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/pqformat.h"
107 #include "libpq/pqsignal.h"
108 #include "miscadmin.h"
109 #include "pg_getopt.h"
110 #include "pgstat.h"
111 #include "postmaster/autovacuum.h"
112 #include "postmaster/bgworker_internals.h"
113 #include "postmaster/fork_process.h"
114 #include "postmaster/pgarch.h"
115 #include "postmaster/postmaster.h"
116 #include "postmaster/syslogger.h"
117 #include "replication/logicallauncher.h"
118 #include "replication/walsender.h"
119 #include "storage/fd.h"
120 #include "storage/ipc.h"
121 #include "storage/pg_shmem.h"
122 #include "storage/pmsignal.h"
123 #include "storage/proc.h"
124 #include "tcop/tcopprot.h"
125 #include "utils/builtins.h"
126 #include "utils/datetime.h"
127 #include "utils/dynamic_loader.h"
128 #include "utils/memutils.h"
129 #include "utils/pidfile.h"
130 #include "utils/ps_status.h"
131 #include "utils/timeout.h"
132 #include "utils/varlena.h"
133
134 #ifdef EXEC_BACKEND
135 #include "storage/spin.h"
136 #endif
137
138
139 /*
140 * Possible types of a backend. Beyond being the possible bkend_type values in
141 * struct bkend, these are OR-able request flag bits for SignalSomeChildren()
142 * and CountChildren().
143 */
144 #define BACKEND_TYPE_NORMAL 0x0001 /* normal backend */
145 #define BACKEND_TYPE_AUTOVAC 0x0002 /* autovacuum worker process */
146 #define BACKEND_TYPE_WALSND 0x0004 /* walsender process */
147 #define BACKEND_TYPE_BGWORKER 0x0008 /* bgworker process */
148 #define BACKEND_TYPE_ALL 0x000F /* OR of all the above */
149
150 /*
151 * List of active backends (or child processes anyway; we don't actually
152 * know whether a given child has become a backend or is still in the
153 * authorization phase). This is used mainly to keep track of how many
154 * children we have and send them appropriate signals when necessary.
155 *
156 * "Special" children such as the startup, bgwriter and autovacuum launcher
157 * tasks are not in this list. Autovacuum worker and walsender are in it.
158 * Also, "dead_end" children are in it: these are children launched just for
159 * the purpose of sending a friendly rejection message to a would-be client.
160 * We must track them because they are attached to shared memory, but we know
161 * they will never become live backends. dead_end children are not assigned a
162 * PMChildSlot.
163 *
164 * Background workers are in this list, too.
165 */
166 typedef struct bkend
167 {
168 pid_t pid; /* process id of backend */
169 int32 cancel_key; /* cancel key for cancels for this backend */
170 int child_slot; /* PMChildSlot for this backend, if any */
171
172 /*
173 * Flavor of backend or auxiliary process. Note that BACKEND_TYPE_WALSND
174 * backends initially announce themselves as BACKEND_TYPE_NORMAL, so if
175 * bkend_type is normal, you should check for a recent transition.
176 */
177 int bkend_type;
178 bool dead_end; /* is it going to send an error and quit? */
179 bool bgworker_notify; /* gets bgworker start/stop notifications */
180 dlist_node elem; /* list link in BackendList */
181 } Backend;
182
183 static dlist_head BackendList = DLIST_STATIC_INIT(BackendList);
184
185 #ifdef EXEC_BACKEND
186 static Backend *ShmemBackendArray;
187 #endif
188
189 BackgroundWorker *MyBgworkerEntry = NULL;
190
191
192
193 /* The socket number we are listening for connections on */
194 int PostPortNumber;
195
196 /* The directory names for Unix socket(s) */
197 char *Unix_socket_directories;
198
199 /* The TCP listen address(es) */
200 char *ListenAddresses;
201
202 /*
203 * ReservedBackends is the number of backends reserved for superuser use.
204 * This number is taken out of the pool size given by MaxBackends so
205 * number of backend slots available to non-superusers is
206 * (MaxBackends - ReservedBackends). Note what this really means is
207 * "if there are <= ReservedBackends connections available, only superusers
208 * can make new connections" --- pre-existing superuser connections don't
209 * count against the limit.
210 */
211 int ReservedBackends;
212
213 /* The socket(s) we're listening to. */
214 #define MAXLISTEN 64
215 static pgsocket ListenSocket[MAXLISTEN];
216
217 /*
218 * Set by the -o option
219 */
220 static char ExtraOptions[MAXPGPATH];
221
222 /*
223 * These globals control the behavior of the postmaster in case some
224 * backend dumps core. Normally, it kills all peers of the dead backend
225 * and reinitializes shared memory. By specifying -s or -n, we can have
226 * the postmaster stop (rather than kill) peers and not reinitialize
227 * shared data structures. (Reinit is currently dead code, though.)
228 */
229 static bool Reinit = true;
230 static int SendStop = false;
231
232 /* still more option variables */
233 bool EnableSSL = false;
234
235 int PreAuthDelay = 0;
236 int AuthenticationTimeout = 60;
237
238 bool log_hostname; /* for ps display and logging */
239 bool Log_connections = false;
240 bool Db_user_namespace = false;
241
242 bool enable_bonjour = false;
243 char *bonjour_name;
244 bool restart_after_crash = true;
245
246 /* PIDs of special child processes; 0 when not running */
247 static pid_t StartupPID = 0,
248 BgWriterPID = 0,
249 CheckpointerPID = 0,
250 WalWriterPID = 0,
251 WalReceiverPID = 0,
252 AutoVacPID = 0,
253 PgArchPID = 0,
254 PgStatPID = 0,
255 SysLoggerPID = 0;
256
257 /* Startup process's status */
258 typedef enum
259 {
260 STARTUP_NOT_RUNNING,
261 STARTUP_RUNNING,
262 STARTUP_SIGNALED, /* we sent it a SIGQUIT or SIGKILL */
263 STARTUP_CRASHED
264 } StartupStatusEnum;
265
266 static StartupStatusEnum StartupStatus = STARTUP_NOT_RUNNING;
267
268 /* Startup/shutdown state */
269 #define NoShutdown 0
270 #define SmartShutdown 1
271 #define FastShutdown 2
272 #define ImmediateShutdown 3
273
274 static int Shutdown = NoShutdown;
275
276 static bool FatalError = false; /* T if recovering from backend crash */
277
278 /*
279 * We use a simple state machine to control startup, shutdown, and
280 * crash recovery (which is rather like shutdown followed by startup).
281 *
282 * After doing all the postmaster initialization work, we enter PM_STARTUP
283 * state and the startup process is launched. The startup process begins by
284 * reading the control file and other preliminary initialization steps.
285 * In a normal startup, or after crash recovery, the startup process exits
286 * with exit code 0 and we switch to PM_RUN state. However, archive recovery
287 * is handled specially since it takes much longer and we would like to support
288 * hot standby during archive recovery.
289 *
290 * When the startup process is ready to start archive recovery, it signals the
291 * postmaster, and we switch to PM_RECOVERY state. The background writer and
292 * checkpointer are launched, while the startup process continues applying WAL.
293 * If Hot Standby is enabled, then, after reaching a consistent point in WAL
294 * redo, startup process signals us again, and we switch to PM_HOT_STANDBY
295 * state and begin accepting connections to perform read-only queries. When
296 * archive recovery is finished, the startup process exits with exit code 0
297 * and we switch to PM_RUN state.
298 *
299 * Normal child backends can only be launched when we are in PM_RUN or
300 * PM_HOT_STANDBY state. (connsAllowed can also restrict launching.)
301 * In other states we handle connection requests by launching "dead_end"
302 * child processes, which will simply send the client an error message and
303 * quit. (We track these in the BackendList so that we can know when they
304 * are all gone; this is important because they're still connected to shared
305 * memory, and would interfere with an attempt to destroy the shmem segment,
306 * possibly leading to SHMALL failure when we try to make a new one.)
307 * In PM_WAIT_DEAD_END state we are waiting for all the dead_end children
308 * to drain out of the system, and therefore stop accepting connection
309 * requests at all until the last existing child has quit (which hopefully
310 * will not be very long).
311 *
312 * Notice that this state variable does not distinguish *why* we entered
313 * states later than PM_RUN --- Shutdown and FatalError must be consulted
314 * to find that out. FatalError is never true in PM_RECOVERY, PM_HOT_STANDBY,
315 * or PM_RUN states, nor in PM_SHUTDOWN states (because we don't enter those
316 * states when trying to recover from a crash). It can be true in PM_STARTUP
317 * state, because we don't clear it until we've successfully started WAL redo.
318 */
319 typedef enum
320 {
321 PM_INIT, /* postmaster starting */
322 PM_STARTUP, /* waiting for startup subprocess */
323 PM_RECOVERY, /* in archive recovery mode */
324 PM_HOT_STANDBY, /* in hot standby mode */
325 PM_RUN, /* normal "database is alive" state */
326 PM_STOP_BACKENDS, /* need to stop remaining backends */
327 PM_WAIT_BACKENDS, /* waiting for live backends to exit */
328 PM_SHUTDOWN, /* waiting for checkpointer to do shutdown
329 * ckpt */
330 PM_SHUTDOWN_2, /* waiting for archiver and walsenders to
331 * finish */
332 PM_WAIT_DEAD_END, /* waiting for dead_end children to exit */
333 PM_NO_CHILDREN /* all important children have exited */
334 } PMState;
335
336 static PMState pmState = PM_INIT;
337
338 /*
339 * While performing a "smart shutdown", we restrict new connections but stay
340 * in PM_RUN or PM_HOT_STANDBY state until all the client backends are gone.
341 * connsAllowed is a sub-state indicator showing the active restriction.
342 * It is of no interest unless pmState is PM_RUN or PM_HOT_STANDBY.
343 */
344 typedef enum
345 {
346 ALLOW_ALL_CONNS, /* normal not-shutting-down state */
347 ALLOW_SUPERUSER_CONNS, /* only superusers can connect */
348 ALLOW_NO_CONNS /* no new connections allowed, period */
349 } ConnsAllowedState;
350
351 static ConnsAllowedState connsAllowed = ALLOW_ALL_CONNS;
352
353 /* Start time of SIGKILL timeout during immediate shutdown or child crash */
354 /* Zero means timeout is not running */
355 static time_t AbortStartTime = 0;
356
357 /* Length of said timeout */
358 #define SIGKILL_CHILDREN_AFTER_SECS 5
359
360 static bool ReachedNormalRunning = false; /* T if we've reached PM_RUN */
361
362 bool ClientAuthInProgress = false; /* T during new-client
363 * authentication */
364
365 bool redirection_done = false; /* stderr redirected for syslogger? */
366
367 /* received START_AUTOVAC_LAUNCHER signal */
368 static volatile sig_atomic_t start_autovac_launcher = false;
369
370 /* the launcher needs to be signalled to communicate some condition */
371 static volatile bool avlauncher_needs_signal = false;
372
373 /* received START_WALRECEIVER signal */
374 static volatile sig_atomic_t WalReceiverRequested = false;
375
376 /* set when there's a worker that needs to be started up */
377 static volatile bool StartWorkerNeeded = true;
378 static volatile bool HaveCrashedWorker = false;
379
380 #ifndef HAVE_STRONG_RANDOM
381 /*
382 * State for assigning cancel keys.
383 * Also, the global MyCancelKey passes the cancel key assigned to a given
384 * backend from the postmaster to that backend (via fork).
385 */
386 static unsigned int random_seed = 0;
387 static struct timeval random_start_time;
388 #endif
389
390 #ifdef USE_SSL
391 /* Set when and if SSL has been initialized properly */
392 static bool LoadedSSL = false;
393 #endif
394
395 #ifdef USE_BONJOUR
396 static DNSServiceRef bonjour_sdref = NULL;
397 #endif
398
399 /*
400 * postmaster.c - function prototypes
401 */
402 static void CloseServerPorts(int status, Datum arg);
403 static void unlink_external_pid_file(int status, Datum arg);
404 static void getInstallationPaths(const char *argv0);
405 static void checkDataDir(void);
406 static Port *ConnCreate(int serverFd);
407 static void ConnFree(Port *port);
408 static void reset_shared(int port);
409 static void SIGHUP_handler(SIGNAL_ARGS);
410 static void pmdie(SIGNAL_ARGS);
411 static void reaper(SIGNAL_ARGS);
412 static void sigusr1_handler(SIGNAL_ARGS);
413 static void process_startup_packet_die(SIGNAL_ARGS);
414 static void process_startup_packet_quickdie(SIGNAL_ARGS);
415 static void dummy_handler(SIGNAL_ARGS);
416 static void StartupPacketTimeoutHandler(void);
417 static void CleanupBackend(int pid, int exitstatus);
418 static bool CleanupBackgroundWorker(int pid, int exitstatus);
419 static void HandleChildCrash(int pid, int exitstatus, const char *procname);
420 static void LogChildExit(int lev, const char *procname,
421 int pid, int exitstatus);
422 static void PostmasterStateMachine(void);
423 static void BackendInitialize(Port *port);
424 static void BackendRun(Port *port) pg_attribute_noreturn();
425 static void ExitPostmaster(int status) pg_attribute_noreturn();
426 static int ServerLoop(void);
427 static int BackendStartup(Port *port);
428 static int ProcessStartupPacket(Port *port, bool SSLdone);
429 static void SendNegotiateProtocolVersion(List *unrecognized_protocol_options);
430 static void processCancelRequest(Port *port, void *pkt);
431 static int initMasks(fd_set *rmask);
432 static void report_fork_failure_to_client(Port *port, int errnum);
433 static CAC_state canAcceptConnections(int backend_type);
434 static bool RandomCancelKey(int32 *cancel_key);
435 static void signal_child(pid_t pid, int signal);
436 static bool SignalSomeChildren(int signal, int targets);
437 static void TerminateChildren(int signal);
438
439 #define SignalChildren(sig) SignalSomeChildren(sig, BACKEND_TYPE_ALL)
440
441 static int CountChildren(int target);
442 static bool assign_backendlist_entry(RegisteredBgWorker *rw);
443 static void maybe_start_bgworkers(void);
444 static bool CreateOptsFile(int argc, char *argv[], char *fullprogname);
445 static pid_t StartChildProcess(AuxProcType type);
446 static void StartAutovacuumWorker(void);
447 static void MaybeStartWalReceiver(void);
448 static void InitPostmasterDeathWatchHandle(void);
449
450 /*
451 * Archiver is allowed to start up at the current postmaster state?
452 *
453 * If WAL archiving is enabled always, we are allowed to start archiver
454 * even during recovery.
455 */
456 #define PgArchStartupAllowed() \
457 ((XLogArchivingActive() && pmState == PM_RUN) || \
458 (XLogArchivingAlways() && \
459 (pmState == PM_RECOVERY || pmState == PM_HOT_STANDBY)))
460
461 #ifdef EXEC_BACKEND
462
463 #ifdef WIN32
464 #define WNOHANG 0 /* ignored, so any integer value will do */
465
466 static pid_t waitpid(pid_t pid, int *exitstatus, int options);
467 static void WINAPI pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired);
468
469 static HANDLE win32ChildQueue;
470
471 typedef struct
472 {
473 HANDLE waitHandle;
474 HANDLE procHandle;
475 DWORD procId;
476 } win32_deadchild_waitinfo;
477 #endif /* WIN32 */
478
479 static pid_t backend_forkexec(Port *port);
480 static pid_t internal_forkexec(int argc, char *argv[], Port *port);
481
482 /* Type for a socket that can be inherited to a client process */
483 #ifdef WIN32
484 typedef struct
485 {
486 SOCKET origsocket; /* Original socket value, or PGINVALID_SOCKET
487 * if not a socket */
488 WSAPROTOCOL_INFO wsainfo;
489 } InheritableSocket;
490 #else
491 typedef int InheritableSocket;
492 #endif
493
494 /*
495 * Structure contains all variables passed to exec:ed backends
496 */
497 typedef struct
498 {
499 Port port;
500 InheritableSocket portsocket;
501 char DataDir[MAXPGPATH];
502 pgsocket ListenSocket[MAXLISTEN];
503 int32 MyCancelKey;
504 int MyPMChildSlot;
505 #ifndef WIN32
506 unsigned long UsedShmemSegID;
507 #else
508 void *ShmemProtectiveRegion;
509 HANDLE UsedShmemSegID;
510 #endif
511 void *UsedShmemSegAddr;
512 slock_t *ShmemLock;
513 VariableCache ShmemVariableCache;
514 Backend *ShmemBackendArray;
515 #ifndef HAVE_SPINLOCKS
516 PGSemaphore *SpinlockSemaArray;
517 #endif
518 int NamedLWLockTrancheRequests;
519 NamedLWLockTranche *NamedLWLockTrancheArray;
520 LWLockPadded *MainLWLockArray;
521 slock_t *ProcStructLock;
522 PROC_HDR *ProcGlobal;
523 PGPROC *AuxiliaryProcs;
524 PGPROC *PreparedXactProcs;
525 PMSignalData *PMSignalState;
526 InheritableSocket pgStatSock;
527 pid_t PostmasterPid;
528 TimestampTz PgStartTime;
529 TimestampTz PgReloadTime;
530 pg_time_t first_syslogger_file_time;
531 bool redirection_done;
532 bool IsBinaryUpgrade;
533 int max_safe_fds;
534 int MaxBackends;
535 #ifdef WIN32
536 HANDLE PostmasterHandle;
537 HANDLE initial_signal_pipe;
538 HANDLE syslogPipe[2];
539 #else
540 int postmaster_alive_fds[2];
541 int syslogPipe[2];
542 #endif
543 char my_exec_path[MAXPGPATH];
544 char pkglib_path[MAXPGPATH];
545 char ExtraOptions[MAXPGPATH];
546 } BackendParameters;
547
548 static void read_backend_variables(char *id, Port *port);
549 static void restore_backend_variables(BackendParameters *param, Port *port);
550
551 #ifndef WIN32
552 static bool save_backend_variables(BackendParameters *param, Port *port);
553 #else
554 static bool save_backend_variables(BackendParameters *param, Port *port,
555 HANDLE childProcess, pid_t childPid);
556 #endif
557
558 static void ShmemBackendArrayAdd(Backend *bn);
559 static void ShmemBackendArrayRemove(Backend *bn);
560 #endif /* EXEC_BACKEND */
561
562 #define StartupDataBase() StartChildProcess(StartupProcess)
563 #define StartBackgroundWriter() StartChildProcess(BgWriterProcess)
564 #define StartCheckpointer() StartChildProcess(CheckpointerProcess)
565 #define StartWalWriter() StartChildProcess(WalWriterProcess)
566 #define StartWalReceiver() StartChildProcess(WalReceiverProcess)
567
568 /* Macros to check exit status of a child process */
569 #define EXIT_STATUS_0(st) ((st) == 0)
570 #define EXIT_STATUS_1(st) (WIFEXITED(st) && WEXITSTATUS(st) == 1)
571 #define EXIT_STATUS_3(st) (WIFEXITED(st) && WEXITSTATUS(st) == 3)
572
573 #ifndef WIN32
574 /*
575 * File descriptors for pipe used to monitor if postmaster is alive.
576 * First is POSTMASTER_FD_WATCH, second is POSTMASTER_FD_OWN.
577 */
578 int postmaster_alive_fds[2] = {-1, -1};
579 #else
580 /* Process handle of postmaster used for the same purpose on Windows */
581 HANDLE PostmasterHandle;
582 #endif
583
584 /*
585 * Postmaster main entry point
586 */
587 void
PostmasterMain(int argc,char * argv[])588 PostmasterMain(int argc, char *argv[])
589 {
590 int opt;
591 int status;
592 char *userDoption = NULL;
593 bool listen_addr_saved = false;
594 int i;
595 char *output_config_variable = NULL;
596
597 MyProcPid = PostmasterPid = getpid();
598
599 MyStartTime = time(NULL);
600
601 IsPostmasterEnvironment = true;
602
603 /*
604 * for security, no dir or file created can be group or other accessible
605 */
606 umask(S_IRWXG | S_IRWXO);
607
608 /*
609 * Initialize random(3) so we don't get the same values in every run.
610 *
611 * Note: the seed is pretty predictable from externally-visible facts such
612 * as postmaster start time, so avoid using random() for security-critical
613 * random values during postmaster startup. At the time of first
614 * connection, PostmasterRandom will select a hopefully-more-random seed.
615 */
616 srandom((unsigned int) (MyProcPid ^ MyStartTime));
617
618 /*
619 * By default, palloc() requests in the postmaster will be allocated in
620 * the PostmasterContext, which is space that can be recycled by backends.
621 * Allocated data that needs to be available to backends should be
622 * allocated in TopMemoryContext.
623 */
624 PostmasterContext = AllocSetContextCreate(TopMemoryContext,
625 "Postmaster",
626 ALLOCSET_DEFAULT_SIZES);
627 MemoryContextSwitchTo(PostmasterContext);
628
629 /* Initialize paths to installation files */
630 getInstallationPaths(argv[0]);
631
632 /*
633 * Set up signal handlers for the postmaster process.
634 *
635 * In the postmaster, we use pqsignal_pm() rather than pqsignal() (which
636 * is used by all child processes and client processes). That has a
637 * couple of special behaviors:
638 *
639 * 1. Except on Windows, we tell sigaction() to block all signals for the
640 * duration of the signal handler. This is faster than our old approach
641 * of blocking/unblocking explicitly in the signal handler, and it should
642 * also prevent excessive stack consumption if signals arrive quickly.
643 *
644 * 2. We do not set the SA_RESTART flag. This is because signals will be
645 * blocked at all times except when ServerLoop is waiting for something to
646 * happen, and during that window, we want signals to exit the select(2)
647 * wait so that ServerLoop can respond if anything interesting happened.
648 * On some platforms, signals marked SA_RESTART would not cause the
649 * select() wait to end.
650 *
651 * Child processes will generally want SA_RESTART, so pqsignal() sets that
652 * flag. We expect children to set up their own handlers before
653 * unblocking signals.
654 *
655 * CAUTION: when changing this list, check for side-effects on the signal
656 * handling setup of child processes. See tcop/postgres.c,
657 * bootstrap/bootstrap.c, postmaster/bgwriter.c, postmaster/walwriter.c,
658 * postmaster/autovacuum.c, postmaster/pgarch.c, postmaster/pgstat.c,
659 * postmaster/syslogger.c, postmaster/bgworker.c and
660 * postmaster/checkpointer.c.
661 */
662 pqinitmask();
663 PG_SETMASK(&BlockSig);
664
665 pqsignal_pm(SIGHUP, SIGHUP_handler); /* reread config file and have
666 * children do same */
667 pqsignal_pm(SIGINT, pmdie); /* send SIGTERM and shut down */
668 pqsignal_pm(SIGQUIT, pmdie); /* send SIGQUIT and die */
669 pqsignal_pm(SIGTERM, pmdie); /* wait for children and shut down */
670 pqsignal_pm(SIGALRM, SIG_IGN); /* ignored */
671 pqsignal_pm(SIGPIPE, SIG_IGN); /* ignored */
672 pqsignal_pm(SIGUSR1, sigusr1_handler); /* message from child process */
673 pqsignal_pm(SIGUSR2, dummy_handler); /* unused, reserve for children */
674 pqsignal_pm(SIGCHLD, reaper); /* handle child termination */
675 pqsignal_pm(SIGTTIN, SIG_IGN); /* ignored */
676 pqsignal_pm(SIGTTOU, SIG_IGN); /* ignored */
677 /* ignore SIGXFSZ, so that ulimit violations work like disk full */
678 #ifdef SIGXFSZ
679 pqsignal_pm(SIGXFSZ, SIG_IGN); /* ignored */
680 #endif
681
682 /*
683 * Options setup
684 */
685 InitializeGUCOptions();
686
687 opterr = 1;
688
689 /*
690 * Parse command-line options. CAUTION: keep this in sync with
691 * tcop/postgres.c (the option sets should not conflict) and with the
692 * common help() function in main/main.c.
