'o':
 750                                 /* Other options to pass to the backend on the command line */
 751                                 snprintf(ExtraOptions + strlen(ExtraOptions),
 752                                                  sizeof(ExtraOptions) - strlen(ExtraOptions),
 753                                                  " %s", optarg);
 754                                 break;
 755
 756                         case 'P':
 757                                 SetConfigOption("ignore_system_indexes", "true", PGC_POSTMASTER, PGC_S_ARGV);
 758                                 break;
 759
 760                         case 'p':
 761                                 SetConfigOption("port", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 762                                 break;
 763
 764                         case 'r':
 765                                 /* only used by single-user backend */
 766                                 break;
 767
 768                         case 'S':
 769                                 SetConfigOption("work_mem", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 770                                 break;
 771
 772                         case 's':
 773                                 SetConfigOption("log_statement_stats", "true", PGC_POSTMASTER, PGC_S_ARGV);
 774                                 break;
 775
 776                         case 'T':
 777
 778                                 /*
 779                                  * In the event that some backend dumps core, send SIGSTOP,
 780                                  * rather than SIGQUIT, to all its peers.  This lets the wily
 781                                  * post_hacker collect core dumps from everyone.
 782                                  */
 783                                 SendStop = true;
 784                                 break;
 785
 786                         case 't':
 787                                 {
 788                                         const char *tmp = get_stats_option_name(optarg);
 789
 790                                         if (tmp)
 791                                         {
 792                                                 SetConfigOption(tmp, "true", PGC_POSTMASTER, PGC_S_ARGV);
 793                                         }
 794                                         else
 795                                         {
 796                                                 write_stderr("%s: invalid argument for option -t: \"%s\"\n",
 797                                                                          progname, optarg);
 798                                                 ExitPostmaster(1);
 799                                         }
 800                                         break;
 801                                 }
 802
 803                         case 'W':
 804                                 SetConfigOption("post_auth_delay", optarg, PGC_POSTMASTER, PGC_S_ARGV);
 805                                 break;
 806
 807                         case 'c':
 808                         case '-':
 809                                 {
 810                                         char       *name,
 811                                                            *value;
 812
 813                                         ParseLongOption(optarg, &name, &value);
 814                                         if (!value)
 815                                         {
 816                                                 if (opt == '-')
 817                                                         ereport(ERROR,
 818                                                                         (errcode(ERRCODE_SYNTAX_ERROR),
 819                                                                          errmsg("--%s requires a value",
 820                                                                                         optarg)));
 821                                                 else
 822                                                         ereport(ERROR,
 823                                                                         (errcode(ERRCODE_SYNTAX_ERROR),
 824                                                                          errmsg("-c %s requires a value",
 825                                                                                         optarg)));
 826                                         }
 827
 828                                         SetConfigOption(name, value, PGC_POSTMASTER, PGC_S_ARGV);
 829                                         free(name);
 830                                         if (value)
 831                                                 free(value);
 832                                         break;
 833                                 }
 834
 835                         default:
 836                                 write_stderr("Try \"%s --help\" for more information.\n",
 837                                                          progname);
 838                                 ExitPostmaster(1);
 839                 }
 840         }
 841
 842         /*
 843          * Postmaster accepts no non-option switch arguments.
 844          */
 845         if (optind < argc)
 846         {
 847                 write_stderr("%s: invalid argument: \"%s\"\n",
 848                                          progname, argv[optind]);
 849                 write_stderr("Try \"%s --help\" for more information.\n",
 850                                          progname);
 851                 ExitPostmaster(1);
 852         }
 853
 854         /*
 855          * Locate the proper configuration files and data directory, and read
 856          * postgresql.conf for the first time.
 857          */
 858         if (!SelectConfigFiles(userDoption, progname))
 859                 ExitPostmaster(2);
 860
 861         if (output_config_variable != NULL)
 862         {
 863                 /*
 864                  * "-C guc" was specified, so print GUC's value and exit.  No extra
 865                  * permission check is needed because the user is reading inside the
 866                  * data dir.
 867                  */
 868                 const char *config_val = GetConfigOption(output_config_variable,
 869                                                                                                  false, false);
 870
 871                 puts(config_val ? config_val : "");
 872                 ExitPostmaster(0);
 873         }
 874
 875         /* Verify that DataDir looks reasonable */
 876         checkDataDir();
 877
 878         /* And switch working directory into it */
 879         ChangeToDataDir();
 880
 881         /*
 882          * Check for invalid combinations of GUC settings.
 883          */
 884         if (ReservedBackends >= MaxConnections)
 885         {
 886                 write_stderr("%s: superuser_reserved_connections must be less than max_connections\n", progname);
 887                 ExitPostmaster(1);
 888         }
 889         if (max_wal_senders >= MaxConnections)
 890         {
 891                 write_stderr("%s: max_wal_senders must be less than max_connections\n", progname);
 892                 ExitPostmaster(1);
 893         }
 894         if (XLogArchiveMode > ARCHIVE_MODE_OFF && wal_level == WAL_LEVEL_MINIMAL)
 895                 ereport(ERROR,
 896                                 (errmsg("WAL archival cannot be enabled when wal_level is \"minimal\"")));
 897         if (max_wal_senders > 0 && wal_level == WAL_LEVEL_MINIMAL)
 898                 ereport(ERROR,
 899                                 (errmsg("WAL streaming (max_wal_senders > 0) requires wal_level \"replica\" or \"logical\"")));
 900
 901         /*
 902          * Other one-time internal sanity checks can go here, if they are fast.
 903          * (Put any slow processing further down, after postmaster.pid creation.)
 904          */
 905         if (!CheckDateTokenTables())
 906         {
 907                 write_stderr("%s: invalid datetoken tables, please fix\n", progname);
 908                 ExitPostmaster(1);
 909         }
 910
 911         /*
 912          * Now that we are done processing the postmaster arguments, reset
 913          * getopt(3) library so that it will work correctly in subprocesses.
 914          */
 915         optind = 1;
 916 #ifdef HAVE_INT_OPTRESET
 917         optreset = 1;                           /* some systems need this too */
 918 #endif
 919
 920         /* For debugging: display postmaster environment */
 921         {
 922                 extern char **environ;
 923                 char      **p;
 924
 925                 ereport(DEBUG3,
 926                                 (errmsg_internal("%s: PostmasterMain: initial environment dump:",
 927                                                                  progname)));
 928                 ereport(DEBUG3,
 929                                 (errmsg_internal("-----------------------------------------")));
 930                 for (p = environ; *p; ++p)
 931                         ereport(DEBUG3,
 932                                         (errmsg_internal("\t%s", *p)));
 933                 ereport(DEBUG3,
 934                                 (errmsg_internal("-----------------------------------------")));
 935         }
 936
 937         /*
 938          * Create lockfile for data directory.
 939          *
 940          * We want to do this before we try to grab the input sockets, because the
 941          * data directory interlock is more reliable than the socket-file
 942          * interlock (thanks to whoever decided to put socket files in /tmp :-().
 943          * For the same reason, it's best to grab the TCP socket(s) before the
 944          * Unix socket(s).
 945          *
 946          * Also note that this internally sets up the on_proc_exit function that
 947          * is responsible for removing both data directory and socket lockfiles;
 948          * so it must happen before opening sockets so that at exit, the socket
 949          * lockfiles go away after CloseServerPorts runs.
 950          */
 951         CreateDataDirLockFile(true);
 952
 953         /* read control file (error checking and contains config) */
 954         LocalProcessControlFile(false);
 955
 956         /*
 957          * Initialize SSL library, if specified.
 958          */
 959 #ifdef USE_SSL
 960         if (EnableSSL)
 961         {
 962                 (void) secure_initialize(true);
 963                 LoadedSSL = true;
 964         }
 965 #endif
 966
 967         /*
 968          * Register the apply launcher.  Since it registers a background worker,
 969          * it needs to be called before InitializeMaxBackends(), and it's probably
 970          * a good idea to call it before any modules had chance to take the
 971          * background worker slots.
 972          */
 973         ApplyLauncherRegister();
 974
 975         /*
 976          * process any libraries that should be preloaded at postmaster start
 977          */
 978         process_shared_preload_libraries();
 979
 980         /*
 981          * Now that loadable modules have had their chance to register background
 982          * workers, calculate MaxBackends.
 983          */
 984         InitializeMaxBackends();
 985
 986         /*
 987          * Establish input sockets.
 988          *
 989          * First, mark them all closed, and set up an on_proc_exit function that's
 990          * charged with closing the sockets again at postmaster shutdown.
 991          */
 992         for (i = 0; i < MAXLISTEN; i++)
 993                 ListenSocket[i] = PGINVALID_SOCKET;
 994
 995         on_proc_exit(CloseServerPorts, 0);
 996
 997         if (ListenAddresses)
 998         {
 999                 char       *rawstring;
1000                 List       *elemlist;
1001                 ListCell   *l;
1002                 int                     success = 0;
1003
1004                 /* Need a modifiable copy of ListenAddresses */
1005                 rawstring = pstrdup(ListenAddresses);
1006
1007                 /* Parse string into list of hostnames */
1008                 if (!SplitIdentifierString(rawstring, ',', &elemlist))
1009                 {
1010                         /* syntax error in list */
1011                         ereport(FATAL,
1012                                         (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1013                                          errmsg("invalid list syntax in parameter \"%s\"",
1014                                                         "listen_addresses")));
1015                 }
1016
1017                 foreach(l, elemlist)
1018                 {
1019                         char       *curhost = (char *) lfirst(l);
1020
1021                         if (strcmp(curhost, "*") == 0)
1022                                 status = StreamServerPort(AF_UNSPEC, NULL,
1023                                                                                   (unsigned short) PostPortNumber,
1024                                                                                   NULL,
1025                                                                                   ListenSocket, MAXLISTEN);
1026                         else
1027                                 status = StreamServerPort(AF_UNSPEC, curhost,
1028                                                                                   (unsigned short) PostPortNumber,
1029                                                                                   NULL,
1030                                                                                   ListenSocket, MAXLISTEN);
1031
1032                         if (status == STATUS_OK)
1033                         {
1034                                 success++;
1035                                 /* record the first successful host addr in lockfile */
1036                                 if (!listen_addr_saved)
1037                                 {
1038                                         AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, curhost);
1039                                         listen_addr_saved = true;
1040                                 }
1041                         }
1042                         else
1043                                 ereport(WARNING,
1044                                                 (errmsg("could not create listen socket for \"%s\"",
1045                                                                 curhost)));
1046                 }
1047
1048                 if (!success && elemlist != NIL)
1049                         ereport(FATAL,
1050                                         (errmsg("could not create any TCP/IP sockets")));
1051
1052                 list_free(elemlist);
1053                 pfree(rawstring);
1054         }
1055
1056 #ifdef USE_BONJOUR
1057         /* Register for Bonjour only if we opened TCP socket(s) */
1058         if (enable_bonjour && ListenSocket[0] != PGINVALID_SOCKET)
1059         {
1060                 DNSServiceErrorType err;
1061
1062                 /*
1063                  * We pass 0 for interface_index, which will result in registering on
1064                  * all "applicable" interfaces.  It's not entirely clear from the
1065                  * DNS-SD docs whether this would be appropriate if we have bound to
1066                  * just a subset of the available network interfaces.
1067                  */
1068                 err = DNSServiceRegister(&bonjour_sdref,
1069                                                                  0,
1070                                                                  0,
1071                                                                  bonjour_name,
1072                                                                  "_postgresql._tcp.",
1073                                                                  NULL,
1074                                                                  NULL,
1075                                                                  htons(PostPortNumber),
1076                                                                  0,
1077                                                                  NULL,
1078                                                                  NULL,
1079                                                                  NULL);
1080                 if (err != kDNSServiceErr_NoError)
1081                         elog(LOG, "DNSServiceRegister() failed: error code %ld",
1082                                  (long) err);
1083
1084                 /*
1085                  * We don't bother to read the mDNS daemon's reply, and we expect that
1086                  * it will automatically terminate our registration when the socket is
1087                  * closed at postmaster termination.  So there's nothing more to be
1088                  * done here.  However, the bonjour_sdref is kept around so that
1089                  * forked children can close their copies of the socket.
1090                  */
1091         }
1092 #endif
1093
1094 #ifdef HAVE_UNIX_SOCKETS
1095         if (Unix_socket_directories)
1096         {
1097                 char       *rawstring;
1098                 List       *elemlist;
1099                 ListCell   *l;
1100                 int                     success = 0;
1101
1102                 /* Need a modifiable copy of Unix_socket_directories */
1103                 rawstring = pstrdup(Unix_socket_directories);
1104
1105                 /* Parse string into list of directories */
1106                 if (!SplitDirectoriesString(rawstring, ',', &elemlist))
1107                 {
1108                         /* syntax error in list */
1109                         ereport(FATAL,
1110                                         (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1111                                          errmsg("invalid list syntax in parameter \"%s\"",
1112                                                         "unix_socket_directories")));
1113                 }
1114
1115                 foreach(l, elemlist)
1116                 {
1117                         char       *socketdir = (char *) lfirst(l);
1118
1119                         status = StreamServerPort(AF_UNIX, NULL,
1120                                                                           (unsigned short) PostPortNumber,
1121                                                                           socketdir,
1122                                                                           ListenSocket, MAXLISTEN);
1123
1124                         if (status == STATUS_OK)
1125                         {
1126                                 success++;
1127                                 /* record the first successful Unix socket in lockfile */
1128                                 if (success == 1)
1129                                         AddToDataDirLockFile(LOCK_FILE_LINE_SOCKET_DIR, socketdir);
1130                         }
1131                         else
1132                                 ereport(WARNING,
1133                                                 (errmsg("could not create Unix-domain socket in directory \"%s\"",
1134                                                                 socketdir)));
1135                 }
1136
1137                 if (!success && elemlist != NIL)
1138                         ereport(FATAL,
1139                                         (errmsg("could not create any Unix-domain sockets")));
1140
1141                 list_free_deep(elemlist);
1142                 pfree(rawstring);
1143         }
1144 #endif
1145
1146         /*
1147          * check that we have some socket to listen on
1148          */
1149         if (ListenSocket[0] == PGINVALID_SOCKET)
1150                 ereport(FATAL,
1151                                 (errmsg("no socket created for listening")));
1152
1153         /*
1154          * If no valid TCP ports, write an empty line for listen address,
1155          * indicating the Unix socket must be used.  Note that this line is not
1156          * added to the lock file until there is a socket backing it.
1157          */
1158         if (!listen_addr_saved)
1159                 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, "");
1160
1161         /*
1162          * Set up shared memory and semaphores.
1163          */
1164         reset_shared(PostPortNumber);
1165
1166         /*
1167          * Estimate number of openable files.  This must happen after setting up
1168          * semaphores, because on some platforms semaphores count as open files.
1169          */
1170         set_max_safe_fds();
1171
1172         /*
1173          * Set reference point for stack-depth checking.
1174          */
1175         set_stack_base();
1176
1177         /*
1178          * Initialize pipe (or process handle on Windows) that allows children to
1179          * wake up from sleep on postmaster death.
1180          */
1181         InitPostmasterDeathWatchHandle();
1182
1183 #ifdef WIN32
1184
1185         /*
1186          * Initialize I/O completion port used to deliver list of dead children.
1187          */
1188         win32ChildQueue = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 1);
1189         if (win32ChildQueue == NULL)
1190                 ereport(FATAL,
1191                                 (errmsg("could not create I/O completion port for child queue")));
1192 #endif
1193
1194         /*
1195          * Record postmaster options.  We delay this till now to avoid recording
1196          * bogus options (eg, NBuffers too high for available memory).
1197          */
1198         if (!CreateOptsFile(argc, argv, my_exec_path))
1199                 ExitPostmaster(1);
1200
1201 #ifdef EXEC_BACKEND
1202         /* Write out nondefault GUC settings for child processes to use */
1203         write_nondefault_variables(PGC_POSTMASTER);
1204 #endif
1205
1206         /*
1207          * Write the external PID file if requested
1208          */
1209         if (external_pid_file)
1210         {
1211                 FILE       *fpidfile = fopen(external_pid_file, "w");
1212
1213                 if (fpidfile)
1214                 {
1215                         fprintf(fpidfile, "%d\n", MyProcPid);
1216                         fclose(fpidfile);
1217
1218                         /* Make PID file world readable */
1219                         if (chmod(external_pid_file, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH) != 0)
1220                                 write_stderr("%s: could not change permissions of external PID file \"%s\": %s\n",
1221                                                          progname, external_pid_file, strerror(errno));
1222                 }
1223                 else
1224                         write_stderr("%s: could not write external PID file \"%s\": %s\n",
1225                                                  progname, external_pid_file, strerror(errno));
1226
1227                 on_proc_exit(unlink_external_pid_file, 0);
1228         }
1229
1230         /*
1231          * Remove old temporary files.  At this point there can be no other
1232          * Postgres processes running in this directory, so this should be safe.
1233          */
1234         RemovePgTempFiles();
1235
1236         /*
1237          * Forcibly remove the files signaling a standby promotion request.
1238          * Otherwise, the existence of those files triggers a promotion too early,
1239          * whether a user wants that or not.
1240          *
1241          * This removal of files is usually unnecessary because they can exist
1242          * only during a few moments during a standby promotion. However there is
1243          * a race condition: if pg_ctl promote is executed and creates the files
1244          * during a promotion, the files can stay around even after the server is
1245          * brought up to new master. Then, if new standby starts by using the
1246          * backup taken from that master, the files can exist at the server
1247          * startup and should be removed in order to avoid an unexpected
1248          * promotion.
1249          *
1250          * Note that promotion signal files need to be removed before the startup
1251          * process is invoked. Because, after that, they can be used by
1252          * postmaster's SIGUSR1 signal handler.
1253          */
1254         RemovePromoteSignalFiles();
1255
1256         /* Remove any outdated file holding the current log filenames. */
1257         if (unlink(LOG_METAINFO_DATAFILE) < 0 && errno != ENOENT)
1258                 ereport(LOG,
1259                                 (errcode_for_file_access(),
1260                                  errmsg("could not remove file \"%s\": %m",
1261                                                 LOG_METAINFO_DATAFILE)));
1262
1263         /*
1264          * If enabled, start up syslogger collection subprocess
1265          */
1266         SysLoggerPID = SysLogger_Start();
1267
1268         /*
1269          * Reset whereToSendOutput from DestDebug (its starting state) to
1270          * DestNone. This stops ereport from sending log messages to stderr unless
1271          * Log_destination permits.  We don't do this until the postmaster is
1272          * fully launched, since startup failures may as well be reported to
1273          * stderr.
1274          *
1275          * If we are in fact disabling logging to stderr, first emit a log message
1276          * saying so, to provide a breadcrumb trail for users who may not remember
1277          * that their logging is configured to go somewhere else.
1278          */
1279         if (!(Log_destination & LOG_DESTINATION_STDERR))
1280                 ereport(LOG,
1281                                 (errmsg("ending log output to stderr"),
1282                                  errhint("Future log output will go to log destination \"%s\".",
1283                                                  Log_destination_string)));
1284
1285         whereToSendOutput = DestNone;
1286
1287         /*
1288          * Initialize stats collection subsystem (this does NOT start the
1289          * collector process!)
1290          */
1291         pgstat_init();
1292
1293         /*
1294          * Initialize the autovacuum subsystem (again, no process start yet)
1295          */
1296         autovac_init();
1297
1298         /*
1299          * Load configuration files for client authentication.
1300          */
1301         if (!load_hba())
1302         {
1303                 /*
1304                  * It makes no sense to continue if we fail to load the HBA file,
1305                  * since there is no way to connect to the database in this case.
1306                  */
1307                 ereport(FATAL,
1308                                 (errmsg("could not load pg_hba.conf")));
1309         }
1310         if (!load_ident())
1311         {
1312                 /*
1313                  * We can start up without the IDENT file, although it means that you
1314                  * cannot log in using any of the authentication methods that need a
1315                  * user name mapping. load_ident() already logged the details of error
1316                  * to the log.
1317                  */
1318         }
1319
1320 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1321
1322         /*
1323          * On macOS, libintl replaces setlocale() with a version that calls
1324          * CFLocaleCopyCurrent() when its second argument is "" and every relevant
1325          * environment variable is unset or empty.  CFLocaleCopyCurrent() makes
1326          * the process multithreaded.  The postmaster calls sigprocmask() and
1327          * calls fork() without an immediate exec(), both of which have undefined
1328          * behavior in a multithreaded program.  A multithreaded postmaster is the
1329          * normal case on Windows, which offers neither fork() nor sigprocmask().
1330          */
1331         if (pthread_is_threaded_np() != 0)
1332                 ereport(FATAL,
1333                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1334                                  errmsg("postmaster became multithreaded during startup"),
1335                                  errhint("Set the LC_ALL environment variable to a valid locale.")));
1336 #endif
1337
1338         /*
1339          * Remember postmaster startup time
1340          */
1341         PgStartTime = GetCurrentTimestamp();
1342 #ifndef HAVE_STRONG_RANDOM
1343         /* RandomCancelKey wants its own copy */
1344         gettimeofday(&random_start_time, NULL);
1345 #endif
1346
1347         /*
1348          * Report postmaster status in the postmaster.pid file, to allow pg_ctl to
1349          * see what's happening.
1350          */
1351         AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STARTING);
1352
1353         /*
1354          * We're ready to rock and roll...
1355          */
1356         StartupPID = StartupDataBase();
1357         Assert(StartupPID != 0);
1358         StartupStatus = STARTUP_RUNNING;
1359         pmState = PM_STARTUP;
1360
1361         /* Some workers may be scheduled to start now */
1362         maybe_start_bgworkers();
1363
1364         status = ServerLoop();
1365
1366         /*
1367          * ServerLoop probably shouldn't ever return, but if it does, close down.
1368          */
1369         ExitPostmaster(status != STATUS_OK);
1370
1371         abort();                                        /* not reached */
1372 }
1373
1374
1375 /*
1376  * on_proc_exit callback to close server's listen sockets
1377  */
1378 static void
1379 CloseServerPorts(int status, Datum arg)
1380 {
1381         int                     i;
1382
1383         /*
1384          * First, explicitly close all the socket FDs.  We used to just let this
1385          * happen implicitly at postmaster exit, but it's better to close them
1386          * before we remove the postmaster.pid lockfile; otherwise there's a race
1387          * condition if a new postmaster wants to re-use the TCP port number.
1388          */
1389         for (i = 0; i < MAXLISTEN; i++)
1390         {
1391                 if (ListenSocket[i] != PGINVALID_SOCKET)
1392                 {
1393                         StreamClose(ListenSocket[i]);
1394                         ListenSocket[i] = PGINVALID_SOCKET;
1395                 }
1396         }
1397
1398         /*
1399          * Next, remove any filesystem entries for Unix sockets.  To avoid race
1400          * conditions against incoming postmasters, this must happen after closing
1401          * the sockets and before removing lock files.
1402          */
1403         RemoveSocketFiles();
1404
1405         /*
1406          * We don't do anything about socket lock files here; those will be
1407          * removed in a later on_proc_exit callback.
1408          */
1409 }
1410
1411 /*
1412  * on_proc_exit callback to delete external_pid_file
1413  */
1414 static void
1415 unlink_external_pid_file(int status, Datum arg)
1416 {
1417         if (external_pid_file)
1418                 unlink(external_pid_file);
1419 }
1420
1421
1422 /*
1423  * Compute and check the directory paths to files that are part of the
1424  * installation (as deduced from the postgres executable's own location)
1425  */
1426 static void
1427 getInstallationPaths(const char *argv0)
1428 {
1429         DIR                *pdir;
1430
1431         /* Locate the postgres executable itself */
1432         if (find_my_exec(argv0, my_exec_path) < 0)
1433                 elog(FATAL, "%s: could not locate my own executable path", argv0);
1434
1435 #ifdef EXEC_BACKEND
1436         /* Locate executable backend before we change working directory */
1437         if (find_other_exec(argv0, "postgres", PG_BACKEND_VERSIONSTR,
1438                                                 postgres_exec_path) < 0)
1439                 ereport(FATAL,
1440                                 (errmsg("%s: could not locate matching postgres executable",
1441                                                 argv0)));
1442 #endif
1443
1444         /*
1445          * Locate the pkglib directory --- this has to be set early in case we try
1446          * to load any modules from it in response to postgresql.conf entries.
1447          */
1448         get_pkglib_path(my_exec_path, pkglib_path);
1449
1450         /*
1451          * Verify that there's a readable directory there; otherwise the Postgres
1452          * installation is incomplete or corrupt.  (A typical cause of this
1453          * failure is that the postgres executable has been moved or hardlinked to
1454          * some directory that's not a sibling of the installation lib/
1455          * directory.)
1456          */
1457         pdir = AllocateDir(pkglib_path);
1458         if (pdir == NULL)
1459                 ereport(ERROR,
1460                                 (errcode_for_file_access(),
1461                                  errmsg("could not open directory \"%s\": %m",
1462                                                 pkglib_path),
1463                                  errhint("This may indicate an incomplete PostgreSQL installation, or that the file \"%s\" has been moved away from its proper location.",
1464                                                  my_exec_path)));
1465         FreeDir(pdir);
1466
1467         /*
1468          * XXX is it worth similarly checking the share/ directory?  If the lib/
1469          * directory is there, then share/ probably is too.