693 */
694 while ((opt = getopt(argc, argv, "B:bc:C:D:d:EeFf:h:ijk:lN:nOo:Pp:r:S:sTt:W:-:")) != -1)
695 {
696 switch (opt)
697 {
698 case 'B':
699 SetConfigOption("shared_buffers", optarg, PGC_POSTMASTER, PGC_S_ARGV);
700 break;
701
702 case 'b':
703 /* Undocumented flag used for binary upgrades */
704 IsBinaryUpgrade = true;
705 break;
706
707 case 'C':
708 output_config_variable = strdup(optarg);
709 break;
710
711 case 'D':
712 userDoption = strdup(optarg);
713 break;
714
715 case 'd':
716 set_debug_options(atoi(optarg), PGC_POSTMASTER, PGC_S_ARGV);
717 break;
718
719 case 'E':
720 SetConfigOption("log_statement", "all", PGC_POSTMASTER, PGC_S_ARGV);
721 break;
722
723 case 'e':
724 SetConfigOption("datestyle", "euro", PGC_POSTMASTER, PGC_S_ARGV);
725 break;
726
727 case 'F':
728 SetConfigOption("fsync", "false", PGC_POSTMASTER, PGC_S_ARGV);
729 break;
730
731 case 'f':
732 if (!set_plan_disabling_options(optarg, PGC_POSTMASTER, PGC_S_ARGV))
733 {
734 write_stderr("%s: invalid argument for option -f: \"%s\"\n",
735 progname, optarg);
736 ExitPostmaster(1);
737 }
738 break;
739
740 case 'h':
741 SetConfigOption("listen_addresses", optarg, PGC_POSTMASTER, PGC_S_ARGV);
742 break;
743
744 case 'i':
745 SetConfigOption("listen_addresses", "*", PGC_POSTMASTER, PGC_S_ARGV);
746 break;
747
748 case 'j':
749 /* only used by interactive backend */
750 break;
751
752 case 'k':
753 SetConfigOption("unix_socket_directories", optarg, PGC_POSTMASTER, PGC_S_ARGV);
754 break;
755
756 case 'l':
757 SetConfigOption("ssl", "true", PGC_POSTMASTER, PGC_S_ARGV);
758 break;
759
760 case 'N':
761 SetConfigOption("max_connections", optarg, PGC_POSTMASTER, PGC_S_ARGV);
762 break;
763
764 case 'n':
765 /* Don't reinit shared mem after abnormal exit */
766 Reinit = false;
767 break;
768
769 case 'O':
770 SetConfigOption("allow_system_table_mods", "true", PGC_POSTMASTER, PGC_S_ARGV);
771 break;
772
773 case 'o':
774 /* Other options to pass to the backend on the command line */
775 snprintf(ExtraOptions + strlen(ExtraOptions),
776 sizeof(ExtraOptions) - strlen(ExtraOptions),
777 " %s", optarg);
778 break;
779
780 case 'P':
781 SetConfigOption("ignore_system_indexes", "true", PGC_POSTMASTER, PGC_S_ARGV);
782 break;
783
784 case 'p':
785 SetConfigOption("port", optarg, PGC_POSTMASTER, PGC_S_ARGV);
786 break;
787
788 case 'r':
789 /* only used by single-user backend */
790 break;
791
792 case 'S':
793 SetConfigOption("work_mem", optarg, PGC_POSTMASTER, PGC_S_ARGV);
794 break;
795
796 case 's':
797 SetConfigOption("log_statement_stats", "true", PGC_POSTMASTER, PGC_S_ARGV);
798 break;
799
800 case 'T':
801
802 /*
803 * In the event that some backend dumps core, send SIGSTOP,
804 * rather than SIGQUIT, to all its peers. This lets the wily
805 * post_hacker collect core dumps from everyone.
806 */
807 SendStop = true;
808 break;
809
810 case 't':
811 {
812 const char *tmp = get_stats_option_name(optarg);
813
814 if (tmp)
815 {
816 SetConfigOption(tmp, "true", PGC_POSTMASTER, PGC_S_ARGV);
817 }
818 else
819 {
820 write_stderr("%s: invalid argument for option -t: \"%s\"\n",
821 progname, optarg);
822 ExitPostmaster(1);
823 }
824 break;
825 }
826
827 case 'W':
828 SetConfigOption("post_auth_delay", optarg, PGC_POSTMASTER, PGC_S_ARGV);
829 break;
830
831 case 'c':
832 case '-':
833 {
834 char *name,
835 *value;
836
837 ParseLongOption(optarg, &name, &value);
838 if (!value)
839 {
840 if (opt == '-')
841 ereport(ERROR,
842 (errcode(ERRCODE_SYNTAX_ERROR),
843 errmsg("--%s requires a value",
844 optarg)));
845 else
846 ereport(ERROR,
847 (errcode(ERRCODE_SYNTAX_ERROR),
848 errmsg("-c %s requires a value",
849 optarg)));
850 }
851
852 SetConfigOption(name, value, PGC_POSTMASTER, PGC_S_ARGV);
853 free(name);
854 if (value)
855 free(value);
856 break;
857 }
858
859 default:
860 write_stderr("Try \"%s --help\" for more information.\n",
861 progname);
862 ExitPostmaster(1);
863 }
864 }
865
866 /*
867 * Postmaster accepts no non-option switch arguments.
868 */
869 if (optind < argc)
870 {
871 write_stderr("%s: invalid argument: \"%s\"\n",
872 progname, argv[optind]);
873 write_stderr("Try \"%s --help\" for more information.\n",
874 progname);
875 ExitPostmaster(1);
876 }
877
878 /*
879 * Locate the proper configuration files and data directory, and read
880 * postgresql.conf for the first time.
881 */
882 if (!SelectConfigFiles(userDoption, progname))
883 ExitPostmaster(2);
884
885 if (output_config_variable != NULL)
886 {
887 /*
888 * "-C guc" was specified, so print GUC's value and exit. No extra
889 * permission check is needed because the user is reading inside the
890 * data dir.
891 */
892 const char *config_val = GetConfigOption(output_config_variable,
893 false, false);
894
895 puts(config_val ? config_val : "");
896 ExitPostmaster(0);
897 }
898
899 /* Verify that DataDir looks reasonable */
900 checkDataDir();
901
902 /* And switch working directory into it */
903 ChangeToDataDir();
904
905 /*
906 * Check for invalid combinations of GUC settings.
907 */
908 if (ReservedBackends >= MaxConnections)
909 {
910 write_stderr("%s: superuser_reserved_connections must be less than max_connections\n", progname);
911 ExitPostmaster(1);
912 }
913 if (max_wal_senders >= MaxConnections)
914 {
915 write_stderr("%s: max_wal_senders must be less than max_connections\n", progname);
916 ExitPostmaster(1);
917 }
918 if (XLogArchiveMode > ARCHIVE_MODE_OFF && wal_level == WAL_LEVEL_MINIMAL)
919 ereport(ERROR,
920 (errmsg("WAL archival cannot be enabled when wal_level is \"minimal\"")));
921 if (max_wal_senders > 0 && wal_level == WAL_LEVEL_MINIMAL)
922 ereport(ERROR,
923 (errmsg("WAL streaming (max_wal_senders > 0) requires wal_level \"replica\" or \"logical\"")));
924
925 /*
926 * Other one-time internal sanity checks can go here, if they are fast.
927 * (Put any slow processing further down, after postmaster.pid creation.)
928 */
929 if (!CheckDateTokenTables())
930 {
931 write_stderr("%s: invalid datetoken tables, please fix\n", progname);
932 ExitPostmaster(1);
933 }
934
935 /*
936 * Now that we are done processing the postmaster arguments, reset
937 * getopt(3) library so that it will work correctly in subprocesses.
938 */
939 optind = 1;
940 #ifdef HAVE_INT_OPTRESET
941 optreset = 1; /* some systems need this too */
942 #endif
943
944 /* For debugging: display postmaster environment */
945 {
946 extern char **environ;
947 char **p;
948
949 ereport(DEBUG3,
950 (errmsg_internal("%s: PostmasterMain: initial environment dump:",
951 progname)));
952 ereport(DEBUG3,
953 (errmsg_internal("-----------------------------------------")));
954 for (p = environ; *p; ++p)
955 ereport(DEBUG3,
956 (errmsg_internal("\t%s", *p)));
957 ereport(DEBUG3,
958 (errmsg_internal("-----------------------------------------")));
959 }
960
961 /*
962 * Create lockfile for data directory.
963 *
964 * We want to do this before we try to grab the input sockets, because the
965 * data directory interlock is more reliable than the socket-file
966 * interlock (thanks to whoever decided to put socket files in /tmp :-().
967 * For the same reason, it's best to grab the TCP socket(s) before the
968 * Unix socket(s).
969 *
970 * Also note that this internally sets up the on_proc_exit function that
971 * is responsible for removing both data directory and socket lockfiles;
972 * so it must happen before opening sockets so that at exit, the socket
973 * lockfiles go away after CloseServerPorts runs.
974 */
975 CreateDataDirLockFile(true);
976
977 /*
978 * Initialize SSL library, if specified.
979 */
980 #ifdef USE_SSL
981 if (EnableSSL)
982 {
983 (void) secure_initialize(true);
984 LoadedSSL = true;
985 }
986 #endif
987
988 /*
989 * Register the apply launcher. Since it registers a background worker,
990 * it needs to be called before InitializeMaxBackends(), and it's probably
991 * a good idea to call it before any modules had chance to take the
992 * background worker slots.
993 */
994 ApplyLauncherRegister();
995
996 /*
997 * process any libraries that should be preloaded at postmaster start
998 */
999 process_shared_preload_libraries();
1000
1001 /*
1002 * Now that loadable modules have had their chance to register background
1003 * workers, calculate MaxBackends.
1004 */
1005 InitializeMaxBackends();
1006
1007 /*
1008 * Establish input sockets.
1009 *
1010 * First, mark them all closed, and set up an on_proc_exit function that's
1011 * charged with closing the sockets again at postmaster shutdown.
1012 */
1013 for (i = 0; i < MAXLISTEN; i++)
1014 ListenSocket[i] = PGINVALID_SOCKET;
1015
1016 on_proc_exit(CloseServerPorts, 0);
1017
1018 if (ListenAddresses)
1019 {
1020 char *rawstring;
1021 List *elemlist;
1022 ListCell *l;
1023 int success = 0;
1024
1025 /* Need a modifiable copy of ListenAddresses */
1026 rawstring = pstrdup(ListenAddresses);
1027
1028 /* Parse string into list of hostnames */
1029 if (!SplitIdentifierString(rawstring, ',', &elemlist))
1030 {
1031 /* syntax error in list */
1032 ereport(FATAL,
1033 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1034 errmsg("invalid list syntax in parameter \"%s\"",
1035 "listen_addresses")));
1036 }
1037
1038 foreach(l, elemlist)
1039 {
1040 char *curhost = (char *) lfirst(l);
1041
1042 if (strcmp(curhost, "*") == 0)
1043 status = StreamServerPort(AF_UNSPEC, NULL,
1044 (unsigned short) PostPortNumber,
1045 NULL,
1046 ListenSocket, MAXLISTEN);
1047 else
1048 status = StreamServerPort(AF_UNSPEC, curhost,
1049 (unsigned short) PostPortNumber,
1050 NULL,
1051 ListenSocket, MAXLISTEN);
1052
1053 if (status == STATUS_OK)
1054 {
1055 success++;
1056 /* record the first successful host addr in lockfile */
1057 if (!listen_addr_saved)
1058 {
1059 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, curhost);
1060 listen_addr_saved = true;
1061 }
1062 }
1063 else
1064 ereport(WARNING,
1065 (errmsg("could not create listen socket for \"%s\"",
1066 curhost)));
1067 }
1068
1069 if (!success && elemlist != NIL)
1070 ereport(FATAL,
1071 (errmsg("could not create any TCP/IP sockets")));
1072
1073 list_free(elemlist);
1074 pfree(rawstring);
1075 }
1076
1077 #ifdef USE_BONJOUR
1078 /* Register for Bonjour only if we opened TCP socket(s) */
1079 if (enable_bonjour && ListenSocket[0] != PGINVALID_SOCKET)
1080 {
1081 DNSServiceErrorType err;
1082
1083 /*
1084 * We pass 0 for interface_index, which will result in registering on
1085 * all "applicable" interfaces. It's not entirely clear from the
1086 * DNS-SD docs whether this would be appropriate if we have bound to
1087 * just a subset of the available network interfaces.
1088 */
1089 err = DNSServiceRegister(&bonjour_sdref,
1090 0,
1091 0,
1092 bonjour_name,
1093 "_postgresql._tcp.",
1094 NULL,
1095 NULL,
1096 htons(PostPortNumber),
1097 0,
1098 NULL,
1099 NULL,
1100 NULL);
1101 if (err != kDNSServiceErr_NoError)
1102 elog(LOG, "DNSServiceRegister() failed: error code %ld",
1103 (long) err);
1104
1105 /*
1106 * We don't bother to read the mDNS daemon's reply, and we expect that
1107 * it will automatically terminate our registration when the socket is
1108 * closed at postmaster termination. So there's nothing more to be
1109 * done here. However, the bonjour_sdref is kept around so that
1110 * forked children can close their copies of the socket.
1111 */
1112 }
1113 #endif
1114
1115 #ifdef HAVE_UNIX_SOCKETS
1116 if (Unix_socket_directories)
1117 {
1118 char *rawstring;
1119 List *elemlist;
1120 ListCell *l;
1121 int success = 0;
1122
1123 /* Need a modifiable copy of Unix_socket_directories */
1124 rawstring = pstrdup(Unix_socket_directories);
1125
1126 /* Parse string into list of directories */
1127 if (!SplitDirectoriesString(rawstring, ',', &elemlist))
1128 {
1129 /* syntax error in list */
1130 ereport(FATAL,
1131 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1132 errmsg("invalid list syntax in parameter \"%s\"",
1133 "unix_socket_directories")));
1134 }
1135
1136 foreach(l, elemlist)
1137 {
1138 char *socketdir = (char *) lfirst(l);
1139
1140 status = StreamServerPort(AF_UNIX, NULL,
1141 (unsigned short) PostPortNumber,
1142 socketdir,
1143 ListenSocket, MAXLISTEN);
1144
1145 if (status == STATUS_OK)
1146 {
1147 success++;
1148 /* record the first successful Unix socket in lockfile */
1149 if (success == 1)
1150 AddToDataDirLockFile(LOCK_FILE_LINE_SOCKET_DIR, socketdir);
1151 }
1152 else
1153 ereport(WARNING,
1154 (errmsg("could not create Unix-domain socket in directory \"%s\"",
1155 socketdir)));
1156 }
1157
1158 if (!success && elemlist != NIL)
1159 ereport(FATAL,
1160 (errmsg("could not create any Unix-domain sockets")));
1161
1162 list_free_deep(elemlist);
1163 pfree(rawstring);
1164 }
1165 #endif
1166
1167 /*
1168 * check that we have some socket to listen on
1169 */
1170 if (ListenSocket[0] == PGINVALID_SOCKET)
1171 ereport(FATAL,
1172 (errmsg("no socket created for listening")));
1173
1174 /*
1175 * If no valid TCP ports, write an empty line for listen address,
1176 * indicating the Unix socket must be used. Note that this line is not
1177 * added to the lock file until there is a socket backing it.
1178 */
1179 if (!listen_addr_saved)
1180 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, "");
1181
1182 /*
1183 * Set up shared memory and semaphores.
1184 */
1185 reset_shared(PostPortNumber);
1186
1187 /*
1188 * Estimate number of openable files. This must happen after setting up
1189 * semaphores, because on some platforms semaphores count as open files.
1190 */
1191 set_max_safe_fds();
1192
1193 /*
1194 * Set reference point for stack-depth checking.
1195 */
1196 set_stack_base();
1197
1198 /*
1199 * Initialize pipe (or process handle on Windows) that allows children to
1200 * wake up from sleep on postmaster death.
1201 */
1202 InitPostmasterDeathWatchHandle();
1203
1204 #ifdef WIN32
1205
1206 /*
1207 * Initialize I/O completion port used to deliver list of dead children.
1208 */
1209 win32ChildQueue = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 1);
1210 if (win32ChildQueue == NULL)
1211 ereport(FATAL,
1212 (errmsg("could not create I/O completion port for child queue")));
1213 #endif
1214
1215 /*
1216 * Record postmaster options. We delay this till now to avoid recording
1217 * bogus options (eg, NBuffers too high for available memory).
1218 */
1219 if (!CreateOptsFile(argc, argv, my_exec_path))
1220 ExitPostmaster(1);
1221
1222 #ifdef EXEC_BACKEND
1223 /* Write out nondefault GUC settings for child processes to use */
1224 write_nondefault_variables(PGC_POSTMASTER);
1225 #endif
1226
1227 /*
1228 * Write the external PID file if requested
1229 */
1230 if (external_pid_file)
1231 {
1232 FILE *fpidfile = fopen(external_pid_file, "w");
1233
1234 if (fpidfile)
1235 {
1236 fprintf(fpidfile, "%d\n", MyProcPid);
1237 fclose(fpidfile);
1238
1239 /* Make PID file world readable */
1240 if (chmod(external_pid_file, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH) != 0)
1241 write_stderr("%s: could not change permissions of external PID file \"%s\": %s\n",
1242 progname, external_pid_file, strerror(errno));
1243 }
1244 else
1245 write_stderr("%s: could not write external PID file \"%s\": %s\n",
1246 progname, external_pid_file, strerror(errno));
1247
1248 on_proc_exit(unlink_external_pid_file, 0);
1249 }
1250
1251 /*
1252 * Remove old temporary files. At this point there can be no other
1253 * Postgres processes running in this directory, so this should be safe.
1254 */
1255 RemovePgTempFiles();
1256
1257 /*
1258 * Forcibly remove the files signaling a standby promotion request.
1259 * Otherwise, the existence of those files triggers a promotion too early,
1260 * whether a user wants that or not.
1261 *
1262 * This removal of files is usually unnecessary because they can exist
1263 * only during a few moments during a standby promotion. However there is
1264 * a race condition: if pg_ctl promote is executed and creates the files
1265 * during a promotion, the files can stay around even after the server is
1266 * brought up to new master. Then, if new standby starts by using the
1267 * backup taken from that master, the files can exist at the server
1268 * startup and should be removed in order to avoid an unexpected
1269 * promotion.
1270 *
1271 * Note that promotion signal files need to be removed before the startup
1272 * process is invoked. Because, after that, they can be used by
1273 * postmaster's SIGUSR1 signal handler.
1274 */
1275 RemovePromoteSignalFiles();
1276
1277 /* Remove any outdated file holding the current log filenames. */
1278 if (unlink(LOG_METAINFO_DATAFILE) < 0 && errno != ENOENT)
1279 ereport(LOG,
1280 (errcode_for_file_access(),
1281 errmsg("could not remove file \"%s\": %m",
1282 LOG_METAINFO_DATAFILE)));
1283
1284 /*
1285 * If enabled, start up syslogger collection subprocess
1286 */
1287 SysLoggerPID = SysLogger_Start();
1288
1289 /*
1290 * Reset whereToSendOutput from DestDebug (its starting state) to
1291 * DestNone. This stops ereport from sending log messages to stderr unless
1292 * Log_destination permits. We don't do this until the postmaster is
1293 * fully launched, since startup failures may as well be reported to
1294 * stderr.
1295 *
1296 * If we are in fact disabling logging to stderr, first emit a log message
1297 * saying so, to provide a breadcrumb trail for users who may not remember
1298 * that their logging is configured to go somewhere else.
1299 */
1300 if (!(Log_destination & LOG_DESTINATION_STDERR))
1301 ereport(LOG,
1302 (errmsg("ending log output to stderr"),
1303 errhint("Future log output will go to log destination \"%s\".",
1304 Log_destination_string)));
1305
1306 whereToSendOutput = DestNone;
1307
1308 /*
1309 * Initialize stats collection subsystem (this does NOT start the
1310 * collector process!)
1311 */
1312 pgstat_init();
1313
1314 /*
1315 * Initialize the autovacuum subsystem (again, no process start yet)
1316 */
1317 autovac_init();
1318
1319 /*
1320 * Load configuration files for client authentication.
1321 */
1322 if (!load_hba())
1323 {
1324 /*
1325 * It makes no sense to continue if we fail to load the HBA file,
1326 * since there is no way to connect to the database in this case.
1327 */
1328 ereport(FATAL,
1329 (errmsg("could not load pg_hba.conf")));
1330 }
1331 if (!load_ident())
1332 {
1333 /*
1334 * We can start up without the IDENT file, although it means that you
1335 * cannot log in using any of the authentication methods that need a
1336 * user name mapping. load_ident() already logged the details of error
1337 * to the log.
1338 */
1339 }
1340
1341 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1342
1343 /*
1344 * On macOS, libintl replaces setlocale() with a version that calls
1345 * CFLocaleCopyCurrent() when its second argument is "" and every relevant
1346 * environment variable is unset or empty. CFLocaleCopyCurrent() makes
1347 * the process multithreaded. The postmaster calls sigprocmask() and
1348 * calls fork() without an immediate exec(), both of which have undefined
1349 * behavior in a multithreaded program. A multithreaded postmaster is the
1350 * normal case on Windows, which offers neither fork() nor sigprocmask().
1351 */
1352 if (pthread_is_threaded_np() != 0)
1353 ereport(FATAL,
1354 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1355 errmsg("postmaster became multithreaded during startup"),
1356 errhint("Set the LC_ALL environment variable to a valid locale.")));
1357 #endif
1358
1359 /*
1360 * Remember postmaster startup time
1361 */
1362 PgStartTime = GetCurrentTimestamp();
1363 #ifndef HAVE_STRONG_RANDOM
1364 /* RandomCancelKey wants its own copy */
1365 gettimeofday(&random_start_time, NULL);
1366 #endif
1367
1368 /*
1369 * Report postmaster status in the postmaster.pid file, to allow pg_ctl to
1370 * see what's happening.
1371 */
1372 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STARTING);
1373
1374 /*
1375 * We're ready to rock and roll...
1376 */
1377 StartupPID = StartupDataBase();
1378 Assert(StartupPID != 0);
1379 StartupStatus = STARTUP_RUNNING;
1380 pmState = PM_STARTUP;
1381
1382 /* Some workers may be scheduled to start now */
1383 maybe_start_bgworkers();
1384
1385 status = ServerLoop();
1386
1387 /*
1388 * ServerLoop probably shouldn't ever return, but if it does, close down.
1389 */
1390 ExitPostmaster(status != STATUS_OK);
1391
1392 abort(); /* not reached */
1393 }
1394
1395
1396 /*
1397 * on_proc_exit callback to close server's listen sockets
1398 */
1399 static void
CloseServerPorts(int status,Datum arg)1400 CloseServerPorts(int status, Datum arg)
1401 {
1402 int i;
1403
1404 /*
1405 * First, explicitly close all the socket FDs. We used to just let this
1406 * happen implicitly at postmaster exit, but it's better to close them
1407 * before we remove the postmaster.pid lockfile; otherwise there's a race
1408 * condition if a new postmaster wants to re-use the TCP port number.
1409 */
1410 for (i = 0; i < MAXLISTEN; i++)
1411 {
1412 if (ListenSocket[i] != PGINVALID_SOCKET)
1413 {
1414 StreamClose(ListenSocket[i]);
1415 ListenSocket[i] = PGINVALID_SOCKET;
1416 }
1417 }
1418
1419 /*
1420 * Next, remove any filesystem entries for Unix sockets. To avoid race
1421 * conditions against incoming postmasters, this must happen after closing
1422 * the sockets and before removing lock files.
1423 */
1424 RemoveSocketFiles();
1425
1426 /*
1427 * We don't do anything about socket lock files here; those will be
1428 * removed in a later on_proc_exit callback.
1429 */
1430 }
1431
1432 /*
1433 * on_proc_exit callback to delete external_pid_file
1434 */
1435 static void
unlink_external_pid_file(int status,Datum arg)1436 unlink_external_pid_file(int status, Datum arg)
1437 {
1438 if (external_pid_file)
1439 unlink(external_pid_file);
1440 }
1441
1442
1443 /*
1444 * Compute and check the directory paths to files that are part of the
1445 * installation (as deduced from the postgres executable's own location)
1446 */
1447 static void
getInstallationPaths(const char * argv0)1448 getInstallationPaths(const char *argv0)
1449 {
1450 DIR *pdir;
1451
1452 /* Locate the postgres executable itself */
1453 if (find_my_exec(argv0, my_exec_path) < 0)
1454 elog(FATAL, "%s: could not locate my own executable path", argv0);
1455
1456 #ifdef EXEC_BACKEND
1457 /* Locate executable backend before we change working directory */
1458 if (find_other_exec(argv0, "postgres", PG_BACKEND_VERSIONSTR,
1459 postgres_exec_path) < 0)
1460 ereport(FATAL,
1461 (errmsg("%s: could not locate matching postgres executable",
1462 argv0)));
1463 #endif
1464
1465 /*
1466 * Locate the pkglib directory --- this has to be set early in case we try
1467 * to load any modules from it in response to postgresql.conf entries.