1470          */
1471 }
1472
1473
1474 /*
1475  * Validate the proposed data directory
1476  */
1477 static void
1478 checkDataDir(void)
1479 {
1480         char            path[MAXPGPATH];
1481         FILE       *fp;
1482         struct stat stat_buf;
1483
1484         Assert(DataDir);
1485
1486         if (stat(DataDir, &stat_buf) != 0)
1487         {
1488                 if (errno == ENOENT)
1489                         ereport(FATAL,
1490                                         (errcode_for_file_access(),
1491                                          errmsg("data directory \"%s\" does not exist",
1492                                                         DataDir)));
1493                 else
1494                         ereport(FATAL,
1495                                         (errcode_for_file_access(),
1496                                          errmsg("could not read permissions of directory \"%s\": %m",
1497                                                         DataDir)));
1498         }
1499
1500         /* eventual chdir would fail anyway, but let's test ... */
1501         if (!S_ISDIR(stat_buf.st_mode))
1502                 ereport(FATAL,
1503                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1504                                  errmsg("specified data directory \"%s\" is not a directory",
1505                                                 DataDir)));
1506
1507         /*
1508          * Check that the directory belongs to my userid; if not, reject.
1509          *
1510          * This check is an essential part of the interlock that prevents two
1511          * postmasters from starting in the same directory (see CreateLockFile()).
1512          * Do not remove or weaken it.
1513          *
1514          * XXX can we safely enable this check on Windows?
1515          */
1516 #if !defined(WIN32) && !defined(__CYGWIN__)
1517         if (stat_buf.st_uid != geteuid())
1518                 ereport(FATAL,
1519                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1520                                  errmsg("data directory \"%s\" has wrong ownership",
1521                                                 DataDir),
1522                                  errhint("The server must be started by the user that owns the data directory.")));
1523 #endif
1524
1525         /*
1526          * Check if the directory has group or world access.  If so, reject.
1527          *
1528          * It would be possible to allow weaker constraints (for example, allow
1529          * group access) but we cannot make a general assumption that that is
1530          * okay; for example there are platforms where nearly all users
1531          * customarily belong to the same group.  Perhaps this test should be
1532          * configurable.
1533          *
1534          * XXX temporarily suppress check when on Windows, because there may not
1535          * be proper support for Unix-y file permissions.  Need to think of a
1536          * reasonable check to apply on Windows.
1537          */
1538 #if !defined(WIN32) && !defined(__CYGWIN__)
1539         if (stat_buf.st_mode & (S_IRWXG | S_IRWXO))
1540                 ereport(FATAL,
1541                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1542                                  errmsg("data directory \"%s\" has group or world access",
1543                                                 DataDir),
1544                                  errdetail("Permissions should be u=rwx (0700).")));
1545 #endif
1546
1547         /* Look for PG_VERSION before looking for pg_control */
1548         ValidatePgVersion(DataDir);
1549
1550         snprintf(path, sizeof(path), "%s/global/pg_control", DataDir);
1551
1552         fp = AllocateFile(path, PG_BINARY_R);
1553         if (fp == NULL)
1554         {
1555                 write_stderr("%s: could not find the database system\n"
1556                                          "Expected to find it in the directory \"%s\",\n"
1557                                          "but could not open file \"%s\": %s\n",
1558                                          progname, DataDir, path, strerror(errno));
1559                 ExitPostmaster(2);
1560         }
1561         FreeFile(fp);
1562 }
1563
1564 /*
1565  * Determine how long should we let ServerLoop sleep.
1566  *
1567  * In normal conditions we wait at most one minute, to ensure that the other
1568  * background tasks handled by ServerLoop get done even when no requests are
1569  * arriving.  However, if there are background workers waiting to be started,
1570  * we don't actually sleep so that they are quickly serviced.  Other exception
1571  * cases are as shown in the code.
1572  */
1573 static void
1574 DetermineSleepTime(struct timeval *timeout)
1575 {
1576         TimestampTz next_wakeup = 0;
1577
1578         /*
1579          * Normal case: either there are no background workers at all, or we're in
1580          * a shutdown sequence (during which we ignore bgworkers altogether).
1581          */
1582         if (Shutdown > NoShutdown ||
1583                 (!StartWorkerNeeded && !HaveCrashedWorker))
1584         {
1585                 if (AbortStartTime != 0)
1586                 {
1587                         /* time left to abort; clamp to 0 in case it already expired */
1588                         timeout->tv_sec = SIGKILL_CHILDREN_AFTER_SECS -
1589                                 (time(NULL) - AbortStartTime);
1590                         timeout->tv_sec = Max(timeout->tv_sec, 0);
1591                         timeout->tv_usec = 0;
1592                 }
1593                 else
1594                 {
1595                         timeout->tv_sec = 60;
1596                         timeout->tv_usec = 0;
1597                 }
1598                 return;
1599         }
1600
1601         if (StartWorkerNeeded)
1602         {
1603                 timeout->tv_sec = 0;
1604                 timeout->tv_usec = 0;
1605                 return;
1606         }
1607
1608         if (HaveCrashedWorker)
1609         {
1610                 slist_mutable_iter siter;
1611
1612                 /*
1613                  * When there are crashed bgworkers, we sleep just long enough that
1614                  * they are restarted when they request to be.  Scan the list to
1615                  * determine the minimum of all wakeup times according to most recent
1616                  * crash time and requested restart interval.
1617                  */
1618                 slist_foreach_modify(siter, &BackgroundWorkerList)
1619                 {
1620                         RegisteredBgWorker *rw;
1621                         TimestampTz this_wakeup;
1622
1623                         rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
1624
1625                         if (rw->rw_crashed_at == 0)
1626                                 continue;
1627
1628                         if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART
1629                                 || rw->rw_terminate)
1630                         {
1631                                 ForgetBackgroundWorker(&siter);
1632                                 continue;
1633                         }
1634
1635                         this_wakeup = TimestampTzPlusMilliseconds(rw->rw_crashed_at,
1636                                                                                                           1000L * rw->rw_worker.bgw_restart_time);
1637                         if (next_wakeup == 0 || this_wakeup < next_wakeup)
1638                                 next_wakeup = this_wakeup;
1639                 }
1640         }
1641
1642         if (next_wakeup != 0)
1643         {
1644                 long            secs;
1645                 int                     microsecs;
1646
1647                 TimestampDifference(GetCurrentTimestamp(), next_wakeup,
1648                                                         &secs, &microsecs);
1649                 timeout->tv_sec = secs;
1650                 timeout->tv_usec = microsecs;
1651
1652                 /* Ensure we don't exceed one minute */
1653                 if (timeout->tv_sec > 60)
1654                 {
1655                         timeout->tv_sec = 60;
1656                         timeout->tv_usec = 0;
1657                 }
1658         }
1659         else
1660         {
1661                 timeout->tv_sec = 60;
1662                 timeout->tv_usec = 0;
1663         }
1664 }
1665
1666 /*
1667  * Main idle loop of postmaster
1668  *
1669  * NB: Needs to be called with signals blocked
1670  */
1671 static int
1672 ServerLoop(void)
1673 {
1674         fd_set          readmask;
1675         int                     nSockets;
1676         time_t          last_lockfile_recheck_time,
1677                                 last_touch_time;
1678
1679         last_lockfile_recheck_time = last_touch_time = time(NULL);
1680
1681         nSockets = initMasks(&readmask);
1682
1683         for (;;)
1684         {
1685                 fd_set          rmask;
1686                 int                     selres;
1687                 time_t          now;
1688
1689                 /*
1690                  * Wait for a connection request to arrive.
1691                  *
1692                  * We block all signals except while sleeping. That makes it safe for
1693                  * signal handlers, which again block all signals while executing, to
1694                  * do nontrivial work.
1695                  *
1696                  * If we are in PM_WAIT_DEAD_END state, then we don't want to accept
1697                  * any new connections, so we don't call select(), and just sleep.
1698                  */
1699                 memcpy((char *) &rmask, (char *) &readmask, sizeof(fd_set));
1700
1701                 if (pmState == PM_WAIT_DEAD_END)
1702                 {
1703                         PG_SETMASK(&UnBlockSig);
1704
1705                         pg_usleep(100000L); /* 100 msec seems reasonable */
1706                         selres = 0;
1707
1708                         PG_SETMASK(&BlockSig);
1709                 }
1710                 else
1711                 {
1712                         /* must set timeout each time; some OSes change it! */
1713                         struct timeval timeout;
1714
1715                         /* Needs to run with blocked signals! */
1716                         DetermineSleepTime(&timeout);
1717
1718                         PG_SETMASK(&UnBlockSig);
1719
1720                         selres = select(nSockets, &rmask, NULL, NULL, &timeout);
1721
1722                         PG_SETMASK(&BlockSig);
1723                 }
1724
1725                 /* Now check the select() result */
1726                 if (selres < 0)
1727                 {
1728                         if (errno != EINTR && errno != EWOULDBLOCK)
1729                         {
1730                                 ereport(LOG,
1731                                                 (errcode_for_socket_access(),
1732                                                  errmsg("select() failed in postmaster: %m")));
1733                                 return STATUS_ERROR;
1734                         }
1735                 }
1736
1737                 /*
1738                  * New connection pending on any of our sockets? If so, fork a child
1739                  * process to deal with it.
1740                  */
1741                 if (selres > 0)
1742                 {
1743                         int                     i;
1744
1745                         for (i = 0; i < MAXLISTEN; i++)
1746                         {
1747                                 if (ListenSocket[i] == PGINVALID_SOCKET)
1748                                         break;
1749                                 if (FD_ISSET(ListenSocket[i], &rmask))
1750                                 {
1751                                         Port       *port;
1752
1753                                         port = ConnCreate(ListenSocket[i]);
1754                                         if (port)
1755                                         {
1756                                                 BackendStartup(port);
1757
1758                                                 /*
1759                                                  * We no longer need the open socket or port structure
1760                                                  * in this process
1761                                                  */
1762                                                 StreamClose(port->sock);
1763                                                 ConnFree(port);
1764                                         }
1765                                 }
1766                         }
1767                 }
1768
1769                 /* If we have lost the log collector, try to start a new one */
1770                 if (SysLoggerPID == 0 && Logging_collector)
1771                         SysLoggerPID = SysLogger_Start();
1772
1773                 /*
1774                  * If no background writer process is running, and we are not in a
1775                  * state that prevents it, start one.  It doesn't matter if this
1776                  * fails, we'll just try again later.  Likewise for the checkpointer.
1777                  */
1778                 if (pmState == PM_RUN || pmState == PM_RECOVERY ||
1779                         pmState == PM_HOT_STANDBY)
1780                 {
1781                         if (CheckpointerPID == 0)
1782                                 CheckpointerPID = StartCheckpointer();
1783                         if (BgWriterPID == 0)
1784                                 BgWriterPID = StartBackgroundWriter();
1785                 }
1786
1787                 /*
1788                  * Likewise, if we have lost the walwriter process, try to start a new
1789                  * one.  But this is needed only in normal operation (else we cannot
1790                  * be writing any new WAL).
1791                  */
1792                 if (WalWriterPID == 0 && pmState == PM_RUN)
1793                         WalWriterPID = StartWalWriter();
1794
1795                 /*
1796                  * If we have lost the autovacuum launcher, try to start a new one. We
1797                  * don't want autovacuum to run in binary upgrade mode because
1798                  * autovacuum might update relfrozenxid for empty tables before the
1799                  * physical files are put in place.
1800                  */
1801                 if (!IsBinaryUpgrade && AutoVacPID == 0 &&
1802                         (AutoVacuumingActive() || start_autovac_launcher) &&
1803                         pmState == PM_RUN)
1804                 {
1805                         AutoVacPID = StartAutoVacLauncher();
1806                         if (AutoVacPID != 0)
1807                                 start_autovac_launcher = false; /* signal processed */
1808                 }
1809
1810                 /* If we have lost the stats collector, try to start a new one */
1811                 if (PgStatPID == 0 &&
1812                         (pmState == PM_RUN || pmState == PM_HOT_STANDBY))
1813                         PgStatPID = pgstat_start();
1814
1815                 /* If we have lost the archiver, try to start a new one. */
1816                 if (PgArchPID == 0 && PgArchStartupAllowed())
1817                         PgArchPID = pgarch_start();
1818
1819                 /* If we need to signal the autovacuum launcher, do so now */
1820                 if (avlauncher_needs_signal)
1821                 {
1822                         avlauncher_needs_signal = false;
1823                         if (AutoVacPID != 0)
1824                                 kill(AutoVacPID, SIGUSR2);
1825                 }
1826
1827                 /* If we need to start a WAL receiver, try to do that now */
1828                 if (WalReceiverRequested)
1829                         MaybeStartWalReceiver();
1830
1831                 /* Get other worker processes running, if needed */
1832                 if (StartWorkerNeeded || HaveCrashedWorker)
1833                         maybe_start_bgworkers();
1834
1835 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1836
1837                 /*
1838                  * With assertions enabled, check regularly for appearance of
1839                  * additional threads.  All builds check at start and exit.
1840                  */
1841                 Assert(pthread_is_threaded_np() == 0);
1842 #endif
1843
1844                 /*
1845                  * Lastly, check to see if it's time to do some things that we don't
1846                  * want to do every single time through the loop, because they're a
1847                  * bit expensive.  Note that there's up to a minute of slop in when
1848                  * these tasks will be performed, since DetermineSleepTime() will let
1849                  * us sleep at most that long; except for SIGKILL timeout which has
1850                  * special-case logic there.
1851                  */
1852                 now = time(NULL);
1853
1854                 /*
1855                  * If we already sent SIGQUIT to children and they are slow to shut
1856                  * down, it's time to send them SIGKILL.  This doesn't happen
1857                  * normally, but under certain conditions backends can get stuck while
1858                  * shutting down.  This is a last measure to get them unwedged.
1859                  *
1860                  * Note we also do this during recovery from a process crash.
1861                  */
1862                 if ((Shutdown >= ImmediateShutdown || (FatalError && !SendStop)) &&
1863                         AbortStartTime != 0 &&
1864                         (now - AbortStartTime) >= SIGKILL_CHILDREN_AFTER_SECS)
1865                 {
1866                         /* We were gentle with them before. Not anymore */
1867                         TerminateChildren(SIGKILL);
1868                         /* reset flag so we don't SIGKILL again */
1869                         AbortStartTime = 0;
1870                 }
1871
1872                 /*
1873                  * Once a minute, verify that postmaster.pid hasn't been removed or
1874                  * overwritten.  If it has, we force a shutdown.  This avoids having
1875                  * postmasters and child processes hanging around after their database
1876                  * is gone, and maybe causing problems if a new database cluster is
1877                  * created in the same place.  It also provides some protection
1878                  * against a DBA foolishly removing postmaster.pid and manually
1879                  * starting a new postmaster.  Data corruption is likely to ensue from
1880                  * that anyway, but we can minimize the damage by aborting ASAP.
1881                  */
1882                 if (now - last_lockfile_recheck_time >= 1 * SECS_PER_MINUTE)
1883                 {
1884                         if (!RecheckDataDirLockFile())
1885                         {
1886                                 ereport(LOG,
1887                                                 (errmsg("performing immediate shutdown because data directory lock file is invalid")));
1888                                 kill(MyProcPid, SIGQUIT);
1889                         }
1890                         last_lockfile_recheck_time = now;
1891                 }
1892
1893                 /*
1894                  * Touch Unix socket and lock files every 58 minutes, to ensure that
1895                  * they are not removed by overzealous /tmp-cleaning tasks.  We assume
1896                  * no one runs cleaners with cutoff times of less than an hour ...
1897                  */
1898                 if (now - last_touch_time >= 58 * SECS_PER_MINUTE)
1899                 {
1900                         TouchSocketFiles();
1901                         TouchSocketLockFiles();
1902                         last_touch_time = now;
1903                 }
1904         }
1905 }
1906
1907 /*
1908  * Initialise the masks for select() for the ports we are listening on.
1909  * Return the number of sockets to listen on.
1910  */
1911 static int
1912 initMasks(fd_set *rmask)
1913 {
1914         int                     maxsock = -1;
1915         int                     i;
1916
1917         FD_ZERO(rmask);
1918
1919         for (i = 0; i < MAXLISTEN; i++)
1920         {
1921                 int                     fd = ListenSocket[i];
1922
1923                 if (fd == PGINVALID_SOCKET)
1924                         break;
1925                 FD_SET(fd, rmask);
1926
1927                 if (fd > maxsock)
1928                         maxsock = fd;
1929         }
1930
1931         return maxsock + 1;
1932 }
1933
1934
1935 /*
1936  * Read a client's startup packet and do something according to it.
1937  *
1938  * Returns STATUS_OK or STATUS_ERROR, or might call ereport(FATAL) and
1939  * not return at all.
1940  *
1941  * (Note that ereport(FATAL) stuff is sent to the client, so only use it
1942  * if that's what you want.  Return STATUS_ERROR if you don't want to
1943  * send anything to the client, which would typically be appropriate
1944  * if we detect a communications failure.)
1945  */
1946 static int
1947 ProcessStartupPacket(Port *port, bool SSLdone)
1948 {
1949         int32           len;
1950         void       *buf;
1951         ProtocolVersion proto;
1952         MemoryContext oldcontext;
1953
1954         pq_startmsgread();
1955         if (pq_getbytes((char *) &len, 4) == EOF)
1956         {
1957                 /*
1958                  * EOF after SSLdone probably means the client didn't like our
1959                  * response to NEGOTIATE_SSL_CODE.  That's not an error condition, so
1960                  * don't clutter the log with a complaint.
1961                  */
1962                 if (!SSLdone)
1963                         ereport(COMMERROR,
1964                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
1965                                          errmsg("incomplete startup packet")));
1966                 return STATUS_ERROR;
1967         }
1968
1969         len = ntohl(len);
1970         len -= 4;
1971
1972         if (len < (int32) sizeof(ProtocolVersion) ||
1973                 len > MAX_STARTUP_PACKET_LENGTH)
1974         {
1975                 ereport(COMMERROR,
1976                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1977                                  errmsg("invalid length of startup packet")));
1978                 return STATUS_ERROR;
1979         }
1980
1981         /*
1982          * Allocate at least the size of an old-style startup packet, plus one
1983          * extra byte, and make sure all are zeroes.  This ensures we will have
1984          * null termination of all strings, in both fixed- and variable-length
1985          * packet layouts.
1986          */
1987         if (len <= (int32) sizeof(StartupPacket))
1988                 buf = palloc0(sizeof(StartupPacket) + 1);
1989         else
1990                 buf = palloc0(len + 1);
1991
1992         if (pq_getbytes(buf, len) == EOF)
1993         {
1994                 ereport(COMMERROR,
1995                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1996                                  errmsg("incomplete startup packet")));
1997                 return STATUS_ERROR;
1998         }
1999         pq_endmsgread();
2000
2001         /*
2002          * The first field is either a protocol version number or a special
2003          * request code.
2004          */
2005         port->proto = proto = ntohl(*((ProtocolVersion *) buf));
2006
2007         if (proto == CANCEL_REQUEST_CODE)
2008         {
2009                 processCancelRequest(port, buf);
2010                 /* Not really an error, but we don't want to proceed further */
2011                 return STATUS_ERROR;
2012         }
2013
2014         if (proto == NEGOTIATE_SSL_CODE && !SSLdone)
2015         {
2016                 char            SSLok;
2017
2018 #ifdef USE_SSL
2019                 /* No SSL when disabled or on Unix sockets */
2020                 if (!LoadedSSL || IS_AF_UNIX(port->laddr.addr.ss_family))
2021                         SSLok = 'N';
2022                 else
2023                         SSLok = 'S';            /* Support for SSL */
2024 #else
2025                 SSLok = 'N';                    /* No support for SSL */
2026 #endif
2027
2028 retry1:
2029                 if (send(port->sock, &SSLok, 1, 0) != 1)
2030                 {
2031                         if (errno == EINTR)
2032                                 goto retry1;    /* if interrupted, just retry */
2033                         ereport(COMMERROR,
2034                                         (errcode_for_socket_access(),
2035                                          errmsg("failed to send SSL negotiation response: %m")));
2036                         return STATUS_ERROR;    /* close the connection */
2037                 }
2038
2039 #ifdef USE_SSL
2040                 if (SSLok == 'S' && secure_open_server(port) == -1)
2041                         return STATUS_ERROR;
2042 #endif
2043                 /* regular startup packet, cancel, etc packet should follow... */
2044                 /* but not another SSL negotiation request */
2045                 return ProcessStartupPacket(port, true);
2046         }
2047
2048         /* Could add additional special packet types here */
2049
2050         /*
2051          * Set FrontendProtocol now so that ereport() knows what format to send if
2052          * we fail during startup.
2053          */
2054         FrontendProtocol = proto;
2055
2056         /* Check we can handle the protocol the frontend is using. */
2057
2058         if (PG_PROTOCOL_MAJOR(proto) < PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST) ||
2059                 PG_PROTOCOL_MAJOR(proto) > PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST) ||
2060                 (PG_PROTOCOL_MAJOR(proto) == PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST) &&
2061                  PG_PROTOCOL_MINOR(proto) > PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST)))
2062                 ereport(FATAL,
2063                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2064                                  errmsg("unsupported frontend protocol %u.%u: server supports %u.0 to %u.%u",
2065                                                 PG_PROTOCOL_MAJOR(proto), PG_PROTOCOL_MINOR(proto),
2066                                                 PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST),
2067                                                 PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST),
2068                                                 PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST))));
2069
2070         /*
2071          * Now fetch parameters out of startup packet and save them into the Port
2072          * structure.  All data structures attached to the Port struct must be
2073          * allocated in TopMemoryContext so that they will remain available in a
2074          * running backend (even after PostmasterContext is destroyed).  We need
2075          * not worry about leaking this storage on failure, since we aren't in the
2076          * postmaster process anymore.
2077          */
2078         oldcontext = MemoryContextSwitchTo(TopMemoryContext);
2079
2080         if (PG_PROTOCOL_MAJOR(proto) >= 3)
2081         {
2082                 int32           offset = sizeof(ProtocolVersion);
2083
2084                 /*
2085                  * Scan packet body for name/option pairs.  We can assume any string
2086                  * beginning within the packet body is null-terminated, thanks to
2087                  * zeroing extra byte above.
2088                  */
2089                 port->guc_options = NIL;
2090
2091                 while (offset < len)
2092                 {
2093                         char       *nameptr = ((char *) buf) + offset;
2094                         int32           valoffset;
2095                         char       *valptr;
2096
2097                         if (*nameptr == '\0')
2098                                 break;                  /* found packet terminator */
2099                         valoffset = offset + strlen(nameptr) + 1;
2100                         if (valoffset >= len)
2101                                 break;                  /* missing value, will complain below */
2102                         valptr = ((char *) buf) + valoffset;
2103
2104                         if (strcmp(nameptr, "database") == 0)
2105                                 port->database_name = pstrdup(valptr);
2106                         else if (strcmp(nameptr, "user") == 0)
2107                                 port->user_name = pstrdup(valptr);
2108                         else if (strcmp(nameptr, "options") == 0)
2109                                 port->cmdline_options = pstrdup(valptr);
2110                         else if (strcmp(nameptr, "replication") == 0)
2111                         {
2112                                 /*
2113                                  * Due to backward compatibility concerns the replication
2114                                  * parameter is a hybrid beast which allows the value to be
2115                                  * either boolean or the string 'database'. The latter
2116                                  * connects to a specific database which is e.g. required for
2117                                  * logical decoding while.
2118                                  */
2119                                 if (strcmp(valptr, "database") == 0)
2120                                 {
2121                                         am_walsender = true;
2122                                         am_db_walsender = true;
2123                                 }
2124                                 else if (!parse_bool(valptr, &am_walsender))
2125                                         ereport(FATAL,
2126                                                         (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
2127                                                          errmsg("invalid value for parameter \"%s\": \"%s\"",
2128                                                                         "replication",
2129                                                                         valptr),
2130                                                          errhint("Valid values are: \"false\", 0, \"true\", 1, \"database\".")));
2131                         }
2132                         else
2133                         {
2134                                 /* Assume it's a generic GUC option */
2135                                 port->guc_options = lappend(port->guc_options,
2136                                                                                         pstrdup(nameptr));
2137                                 port->guc_options = lappend(port->guc_options,
2138                                                                                         pstrdup(valptr));
2139                         }
2140                         offset = valoffset + strlen(valptr) + 1;
2141                 }
2142
2143                 /*
2144                  * If we didn't find a packet terminator exactly at the end of the
2145                  * given packet length, complain.
2146                  */
2147                 if (offset != len - 1)
2148                         ereport(FATAL,
2149                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
2150                                          errmsg("invalid startup packet layout: expected terminator as last byte")));
2151         }
2152         else
2153         {
2154                 /*
2155                  * Get the parameters from the old-style, fixed-width-fields startup
2156                  * packet as C strings.  The packet destination was cleared first so a
2157                  * short packet has zeros silently added.  We have to be prepared to
2158                  * truncate the pstrdup result for oversize fields, though.
2159                  */
2160                 StartupPacket *packet = (StartupPacket *) buf;
2161
2162                 port->database_name = pstrdup(packet->database);
2163                 if (strlen(port->database_name) > sizeof(packet->database))
2164                         port->database_name[sizeof(packet->database)] = '\0';
2165                 port->user_name = pstrdup(packet->user);
2166                 if (strlen(port->user_name) > sizeof(packet->user))
2167                         port->user_name[sizeof(packet->user)] = '\0';
2168                 port->cmdline_options = pstrdup(packet->options);
2169                 if (strlen(port->cmdline_options) > sizeof(packet->options))
2170                         port->cmdline_options[sizeof(packet->options)] = '\0';
2171                 port->guc_options = NIL;
2172         }
2173
2174         /* Check a user name was given. */
2175         if (port->user_name == NULL || port->user_name[0] == '\0')
2176                 ereport(FATAL,
2177                                 (errcode(ERRCODE_INVALID_AUTHORIZATION_SPECIFICATION),
2178                                  errmsg("no PostgreSQL user name specified in startup packet")));
2179
2180         /* The database defaults to the user name. */
2181         if (port->database_name == NULL || port->database_name[0] == '\0')
2182                 port->database_name = pstrdup(port->user_name);
2183
2184         if (Db_user_namespace)
2185         {
2186                 /*
2187                  * If user@, it is a global user, remove '@'. We only want to do this
2188                  * if there is an '@' at the end and no earlier in the user string or
2189                  * they may fake as a local user of another database attaching to this
2190                  * database.