1468 */
1469 get_pkglib_path(my_exec_path, pkglib_path);
1470
1471 /*
1472 * Verify that there's a readable directory there; otherwise the Postgres
1473 * installation is incomplete or corrupt. (A typical cause of this
1474 * failure is that the postgres executable has been moved or hardlinked to
1475 * some directory that's not a sibling of the installation lib/
1476 * directory.)
1477 */
1478 pdir = AllocateDir(pkglib_path);
1479 if (pdir == NULL)
1480 ereport(ERROR,
1481 (errcode_for_file_access(),
1482 errmsg("could not open directory \"%s\": %m",
1483 pkglib_path),
1484 errhint("This may indicate an incomplete PostgreSQL installation, or that the file \"%s\" has been moved away from its proper location.",
1485 my_exec_path)));
1486 FreeDir(pdir);
1487
1488 /*
1489 * XXX is it worth similarly checking the share/ directory? If the lib/
1490 * directory is there, then share/ probably is too.
1491 */
1492 }
1493
1494
1495 /*
1496 * Validate the proposed data directory
1497 */
1498 static void
checkDataDir(void)1499 checkDataDir(void)
1500 {
1501 char path[MAXPGPATH];
1502 FILE *fp;
1503 struct stat stat_buf;
1504
1505 Assert(DataDir);
1506
1507 if (stat(DataDir, &stat_buf) != 0)
1508 {
1509 if (errno == ENOENT)
1510 ereport(FATAL,
1511 (errcode_for_file_access(),
1512 errmsg("data directory \"%s\" does not exist",
1513 DataDir)));
1514 else
1515 ereport(FATAL,
1516 (errcode_for_file_access(),
1517 errmsg("could not read permissions of directory \"%s\": %m",
1518 DataDir)));
1519 }
1520
1521 /* eventual chdir would fail anyway, but let's test ... */
1522 if (!S_ISDIR(stat_buf.st_mode))
1523 ereport(FATAL,
1524 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1525 errmsg("specified data directory \"%s\" is not a directory",
1526 DataDir)));
1527
1528 /*
1529 * Check that the directory belongs to my userid; if not, reject.
1530 *
1531 * This check is an essential part of the interlock that prevents two
1532 * postmasters from starting in the same directory (see CreateLockFile()).
1533 * Do not remove or weaken it.
1534 *
1535 * XXX can we safely enable this check on Windows?
1536 */
1537 #if !defined(WIN32) && !defined(__CYGWIN__)
1538 if (stat_buf.st_uid != geteuid())
1539 ereport(FATAL,
1540 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1541 errmsg("data directory \"%s\" has wrong ownership",
1542 DataDir),
1543 errhint("The server must be started by the user that owns the data directory.")));
1544 #endif
1545
1546 /*
1547 * Check if the directory has group or world access. If so, reject.
1548 *
1549 * It would be possible to allow weaker constraints (for example, allow
1550 * group access) but we cannot make a general assumption that that is
1551 * okay; for example there are platforms where nearly all users
1552 * customarily belong to the same group. Perhaps this test should be
1553 * configurable.
1554 *
1555 * XXX temporarily suppress check when on Windows, because there may not
1556 * be proper support for Unix-y file permissions. Need to think of a
1557 * reasonable check to apply on Windows.
1558 */
1559 #if !defined(WIN32) && !defined(__CYGWIN__)
1560 if (stat_buf.st_mode & (S_IRWXG | S_IRWXO))
1561 ereport(FATAL,
1562 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1563 errmsg("data directory \"%s\" has group or world access",
1564 DataDir),
1565 errdetail("Permissions should be u=rwx (0700).")));
1566 #endif
1567
1568 /* Look for PG_VERSION before looking for pg_control */
1569 ValidatePgVersion(DataDir);
1570
1571 snprintf(path, sizeof(path), "%s/global/pg_control", DataDir);
1572
1573 fp = AllocateFile(path, PG_BINARY_R);
1574 if (fp == NULL)
1575 {
1576 write_stderr("%s: could not find the database system\n"
1577 "Expected to find it in the directory \"%s\",\n"
1578 "but could not open file \"%s\": %s\n",
1579 progname, DataDir, path, strerror(errno));
1580 ExitPostmaster(2);
1581 }
1582 FreeFile(fp);
1583 }
1584
1585 /*
1586 * Determine how long should we let ServerLoop sleep.
1587 *
1588 * In normal conditions we wait at most one minute, to ensure that the other
1589 * background tasks handled by ServerLoop get done even when no requests are
1590 * arriving. However, if there are background workers waiting to be started,
1591 * we don't actually sleep so that they are quickly serviced. Other exception
1592 * cases are as shown in the code.
1593 */
1594 static void
DetermineSleepTime(struct timeval * timeout)1595 DetermineSleepTime(struct timeval *timeout)
1596 {
1597 TimestampTz next_wakeup = 0;
1598
1599 /*
1600 * Normal case: either there are no background workers at all, or we're in
1601 * a shutdown sequence (during which we ignore bgworkers altogether).
1602 */
1603 if (Shutdown > NoShutdown ||
1604 (!StartWorkerNeeded && !HaveCrashedWorker))
1605 {
1606 if (AbortStartTime != 0)
1607 {
1608 /* time left to abort; clamp to 0 in case it already expired */
1609 timeout->tv_sec = SIGKILL_CHILDREN_AFTER_SECS -
1610 (time(NULL) - AbortStartTime);
1611 timeout->tv_sec = Max(timeout->tv_sec, 0);
1612 timeout->tv_usec = 0;
1613 }
1614 else
1615 {
1616 timeout->tv_sec = 60;
1617 timeout->tv_usec = 0;
1618 }
1619 return;
1620 }
1621
1622 if (StartWorkerNeeded)
1623 {
1624 timeout->tv_sec = 0;
1625 timeout->tv_usec = 0;
1626 return;
1627 }
1628
1629 if (HaveCrashedWorker)
1630 {
1631 slist_mutable_iter siter;
1632
1633 /*
1634 * When there are crashed bgworkers, we sleep just long enough that
1635 * they are restarted when they request to be. Scan the list to
1636 * determine the minimum of all wakeup times according to most recent
1637 * crash time and requested restart interval.
1638 */
1639 slist_foreach_modify(siter, &BackgroundWorkerList)
1640 {
1641 RegisteredBgWorker *rw;
1642 TimestampTz this_wakeup;
1643
1644 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
1645
1646 if (rw->rw_crashed_at == 0)
1647 continue;
1648
1649 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART
1650 || rw->rw_terminate)
1651 {
1652 ForgetBackgroundWorker(&siter);
1653 continue;
1654 }
1655
1656 this_wakeup = TimestampTzPlusMilliseconds(rw->rw_crashed_at,
1657 1000L * rw->rw_worker.bgw_restart_time);
1658 if (next_wakeup == 0 || this_wakeup < next_wakeup)
1659 next_wakeup = this_wakeup;
1660 }
1661 }
1662
1663 if (next_wakeup != 0)
1664 {
1665 long secs;
1666 int microsecs;
1667
1668 TimestampDifference(GetCurrentTimestamp(), next_wakeup,
1669 &secs, µsecs);
1670 timeout->tv_sec = secs;
1671 timeout->tv_usec = microsecs;
1672
1673 /* Ensure we don't exceed one minute */
1674 if (timeout->tv_sec > 60)
1675 {
1676 timeout->tv_sec = 60;
1677 timeout->tv_usec = 0;
1678 }
1679 }
1680 else
1681 {
1682 timeout->tv_sec = 60;
1683 timeout->tv_usec = 0;
1684 }
1685 }
1686
1687 /*
1688 * Main idle loop of postmaster
1689 *
1690 * NB: Needs to be called with signals blocked
1691 */
1692 static int
ServerLoop(void)1693 ServerLoop(void)
1694 {
1695 fd_set readmask;
1696 int nSockets;
1697 time_t last_lockfile_recheck_time,
1698 last_touch_time;
1699
1700 last_lockfile_recheck_time = last_touch_time = time(NULL);
1701
1702 nSockets = initMasks(&readmask);
1703
1704 for (;;)
1705 {
1706 fd_set rmask;
1707 int selres;
1708 time_t now;
1709
1710 /*
1711 * Wait for a connection request to arrive.
1712 *
1713 * We block all signals except while sleeping. That makes it safe for
1714 * signal handlers, which again block all signals while executing, to
1715 * do nontrivial work.
1716 *
1717 * If we are in PM_WAIT_DEAD_END state, then we don't want to accept
1718 * any new connections, so we don't call select(), and just sleep.
1719 */
1720 memcpy((char *) &rmask, (char *) &readmask, sizeof(fd_set));
1721
1722 if (pmState == PM_WAIT_DEAD_END)
1723 {
1724 PG_SETMASK(&UnBlockSig);
1725
1726 pg_usleep(100000L); /* 100 msec seems reasonable */
1727 selres = 0;
1728
1729 PG_SETMASK(&BlockSig);
1730 }
1731 else
1732 {
1733 /* must set timeout each time; some OSes change it! */
1734 struct timeval timeout;
1735
1736 /* Needs to run with blocked signals! */
1737 DetermineSleepTime(&timeout);
1738
1739 PG_SETMASK(&UnBlockSig);
1740
1741 selres = select(nSockets, &rmask, NULL, NULL, &timeout);
1742
1743 PG_SETMASK(&BlockSig);
1744 }
1745
1746 /* Now check the select() result */
1747 if (selres < 0)
1748 {
1749 if (errno != EINTR && errno != EWOULDBLOCK)
1750 {
1751 ereport(LOG,
1752 (errcode_for_socket_access(),
1753 errmsg("select() failed in postmaster: %m")));
1754 return STATUS_ERROR;
1755 }
1756 }
1757
1758 /*
1759 * New connection pending on any of our sockets? If so, fork a child
1760 * process to deal with it.
1761 */
1762 if (selres > 0)
1763 {
1764 int i;
1765
1766 for (i = 0; i < MAXLISTEN; i++)
1767 {
1768 if (ListenSocket[i] == PGINVALID_SOCKET)
1769 break;
1770 if (FD_ISSET(ListenSocket[i], &rmask))
1771 {
1772 Port *port;
1773
1774 port = ConnCreate(ListenSocket[i]);
1775 if (port)
1776 {
1777 BackendStartup(port);
1778
1779 /*
1780 * We no longer need the open socket or port structure
1781 * in this process
1782 */
1783 StreamClose(port->sock);
1784 ConnFree(port);
1785 }
1786 }
1787 }
1788 }
1789
1790 /* If we have lost the log collector, try to start a new one */
1791 if (SysLoggerPID == 0 && Logging_collector)
1792 SysLoggerPID = SysLogger_Start();
1793
1794 /*
1795 * If no background writer process is running, and we are not in a
1796 * state that prevents it, start one. It doesn't matter if this
1797 * fails, we'll just try again later. Likewise for the checkpointer.
1798 */
1799 if (pmState == PM_RUN || pmState == PM_RECOVERY ||
1800 pmState == PM_HOT_STANDBY)
1801 {
1802 if (CheckpointerPID == 0)
1803 CheckpointerPID = StartCheckpointer();
1804 if (BgWriterPID == 0)
1805 BgWriterPID = StartBackgroundWriter();
1806 }
1807
1808 /*
1809 * Likewise, if we have lost the walwriter process, try to start a new
1810 * one. But this is needed only in normal operation (else we cannot
1811 * be writing any new WAL).
1812 */
1813 if (WalWriterPID == 0 && pmState == PM_RUN)
1814 WalWriterPID = StartWalWriter();
1815
1816 /*
1817 * If we have lost the autovacuum launcher, try to start a new one. We
1818 * don't want autovacuum to run in binary upgrade mode because
1819 * autovacuum might update relfrozenxid for empty tables before the
1820 * physical files are put in place.
1821 */
1822 if (!IsBinaryUpgrade && AutoVacPID == 0 &&
1823 (AutoVacuumingActive() || start_autovac_launcher) &&
1824 pmState == PM_RUN)
1825 {
1826 AutoVacPID = StartAutoVacLauncher();
1827 if (AutoVacPID != 0)
1828 start_autovac_launcher = false; /* signal processed */
1829 }
1830
1831 /* If we have lost the stats collector, try to start a new one */
1832 if (PgStatPID == 0 &&
1833 (pmState == PM_RUN || pmState == PM_HOT_STANDBY))
1834 PgStatPID = pgstat_start();
1835
1836 /* If we have lost the archiver, try to start a new one. */
1837 if (PgArchPID == 0 && PgArchStartupAllowed())
1838 PgArchPID = pgarch_start();
1839
1840 /* If we need to signal the autovacuum launcher, do so now */
1841 if (avlauncher_needs_signal)
1842 {
1843 avlauncher_needs_signal = false;
1844 if (AutoVacPID != 0)
1845 kill(AutoVacPID, SIGUSR2);
1846 }
1847
1848 /* If we need to start a WAL receiver, try to do that now */
1849 if (WalReceiverRequested)
1850 MaybeStartWalReceiver();
1851
1852 /* Get other worker processes running, if needed */
1853 if (StartWorkerNeeded || HaveCrashedWorker)
1854 maybe_start_bgworkers();
1855
1856 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1857
1858 /*
1859 * With assertions enabled, check regularly for appearance of
1860 * additional threads. All builds check at start and exit.
1861 */
1862 Assert(pthread_is_threaded_np() == 0);
1863 #endif
1864
1865 /*
1866 * Lastly, check to see if it's time to do some things that we don't
1867 * want to do every single time through the loop, because they're a
1868 * bit expensive. Note that there's up to a minute of slop in when
1869 * these tasks will be performed, since DetermineSleepTime() will let
1870 * us sleep at most that long; except for SIGKILL timeout which has
1871 * special-case logic there.
1872 */
1873 now = time(NULL);
1874
1875 /*
1876 * If we already sent SIGQUIT to children and they are slow to shut
1877 * down, it's time to send them SIGKILL. This doesn't happen
1878 * normally, but under certain conditions backends can get stuck while
1879 * shutting down. This is a last measure to get them unwedged.
1880 *
1881 * Note we also do this during recovery from a process crash.
1882 */
1883 if ((Shutdown >= ImmediateShutdown || (FatalError && !SendStop)) &&
1884 AbortStartTime != 0 &&
1885 (now - AbortStartTime) >= SIGKILL_CHILDREN_AFTER_SECS)
1886 {
1887 /* We were gentle with them before. Not anymore */
1888 TerminateChildren(SIGKILL);
1889 /* reset flag so we don't SIGKILL again */
1890 AbortStartTime = 0;
1891 }
1892
1893 /*
1894 * Once a minute, verify that postmaster.pid hasn't been removed or
1895 * overwritten. If it has, we force a shutdown. This avoids having
1896 * postmasters and child processes hanging around after their database
1897 * is gone, and maybe causing problems if a new database cluster is
1898 * created in the same place. It also provides some protection
1899 * against a DBA foolishly removing postmaster.pid and manually
1900 * starting a new postmaster. Data corruption is likely to ensue from
1901 * that anyway, but we can minimize the damage by aborting ASAP.
1902 */
1903 if (now - last_lockfile_recheck_time >= 1 * SECS_PER_MINUTE)
1904 {
1905 if (!RecheckDataDirLockFile())
1906 {
1907 ereport(LOG,
1908 (errmsg("performing immediate shutdown because data directory lock file is invalid")));
1909 kill(MyProcPid, SIGQUIT);
1910 }
1911 last_lockfile_recheck_time = now;
1912 }
1913
1914 /*
1915 * Touch Unix socket and lock files every 58 minutes, to ensure that
1916 * they are not removed by overzealous /tmp-cleaning tasks. We assume
1917 * no one runs cleaners with cutoff times of less than an hour ...
1918 */
1919 if (now - last_touch_time >= 58 * SECS_PER_MINUTE)
1920 {
1921 TouchSocketFiles();
1922 TouchSocketLockFiles();
1923 last_touch_time = now;
1924 }
1925 }
1926 }
1927
1928 /*
1929 * Initialise the masks for select() for the ports we are listening on.
1930 * Return the number of sockets to listen on.
1931 */
1932 static int
initMasks(fd_set * rmask)1933 initMasks(fd_set *rmask)
1934 {
1935 int maxsock = -1;
1936 int i;
1937
1938 FD_ZERO(rmask);
1939
1940 for (i = 0; i < MAXLISTEN; i++)
1941 {
1942 int fd = ListenSocket[i];
1943
1944 if (fd == PGINVALID_SOCKET)
1945 break;
1946 FD_SET(fd, rmask);
1947
1948 if (fd > maxsock)
1949 maxsock = fd;
1950 }
1951
1952 return maxsock + 1;
1953 }
1954
1955
1956 /*
1957 * Read a client's startup packet and do something according to it.
1958 *
1959 * Returns STATUS_OK or STATUS_ERROR, or might call ereport(FATAL) and
1960 * not return at all.
1961 *
1962 * (Note that ereport(FATAL) stuff is sent to the client, so only use it
1963 * if that's what you want. Return STATUS_ERROR if you don't want to
1964 * send anything to the client, which would typically be appropriate
1965 * if we detect a communications failure.)
1966 */
1967 static int
ProcessStartupPacket(Port * port,bool SSLdone)1968 ProcessStartupPacket(Port *port, bool SSLdone)
1969 {
1970 int32 len;
1971 void *buf;
1972 ProtocolVersion proto;
1973 MemoryContext oldcontext;
1974
1975 pq_startmsgread();
1976 if (pq_getbytes((char *) &len, 4) == EOF)
1977 {
1978 /*
1979 * EOF after SSLdone probably means the client didn't like our
1980 * response to NEGOTIATE_SSL_CODE. That's not an error condition, so
1981 * don't clutter the log with a complaint.
1982 */
1983 if (!SSLdone)
1984 ereport(COMMERROR,
1985 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1986 errmsg("incomplete startup packet")));
1987 return STATUS_ERROR;
1988 }
1989
1990 len = ntohl(len);
1991 len -= 4;
1992
1993 if (len < (int32) sizeof(ProtocolVersion) ||
1994 len > MAX_STARTUP_PACKET_LENGTH)
1995 {
1996 ereport(COMMERROR,
1997 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1998 errmsg("invalid length of startup packet")));
1999 return STATUS_ERROR;
2000 }
2001
2002 /*
2003 * Allocate at least the size of an old-style startup packet, plus one
2004 * extra byte, and make sure all are zeroes. This ensures we will have
2005 * null termination of all strings, in both fixed- and variable-length
2006 * packet layouts.
2007 */
2008 if (len <= (int32) sizeof(StartupPacket))
2009 buf = palloc0(sizeof(StartupPacket) + 1);
2010 else
2011 buf = palloc0(len + 1);
2012
2013 if (pq_getbytes(buf, len) == EOF)
2014 {
2015 ereport(COMMERROR,
2016 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2017 errmsg("incomplete startup packet")));
2018 return STATUS_ERROR;
2019 }
2020 pq_endmsgread();
2021
2022 /*
2023 * The first field is either a protocol version number or a special
2024 * request code.
2025 */
2026 port->proto = proto = ntohl(*((ProtocolVersion *) buf));
2027
2028 if (proto == CANCEL_REQUEST_CODE)
2029 {
2030 processCancelRequest(port, buf);
2031 /* Not really an error, but we don't want to proceed further */
2032 return STATUS_ERROR;
2033 }
2034
2035 if (proto == NEGOTIATE_SSL_CODE && !SSLdone)
2036 {
2037 char SSLok;
2038
2039 #ifdef USE_SSL
2040 /* No SSL when disabled or on Unix sockets */
2041 if (!LoadedSSL || IS_AF_UNIX(port->laddr.addr.ss_family))
2042 SSLok = 'N';
2043 else
2044 SSLok = 'S'; /* Support for SSL */
2045 #else
2046 SSLok = 'N'; /* No support for SSL */
2047 #endif
2048
2049 retry1:
2050 if (send(port->sock, &SSLok, 1, 0) != 1)
2051 {
2052 if (errno == EINTR)
2053 goto retry1; /* if interrupted, just retry */
2054 ereport(COMMERROR,
2055 (errcode_for_socket_access(),
2056 errmsg("failed to send SSL negotiation response: %m")));
2057 return STATUS_ERROR; /* close the connection */
2058 }
2059
2060 #ifdef USE_SSL
2061 if (SSLok == 'S' && secure_open_server(port) == -1)
2062 return STATUS_ERROR;
2063 #endif
2064
2065 /*
2066 * At this point we should have no data already buffered. If we do,
2067 * it was received before we performed the SSL handshake, so it wasn't
2068 * encrypted and indeed may have been injected by a man-in-the-middle.
2069 * We report this case to the client.
2070 */
2071 if (pq_buffer_has_data())
2072 ereport(FATAL,
2073 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2074 errmsg("received unencrypted data after SSL request"),
2075 errdetail("This could be either a client-software bug or evidence of an attempted man-in-the-middle attack.")));
2076
2077 /* regular startup packet, cancel, etc packet should follow... */
2078 /* but not another SSL negotiation request */
2079 return ProcessStartupPacket(port, true);
2080 }
2081
2082 /* Could add additional special packet types here */
2083
2084 /*
2085 * Set FrontendProtocol now so that ereport() knows what format to send if
2086 * we fail during startup.
2087 */
2088 FrontendProtocol = proto;
2089
2090 /* Check that the major protocol version is in range. */
2091 if (PG_PROTOCOL_MAJOR(proto) < PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST) ||
2092 PG_PROTOCOL_MAJOR(proto) > PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST))
2093 ereport(FATAL,
2094 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2095 errmsg("unsupported frontend protocol %u.%u: server supports %u.0 to %u.%u",
2096 PG_PROTOCOL_MAJOR(proto), PG_PROTOCOL_MINOR(proto),
2097 PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST),
2098 PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST),
2099 PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST))));
2100
2101 /*
2102 * Now fetch parameters out of startup packet and save them into the Port
2103 * structure. All data structures attached to the Port struct must be
2104 * allocated in TopMemoryContext so that they will remain available in a
2105 * running backend (even after PostmasterContext is destroyed). We need
2106 * not worry about leaking this storage on failure, since we aren't in the
2107 * postmaster process anymore.
2108 */
2109 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
2110
2111 if (PG_PROTOCOL_MAJOR(proto) >= 3)
2112 {
2113 int32 offset = sizeof(ProtocolVersion);
2114 List *unrecognized_protocol_options = NIL;
2115
2116 /*
2117 * Scan packet body for name/option pairs. We can assume any string
2118 * beginning within the packet body is null-terminated, thanks to
2119 * zeroing extra byte above.
2120 */
2121 port->guc_options = NIL;
2122
2123 while (offset < len)
2124 {
2125 char *nameptr = ((char *) buf) + offset;
2126 int32 valoffset;
2127 char *valptr;
2128
2129 if (*nameptr == '\0')
2130 break; /* found packet terminator */
2131 valoffset = offset + strlen(nameptr) + 1;
2132 if (valoffset >= len)
2133 break; /* missing value, will complain below */
2134 valptr = ((char *) buf) + valoffset;
2135
2136 if (strcmp(nameptr, "database") == 0)
2137 port->database_name = pstrdup(valptr);
2138 else if (strcmp(nameptr, "user") == 0)
2139 port->user_name = pstrdup(valptr);
2140 else if (strcmp(nameptr, "options") == 0)
2141 port->cmdline_options = pstrdup(valptr);
2142 else if (strcmp(nameptr, "replication") == 0)
2143 {
2144 /*
2145 * Due to backward compatibility concerns the replication
2146 * parameter is a hybrid beast which allows the value to be
2147 * either boolean or the string 'database'. The latter
2148 * connects to a specific database which is e.g. required for
2149 * logical decoding while.
2150 */
2151 if (strcmp(valptr, "database") == 0)
2152 {
2153 am_walsender = true;
2154 am_db_walsender = true;
2155 }
2156 else if (!parse_bool(valptr, &am_walsender))
2157 ereport(FATAL,
2158 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
2159 errmsg("invalid value for parameter \"%s\": \"%s\"",
2160 "replication",
2161 valptr),
2162 errhint("Valid values are: \"false\", 0, \"true\", 1, \"database\".")));
2163 }
2164 else if (strncmp(nameptr, "_pq_.", 5) == 0)
2165 {
2166 /*
2167 * Any option beginning with _pq_. is reserved for use as a
2168 * protocol-level option, but at present no such options are
2169 * defined.