2191                  */
2192                 if (strchr(port->user_name, '@') ==
2193                         port->user_name + strlen(port->user_name) - 1)
2194                         *strchr(port->user_name, '@') = '\0';
2195                 else
2196                 {
2197                         /* Append '@' and dbname */
2198                         port->user_name = psprintf("%s@%s", port->user_name, port->database_name);
2199                 }
2200         }
2201
2202         /*
2203          * Truncate given database and user names to length of a Postgres name.
2204          * This avoids lookup failures when overlength names are given.
2205          */
2206         if (strlen(port->database_name) >= NAMEDATALEN)
2207                 port->database_name[NAMEDATALEN - 1] = '\0';
2208         if (strlen(port->user_name) >= NAMEDATALEN)
2209                 port->user_name[NAMEDATALEN - 1] = '\0';
2210
2211         /*
2212          * Normal walsender backends, e.g. for streaming replication, are not
2213          * connected to a particular database. But walsenders used for logical
2214          * replication need to connect to a specific database. We allow streaming
2215          * replication commands to be issued even if connected to a database as it
2216          * can make sense to first make a basebackup and then stream changes
2217          * starting from that.
2218          */
2219         if (am_walsender && !am_db_walsender)
2220                 port->database_name[0] = '\0';
2221
2222         /*
2223          * Done putting stuff in TopMemoryContext.
2224          */
2225         MemoryContextSwitchTo(oldcontext);
2226
2227         /*
2228          * If we're going to reject the connection due to database state, say so
2229          * now instead of wasting cycles on an authentication exchange. (This also
2230          * allows a pg_ping utility to be written.)
2231          */
2232         switch (port->canAcceptConnections)
2233         {
2234                 case CAC_STARTUP:
2235                         ereport(FATAL,
2236                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2237                                          errmsg("the database system is starting up")));
2238                         break;
2239                 case CAC_SHUTDOWN:
2240                         ereport(FATAL,
2241                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2242                                          errmsg("the database system is shutting down")));
2243                         break;
2244                 case CAC_RECOVERY:
2245                         ereport(FATAL,
2246                                         (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2247                                          errmsg("the database system is in recovery mode")));
2248                         break;
2249                 case CAC_TOOMANY:
2250                         ereport(FATAL,
2251                                         (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
2252                                          errmsg("sorry, too many clients already")));
2253                         break;
2254                 case CAC_WAITBACKUP:
2255                         /* OK for now, will check in InitPostgres */
2256                         break;
2257                 case CAC_OK:
2258                         break;
2259         }
2260
2261         return STATUS_OK;
2262 }
2263
2264
2265 /*
2266  * The client has sent a cancel request packet, not a normal
2267  * start-a-new-connection packet.  Perform the necessary processing.
2268  * Nothing is sent back to the client.
2269  */
2270 static void
2271 processCancelRequest(Port *port, void *pkt)
2272 {
2273         CancelRequestPacket *canc = (CancelRequestPacket *) pkt;
2274         int                     backendPID;
2275         int32           cancelAuthCode;
2276         Backend    *bp;
2277
2278 #ifndef EXEC_BACKEND
2279         dlist_iter      iter;
2280 #else
2281         int                     i;
2282 #endif
2283
2284         backendPID = (int) ntohl(canc->backendPID);
2285         cancelAuthCode = (int32) ntohl(canc->cancelAuthCode);
2286
2287         /*
2288          * See if we have a matching backend.  In the EXEC_BACKEND case, we can no
2289          * longer access the postmaster's own backend list, and must rely on the
2290          * duplicate array in shared memory.
2291          */
2292 #ifndef EXEC_BACKEND
2293         dlist_foreach(iter, &BackendList)
2294         {
2295                 bp = dlist_container(Backend, elem, iter.cur);
2296 #else
2297         for (i = MaxLivePostmasterChildren() - 1; i >= 0; i--)
2298         {
2299                 bp = (Backend *) &ShmemBackendArray[i];
2300 #endif
2301                 if (bp->pid == backendPID)
2302                 {
2303                         if (bp->cancel_key == cancelAuthCode)
2304                         {
2305                                 /* Found a match; signal that backend to cancel current op */
2306                                 ereport(DEBUG2,
2307                                                 (errmsg_internal("processing cancel request: sending SIGINT to process %d",
2308                                                                                  backendPID)));
2309                                 signal_child(bp->pid, SIGINT);
2310                         }
2311                         else
2312                                 /* Right PID, wrong key: no way, Jose */
2313                                 ereport(LOG,
2314                                                 (errmsg("wrong key in cancel request for process %d",
2315                                                                 backendPID)));
2316                         return;
2317                 }
2318         }
2319
2320         /* No matching backend */
2321         ereport(LOG,
2322                         (errmsg("PID %d in cancel request did not match any process",
2323                                         backendPID)));
2324 }
2325
2326 /*
2327  * canAcceptConnections --- check to see if database state allows connections.
2328  */
2329 static CAC_state
2330 canAcceptConnections(void)
2331 {
2332         CAC_state       result = CAC_OK;
2333
2334         /*
2335          * Can't start backends when in startup/shutdown/inconsistent recovery
2336          * state.
2337          *
2338          * In state PM_WAIT_BACKUP only superusers can connect (this must be
2339          * allowed so that a superuser can end online backup mode); we return
2340          * CAC_WAITBACKUP code to indicate that this must be checked later. Note
2341          * that neither CAC_OK nor CAC_WAITBACKUP can safely be returned until we
2342          * have checked for too many children.
2343          */
2344         if (pmState != PM_RUN)
2345         {
2346                 if (pmState == PM_WAIT_BACKUP)
2347                         result = CAC_WAITBACKUP;        /* allow superusers only */
2348                 else if (Shutdown > NoShutdown)
2349                         return CAC_SHUTDOWN;    /* shutdown is pending */
2350                 else if (!FatalError &&
2351                                  (pmState == PM_STARTUP ||
2352                                   pmState == PM_RECOVERY))
2353                         return CAC_STARTUP; /* normal startup */
2354                 else if (!FatalError &&
2355                                  pmState == PM_HOT_STANDBY)
2356                         result = CAC_OK;        /* connection OK during hot standby */
2357                 else
2358                         return CAC_RECOVERY;    /* else must be crash recovery */
2359         }
2360
2361         /*
2362          * Don't start too many children.
2363          *
2364          * We allow more connections than we can have backends here because some
2365          * might still be authenticating; they might fail auth, or some existing
2366          * backend might exit before the auth cycle is completed. The exact
2367          * MaxBackends limit is enforced when a new backend tries to join the
2368          * shared-inval backend array.
2369          *
2370          * The limit here must match the sizes of the per-child-process arrays;
2371          * see comments for MaxLivePostmasterChildren().
2372          */
2373         if (CountChildren(BACKEND_TYPE_ALL) >= MaxLivePostmasterChildren())
2374                 result = CAC_TOOMANY;
2375
2376         return result;
2377 }
2378
2379
2380 /*
2381  * ConnCreate -- create a local connection data structure
2382  *
2383  * Returns NULL on failure, other than out-of-memory which is fatal.
2384  */
2385 static Port *
2386 ConnCreate(int serverFd)
2387 {
2388         Port       *port;
2389
2390         if (!(port = (Port *) calloc(1, sizeof(Port))))
2391         {
2392                 ereport(LOG,
2393                                 (errcode(ERRCODE_OUT_OF_MEMORY),
2394                                  errmsg("out of memory")));
2395                 ExitPostmaster(1);
2396         }
2397
2398         if (StreamConnection(serverFd, port) != STATUS_OK)
2399         {
2400                 if (port->sock != PGINVALID_SOCKET)
2401                         StreamClose(port->sock);
2402                 ConnFree(port);
2403                 return NULL;
2404         }
2405
2406         /*
2407          * Allocate GSSAPI specific state struct
2408          */
2409 #ifndef EXEC_BACKEND
2410 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
2411         port->gss = (pg_gssinfo *) calloc(1, sizeof(pg_gssinfo));
2412         if (!port->gss)
2413         {
2414                 ereport(LOG,
2415                                 (errcode(ERRCODE_OUT_OF_MEMORY),
2416                                  errmsg("out of memory")));
2417                 ExitPostmaster(1);
2418         }
2419 #endif
2420 #endif
2421
2422         return port;
2423 }
2424
2425
2426 /*
2427  * ConnFree -- free a local connection data structure
2428  */
2429 static void
2430 ConnFree(Port *conn)
2431 {
2432 #ifdef USE_SSL
2433         secure_close(conn);
2434 #endif
2435         if (conn->gss)
2436                 free(conn->gss);
2437         free(conn);
2438 }
2439
2440
2441 /*
2442  * ClosePostmasterPorts -- close all the postmaster's open sockets
2443  *
2444  * This is called during child process startup to release file descriptors
2445  * that are not needed by that child process.  The postmaster still has
2446  * them open, of course.
2447  *
2448  * Note: we pass am_syslogger as a boolean because we don't want to set
2449  * the global variable yet when this is called.
2450  */
2451 void
2452 ClosePostmasterPorts(bool am_syslogger)
2453 {
2454         int                     i;
2455
2456 #ifndef WIN32
2457
2458         /*
2459          * Close the write end of postmaster death watch pipe. It's important to
2460          * do this as early as possible, so that if postmaster dies, others won't
2461          * think that it's still running because we're holding the pipe open.
2462          */
2463         if (close(postmaster_alive_fds[POSTMASTER_FD_OWN]))
2464                 ereport(FATAL,
2465                                 (errcode_for_file_access(),
2466                                  errmsg_internal("could not close postmaster death monitoring pipe in child process: %m")));
2467         postmaster_alive_fds[POSTMASTER_FD_OWN] = -1;
2468 #endif
2469
2470         /* Close the listen sockets */
2471         for (i = 0; i < MAXLISTEN; i++)
2472         {
2473                 if (ListenSocket[i] != PGINVALID_SOCKET)
2474                 {
2475                         StreamClose(ListenSocket[i]);
2476                         ListenSocket[i] = PGINVALID_SOCKET;
2477                 }
2478         }
2479
2480         /* If using syslogger, close the read side of the pipe */
2481         if (!am_syslogger)
2482         {
2483 #ifndef WIN32
2484                 if (syslogPipe[0] >= 0)
2485                         close(syslogPipe[0]);
2486                 syslogPipe[0] = -1;
2487 #else
2488                 if (syslogPipe[0])
2489                         CloseHandle(syslogPipe[0]);
2490                 syslogPipe[0] = 0;
2491 #endif
2492         }
2493
2494 #ifdef USE_BONJOUR
2495         /* If using Bonjour, close the connection to the mDNS daemon */
2496         if (bonjour_sdref)
2497                 close(DNSServiceRefSockFD(bonjour_sdref));
2498 #endif
2499 }
2500
2501
2502 /*
2503  * reset_shared -- reset shared memory and semaphores
2504  */
2505 static void
2506 reset_shared(int port)
2507 {
2508         /*
2509          * Create or re-create shared memory and semaphores.
2510          *
2511          * Note: in each "cycle of life" we will normally assign the same IPC keys
2512          * (if using SysV shmem and/or semas), since the port number is used to
2513          * determine IPC keys.  This helps ensure that we will clean up dead IPC
2514          * objects if the postmaster crashes and is restarted.
2515          */
2516         CreateSharedMemoryAndSemaphores(false, port);
2517 }
2518
2519
2520 /*
2521  * SIGHUP -- reread config files, and tell children to do same
2522  */
2523 static void
2524 SIGHUP_handler(SIGNAL_ARGS)
2525 {
2526         int                     save_errno = errno;
2527
2528         PG_SETMASK(&BlockSig);
2529
2530         if (Shutdown <= SmartShutdown)
2531         {
2532                 ereport(LOG,
2533                                 (errmsg("received SIGHUP, reloading configuration files")));
2534                 ProcessConfigFile(PGC_SIGHUP);
2535                 SignalChildren(SIGHUP);
2536                 if (StartupPID != 0)
2537                         signal_child(StartupPID, SIGHUP);
2538                 if (BgWriterPID != 0)
2539                         signal_child(BgWriterPID, SIGHUP);
2540                 if (CheckpointerPID != 0)
2541                         signal_child(CheckpointerPID, SIGHUP);
2542                 if (WalWriterPID != 0)
2543                         signal_child(WalWriterPID, SIGHUP);
2544                 if (WalReceiverPID != 0)
2545                         signal_child(WalReceiverPID, SIGHUP);
2546                 if (AutoVacPID != 0)
2547                         signal_child(AutoVacPID, SIGHUP);
2548                 if (PgArchPID != 0)
2549                         signal_child(PgArchPID, SIGHUP);
2550                 if (SysLoggerPID != 0)
2551                         signal_child(SysLoggerPID, SIGHUP);
2552                 if (PgStatPID != 0)
2553                         signal_child(PgStatPID, SIGHUP);
2554
2555                 /* Reload authentication config files too */
2556                 if (!load_hba())
2557                         ereport(LOG,
2558                                         (errmsg("pg_hba.conf was not reloaded")));
2559
2560                 if (!load_ident())
2561                         ereport(LOG,
2562                                         (errmsg("pg_ident.conf was not reloaded")));
2563
2564 #ifdef USE_SSL
2565                 /* Reload SSL configuration as well */
2566                 if (EnableSSL)
2567                 {
2568                         if (secure_initialize(false) == 0)
2569                                 LoadedSSL = true;
2570                         else
2571                                 ereport(LOG,
2572                                                 (errmsg("SSL configuration was not reloaded")));
2573                 }
2574                 else
2575                 {
2576                         secure_destroy();
2577                         LoadedSSL = false;
2578                 }
2579 #endif
2580
2581 #ifdef EXEC_BACKEND
2582                 /* Update the starting-point file for future children */
2583                 write_nondefault_variables(PGC_SIGHUP);
2584 #endif
2585         }
2586
2587         PG_SETMASK(&UnBlockSig);
2588
2589         errno = save_errno;
2590 }
2591
2592
2593 /*
2594  * pmdie -- signal handler for processing various postmaster signals.
2595  */
2596 static void
2597 pmdie(SIGNAL_ARGS)
2598 {
2599         int                     save_errno = errno;
2600
2601         PG_SETMASK(&BlockSig);
2602
2603         ereport(DEBUG2,
2604                         (errmsg_internal("postmaster received signal %d",
2605                                                          postgres_signal_arg)));
2606
2607         switch (postgres_signal_arg)
2608         {
2609                 case SIGTERM:
2610
2611                         /*
2612                          * Smart Shutdown:
2613                          *
2614                          * Wait for children to end their work, then shut down.
2615                          */
2616                         if (Shutdown >= SmartShutdown)
2617                                 break;
2618                         Shutdown = SmartShutdown;
2619                         ereport(LOG,
2620                                         (errmsg("received smart shutdown request")));
2621
2622                         /* Report status */
2623                         AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2624 #ifdef USE_SYSTEMD
2625                         sd_notify(0, "STOPPING=1");
2626 #endif
2627
2628                         if (pmState == PM_RUN || pmState == PM_RECOVERY ||
2629                                 pmState == PM_HOT_STANDBY || pmState == PM_STARTUP)
2630                         {
2631                                 /* autovac workers are told to shut down immediately */
2632                                 /* and bgworkers too; does this need tweaking? */
2633                                 SignalSomeChildren(SIGTERM,
2634                                                                    BACKEND_TYPE_AUTOVAC | BACKEND_TYPE_BGWORKER);
2635                                 /* and the autovac launcher too */
2636                                 if (AutoVacPID != 0)
2637                                         signal_child(AutoVacPID, SIGTERM);
2638                                 /* and the bgwriter too */
2639                                 if (BgWriterPID != 0)
2640                                         signal_child(BgWriterPID, SIGTERM);
2641                                 /* and the walwriter too */
2642                                 if (WalWriterPID != 0)
2643                                         signal_child(WalWriterPID, SIGTERM);
2644
2645                                 /*
2646                                  * If we're in recovery, we can't kill the startup process
2647                                  * right away, because at present doing so does not release
2648                                  * its locks.  We might want to change this in a future
2649                                  * release.  For the time being, the PM_WAIT_READONLY state
2650                                  * indicates that we're waiting for the regular (read only)
2651                                  * backends to die off; once they do, we'll kill the startup
2652                                  * and walreceiver processes.
2653                                  */
2654                                 pmState = (pmState == PM_RUN) ?
2655                                         PM_WAIT_BACKUP : PM_WAIT_READONLY;
2656                         }
2657
2658                         /*
2659                          * Now wait for online backup mode to end and backends to exit. If
2660                          * that is already the case, PostmasterStateMachine will take the
2661                          * next step.
2662                          */
2663                         PostmasterStateMachine();
2664                         break;
2665
2666                 case SIGINT:
2667
2668                         /*
2669                          * Fast Shutdown:
2670                          *
2671                          * Abort all children with SIGTERM (rollback active transactions
2672                          * and exit) and shut down when they are gone.
2673                          */
2674                         if (Shutdown >= FastShutdown)
2675                                 break;
2676                         Shutdown = FastShutdown;
2677                         ereport(LOG,
2678                                         (errmsg("received fast shutdown request")));
2679
2680                         /* Report status */
2681                         AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2682 #ifdef USE_SYSTEMD
2683                         sd_notify(0, "STOPPING=1");
2684 #endif
2685
2686                         if (StartupPID != 0)
2687                                 signal_child(StartupPID, SIGTERM);
2688                         if (BgWriterPID != 0)
2689                                 signal_child(BgWriterPID, SIGTERM);
2690                         if (WalReceiverPID != 0)
2691                                 signal_child(WalReceiverPID, SIGTERM);
2692                         if (pmState == PM_RECOVERY)
2693                         {
2694                                 SignalSomeChildren(SIGTERM, BACKEND_TYPE_BGWORKER);
2695
2696                                 /*
2697                                  * Only startup, bgwriter, walreceiver, possibly bgworkers,
2698                                  * and/or checkpointer should be active in this state; we just
2699                                  * signaled the first four, and we don't want to kill
2700                                  * checkpointer yet.
2701                                  */
2702                                 pmState = PM_WAIT_BACKENDS;
2703                         }
2704                         else if (pmState == PM_RUN ||
2705                                          pmState == PM_WAIT_BACKUP ||
2706                                          pmState == PM_WAIT_READONLY ||
2707                                          pmState == PM_WAIT_BACKENDS ||
2708                                          pmState == PM_HOT_STANDBY)
2709                         {
2710                                 ereport(LOG,
2711                                                 (errmsg("aborting any active transactions")));
2712                                 /* shut down all backends and workers */
2713                                 SignalSomeChildren(SIGTERM,
2714                                                                    BACKEND_TYPE_NORMAL | BACKEND_TYPE_AUTOVAC |
2715                                                                    BACKEND_TYPE_BGWORKER);
2716                                 /* and the autovac launcher too */
2717                                 if (AutoVacPID != 0)
2718                                         signal_child(AutoVacPID, SIGTERM);
2719                                 /* and the walwriter too */
2720                                 if (WalWriterPID != 0)
2721                                         signal_child(WalWriterPID, SIGTERM);
2722                                 pmState = PM_WAIT_BACKENDS;
2723                         }
2724
2725                         /*
2726                          * Now wait for backends to exit.  If there are none,
2727                          * PostmasterStateMachine will take the next step.
2728                          */
2729                         PostmasterStateMachine();
2730                         break;
2731
2732                 case SIGQUIT:
2733
2734                         /*
2735                          * Immediate Shutdown:
2736                          *
2737                          * abort all children with SIGQUIT, wait for them to exit,
2738                          * terminate remaining ones with SIGKILL, then exit without
2739                          * attempt to properly shut down the data base system.
2740                          */
2741                         if (Shutdown >= ImmediateShutdown)
2742                                 break;
2743                         Shutdown = ImmediateShutdown;
2744                         ereport(LOG,
2745                                         (errmsg("received immediate shutdown request")));
2746
2747                         /* Report status */
2748                         AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2749 #ifdef USE_SYSTEMD
2750                         sd_notify(0, "STOPPING=1");
2751 #endif
2752
2753                         TerminateChildren(SIGQUIT);
2754                         pmState = PM_WAIT_BACKENDS;
2755
2756                         /* set stopwatch for them to die */
2757                         AbortStartTime = time(NULL);
2758
2759                         /*
2760                          * Now wait for backends to exit.  If there are none,
2761                          * PostmasterStateMachine will take the next step.
2762                          */
2763                         PostmasterStateMachine();
2764                         break;
2765         }
2766
2767         PG_SETMASK(&UnBlockSig);
2768
2769         errno = save_errno;
2770 }
2771
2772 /*
2773  * Reaper -- signal handler to cleanup after a child process dies.
2774  */
2775 static void
2776 reaper(SIGNAL_ARGS)
2777 {
2778         int                     save_errno = errno;
2779         int                     pid;                    /* process id of dead child process */
2780         int                     exitstatus;             /* its exit status */
2781
2782         PG_SETMASK(&BlockSig);
2783
2784         ereport(DEBUG4,
2785                         (errmsg_internal("reaping dead processes")));
2786
2787         while ((pid = waitpid(-1, &exitstatus, WNOHANG)) > 0)
2788         {
2789                 /*
2790                  * Check if this child was a startup process.
2791                  */
2792                 if (pid == StartupPID)
2793                 {
2794                         StartupPID = 0;
2795
2796                         /*
2797                          * Startup process exited in response to a shutdown request (or it
2798                          * completed normally regardless of the shutdown request).
2799                          */
2800                         if (Shutdown > NoShutdown &&
2801                                 (EXIT_STATUS_0(exitstatus) || EXIT_STATUS_1(exitstatus)))
2802                         {
2803                                 StartupStatus = STARTUP_NOT_RUNNING;
2804                                 pmState = PM_WAIT_BACKENDS;
2805                                 /* PostmasterStateMachine logic does the rest */
2806                                 continue;
2807                         }
2808
2809                         if (EXIT_STATUS_3(exitstatus))
2810                         {
2811                                 ereport(LOG,
2812                                                 (errmsg("shutdown at recovery target")));
2813                                 StartupStatus = STARTUP_NOT_RUNNING;
2814                                 Shutdown = SmartShutdown;
2815                                 TerminateChildren(SIGTERM);
2816                                 pmState = PM_WAIT_BACKENDS;
2817                                 /* PostmasterStateMachine logic does the rest */
2818                                 continue;
2819                         }
2820
2821                         /*
2822                          * Unexpected exit of startup process (including FATAL exit)
2823                          * during PM_STARTUP is treated as catastrophic. There are no
2824                          * other processes running yet, so we can just exit.
2825                          */
2826                         if (pmState == PM_STARTUP && !EXIT_STATUS_0(exitstatus))
2827                         {
2828                                 LogChildExit(LOG, _("startup process"),
2829                                                          pid, exitstatus);
2830                                 ereport(LOG,
2831                                                 (errmsg("aborting startup due to startup process failure")));
2832                                 ExitPostmaster(1);
2833                         }
2834
2835                         /*
2836                          * After PM_STARTUP, any unexpected exit (including FATAL exit) of
2837                          * the startup process is catastrophic, so kill other children,
2838                          * and set StartupStatus so we don't try to reinitialize after
2839                          * they're gone.  Exception: if StartupStatus is STARTUP_SIGNALED,
2840                          * then we previously sent the startup process a SIGQUIT; so
2841                          * that's probably the reason it died, and we do want to try to
2842                          * restart in that case.
2843                          */
2844                         if (!EXIT_STATUS_0(exitstatus))
2845                         {
2846                                 if (StartupStatus == STARTUP_SIGNALED)
2847                                         StartupStatus = STARTUP_NOT_RUNNING;
2848                                 else
2849                                         StartupStatus = STARTUP_CRASHED;
2850                                 HandleChildCrash(pid, exitstatus,
2851                                                                  _("startup process"));
2852                                 continue;
2853                         }
2854
2855                         /*
2856                          * Startup succeeded, commence normal operations
2857                          */
2858                         StartupStatus = STARTUP_NOT_RUNNING;
2859                         FatalError = false;
2860                         Assert(AbortStartTime == 0);
2861                         ReachedNormalRunning = true;
2862                         pmState = PM_RUN;
2863
2864                         /*
2865                          * Crank up the background tasks, if we didn't do that already
2866                          * when we entered consistent recovery state.  It doesn't matter
2867                          * if this fails, we'll just try again later.
2868                          */
2869                         if (CheckpointerPID == 0)
2870                                 CheckpointerPID = StartCheckpointer();
2871                         if (BgWriterPID == 0)
2872                                 BgWriterPID = StartBackgroundWriter();
2873                         if (WalWriterPID == 0)
2874                                 WalWriterPID = StartWalWriter();
2875
2876                         /*
2877                          * Likewise, start other special children as needed.  In a restart
2878                          * situation, some of them may be alive already.
2879                          */
2880                         if (!IsBinaryUpgrade && AutoVacuumingActive() && AutoVacPID == 0)
2881                                 AutoVacPID = StartAutoVacLauncher();
2882                         if (PgArchStartupAllowed() && PgArchPID == 0)
2883                                 PgArchPID = pgarch_start();
2884                         if (PgStatPID == 0)
2885                                 PgStatPID = pgstat_start();
2886
2887                         /* workers may be scheduled to start now */
2888                         maybe_start_bgworkers();
2889
2890                         /* at this point we are really open for business */
2891                         ereport(LOG,
2892                                         (errmsg("database system is ready to accept connections")));
2893
2894                         /* Report status */
2895                         AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
2896 #ifdef USE_SYSTEMD
2897                         sd_notify(0, "READY=1");
2898 #endif
2899
2900                         continue;
2901                 }
2902
2903                 /*
2904                  * Was it the bgwriter?  Normal exit can be ignored; we'll start a new
2905                  * one at the next iteration of the postmaster's main loop, if
2906                  * necessary.  Any other exit condition is treated as a crash.
2907                  */
2908                 if (pid == BgWriterPID)
2909                 {
2910                         BgWriterPID = 0;
2911                         if (!EXIT_STATUS_0(exitstatus))
2912                                 HandleChildCrash(pid, exitstatus,
2913                                                                  _("background writer process"));
2914                         continue;
2915                 }
2916
2917                 /*
2918                  * Was it the checkpointer?
2919                  */
2920                 if (pid == CheckpointerPID)
2921                 {
2922                         CheckpointerPID = 0;
2923                         if (EXIT_STATUS_0(exitstatus) && pmState == PM_SHUTDOWN)
2924                         {
2925                                 /*
2926                                  * OK, we saw normal exit of the checkpointer after it's been
2927                                  * told to shut down.  We expect that it wrote a shutdown
2928                                  * checkpoint.  (If for some reason it didn't, recovery will
2929                                  * occur on next postmaster start.)
2930                                  *
2931                                  * At this point we should have no normal backend children
2932                                  * left (else we'd not be in PM_SHUTDOWN state) but we might
2933                                  * have dead_end children to wait for.
2934                                  *
2935                                  * If we have an archiver subprocess, tell it to do a last
2936                                  * archive cycle and quit. Likewise, if we have walsender
2937                                  * processes, tell them to send any remaining WAL and quit.
2938                                  */
2939                                 Assert(Shutdown > NoShutdown);
2940
2941                                 /* Waken archiver for the last time */
2942                                 if (PgArchPID != 0)
2943                                         signal_child(PgArchPID, SIGUSR2);
2944
2945                                 /*
2946                                  * Waken walsenders for the last time. No regular backends
2947                                  * should be around anymore.
2948                                  */
2949                                 SignalChildren(SIGUSR2);
2950
2951                                 pmState = PM_SHUTDOWN_2;
2952
2953                                 /*
2954                                  * We can also shut down the stats collector now; there's
2955                                  * nothing left for it to do.
2956                                  */
2957                                 if (PgStatPID != 0)
2958                                         signal_child(PgStatPID, SIGQUIT);
2959                         }
2960                         else
2961                         {
2962                                 /*
2963                                  * Any unexpected exit of the checkpointer (including FATAL
2964                                  * exit) is treated as a crash.
2965                                  */
2966                                 HandleChildCrash(pid, exitstatus,
2967                                                                  _("checkpointer process"));
2968                         }
2969
2970                         continue;
2971                 }
2972
2973                 /*
2974                  * Was it the wal writer?  Normal exit can be ignored; we'll start a
2975                  * new one at the next iteration of the postmaster's main loop, if
2976                  * necessary.  Any other exit condition is treated as a crash.
2977                  */
2978                 if (pid == WalWriterPID)
2979                 {
2980                         WalWriterPID = 0;
2981                         if (!EXIT_STATUS_0(exitstatus))
2982                                 HandleChildCrash(pid, exitstatus,
2983                                                                  _("WAL writer process"));
2984                         continue;
2985                 }
2986
2987                 /*
2988                  * Was it the wal receiver?  If exit status is zero (normal) or one
2989                  * (FATAL exit), we assume everything is all right just like normal
2990                  * backends.  (If we need a new wal receiver, we'll start one at the
2991                  * next iteration of the postmaster's main loop.)
2992                  */
2993                 if (pid == WalReceiverPID)
2994                 {
2995                         WalReceiverPID = 0;
2996                         if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
2997                                 HandleChildCrash(pid, exitstatus,
2998                                                                  _("WAL receiver process"));
2999                         continue;
3000                 }
3001
3002                 /*
3003                  * Was it the autovacuum launcher?      Normal exit can be ignored; we'll
3004                  * start a new one at the next iteration of the postmaster's main
3005                  * loop, if necessary.  Any other exit condition is treated as a
3006                  * crash.
3007                  */
3008                 if (pid == AutoVacPID)
3009                 {
3010                         AutoVacPID = 0;
3011                         if (!EXIT_STATUS_0(exitstatus))
3012                                 HandleChildCrash(pid, exitstatus,
3013                                                                  _("autovacuum launcher process"));
3014                         continue;
3015                 }
3016
3017                 /*
3018                  * Was it the archiver?  If so, just try to start a new one; no need
3019                  * to force reset of the rest of the system.  (If fail, we'll try
3020                  * again in future cycles of the main loop.).  Unless we were waiting
3021                  * for it to shut down; don't restart it in that case, and
3022                  * PostmasterStateMachine() will advance to the next shutdown step.
3023                  */
3024                 if (pid == PgArchPID)
3025                 {
3026                         PgArchPID = 0;
3027                         if (!EXIT_STATUS_0(exitstatus))
3028                                 LogChildExit(LOG, _("archiver process"),
3029                                                          pid, exitstatus);
3030                         if (PgArchStartupAllowed())
3031                                 PgArchPID = pgarch_start();
3032                         continue;
3033                 }
3034
3035                 /*
3036                  * Was it the statistics collector?  If so, just try to start a new
3037                  * one; no need to force reset of the rest of the system.  (If fail,
3038                  * we'll try again in future cycles of the main loop.)
3039                  */
3040                 if (pid == PgStatPID)
3041                 {
3042                         PgStatPID = 0;
3043                         if (!EXIT_STATUS_0(exitstatus))
3044                                 LogChildExit(LOG, _("statistics collector process"),
3045                                                          pid, exitstatus);
3046                         if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3047                                 PgStatPID = pgstat_start();
3048                         continue;
3049                 }
3050
3051                 /* Was it the system logger?  If so, try to start a new one */
3052                 if (pid == SysLoggerPID)
3053                 {
3054                         SysLoggerPID = 0;
3055                         /* for safety's sake, launch new logger *first* */
3056                         SysLoggerPID = SysLogger_Start();
3057                         if (!EXIT_STATUS_0(exitstatus))
3058                                 LogChildExit(LOG, _("system logger process"),
3059                                                          pid, exitstatus);
3060                         continue;
3061                 }
3062
3063                 /* Was it one of our background workers? */
3064                 if (CleanupBackgroundWorker(pid, exitstatus))
3065                 {
3066                         /* have it be restarted */
3067                         HaveCrashedWorker = true;
3068                         continue;
3069                 }
3070
3071                 /*
3072                  * Else do standard backend child cleanup.
3073                  */
3074                 CleanupBackend(pid, exitstatus);
3075         }                                                       /* loop over pending child-death reports */
3076
3077         /*
3078          * After cleaning out the SIGCHLD queue, see if we have any state changes
3079          * or actions to make.
3080          */
3081         PostmasterStateMachine();
3082
3083         /* Done with signal handler */
3084         PG_SETMASK(&UnBlockSig);
3085
3086         errno = save_errno;
3087 }
3088
3089 /*
3090  * Scan the bgworkers list and see if the given PID (which has just stopped
3091  * or crashed) is in it.  Handle its shutdown if so, and return true.  If not a
3092  * bgworker, return false.
3093  *
3094  * This is heavily based on CleanupBackend.  One important difference is that
3095  * we don't know yet that the dying process is a bgworker, so we must be silent
3096  * until we're sure it is.
3097  */
3098 static bool
3099 CleanupBackgroundWorker(int pid,
3100                                                 int exitstatus) /* child's exit status */
3101 {
3102         char            namebuf[MAXPGPATH];
3103         slist_mutable_iter iter;
3104
3105         slist_foreach_modify(iter, &BackgroundWorkerList)
3106         {
3107                 RegisteredBgWorker *rw;
3108
3109                 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
3110
3111                 if (rw->rw_pid != pid)
3112                         continue;
3113
3114 #ifdef WIN32
3115                 /* see CleanupBackend */
3116                 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3117                         exitstatus = 0;
3118 #endif
3119
3120                 snprintf(namebuf, MAXPGPATH, "%s: %s", _("worker process"),
3121                                  rw->rw_worker.bgw_name);
3122
3123                 if (!EXIT_STATUS_0(exitstatus))
3124                 {
3125                         /* Record timestamp, so we know when to restart the worker. */
3126                         rw->rw_crashed_at = GetCurrentTimestamp();
3127                 }
3128                 else
3129                 {
3130                         /* Zero exit status means terminate */
3131                         rw->rw_crashed_at = 0;
3132                         rw->rw_terminate = true;
3133                 }
3134
3135                 /*
3136                  * Additionally, for shared-memory-connected workers, just like a
3137                  * backend, any exit status other than 0 or 1 is considered a crash
3138                  * and causes a system-wide restart.
3139                  */
3140                 if ((rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3141                 {
3142                         if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3143                         {
3144                                 HandleChildCrash(pid, exitstatus, namebuf);
3145                                 return true;
3146                         }
3147                 }
3148
3149                 /*
3150                  * We must release the postmaster child slot whether this worker is
3151                  * connected to shared memory or not, but we only treat it as a crash
3152                  * if it is in fact connected.
3153                  */
3154                 if (!ReleasePostmasterChildSlot(rw->rw_child_slot) &&
3155                         (rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3156                 {
3157                         HandleChildCrash(pid, exitstatus, namebuf);
3158                         return true;
3159                 }
3160
3161                 /* Get it out of the BackendList and clear out remaining data */
3162                 dlist_delete(&rw->rw_backend->elem);
3163 #ifdef EXEC_BACKEND
3164                 ShmemBackendArrayRemove(rw->rw_backend);
3165 #endif
3166
3167                 /*
3168                  * It's possible that this background worker started some OTHER
3169                  * background worker and asked to be notified when that worker started
3170                  * or stopped.  If so, cancel any notifications destined for the
3171                  * now-dead backend.
3172                  */
3173                 if (rw->rw_backend->bgworker_notify)
3174                         BackgroundWorkerStopNotifications(rw->rw_pid);
3175                 free(rw->rw_backend);
3176                 rw->rw_backend = NULL;
3177                 rw->rw_pid = 0;
3178                 rw->rw_child_slot = 0;
3179                 ReportBackgroundWorkerExit(&iter);      /* report child death */
3180
3181                 LogChildExit(EXIT_STATUS_0(exitstatus) ? DEBUG1 : LOG,
3182                                          namebuf, pid, exitstatus);
3183
3184                 return true;
3185         }
3186
3187         return false;
3188 }
3189
3190 /*
3191  * CleanupBackend -- cleanup after terminated backend.
3192  *
3193  * Remove all local state associated with backend.
3194  *
3195  * If you change this, see also CleanupBackgroundWorker.
3196  */
3197 static void
3198 CleanupBackend(int pid,
3199                            int exitstatus)      /* child's exit status. */
3200 {
3201         dlist_mutable_iter iter;
3202
3203         LogChildExit(DEBUG2, _("server process"), pid, exitstatus);
3204
3205         /*
3206          * If a backend dies in an ugly way then we must signal all other backends
3207          * to quickdie.  If exit status is zero (normal) or one (FATAL exit), we
3208          * assume everything is all right and proceed to remove the backend from
3209          * the active backend list.
3210          */
3211
3212 #ifdef WIN32
3213
3214         /*
3215          * On win32, also treat ERROR_WAIT_NO_CHILDREN (128) as nonfatal case,
3216          * since that sometimes happens under load when the process fails to start
3217          * properly (long before it starts using shared memory). Microsoft reports
3218          * it is related to mutex failure:
3219          * http://archives.postgresql.org/pgsql-hackers/2010-09/msg00790.php
3220          */
3221         if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3222         {
3223                 LogChildExit(LOG, _("server process"), pid, exitstatus);
3224                 exitstatus = 0;
3225         }
3226 #endif
3227
3228         if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3229         {
3230                 HandleChildCrash(pid, exitstatus, _("server process"));
3231                 return;
3232         }
3233
3234         dlist_foreach_modify(iter, &BackendList)
3235         {
3236                 Backend    *bp = dlist_container(Backend, elem, iter.cur);
3237
3238                 if (bp->pid == pid)
3239                 {
3240                         if (!bp->dead_end)
3241                         {
3242                                 if (!ReleasePostmasterChildSlot(bp->child_slot))
3243                                 {
3244                                         /*
3245                                          * Uh-oh, the child failed to clean itself up.  Treat as a
3246                                          * crash after all.
3247                                          */
3248                                         HandleChildCrash(pid, exitstatus, _("server process"));
3249                                         return;
3250                                 }
3251 #ifdef EXEC_BACKEND
3252                                 ShmemBackendArrayRemove(bp);
3253 #endif
3254                         }
3255                         if (bp->bgworker_notify)
3256                         {
3257                                 /*
3258                                  * This backend may have been slated to receive SIGUSR1 when
3259                                  * some background worker started or stopped.  Cancel those
3260                                  * notifications, as we don't want to signal PIDs that are not
3261                                  * PostgreSQL backends.  This gets skipped in the (probably
3262                                  * very common) case where the backend has never requested any
3263                                  * such notifications.
3264                                  */
3265                                 BackgroundWorkerStopNotifications(bp->pid);
3266                         }
3267                         dlist_delete(iter.cur);
3268                         free(bp);
3269                         break;
3270                 }
3271         }
3272 }
3273
3274 /*
3275  * HandleChildCrash -- cleanup after failed backend, bgwriter, checkpointer,
3276  * walwriter, autovacuum, or background worker.
3277  *
3278  * The objectives here are to clean up our local state about the child
3279  * process, and to signal all other remaining children to quickdie.
3280  */
3281 static void
3282 HandleChildCrash(int pid, int exitstatus, const char *procname)
3283 {
3284         dlist_mutable_iter iter;
3285         slist_iter      siter;
3286         Backend    *bp;
3287         bool            take_action;
3288
3289         /*
3290          * We only log messages and send signals if this is the first process
3291          * crash and we're not doing an immediate shutdown; otherwise, we're only
3292          * here to update postmaster's idea of live processes.  If we have already
3293          * signalled children, nonzero exit status is to be expected, so don't
3294          * clutter log.
3295          */
3296         take_action = !FatalError && Shutdown != ImmediateShutdown;
3297
3298         if (take_action)
3299         {
3300                 LogChildExit(LOG, procname, pid, exitstatus);
3301                 ereport(LOG,
3302                                 (errmsg("terminating any other active server processes")));
3303         }
3304
3305         /* Process background workers. */
3306         slist_foreach(siter, &BackgroundWorkerList)
3307         {
3308                 RegisteredBgWorker *rw;
3309
3310                 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
3311                 if (rw->rw_pid == 0)
3312                         continue;                       /* not running */
3313                 if (rw->rw_pid == pid)
3314                 {
3315                         /*
3316                          * Found entry for freshly-dead worker, so remove it.
3317                          */
3318                         (void) ReleasePostmasterChildSlot(rw->rw_child_slot);
3319                         dlist_delete(&rw->rw_backend->elem);
3320 #ifdef EXEC_BACKEND
3321                         ShmemBackendArrayRemove(rw->rw_backend);
3322 #endif
3323                         free(rw->rw_backend);
3324                         rw->rw_backend = NULL;
3325                         rw->rw_pid = 0;
3326                         rw->rw_child_slot = 0;
3327                         /* don't reset crashed_at */
3328                         /* don't report child stop, either */
3329                         /* Keep looping so we can signal remaining workers */
3330                 }
3331                 else
3332                 {
3333                         /*
3334                          * This worker is still alive.  Unless we did so already, tell it
3335                          * to commit hara-kiri.
3336                          *
3337                          * SIGQUIT is the special signal that says exit without proc_exit
3338                          * and let the user know what's going on. But if SendStop is set
3339                          * (-s on command line), then we send SIGSTOP instead, so that we
3340                          * can get core dumps from all backends by hand.
3341                          */
3342                         if (take_action)
3343                         {
3344                                 ereport(DEBUG2,
3345                                                 (errmsg_internal("sending %s to process %d",
3346                                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
3347                                                                                  (int) rw->rw_pid)));
3348                                 signal_child(rw->rw_pid, (SendStop ? SIGSTOP : SIGQUIT));
3349                         }
3350                 }
3351         }
3352
3353         /* Process regular backends */
3354         dlist_foreach_modify(iter, &BackendList)
3355         {
3356                 bp = dlist_container(Backend, elem, iter.cur);
3357
3358                 if (bp->pid == pid)
3359                 {
3360                         /*
3361                          * Found entry for freshly-dead backend, so remove it.
3362                          */
3363                         if (!bp->dead_end)
3364                         {
3365                                 (void) ReleasePostmasterChildSlot(bp->child_slot);
3366 #ifdef EXEC_BACKEND
3367                                 ShmemBackendArrayRemove(bp);
3368 #endif
3369                         }
3370                         dlist_delete(iter.cur);
3371                         free(bp);
3372                         /* Keep looping so we can signal remaining backends */
3373                 }
3374                 else
3375                 {
3376                         /*
3377                          * This backend is still alive.  Unless we did so already, tell it
3378                          * to commit hara-kiri.
3379                          *
3380                          * SIGQUIT is the special signal that says exit without proc_exit
3381                          * and let the user know what's going on. But if SendStop is set
3382                          * (-s on command line), then we send SIGSTOP instead, so that we
3383                          * can get core dumps from all backends by hand.
3384                          *
3385                          * We could exclude dead_end children here, but at least in the
3386                          * SIGSTOP case it seems better to include them.
3387                          *
3388                          * Background workers were already processed above; ignore them
3389                          * here.
3390                          */
3391                         if (bp->bkend_type == BACKEND_TYPE_BGWORKER)
3392                                 continue;
3393
3394                         if (take_action)
3395                         {
3396                                 ereport(DEBUG2,
3397                                                 (errmsg_internal("sending %s to process %d",
3398                                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
3399                                                                                  (int) bp->pid)));
3400                                 signal_child(bp->pid, (SendStop ? SIGSTOP : SIGQUIT));
3401                         }
3402                 }
3403         }
3404
3405         /* Take care of the startup process too */
3406         if (pid == StartupPID)
3407         {
3408                 StartupPID = 0;
3409                 StartupStatus = STARTUP_CRASHED;
3410         }
3411         else if (StartupPID != 0 && take_action)
3412         {
3413                 ereport(DEBUG2,
3414                                 (errmsg_internal("sending %s to process %d",
3415                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
3416                                                                  (int) StartupPID)));
3417                 signal_child(StartupPID, (SendStop ? SIGSTOP : SIGQUIT));
3418                 StartupStatus = STARTUP_SIGNALED;
3419         }
3420
3421         /* Take care of the bgwriter too */
3422         if (pid == BgWriterPID)
3423                 BgWriterPID = 0;
3424         else if (BgWriterPID != 0 && take_action)
3425         {
3426                 ereport(DEBUG2,
3427                                 (errmsg_internal("sending %s to process %d",
3428                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
3429                                                                  (int) BgWriterPID)));
3430                 signal_child(BgWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3431         }
3432
3433         /* Take care of the checkpointer too */
3434         if (pid == CheckpointerPID)
3435                 CheckpointerPID = 0;
3436         else if (CheckpointerPID != 0 && take_action)
3437         {
3438                 ereport(DEBUG2,
3439                                 (errmsg_internal("sending %s to process %d",
3440                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
3441                                                                  (int) CheckpointerPID)));
3442                 signal_child(CheckpointerPID, (SendStop ? SIGSTOP : SIGQUIT));
3443         }
3444
3445         /* Take care of the walwriter too */
3446         if (pid == WalWriterPID)
3447                 WalWriterPID = 0;
3448         else if (WalWriterPID != 0 && take_action)
3449         {
3450                 ereport(DEBUG2,
3451                                 (errmsg_internal("sending %s to process %d",
3452                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
3453                                                                  (int) WalWriterPID)));
3454                 signal_child(WalWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3455         }
3456
3457         /* Take care of the walreceiver too */
3458         if (pid == WalReceiverPID)
3459                 WalReceiverPID = 0;
3460         else if (WalReceiverPID != 0 && take_action)
3461         {
3462                 ereport(DEBUG2,
3463                                 (errmsg_internal("sending %s to process %d",
3464                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
3465                                                                  (int) WalReceiverPID)));
3466                 signal_child(WalReceiverPID, (SendStop ? SIGSTOP : SIGQUIT));
3467         }
3468
3469         /* Take care of the autovacuum launcher too */
3470         if (pid == AutoVacPID)
3471                 AutoVacPID = 0;
3472         else if (AutoVacPID != 0 && take_action)
3473         {
3474                 ereport(DEBUG2,
3475                                 (errmsg_internal("sending %s to process %d",
3476                                                                  (SendStop ? "SIGSTOP" : "SIGQUIT"),
3477                                                                  (int) AutoVacPID)));
3478                 signal_child(AutoVacPID, (SendStop ? SIGSTOP : SIGQUIT));
3479         }
3480
3481         /*
3482          * Force a power-cycle of the pgarch process too.  (This isn't absolutely
3483          * necessary, but it seems like a good idea for robustness, and it
3484          * simplifies the state-machine logic in the case where a shutdown request
3485          * arrives during crash processing.)
3486          */
3487         if (PgArchPID != 0 && take_action)
3488         {
3489                 ereport(DEBUG2,
3490                                 (errmsg_internal("sending %s to process %d",
3491                                                                  "SIGQUIT",
3492                                                                  (int) PgArchPID)));
3493                 signal_child(PgArchPID, SIGQUIT);
3494         }
3495
3496         /*
3497          * Force a power-cycle of the pgstat process too.  (This isn't absolutely
3498          * necessary, but it seems like a good idea for robustness, and it
3499          * simplifies the state-machine logic in the case where a shutdown request
3500          * arrives during crash processing.)
3501          */
3502         if (PgStatPID != 0 && take_action)
3503         {
3504                 ereport(DEBUG2,
3505                                 (errmsg_internal("sending %s to process %d",
3506                                                                  "SIGQUIT",
3507                                                                  (int) PgStatPID)));
3508                 signal_child(PgStatPID, SIGQUIT);
3509                 allow_immediate_pgstat_restart();
3510         }
3511
3512         /* We do NOT restart the syslogger */
3513
3514         if (Shutdown != ImmediateShutdown)
3515                 FatalError = true;
3516
3517         /* We now transit into a state of waiting for children to die */
3518         if (pmState == PM_RECOVERY ||
3519                 pmState == PM_HOT_STANDBY ||
3520                 pmState == PM_RUN ||
3521                 pmState == PM_WAIT_BACKUP ||
3522                 pmState == PM_WAIT_READONLY ||
3523                 pmState == PM_SHUTDOWN)
3524                 pmState = PM_WAIT_BACKENDS;
3525
3526         /*
3527          * .. and if this doesn't happen quickly enough, now the clock is ticking
3528          * for us to kill them without mercy.
3529          */
3530         if (AbortStartTime == 0)
3531                 AbortStartTime = time(NULL);
3532 }
3533
3534 /*
3535  * Log the death of a child process.
3536  */
3537 static void
3538 LogChildExit(int lev, const char *procname, int pid, int exitstatus)
3539 {
3540         /*
3541          * size of activity_buffer is arbitrary, but set equal to default
3542          * track_activity_query_size
3543          */
3544         char            activity_buffer[1024];
3545         const char *activity = NULL;
3546
3547         if (!EXIT_STATUS_0(exitstatus))
3548                 activity = pgstat_get_crashed_backend_activity(pid,
3549                                                                                                            activity_buffer,
3550                                                                                                            sizeof(activity_buffer));
3551
3552         if (WIFEXITED(exitstatus))
3553                 ereport(lev,
3554
3555                 /*------
3556                   translator: %s is a noun phrase describing a child process, such as
3557                   "server process" */
3558                                 (errmsg("%s (PID %d) exited with exit code %d",
3559                                                 procname, pid, WEXITSTATUS(exitstatus)),
3560                                  activity ? errdetail("Failed process was running: %s", activity) : 0));
3561         else if (WIFSIGNALED(exitstatus))
3562 #if defined(WIN32)
3563                 ereport(lev,
3564
3565                 /*------
3566                   translator: %s is a noun phrase describing a child process, such as
3567                   "server process" */
3568                                 (errmsg("%s (PID %d) was terminated by exception 0x%X",
3569                                                 procname, pid, WTERMSIG(exitstatus)),
3570                                  errhint("See C include file \"ntstatus.h\" for a description of the hexadecimal value."),
3571                                  activity ? errdetail("Failed process was running: %s", activity) : 0));
3572 #elif defined(HAVE_DECL_SYS_SIGLIST) && HAVE_DECL_SYS_SIGLIST
3573                 ereport(lev,
3574
3575                 /*------
3576                   translator: %s is a noun phrase describing a child process, such as
3577                   "server process" */
3578                                 (errmsg("%s (PID %d) was terminated by signal %d: %s",
3579                                                 procname, pid, WTERMSIG(exitstatus),
3580                                                 WTERMSIG(exitstatus) < NSIG ?
3581                                                 sys_siglist[WTERMSIG(exitstatus)] : "(unknown)"),
3582                                  activity ? errdetail("Failed process was running: %s", activity) : 0));
3583 #else
3584                 ereport(lev,
3585
3586                 /*------
3587                   translator: %s is a noun phrase describing a child process, such as
3588                   "server process" */
3589                                 (errmsg("%s (PID %d) was terminated by signal %d",
3590                                                 procname, pid, WTERMSIG(exitstatus)),
3591                                  activity ? errdetail("Failed process was running: %s", activity) : 0));
3592 #endif
3593         else
3594                 ereport(lev,
3595
3596                 /*------
3597                   translator: %s is a noun phrase describing a child process, such as
3598                   "server process" */
3599                                 (errmsg("%s (PID %d) exited with unrecognized status %d",
3600                                                 procname, pid, exitstatus),
3601                                  activity ? errdetail("Failed process was running: %s", activity) : 0));
3602 }
3603
3604 /*
3605  * Advance the postmaster's state machine and take actions as appropriate
3606  *
3607  * This is common code for pmdie(), reaper() and sigusr1_handler(), which
3608  * receive the signals that might mean we need to change state.