2170 */
2171 unrecognized_protocol_options =
2172 lappend(unrecognized_protocol_options, pstrdup(nameptr));
2173 }
2174 else
2175 {
2176 /* Assume it's a generic GUC option */
2177 port->guc_options = lappend(port->guc_options,
2178 pstrdup(nameptr));
2179 port->guc_options = lappend(port->guc_options,
2180 pstrdup(valptr));
2181 }
2182 offset = valoffset + strlen(valptr) + 1;
2183 }
2184
2185 /*
2186 * If we didn't find a packet terminator exactly at the end of the
2187 * given packet length, complain.
2188 */
2189 if (offset != len - 1)
2190 ereport(FATAL,
2191 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2192 errmsg("invalid startup packet layout: expected terminator as last byte")));
2193
2194 /*
2195 * If the client requested a newer protocol version or if the client
2196 * requested any protocol options we didn't recognize, let them know
2197 * the newest minor protocol version we do support and the names of any
2198 * unrecognized options.
2199 */
2200 if (PG_PROTOCOL_MINOR(proto) > PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST) ||
2201 unrecognized_protocol_options != NIL)
2202 SendNegotiateProtocolVersion(unrecognized_protocol_options);
2203 }
2204 else
2205 {
2206 /*
2207 * Get the parameters from the old-style, fixed-width-fields startup
2208 * packet as C strings. The packet destination was cleared first so a
2209 * short packet has zeros silently added. We have to be prepared to
2210 * truncate the pstrdup result for oversize fields, though.
2211 */
2212 StartupPacket *packet = (StartupPacket *) buf;
2213
2214 port->database_name = pstrdup(packet->database);
2215 if (strlen(port->database_name) > sizeof(packet->database))
2216 port->database_name[sizeof(packet->database)] = '\0';
2217 port->user_name = pstrdup(packet->user);
2218 if (strlen(port->user_name) > sizeof(packet->user))
2219 port->user_name[sizeof(packet->user)] = '\0';
2220 port->cmdline_options = pstrdup(packet->options);
2221 if (strlen(port->cmdline_options) > sizeof(packet->options))
2222 port->cmdline_options[sizeof(packet->options)] = '\0';
2223 port->guc_options = NIL;
2224 }
2225
2226 /* Check a user name was given. */
2227 if (port->user_name == NULL || port->user_name[0] == '\0')
2228 ereport(FATAL,
2229 (errcode(ERRCODE_INVALID_AUTHORIZATION_SPECIFICATION),
2230 errmsg("no PostgreSQL user name specified in startup packet")));
2231
2232 /* The database defaults to the user name. */
2233 if (port->database_name == NULL || port->database_name[0] == '\0')
2234 port->database_name = pstrdup(port->user_name);
2235
2236 if (Db_user_namespace)
2237 {
2238 /*
2239 * If user@, it is a global user, remove '@'. We only want to do this
2240 * if there is an '@' at the end and no earlier in the user string or
2241 * they may fake as a local user of another database attaching to this
2242 * database.
2243 */
2244 if (strchr(port->user_name, '@') ==
2245 port->user_name + strlen(port->user_name) - 1)
2246 *strchr(port->user_name, '@') = '\0';
2247 else
2248 {
2249 /* Append '@' and dbname */
2250 port->user_name = psprintf("%s@%s", port->user_name, port->database_name);
2251 }
2252 }
2253
2254 /*
2255 * Truncate given database and user names to length of a Postgres name.
2256 * This avoids lookup failures when overlength names are given.
2257 */
2258 if (strlen(port->database_name) >= NAMEDATALEN)
2259 port->database_name[NAMEDATALEN - 1] = '\0';
2260 if (strlen(port->user_name) >= NAMEDATALEN)
2261 port->user_name[NAMEDATALEN - 1] = '\0';
2262
2263 /*
2264 * Normal walsender backends, e.g. for streaming replication, are not
2265 * connected to a particular database. But walsenders used for logical
2266 * replication need to connect to a specific database. We allow streaming
2267 * replication commands to be issued even if connected to a database as it
2268 * can make sense to first make a basebackup and then stream changes
2269 * starting from that.
2270 */
2271 if (am_walsender && !am_db_walsender)
2272 port->database_name[0] = '\0';
2273
2274 /*
2275 * Done putting stuff in TopMemoryContext.
2276 */
2277 MemoryContextSwitchTo(oldcontext);
2278
2279 /*
2280 * If we're going to reject the connection due to database state, say so
2281 * now instead of wasting cycles on an authentication exchange. (This also
2282 * allows a pg_ping utility to be written.)
2283 */
2284 switch (port->canAcceptConnections)
2285 {
2286 case CAC_STARTUP:
2287 ereport(FATAL,
2288 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2289 errmsg("the database system is starting up")));
2290 break;
2291 case CAC_SHUTDOWN:
2292 ereport(FATAL,
2293 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2294 errmsg("the database system is shutting down")));
2295 break;
2296 case CAC_RECOVERY:
2297 ereport(FATAL,
2298 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2299 errmsg("the database system is in recovery mode")));
2300 break;
2301 case CAC_TOOMANY:
2302 ereport(FATAL,
2303 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
2304 errmsg("sorry, too many clients already")));
2305 break;
2306 case CAC_SUPERUSER:
2307 /* OK for now, will check in InitPostgres */
2308 break;
2309 case CAC_OK:
2310 break;
2311 }
2312
2313 return STATUS_OK;
2314 }
2315
2316 /*
2317 * Send a NegotiateProtocolVersion to the client. This lets the client know
2318 * that they have requested a newer minor protocol version than we are able
2319 * to speak. We'll speak the highest version we know about; the client can,
2320 * of course, abandon the connection if that's a problem.
2321 *
2322 * We also include in the response a list of protocol options we didn't
2323 * understand. This allows clients to include optional parameters that might
2324 * be present either in newer protocol versions or third-party protocol
2325 * extensions without fear of having to reconnect if those options are not
2326 * understood, while at the same time making certain that the client is aware
2327 * of which options were actually accepted.
2328 */
2329 static void
SendNegotiateProtocolVersion(List * unrecognized_protocol_options)2330 SendNegotiateProtocolVersion(List *unrecognized_protocol_options)
2331 {
2332 StringInfoData buf;
2333 ListCell *lc;
2334
2335 pq_beginmessage(&buf, 'v'); /* NegotiateProtocolVersion */
2336 pq_sendint(&buf, PG_PROTOCOL_LATEST, 4);
2337 pq_sendint(&buf, list_length(unrecognized_protocol_options), 4);
2338 foreach(lc, unrecognized_protocol_options)
2339 pq_sendstring(&buf, lfirst(lc));
2340 pq_endmessage(&buf);
2341
2342 /* no need to flush, some other message will follow */
2343 }
2344
2345 /*
2346 * The client has sent a cancel request packet, not a normal
2347 * start-a-new-connection packet. Perform the necessary processing.
2348 * Nothing is sent back to the client.
2349 */
2350 static void
processCancelRequest(Port * port,void * pkt)2351 processCancelRequest(Port *port, void *pkt)
2352 {
2353 CancelRequestPacket *canc = (CancelRequestPacket *) pkt;
2354 int backendPID;
2355 int32 cancelAuthCode;
2356 Backend *bp;
2357
2358 #ifndef EXEC_BACKEND
2359 dlist_iter iter;
2360 #else
2361 int i;
2362 #endif
2363
2364 backendPID = (int) ntohl(canc->backendPID);
2365 cancelAuthCode = (int32) ntohl(canc->cancelAuthCode);
2366
2367 /*
2368 * See if we have a matching backend. In the EXEC_BACKEND case, we can no
2369 * longer access the postmaster's own backend list, and must rely on the
2370 * duplicate array in shared memory.
2371 */
2372 #ifndef EXEC_BACKEND
2373 dlist_foreach(iter, &BackendList)
2374 {
2375 bp = dlist_container(Backend, elem, iter.cur);
2376 #else
2377 for (i = MaxLivePostmasterChildren() - 1; i >= 0; i--)
2378 {
2379 bp = (Backend *) &ShmemBackendArray[i];
2380 #endif
2381 if (bp->pid == backendPID)
2382 {
2383 if (bp->cancel_key == cancelAuthCode)
2384 {
2385 /* Found a match; signal that backend to cancel current op */
2386 ereport(DEBUG2,
2387 (errmsg_internal("processing cancel request: sending SIGINT to process %d",
2388 backendPID)));
2389 signal_child(bp->pid, SIGINT);
2390 }
2391 else
2392 /* Right PID, wrong key: no way, Jose */
2393 ereport(LOG,
2394 (errmsg("wrong key in cancel request for process %d",
2395 backendPID)));
2396 return;
2397 }
2398 }
2399
2400 /* No matching backend */
2401 ereport(LOG,
2402 (errmsg("PID %d in cancel request did not match any process",
2403 backendPID)));
2404 }
2405
2406 /*
2407 * canAcceptConnections --- check to see if database state allows connections
2408 * of the specified type. backend_type can be BACKEND_TYPE_NORMAL,
2409 * BACKEND_TYPE_AUTOVAC, or BACKEND_TYPE_BGWORKER. (Note that we don't yet
2410 * know whether a NORMAL connection might turn into a walsender.)
2411 */
2412 static CAC_state
2413 canAcceptConnections(int backend_type)
2414 {
2415 CAC_state result = CAC_OK;
2416
2417 /*
2418 * Can't start backends when in startup/shutdown/inconsistent recovery
2419 * state. We treat autovac workers the same as user backends for this
2420 * purpose. However, bgworkers are excluded from this test; we expect
2421 * bgworker_should_start_now() decided whether the DB state allows them.
2422 */
2423 if (pmState != PM_RUN && pmState != PM_HOT_STANDBY &&
2424 backend_type != BACKEND_TYPE_BGWORKER)
2425 {
2426 if (Shutdown > NoShutdown)
2427 return CAC_SHUTDOWN; /* shutdown is pending */
2428 else if (!FatalError &&
2429 (pmState == PM_STARTUP ||
2430 pmState == PM_RECOVERY))
2431 return CAC_STARTUP; /* normal startup */
2432 else
2433 return CAC_RECOVERY; /* else must be crash recovery */
2434 }
2435
2436 /*
2437 * "Smart shutdown" restrictions are applied only to normal connections,
2438 * not to autovac workers or bgworkers. When only superusers can connect,
2439 * we return CAC_SUPERUSER to indicate that superuserness must be checked
2440 * later. Note that neither CAC_OK nor CAC_SUPERUSER can safely be
2441 * returned until we have checked for too many children.
2442 */
2443 if (connsAllowed != ALLOW_ALL_CONNS &&
2444 backend_type == BACKEND_TYPE_NORMAL)
2445 {
2446 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
2447 result = CAC_SUPERUSER; /* allow superusers only */
2448 else
2449 return CAC_SHUTDOWN; /* shutdown is pending */
2450 }
2451
2452 /*
2453 * Don't start too many children.
2454 *
2455 * We allow more connections here than we can have backends because some
2456 * might still be authenticating; they might fail auth, or some existing
2457 * backend might exit before the auth cycle is completed. The exact
2458 * MaxBackends limit is enforced when a new backend tries to join the
2459 * shared-inval backend array.
2460 *
2461 * The limit here must match the sizes of the per-child-process arrays;
2462 * see comments for MaxLivePostmasterChildren().
2463 */
2464 if (CountChildren(BACKEND_TYPE_ALL) >= MaxLivePostmasterChildren())
2465 result = CAC_TOOMANY;
2466
2467 return result;
2468 }
2469
2470
2471 /*
2472 * ConnCreate -- create a local connection data structure
2473 *
2474 * Returns NULL on failure, other than out-of-memory which is fatal.
2475 */
2476 static Port *
2477 ConnCreate(int serverFd)
2478 {
2479 Port *port;
2480
2481 if (!(port = (Port *) calloc(1, sizeof(Port))))
2482 {
2483 ereport(LOG,
2484 (errcode(ERRCODE_OUT_OF_MEMORY),
2485 errmsg("out of memory")));
2486 ExitPostmaster(1);
2487 }
2488
2489 if (StreamConnection(serverFd, port) != STATUS_OK)
2490 {
2491 if (port->sock != PGINVALID_SOCKET)
2492 StreamClose(port->sock);
2493 ConnFree(port);
2494 return NULL;
2495 }
2496
2497 /*
2498 * Allocate GSSAPI specific state struct
2499 */
2500 #ifndef EXEC_BACKEND
2501 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
2502 port->gss = (pg_gssinfo *) calloc(1, sizeof(pg_gssinfo));
2503 if (!port->gss)
2504 {
2505 ereport(LOG,
2506 (errcode(ERRCODE_OUT_OF_MEMORY),
2507 errmsg("out of memory")));
2508 ExitPostmaster(1);
2509 }
2510 #endif
2511 #endif
2512
2513 return port;
2514 }
2515
2516
2517 /*
2518 * ConnFree -- free a local connection data structure
2519 */
2520 static void
2521 ConnFree(Port *conn)
2522 {
2523 #ifdef USE_SSL
2524 secure_close(conn);
2525 #endif
2526 if (conn->gss)
2527 free(conn->gss);
2528 free(conn);
2529 }
2530
2531
2532 /*
2533 * ClosePostmasterPorts -- close all the postmaster's open sockets
2534 *
2535 * This is called during child process startup to release file descriptors
2536 * that are not needed by that child process. The postmaster still has
2537 * them open, of course.
2538 *
2539 * Note: we pass am_syslogger as a boolean because we don't want to set
2540 * the global variable yet when this is called.
2541 */
2542 void
2543 ClosePostmasterPorts(bool am_syslogger)
2544 {
2545 int i;
2546
2547 #ifndef WIN32
2548
2549 /*
2550 * Close the write end of postmaster death watch pipe. It's important to
2551 * do this as early as possible, so that if postmaster dies, others won't
2552 * think that it's still running because we're holding the pipe open.
2553 */
2554 if (close(postmaster_alive_fds[POSTMASTER_FD_OWN]))
2555 ereport(FATAL,
2556 (errcode_for_file_access(),
2557 errmsg_internal("could not close postmaster death monitoring pipe in child process: %m")));
2558 postmaster_alive_fds[POSTMASTER_FD_OWN] = -1;
2559 #endif
2560
2561 /* Close the listen sockets */
2562 for (i = 0; i < MAXLISTEN; i++)
2563 {
2564 if (ListenSocket[i] != PGINVALID_SOCKET)
2565 {
2566 StreamClose(ListenSocket[i]);
2567 ListenSocket[i] = PGINVALID_SOCKET;
2568 }
2569 }
2570
2571 /* If using syslogger, close the read side of the pipe */
2572 if (!am_syslogger)
2573 {
2574 #ifndef WIN32
2575 if (syslogPipe[0] >= 0)
2576 close(syslogPipe[0]);
2577 syslogPipe[0] = -1;
2578 #else
2579 if (syslogPipe[0])
2580 CloseHandle(syslogPipe[0]);
2581 syslogPipe[0] = 0;
2582 #endif
2583 }
2584
2585 #ifdef USE_BONJOUR
2586 /* If using Bonjour, close the connection to the mDNS daemon */
2587 if (bonjour_sdref)
2588 close(DNSServiceRefSockFD(bonjour_sdref));
2589 #endif
2590 }
2591
2592
2593 /*
2594 * reset_shared -- reset shared memory and semaphores
2595 */
2596 static void
2597 reset_shared(int port)
2598 {
2599 /*
2600 * Create or re-create shared memory and semaphores.
2601 *
2602 * Note: in each "cycle of life" we will normally assign the same IPC keys
2603 * (if using SysV shmem and/or semas), since the port number is used to
2604 * determine IPC keys. This helps ensure that we will clean up dead IPC
2605 * objects if the postmaster crashes and is restarted.
2606 */
2607 CreateSharedMemoryAndSemaphores(port);
2608 }
2609
2610
2611 /*
2612 * SIGHUP -- reread config files, and tell children to do same
2613 */
2614 static void
2615 SIGHUP_handler(SIGNAL_ARGS)
2616 {
2617 int save_errno = errno;
2618
2619 /*
2620 * We rely on the signal mechanism to have blocked all signals ... except
2621 * on Windows, which lacks sigaction(), so we have to do it manually.
2622 */
2623 #ifdef WIN32
2624 PG_SETMASK(&BlockSig);
2625 #endif
2626
2627 if (Shutdown <= SmartShutdown)
2628 {
2629 ereport(LOG,
2630 (errmsg("received SIGHUP, reloading configuration files")));
2631 ProcessConfigFile(PGC_SIGHUP);
2632 SignalChildren(SIGHUP);
2633 if (StartupPID != 0)
2634 signal_child(StartupPID, SIGHUP);
2635 if (BgWriterPID != 0)
2636 signal_child(BgWriterPID, SIGHUP);
2637 if (CheckpointerPID != 0)
2638 signal_child(CheckpointerPID, SIGHUP);
2639 if (WalWriterPID != 0)
2640 signal_child(WalWriterPID, SIGHUP);
2641 if (WalReceiverPID != 0)
2642 signal_child(WalReceiverPID, SIGHUP);
2643 if (AutoVacPID != 0)
2644 signal_child(AutoVacPID, SIGHUP);
2645 if (PgArchPID != 0)
2646 signal_child(PgArchPID, SIGHUP);
2647 if (SysLoggerPID != 0)
2648 signal_child(SysLoggerPID, SIGHUP);
2649 if (PgStatPID != 0)
2650 signal_child(PgStatPID, SIGHUP);
2651
2652 /* Reload authentication config files too */
2653 if (!load_hba())
2654 ereport(LOG,
2655 (errmsg("pg_hba.conf was not reloaded")));
2656
2657 if (!load_ident())
2658 ereport(LOG,
2659 (errmsg("pg_ident.conf was not reloaded")));
2660
2661 #ifdef USE_SSL
2662 /* Reload SSL configuration as well */
2663 if (EnableSSL)
2664 {
2665 if (secure_initialize(false) == 0)
2666 LoadedSSL = true;
2667 else
2668 ereport(LOG,
2669 (errmsg("SSL configuration was not reloaded")));
2670 }
2671 else
2672 {
2673 secure_destroy();
2674 LoadedSSL = false;
2675 }
2676 #endif
2677
2678 #ifdef EXEC_BACKEND
2679 /* Update the starting-point file for future children */
2680 write_nondefault_variables(PGC_SIGHUP);
2681 #endif
2682 }
2683
2684 #ifdef WIN32
2685 PG_SETMASK(&UnBlockSig);
2686 #endif
2687
2688 errno = save_errno;
2689 }
2690
2691
2692 /*
2693 * pmdie -- signal handler for processing various postmaster signals.
2694 */
2695 static void
2696 pmdie(SIGNAL_ARGS)
2697 {
2698 int save_errno = errno;
2699
2700 /*
2701 * We rely on the signal mechanism to have blocked all signals ... except
2702 * on Windows, which lacks sigaction(), so we have to do it manually.
2703 */
2704 #ifdef WIN32
2705 PG_SETMASK(&BlockSig);
2706 #endif
2707
2708 ereport(DEBUG2,
2709 (errmsg_internal("postmaster received signal %d",
2710 postgres_signal_arg)));
2711
2712 switch (postgres_signal_arg)
2713 {
2714 case SIGTERM:
2715
2716 /*
2717 * Smart Shutdown:
2718 *
2719 * Wait for children to end their work, then shut down.
2720 */
2721 if (Shutdown >= SmartShutdown)
2722 break;
2723 Shutdown = SmartShutdown;
2724 ereport(LOG,
2725 (errmsg("received smart shutdown request")));
2726
2727 /* Report status */
2728 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2729 #ifdef USE_SYSTEMD
2730 sd_notify(0, "STOPPING=1");
2731 #endif
2732
2733 /*
2734 * If we reached normal running, we have to wait for any online
2735 * backup mode to end; otherwise go straight to waiting for client
2736 * backends to exit. (The difference is that in the former state,
2737 * we'll still let in new superuser clients, so that somebody can
2738 * end the online backup mode.) If already in PM_STOP_BACKENDS or
2739 * a later state, do not change it.
2740 */
2741 if (pmState == PM_RUN)
2742 connsAllowed = ALLOW_SUPERUSER_CONNS;
2743 else if (pmState == PM_HOT_STANDBY)
2744 connsAllowed = ALLOW_NO_CONNS;
2745 else if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2746 {
2747 /* There should be no clients, so proceed to stop children */
2748 pmState = PM_STOP_BACKENDS;
2749 }
2750
2751 /*
2752 * Now wait for online backup mode to end and backends to exit. If
2753 * that is already the case, PostmasterStateMachine will take the
2754 * next step.
2755 */
2756 PostmasterStateMachine();
2757 break;
2758
2759 case SIGINT:
2760
2761 /*
2762 * Fast Shutdown:
2763 *
2764 * Abort all children with SIGTERM (rollback active transactions
2765 * and exit) and shut down when they are gone.
2766 */
2767 if (Shutdown >= FastShutdown)
2768 break;
2769 Shutdown = FastShutdown;
2770 ereport(LOG,
2771 (errmsg("received fast shutdown request")));
2772
2773 /* Report status */
2774 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2775 #ifdef USE_SYSTEMD
2776 sd_notify(0, "STOPPING=1");
2777 #endif
2778
2779 if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2780 {
2781 /* Just shut down background processes silently */
2782 pmState = PM_STOP_BACKENDS;
2783 }
2784 else if (pmState == PM_RUN ||
2785 pmState == PM_HOT_STANDBY)
2786 {
2787 /* Report that we're about to zap live client sessions */
2788 ereport(LOG,
2789 (errmsg("aborting any active transactions")));
2790 pmState = PM_STOP_BACKENDS;
2791 }
2792
2793 /*
2794 * PostmasterStateMachine will issue any necessary signals, or
2795 * take the next step if no child processes need to be killed.
2796 */
2797 PostmasterStateMachine();
2798 break;
2799
2800 case SIGQUIT:
2801
2802 /*
2803 * Immediate Shutdown:
2804 *
2805 * abort all children with SIGQUIT, wait for them to exit,
2806 * terminate remaining ones with SIGKILL, then exit without
2807 * attempt to properly shut down the data base system.
2808 */
2809 if (Shutdown >= ImmediateShutdown)
2810 break;
2811 Shutdown = ImmediateShutdown;
2812 ereport(LOG,
2813 (errmsg("received immediate shutdown request")));
2814
2815 /* Report status */
2816 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2817 #ifdef USE_SYSTEMD
2818 sd_notify(0, "STOPPING=1");
2819 #endif
2820
2821 TerminateChildren(SIGQUIT);
2822 pmState = PM_WAIT_BACKENDS;
2823
2824 /* set stopwatch for them to die */
2825 AbortStartTime = time(NULL);
2826
2827 /*
2828 * Now wait for backends to exit. If there are none,
2829 * PostmasterStateMachine will take the next step.
2830 */
2831 PostmasterStateMachine();
2832 break;
2833 }
2834
2835 #ifdef WIN32
2836 PG_SETMASK(&UnBlockSig);
2837 #endif
2838
2839 errno = save_errno;
2840 }
2841
2842 /*
2843 * Reaper -- signal handler to cleanup after a child process dies.
2844 */
2845 static void
2846 reaper(SIGNAL_ARGS)
2847 {
2848 int save_errno = errno;
2849 int pid; /* process id of dead child process */
2850 int exitstatus; /* its exit status */
2851
2852 /*
2853 * We rely on the signal mechanism to have blocked all signals ... except
2854 * on Windows, which lacks sigaction(), so we have to do it manually.
2855 */
2856 #ifdef WIN32
2857 PG_SETMASK(&BlockSig);
2858 #endif
2859
2860 ereport(DEBUG4,
2861 (errmsg_internal("reaping dead processes")));
2862
2863 while ((pid = waitpid(-1, &exitstatus, WNOHANG)) > 0)
2864 {
2865 /*
2866 * Check if this child was a startup process.
2867 */
2868 if (pid == StartupPID)
2869 {
2870 StartupPID = 0;
2871
2872 /*
2873 * Startup process exited in response to a shutdown request (or it
2874 * completed normally regardless of the shutdown request).