3609  */
3610 static void
3611 PostmasterStateMachine(void)
3612 {
3613         if (pmState == PM_WAIT_BACKUP)
3614         {
3615                 /*
3616                  * PM_WAIT_BACKUP state ends when online backup mode is not active.
3617                  */
3618                 if (!BackupInProgress())
3619                         pmState = PM_WAIT_BACKENDS;
3620         }
3621
3622         if (pmState == PM_WAIT_READONLY)
3623         {
3624                 /*
3625                  * PM_WAIT_READONLY state ends when we have no regular backends that
3626                  * have been started during recovery.  We kill the startup and
3627                  * walreceiver processes and transition to PM_WAIT_BACKENDS.  Ideally,
3628                  * we might like to kill these processes first and then wait for
3629                  * backends to die off, but that doesn't work at present because
3630                  * killing the startup process doesn't release its locks.
3631                  */
3632                 if (CountChildren(BACKEND_TYPE_NORMAL) == 0)
3633                 {
3634                         if (StartupPID != 0)
3635                                 signal_child(StartupPID, SIGTERM);
3636                         if (WalReceiverPID != 0)
3637                                 signal_child(WalReceiverPID, SIGTERM);
3638                         pmState = PM_WAIT_BACKENDS;
3639                 }
3640         }
3641
3642         /*
3643          * If we are in a state-machine state that implies waiting for backends to
3644          * exit, see if they're all gone, and change state if so.
3645          */
3646         if (pmState == PM_WAIT_BACKENDS)
3647         {
3648                 /*
3649                  * PM_WAIT_BACKENDS state ends when we have no regular backends
3650                  * (including autovac workers), no bgworkers (including unconnected
3651                  * ones), and no walwriter, autovac launcher or bgwriter.  If we are
3652                  * doing crash recovery or an immediate shutdown then we expect the
3653                  * checkpointer to exit as well, otherwise not. The archiver, stats,
3654                  * and syslogger processes are disregarded since they are not
3655                  * connected to shared memory; we also disregard dead_end children
3656                  * here. Walsenders are also disregarded, they will be terminated
3657                  * later after writing the checkpoint record, like the archiver
3658                  * process.
3659                  */
3660                 if (CountChildren(BACKEND_TYPE_NORMAL | BACKEND_TYPE_WORKER) == 0 &&
3661                         StartupPID == 0 &&
3662                         WalReceiverPID == 0 &&
3663                         BgWriterPID == 0 &&
3664                         (CheckpointerPID == 0 ||
3665                          (!FatalError && Shutdown < ImmediateShutdown)) &&
3666                         WalWriterPID == 0 &&
3667                         AutoVacPID == 0)
3668                 {
3669                         if (Shutdown >= ImmediateShutdown || FatalError)
3670                         {
3671                                 /*
3672                                  * Start waiting for dead_end children to die.  This state
3673                                  * change causes ServerLoop to stop creating new ones.
3674                                  */
3675                                 pmState = PM_WAIT_DEAD_END;
3676
3677                                 /*
3678                                  * We already SIGQUIT'd the archiver and stats processes, if
3679                                  * any, when we started immediate shutdown or entered
3680                                  * FatalError state.
3681                                  */
3682                         }
3683                         else
3684                         {
3685                                 /*
3686                                  * If we get here, we are proceeding with normal shutdown. All
3687                                  * the regular children are gone, and it's time to tell the
3688                                  * checkpointer to do a shutdown checkpoint.
3689                                  */
3690                                 Assert(Shutdown > NoShutdown);
3691                                 /* Start the checkpointer if not running */
3692                                 if (CheckpointerPID == 0)
3693                                         CheckpointerPID = StartCheckpointer();
3694                                 /* And tell it to shut down */
3695                                 if (CheckpointerPID != 0)
3696                                 {
3697                                         signal_child(CheckpointerPID, SIGUSR2);
3698                                         pmState = PM_SHUTDOWN;
3699                                 }
3700                                 else
3701                                 {
3702                                         /*
3703                                          * If we failed to fork a checkpointer, just shut down.
3704                                          * Any required cleanup will happen at next restart. We
3705                                          * set FatalError so that an "abnormal shutdown" message
3706                                          * gets logged when we exit.
3707                                          */
3708                                         FatalError = true;
3709                                         pmState = PM_WAIT_DEAD_END;
3710
3711                                         /* Kill the walsenders, archiver and stats collector too */
3712                                         SignalChildren(SIGQUIT);
3713                                         if (PgArchPID != 0)
3714                                                 signal_child(PgArchPID, SIGQUIT);
3715                                         if (PgStatPID != 0)
3716                                                 signal_child(PgStatPID, SIGQUIT);
3717                                 }
3718                         }
3719                 }
3720         }
3721
3722         if (pmState == PM_SHUTDOWN_2)
3723         {
3724                 /*
3725                  * PM_SHUTDOWN_2 state ends when there's no other children than
3726                  * dead_end children left. There shouldn't be any regular backends
3727                  * left by now anyway; what we're really waiting for is walsenders and
3728                  * archiver.
3729                  *
3730                  * Walreceiver should normally be dead by now, but not when a fast
3731                  * shutdown is performed during recovery.
3732                  */
3733                 if (PgArchPID == 0 && CountChildren(BACKEND_TYPE_ALL) == 0 &&
3734                         WalReceiverPID == 0)
3735                 {
3736                         pmState = PM_WAIT_DEAD_END;
3737                 }
3738         }
3739
3740         if (pmState == PM_WAIT_DEAD_END)
3741         {
3742                 /*
3743                  * PM_WAIT_DEAD_END state ends when the BackendList is entirely empty
3744                  * (ie, no dead_end children remain), and the archiver and stats
3745                  * collector are gone too.
3746                  *
3747                  * The reason we wait for those two is to protect them against a new
3748                  * postmaster starting conflicting subprocesses; this isn't an
3749                  * ironclad protection, but it at least helps in the
3750                  * shutdown-and-immediately-restart scenario.  Note that they have
3751                  * already been sent appropriate shutdown signals, either during a
3752                  * normal state transition leading up to PM_WAIT_DEAD_END, or during
3753                  * FatalError processing.
3754                  */
3755                 if (dlist_is_empty(&BackendList) &&
3756                         PgArchPID == 0 && PgStatPID == 0)
3757                 {
3758                         /* These other guys should be dead already */
3759                         Assert(StartupPID == 0);
3760                         Assert(WalReceiverPID == 0);
3761                         Assert(BgWriterPID == 0);
3762                         Assert(CheckpointerPID == 0);
3763                         Assert(WalWriterPID == 0);
3764                         Assert(AutoVacPID == 0);
3765                         /* syslogger is not considered here */
3766                         pmState = PM_NO_CHILDREN;
3767                 }
3768         }
3769
3770         /*
3771          * If we've been told to shut down, we exit as soon as there are no
3772          * remaining children.  If there was a crash, cleanup will occur at the
3773          * next startup.  (Before PostgreSQL 8.3, we tried to recover from the
3774          * crash before exiting, but that seems unwise if we are quitting because
3775          * we got SIGTERM from init --- there may well not be time for recovery
3776          * before init decides to SIGKILL us.)
3777          *
3778          * Note that the syslogger continues to run.  It will exit when it sees
3779          * EOF on its input pipe, which happens when there are no more upstream
3780          * processes.
3781          */
3782         if (Shutdown > NoShutdown && pmState == PM_NO_CHILDREN)
3783         {
3784                 if (FatalError)
3785                 {
3786                         ereport(LOG, (errmsg("abnormal database system shutdown")));
3787                         ExitPostmaster(1);
3788                 }
3789                 else
3790                 {
3791                         /*
3792                          * Terminate exclusive backup mode to avoid recovery after a clean
3793                          * fast shutdown.  Since an exclusive backup can only be taken
3794                          * during normal running (and not, for example, while running
3795                          * under Hot Standby) it only makes sense to do this if we reached
3796                          * normal running. If we're still in recovery, the backup file is
3797                          * one we're recovering *from*, and we must keep it around so that
3798                          * recovery restarts from the right place.
3799                          */
3800                         if (ReachedNormalRunning)
3801                                 CancelBackup();
3802
3803                         /* Normal exit from the postmaster is here */
3804                         ExitPostmaster(0);
3805                 }
3806         }
3807
3808         /*
3809          * If the startup process failed, or the user does not want an automatic
3810          * restart after backend crashes, wait for all non-syslogger children to
3811          * exit, and then exit postmaster.  We don't try to reinitialize when the
3812          * startup process fails, because more than likely it will just fail again
3813          * and we will keep trying forever.
3814          */
3815         if (pmState == PM_NO_CHILDREN &&
3816                 (StartupStatus == STARTUP_CRASHED || !restart_after_crash))
3817                 ExitPostmaster(1);
3818
3819         /*
3820          * If we need to recover from a crash, wait for all non-syslogger children
3821          * to exit, then reset shmem and StartupDataBase.
3822          */
3823         if (FatalError && pmState == PM_NO_CHILDREN)
3824         {
3825                 ereport(LOG,
3826                                 (errmsg("all server processes terminated; reinitializing")));
3827
3828                 /* allow background workers to immediately restart */
3829                 ResetBackgroundWorkerCrashTimes();
3830
3831                 shmem_exit(1);
3832
3833                 /* re-read control file into local memory */
3834                 LocalProcessControlFile(true);
3835
3836                 reset_shared(PostPortNumber);
3837
3838                 StartupPID = StartupDataBase();
3839                 Assert(StartupPID != 0);
3840                 StartupStatus = STARTUP_RUNNING;
3841                 pmState = PM_STARTUP;
3842                 /* crash recovery started, reset SIGKILL flag */
3843                 AbortStartTime = 0;
3844         }
3845 }
3846
3847
3848 /*
3849  * Send a signal to a postmaster child process
3850  *
3851  * On systems that have setsid(), each child process sets itself up as a
3852  * process group leader.  For signals that are generally interpreted in the
3853  * appropriate fashion, we signal the entire process group not just the
3854  * direct child process.  This allows us to, for example, SIGQUIT a blocked
3855  * archive_recovery script, or SIGINT a script being run by a backend via
3856  * system().
3857  *
3858  * There is a race condition for recently-forked children: they might not
3859  * have executed setsid() yet.  So we signal the child directly as well as
3860  * the group.  We assume such a child will handle the signal before trying
3861  * to spawn any grandchild processes.  We also assume that signaling the
3862  * child twice will not cause any problems.
3863  */
3864 static void
3865 signal_child(pid_t pid, int signal)
3866 {
3867         if (kill(pid, signal) < 0)
3868                 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) pid, signal);
3869 #ifdef HAVE_SETSID
3870         switch (signal)
3871         {
3872                 case SIGINT:
3873                 case SIGTERM:
3874                 case SIGQUIT:
3875                 case SIGSTOP:
3876                 case SIGKILL:
3877                         if (kill(-pid, signal) < 0)
3878                                 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) (-pid), signal);
3879                         break;
3880                 default:
3881                         break;
3882         }
3883 #endif
3884 }
3885
3886 /*
3887  * Send a signal to the targeted children (but NOT special children;
3888  * dead_end children are never signaled, either).
3889  */
3890 static bool
3891 SignalSomeChildren(int signal, int target)
3892 {
3893         dlist_iter      iter;
3894         bool            signaled = false;
3895
3896         dlist_foreach(iter, &BackendList)
3897         {
3898                 Backend    *bp = dlist_container(Backend, elem, iter.cur);
3899
3900                 if (bp->dead_end)
3901                         continue;
3902
3903                 /*
3904                  * Since target == BACKEND_TYPE_ALL is the most common case, we test
3905                  * it first and avoid touching shared memory for every child.
3906                  */
3907                 if (target != BACKEND_TYPE_ALL)
3908                 {
3909                         /*
3910                          * Assign bkend_type for any recently announced WAL Sender
3911                          * processes.
3912                          */
3913                         if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
3914                                 IsPostmasterChildWalSender(bp->child_slot))
3915                                 bp->bkend_type = BACKEND_TYPE_WALSND;
3916
3917                         if (!(target & bp->bkend_type))
3918                                 continue;
3919                 }
3920
3921                 ereport(DEBUG4,
3922                                 (errmsg_internal("sending signal %d to process %d",
3923                                                                  signal, (int) bp->pid)));
3924                 signal_child(bp->pid, signal);
3925                 signaled = true;
3926         }
3927         return signaled;
3928 }
3929
3930 /*
3931  * Send a termination signal to children.  This considers all of our children
3932  * processes, except syslogger and dead_end backends.
3933  */
3934 static void
3935 TerminateChildren(int signal)
3936 {
3937         SignalChildren(signal);
3938         if (StartupPID != 0)
3939         {
3940                 signal_child(StartupPID, signal);
3941                 if (signal == SIGQUIT || signal == SIGKILL)
3942                         StartupStatus = STARTUP_SIGNALED;
3943         }
3944         if (BgWriterPID != 0)
3945                 signal_child(BgWriterPID, signal);
3946         if (CheckpointerPID != 0)
3947                 signal_child(CheckpointerPID, signal);
3948         if (WalWriterPID != 0)
3949                 signal_child(WalWriterPID, signal);
3950         if (WalReceiverPID != 0)
3951                 signal_child(WalReceiverPID, signal);
3952         if (AutoVacPID != 0)
3953                 signal_child(AutoVacPID, signal);
3954         if (PgArchPID != 0)
3955                 signal_child(PgArchPID, signal);
3956         if (PgStatPID != 0)
3957                 signal_child(PgStatPID, signal);
3958 }
3959
3960 /*
3961  * BackendStartup -- start backend process
3962  *
3963  * returns: STATUS_ERROR if the fork failed, STATUS_OK otherwise.
3964  *
3965  * Note: if you change this code, also consider StartAutovacuumWorker.
3966  */
3967 static int
3968 BackendStartup(Port *port)
3969 {
3970         Backend    *bn;                         /* for backend cleanup */
3971         pid_t           pid;
3972
3973         /*
3974          * Create backend data structure.  Better before the fork() so we can
3975          * handle failure cleanly.
3976          */
3977         bn = (Backend *) malloc(sizeof(Backend));
3978         if (!bn)
3979         {
3980                 ereport(LOG,
3981                                 (errcode(ERRCODE_OUT_OF_MEMORY),
3982                                  errmsg("out of memory")));
3983                 return STATUS_ERROR;
3984         }
3985
3986         /*
3987          * Compute the cancel key that will be assigned to this backend. The
3988          * backend will have its own copy in the forked-off process' value of
3989          * MyCancelKey, so that it can transmit the key to the frontend.
3990          */
3991         if (!RandomCancelKey(&MyCancelKey))
3992         {
3993                 free(bn);
3994                 ereport(LOG,
3995                                 (errcode(ERRCODE_INTERNAL_ERROR),
3996                                  errmsg("could not generate random cancel key")));
3997                 return STATUS_ERROR;
3998         }
3999
4000         bn->cancel_key = MyCancelKey;
4001
4002         /* Pass down canAcceptConnections state */
4003         port->canAcceptConnections = canAcceptConnections();
4004         bn->dead_end = (port->canAcceptConnections != CAC_OK &&
4005                                         port->canAcceptConnections != CAC_WAITBACKUP);
4006
4007         /*
4008          * Unless it's a dead_end child, assign it a child slot number
4009          */
4010         if (!bn->dead_end)
4011                 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
4012         else
4013                 bn->child_slot = 0;
4014
4015         /* Hasn't asked to be notified about any bgworkers yet */
4016         bn->bgworker_notify = false;
4017
4018 #ifdef EXEC_BACKEND
4019         pid = backend_forkexec(port);
4020 #else                                                   /* !EXEC_BACKEND */
4021         pid = fork_process();
4022         if (pid == 0)                           /* child */
4023         {
4024                 free(bn);
4025
4026                 /* Detangle from postmaster */
4027                 InitPostmasterChild();
4028
4029                 /* Close the postmaster's sockets */
4030                 ClosePostmasterPorts(false);
4031
4032                 /* Perform additional initialization and collect startup packet */
4033                 BackendInitialize(port);
4034
4035                 /* And run the backend */
4036                 BackendRun(port);
4037         }
4038 #endif                                                  /* EXEC_BACKEND */
4039
4040         if (pid < 0)
4041         {
4042                 /* in parent, fork failed */
4043                 int                     save_errno = errno;
4044
4045                 if (!bn->dead_end)
4046                         (void) ReleasePostmasterChildSlot(bn->child_slot);
4047                 free(bn);
4048                 errno = save_errno;
4049                 ereport(LOG,
4050                                 (errmsg("could not fork new process for connection: %m")));
4051                 report_fork_failure_to_client(port, save_errno);
4052                 return STATUS_ERROR;
4053         }
4054
4055         /* in parent, successful fork */
4056         ereport(DEBUG2,
4057                         (errmsg_internal("forked new backend, pid=%d socket=%d",
4058                                                          (int) pid, (int) port->sock)));
4059
4060         /*
4061          * Everything's been successful, it's safe to add this backend to our list
4062          * of backends.
4063          */
4064         bn->pid = pid;
4065         bn->bkend_type = BACKEND_TYPE_NORMAL;   /* Can change later to WALSND */
4066         dlist_push_head(&BackendList, &bn->elem);
4067
4068 #ifdef EXEC_BACKEND
4069         if (!bn->dead_end)
4070                 ShmemBackendArrayAdd(bn);
4071 #endif
4072
4073         return STATUS_OK;
4074 }
4075
4076 /*
4077  * Try to report backend fork() failure to client before we close the
4078  * connection.  Since we do not care to risk blocking the postmaster on
4079  * this connection, we set the connection to non-blocking and try only once.
4080  *
4081  * This is grungy special-purpose code; we cannot use backend libpq since
4082  * it's not up and running.
4083  */
4084 static void
4085 report_fork_failure_to_client(Port *port, int errnum)
4086 {
4087         char            buffer[1000];
4088         int                     rc;
4089
4090         /* Format the error message packet (always V2 protocol) */
4091         snprintf(buffer, sizeof(buffer), "E%s%s\n",
4092                          _("could not fork new process for connection: "),
4093                          strerror(errnum));
4094
4095         /* Set port to non-blocking.  Don't do send() if this fails */
4096         if (!pg_set_noblock(port->sock))
4097                 return;
4098
4099         /* We'll retry after EINTR, but ignore all other failures */
4100         do
4101         {
4102                 rc = send(port->sock, buffer, strlen(buffer) + 1, 0);
4103         } while (rc < 0 && errno == EINTR);
4104 }
4105
4106
4107 /*
4108  * BackendInitialize -- initialize an interactive (postmaster-child)
4109  *                              backend process, and collect the client's startup packet.
4110  *
4111  * returns: nothing.  Will not return at all if there's any failure.
4112  *
4113  * Note: this code does not depend on having any access to shared memory.
4114  * In the EXEC_BACKEND case, we are physically attached to shared memory
4115  * but have not yet set up most of our local pointers to shmem structures.
4116  */
4117 static void
4118 BackendInitialize(Port *port)
4119 {
4120         int                     status;
4121         int                     ret;
4122         char            remote_host[NI_MAXHOST];
4123         char            remote_port[NI_MAXSERV];
4124         char            remote_ps_data[NI_MAXHOST];
4125
4126         /* Save port etc. for ps status */
4127         MyProcPort = port;
4128
4129         /*
4130          * PreAuthDelay is a debugging aid for investigating problems in the
4131          * authentication cycle: it can be set in postgresql.conf to allow time to
4132          * attach to the newly-forked backend with a debugger.  (See also
4133          * PostAuthDelay, which we allow clients to pass through PGOPTIONS, but it
4134          * is not honored until after authentication.)
4135          */
4136         if (PreAuthDelay > 0)
4137                 pg_usleep(PreAuthDelay * 1000000L);
4138
4139         /* This flag will remain set until InitPostgres finishes authentication */
4140         ClientAuthInProgress = true;    /* limit visibility of log messages */
4141
4142         /* save process start time */
4143         port->SessionStartTime = GetCurrentTimestamp();
4144         MyStartTime = timestamptz_to_time_t(port->SessionStartTime);
4145
4146         /* set these to empty in case they are needed before we set them up */
4147         port->remote_host = "";
4148         port->remote_port = "";
4149
4150         /*
4151          * Initialize libpq and enable reporting of ereport errors to the client.
4152          * Must do this now because authentication uses libpq to send messages.
4153          */
4154         pq_init();                                      /* initialize libpq to talk to client */
4155         whereToSendOutput = DestRemote; /* now safe to ereport to client */
4156
4157         /*
4158          * We arrange for a simple exit(1) if we receive SIGTERM or SIGQUIT or
4159          * timeout while trying to collect the startup packet.  Otherwise the
4160          * postmaster cannot shutdown the database FAST or IMMED cleanly if a
4161          * buggy client fails to send the packet promptly.  XXX it follows that
4162          * the remainder of this function must tolerate losing control at any
4163          * instant.  Likewise, any pg_on_exit_callback registered before or during
4164          * this function must be prepared to execute at any instant between here
4165          * and the end of this function.  Furthermore, affected callbacks execute
4166          * partially or not at all when a second exit-inducing signal arrives
4167          * after proc_exit_prepare() decrements on_proc_exit_index.  (Thanks to
4168          * that mechanic, callbacks need not anticipate more than one call.)  This
4169          * is fragile; it ought to instead follow the norm of handling interrupts
4170          * at selected, safe opportunities.
4171          */
4172         pqsignal(SIGTERM, startup_die);
4173         pqsignal(SIGQUIT, startup_die);
4174         InitializeTimeouts();           /* establishes SIGALRM handler */
4175         PG_SETMASK(&StartupBlockSig);
4176
4177         /*
4178          * Get the remote host name and port for logging and status display.
4179          */
4180         remote_host[0] = '\0';
4181         remote_port[0] = '\0';
4182         if ((ret = pg_getnameinfo_all(&port->raddr.addr, port->raddr.salen,
4183                                                                   remote_host, sizeof(remote_host),
4184                                                                   remote_port, sizeof(remote_port),
4185                                                                   (log_hostname ? 0 : NI_NUMERICHOST) | NI_NUMERICSERV)) != 0)
4186                 ereport(WARNING,
4187                                 (errmsg_internal("pg_getnameinfo_all() failed: %s",
4188                                                                  gai_strerror(ret))));
4189         if (remote_port[0] == '\0')
4190                 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s", remote_host);
4191         else
4192                 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s(%s)", remote_host, remote_port);
4193
4194         /*
4195          * Save remote_host and remote_port in port structure (after this, they
4196          * will appear in log_line_prefix data for log messages).
4197          */
4198         port->remote_host = strdup(remote_host);
4199         port->remote_port = strdup(remote_port);
4200
4201         /* And now we can issue the Log_connections message, if wanted */
4202         if (Log_connections)
4203         {
4204                 if (remote_port[0])
4205                         ereport(LOG,
4206                                         (errmsg("connection received: host=%s port=%s",
4207                                                         remote_host,
4208                                                         remote_port)));
4209                 else
4210                         ereport(LOG,
4211                                         (errmsg("connection received: host=%s",
4212                                                         remote_host)));
4213         }
4214
4215         /*
4216          * If we did a reverse lookup to name, we might as well save the results
4217          * rather than possibly repeating the lookup during authentication.
4218          *
4219          * Note that we don't want to specify NI_NAMEREQD above, because then we'd
4220          * get nothing useful for a client without an rDNS entry.  Therefore, we
4221          * must check whether we got a numeric IPv4 or IPv6 address, and not save
4222          * it into remote_hostname if so.  (This test is conservative and might
4223          * sometimes classify a hostname as numeric, but an error in that
4224          * direction is safe; it only results in a possible extra lookup.)
4225          */
4226         if (log_hostname &&
4227                 ret == 0 &&
4228                 strspn(remote_host, "0123456789.") < strlen(remote_host) &&
4229                 strspn(remote_host, "0123456789ABCDEFabcdef:") < strlen(remote_host))
4230                 port->remote_hostname = strdup(remote_host);
4231
4232         /*
4233          * Ready to begin client interaction.  We will give up and exit(1) after a
4234          * time delay, so that a broken client can't hog a connection
4235          * indefinitely.  PreAuthDelay and any DNS interactions above don't count
4236          * against the time limit.
4237          *
4238          * Note: AuthenticationTimeout is applied here while waiting for the
4239          * startup packet, and then again in InitPostgres for the duration of any
4240          * authentication operations.  So a hostile client could tie up the
4241          * process for nearly twice AuthenticationTimeout before we kick him off.
4242          *
4243          * Note: because PostgresMain will call InitializeTimeouts again, the
4244          * registration of STARTUP_PACKET_TIMEOUT will be lost.  This is okay
4245          * since we never use it again after this function.
4246          */
4247         RegisterTimeout(STARTUP_PACKET_TIMEOUT, StartupPacketTimeoutHandler);
4248         enable_timeout_after(STARTUP_PACKET_TIMEOUT, AuthenticationTimeout * 1000);
4249
4250         /*
4251          * Receive the startup packet (which might turn out to be a cancel request
4252          * packet).