2875 */
2876 if (Shutdown > NoShutdown &&
2877 (EXIT_STATUS_0(exitstatus) || EXIT_STATUS_1(exitstatus)))
2878 {
2879 StartupStatus = STARTUP_NOT_RUNNING;
2880 pmState = PM_WAIT_BACKENDS;
2881 /* PostmasterStateMachine logic does the rest */
2882 continue;
2883 }
2884
2885 if (EXIT_STATUS_3(exitstatus))
2886 {
2887 ereport(LOG,
2888 (errmsg("shutdown at recovery target")));
2889 StartupStatus = STARTUP_NOT_RUNNING;
2890 Shutdown = Max(Shutdown, SmartShutdown);
2891 TerminateChildren(SIGTERM);
2892 pmState = PM_WAIT_BACKENDS;
2893 /* PostmasterStateMachine logic does the rest */
2894 continue;
2895 }
2896
2897 /*
2898 * Unexpected exit of startup process (including FATAL exit)
2899 * during PM_STARTUP is treated as catastrophic. There are no
2900 * other processes running yet, so we can just exit.
2901 */
2902 if (pmState == PM_STARTUP && !EXIT_STATUS_0(exitstatus))
2903 {
2904 LogChildExit(LOG, _("startup process"),
2905 pid, exitstatus);
2906 ereport(LOG,
2907 (errmsg("aborting startup due to startup process failure")));
2908 ExitPostmaster(1);
2909 }
2910
2911 /*
2912 * After PM_STARTUP, any unexpected exit (including FATAL exit) of
2913 * the startup process is catastrophic, so kill other children,
2914 * and set StartupStatus so we don't try to reinitialize after
2915 * they're gone. Exception: if StartupStatus is STARTUP_SIGNALED,
2916 * then we previously sent the startup process a SIGQUIT; so
2917 * that's probably the reason it died, and we do want to try to
2918 * restart in that case.
2919 */
2920 if (!EXIT_STATUS_0(exitstatus))
2921 {
2922 if (StartupStatus == STARTUP_SIGNALED)
2923 StartupStatus = STARTUP_NOT_RUNNING;
2924 else
2925 StartupStatus = STARTUP_CRASHED;
2926 HandleChildCrash(pid, exitstatus,
2927 _("startup process"));
2928 continue;
2929 }
2930
2931 /*
2932 * Startup succeeded, commence normal operations
2933 */
2934 StartupStatus = STARTUP_NOT_RUNNING;
2935 FatalError = false;
2936 Assert(AbortStartTime == 0);
2937 ReachedNormalRunning = true;
2938 pmState = PM_RUN;
2939 connsAllowed = ALLOW_ALL_CONNS;
2940
2941 /*
2942 * Crank up the background tasks, if we didn't do that already
2943 * when we entered consistent recovery state. It doesn't matter
2944 * if this fails, we'll just try again later.
2945 */
2946 if (CheckpointerPID == 0)
2947 CheckpointerPID = StartCheckpointer();
2948 if (BgWriterPID == 0)
2949 BgWriterPID = StartBackgroundWriter();
2950 if (WalWriterPID == 0)
2951 WalWriterPID = StartWalWriter();
2952
2953 /*
2954 * Likewise, start other special children as needed. In a restart
2955 * situation, some of them may be alive already.
2956 */
2957 if (!IsBinaryUpgrade && AutoVacuumingActive() && AutoVacPID == 0)
2958 AutoVacPID = StartAutoVacLauncher();
2959 if (PgArchStartupAllowed() && PgArchPID == 0)
2960 PgArchPID = pgarch_start();
2961 if (PgStatPID == 0)
2962 PgStatPID = pgstat_start();
2963
2964 /* workers may be scheduled to start now */
2965 maybe_start_bgworkers();
2966
2967 /* at this point we are really open for business */
2968 ereport(LOG,
2969 (errmsg("database system is ready to accept connections")));
2970
2971 /* Report status */
2972 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
2973 #ifdef USE_SYSTEMD
2974 sd_notify(0, "READY=1");
2975 #endif
2976
2977 continue;
2978 }
2979
2980 /*
2981 * Was it the bgwriter? Normal exit can be ignored; we'll start a new
2982 * one at the next iteration of the postmaster's main loop, if
2983 * necessary. Any other exit condition is treated as a crash.
2984 */
2985 if (pid == BgWriterPID)
2986 {
2987 BgWriterPID = 0;
2988 if (!EXIT_STATUS_0(exitstatus))
2989 HandleChildCrash(pid, exitstatus,
2990 _("background writer process"));
2991 continue;
2992 }
2993
2994 /*
2995 * Was it the checkpointer?
2996 */
2997 if (pid == CheckpointerPID)
2998 {
2999 CheckpointerPID = 0;
3000 if (EXIT_STATUS_0(exitstatus) && pmState == PM_SHUTDOWN)
3001 {
3002 /*
3003 * OK, we saw normal exit of the checkpointer after it's been
3004 * told to shut down. We expect that it wrote a shutdown
3005 * checkpoint. (If for some reason it didn't, recovery will
3006 * occur on next postmaster start.)
3007 *
3008 * At this point we should have no normal backend children
3009 * left (else we'd not be in PM_SHUTDOWN state) but we might
3010 * have dead_end children to wait for.
3011 *
3012 * If we have an archiver subprocess, tell it to do a last
3013 * archive cycle and quit. Likewise, if we have walsender
3014 * processes, tell them to send any remaining WAL and quit.
3015 */
3016 Assert(Shutdown > NoShutdown);
3017
3018 /* Waken archiver for the last time */
3019 if (PgArchPID != 0)
3020 signal_child(PgArchPID, SIGUSR2);
3021
3022 /*
3023 * Waken walsenders for the last time. No regular backends
3024 * should be around anymore.
3025 */
3026 SignalChildren(SIGUSR2);
3027
3028 pmState = PM_SHUTDOWN_2;
3029
3030 /*
3031 * We can also shut down the stats collector now; there's
3032 * nothing left for it to do.
3033 */
3034 if (PgStatPID != 0)
3035 signal_child(PgStatPID, SIGQUIT);
3036 }
3037 else
3038 {
3039 /*
3040 * Any unexpected exit of the checkpointer (including FATAL
3041 * exit) is treated as a crash.
3042 */
3043 HandleChildCrash(pid, exitstatus,
3044 _("checkpointer process"));
3045 }
3046
3047 continue;
3048 }
3049
3050 /*
3051 * Was it the wal writer? Normal exit can be ignored; we'll start a
3052 * new one at the next iteration of the postmaster's main loop, if
3053 * necessary. Any other exit condition is treated as a crash.
3054 */
3055 if (pid == WalWriterPID)
3056 {
3057 WalWriterPID = 0;
3058 if (!EXIT_STATUS_0(exitstatus))
3059 HandleChildCrash(pid, exitstatus,
3060 _("WAL writer process"));
3061 continue;
3062 }
3063
3064 /*
3065 * Was it the wal receiver? If exit status is zero (normal) or one
3066 * (FATAL exit), we assume everything is all right just like normal
3067 * backends. (If we need a new wal receiver, we'll start one at the
3068 * next iteration of the postmaster's main loop.)
3069 */
3070 if (pid == WalReceiverPID)
3071 {
3072 WalReceiverPID = 0;
3073 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3074 HandleChildCrash(pid, exitstatus,
3075 _("WAL receiver process"));
3076 continue;
3077 }
3078
3079 /*
3080 * Was it the autovacuum launcher? Normal exit can be ignored; we'll
3081 * start a new one at the next iteration of the postmaster's main
3082 * loop, if necessary. Any other exit condition is treated as a
3083 * crash.
3084 */
3085 if (pid == AutoVacPID)
3086 {
3087 AutoVacPID = 0;
3088 if (!EXIT_STATUS_0(exitstatus))
3089 HandleChildCrash(pid, exitstatus,
3090 _("autovacuum launcher process"));
3091 continue;
3092 }
3093
3094 /*
3095 * Was it the archiver? If so, just try to start a new one; no need
3096 * to force reset of the rest of the system. (If fail, we'll try
3097 * again in future cycles of the main loop.). Unless we were waiting
3098 * for it to shut down; don't restart it in that case, and
3099 * PostmasterStateMachine() will advance to the next shutdown step.
3100 */
3101 if (pid == PgArchPID)
3102 {
3103 PgArchPID = 0;
3104 if (!EXIT_STATUS_0(exitstatus))
3105 LogChildExit(LOG, _("archiver process"),
3106 pid, exitstatus);
3107 if (PgArchStartupAllowed())
3108 PgArchPID = pgarch_start();
3109 continue;
3110 }
3111
3112 /*
3113 * Was it the statistics collector? If so, just try to start a new
3114 * one; no need to force reset of the rest of the system. (If fail,
3115 * we'll try again in future cycles of the main loop.)
3116 */
3117 if (pid == PgStatPID)
3118 {
3119 PgStatPID = 0;
3120 if (!EXIT_STATUS_0(exitstatus))
3121 LogChildExit(LOG, _("statistics collector process"),
3122 pid, exitstatus);
3123 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3124 PgStatPID = pgstat_start();
3125 continue;
3126 }
3127
3128 /* Was it the system logger? If so, try to start a new one */
3129 if (pid == SysLoggerPID)
3130 {
3131 SysLoggerPID = 0;
3132 /* for safety's sake, launch new logger *first* */
3133 SysLoggerPID = SysLogger_Start();
3134 if (!EXIT_STATUS_0(exitstatus))
3135 LogChildExit(LOG, _("system logger process"),
3136 pid, exitstatus);
3137 continue;
3138 }
3139
3140 /* Was it one of our background workers? */
3141 if (CleanupBackgroundWorker(pid, exitstatus))
3142 {
3143 /* have it be restarted */
3144 HaveCrashedWorker = true;
3145 continue;
3146 }
3147
3148 /*
3149 * Else do standard backend child cleanup.
3150 */
3151 CleanupBackend(pid, exitstatus);
3152 } /* loop over pending child-death reports */
3153
3154 /*
3155 * After cleaning out the SIGCHLD queue, see if we have any state changes
3156 * or actions to make.
3157 */
3158 PostmasterStateMachine();
3159
3160 /* Done with signal handler */
3161 #ifdef WIN32
3162 PG_SETMASK(&UnBlockSig);
3163 #endif
3164
3165 errno = save_errno;
3166 }
3167
3168 /*
3169 * Scan the bgworkers list and see if the given PID (which has just stopped
3170 * or crashed) is in it. Handle its shutdown if so, and return true. If not a
3171 * bgworker, return false.
3172 *
3173 * This is heavily based on CleanupBackend. One important difference is that
3174 * we don't know yet that the dying process is a bgworker, so we must be silent
3175 * until we're sure it is.
3176 */
3177 static bool
3178 CleanupBackgroundWorker(int pid,
3179 int exitstatus) /* child's exit status */
3180 {
3181 char namebuf[MAXPGPATH];
3182 slist_mutable_iter iter;
3183
3184 slist_foreach_modify(iter, &BackgroundWorkerList)
3185 {
3186 RegisteredBgWorker *rw;
3187
3188 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
3189
3190 if (rw->rw_pid != pid)
3191 continue;
3192
3193 #ifdef WIN32
3194 /* see CleanupBackend */
3195 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3196 exitstatus = 0;
3197 #endif
3198
3199 snprintf(namebuf, MAXPGPATH, "%s: %s", _("worker process"),
3200 rw->rw_worker.bgw_name);
3201
3202 if (!EXIT_STATUS_0(exitstatus))
3203 {
3204 /* Record timestamp, so we know when to restart the worker. */
3205 rw->rw_crashed_at = GetCurrentTimestamp();
3206 }
3207 else
3208 {
3209 /* Zero exit status means terminate */
3210 rw->rw_crashed_at = 0;
3211 rw->rw_terminate = true;
3212 }
3213
3214 /*
3215 * Additionally, for shared-memory-connected workers, just like a
3216 * backend, any exit status other than 0 or 1 is considered a crash
3217 * and causes a system-wide restart.
3218 */
3219 if ((rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3220 {
3221 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3222 {
3223 HandleChildCrash(pid, exitstatus, namebuf);
3224 return true;
3225 }
3226 }
3227
3228 /*
3229 * We must release the postmaster child slot whether this worker is
3230 * connected to shared memory or not, but we only treat it as a crash
3231 * if it is in fact connected.
3232 */
3233 if (!ReleasePostmasterChildSlot(rw->rw_child_slot) &&
3234 (rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3235 {
3236 HandleChildCrash(pid, exitstatus, namebuf);
3237 return true;
3238 }
3239
3240 /* Get it out of the BackendList and clear out remaining data */
3241 dlist_delete(&rw->rw_backend->elem);
3242 #ifdef EXEC_BACKEND
3243 ShmemBackendArrayRemove(rw->rw_backend);
3244 #endif
3245
3246 /*
3247 * It's possible that this background worker started some OTHER
3248 * background worker and asked to be notified when that worker started
3249 * or stopped. If so, cancel any notifications destined for the
3250 * now-dead backend.
3251 */
3252 if (rw->rw_backend->bgworker_notify)
3253 BackgroundWorkerStopNotifications(rw->rw_pid);
3254 free(rw->rw_backend);
3255 rw->rw_backend = NULL;
3256 rw->rw_pid = 0;
3257 rw->rw_child_slot = 0;
3258 ReportBackgroundWorkerExit(&iter); /* report child death */
3259
3260 LogChildExit(EXIT_STATUS_0(exitstatus) ? DEBUG1 : LOG,
3261 namebuf, pid, exitstatus);
3262
3263 return true;
3264 }
3265
3266 return false;
3267 }
3268
3269 /*
3270 * CleanupBackend -- cleanup after terminated backend.
3271 *
3272 * Remove all local state associated with backend.
3273 *
3274 * If you change this, see also CleanupBackgroundWorker.
3275 */
3276 static void
3277 CleanupBackend(int pid,
3278 int exitstatus) /* child's exit status. */
3279 {
3280 dlist_mutable_iter iter;
3281
3282 LogChildExit(DEBUG2, _("server process"), pid, exitstatus);
3283
3284 /*
3285 * If a backend dies in an ugly way then we must signal all other backends
3286 * to quickdie. If exit status is zero (normal) or one (FATAL exit), we
3287 * assume everything is all right and proceed to remove the backend from
3288 * the active backend list.
3289 */
3290
3291 #ifdef WIN32
3292
3293 /*
3294 * On win32, also treat ERROR_WAIT_NO_CHILDREN (128) as nonfatal case,
3295 * since that sometimes happens under load when the process fails to start
3296 * properly (long before it starts using shared memory). Microsoft reports
3297 * it is related to mutex failure:
3298 * http://archives.postgresql.org/pgsql-hackers/2010-09/msg00790.php
3299 */
3300 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3301 {
3302 LogChildExit(LOG, _("server process"), pid, exitstatus);
3303 exitstatus = 0;
3304 }
3305 #endif
3306
3307 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3308 {
3309 HandleChildCrash(pid, exitstatus, _("server process"));
3310 return;
3311 }
3312
3313 dlist_foreach_modify(iter, &BackendList)
3314 {
3315 Backend *bp = dlist_container(Backend, elem, iter.cur);
3316
3317 if (bp->pid == pid)
3318 {
3319 if (!bp->dead_end)
3320 {
3321 if (!ReleasePostmasterChildSlot(bp->child_slot))
3322 {
3323 /*
3324 * Uh-oh, the child failed to clean itself up. Treat as a
3325 * crash after all.
3326 */
3327 HandleChildCrash(pid, exitstatus, _("server process"));
3328 return;
3329 }
3330 #ifdef EXEC_BACKEND
3331 ShmemBackendArrayRemove(bp);
3332 #endif
3333 }
3334 if (bp->bgworker_notify)
3335 {
3336 /*
3337 * This backend may have been slated to receive SIGUSR1 when
3338 * some background worker started or stopped. Cancel those
3339 * notifications, as we don't want to signal PIDs that are not
3340 * PostgreSQL backends. This gets skipped in the (probably
3341 * very common) case where the backend has never requested any
3342 * such notifications.
3343 */
3344 BackgroundWorkerStopNotifications(bp->pid);
3345 }
3346 dlist_delete(iter.cur);
3347 free(bp);
3348 break;
3349 }
3350 }
3351 }
3352
3353 /*
3354 * HandleChildCrash -- cleanup after failed backend, bgwriter, checkpointer,
3355 * walwriter, autovacuum, or background worker.
3356 *
3357 * The objectives here are to clean up our local state about the child
3358 * process, and to signal all other remaining children to quickdie.
3359 */
3360 static void
3361 HandleChildCrash(int pid, int exitstatus, const char *procname)
3362 {
3363 dlist_mutable_iter iter;
3364 slist_iter siter;
3365 Backend *bp;
3366 bool take_action;
3367
3368 /*
3369 * We only log messages and send signals if this is the first process
3370 * crash and we're not doing an immediate shutdown; otherwise, we're only
3371 * here to update postmaster's idea of live processes. If we have already
3372 * signalled children, nonzero exit status is to be expected, so don't
3373 * clutter log.
3374 */
3375 take_action = !FatalError && Shutdown != ImmediateShutdown;
3376
3377 if (take_action)
3378 {
3379 LogChildExit(LOG, procname, pid, exitstatus);
3380 ereport(LOG,
3381 (errmsg("terminating any other active server processes")));
3382 }
3383
3384 /* Process background workers. */
3385 slist_foreach(siter, &BackgroundWorkerList)
3386 {
3387 RegisteredBgWorker *rw;
3388
3389 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
3390 if (rw->rw_pid == 0)
3391 continue; /* not running */
3392 if (rw->rw_pid == pid)
3393 {
3394 /*
3395 * Found entry for freshly-dead worker, so remove it.
3396 */
3397 (void) ReleasePostmasterChildSlot(rw->rw_child_slot);
3398 dlist_delete(&rw->rw_backend->elem);
3399 #ifdef EXEC_BACKEND
3400 ShmemBackendArrayRemove(rw->rw_backend);
3401 #endif
3402 free(rw->rw_backend);
3403 rw->rw_backend = NULL;
3404 rw->rw_pid = 0;
3405 rw->rw_child_slot = 0;
3406 /* don't reset crashed_at */
3407 /* don't report child stop, either */
3408 /* Keep looping so we can signal remaining workers */
3409 }
3410 else
3411 {
3412 /*
3413 * This worker is still alive. Unless we did so already, tell it
3414 * to commit hara-kiri.
3415 *
3416 * SIGQUIT is the special signal that says exit without proc_exit
3417 * and let the user know what's going on. But if SendStop is set
3418 * (-s on command line), then we send SIGSTOP instead, so that we
3419 * can get core dumps from all backends by hand.
3420 */
3421 if (take_action)
3422 {
3423 ereport(DEBUG2,
3424 (errmsg_internal("sending %s to process %d",
3425 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3426 (int) rw->rw_pid)));
3427 signal_child(rw->rw_pid, (SendStop ? SIGSTOP : SIGQUIT));
3428 }
3429 }
3430 }
3431
3432 /* Process regular backends */
3433 dlist_foreach_modify(iter, &BackendList)
3434 {
3435 bp = dlist_container(Backend, elem, iter.cur);
3436
3437 if (bp->pid == pid)
3438 {
3439 /*
3440 * Found entry for freshly-dead backend, so remove it.
3441 */
3442 if (!bp->dead_end)
3443 {
3444 (void) ReleasePostmasterChildSlot(bp->child_slot);
3445 #ifdef EXEC_BACKEND
3446 ShmemBackendArrayRemove(bp);
3447 #endif
3448 }
3449 dlist_delete(iter.cur);
3450 free(bp);
3451 /* Keep looping so we can signal remaining backends */
3452 }
3453 else
3454 {
3455 /*
3456 * This backend is still alive. Unless we did so already, tell it
3457 * to commit hara-kiri.
3458 *
3459 * SIGQUIT is the special signal that says exit without proc_exit
3460 * and let the user know what's going on. But if SendStop is set
3461 * (-s on command line), then we send SIGSTOP instead, so that we
3462 * can get core dumps from all backends by hand.
3463 *
3464 * We could exclude dead_end children here, but at least in the
3465 * SIGSTOP case it seems better to include them.
3466 *
3467 * Background workers were already processed above; ignore them
3468 * here.
3469 */
3470 if (bp->bkend_type == BACKEND_TYPE_BGWORKER)
3471 continue;
3472
3473 if (take_action)
3474 {
3475 ereport(DEBUG2,
3476 (errmsg_internal("sending %s to process %d",
3477 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3478 (int) bp->pid)));
3479 signal_child(bp->pid, (SendStop ? SIGSTOP : SIGQUIT));
3480 }
3481 }
3482 }
3483
3484 /* Take care of the startup process too */
3485 if (pid == StartupPID)
3486 {
3487 StartupPID = 0;
3488 StartupStatus = STARTUP_CRASHED;
3489 }
3490 else if (StartupPID != 0 && take_action)
3491 {
3492 ereport(DEBUG2,
3493 (errmsg_internal("sending %s to process %d",
3494 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3495 (int) StartupPID)));
3496 signal_child(StartupPID, (SendStop ? SIGSTOP : SIGQUIT));
3497 StartupStatus = STARTUP_SIGNALED;
3498 }
3499
3500 /* Take care of the bgwriter too */
3501 if (pid == BgWriterPID)
3502 BgWriterPID = 0;
3503 else if (BgWriterPID != 0 && take_action)
3504 {
3505 ereport(DEBUG2,
3506 (errmsg_internal("sending %s to process %d",
3507 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3508 (int) BgWriterPID)));
3509 signal_child(BgWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3510 }
3511
3512 /* Take care of the checkpointer too */
3513 if (pid == CheckpointerPID)
3514 CheckpointerPID = 0;
3515 else if (CheckpointerPID != 0 && take_action)
3516 {
3517 ereport(DEBUG2,
3518 (errmsg_internal("sending %s to process %d",
3519 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3520 (int) CheckpointerPID)));
3521 signal_child(CheckpointerPID, (SendStop ? SIGSTOP : SIGQUIT));
3522 }
3523
3524 /* Take care of the walwriter too */
3525 if (pid == WalWriterPID)
3526 WalWriterPID = 0;
3527 else if (WalWriterPID != 0 && take_action)
3528 {
3529 ereport(DEBUG2,
3530 (errmsg_internal("sending %s to process %d",
3531 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3532 (int) WalWriterPID)));
3533 signal_child(WalWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3534 }
3535
3536 /* Take care of the walreceiver too */
3537 if (pid == WalReceiverPID)
3538 WalReceiverPID = 0;
3539 else if (WalReceiverPID != 0 && take_action)
3540 {
3541 ereport(DEBUG2,
3542 (errmsg_internal("sending %s to process %d",
3543 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3544 (int) WalReceiverPID)));
3545 signal_child(WalReceiverPID, (SendStop ? SIGSTOP : SIGQUIT));
3546 }
3547
3548 /* Take care of the autovacuum launcher too */
3549 if (pid == AutoVacPID)
3550 AutoVacPID = 0;
3551 else if (AutoVacPID != 0 && take_action)
3552 {
3553 ereport(DEBUG2,
3554 (errmsg_internal("sending %s to process %d",
3555 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3556 (int) AutoVacPID)));
3557 signal_child(AutoVacPID, (SendStop ? SIGSTOP : SIGQUIT));
3558 }
3559
3560 /*
3561 * Force a power-cycle of the pgarch process too. (This isn't absolutely
3562 * necessary, but it seems like a good idea for robustness, and it
3563 * simplifies the state-machine logic in the case where a shutdown request
3564 * arrives during crash processing.)
3565 */
3566 if (PgArchPID != 0 && take_action)
3567 {
3568 ereport(DEBUG2,
3569 (errmsg_internal("sending %s to process %d",
3570 "SIGQUIT",
3571 (int) PgArchPID)));
3572 signal_child(PgArchPID, SIGQUIT);
3573 }
3574
3575 /*
3576 * Force a power-cycle of the pgstat process too. (This isn't absolutely
3577 * necessary, but it seems like a good idea for robustness, and it
3578 * simplifies the state-machine logic in the case where a shutdown request
3579 * arrives during crash processing.)
3580 */
3581 if (PgStatPID != 0 && take_action)
3582 {
3583 ereport(DEBUG2,
3584 (errmsg_internal("sending %s to process %d",
3585 "SIGQUIT",
3586 (int) PgStatPID)));
3587 signal_child(PgStatPID, SIGQUIT);
3588 allow_immediate_pgstat_restart();
3589 }
3590
3591 /* We do NOT restart the syslogger */
3592
3593 if (Shutdown != ImmediateShutdown)
3594 FatalError = true;
3595
3596 /* We now transit into a state of waiting for children to die */
3597 if (pmState == PM_RECOVERY ||
3598 pmState == PM_HOT_STANDBY ||
3599 pmState == PM_RUN ||
3600 pmState == PM_STOP_BACKENDS ||
3601 pmState == PM_SHUTDOWN)
3602 pmState = PM_WAIT_BACKENDS;
3603
3604 /*
3605 * .. and if this doesn't happen quickly enough, now the clock is ticking
3606 * for us to kill them without mercy.
3607 */
3608 if (AbortStartTime == 0)
3609 AbortStartTime = time(NULL);
3610 }
3611
3612 /*
3613 * Log the death of a child process.
3614 */
3615 static void
3616 LogChildExit(int lev, const char *procname, int pid, int exitstatus)
3617 {
3618 /*
3619 * size of activity_buffer is arbitrary, but set equal to default
3620 * track_activity_query_size
3621 */
3622 char activity_buffer[1024];
3623 const char *activity = NULL;
3624
3625 if (!EXIT_STATUS_0(exitstatus))
3626 activity = pgstat_get_crashed_backend_activity(pid,
3627 activity_buffer,
3628 sizeof(activity_buffer));
3629
3630 if (WIFEXITED(exitstatus))
3631 ereport(lev,
3632
3633 /*------
3634 translator: %s is a noun phrase describing a child process, such as
3635 "server process" */
3636 (errmsg("%s (PID %d) exited with exit code %d",
3637 procname, pid, WEXITSTATUS(exitstatus)),
3638 activity ? errdetail("Failed process was running: %s", activity) : 0));
3639 else if (WIFSIGNALED(exitstatus))
3640 {
3641 #if defined(WIN32)
3642 ereport(lev,
3643
3644 /*------
3645 translator: %s is a noun phrase describing a child process, such as
3646 "server process" */
3647 (errmsg("%s (PID %d) was terminated by exception 0x%X",
3648 procname, pid, WTERMSIG(exitstatus)),
3649 errhint("See C include file \"ntstatus.h\" for a description of the hexadecimal value."),
3650 activity ? errdetail("Failed process was running: %s", activity) : 0));
3651 #else
3652 ereport(lev,
3653
3654 /*------
3655 translator: %s is a noun phrase describing a child process, such as
3656 "server process" */
3657 (errmsg("%s (PID %d) was terminated by signal %d: %s",
3658 procname, pid, WTERMSIG(exitstatus),
3659 pg_strsignal(WTERMSIG(exitstatus))),
3660 activity ? errdetail("Failed process was running: %s", activity) : 0));
3661 #endif
3662 }
3663 else
3664 ereport(lev,
3665
3666 /*------
3667 translator: %s is a noun phrase describing a child process, such as
3668 "server process" */
3669 (errmsg("%s (PID %d) exited with unrecognized status %d",
3670 procname, pid, exitstatus),
3671 activity ? errdetail("Failed process was running: %s", activity) : 0));
3672 }
3673
3674 /*
3675 * Advance the postmaster's state machine and take actions as appropriate
3676 *
3677 * This is common code for pmdie(), reaper() and sigusr1_handler(), which
3678 * receive the signals that might mean we need to change state.
3679 */
3680 static void
3681 PostmasterStateMachine(void)
3682 {
3683 /* If we're doing a smart shutdown, try to advance that state. */
3684 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3685 {
3686 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
3687 {
3688 /*
3689 * ALLOW_SUPERUSER_CONNS state ends as soon as online backup mode
3690 * is not active.
3691 */
3692 if (!BackupInProgress())
3693 connsAllowed = ALLOW_NO_CONNS;
3694 }
3695
3696 if (connsAllowed == ALLOW_NO_CONNS)
3697 {
3698 /*
3699 * ALLOW_NO_CONNS state ends when we have no normal client
3700 * backends running. Then we're ready to stop other children.
3701 */
3702 if (CountChildren(BACKEND_TYPE_NORMAL) == 0)
3703 pmState = PM_STOP_BACKENDS;
3704 }
3705 }
3706
3707 /*
3708 * If we're ready to do so, signal child processes to shut down. (This
3709 * isn't a persistent state, but treating it as a distinct pmState allows
3710 * us to share this code across multiple shutdown code paths.)
3711 */
3712 if (pmState == PM_STOP_BACKENDS)
3713 {
3714 /*
3715 * Forget any pending requests for background workers, since we're no
3716 * longer willing to launch any new workers. (If additional requests
3717 * arrive, BackgroundWorkerStateChange will reject them.)
3718 */
3719 ForgetUnstartedBackgroundWorkers();
3720
3721 /* Signal all backend children except walsenders */
3722 SignalSomeChildren(SIGTERM,
3723 BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND);
3724 /* and the autovac launcher too */
3725 if (AutoVacPID != 0)
3726 signal_child(AutoVacPID, SIGTERM);
3727 /* and the bgwriter too */
3728 if (BgWriterPID != 0)
3729 signal_child(BgWriterPID, SIGTERM);
3730 /* and the walwriter too */
3731 if (WalWriterPID != 0)
3732 signal_child(WalWriterPID, SIGTERM);
3733 /* If we're in recovery, also stop startup and walreceiver procs */
3734 if (StartupPID != 0)
3735 signal_child(StartupPID, SIGTERM);
3736 if (WalReceiverPID != 0)
3737 signal_child(WalReceiverPID, SIGTERM);
3738 /* checkpointer, archiver, stats, and syslogger may continue for now */
3739
3740 /* Now transition to PM_WAIT_BACKENDS state to wait for them to die */
3741 pmState = PM_WAIT_BACKENDS;
3742 }
3743
3744 /*
3745 * If we are in a state-machine state that implies waiting for backends to
3746 * exit, see if they're all gone, and change state if so.
3747 */
3748 if (pmState == PM_WAIT_BACKENDS)
3749 {
3750 /*
3751 * PM_WAIT_BACKENDS state ends when we have no regular backends
3752 * (including autovac workers), no bgworkers (including unconnected
3753 * ones), and no walwriter, autovac launcher or bgwriter. If we are
3754 * doing crash recovery or an immediate shutdown then we expect the
3755 * checkpointer to exit as well, otherwise not. The archiver, stats,
3756 * and syslogger processes are disregarded since they are not
3757 * connected to shared memory; we also disregard dead_end children
3758 * here. Walsenders are also disregarded, they will be terminated
3759 * later after writing the checkpoint record, like the archiver
3760 * process.
3761 */
3762 if (CountChildren(BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND) == 0 &&
3763 StartupPID == 0 &&
3764 WalReceiverPID == 0 &&
3765 BgWriterPID == 0 &&
3766 (CheckpointerPID == 0 ||
3767 (!FatalError && Shutdown < ImmediateShutdown)) &&
3768 WalWriterPID == 0 &&
3769 AutoVacPID == 0)
3770 {
3771 if (Shutdown >= ImmediateShutdown || FatalError)
3772 {
3773 /*
3774 * Start waiting for dead_end children to die. This state
3775 * change causes ServerLoop to stop creating new ones.
3776 */
3777 pmState = PM_WAIT_DEAD_END;
3778
3779 /*
3780 * We already SIGQUIT'd the archiver and stats processes, if
3781 * any, when we started immediate shutdown or entered
3782 * FatalError state.
3783 */
3784 }
3785 else
3786 {
3787 /*
3788 * If we get here, we are proceeding with normal shutdown. All
3789 * the regular children are gone, and it's time to tell the
3790 * checkpointer to do a shutdown checkpoint.
3791 */
3792 Assert(Shutdown > NoShutdown);
3793 /* Start the checkpointer if not running */
3794 if (CheckpointerPID == 0)
3795 CheckpointerPID = StartCheckpointer();
3796 /* And tell it to shut down */
3797 if (CheckpointerPID != 0)
3798 {
3799 signal_child(CheckpointerPID, SIGUSR2);
3800 pmState = PM_SHUTDOWN;
3801 }
3802 else
3803 {
3804 /*
3805 * If we failed to fork a checkpointer, just shut down.
3806 * Any required cleanup will happen at next restart. We
3807 * set FatalError so that an "abnormal shutdown" message
3808 * gets logged when we exit.
3809 */
3810 FatalError = true;
3811 pmState = PM_WAIT_DEAD_END;
3812
3813 /* Kill the walsenders, archiver and stats collector too */
3814 SignalChildren(SIGQUIT);
3815 if (PgArchPID != 0)
3816 signal_child(PgArchPID, SIGQUIT);
3817 if (PgStatPID != 0)
3818 signal_child(PgStatPID, SIGQUIT);
3819 }
3820 }
3821 }
3822 }
3823
3824 if (pmState == PM_SHUTDOWN_2)
3825 {
3826 /*
3827 * PM_SHUTDOWN_2 state ends when there's no other children than
3828 * dead_end children left. There shouldn't be any regular backends
3829 * left by now anyway; what we're really waiting for is walsenders and
3830 * archiver.
3831 */
3832 if (PgArchPID == 0 && CountChildren(BACKEND_TYPE_ALL) == 0)
3833 {
3834 pmState = PM_WAIT_DEAD_END;
3835 }
3836 }
3837
3838 if (pmState == PM_WAIT_DEAD_END)
3839 {
3840 /*
3841 * PM_WAIT_DEAD_END state ends when the BackendList is entirely empty
3842 * (ie, no dead_end children remain), and the archiver and stats
3843 * collector are gone too.
3844 *
3845 * The reason we wait for those two is to protect them against a new
3846 * postmaster starting conflicting subprocesses; this isn't an
3847 * ironclad protection, but it at least helps in the
3848 * shutdown-and-immediately-restart scenario. Note that they have
3849 * already been sent appropriate shutdown signals, either during a
3850 * normal state transition leading up to PM_WAIT_DEAD_END, or during
3851 * FatalError processing.
3852 */
3853 if (dlist_is_empty(&BackendList) &&
3854 PgArchPID == 0 && PgStatPID == 0)
3855 {
3856 /* These other guys should be dead already */
3857 Assert(StartupPID == 0);
3858 Assert(WalReceiverPID == 0);
3859 Assert(BgWriterPID == 0);
3860 Assert(CheckpointerPID == 0);
3861 Assert(WalWriterPID == 0);
3862 Assert(AutoVacPID == 0);
3863 /* syslogger is not considered here */
3864 pmState = PM_NO_CHILDREN;
3865 }
3866 }
3867
3868 /*
3869 * If we've been told to shut down, we exit as soon as there are no
3870 * remaining children. If there was a crash, cleanup will occur at the
3871 * next startup. (Before PostgreSQL 8.3, we tried to recover from the
3872 * crash before exiting, but that seems unwise if we are quitting because
3873 * we got SIGTERM from init --- there may well not be time for recovery
3874 * before init decides to SIGKILL us.)
3875 *
3876 * Note that the syslogger continues to run. It will exit when it sees
3877 * EOF on its input pipe, which happens when there are no more upstream
3878 * processes.
3879 */
3880 if (Shutdown > NoShutdown && pmState == PM_NO_CHILDREN)
3881 {
3882 if (FatalError)
3883 {
3884 ereport(LOG, (errmsg("abnormal database system shutdown")));
3885 ExitPostmaster(1);
3886 }
3887 else
3888 {
3889 /*
3890 * Terminate exclusive backup mode to avoid recovery after a clean
3891 * fast shutdown. Since an exclusive backup can only be taken
3892 * during normal running (and not, for example, while running
3893 * under Hot Standby) it only makes sense to do this if we reached
3894 * normal running. If we're still in recovery, the backup file is
3895 * one we're recovering *from*, and we must keep it around so that
3896 * recovery restarts from the right place.
3897 */
3898 if (ReachedNormalRunning)
3899 CancelBackup();
3900
3901 /* Normal exit from the postmaster is here */
3902 ExitPostmaster(0);
3903 }
3904 }
3905
3906 /*
3907 * If the startup process failed, or the user does not want an automatic
3908 * restart after backend crashes, wait for all non-syslogger children to
3909 * exit, and then exit postmaster. We don't try to reinitialize when the
3910 * startup process fails, because more than likely it will just fail again
3911 * and we will keep trying forever.
3912 */
3913 if (pmState == PM_NO_CHILDREN &&
3914 (StartupStatus == STARTUP_CRASHED || !restart_after_crash))
3915 ExitPostmaster(1);
3916
3917 /*
3918 * If we need to recover from a crash, wait for all non-syslogger children
3919 * to exit, then reset shmem and StartupDataBase.
3920 */
3921 if (FatalError && pmState == PM_NO_CHILDREN)
3922 {
3923 ereport(LOG,
3924 (errmsg("all server processes terminated; reinitializing")));
3925
3926 /* allow background workers to immediately restart */
3927 ResetBackgroundWorkerCrashTimes();
3928
3929 shmem_exit(1);
3930 reset_shared(PostPortNumber);
3931
3932 StartupPID = StartupDataBase();
3933 Assert(StartupPID != 0);
3934 StartupStatus = STARTUP_RUNNING;
3935 pmState = PM_STARTUP;
3936 /* crash recovery started, reset SIGKILL flag */
3937 AbortStartTime = 0;
3938 }
3939 }
3940
3941
3942 /*
3943 * Send a signal to a postmaster child process
3944 *
3945 * On systems that have setsid(), each child process sets itself up as a
3946 * process group leader. For signals that are generally interpreted in the
3947 * appropriate fashion, we signal the entire process group not just the
3948 * direct child process. This allows us to, for example, SIGQUIT a blocked
3949 * archive_recovery script, or SIGINT a script being run by a backend via
3950 * system().
3951 *
3952 * There is a race condition for recently-forked children: they might not
3953 * have executed setsid() yet. So we signal the child directly as well as
3954 * the group. We assume such a child will handle the signal before trying
3955 * to spawn any grandchild processes. We also assume that signaling the
3956 * child twice will not cause any problems.
3957 */
3958 static void
3959 signal_child(pid_t pid, int signal)
3960 {
3961 if (kill(pid, signal) < 0)
3962 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) pid, signal);
3963 #ifdef HAVE_SETSID
3964 switch (signal)
3965 {
3966 case SIGINT:
3967 case SIGTERM:
3968 case SIGQUIT:
3969 case SIGSTOP:
3970 case SIGKILL:
3971 if (kill(-pid, signal) < 0)
3972 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) (-pid), signal);
3973 break;
3974 default:
3975 break;
3976 }
3977 #endif
3978 }
3979
3980 /*
3981 * Send a signal to the targeted children (but NOT special children;
3982 * dead_end children are never signaled, either).
3983 */
3984 static bool
3985 SignalSomeChildren(int signal, int target)
3986 {
3987 dlist_iter iter;
3988 bool signaled = false;
3989
3990 dlist_foreach(iter, &BackendList)
3991 {
3992 Backend *bp = dlist_container(Backend, elem, iter.cur);
3993
3994 if (bp->dead_end)
3995 continue;
3996
3997 /*
3998 * Since target == BACKEND_TYPE_ALL is the most common case, we test
3999 * it first and avoid touching shared memory for every child.
4000 */
4001 if (target != BACKEND_TYPE_ALL)
4002 {
4003 /*
4004 * Assign bkend_type for any recently announced WAL Sender
4005 * processes.
4006 */
4007 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
4008 IsPostmasterChildWalSender(bp->child_slot))
4009 bp->bkend_type = BACKEND_TYPE_WALSND;
4010
4011 if (!(target & bp->bkend_type))
4012 continue;
4013 }
4014
4015 ereport(DEBUG4,
4016 (errmsg_internal("sending signal %d to process %d",
4017 signal, (int) bp->pid)));
4018 signal_child(bp->pid, signal);
4019 signaled = true;
4020 }
4021 return signaled;
4022 }
4023
4024 /*
4025 * Send a termination signal to children. This considers all of our children
4026 * processes, except syslogger and dead_end backends.
4027 */
4028 static void
4029 TerminateChildren(int signal)
4030 {
4031 SignalChildren(signal);
4032 if (StartupPID != 0)
4033 {
4034 signal_child(StartupPID, signal);
4035 if (signal == SIGQUIT || signal == SIGKILL)
4036 StartupStatus = STARTUP_SIGNALED;
4037 }
4038 if (BgWriterPID != 0)
4039 signal_child(BgWriterPID, signal);
4040 if (CheckpointerPID != 0)
4041 signal_child(CheckpointerPID, signal);
4042 if (WalWriterPID != 0)
4043 signal_child(WalWriterPID, signal);
4044 if (WalReceiverPID != 0)
4045 signal_child(WalReceiverPID, signal);
4046 if (AutoVacPID != 0)
4047 signal_child(AutoVacPID, signal);
4048 if (PgArchPID != 0)
4049 signal_child(PgArchPID, signal);
4050 if (PgStatPID != 0)
4051 signal_child(PgStatPID, signal);
4052 }
4053
4054 /*
4055 * BackendStartup -- start backend process
4056 *
4057 * returns: STATUS_ERROR if the fork failed, STATUS_OK otherwise.
4058 *
4059 * Note: if you change this code, also consider StartAutovacuumWorker.
4060 */
4061 static int
4062 BackendStartup(Port *port)
4063 {
4064 Backend *bn; /* for backend cleanup */
4065 pid_t pid;
4066
4067 /*
4068 * Create backend data structure. Better before the fork() so we can
4069 * handle failure cleanly.
4070 */
4071 bn = (Backend *) malloc(sizeof(Backend));
4072 if (!bn)
4073 {
4074 ereport(LOG,
4075 (errcode(ERRCODE_OUT_OF_MEMORY),
4076 errmsg("out of memory")));
4077 return STATUS_ERROR;
4078 }
4079
4080 /*
4081 * Compute the cancel key that will be assigned to this backend. The
4082 * backend will have its own copy in the forked-off process' value of
4083 * MyCancelKey, so that it can transmit the key to the frontend.
4084 */
4085 if (!RandomCancelKey(&MyCancelKey))
4086 {
4087 free(bn);
4088 ereport(LOG,
4089 (errcode(ERRCODE_INTERNAL_ERROR),
4090 errmsg("could not generate random cancel key")));
4091 return STATUS_ERROR;
4092 }
4093
4094 bn->cancel_key = MyCancelKey;
4095
4096 /* Pass down canAcceptConnections state */
4097 port->canAcceptConnections = canAcceptConnections(BACKEND_TYPE_NORMAL);
4098 bn->dead_end = (port->canAcceptConnections != CAC_OK &&
4099 port->canAcceptConnections != CAC_SUPERUSER);
4100
4101 /*
4102 * Unless it's a dead_end child, assign it a child slot number
4103 */
4104 if (!bn->dead_end)
4105 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
4106 else
4107 bn->child_slot = 0;
4108
4109 /* Hasn't asked to be notified about any bgworkers yet */
4110 bn->bgworker_notify = false;
4111
4112 #ifdef EXEC_BACKEND
4113 pid = backend_forkexec(port);
4114 #else /* !EXEC_BACKEND */
4115 pid = fork_process();
4116 if (pid == 0) /* child */
4117 {
4118 free(bn);
4119
4120 /* Detangle from postmaster */
4121 InitPostmasterChild();
4122
4123 /* Close the postmaster's sockets */
4124 ClosePostmasterPorts(false);
4125
4126 /* Perform additional initialization and collect startup packet */
4127 BackendInitialize(port);
4128
4129 /* And run the backend */
4130 BackendRun(port);
4131 }
4132 #endif /* EXEC_BACKEND */
4133
4134 if (pid < 0)
4135 {
4136 /* in parent, fork failed */
4137 int save_errno = errno;
4138
4139 if (!bn->dead_end)
4140 (void) ReleasePostmasterChildSlot(bn->child_slot);
4141 free(bn);
4142 errno = save_errno;
4143 ereport(LOG,
4144 (errmsg("could not fork new process for connection: %m")));
4145 report_fork_failure_to_client(port, save_errno);
4146 return STATUS_ERROR;
4147 }
4148
4149 /* in parent, successful fork */
4150 ereport(DEBUG2,
4151 (errmsg_internal("forked new backend, pid=%d socket=%d",
4152 (int) pid, (int) port->sock)));
4153
4154 /*
4155 * Everything's been successful, it's safe to add this backend to our list
4156 * of backends.
4157 */
4158 bn->pid = pid;
4159 bn->bkend_type = BACKEND_TYPE_NORMAL; /* Can change later to WALSND */
4160 dlist_push_head(&BackendList, &bn->elem);
4161
4162 #ifdef EXEC_BACKEND
4163 if (!bn->dead_end)
4164 ShmemBackendArrayAdd(bn);
4165 #endif
4166
4167 return STATUS_OK;
4168 }
4169
4170 /*
4171 * Try to report backend fork() failure to client before we close the
4172 * connection. Since we do not care to risk blocking the postmaster on
4173 * this connection, we set the connection to non-blocking and try only once.
4174 *
4175 * This is grungy special-purpose code; we cannot use backend libpq since
4176 * it's not up and running.
4177 */
4178 static void
4179 report_fork_failure_to_client(Port *port, int errnum)
4180 {
4181 char buffer[1000];
4182 int rc;
4183
4184 /* Format the error message packet (always V2 protocol) */
4185 snprintf(buffer, sizeof(buffer), "E%s%s\n",
4186 _("could not fork new process for connection: "),
4187 strerror(errnum));
4188
4189 /* Set port to non-blocking. Don't do send() if this fails */
4190 if (!pg_set_noblock(port->sock))
4191 return;
4192
4193 /* We'll retry after EINTR, but ignore all other failures */
4194 do
4195 {
4196 rc = send(port->sock, buffer, strlen(buffer) + 1, 0);
4197 } while (rc < 0 && errno == EINTR);
4198 }
4199
4200
4201 /*
4202 * BackendInitialize -- initialize an interactive (postmaster-child)
4203 * backend process, and collect the client's startup packet.
4204 *
4205 * returns: nothing. Will not return at all if there's any failure.
4206 *
4207 * Note: this code does not depend on having any access to shared memory.
4208 * In the EXEC_BACKEND case, we are physically attached to shared memory
4209 * but have not yet set up most of our local pointers to shmem structures.
4210 */
4211 static void
4212 BackendInitialize(Port *port)
4213 {
4214 int status;
4215 int ret;
4216 char remote_host[NI_MAXHOST];
4217 char remote_port[NI_MAXSERV];
4218 char remote_ps_data[NI_MAXHOST];
4219
4220 /* Save port etc. for ps status */
4221 MyProcPort = port;
4222
4223 /*
4224 * PreAuthDelay is a debugging aid for investigating problems in the
4225 * authentication cycle: it can be set in postgresql.conf to allow time to
4226 * attach to the newly-forked backend with a debugger. (See also
4227 * PostAuthDelay, which we allow clients to pass through PGOPTIONS, but it
4228 * is not honored until after authentication.)
4229 */
4230 if (PreAuthDelay > 0)
4231 pg_usleep(PreAuthDelay * 1000000L);
4232
4233 /* This flag will remain set until InitPostgres finishes authentication */
4234 ClientAuthInProgress = true; /* limit visibility of log messages */
4235
4236 /* save process start time */
4237 port->SessionStartTime = GetCurrentTimestamp();
4238 MyStartTime = timestamptz_to_time_t(port->SessionStartTime);
4239
4240 /* set these to empty in case they are needed before we set them up */
4241 port->remote_host = "";
4242 port->remote_port = "";
4243
4244 /*
4245 * Initialize libpq and enable reporting of ereport errors to the client.
4246 * Must do this now because authentication uses libpq to send messages.
4247 */
4248 pq_init(); /* initialize libpq to talk to client */
4249 whereToSendOutput = DestRemote; /* now safe to ereport to client */
4250
4251 /*
4252 * We arrange to do proc_exit(1) if we receive SIGTERM or timeout while
4253 * trying to collect the startup packet; while SIGQUIT results in
4254 * _exit(2). Otherwise the postmaster cannot shutdown the database FAST
4255 * or IMMED cleanly if a buggy client fails to send the packet promptly.
4256 *
4257 * XXX this is pretty dangerous; signal handlers should not call anything
4258 * as complex as proc_exit() directly. We minimize the hazard by not
4259 * keeping these handlers active for longer than we must. However, it
4260 * seems necessary to be able to escape out of DNS lookups as well as the
4261 * startup packet reception proper, so we can't narrow the scope further
4262 * than is done here.