4253          */
4254         status = ProcessStartupPacket(port, false);
4255
4256         /*
4257          * Stop here if it was bad or a cancel packet.  ProcessStartupPacket
4258          * already did any appropriate error reporting.
4259          */
4260         if (status != STATUS_OK)
4261                 proc_exit(0);
4262
4263         /*
4264          * Now that we have the user and database name, we can set the process
4265          * title for ps.  It's good to do this as early as possible in startup.
4266          *
4267          * For a walsender, the ps display is set in the following form:
4268          *
4269          * postgres: walsender <user> <host> <activity>
4270          *
4271          * To achieve that, we pass "walsender" as username and username
4272          * as dbname to init_ps_display(). XXX: should add a new variant of
4273          * init_ps_display() to avoid abusing the parameters like this.
4274          */
4275         if (am_walsender)
4276                 init_ps_display(pgstat_get_backend_desc(B_WAL_SENDER), port->user_name, remote_ps_data,
4277                                                 update_process_title ? "authentication" : "");
4278         else
4279                 init_ps_display(port->user_name, port->database_name, remote_ps_data,
4280                                                 update_process_title ? "authentication" : "");
4281
4282         /*
4283          * Disable the timeout, and prevent SIGTERM/SIGQUIT again.
4284          */
4285         disable_timeout(STARTUP_PACKET_TIMEOUT, false);
4286         PG_SETMASK(&BlockSig);
4287 }
4288
4289
4290 /*
4291  * BackendRun -- set up the backend's argument list and invoke PostgresMain()
4292  *
4293  * returns:
4294  *              Shouldn't return at all.
4295  *              If PostgresMain() fails, return status.
4296  */
4297 static void
4298 BackendRun(Port *port)
4299 {
4300         char      **av;
4301         int                     maxac;
4302         int                     ac;
4303         long            secs;
4304         int                     usecs;
4305         int                     i;
4306
4307         /*
4308          * Don't want backend to be able to see the postmaster random number
4309          * generator state.  We have to clobber the static random_seed *and* start
4310          * a new random sequence in the random() library function.
4311          */
4312 #ifndef HAVE_STRONG_RANDOM
4313         random_seed = 0;
4314         random_start_time.tv_usec = 0;
4315 #endif
4316         /* slightly hacky way to convert timestamptz into integers */
4317         TimestampDifference(0, port->SessionStartTime, &secs, &usecs);
4318         srandom((unsigned int) (MyProcPid ^ (usecs << 12) ^ secs));
4319
4320         /*
4321          * Now, build the argv vector that will be given to PostgresMain.
4322          *
4323          * The maximum possible number of commandline arguments that could come
4324          * from ExtraOptions is (strlen(ExtraOptions) + 1) / 2; see
4325          * pg_split_opts().
4326          */
4327         maxac = 2;                                      /* for fixed args supplied below */
4328         maxac += (strlen(ExtraOptions) + 1) / 2;
4329
4330         av = (char **) MemoryContextAlloc(TopMemoryContext,
4331                                                                           maxac * sizeof(char *));
4332         ac = 0;
4333
4334         av[ac++] = "postgres";
4335
4336         /*
4337          * Pass any backend switches specified with -o on the postmaster's own
4338          * command line.  We assume these are secure.
4339          */
4340         pg_split_opts(av, &ac, ExtraOptions);
4341
4342         av[ac] = NULL;
4343
4344         Assert(ac < maxac);
4345
4346         /*
4347          * Debug: print arguments being passed to backend
4348          */
4349         ereport(DEBUG3,
4350                         (errmsg_internal("%s child[%d]: starting with (",
4351                                                          progname, (int) getpid())));
4352         for (i = 0; i < ac; ++i)
4353                 ereport(DEBUG3,
4354                                 (errmsg_internal("\t%s", av[i])));
4355         ereport(DEBUG3,
4356                         (errmsg_internal(")")));
4357
4358         /*
4359          * Make sure we aren't in PostmasterContext anymore.  (We can't delete it
4360          * just yet, though, because InitPostgres will need the HBA data.)
4361          */
4362         MemoryContextSwitchTo(TopMemoryContext);
4363
4364         PostgresMain(ac, av, port->database_name, port->user_name);
4365 }
4366
4367
4368 #ifdef EXEC_BACKEND
4369
4370 /*
4371  * postmaster_forkexec -- fork and exec a postmaster subprocess
4372  *
4373  * The caller must have set up the argv array already, except for argv[2]
4374  * which will be filled with the name of the temp variable file.
4375  *
4376  * Returns the child process PID, or -1 on fork failure (a suitable error
4377  * message has been logged on failure).
4378  *
4379  * All uses of this routine will dispatch to SubPostmasterMain in the
4380  * child process.
4381  */
4382 pid_t
4383 postmaster_forkexec(int argc, char *argv[])
4384 {
4385         Port            port;
4386
4387         /* This entry point passes dummy values for the Port variables */
4388         memset(&port, 0, sizeof(port));
4389         return internal_forkexec(argc, argv, &port);
4390 }
4391
4392 /*
4393  * backend_forkexec -- fork/exec off a backend process
4394  *
4395  * Some operating systems (WIN32) don't have fork() so we have to simulate
4396  * it by storing parameters that need to be passed to the child and
4397  * then create a new child process.
4398  *
4399  * returns the pid of the fork/exec'd process, or -1 on failure
4400  */
4401 static pid_t
4402 backend_forkexec(Port *port)
4403 {
4404         char       *av[4];
4405         int                     ac = 0;
4406
4407         av[ac++] = "postgres";
4408         av[ac++] = "--forkbackend";
4409         av[ac++] = NULL;                        /* filled in by internal_forkexec */
4410
4411         av[ac] = NULL;
4412         Assert(ac < lengthof(av));
4413
4414         return internal_forkexec(ac, av, port);
4415 }
4416
4417 #ifndef WIN32
4418
4419 /*
4420  * internal_forkexec non-win32 implementation
4421  *
4422  * - writes out backend variables to the parameter file
4423  * - fork():s, and then exec():s the child process
4424  */
4425 static pid_t
4426 internal_forkexec(int argc, char *argv[], Port *port)
4427 {
4428         static unsigned long tmpBackendFileNum = 0;
4429         pid_t           pid;
4430         char            tmpfilename[MAXPGPATH];
4431         BackendParameters param;
4432         FILE       *fp;
4433
4434         if (!save_backend_variables(&param, port))
4435                 return -1;                              /* log made by save_backend_variables */
4436
4437         /* Calculate name for temp file */
4438         snprintf(tmpfilename, MAXPGPATH, "%s/%s.backend_var.%d.%lu",
4439                          PG_TEMP_FILES_DIR, PG_TEMP_FILE_PREFIX,
4440                          MyProcPid, ++tmpBackendFileNum);
4441
4442         /* Open file */
4443         fp = AllocateFile(tmpfilename, PG_BINARY_W);
4444         if (!fp)
4445         {
4446                 /*
4447                  * As in OpenTemporaryFileInTablespace, try to make the temp-file
4448                  * directory
4449                  */
4450                 mkdir(PG_TEMP_FILES_DIR, S_IRWXU);
4451
4452                 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4453                 if (!fp)
4454                 {
4455                         ereport(LOG,
4456                                         (errcode_for_file_access(),
4457                                          errmsg("could not create file \"%s\": %m",
4458                                                         tmpfilename)));
4459                         return -1;
4460                 }
4461         }
4462
4463         if (fwrite(&param, sizeof(param), 1, fp) != 1)
4464         {
4465                 ereport(LOG,
4466                                 (errcode_for_file_access(),
4467                                  errmsg("could not write to file \"%s\": %m", tmpfilename)));
4468                 FreeFile(fp);
4469                 return -1;
4470         }
4471
4472         /* Release file */
4473         if (FreeFile(fp))
4474         {
4475                 ereport(LOG,
4476                                 (errcode_for_file_access(),
4477                                  errmsg("could not write to file \"%s\": %m", tmpfilename)));
4478                 return -1;
4479         }
4480
4481         /* Make sure caller set up argv properly */
4482         Assert(argc >= 3);
4483         Assert(argv[argc] == NULL);
4484         Assert(strncmp(argv[1], "--fork", 6) == 0);
4485         Assert(argv[2] == NULL);
4486
4487         /* Insert temp file name after --fork argument */
4488         argv[2] = tmpfilename;
4489
4490         /* Fire off execv in child */
4491         if ((pid = fork_process()) == 0)
4492         {
4493                 if (execv(postgres_exec_path, argv) < 0)
4494                 {
4495                         ereport(LOG,
4496                                         (errmsg("could not execute server process \"%s\": %m",
4497                                                         postgres_exec_path)));
4498                         /* We're already in the child process here, can't return */
4499                         exit(1);
4500                 }
4501         }
4502
4503         return pid;                                     /* Parent returns pid, or -1 on fork failure */
4504 }
4505 #else                                                   /* WIN32 */
4506
4507 /*
4508  * internal_forkexec win32 implementation
4509  *
4510  * - starts backend using CreateProcess(), in suspended state
4511  * - writes out backend variables to the parameter file
4512  *      - during this, duplicates handles and sockets required for
4513  *        inheritance into the new process
4514  * - resumes execution of the new process once the backend parameter
4515  *       file is complete.
4516  */
4517 static pid_t
4518 internal_forkexec(int argc, char *argv[], Port *port)
4519 {
4520         int                     retry_count = 0;
4521         STARTUPINFO si;
4522         PROCESS_INFORMATION pi;
4523         int                     i;
4524         int                     j;
4525         char            cmdLine[MAXPGPATH * 2];
4526         HANDLE          paramHandle;
4527         BackendParameters *param;
4528         SECURITY_ATTRIBUTES sa;
4529         char            paramHandleStr[32];
4530         win32_deadchild_waitinfo *childinfo;
4531
4532         /* Make sure caller set up argv properly */
4533         Assert(argc >= 3);
4534         Assert(argv[argc] == NULL);
4535         Assert(strncmp(argv[1], "--fork", 6) == 0);
4536         Assert(argv[2] == NULL);
4537
4538         /* Resume here if we need to retry */
4539 retry:
4540
4541         /* Set up shared memory for parameter passing */
4542         ZeroMemory(&sa, sizeof(sa));
4543         sa.nLength = sizeof(sa);
4544         sa.bInheritHandle = TRUE;
4545         paramHandle = CreateFileMapping(INVALID_HANDLE_VALUE,
4546                                                                         &sa,
4547                                                                         PAGE_READWRITE,
4548                                                                         0,
4549                                                                         sizeof(BackendParameters),
4550                                                                         NULL);
4551         if (paramHandle == INVALID_HANDLE_VALUE)
4552         {
4553                 elog(LOG, "could not create backend parameter file mapping: error code %lu",
4554                          GetLastError());
4555                 return -1;
4556         }
4557
4558         param = MapViewOfFile(paramHandle, FILE_MAP_WRITE, 0, 0, sizeof(BackendParameters));
4559         if (!param)
4560         {
4561                 elog(LOG, "could not map backend parameter memory: error code %lu",
4562                          GetLastError());
4563                 CloseHandle(paramHandle);
4564                 return -1;
4565         }
4566
4567         /* Insert temp file name after --fork argument */
4568 #ifdef _WIN64
4569         sprintf(paramHandleStr, "%llu", (LONG_PTR) paramHandle);
4570 #else
4571         sprintf(paramHandleStr, "%lu", (DWORD) paramHandle);
4572 #endif
4573         argv[2] = paramHandleStr;
4574
4575         /* Format the cmd line */
4576         cmdLine[sizeof(cmdLine) - 1] = '\0';
4577         cmdLine[sizeof(cmdLine) - 2] = '\0';
4578         snprintf(cmdLine, sizeof(cmdLine) - 1, "\"%s\"", postgres_exec_path);
4579         i = 0;
4580         while (argv[++i] != NULL)
4581         {
4582                 j = strlen(cmdLine);
4583                 snprintf(cmdLine + j, sizeof(cmdLine) - 1 - j, " \"%s\"", argv[i]);
4584         }
4585         if (cmdLine[sizeof(cmdLine) - 2] != '\0')
4586         {
4587                 elog(LOG, "subprocess command line too long");
4588                 return -1;
4589         }
4590
4591         memset(&pi, 0, sizeof(pi));
4592         memset(&si, 0, sizeof(si));
4593         si.cb = sizeof(si);
4594
4595         /*
4596          * Create the subprocess in a suspended state. This will be resumed later,
4597          * once we have written out the parameter file.
4598          */
4599         if (!CreateProcess(NULL, cmdLine, NULL, NULL, TRUE, CREATE_SUSPENDED,
4600                                            NULL, NULL, &si, &pi))
4601         {
4602                 elog(LOG, "CreateProcess call failed: %m (error code %lu)",
4603                          GetLastError());
4604                 return -1;
4605         }
4606
4607         if (!save_backend_variables(param, port, pi.hProcess, pi.dwProcessId))
4608         {
4609                 /*
4610                  * log made by save_backend_variables, but we have to clean up the
4611                  * mess with the half-started process
4612                  */
4613                 if (!TerminateProcess(pi.hProcess, 255))
4614                         ereport(LOG,
4615                                         (errmsg_internal("could not terminate unstarted process: error code %lu",
4616                                                                          GetLastError())));
4617                 CloseHandle(pi.hProcess);
4618                 CloseHandle(pi.hThread);
4619                 return -1;                              /* log made by save_backend_variables */
4620         }
4621
4622         /* Drop the parameter shared memory that is now inherited to the backend */
4623         if (!UnmapViewOfFile(param))
4624                 elog(LOG, "could not unmap view of backend parameter file: error code %lu",
4625                          GetLastError());
4626         if (!CloseHandle(paramHandle))
4627                 elog(LOG, "could not close handle to backend parameter file: error code %lu",
4628                          GetLastError());
4629
4630         /*
4631          * Reserve the memory region used by our main shared memory segment before
4632          * we resume the child process.  Normally this should succeed, but if ASLR
4633          * is active then it might sometimes fail due to the stack or heap having
4634          * gotten mapped into that range.  In that case, just terminate the
4635          * process and retry.
4636          */
4637         if (!pgwin32_ReserveSharedMemoryRegion(pi.hProcess))
4638         {
4639                 /* pgwin32_ReserveSharedMemoryRegion already made a log entry */
4640                 if (!TerminateProcess(pi.hProcess, 255))
4641                         ereport(LOG,
4642                                         (errmsg_internal("could not terminate process that failed to reserve memory: error code %lu",
4643                                                                          GetLastError())));
4644                 CloseHandle(pi.hProcess);
4645                 CloseHandle(pi.hThread);
4646                 if (++retry_count < 100)
4647                         goto retry;
4648                 ereport(LOG,
4649                                 (errmsg("giving up after too many tries to reserve shared memory"),
4650                                  errhint("This might be caused by ASLR or antivirus software.")));
4651                 return -1;
4652         }
4653
4654         /*
4655          * Now that the backend variables are written out, we start the child
4656          * thread so it can start initializing while we set up the rest of the
4657          * parent state.
4658          */
4659         if (ResumeThread(pi.hThread) == -1)
4660         {
4661                 if (!TerminateProcess(pi.hProcess, 255))
4662                 {
4663                         ereport(LOG,
4664                                         (errmsg_internal("could not terminate unstartable process: error code %lu",
4665                                                                          GetLastError())));
4666                         CloseHandle(pi.hProcess);
4667                         CloseHandle(pi.hThread);
4668                         return -1;
4669                 }
4670                 CloseHandle(pi.hProcess);
4671                 CloseHandle(pi.hThread);
4672                 ereport(LOG,
4673                                 (errmsg_internal("could not resume thread of unstarted process: error code %lu",
4674                                                                  GetLastError())));
4675                 return -1;
4676         }
4677
4678         /*
4679          * Queue a waiter for to signal when this child dies. The wait will be
4680          * handled automatically by an operating system thread pool.
4681          *
4682          * Note: use malloc instead of palloc, since it needs to be thread-safe.
4683          * Struct will be free():d from the callback function that runs on a
4684          * different thread.
4685          */
4686         childinfo = malloc(sizeof(win32_deadchild_waitinfo));
4687         if (!childinfo)
4688                 ereport(FATAL,
4689                                 (errcode(ERRCODE_OUT_OF_MEMORY),
4690                                  errmsg("out of memory")));
4691
4692         childinfo->procHandle = pi.hProcess;
4693         childinfo->procId = pi.dwProcessId;
4694
4695         if (!RegisterWaitForSingleObject(&childinfo->waitHandle,
4696                                                                          pi.hProcess,
4697                                                                          pgwin32_deadchild_callback,
4698                                                                          childinfo,
4699                                                                          INFINITE,
4700                                                                          WT_EXECUTEONLYONCE | WT_EXECUTEINWAITTHREAD))
4701                 ereport(FATAL,
4702                                 (errmsg_internal("could not register process for wait: error code %lu",
4703                                                                  GetLastError())));
4704
4705         /* Don't close pi.hProcess here - the wait thread needs access to it */
4706
4707         CloseHandle(pi.hThread);
4708
4709         return pi.dwProcessId;
4710 }
4711 #endif                                                  /* WIN32 */
4712
4713
4714 /*
4715  * SubPostmasterMain -- Get the fork/exec'd process into a state equivalent
4716  *                      to what it would be if we'd simply forked on Unix, and then
4717  *                      dispatch to the appropriate place.
4718  *
4719  * The first two command line arguments are expected to be "--forkFOO"
4720  * (where FOO indicates which postmaster child we are to become), and
4721  * the name of a variables file that we can read to load data that would
4722  * have been inherited by fork() on Unix.  Remaining arguments go to the
4723  * subprocess FooMain() routine.
4724  */
4725 void
4726 SubPostmasterMain(int argc, char *argv[])
4727 {
4728         Port            port;
4729
4730         /* In EXEC_BACKEND case we will not have inherited these settings */
4731         IsPostmasterEnvironment = true;
4732         whereToSendOutput = DestNone;
4733
4734         /* Setup as postmaster child */
4735         InitPostmasterChild();
4736
4737         /* Setup essential subsystems (to ensure elog() behaves sanely) */
4738         InitializeGUCOptions();
4739
4740         /* Check we got appropriate args */
4741         if (argc < 3)
4742                 elog(FATAL, "invalid subpostmaster invocation");
4743
4744         /* Read in the variables file */
4745         memset(&port, 0, sizeof(Port));
4746         read_backend_variables(argv[2], &port);
4747
4748         /* Close the postmaster's sockets (as soon as we know them) */
4749         ClosePostmasterPorts(strcmp(argv[1], "--forklog") == 0);
4750
4751         /*
4752          * Set reference point for stack-depth checking
4753          */
4754         set_stack_base();
4755
4756         /*
4757          * Set up memory area for GSS information. Mirrors the code in ConnCreate
4758          * for the non-exec case.
4759          */
4760 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
4761         port.gss = (pg_gssinfo *) calloc(1, sizeof(pg_gssinfo));
4762         if (!port.gss)
4763                 ereport(FATAL,
4764                                 (errcode(ERRCODE_OUT_OF_MEMORY),
4765                                  errmsg("out of memory")));
4766 #endif
4767
4768         /*
4769          * If appropriate, physically re-attach to shared memory segment. We want
4770          * to do this before going any further to ensure that we can attach at the
4771          * same address the postmaster used.  On the other hand, if we choose not
4772          * to re-attach, we may have other cleanup to do.
4773          *
4774          * If testing EXEC_BACKEND on Linux, you should run this as root before
4775          * starting the postmaster:
4776          *
4777          * echo 0 >/proc/sys/kernel/randomize_va_space
4778          *
4779          * This prevents using randomized stack and code addresses that cause the
4780          * child process's memory map to be different from the parent's, making it
4781          * sometimes impossible to attach to shared memory at the desired address.
4782          * Return the setting to its old value (usually '1' or '2') when finished.
4783          */
4784         if (strcmp(argv[1], "--forkbackend") == 0 ||
4785                 strcmp(argv[1], "--forkavlauncher") == 0 ||
4786                 strcmp(argv[1], "--forkavworker") == 0 ||
4787                 strcmp(argv[1], "--forkboot") == 0 ||
4788                 strncmp(argv[1], "--forkbgworker=", 15) == 0)
4789                 PGSharedMemoryReAttach();
4790         else
4791                 PGSharedMemoryNoReAttach();
4792
4793         /* autovacuum needs this set before calling InitProcess */
4794         if (strcmp(argv[1], "--forkavlauncher") == 0)
4795                 AutovacuumLauncherIAm();
4796         if (strcmp(argv[1], "--forkavworker") == 0)
4797                 AutovacuumWorkerIAm();
4798
4799         /*
4800          * Start our win32 signal implementation. This has to be done after we
4801          * read the backend variables, because we need to pick up the signal pipe
4802          * from the parent process.
4803          */
4804 #ifdef WIN32
4805         pgwin32_signal_initialize();
4806 #endif
4807
4808         /* In EXEC_BACKEND case we will not have inherited these settings */
4809         pqinitmask();
4810         PG_SETMASK(&BlockSig);
4811
4812         /* Read in remaining GUC variables */
4813         read_nondefault_variables();
4814
4815         /*
4816          * (re-)read control file, as it contains config. The postmaster will
4817          * already have read this, but this process doesn't know about that.
4818          */
4819         LocalProcessControlFile(false);
4820
4821         /*
4822          * Reload any libraries that were preloaded by the postmaster.  Since we
4823          * exec'd this process, those libraries didn't come along with us; but we
4824          * should load them into all child processes to be consistent with the
4825          * non-EXEC_BACKEND behavior.
4826          */
4827         process_shared_preload_libraries();
4828
4829         /* Run backend or appropriate child */
4830         if (strcmp(argv[1], "--forkbackend") == 0)
4831         {
4832                 Assert(argc == 3);              /* shouldn't be any more args */
4833
4834                 /*
4835                  * Need to reinitialize the SSL library in the backend, since the
4836                  * context structures contain function pointers and cannot be passed
4837                  * through the parameter file.
4838                  *
4839                  * If for some reason reload fails (maybe the user installed broken
4840                  * key files), soldier on without SSL; that's better than all
4841                  * connections becoming impossible.
4842                  *
4843                  * XXX should we do this in all child processes?  For the moment it's
4844                  * enough to do it in backend children.
4845                  */
4846 #ifdef USE_SSL
4847                 if (EnableSSL)
4848                 {
4849                         if (secure_initialize(false) == 0)
4850                                 LoadedSSL = true;
4851                         else
4852                                 ereport(LOG,
4853                                                 (errmsg("SSL configuration could not be loaded in child process")));
4854                 }
4855 #endif
4856
4857                 /*
4858                  * Perform additional initialization and collect startup packet.
4859                  *
4860                  * We want to do this before InitProcess() for a couple of reasons: 1.
4861                  * so that we aren't eating up a PGPROC slot while waiting on the
4862                  * client. 2. so that if InitProcess() fails due to being out of
4863                  * PGPROC slots, we have already initialized libpq and are able to
4864                  * report the error to the client.
4865                  */
4866                 BackendInitialize(&port);
4867
4868                 /* Restore basic shared memory pointers */
4869                 InitShmemAccess(UsedShmemSegAddr);
4870
4871                 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4872                 InitProcess();
4873
4874                 /* Attach process to shared data structures */
4875                 CreateSharedMemoryAndSemaphores(false, 0);
4876
4877                 /* And run the backend */
4878                 BackendRun(&port);              /* does not return */
4879         }
4880         if (strcmp(argv[1], "--forkboot") == 0)
4881         {
4882                 /* Restore basic shared memory pointers */
4883                 InitShmemAccess(UsedShmemSegAddr);
4884
4885                 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4886                 InitAuxiliaryProcess();
4887
4888                 /* Attach process to shared data structures */
4889                 CreateSharedMemoryAndSemaphores(false, 0);
4890
4891                 AuxiliaryProcessMain(argc - 2, argv + 2);       /* does not return */
4892         }
4893         if (strcmp(argv[1], "--forkavlauncher") == 0)
4894         {
4895                 /* Restore basic shared memory pointers */
4896                 InitShmemAccess(UsedShmemSegAddr);
4897
4898                 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4899                 InitProcess();
4900
4901                 /* Attach process to shared data structures */
4902                 CreateSharedMemoryAndSemaphores(false, 0);
4903
4904                 AutoVacLauncherMain(argc - 2, argv + 2);        /* does not return */
4905         }
4906         if (strcmp(argv[1], "--forkavworker") == 0)
4907         {
4908                 /* Restore basic shared memory pointers */
4909                 InitShmemAccess(UsedShmemSegAddr);
4910
4911                 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4912                 InitProcess();
4913
4914                 /* Attach process to shared data structures */
4915                 CreateSharedMemoryAndSemaphores(false, 0);
4916
4917                 AutoVacWorkerMain(argc - 2, argv + 2);  /* does not return */
4918         }
4919         if (strncmp(argv[1], "--forkbgworker=", 15) == 0)
4920         {
4921                 int                     shmem_slot;
4922
4923                 /* do this as early as possible; in particular, before InitProcess() */
4924                 IsBackgroundWorker = true;
4925
4926                 /* Restore basic shared memory pointers */
4927                 InitShmemAccess(UsedShmemSegAddr);
4928
4929                 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4930                 InitProcess();
4931
4932                 /* Attach process to shared data structures */
4933                 CreateSharedMemoryAndSemaphores(false, 0);
4934
4935                 /* Fetch MyBgworkerEntry from shared memory */
4936                 shmem_slot = atoi(argv[1] + 15);
4937                 MyBgworkerEntry = BackgroundWorkerEntry(shmem_slot);
4938
4939                 StartBackgroundWorker();
4940         }
4941         if (strcmp(argv[1], "--forkarch") == 0)
4942         {
4943                 /* Do not want to attach to shared memory */
4944
4945                 PgArchiverMain(argc, argv); /* does not return */
4946         }
4947         if (strcmp(argv[1], "--forkcol") == 0)
4948         {
4949                 /* Do not want to attach to shared memory */
4950
4951                 PgstatCollectorMain(argc, argv);        /* does not return */
4952         }
4953         if (strcmp(argv[1], "--forklog") == 0)
4954         {
4955                 /* Do not want to attach to shared memory */
4956
4957                 SysLoggerMain(argc, argv);      /* does not return */
4958         }
4959
4960         abort();                                        /* shouldn't get here */
4961 }
4962 #endif                                                  /* EXEC_BACKEND */
4963
4964
4965 /*
4966  * ExitPostmaster -- cleanup
4967  *
4968  * Do NOT call exit() directly --- always go through here!
4969  */
4970 static void
4971 ExitPostmaster(int status)
4972 {
4973 #ifdef HAVE_PTHREAD_IS_THREADED_NP
4974
4975         /*
4976          * There is no known cause for a postmaster to become multithreaded after
4977          * startup.  Recheck to account for the possibility of unknown causes.
4978          * This message uses LOG level, because an unclean shutdown at this point
4979          * would usually not look much different from a clean shutdown.
4980          */
4981         if (pthread_is_threaded_np() != 0)
4982                 ereport(LOG,
4983                                 (errcode(ERRCODE_INTERNAL_ERROR),
4984                                  errmsg_internal("postmaster became multithreaded"),
4985                                  errdetail("Please report this to <pgsql-bugs@postgresql.org>.")));
4986 #endif
4987
4988         /* should cleanup shared memory and kill all backends */
4989
4990         /*
4991          * Not sure of the semantics here.  When the Postmaster dies, should the
4992          * backends all be killed? probably not.
4993          *
4994          * MUST         -- vadim 05-10-1999
4995          */
4996
4997         proc_exit(status);
4998 }
4999
5000 /*
5001  * sigusr1_handler - handle signal conditions from child processes
5002  */
5003 static void
5004 sigusr1_handler(SIGNAL_ARGS)
5005 {
5006         int                     save_errno = errno;
5007
5008         PG_SETMASK(&BlockSig);
5009
5010         /* Process background worker state change. */
5011         if (CheckPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE))
5012         {
5013                 BackgroundWorkerStateChange();
5014                 StartWorkerNeeded = true;
5015         }
5016
5017         /*
5018          * RECOVERY_STARTED and BEGIN_HOT_STANDBY signals are ignored in
5019          * unexpected states. If the startup process quickly starts up, completes
5020          * recovery, exits, we might process the death of the startup process
5021          * first. We don't want to go back to recovery in that case.
5022          */
5023         if (CheckPostmasterSignal(PMSIGNAL_RECOVERY_STARTED) &&
5024                 pmState == PM_STARTUP && Shutdown == NoShutdown)
5025         {
5026                 /* WAL redo has started. We're out of reinitialization. */
5027                 FatalError = false;
5028                 Assert(AbortStartTime == 0);
5029
5030                 /*
5031                  * Crank up the background tasks.  It doesn't matter if this fails,
5032                  * we'll just try again later.
5033                  */
5034                 Assert(CheckpointerPID == 0);
5035                 CheckpointerPID = StartCheckpointer();
5036                 Assert(BgWriterPID == 0);
5037                 BgWriterPID = StartBackgroundWriter();
5038
5039                 /*
5040                  * Start the archiver if we're responsible for (re-)archiving received
5041                  * files.
5042                  */
5043                 Assert(PgArchPID == 0);
5044                 if (XLogArchivingAlways())
5045                         PgArchPID = pgarch_start();
5046
5047                 /*
5048                  * If we aren't planning to enter hot standby mode later, treat
5049                  * RECOVERY_STARTED as meaning we're out of startup, and report status
5050                  * accordingly.
5051                  */
5052                 if (!EnableHotStandby)
5053                 {
5054                         AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STANDBY);
5055 #ifdef USE_SYSTEMD
5056                         sd_notify(0, "READY=1");
5057 #endif
5058                 }
5059
5060                 pmState = PM_RECOVERY;
5061         }
5062         if (CheckPostmasterSignal(PMSIGNAL_BEGIN_HOT_STANDBY) &&
5063                 pmState == PM_RECOVERY && Shutdown == NoShutdown)
5064         {
5065                 /*
5066                  * Likewise, start other special children as needed.
5067                  */
5068                 Assert(PgStatPID == 0);
5069                 PgStatPID = pgstat_start();
5070
5071                 ereport(LOG,
5072                                 (errmsg("database system is ready to accept read only connections")));
5073
5074                 /* Report status */
5075                 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
5076 #ifdef USE_SYSTEMD
5077                 sd_notify(0, "READY=1");
5078 #endif
5079
5080                 pmState = PM_HOT_STANDBY;
5081                 /* Some workers may be scheduled to start now */
5082                 StartWorkerNeeded = true;
5083         }
5084
5085         if (StartWorkerNeeded || HaveCrashedWorker)
5086                 maybe_start_bgworkers();
5087
5088         if (CheckPostmasterSignal(PMSIGNAL_WAKEN_ARCHIVER) &&
5089                 PgArchPID != 0)
5090         {
5091                 /*
5092                  * Send SIGUSR1 to archiver process, to wake it up and begin archiving
5093                  * next WAL file.
5094                  */
5095                 signal_child(PgArchPID, SIGUSR1);
5096         }
5097
5098         if (CheckPostmasterSignal(PMSIGNAL_ROTATE_LOGFILE) &&
5099                 SysLoggerPID != 0)
5100         {
5101                 /* Tell syslogger to rotate logfile */
5102                 signal_child(SysLoggerPID, SIGUSR1);
5103         }
5104
5105         if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER) &&
5106                 Shutdown == NoShutdown)
5107         {
5108                 /*
5109                  * Start one iteration of the autovacuum daemon, even if autovacuuming
5110                  * is nominally not enabled.  This is so we can have an active defense
5111                  * against transaction ID wraparound.  We set a flag for the main loop
5112                  * to do it rather than trying to do it here --- this is because the
5113                  * autovac process itself may send the signal, and we want to handle
5114                  * that by launching another iteration as soon as the current one
5115                  * completes.
5116                  */
5117                 start_autovac_launcher = true;
5118         }
5119
5120         if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER) &&
5121                 Shutdown == NoShutdown)
5122         {
5123                 /* The autovacuum launcher wants us to start a worker process. */
5124                 StartAutovacuumWorker();
5125         }
5126
5127         if (CheckPostmasterSignal(PMSIGNAL_START_WALRECEIVER))
5128         {
5129                 /* Startup Process wants us to start the walreceiver process. */
5130                 /* Start immediately if possible, else remember request for later. */
5131                 WalReceiverRequested = true;
5132                 MaybeStartWalReceiver();
5133         }
5134
5135         if (CheckPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE) &&
5136                 (pmState == PM_WAIT_BACKUP || pmState == PM_WAIT_BACKENDS))
5137         {
5138                 /* Advance postmaster's state machine */
5139                 PostmasterStateMachine();
5140         }
5141
5142         if (CheckPromoteSignal() && StartupPID != 0 &&
5143                 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5144                  pmState == PM_HOT_STANDBY || pmState == PM_WAIT_READONLY))
5145         {
5146                 /* Tell startup process to finish recovery */
5147                 signal_child(StartupPID, SIGUSR2);
5148         }
5149
5150         PG_SETMASK(&UnBlockSig);
5151
5152         errno = save_errno;
5153 }
5154
5155 /*
5156  * SIGTERM or SIGQUIT while processing startup packet.
5157  * Clean up and exit(1).
5158  *
5159  * XXX: possible future improvement: try to send a message indicating
5160  * why we are disconnecting.  Problem is to be sure we don't block while
5161  * doing so, nor mess up SSL initialization.  In practice, if the client
5162  * has wedged here, it probably couldn't do anything with the message anyway.
5163  */
5164 static void
5165 startup_die(SIGNAL_ARGS)
5166 {
5167         proc_exit(1);
5168 }
5169
5170 /*
5171  * Dummy signal handler
5172  *
5173  * We use this for signals that we don't actually use in the postmaster,
5174  * but we do use in backends.  If we were to SIG_IGN such signals in the
5175  * postmaster, then a newly started backend might drop a signal that arrives
5176  * before it's able to reconfigure its signal processing.  (See notes in
5177  * tcop/postgres.c.)
5178  */
5179 static void
5180 dummy_handler(SIGNAL_ARGS)
5181 {
5182 }
5183
5184 /*
5185  * Timeout while processing startup packet.
5186  * As for startup_die(), we clean up and exit(1).
5187  */
5188 static void
5189 StartupPacketTimeoutHandler(void)
5190 {
5191         proc_exit(1);
5192 }
5193
5194
5195 /*
5196  * Generate a random cancel key.
5197  */
5198 static bool
5199 RandomCancelKey(int32 *cancel_key)
5200 {
5201 #ifdef HAVE_STRONG_RANDOM
5202         return pg_strong_random((char *) cancel_key, sizeof(int32));
5203 #else
5204
5205         /*
5206          * If built with --disable-strong-random, use plain old erand48.
5207          *
5208          * We cannot use pg_backend_random() in postmaster, because it stores its
5209          * state in shared memory.
5210          */
5211         static unsigned short seed[3];
5212
5213         /*
5214          * Select a random seed at the time of first receiving a request.
5215          */
5216         if (random_seed == 0)
5217         {
5218                 struct timeval random_stop_time;
5219
5220                 gettimeofday(&random_stop_time, NULL);
5221
5222                 seed[0] = (unsigned short) random_start_time.tv_usec;
5223                 seed[1] = (unsigned short) (random_stop_time.tv_usec) ^ (random_start_time.tv_usec >> 16);
5224                 seed[2] = (unsigned short) (random_stop_time.tv_usec >> 16);
5225
5226                 random_seed = 1;
5227         }
5228
5229         *cancel_key = pg_jrand48(seed);
5230
5231         return true;
5232 #endif
5233 }
5234
5235 /*
5236  * Count up number of child processes of specified types (dead_end children
5237  * are always excluded).
5238  */
5239 static int
5240 CountChildren(int target)
5241 {
5242         dlist_iter      iter;
5243         int                     cnt = 0;
5244
5245         dlist_foreach(iter, &BackendList)
5246         {
5247                 Backend    *bp = dlist_container(Backend, elem, iter.cur);
5248
5249                 if (bp->dead_end)
5250                         continue;
5251
5252                 /*
5253                  * Since target == BACKEND_TYPE_ALL is the most common case, we test
5254                  * it first and avoid touching shared memory for every child.
5255                  */
5256                 if (target != BACKEND_TYPE_ALL)
5257                 {
5258                         /*
5259                          * Assign bkend_type for any recently announced WAL Sender
5260                          * processes.
5261                          */
5262                         if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
5263                                 IsPostmasterChildWalSender(bp->child_slot))
5264                                 bp->bkend_type = BACKEND_TYPE_WALSND;
5265
5266                         if (!(target & bp->bkend_type))
5267                                 continue;
5268                 }
5269
5270                 cnt++;
5271         }
5272         return cnt;
5273 }
5274
5275
5276 /*
5277  * StartChildProcess -- start an auxiliary process for the postmaster
5278  *
5279  * "type" determines what kind of child will be started.  All child types
5280  * initially go to AuxiliaryProcessMain, which will handle common setup.
5281  *
5282  * Return value of StartChildProcess is subprocess' PID, or 0 if failed
5283  * to start subprocess.
5284  */
5285 static pid_t
5286 StartChildProcess(AuxProcType type)
5287 {
5288         pid_t           pid;
5289         char       *av[10];
5290         int                     ac = 0;
5291         char            typebuf[32];
5292
5293         /*
5294          * Set up command-line arguments for subprocess
5295          */
5296         av[ac++] = "postgres";
5297
5298 #ifdef EXEC_BACKEND
5299         av[ac++] = "--forkboot";
5300         av[ac++] = NULL;                        /* filled in by postmaster_forkexec */
5301 #endif
5302
5303         snprintf(typebuf, sizeof(typebuf), "-x%d", type);
5304         av[ac++] = typebuf;
5305
5306         av[ac] = NULL;
5307         Assert(ac < lengthof(av));
5308
5309 #ifdef EXEC_BACKEND
5310         pid = postmaster_forkexec(ac, av);
5311 #else                                                   /* !EXEC_BACKEND */
5312         pid = fork_process();
5313
5314         if (pid == 0)                           /* child */
5315         {
5316                 InitPostmasterChild();
5317
5318                 /* Close the postmaster's sockets */
5319                 ClosePostmasterPorts(false);
5320
5321                 /* Release postmaster's working memory context */
5322                 MemoryContextSwitchTo(TopMemoryContext);
5323                 MemoryContextDelete(PostmasterContext);
5324                 PostmasterContext = NULL;
5325
5326                 AuxiliaryProcessMain(ac, av);
5327                 ExitPostmaster(0);
5328         }
5329 #endif                                                  /* EXEC_BACKEND */
5330
5331         if (pid < 0)
5332         {
5333                 /* in parent, fork failed */
5334                 int                     save_errno = errno;
5335
5336                 errno = save_errno;
5337                 switch (type)
5338                 {
5339                         case StartupProcess:
5340                                 ereport(LOG,
5341                                                 (errmsg("could not fork startup process: %m")));
5342                                 break;
5343                         case BgWriterProcess:
5344                                 ereport(LOG,
5345                                                 (errmsg("could not fork background writer process: %m")));
5346                                 break;
5347                         case CheckpointerProcess:
5348                                 ereport(LOG,
5349                                                 (errmsg("could not fork checkpointer process: %m")));
5350                                 break;
5351                         case WalWriterProcess:
5352                                 ereport(LOG,
5353                                                 (errmsg("could not fork WAL writer process: %m")));
5354                                 break;
5355                         case WalReceiverProcess:
5356                                 ereport(LOG,
5357                                                 (errmsg("could not fork WAL receiver process: %m")));
5358                                 break;
5359                         default:
5360                                 ereport(LOG,
5361                                                 (errmsg("could not fork process: %m")));
5362                                 break;
5363                 }
5364
5365                 /*
5366                  * fork failure is fatal during startup, but there's no need to choke
5367                  * immediately if starting other child types fails.
5368                  */
5369                 if (type == StartupProcess)
5370                         ExitPostmaster(1);
5371                 return 0;
5372         }
5373
5374         /*
5375          * in parent, successful fork
5376          */
5377         return pid;
5378 }
5379
5380 /*
5381  * StartAutovacuumWorker
5382  *              Start an autovac worker process.
5383  *
5384  * This function is here because it enters the resulting PID into the
5385  * postmaster's private backends list.
5386  *
5387  * NB -- this code very roughly matches BackendStartup.
5388  */
5389 static void
5390 StartAutovacuumWorker(void)
5391 {
5392         Backend    *bn;
5393
5394         /*
5395          * If not in condition to run a process, don't try, but handle it like a
5396          * fork failure.  This does not normally happen, since the signal is only
5397          * supposed to be sent by autovacuum launcher when it's OK to do it, but
5398          * we have to check to avoid race-condition problems during DB state
5399          * changes.
5400          */
5401         if (canAcceptConnections() == CAC_OK)
5402         {
5403                 /*
5404                  * Compute the cancel key that will be assigned to this session. We
5405                  * probably don't need cancel keys for autovac workers, but we'd
5406                  * better have something random in the field to prevent unfriendly
5407                  * people from sending cancels to them.
5408                  */
5409                 if (!RandomCancelKey(&MyCancelKey))
5410                 {
5411                         ereport(LOG,
5412                                         (errcode(ERRCODE_INTERNAL_ERROR),
5413                                          errmsg("could not generate random cancel key")));
5414                         return;
5415                 }
5416
5417                 bn = (Backend *) malloc(sizeof(Backend));
5418                 if (bn)
5419                 {
5420                         bn->cancel_key = MyCancelKey;
5421
5422                         /* Autovac workers are not dead_end and need a child slot */
5423                         bn->dead_end = false;
5424                         bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5425                         bn->bgworker_notify = false;
5426
5427                         bn->pid = StartAutoVacWorker();
5428                         if (bn->pid > 0)
5429                         {
5430                                 bn->bkend_type = BACKEND_TYPE_AUTOVAC;
5431                                 dlist_push_head(&BackendList, &bn->elem);
5432 #ifdef EXEC_BACKEND
5433                                 ShmemBackendArrayAdd(bn);
5434 #endif
5435                                 /* all OK */
5436                                 return;
5437                         }
5438
5439                         /*
5440                          * fork failed, fall through to report -- actual error message was
5441                          * logged by StartAutoVacWorker
5442                          */
5443                         (void) ReleasePostmasterChildSlot(bn->child_slot);
5444                         free(bn);
5445                 }
5446                 else
5447                         ereport(LOG,
5448                                         (errcode(ERRCODE_OUT_OF_MEMORY),
5449                                          errmsg("out of memory")));
5450         }
5451
5452         /*
5453          * Report the failure to the launcher, if it's running.  (If it's not, we
5454          * might not even be connected to shared memory, so don't try to call
5455          * AutoVacWorkerFailed.)  Note that we also need to signal it so that it
5456          * responds to the condition, but we don't do that here, instead waiting
5457          * for ServerLoop to do it.  This way we avoid a ping-pong signalling in
5458          * quick succession between the autovac launcher and postmaster in case
5459          * things get ugly.
5460          */
5461         if (AutoVacPID != 0)
5462         {
5463                 AutoVacWorkerFailed();
5464                 avlauncher_needs_signal = true;
5465         }
5466 }
5467
5468 /*
5469  * MaybeStartWalReceiver
5470  *              Start the WAL receiver process, if not running and our state allows.
5471  */
5472 static void
5473 MaybeStartWalReceiver(void)
5474 {
5475         if (WalReceiverPID == 0 &&
5476                 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5477                  pmState == PM_HOT_STANDBY || pmState == PM_WAIT_READONLY) &&
5478                 Shutdown == NoShutdown)
5479         {
5480                 WalReceiverPID = StartWalReceiver();
5481                 WalReceiverRequested = false;
5482         }
5483 }
5484
5485
5486 /*
5487  * Create the opts file
5488  */
5489 static bool
5490 CreateOptsFile(int argc, char *argv[], char *fullprogname)
5491 {
5492         FILE       *fp;
5493         int                     i;
5494
5495 #define OPTS_FILE       "postmaster.opts"
5496
5497         if ((fp = fopen(OPTS_FILE, "w")) == NULL)
5498         {
5499                 elog(LOG, "could not create file \"%s\": %m", OPTS_FILE);
5500                 return false;
5501         }
5502
5503         fprintf(fp, "%s", fullprogname);
5504         for (i = 1; i < argc; i++)
5505                 fprintf(fp, " \"%s\"", argv[i]);
5506         fputs("\n", fp);
5507
5508         if (fclose(fp))
5509         {
5510                 elog(LOG, "could not write file \"%s\": %m", OPTS_FILE);
5511                 return false;
5512         }
5513
5514         return true;
5515 }
5516
5517
5518 /*
5519  * MaxLivePostmasterChildren
5520  *
5521  * This reports the number of entries needed in per-child-process arrays
5522  * (the PMChildFlags array, and if EXEC_BACKEND the ShmemBackendArray).
5523  * These arrays include regular backends, autovac workers, walsenders
5524  * and background workers, but not special children nor dead_end children.
5525  * This allows the arrays to have a fixed maximum size, to wit the same
5526  * too-many-children limit enforced by canAcceptConnections().  The exact value
5527  * isn't too critical as long as it's more than MaxBackends.
5528  */
5529 int
5530 MaxLivePostmasterChildren(void)
5531 {
5532         return 2 * (MaxConnections + autovacuum_max_workers + 1 +
5533                                 max_worker_processes);
5534 }
5535
5536 /*
5537  * Connect background worker to a database.
5538  */
5539 void
5540 BackgroundWorkerInitializeConnection(char *dbname, char *username)
5541 {
5542         BackgroundWorker *worker = MyBgworkerEntry;
5543
5544         /* XXX is this the right errcode? */
5545         if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5546                 ereport(FATAL,
5547                                 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5548                                  errmsg("database connection requirement not indicated during registration")));
5549
5550         InitPostgres(dbname, InvalidOid, username, InvalidOid, NULL);
5551
5552         /* it had better not gotten out of "init" mode yet */
5553         if (!IsInitProcessingMode())
5554                 ereport(ERROR,
5555                                 (errmsg("invalid processing mode in background worker")));
5556         SetProcessingMode(NormalProcessing);
5557 }
5558
5559 /*
5560  * Connect background worker to a database using OIDs.
5561  */
5562 void
5563 BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid)
5564 {
5565         BackgroundWorker *worker = MyBgworkerEntry;
5566
5567         /* XXX is this the right errcode? */
5568         if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5569                 ereport(FATAL,
5570                                 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5571                                  errmsg("database connection requirement not indicated during registration")));
5572
5573         InitPostgres(NULL, dboid, NULL, useroid, NULL);
5574
5575         /* it had better not gotten out of "init" mode yet */
5576         if (!IsInitProcessingMode())
5577                 ereport(ERROR,
5578                                 (errmsg("invalid processing mode in background worker")));
5579         SetProcessingMode(NormalProcessing);
5580 }
5581
5582 /*
5583  * Block/unblock signals in a background worker
5584  */
5585 void
5586 BackgroundWorkerBlockSignals(void)
5587 {
5588         PG_SETMASK(&BlockSig);
5589 }
5590
5591 void
5592 BackgroundWorkerUnblockSignals(void)
5593 {
5594         PG_SETMASK(&UnBlockSig);
5595 }
5596
5597 #ifdef EXEC_BACKEND
5598 static pid_t
5599 bgworker_forkexec(int shmem_slot)
5600 {
5601         char       *av[10];
5602         int                     ac = 0;
5603         char            forkav[MAXPGPATH];
5604
5605         snprintf(forkav, MAXPGPATH, "--forkbgworker=%d", shmem_slot);
5606
5607         av[ac++] = "postgres";
5608         av[ac++] = forkav;
5609         av[ac++] = NULL;                        /* filled in by postmaster_forkexec */
5610         av[ac] = NULL;
5611
5612         Assert(ac < lengthof(av));
5613
5614         return postmaster_forkexec(ac, av);
5615 }
5616 #endif
5617
5618 /*
5619  * Start a new bgworker.
5620  * Starting time conditions must have been checked already.
5621  *
5622  * Returns true on success, false on failure.
5623  * In either case, update the RegisteredBgWorker's state appropriately.
5624  *
5625  * This code is heavily based on autovacuum.c, q.v.
5626  */
5627 static bool
5628 do_start_bgworker(RegisteredBgWorker *rw)
5629 {
5630         pid_t           worker_pid;
5631
5632         Assert(rw->rw_pid == 0);
5633
5634         /*
5635          * Allocate and assign the Backend element.  Note we must do this before
5636          * forking, so that we can handle out of memory properly.
5637          *
5638          * Treat failure as though the worker had crashed.  That way, the
5639          * postmaster will wait a bit before attempting to start it again; if it
5640          * tried again right away, most likely it'd find itself repeating the
5641          * out-of-memory or fork failure condition.
5642          */
5643         if (!assign_backendlist_entry(rw))
5644         {
5645                 rw->rw_crashed_at = GetCurrentTimestamp();
5646                 return false;
5647         }
5648
5649         ereport(DEBUG1,
5650                         (errmsg("starting background worker process \"%s\"",
5651                                         rw->rw_worker.bgw_name)));
5652
5653 #ifdef EXEC_BACKEND
5654         switch ((worker_pid = bgworker_forkexec(rw->rw_shmem_slot)))
5655 #else
5656         switch ((worker_pid = fork_process()))
5657 #endif
5658         {
5659                 case -1:
5660                         /* in postmaster, fork failed ... */
5661                         ereport(LOG,
5662                                         (errmsg("could not fork worker process: %m")));
5663                         /* undo what assign_backendlist_entry did */
5664                         ReleasePostmasterChildSlot(rw->rw_child_slot);
5665                         rw->rw_child_slot = 0;
5666                         free(rw->rw_backend);
5667                         rw->rw_backend = NULL;
5668                         /* mark entry as crashed, so we'll try again later */
5669                         rw->rw_crashed_at = GetCurrentTimestamp();
5670                         break;
5671
5672 #ifndef EXEC_BACKEND
5673                 case 0:
5674                         /* in postmaster child ... */
5675                         InitPostmasterChild();
5676
5677                         /* Close the postmaster's sockets */
5678                         ClosePostmasterPorts(false);
5679
5680                         /*
5681                          * Before blowing away PostmasterContext, save this bgworker's
5682                          * data where it can find it.
5683                          */
5684                         MyBgworkerEntry = (BackgroundWorker *)
5685                                 MemoryContextAlloc(TopMemoryContext, sizeof(BackgroundWorker));
5686                         memcpy(MyBgworkerEntry, &rw->rw_worker, sizeof(BackgroundWorker));
5687
5688                         /* Release postmaster's working memory context */
5689                         MemoryContextSwitchTo(TopMemoryContext);
5690                         MemoryContextDelete(PostmasterContext);
5691                         PostmasterContext = NULL;
5692
5693                         StartBackgroundWorker();
5694
5695                         exit(1);                        /* should not get here */
5696                         break;
5697 #endif
5698                 default:
5699                         /* in postmaster, fork successful ... */
5700                         rw->rw_pid = worker_pid;
5701                         rw->rw_backend->pid = rw->rw_pid;
5702                         ReportBackgroundWorkerPID(rw);
5703                         /* add new worker to lists of backends */
5704                         dlist_push_head(&BackendList, &rw->rw_backend->elem);
5705 #ifdef EXEC_BACKEND
5706                         ShmemBackendArrayAdd(rw->rw_backend);
5707 #endif
5708                         return true;
5709         }
5710
5711         return false;
5712 }
5713
5714 /*
5715  * Does the current postmaster state require starting a worker with the
5716  * specified start_time?