4263 *
4264 * XXX it follows that the remainder of this function must tolerate losing
4265 * control at any instant. Likewise, any pg_on_exit_callback registered
4266 * before or during this function must be prepared to execute at any
4267 * instant between here and the end of this function. Furthermore,
4268 * affected callbacks execute partially or not at all when a second
4269 * exit-inducing signal arrives after proc_exit_prepare() decrements
4270 * on_proc_exit_index. (Thanks to that mechanic, callbacks need not
4271 * anticipate more than one call.) This is fragile; it ought to instead
4272 * follow the norm of handling interrupts at selected, safe opportunities.
4273 */
4274 pqsignal(SIGTERM, process_startup_packet_die);
4275 pqsignal(SIGQUIT, process_startup_packet_quickdie);
4276 InitializeTimeouts(); /* establishes SIGALRM handler */
4277 PG_SETMASK(&StartupBlockSig);
4278
4279 /*
4280 * Get the remote host name and port for logging and status display.
4281 */
4282 remote_host[0] = '\0';
4283 remote_port[0] = '\0';
4284 if ((ret = pg_getnameinfo_all(&port->raddr.addr, port->raddr.salen,
4285 remote_host, sizeof(remote_host),
4286 remote_port, sizeof(remote_port),
4287 (log_hostname ? 0 : NI_NUMERICHOST) | NI_NUMERICSERV)) != 0)
4288 ereport(WARNING,
4289 (errmsg_internal("pg_getnameinfo_all() failed: %s",
4290 gai_strerror(ret))));
4291 if (remote_port[0] == '\0')
4292 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s", remote_host);
4293 else
4294 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s(%s)", remote_host, remote_port);
4295
4296 /*
4297 * Save remote_host and remote_port in port structure (after this, they
4298 * will appear in log_line_prefix data for log messages).
4299 */
4300 port->remote_host = strdup(remote_host);
4301 port->remote_port = strdup(remote_port);
4302
4303 /* And now we can issue the Log_connections message, if wanted */
4304 if (Log_connections)
4305 {
4306 if (remote_port[0])
4307 ereport(LOG,
4308 (errmsg("connection received: host=%s port=%s",
4309 remote_host,
4310 remote_port)));
4311 else
4312 ereport(LOG,
4313 (errmsg("connection received: host=%s",
4314 remote_host)));
4315 }
4316
4317 /*
4318 * If we did a reverse lookup to name, we might as well save the results
4319 * rather than possibly repeating the lookup during authentication.
4320 *
4321 * Note that we don't want to specify NI_NAMEREQD above, because then we'd
4322 * get nothing useful for a client without an rDNS entry. Therefore, we
4323 * must check whether we got a numeric IPv4 or IPv6 address, and not save
4324 * it into remote_hostname if so. (This test is conservative and might
4325 * sometimes classify a hostname as numeric, but an error in that
4326 * direction is safe; it only results in a possible extra lookup.)
4327 */
4328 if (log_hostname &&
4329 ret == 0 &&
4330 strspn(remote_host, "0123456789.") < strlen(remote_host) &&
4331 strspn(remote_host, "0123456789ABCDEFabcdef:") < strlen(remote_host))
4332 port->remote_hostname = strdup(remote_host);
4333
4334 /*
4335 * Ready to begin client interaction. We will give up and proc_exit(1)
4336 * after a time delay, so that a broken client can't hog a connection
4337 * indefinitely. PreAuthDelay and any DNS interactions above don't count
4338 * against the time limit.
4339 *
4340 * Note: AuthenticationTimeout is applied here while waiting for the
4341 * startup packet, and then again in InitPostgres for the duration of any
4342 * authentication operations. So a hostile client could tie up the
4343 * process for nearly twice AuthenticationTimeout before we kick him off.
4344 *
4345 * Note: because PostgresMain will call InitializeTimeouts again, the
4346 * registration of STARTUP_PACKET_TIMEOUT will be lost. This is okay
4347 * since we never use it again after this function.
4348 */
4349 RegisterTimeout(STARTUP_PACKET_TIMEOUT, StartupPacketTimeoutHandler);
4350 enable_timeout_after(STARTUP_PACKET_TIMEOUT, AuthenticationTimeout * 1000);
4351
4352 /*
4353 * Receive the startup packet (which might turn out to be a cancel request
4354 * packet).
4355 */
4356 status = ProcessStartupPacket(port, false);
4357
4358 /*
4359 * Disable the timeout, and prevent SIGTERM/SIGQUIT again.
4360 */
4361 disable_timeout(STARTUP_PACKET_TIMEOUT, false);
4362 PG_SETMASK(&BlockSig);
4363
4364 /*
4365 * Stop here if it was bad or a cancel packet. ProcessStartupPacket
4366 * already did any appropriate error reporting.
4367 */
4368 if (status != STATUS_OK)
4369 proc_exit(0);
4370
4371 /*
4372 * Now that we have the user and database name, we can set the process
4373 * title for ps. It's good to do this as early as possible in startup.
4374 *
4375 * For a walsender, the ps display is set in the following form:
4376 *
4377 * postgres: wal sender process <user> <host> <activity>
4378 *
4379 * To achieve that, we pass "wal sender process" as username and username
4380 * as dbname to init_ps_display(). XXX: should add a new variant of
4381 * init_ps_display() to avoid abusing the parameters like this.
4382 */
4383 if (am_walsender)
4384 init_ps_display("wal sender process", port->user_name, remote_ps_data,
4385 update_process_title ? "authentication" : "");
4386 else
4387 init_ps_display(port->user_name, port->database_name, remote_ps_data,
4388 update_process_title ? "authentication" : "");
4389 }
4390
4391
4392 /*
4393 * BackendRun -- set up the backend's argument list and invoke PostgresMain()
4394 *
4395 * returns:
4396 * Shouldn't return at all.
4397 * If PostgresMain() fails, return status.
4398 */
4399 static void
4400 BackendRun(Port *port)
4401 {
4402 char **av;
4403 int maxac;
4404 int ac;
4405 long secs;
4406 int usecs;
4407 int i;
4408
4409 /*
4410 * Don't want backend to be able to see the postmaster random number
4411 * generator state. We have to clobber the static random_seed *and* start
4412 * a new random sequence in the random() library function.
4413 */
4414 #ifndef HAVE_STRONG_RANDOM
4415 random_seed = 0;
4416 random_start_time.tv_usec = 0;
4417 #endif
4418 /* slightly hacky way to convert timestamptz into integers */
4419 TimestampDifference(0, port->SessionStartTime, &secs, &usecs);
4420 srandom((unsigned int) (MyProcPid ^ (usecs << 12) ^ secs));
4421
4422 /*
4423 * Now, build the argv vector that will be given to PostgresMain.
4424 *
4425 * The maximum possible number of commandline arguments that could come
4426 * from ExtraOptions is (strlen(ExtraOptions) + 1) / 2; see
4427 * pg_split_opts().
4428 */
4429 maxac = 2; /* for fixed args supplied below */
4430 maxac += (strlen(ExtraOptions) + 1) / 2;
4431
4432 av = (char **) MemoryContextAlloc(TopMemoryContext,
4433 maxac * sizeof(char *));
4434 ac = 0;
4435
4436 av[ac++] = "postgres";
4437
4438 /*
4439 * Pass any backend switches specified with -o on the postmaster's own
4440 * command line. We assume these are secure.
4441 */
4442 pg_split_opts(av, &ac, ExtraOptions);
4443
4444 av[ac] = NULL;
4445
4446 Assert(ac < maxac);
4447
4448 /*
4449 * Debug: print arguments being passed to backend
4450 */
4451 ereport(DEBUG3,
4452 (errmsg_internal("%s child[%d]: starting with (",
4453 progname, (int) getpid())));
4454 for (i = 0; i < ac; ++i)
4455 ereport(DEBUG3,
4456 (errmsg_internal("\t%s", av[i])));
4457 ereport(DEBUG3,
4458 (errmsg_internal(")")));
4459
4460 /*
4461 * Make sure we aren't in PostmasterContext anymore. (We can't delete it
4462 * just yet, though, because InitPostgres will need the HBA data.)
4463 */
4464 MemoryContextSwitchTo(TopMemoryContext);
4465
4466 PostgresMain(ac, av, port->database_name, port->user_name);
4467 }
4468
4469
4470 #ifdef EXEC_BACKEND
4471
4472 /*
4473 * postmaster_forkexec -- fork and exec a postmaster subprocess
4474 *
4475 * The caller must have set up the argv array already, except for argv[2]
4476 * which will be filled with the name of the temp variable file.
4477 *
4478 * Returns the child process PID, or -1 on fork failure (a suitable error
4479 * message has been logged on failure).
4480 *
4481 * All uses of this routine will dispatch to SubPostmasterMain in the
4482 * child process.
4483 */
4484 pid_t
4485 postmaster_forkexec(int argc, char *argv[])
4486 {
4487 Port port;
4488
4489 /* This entry point passes dummy values for the Port variables */
4490 memset(&port, 0, sizeof(port));
4491 return internal_forkexec(argc, argv, &port);
4492 }
4493
4494 /*
4495 * backend_forkexec -- fork/exec off a backend process
4496 *
4497 * Some operating systems (WIN32) don't have fork() so we have to simulate
4498 * it by storing parameters that need to be passed to the child and
4499 * then create a new child process.
4500 *
4501 * returns the pid of the fork/exec'd process, or -1 on failure
4502 */
4503 static pid_t
4504 backend_forkexec(Port *port)
4505 {
4506 char *av[4];
4507 int ac = 0;
4508
4509 av[ac++] = "postgres";
4510 av[ac++] = "--forkbackend";
4511 av[ac++] = NULL; /* filled in by internal_forkexec */
4512
4513 av[ac] = NULL;
4514 Assert(ac < lengthof(av));
4515
4516 return internal_forkexec(ac, av, port);
4517 }
4518
4519 #ifndef WIN32
4520
4521 /*
4522 * internal_forkexec non-win32 implementation
4523 *
4524 * - writes out backend variables to the parameter file
4525 * - fork():s, and then exec():s the child process
4526 */
4527 static pid_t
4528 internal_forkexec(int argc, char *argv[], Port *port)
4529 {
4530 static unsigned long tmpBackendFileNum = 0;
4531 pid_t pid;
4532 char tmpfilename[MAXPGPATH];
4533 BackendParameters param;
4534 FILE *fp;
4535
4536 if (!save_backend_variables(¶m, port))
4537 return -1; /* log made by save_backend_variables */
4538
4539 /* Calculate name for temp file */
4540 snprintf(tmpfilename, MAXPGPATH, "%s/%s.backend_var.%d.%lu",
4541 PG_TEMP_FILES_DIR, PG_TEMP_FILE_PREFIX,
4542 MyProcPid, ++tmpBackendFileNum);
4543
4544 /* Open file */
4545 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4546 if (!fp)
4547 {
4548 /*
4549 * As in OpenTemporaryFileInTablespace, try to make the temp-file
4550 * directory
4551 */
4552 mkdir(PG_TEMP_FILES_DIR, S_IRWXU);
4553
4554 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4555 if (!fp)
4556 {
4557 ereport(LOG,
4558 (errcode_for_file_access(),
4559 errmsg("could not create file \"%s\": %m",
4560 tmpfilename)));
4561 return -1;
4562 }
4563 }
4564
4565 if (fwrite(¶m, sizeof(param), 1, fp) != 1)
4566 {
4567 ereport(LOG,
4568 (errcode_for_file_access(),
4569 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4570 FreeFile(fp);
4571 return -1;
4572 }
4573
4574 /* Release file */
4575 if (FreeFile(fp))
4576 {
4577 ereport(LOG,
4578 (errcode_for_file_access(),
4579 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4580 return -1;
4581 }
4582
4583 /* Make sure caller set up argv properly */
4584 Assert(argc >= 3);
4585 Assert(argv[argc] == NULL);
4586 Assert(strncmp(argv[1], "--fork", 6) == 0);
4587 Assert(argv[2] == NULL);
4588
4589 /* Insert temp file name after --fork argument */
4590 argv[2] = tmpfilename;
4591
4592 /* Fire off execv in child */
4593 if ((pid = fork_process()) == 0)
4594 {
4595 if (execv(postgres_exec_path, argv) < 0)
4596 {
4597 ereport(LOG,
4598 (errmsg("could not execute server process \"%s\": %m",
4599 postgres_exec_path)));
4600 /* We're already in the child process here, can't return */
4601 exit(1);
4602 }
4603 }
4604
4605 return pid; /* Parent returns pid, or -1 on fork failure */
4606 }
4607 #else /* WIN32 */
4608
4609 /*
4610 * internal_forkexec win32 implementation
4611 *
4612 * - starts backend using CreateProcess(), in suspended state
4613 * - writes out backend variables to the parameter file
4614 * - during this, duplicates handles and sockets required for
4615 * inheritance into the new process
4616 * - resumes execution of the new process once the backend parameter
4617 * file is complete.
4618 */
4619 static pid_t
4620 internal_forkexec(int argc, char *argv[], Port *port)
4621 {
4622 int retry_count = 0;
4623 STARTUPINFO si;
4624 PROCESS_INFORMATION pi;
4625 int i;
4626 int j;
4627 char cmdLine[MAXPGPATH * 2];
4628 HANDLE paramHandle;
4629 BackendParameters *param;
4630 SECURITY_ATTRIBUTES sa;
4631 char paramHandleStr[32];
4632 win32_deadchild_waitinfo *childinfo;
4633
4634 /* Make sure caller set up argv properly */
4635 Assert(argc >= 3);
4636 Assert(argv[argc] == NULL);
4637 Assert(strncmp(argv[1], "--fork", 6) == 0);
4638 Assert(argv[2] == NULL);
4639
4640 /* Resume here if we need to retry */
4641 retry:
4642
4643 /* Set up shared memory for parameter passing */
4644 ZeroMemory(&sa, sizeof(sa));
4645 sa.nLength = sizeof(sa);
4646 sa.bInheritHandle = TRUE;
4647 paramHandle = CreateFileMapping(INVALID_HANDLE_VALUE,
4648 &sa,
4649 PAGE_READWRITE,
4650 0,
4651 sizeof(BackendParameters),
4652 NULL);
4653 if (paramHandle == INVALID_HANDLE_VALUE)
4654 {
4655 elog(LOG, "could not create backend parameter file mapping: error code %lu",
4656 GetLastError());
4657 return -1;
4658 }
4659
4660 param = MapViewOfFile(paramHandle, FILE_MAP_WRITE, 0, 0, sizeof(BackendParameters));
4661 if (!param)
4662 {
4663 elog(LOG, "could not map backend parameter memory: error code %lu",
4664 GetLastError());
4665 CloseHandle(paramHandle);
4666 return -1;
4667 }
4668
4669 /* Insert temp file name after --fork argument */
4670 #ifdef _WIN64
4671 sprintf(paramHandleStr, "%llu", (LONG_PTR) paramHandle);
4672 #else
4673 sprintf(paramHandleStr, "%lu", (DWORD) paramHandle);
4674 #endif
4675 argv[2] = paramHandleStr;
4676
4677 /* Format the cmd line */
4678 cmdLine[sizeof(cmdLine) - 1] = '\0';
4679 cmdLine[sizeof(cmdLine) - 2] = '\0';
4680 snprintf(cmdLine, sizeof(cmdLine) - 1, "\"%s\"", postgres_exec_path);
4681 i = 0;
4682 while (argv[++i] != NULL)
4683 {
4684 j = strlen(cmdLine);
4685 snprintf(cmdLine + j, sizeof(cmdLine) - 1 - j, " \"%s\"", argv[i]);
4686 }
4687 if (cmdLine[sizeof(cmdLine) - 2] != '\0')
4688 {
4689 elog(LOG, "subprocess command line too long");
4690 return -1;
4691 }
4692
4693 memset(&pi, 0, sizeof(pi));
4694 memset(&si, 0, sizeof(si));
4695 si.cb = sizeof(si);
4696
4697 /*
4698 * Create the subprocess in a suspended state. This will be resumed later,
4699 * once we have written out the parameter file.
4700 */
4701 if (!CreateProcess(NULL, cmdLine, NULL, NULL, TRUE, CREATE_SUSPENDED,
4702 NULL, NULL, &si, &pi))
4703 {
4704 elog(LOG, "CreateProcess call failed: %m (error code %lu)",
4705 GetLastError());
4706 return -1;
4707 }
4708
4709 if (!save_backend_variables(param, port, pi.hProcess, pi.dwProcessId))
4710 {
4711 /*
4712 * log made by save_backend_variables, but we have to clean up the
4713 * mess with the half-started process
4714 */
4715 if (!TerminateProcess(pi.hProcess, 255))
4716 ereport(LOG,
4717 (errmsg_internal("could not terminate unstarted process: error code %lu",
4718 GetLastError())));
4719 CloseHandle(pi.hProcess);
4720 CloseHandle(pi.hThread);
4721 return -1; /* log made by save_backend_variables */
4722 }
4723
4724 /* Drop the parameter shared memory that is now inherited to the backend */
4725 if (!UnmapViewOfFile(param))
4726 elog(LOG, "could not unmap view of backend parameter file: error code %lu",
4727 GetLastError());
4728 if (!CloseHandle(paramHandle))
4729 elog(LOG, "could not close handle to backend parameter file: error code %lu",
4730 GetLastError());
4731
4732 /*
4733 * Reserve the memory region used by our main shared memory segment before
4734 * we resume the child process. Normally this should succeed, but if ASLR
4735 * is active then it might sometimes fail due to the stack or heap having
4736 * gotten mapped into that range. In that case, just terminate the
4737 * process and retry.
4738 */
4739 if (!pgwin32_ReserveSharedMemoryRegion(pi.hProcess))
4740 {
4741 /* pgwin32_ReserveSharedMemoryRegion already made a log entry */
4742 if (!TerminateProcess(pi.hProcess, 255))
4743 ereport(LOG,
4744 (errmsg_internal("could not terminate process that failed to reserve memory: error code %lu",
4745 GetLastError())));
4746 CloseHandle(pi.hProcess);
4747 CloseHandle(pi.hThread);
4748 if (++retry_count < 100)
4749 goto retry;
4750 ereport(LOG,
4751 (errmsg("giving up after too many tries to reserve shared memory"),
4752 errhint("This might be caused by ASLR or antivirus software.")));
4753 return -1;
4754 }
4755
4756 /*
4757 * Now that the backend variables are written out, we start the child
4758 * thread so it can start initializing while we set up the rest of the
4759 * parent state.
4760 */
4761 if (ResumeThread(pi.hThread) == -1)
4762 {
4763 if (!TerminateProcess(pi.hProcess, 255))
4764 {
4765 ereport(LOG,
4766 (errmsg_internal("could not terminate unstartable process: error code %lu",
4767 GetLastError())));
4768 CloseHandle(pi.hProcess);
4769 CloseHandle(pi.hThread);
4770 return -1;
4771 }
4772 CloseHandle(pi.hProcess);
4773 CloseHandle(pi.hThread);
4774 ereport(LOG,
4775 (errmsg_internal("could not resume thread of unstarted process: error code %lu",
4776 GetLastError())));
4777 return -1;
4778 }
4779
4780 /*
4781 * Queue a waiter for to signal when this child dies. The wait will be
4782 * handled automatically by an operating system thread pool.
4783 *
4784 * Note: use malloc instead of palloc, since it needs to be thread-safe.
4785 * Struct will be free():d from the callback function that runs on a
4786 * different thread.
4787 */
4788 childinfo = malloc(sizeof(win32_deadchild_waitinfo));
4789 if (!childinfo)
4790 ereport(FATAL,
4791 (errcode(ERRCODE_OUT_OF_MEMORY),
4792 errmsg("out of memory")));
4793
4794 childinfo->procHandle = pi.hProcess;
4795 childinfo->procId = pi.dwProcessId;
4796
4797 if (!RegisterWaitForSingleObject(&childinfo->waitHandle,
4798 pi.hProcess,
4799 pgwin32_deadchild_callback,
4800 childinfo,
4801 INFINITE,
4802 WT_EXECUTEONLYONCE | WT_EXECUTEINWAITTHREAD))
4803 ereport(FATAL,
4804 (errmsg_internal("could not register process for wait: error code %lu",
4805 GetLastError())));
4806
4807 /* Don't close pi.hProcess here - the wait thread needs access to it */
4808
4809 CloseHandle(pi.hThread);
4810
4811 return pi.dwProcessId;
4812 }
4813 #endif /* WIN32 */
4814
4815
4816 /*
4817 * SubPostmasterMain -- Get the fork/exec'd process into a state equivalent
4818 * to what it would be if we'd simply forked on Unix, and then
4819 * dispatch to the appropriate place.
4820 *
4821 * The first two command line arguments are expected to be "--forkFOO"
4822 * (where FOO indicates which postmaster child we are to become), and
4823 * the name of a variables file that we can read to load data that would
4824 * have been inherited by fork() on Unix. Remaining arguments go to the
4825 * subprocess FooMain() routine.
4826 */
4827 void
4828 SubPostmasterMain(int argc, char *argv[])
4829 {
4830 Port port;
4831
4832 /* In EXEC_BACKEND case we will not have inherited these settings */
4833 IsPostmasterEnvironment = true;
4834 whereToSendOutput = DestNone;
4835
4836 /* Setup as postmaster child */
4837 InitPostmasterChild();
4838
4839 /* Setup essential subsystems (to ensure elog() behaves sanely) */
4840 InitializeGUCOptions();
4841
4842 /* Check we got appropriate args */
4843 if (argc < 3)
4844 elog(FATAL, "invalid subpostmaster invocation");
4845
4846 /* Read in the variables file */
4847 memset(&port, 0, sizeof(Port));
4848 read_backend_variables(argv[2], &port);
4849
4850 /* Close the postmaster's sockets (as soon as we know them) */
4851 ClosePostmasterPorts(strcmp(argv[1], "--forklog") == 0);
4852
4853 /*
4854 * Set reference point for stack-depth checking
4855 */
4856 set_stack_base();
4857
4858 /*
4859 * Set up memory area for GSS information. Mirrors the code in ConnCreate
4860 * for the non-exec case.
4861 */
4862 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
4863 port.gss = (pg_gssinfo *) calloc(1, sizeof(pg_gssinfo));
4864 if (!port.gss)
4865 ereport(FATAL,
4866 (errcode(ERRCODE_OUT_OF_MEMORY),
4867 errmsg("out of memory")));
4868 #endif
4869
4870 /*
4871 * If appropriate, physically re-attach to shared memory segment. We want
4872 * to do this before going any further to ensure that we can attach at the
4873 * same address the postmaster used. On the other hand, if we choose not
4874 * to re-attach, we may have other cleanup to do.
4875 *
4876 * If testing EXEC_BACKEND on Linux, you should run this as root before
4877 * starting the postmaster:
4878 *
4879 * echo 0 >/proc/sys/kernel/randomize_va_space
4880 *
4881 * This prevents using randomized stack and code addresses that cause the
4882 * child process's memory map to be different from the parent's, making it
4883 * sometimes impossible to attach to shared memory at the desired address.
4884 * Return the setting to its old value (usually '1' or '2') when finished.
4885 */
4886 if (strcmp(argv[1], "--forkbackend") == 0 ||
4887 strcmp(argv[1], "--forkavlauncher") == 0 ||
4888 strcmp(argv[1], "--forkavworker") == 0 ||
4889 strcmp(argv[1], "--forkboot") == 0 ||
4890 strncmp(argv[1], "--forkbgworker=", 15) == 0)
4891 PGSharedMemoryReAttach();
4892 else
4893 PGSharedMemoryNoReAttach();
4894
4895 /* autovacuum needs this set before calling InitProcess */
4896 if (strcmp(argv[1], "--forkavlauncher") == 0)
4897 AutovacuumLauncherIAm();
4898 if (strcmp(argv[1], "--forkavworker") == 0)
4899 AutovacuumWorkerIAm();
4900
4901 /*
4902 * Start our win32 signal implementation. This has to be done after we
4903 * read the backend variables, because we need to pick up the signal pipe
4904 * from the parent process.
4905 */
4906 #ifdef WIN32
4907 pgwin32_signal_initialize();
4908 #endif
4909
4910 /* In EXEC_BACKEND case we will not have inherited these settings */
4911 pqinitmask();
4912 PG_SETMASK(&BlockSig);
4913
4914 /* Read in remaining GUC variables */
4915 read_nondefault_variables();
4916
4917 /*
4918 * Reload any libraries that were preloaded by the postmaster. Since we
4919 * exec'd this process, those libraries didn't come along with us; but we
4920 * should load them into all child processes to be consistent with the
4921 * non-EXEC_BACKEND behavior.