5717  */
5718 static bool
5719 bgworker_should_start_now(BgWorkerStartTime start_time)
5720 {
5721         switch (pmState)
5722         {
5723                 case PM_NO_CHILDREN:
5724                 case PM_WAIT_DEAD_END:
5725                 case PM_SHUTDOWN_2:
5726                 case PM_SHUTDOWN:
5727                 case PM_WAIT_BACKENDS:
5728                 case PM_WAIT_READONLY:
5729                 case PM_WAIT_BACKUP:
5730                         break;
5731
5732                 case PM_RUN:
5733                         if (start_time == BgWorkerStart_RecoveryFinished)
5734                                 return true;
5735                         /* fall through */
5736
5737                 case PM_HOT_STANDBY:
5738                         if (start_time == BgWorkerStart_ConsistentState)
5739                                 return true;
5740                         /* fall through */
5741
5742                 case PM_RECOVERY:
5743                 case PM_STARTUP:
5744                 case PM_INIT:
5745                         if (start_time == BgWorkerStart_PostmasterStart)
5746                                 return true;
5747                         /* fall through */
5748
5749         }
5750
5751         return false;
5752 }
5753
5754 /*
5755  * Allocate the Backend struct for a connected background worker, but don't
5756  * add it to the list of backends just yet.
5757  *
5758  * On failure, return false without changing any worker state.
5759  *
5760  * Some info from the Backend is copied into the passed rw.
5761  */
5762 static bool
5763 assign_backendlist_entry(RegisteredBgWorker *rw)
5764 {
5765         Backend    *bn;
5766
5767         /*
5768          * Compute the cancel key that will be assigned to this session. We
5769          * probably don't need cancel keys for background workers, but we'd better
5770          * have something random in the field to prevent unfriendly people from
5771          * sending cancels to them.
5772          */
5773         if (!RandomCancelKey(&MyCancelKey))
5774         {
5775                 ereport(LOG,
5776                                 (errcode(ERRCODE_INTERNAL_ERROR),
5777                                  errmsg("could not generate random cancel key")));
5778                 return false;
5779         }
5780
5781         bn = malloc(sizeof(Backend));
5782         if (bn == NULL)
5783         {
5784                 ereport(LOG,
5785                                 (errcode(ERRCODE_OUT_OF_MEMORY),
5786                                  errmsg("out of memory")));
5787                 return false;
5788         }
5789
5790         bn->cancel_key = MyCancelKey;
5791         bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5792         bn->bkend_type = BACKEND_TYPE_BGWORKER;
5793         bn->dead_end = false;
5794         bn->bgworker_notify = false;
5795
5796         rw->rw_backend = bn;
5797         rw->rw_child_slot = bn->child_slot;
5798
5799         return true;
5800 }
5801
5802 /*
5803  * If the time is right, start background worker(s).
5804  *
5805  * As a side effect, the bgworker control variables are set or reset
5806  * depending on whether more workers may need to be started.
5807  *
5808  * We limit the number of workers started per call, to avoid consuming the
5809  * postmaster's attention for too long when many such requests are pending.
5810  * As long as StartWorkerNeeded is true, ServerLoop will not block and will
5811  * call this function again after dealing with any other issues.
5812  */
5813 static void
5814 maybe_start_bgworkers(void)
5815 {
5816 #define MAX_BGWORKERS_TO_LAUNCH 100
5817         int                     num_launched = 0;
5818         TimestampTz now = 0;
5819         slist_mutable_iter iter;
5820
5821         /*
5822          * During crash recovery, we have no need to be called until the state
5823          * transition out of recovery.
5824          */
5825         if (FatalError)
5826         {
5827                 StartWorkerNeeded = false;
5828                 HaveCrashedWorker = false;
5829                 return;
5830         }
5831
5832         /* Don't need to be called again unless we find a reason for it below */
5833         StartWorkerNeeded = false;
5834         HaveCrashedWorker = false;
5835
5836         slist_foreach_modify(iter, &BackgroundWorkerList)
5837         {
5838                 RegisteredBgWorker *rw;
5839
5840                 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
5841
5842                 /* ignore if already running */
5843                 if (rw->rw_pid != 0)
5844                         continue;
5845
5846                 /* if marked for death, clean up and remove from list */
5847                 if (rw->rw_terminate)
5848                 {
5849                         ForgetBackgroundWorker(&iter);
5850                         continue;
5851                 }
5852
5853                 /*
5854                  * If this worker has crashed previously, maybe it needs to be
5855                  * restarted (unless on registration it specified it doesn't want to
5856                  * be restarted at all).  Check how long ago did a crash last happen.
5857                  * If the last crash is too recent, don't start it right away; let it
5858                  * be restarted once enough time has passed.
5859                  */
5860                 if (rw->rw_crashed_at != 0)
5861                 {
5862                         if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
5863                         {
5864                                 ForgetBackgroundWorker(&iter);
5865                                 continue;
5866                         }
5867
5868                         /* read system time only when needed */
5869                         if (now == 0)
5870                                 now = GetCurrentTimestamp();
5871
5872                         if (!TimestampDifferenceExceeds(rw->rw_crashed_at, now,
5873                                                                                         rw->rw_worker.bgw_restart_time * 1000))
5874                         {
5875                                 /* Set flag to remember that we have workers to start later */
5876                                 HaveCrashedWorker = true;
5877                                 continue;
5878                         }
5879                 }
5880
5881                 if (bgworker_should_start_now(rw->rw_worker.bgw_start_time))
5882                 {
5883                         /* reset crash time before trying to start worker */
5884                         rw->rw_crashed_at = 0;
5885
5886                         /*
5887                          * Try to start the worker.
5888                          *
5889                          * On failure, give up processing workers for now, but set
5890                          * StartWorkerNeeded so we'll come back here on the next iteration
5891                          * of ServerLoop to try again.  (We don't want to wait, because
5892                          * there might be additional ready-to-run workers.)  We could set
5893                          * HaveCrashedWorker as well, since this worker is now marked
5894                          * crashed, but there's no need because the next run of this
5895                          * function will do that.
5896                          */
5897                         if (!do_start_bgworker(rw))
5898                         {
5899                                 StartWorkerNeeded = true;
5900                                 return;
5901                         }
5902
5903                         /*
5904                          * If we've launched as many workers as allowed, quit, but have
5905                          * ServerLoop call us again to look for additional ready-to-run
5906                          * workers.  There might not be any, but we'll find out the next
5907                          * time we run.
5908                          */
5909                         if (++num_launched >= MAX_BGWORKERS_TO_LAUNCH)
5910                         {
5911                                 StartWorkerNeeded = true;
5912                                 return;
5913                         }
5914                 }
5915         }
5916 }
5917
5918 /*
5919  * When a backend asks to be notified about worker state changes, we
5920  * set a flag in its backend entry.  The background worker machinery needs
5921  * to know when such backends exit.
5922  */
5923 bool
5924 PostmasterMarkPIDForWorkerNotify(int pid)
5925 {
5926         dlist_iter      iter;
5927         Backend    *bp;
5928
5929         dlist_foreach(iter, &BackendList)
5930         {
5931                 bp = dlist_container(Backend, elem, iter.cur);
5932                 if (bp->pid == pid)
5933                 {
5934                         bp->bgworker_notify = true;
5935                         return true;
5936                 }
5937         }
5938         return false;
5939 }
5940
5941 #ifdef EXEC_BACKEND
5942
5943 /*
5944  * The following need to be available to the save/restore_backend_variables
5945  * functions.  They are marked NON_EXEC_STATIC in their home modules.
5946  */
5947 extern slock_t *ShmemLock;
5948 extern slock_t *ProcStructLock;
5949 extern PGPROC *AuxiliaryProcs;
5950 extern PMSignalData *PMSignalState;
5951 extern pgsocket pgStatSock;
5952 extern pg_time_t first_syslogger_file_time;
5953
5954 #ifndef WIN32
5955 #define write_inheritable_socket(dest, src, childpid) ((*(dest) = (src)), true)
5956 #define read_inheritable_socket(dest, src) (*(dest) = *(src))
5957 #else
5958 static bool write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE child);
5959 static bool write_inheritable_socket(InheritableSocket *dest, SOCKET src,
5960                                                  pid_t childPid);
5961 static void read_inheritable_socket(SOCKET *dest, InheritableSocket *src);
5962 #endif
5963
5964
5965 /* Save critical backend variables into the BackendParameters struct */
5966 #ifndef WIN32
5967 static bool
5968 save_backend_variables(BackendParameters *param, Port *port)
5969 #else
5970 static bool
5971 save_backend_variables(BackendParameters *param, Port *port,
5972                                            HANDLE childProcess, pid_t childPid)
5973 #endif
5974 {
5975         memcpy(&param->port, port, sizeof(Port));
5976         if (!write_inheritable_socket(&param->portsocket, port->sock, childPid))
5977                 return false;
5978
5979         strlcpy(param->DataDir, DataDir, MAXPGPATH);
5980
5981         memcpy(&param->ListenSocket, &ListenSocket, sizeof(ListenSocket));
5982
5983         param->MyCancelKey = MyCancelKey;
5984         param->MyPMChildSlot = MyPMChildSlot;
5985
5986         param->UsedShmemSegID = UsedShmemSegID;
5987         param->UsedShmemSegAddr = UsedShmemSegAddr;
5988
5989         param->ShmemLock = ShmemLock;
5990         param->ShmemVariableCache = ShmemVariableCache;
5991         param->ShmemBackendArray = ShmemBackendArray;
5992
5993 #ifndef HAVE_SPINLOCKS
5994         param->SpinlockSemaArray = SpinlockSemaArray;
5995 #endif
5996         param->NamedLWLockTrancheRequests = NamedLWLockTrancheRequests;
5997         param->NamedLWLockTrancheArray = NamedLWLockTrancheArray;
5998         param->MainLWLockArray = MainLWLockArray;
5999         param->ProcStructLock = ProcStructLock;
6000         param->ProcGlobal = ProcGlobal;
6001         param->AuxiliaryProcs = AuxiliaryProcs;
6002         param->PreparedXactProcs = PreparedXactProcs;
6003         param->PMSignalState = PMSignalState;
6004         if (!write_inheritable_socket(&param->pgStatSock, pgStatSock, childPid))
6005                 return false;
6006
6007         param->PostmasterPid = PostmasterPid;
6008         param->PgStartTime = PgStartTime;
6009         param->PgReloadTime = PgReloadTime;
6010         param->first_syslogger_file_time = first_syslogger_file_time;
6011
6012         param->redirection_done = redirection_done;
6013         param->IsBinaryUpgrade = IsBinaryUpgrade;
6014         param->max_safe_fds = max_safe_fds;
6015
6016         param->MaxBackends = MaxBackends;
6017
6018 #ifdef WIN32
6019         param->PostmasterHandle = PostmasterHandle;
6020         if (!write_duplicated_handle(&param->initial_signal_pipe,
6021                                                                  pgwin32_create_signal_listener(childPid),
6022                                                                  childProcess))
6023                 return false;
6024 #else
6025         memcpy(&param->postmaster_alive_fds, &postmaster_alive_fds,
6026                    sizeof(postmaster_alive_fds));
6027 #endif
6028
6029         memcpy(&param->syslogPipe, &syslogPipe, sizeof(syslogPipe));
6030
6031         strlcpy(param->my_exec_path, my_exec_path, MAXPGPATH);
6032
6033         strlcpy(param->pkglib_path, pkglib_path, MAXPGPATH);
6034
6035         strlcpy(param->ExtraOptions, ExtraOptions, MAXPGPATH);
6036
6037         return true;
6038 }
6039
6040
6041 #ifdef WIN32
6042 /*
6043  * Duplicate a handle for usage in a child process, and write the child
6044  * process instance of the handle to the parameter file.
6045  */
6046 static bool
6047 write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE childProcess)
6048 {
6049         HANDLE          hChild = INVALID_HANDLE_VALUE;
6050
6051         if (!DuplicateHandle(GetCurrentProcess(),
6052                                                  src,
6053                                                  childProcess,
6054                                                  &hChild,
6055                                                  0,
6056                                                  TRUE,
6057                                                  DUPLICATE_CLOSE_SOURCE | DUPLICATE_SAME_ACCESS))
6058         {
6059                 ereport(LOG,
6060                                 (errmsg_internal("could not duplicate handle to be written to backend parameter file: error code %lu",
6061                                                                  GetLastError())));
6062                 return false;
6063         }
6064
6065         *dest = hChild;
6066         return true;
6067 }
6068
6069 /*
6070  * Duplicate a socket for usage in a child process, and write the resulting
6071  * structure to the parameter file.
6072  * This is required because a number of LSPs (Layered Service Providers) very
6073  * common on Windows (antivirus, firewalls, download managers etc) break
6074  * straight socket inheritance.
6075  */
6076 static bool
6077 write_inheritable_socket(InheritableSocket *dest, SOCKET src, pid_t childpid)
6078 {
6079         dest->origsocket = src;
6080         if (src != 0 && src != PGINVALID_SOCKET)
6081         {
6082                 /* Actual socket */
6083                 if (WSADuplicateSocket(src, childpid, &dest->wsainfo) != 0)
6084                 {
6085                         ereport(LOG,
6086                                         (errmsg("could not duplicate socket %d for use in backend: error code %d",
6087                                                         (int) src, WSAGetLastError())));
6088                         return false;
6089                 }
6090         }
6091         return true;
6092 }
6093
6094 /*
6095  * Read a duplicate socket structure back, and get the socket descriptor.
6096  */
6097 static void
6098 read_inheritable_socket(SOCKET *dest, InheritableSocket *src)
6099 {
6100         SOCKET          s;
6101
6102         if (src->origsocket == PGINVALID_SOCKET || src->origsocket == 0)
6103         {
6104                 /* Not a real socket! */
6105                 *dest = src->origsocket;
6106         }
6107         else
6108         {
6109                 /* Actual socket, so create from structure */
6110                 s = WSASocket(FROM_PROTOCOL_INFO,
6111                                           FROM_PROTOCOL_INFO,
6112                                           FROM_PROTOCOL_INFO,
6113                                           &src->wsainfo,
6114                                           0,
6115                                           0);
6116                 if (s == INVALID_SOCKET)
6117                 {
6118                         write_stderr("could not create inherited socket: error code %d\n",
6119                                                  WSAGetLastError());
6120                         exit(1);
6121                 }
6122                 *dest = s;
6123
6124                 /*
6125                  * To make sure we don't get two references to the same socket, close
6126                  * the original one. (This would happen when inheritance actually
6127                  * works..
6128                  */
6129                 closesocket(src->origsocket);
6130         }
6131 }
6132 #endif
6133
6134 static void
6135 read_backend_variables(char *id, Port *port)
6136 {
6137         BackendParameters param;
6138
6139 #ifndef WIN32
6140         /* Non-win32 implementation reads from file */
6141         FILE       *fp;
6142
6143         /* Open file */
6144         fp = AllocateFile(id, PG_BINARY_R);
6145         if (!fp)
6146         {
6147                 write_stderr("could not open backend variables file \"%s\": %s\n",
6148                                          id, strerror(errno));
6149                 exit(1);
6150         }
6151
6152         if (fread(&param, sizeof(param), 1, fp) != 1)
6153         {
6154                 write_stderr("could not read from backend variables file \"%s\": %s\n",
6155                                          id, strerror(errno));
6156                 exit(1);
6157         }
6158
6159         /* Release file */
6160         FreeFile(fp);
6161         if (unlink(id) != 0)
6162         {
6163                 write_stderr("could not remove file \"%s\": %s\n",
6164                                          id, strerror(errno));
6165                 exit(1);
6166         }
6167 #else
6168         /* Win32 version uses mapped file */
6169         HANDLE          paramHandle;
6170         BackendParameters *paramp;
6171
6172 #ifdef _WIN64
6173         paramHandle = (HANDLE) _atoi64(id);
6174 #else
6175         paramHandle = (HANDLE) atol(id);
6176 #endif
6177         paramp = MapViewOfFile(paramHandle, FILE_MAP_READ, 0, 0, 0);
6178         if (!paramp)
6179         {
6180                 write_stderr("could not map view of backend variables: error code %lu\n",
6181                                          GetLastError());
6182                 exit(1);
6183         }
6184
6185         memcpy(&param, paramp, sizeof(BackendParameters));
6186
6187         if (!UnmapViewOfFile(paramp))
6188         {
6189                 write_stderr("could not unmap view of backend variables: error code %lu\n",
6190                                          GetLastError());
6191                 exit(1);
6192         }
6193
6194         if (!CloseHandle(paramHandle))
6195         {
6196                 write_stderr("could not close handle to backend parameter variables: error code %lu\n",
6197                                          GetLastError());
6198                 exit(1);
6199         }
6200 #endif
6201
6202         restore_backend_variables(&param, port);
6203 }
6204
6205 /* Restore critical backend variables from the BackendParameters struct */
6206 static void
6207 restore_backend_variables(BackendParameters *param, Port *port)
6208 {
6209         memcpy(port, &param->port, sizeof(Port));
6210         read_inheritable_socket(&port->sock, &param->portsocket);
6211
6212         SetDataDir(param->DataDir);
6213
6214         memcpy(&ListenSocket, &param->ListenSocket, sizeof(ListenSocket));
6215
6216         MyCancelKey = param->MyCancelKey;
6217         MyPMChildSlot = param->MyPMChildSlot;
6218
6219         UsedShmemSegID = param->UsedShmemSegID;
6220         UsedShmemSegAddr = param->UsedShmemSegAddr;
6221
6222         ShmemLock = param->ShmemLock;
6223         ShmemVariableCache = param->ShmemVariableCache;
6224         ShmemBackendArray = param->ShmemBackendArray;
6225
6226 #ifndef HAVE_SPINLOCKS
6227         SpinlockSemaArray = param->SpinlockSemaArray;
6228 #endif
6229         NamedLWLockTrancheRequests = param->NamedLWLockTrancheRequests;
6230         NamedLWLockTrancheArray = param->NamedLWLockTrancheArray;
6231         MainLWLockArray = param->MainLWLockArray;
6232         ProcStructLock = param->ProcStructLock;
6233         ProcGlobal = param->ProcGlobal;
6234         AuxiliaryProcs = param->AuxiliaryProcs;
6235         PreparedXactProcs = param->PreparedXactProcs;
6236         PMSignalState = param->PMSignalState;
6237         read_inheritable_socket(&pgStatSock, &param->pgStatSock);
6238
6239         PostmasterPid = param->PostmasterPid;
6240         PgStartTime = param->PgStartTime;
6241         PgReloadTime = param->PgReloadTime;
6242         first_syslogger_file_time = param->first_syslogger_file_time;
6243
6244         redirection_done = param->redirection_done;
6245         IsBinaryUpgrade = param->IsBinaryUpgrade;
6246         max_safe_fds = param->max_safe_fds;
6247
6248         MaxBackends = param->MaxBackends;
6249
6250 #ifdef WIN32
6251         PostmasterHandle = param->PostmasterHandle;
6252         pgwin32_initial_signal_pipe = param->initial_signal_pipe;
6253 #else
6254         memcpy(&postmaster_alive_fds, &param->postmaster_alive_fds,
6255                    sizeof(postmaster_alive_fds));
6256 #endif
6257
6258         memcpy(&syslogPipe, &param->syslogPipe, sizeof(syslogPipe));
6259
6260         strlcpy(my_exec_path, param->my_exec_path, MAXPGPATH);
6261
6262         strlcpy(pkglib_path, param->pkglib_path, MAXPGPATH);
6263
6264         strlcpy(ExtraOptions, param->ExtraOptions, MAXPGPATH);
6265 }
6266
6267
6268 Size
6269 ShmemBackendArraySize(void)
6270 {
6271         return mul_size(MaxLivePostmasterChildren(), sizeof(Backend));
6272 }
6273
6274 void
6275 ShmemBackendArrayAllocation(void)
6276 {
6277         Size            size = ShmemBackendArraySize();
6278
6279         ShmemBackendArray = (Backend *) ShmemAlloc(size);
6280         /* Mark all slots as empty */
6281         memset(ShmemBackendArray, 0, size);
6282 }
6283
6284 static void
6285 ShmemBackendArrayAdd(Backend *bn)
6286 {
6287         /* The array slot corresponding to my PMChildSlot should be free */
6288         int                     i = bn->child_slot - 1;
6289
6290         Assert(ShmemBackendArray[i].pid == 0);
6291         ShmemBackendArray[i] = *bn;
6292 }
6293
6294 static void
6295 ShmemBackendArrayRemove(Backend *bn)
6296 {
6297         int                     i = bn->child_slot - 1;
6298
6299         Assert(ShmemBackendArray[i].pid == bn->pid);
6300         /* Mark the slot as empty */
6301         ShmemBackendArray[i].pid = 0;
6302 }
6303 #endif                                                  /* EXEC_BACKEND */
6304
6305
6306 #ifdef WIN32
6307
6308 /*
6309  * Subset implementation of waitpid() for Windows.  We assume pid is -1
6310  * (that is, check all child processes) and options is WNOHANG (don't wait).
6311  */
6312 static pid_t
6313 waitpid(pid_t pid, int *exitstatus, int options)
6314 {
6315         DWORD           dwd;
6316         ULONG_PTR       key;
6317         OVERLAPPED *ovl;
6318
6319         /*
6320          * Check if there are any dead children. If there are, return the pid of
6321          * the first one that died.
6322          */
6323         if (GetQueuedCompletionStatus(win32ChildQueue, &dwd, &key, &ovl, 0))
6324         {
6325                 *exitstatus = (int) key;
6326                 return dwd;
6327         }
6328
6329         return -1;
6330 }
6331
6332 /*
6333  * Note! Code below executes on a thread pool! All operations must
6334  * be thread safe! Note that elog() and friends must *not* be used.
6335  */
6336 static void WINAPI
6337 pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired)
6338 {
6339         win32_deadchild_waitinfo *childinfo = (win32_deadchild_waitinfo *) lpParameter;
6340         DWORD           exitcode;
6341
6342         if (TimerOrWaitFired)
6343                 return;                                 /* timeout. Should never happen, since we use
6344                                                                  * INFINITE as timeout value. */
6345
6346         /*
6347          * Remove handle from wait - required even though it's set to wait only
6348          * once
6349          */
6350         UnregisterWaitEx(childinfo->waitHandle, NULL);
6351
6352         if (!GetExitCodeProcess(childinfo->procHandle, &exitcode))
6353         {
6354                 /*
6355                  * Should never happen. Inform user and set a fixed exitcode.
6356                  */
6357                 write_stderr("could not read exit code for process\n");
6358                 exitcode = 255;
6359         }
6360
6361         if (!PostQueuedCompletionStatus(win32ChildQueue, childinfo->procId, (ULONG_PTR) exitcode, NULL))
6362                 write_stderr("could not post child completion status\n");
6363
6364         /*
6365          * Handle is per-process, so we close it here instead of in the
6366          * originating thread
6367          */
6368         CloseHandle(childinfo->procHandle);
6369
6370         /*
6371          * Free struct that was allocated before the call to
6372          * RegisterWaitForSingleObject()
6373          */
6374         free(childinfo);
6375
6376         /* Queue SIGCHLD signal */
6377         pg_queue_signal(SIGCHLD);
6378 }
6379 #endif                                                  /* WIN32 */
6380
6381 /*
6382  * Initialize one and only handle for monitoring postmaster death.
6383  *
6384  * Called once in the postmaster, so that child processes can subsequently
6385  * monitor if their parent is dead.
6386  */
6387 static void
6388 InitPostmasterDeathWatchHandle(void)
6389 {
6390 #ifndef WIN32
6391
6392         /*
6393          * Create a pipe. Postmaster holds the write end of the pipe open
6394          * (POSTMASTER_FD_OWN), and children hold the read end. Children can pass
6395          * the read file descriptor to select() to wake up in case postmaster
6396          * dies, or check for postmaster death with a (read() == 0). Children must
6397          * close the write end as soon as possible after forking, because EOF
6398          * won't be signaled in the read end until all processes have closed the
6399          * write fd. That is taken care of in ClosePostmasterPorts().
6400          */
6401         Assert(MyProcPid == PostmasterPid);
6402         if (pipe(postmaster_alive_fds) < 0)
6403                 ereport(FATAL,
6404                                 (errcode_for_file_access(),
6405                                  errmsg_internal("could not create pipe to monitor postmaster death: %m")));
6406
6407         /*
6408          * Set O_NONBLOCK to allow testing for the fd's presence with a read()
6409          * call.
6410          */
6411         if (fcntl(postmaster_alive_fds[POSTMASTER_FD_WATCH], F_SETFL, O_NONBLOCK) == -1)
6412                 ereport(FATAL,
6413                                 (errcode_for_socket_access(),
6414                                  errmsg_internal("could not set postmaster death monitoring pipe to nonblocking mode: %m")));
6415 #else
6416
6417         /*
6418          * On Windows, we use a process handle for the same purpose.
6419          */
6420         if (DuplicateHandle(GetCurrentProcess(),
6421                                                 GetCurrentProcess(),
6422                                                 GetCurrentProcess(),
6423                                                 &PostmasterHandle,
6424                                                 0,
6425                                                 TRUE,
6426                                                 DUPLICATE_SAME_ACCESS) == 0)
6427                 ereport(FATAL,
6428                                 (errmsg_internal("could not duplicate postmaster handle: error code %lu",
6429                                                                  GetLastError())));
6430 #endif                                                  /* WIN32 */
6431 }