4922 */
4923 process_shared_preload_libraries();
4924
4925 /* Run backend or appropriate child */
4926 if (strcmp(argv[1], "--forkbackend") == 0)
4927 {
4928 Assert(argc == 3); /* shouldn't be any more args */
4929
4930 /*
4931 * Need to reinitialize the SSL library in the backend, since the
4932 * context structures contain function pointers and cannot be passed
4933 * through the parameter file.
4934 *
4935 * If for some reason reload fails (maybe the user installed broken
4936 * key files), soldier on without SSL; that's better than all
4937 * connections becoming impossible.
4938 *
4939 * XXX should we do this in all child processes? For the moment it's
4940 * enough to do it in backend children.
4941 */
4942 #ifdef USE_SSL
4943 if (EnableSSL)
4944 {
4945 if (secure_initialize(false) == 0)
4946 LoadedSSL = true;
4947 else
4948 ereport(LOG,
4949 (errmsg("SSL configuration could not be loaded in child process")));
4950 }
4951 #endif
4952
4953 /*
4954 * Perform additional initialization and collect startup packet.
4955 *
4956 * We want to do this before InitProcess() for a couple of reasons: 1.
4957 * so that we aren't eating up a PGPROC slot while waiting on the
4958 * client. 2. so that if InitProcess() fails due to being out of
4959 * PGPROC slots, we have already initialized libpq and are able to
4960 * report the error to the client.
4961 */
4962 BackendInitialize(&port);
4963
4964 /* Restore basic shared memory pointers */
4965 InitShmemAccess(UsedShmemSegAddr);
4966
4967 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4968 InitProcess();
4969
4970 /* Attach process to shared data structures */
4971 CreateSharedMemoryAndSemaphores(0);
4972
4973 /* And run the backend */
4974 BackendRun(&port); /* does not return */
4975 }
4976 if (strcmp(argv[1], "--forkboot") == 0)
4977 {
4978 /* Restore basic shared memory pointers */
4979 InitShmemAccess(UsedShmemSegAddr);
4980
4981 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4982 InitAuxiliaryProcess();
4983
4984 /* Attach process to shared data structures */
4985 CreateSharedMemoryAndSemaphores(0);
4986
4987 AuxiliaryProcessMain(argc - 2, argv + 2); /* does not return */
4988 }
4989 if (strcmp(argv[1], "--forkavlauncher") == 0)
4990 {
4991 /* Restore basic shared memory pointers */
4992 InitShmemAccess(UsedShmemSegAddr);
4993
4994 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4995 InitProcess();
4996
4997 /* Attach process to shared data structures */
4998 CreateSharedMemoryAndSemaphores(0);
4999
5000 AutoVacLauncherMain(argc - 2, argv + 2); /* does not return */
5001 }
5002 if (strcmp(argv[1], "--forkavworker") == 0)
5003 {
5004 /* Restore basic shared memory pointers */
5005 InitShmemAccess(UsedShmemSegAddr);
5006
5007 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5008 InitProcess();
5009
5010 /* Attach process to shared data structures */
5011 CreateSharedMemoryAndSemaphores(0);
5012
5013 AutoVacWorkerMain(argc - 2, argv + 2); /* does not return */
5014 }
5015 if (strncmp(argv[1], "--forkbgworker=", 15) == 0)
5016 {
5017 int shmem_slot;
5018
5019 /* do this as early as possible; in particular, before InitProcess() */
5020 IsBackgroundWorker = true;
5021
5022 /* Restore basic shared memory pointers */
5023 InitShmemAccess(UsedShmemSegAddr);
5024
5025 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5026 InitProcess();
5027
5028 /* Attach process to shared data structures */
5029 CreateSharedMemoryAndSemaphores(0);
5030
5031 /* Fetch MyBgworkerEntry from shared memory */
5032 shmem_slot = atoi(argv[1] + 15);
5033 MyBgworkerEntry = BackgroundWorkerEntry(shmem_slot);
5034
5035 StartBackgroundWorker();
5036 }
5037 if (strcmp(argv[1], "--forkarch") == 0)
5038 {
5039 /* Do not want to attach to shared memory */
5040
5041 PgArchiverMain(argc, argv); /* does not return */
5042 }
5043 if (strcmp(argv[1], "--forkcol") == 0)
5044 {
5045 /* Do not want to attach to shared memory */
5046
5047 PgstatCollectorMain(argc, argv); /* does not return */
5048 }
5049 if (strcmp(argv[1], "--forklog") == 0)
5050 {
5051 /* Do not want to attach to shared memory */
5052
5053 SysLoggerMain(argc, argv); /* does not return */
5054 }
5055
5056 abort(); /* shouldn't get here */
5057 }
5058 #endif /* EXEC_BACKEND */
5059
5060
5061 /*
5062 * ExitPostmaster -- cleanup
5063 *
5064 * Do NOT call exit() directly --- always go through here!
5065 */
5066 static void
5067 ExitPostmaster(int status)
5068 {
5069 #ifdef HAVE_PTHREAD_IS_THREADED_NP
5070
5071 /*
5072 * There is no known cause for a postmaster to become multithreaded after
5073 * startup. Recheck to account for the possibility of unknown causes.
5074 * This message uses LOG level, because an unclean shutdown at this point
5075 * would usually not look much different from a clean shutdown.
5076 */
5077 if (pthread_is_threaded_np() != 0)
5078 ereport(LOG,
5079 (errcode(ERRCODE_INTERNAL_ERROR),
5080 errmsg_internal("postmaster became multithreaded"),
5081 errdetail("Please report this to <pgsql-bugs@postgresql.org>.")));
5082 #endif
5083
5084 /* should cleanup shared memory and kill all backends */
5085
5086 /*
5087 * Not sure of the semantics here. When the Postmaster dies, should the
5088 * backends all be killed? probably not.
5089 *
5090 * MUST -- vadim 05-10-1999
5091 */
5092
5093 proc_exit(status);
5094 }
5095
5096 /*
5097 * sigusr1_handler - handle signal conditions from child processes
5098 */
5099 static void
5100 sigusr1_handler(SIGNAL_ARGS)
5101 {
5102 int save_errno = errno;
5103
5104 /*
5105 * We rely on the signal mechanism to have blocked all signals ... except
5106 * on Windows, which lacks sigaction(), so we have to do it manually.
5107 */
5108 #ifdef WIN32
5109 PG_SETMASK(&BlockSig);
5110 #endif
5111
5112 /*
5113 * RECOVERY_STARTED and BEGIN_HOT_STANDBY signals are ignored in
5114 * unexpected states. If the startup process quickly starts up, completes
5115 * recovery, exits, we might process the death of the startup process
5116 * first. We don't want to go back to recovery in that case.
5117 */
5118 if (CheckPostmasterSignal(PMSIGNAL_RECOVERY_STARTED) &&
5119 pmState == PM_STARTUP && Shutdown == NoShutdown)
5120 {
5121 /* WAL redo has started. We're out of reinitialization. */
5122 FatalError = false;
5123 Assert(AbortStartTime == 0);
5124
5125 /*
5126 * Crank up the background tasks. It doesn't matter if this fails,
5127 * we'll just try again later.
5128 */
5129 Assert(CheckpointerPID == 0);
5130 CheckpointerPID = StartCheckpointer();
5131 Assert(BgWriterPID == 0);
5132 BgWriterPID = StartBackgroundWriter();
5133
5134 /*
5135 * Start the archiver if we're responsible for (re-)archiving received
5136 * files.
5137 */
5138 Assert(PgArchPID == 0);
5139 if (XLogArchivingAlways())
5140 PgArchPID = pgarch_start();
5141
5142 /*
5143 * If we aren't planning to enter hot standby mode later, treat
5144 * RECOVERY_STARTED as meaning we're out of startup, and report status
5145 * accordingly.
5146 */
5147 if (!EnableHotStandby)
5148 {
5149 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STANDBY);
5150 #ifdef USE_SYSTEMD
5151 sd_notify(0, "READY=1");
5152 #endif
5153 }
5154
5155 pmState = PM_RECOVERY;
5156 }
5157
5158 if (CheckPostmasterSignal(PMSIGNAL_BEGIN_HOT_STANDBY) &&
5159 pmState == PM_RECOVERY && Shutdown == NoShutdown)
5160 {
5161 /*
5162 * Likewise, start other special children as needed.
5163 */
5164 Assert(PgStatPID == 0);
5165 PgStatPID = pgstat_start();
5166
5167 ereport(LOG,
5168 (errmsg("database system is ready to accept read only connections")));
5169
5170 /* Report status */
5171 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
5172 #ifdef USE_SYSTEMD
5173 sd_notify(0, "READY=1");
5174 #endif
5175
5176 pmState = PM_HOT_STANDBY;
5177 connsAllowed = ALLOW_ALL_CONNS;
5178
5179 /* Some workers may be scheduled to start now */
5180 StartWorkerNeeded = true;
5181 }
5182
5183 /* Process background worker state changes. */
5184 if (CheckPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE))
5185 {
5186 /* Accept new worker requests only if not stopping. */
5187 BackgroundWorkerStateChange(pmState < PM_STOP_BACKENDS);
5188 StartWorkerNeeded = true;
5189 }
5190
5191 if (StartWorkerNeeded || HaveCrashedWorker)
5192 maybe_start_bgworkers();
5193
5194 if (CheckPostmasterSignal(PMSIGNAL_WAKEN_ARCHIVER) &&
5195 PgArchPID != 0)
5196 {
5197 /*
5198 * Send SIGUSR1 to archiver process, to wake it up and begin archiving
5199 * next WAL file.
5200 */
5201 signal_child(PgArchPID, SIGUSR1);
5202 }
5203
5204 if (CheckPostmasterSignal(PMSIGNAL_ROTATE_LOGFILE) &&
5205 SysLoggerPID != 0)
5206 {
5207 /* Tell syslogger to rotate logfile */
5208 signal_child(SysLoggerPID, SIGUSR1);
5209 }
5210
5211 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER) &&
5212 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5213 {
5214 /*
5215 * Start one iteration of the autovacuum daemon, even if autovacuuming
5216 * is nominally not enabled. This is so we can have an active defense
5217 * against transaction ID wraparound. We set a flag for the main loop
5218 * to do it rather than trying to do it here --- this is because the
5219 * autovac process itself may send the signal, and we want to handle
5220 * that by launching another iteration as soon as the current one
5221 * completes.
5222 */
5223 start_autovac_launcher = true;
5224 }
5225
5226 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER) &&
5227 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5228 {
5229 /* The autovacuum launcher wants us to start a worker process. */
5230 StartAutovacuumWorker();
5231 }
5232
5233 if (CheckPostmasterSignal(PMSIGNAL_START_WALRECEIVER))
5234 {
5235 /* Startup Process wants us to start the walreceiver process. */
5236 /* Start immediately if possible, else remember request for later. */
5237 WalReceiverRequested = true;
5238 MaybeStartWalReceiver();
5239 }
5240
5241 /*
5242 * Try to advance postmaster's state machine, if a child requests it.
5243 *
5244 * Be careful about the order of this action relative to sigusr1_handler's
5245 * other actions. Generally, this should be after other actions, in case
5246 * they have effects PostmasterStateMachine would need to know about.
5247 * However, we should do it before the CheckPromoteSignal step, which
5248 * cannot have any (immediate) effect on the state machine, but does
5249 * depend on what state we're in now.
5250 */
5251 if (CheckPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE))
5252 {
5253 PostmasterStateMachine();
5254 }
5255
5256 if (StartupPID != 0 &&
5257 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5258 pmState == PM_HOT_STANDBY) &&
5259 CheckPromoteSignal())
5260 {
5261 /* Tell startup process to finish recovery */
5262 signal_child(StartupPID, SIGUSR2);
5263 }
5264
5265 #ifdef WIN32
5266 PG_SETMASK(&UnBlockSig);
5267 #endif
5268
5269 errno = save_errno;
5270 }
5271
5272 /*
5273 * SIGTERM while processing startup packet.
5274 * Clean up and exit(1).
5275 *
5276 * Running proc_exit() from a signal handler is pretty unsafe, since we
5277 * can't know what code we've interrupted. But the alternative of using
5278 * _exit(2) is also unpalatable, since it'd mean that a "fast shutdown"
5279 * would cause a database crash cycle (forcing WAL replay at restart)
5280 * if any sessions are in authentication. So we live with it for now.
5281 *
5282 * One might be tempted to try to send a message indicating why we are
5283 * disconnecting. However, that would make this even more unsafe. Also,
5284 * it seems undesirable to provide clues about the database's state to
5285 * a client that has not yet completed authentication.
5286 */
5287 static void
5288 process_startup_packet_die(SIGNAL_ARGS)
5289 {
5290 proc_exit(1);
5291 }
5292
5293 /*
5294 * SIGQUIT while processing startup packet.
5295 *
5296 * Some backend has bought the farm,
5297 * so we need to stop what we're doing and exit.
5298 */
5299 static void
5300 process_startup_packet_quickdie(SIGNAL_ARGS)
5301 {
5302 /*
5303 * We DO NOT want to run proc_exit() or atexit() callbacks; they wouldn't
5304 * be safe to run from a signal handler. Just nail the windows shut and
5305 * get out of town.
5306 *
5307 * Note we do _exit(2) not _exit(1). This is to force the postmaster into
5308 * a system reset cycle if someone sends a manual SIGQUIT to a random
5309 * backend. (While it might be safe to do _exit(1), since this session
5310 * shouldn't have touched shared memory yet, there seems little point in
5311 * taking any risks.)
5312 */
5313 _exit(2);
5314 }
5315
5316 /*
5317 * Dummy signal handler
5318 *
5319 * We use this for signals that we don't actually use in the postmaster,
5320 * but we do use in backends. If we were to SIG_IGN such signals in the
5321 * postmaster, then a newly started backend might drop a signal that arrives
5322 * before it's able to reconfigure its signal processing. (See notes in
5323 * tcop/postgres.c.)
5324 */
5325 static void
5326 dummy_handler(SIGNAL_ARGS)
5327 {
5328 }
5329
5330 /*
5331 * Timeout while processing startup packet.
5332 * As for process_startup_packet_die(), we clean up and exit(1).
5333 *
5334 * This is theoretically just as hazardous as in process_startup_packet_die(),
5335 * although in practice we're almost certainly waiting for client input,
5336 * which greatly reduces the risk.
5337 */
5338 static void
5339 StartupPacketTimeoutHandler(void)
5340 {
5341 proc_exit(1);
5342 }
5343
5344
5345 /*
5346 * Generate a random cancel key.
5347 */
5348 static bool
5349 RandomCancelKey(int32 *cancel_key)
5350 {
5351 #ifdef HAVE_STRONG_RANDOM
5352 return pg_strong_random((char *) cancel_key, sizeof(int32));
5353 #else
5354
5355 /*
5356 * If built with --disable-strong-random, use plain old erand48.
5357 *
5358 * We cannot use pg_backend_random() in postmaster, because it stores its
5359 * state in shared memory.
5360 */
5361 static unsigned short seed[3];
5362
5363 /*
5364 * Select a random seed at the time of first receiving a request.
5365 */
5366 if (random_seed == 0)
5367 {
5368 struct timeval random_stop_time;
5369
5370 gettimeofday(&random_stop_time, NULL);
5371
5372 seed[0] = (unsigned short) random_start_time.tv_usec;
5373 seed[1] = (unsigned short) (random_stop_time.tv_usec) ^ (random_start_time.tv_usec >> 16);
5374 seed[2] = (unsigned short) (random_stop_time.tv_usec >> 16);
5375
5376 random_seed = 1;
5377 }
5378
5379 *cancel_key = pg_jrand48(seed);
5380
5381 return true;
5382 #endif
5383 }
5384
5385 /*
5386 * Count up number of child processes of specified types (dead_end children
5387 * are always excluded).
5388 */
5389 static int
5390 CountChildren(int target)
5391 {
5392 dlist_iter iter;
5393 int cnt = 0;
5394
5395 dlist_foreach(iter, &BackendList)
5396 {
5397 Backend *bp = dlist_container(Backend, elem, iter.cur);
5398
5399 if (bp->dead_end)
5400 continue;
5401
5402 /*
5403 * Since target == BACKEND_TYPE_ALL is the most common case, we test
5404 * it first and avoid touching shared memory for every child.
5405 */
5406 if (target != BACKEND_TYPE_ALL)
5407 {
5408 /*
5409 * Assign bkend_type for any recently announced WAL Sender
5410 * processes.
5411 */
5412 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
5413 IsPostmasterChildWalSender(bp->child_slot))
5414 bp->bkend_type = BACKEND_TYPE_WALSND;
5415
5416 if (!(target & bp->bkend_type))
5417 continue;
5418 }
5419
5420 cnt++;
5421 }
5422 return cnt;
5423 }
5424
5425
5426 /*
5427 * StartChildProcess -- start an auxiliary process for the postmaster
5428 *
5429 * "type" determines what kind of child will be started. All child types
5430 * initially go to AuxiliaryProcessMain, which will handle common setup.
5431 *
5432 * Return value of StartChildProcess is subprocess' PID, or 0 if failed
5433 * to start subprocess.
5434 */
5435 static pid_t
5436 StartChildProcess(AuxProcType type)
5437 {
5438 pid_t pid;
5439 char *av[10];
5440 int ac = 0;
5441 char typebuf[32];
5442
5443 /*
5444 * Set up command-line arguments for subprocess
5445 */
5446 av[ac++] = "postgres";
5447
5448 #ifdef EXEC_BACKEND
5449 av[ac++] = "--forkboot";
5450 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5451 #endif
5452
5453 snprintf(typebuf, sizeof(typebuf), "-x%d", type);
5454 av[ac++] = typebuf;
5455
5456 av[ac] = NULL;
5457 Assert(ac < lengthof(av));
5458
5459 #ifdef EXEC_BACKEND
5460 pid = postmaster_forkexec(ac, av);
5461 #else /* !EXEC_BACKEND */
5462 pid = fork_process();
5463
5464 if (pid == 0) /* child */
5465 {
5466 InitPostmasterChild();
5467
5468 /* Close the postmaster's sockets */
5469 ClosePostmasterPorts(false);
5470
5471 /* Release postmaster's working memory context */
5472 MemoryContextSwitchTo(TopMemoryContext);
5473 MemoryContextDelete(PostmasterContext);
5474 PostmasterContext = NULL;
5475
5476 AuxiliaryProcessMain(ac, av);
5477 ExitPostmaster(0);
5478 }
5479 #endif /* EXEC_BACKEND */
5480
5481 if (pid < 0)
5482 {
5483 /* in parent, fork failed */
5484 int save_errno = errno;
5485
5486 errno = save_errno;
5487 switch (type)
5488 {
5489 case StartupProcess:
5490 ereport(LOG,
5491 (errmsg("could not fork startup process: %m")));
5492 break;
5493 case BgWriterProcess:
5494 ereport(LOG,
5495 (errmsg("could not fork background writer process: %m")));
5496 break;
5497 case CheckpointerProcess:
5498 ereport(LOG,
5499 (errmsg("could not fork checkpointer process: %m")));
5500 break;
5501 case WalWriterProcess:
5502 ereport(LOG,
5503 (errmsg("could not fork WAL writer process: %m")));
5504 break;
5505 case WalReceiverProcess:
5506 ereport(LOG,
5507 (errmsg("could not fork WAL receiver process: %m")));
5508 break;
5509 default:
5510 ereport(LOG,
5511 (errmsg("could not fork process: %m")));
5512 break;
5513 }
5514
5515 /*
5516 * fork failure is fatal during startup, but there's no need to choke
5517 * immediately if starting other child types fails.
5518 */
5519 if (type == StartupProcess)
5520 ExitPostmaster(1);
5521 return 0;
5522 }
5523
5524 /*
5525 * in parent, successful fork
5526 */
5527 return pid;
5528 }
5529
5530 /*
5531 * StartAutovacuumWorker
5532 * Start an autovac worker process.
5533 *
5534 * This function is here because it enters the resulting PID into the
5535 * postmaster's private backends list.
5536 *
5537 * NB -- this code very roughly matches BackendStartup.
5538 */
5539 static void
5540 StartAutovacuumWorker(void)
5541 {
5542 Backend *bn;
5543
5544 /*
5545 * If not in condition to run a process, don't try, but handle it like a
5546 * fork failure. This does not normally happen, since the signal is only
5547 * supposed to be sent by autovacuum launcher when it's OK to do it, but
5548 * we have to check to avoid race-condition problems during DB state
5549 * changes.
5550 */
5551 if (canAcceptConnections(BACKEND_TYPE_AUTOVAC) == CAC_OK)
5552 {
5553 /*
5554 * Compute the cancel key that will be assigned to this session. We
5555 * probably don't need cancel keys for autovac workers, but we'd
5556 * better have something random in the field to prevent unfriendly
5557 * people from sending cancels to them.
5558 */
5559 if (!RandomCancelKey(&MyCancelKey))
5560 {
5561 ereport(LOG,
5562 (errcode(ERRCODE_INTERNAL_ERROR),
5563 errmsg("could not generate random cancel key")));
5564 return;
5565 }
5566
5567 bn = (Backend *) malloc(sizeof(Backend));
5568 if (bn)
5569 {
5570 bn->cancel_key = MyCancelKey;
5571
5572 /* Autovac workers are not dead_end and need a child slot */
5573 bn->dead_end = false;
5574 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5575 bn->bgworker_notify = false;
5576
5577 bn->pid = StartAutoVacWorker();
5578 if (bn->pid > 0)
5579 {
5580 bn->bkend_type = BACKEND_TYPE_AUTOVAC;
5581 dlist_push_head(&BackendList, &bn->elem);
5582 #ifdef EXEC_BACKEND
5583 ShmemBackendArrayAdd(bn);
5584 #endif
5585 /* all OK */
5586 return;
5587 }
5588
5589 /*
5590 * fork failed, fall through to report -- actual error message was
5591 * logged by StartAutoVacWorker
5592 */
5593 (void) ReleasePostmasterChildSlot(bn->child_slot);
5594 free(bn);
5595 }
5596 else
5597 ereport(LOG,
5598 (errcode(ERRCODE_OUT_OF_MEMORY),
5599 errmsg("out of memory")));
5600 }
5601
5602 /*
5603 * Report the failure to the launcher, if it's running. (If it's not, we
5604 * might not even be connected to shared memory, so don't try to call
5605 * AutoVacWorkerFailed.) Note that we also need to signal it so that it
5606 * responds to the condition, but we don't do that here, instead waiting
5607 * for ServerLoop to do it. This way we avoid a ping-pong signalling in
5608 * quick succession between the autovac launcher and postmaster in case
5609 * things get ugly.
5610 */
5611 if (AutoVacPID != 0)
5612 {
5613 AutoVacWorkerFailed();
5614 avlauncher_needs_signal = true;
5615 }
5616 }
5617
5618 /*
5619 * MaybeStartWalReceiver
5620 * Start the WAL receiver process, if not running and our state allows.
5621 *
5622 * Note: if WalReceiverPID is already nonzero, it might seem that we should
5623 * clear WalReceiverRequested. However, there's a race condition if the
5624 * walreceiver terminates and the startup process immediately requests a new
5625 * one: it's quite possible to get the signal for the request before reaping
5626 * the dead walreceiver process. Better to risk launching an extra
5627 * walreceiver than to miss launching one we need. (The walreceiver code
5628 * has logic to recognize that it should go away if not needed.)
5629 */
5630 static void
5631 MaybeStartWalReceiver(void)
5632 {
5633 if (WalReceiverPID == 0 &&
5634 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5635 pmState == PM_HOT_STANDBY) &&
5636 Shutdown <= SmartShutdown)
5637 {
5638 WalReceiverPID = StartWalReceiver();
5639 if (WalReceiverPID != 0)
5640 WalReceiverRequested = false;
5641 /* else leave the flag set, so we'll try again later */
5642 }
5643 }
5644
5645
5646 /*
5647 * Create the opts file
5648 */
5649 static bool
5650 CreateOptsFile(int argc, char *argv[], char *fullprogname)
5651 {
5652 FILE *fp;
5653 int i;
5654
5655 #define OPTS_FILE "postmaster.opts"
5656
5657 if ((fp = fopen(OPTS_FILE, "w")) == NULL)
5658 {
5659 elog(LOG, "could not create file \"%s\": %m", OPTS_FILE);
5660 return false;
5661 }
5662
5663 fprintf(fp, "%s", fullprogname);
5664 for (i = 1; i < argc; i++)
5665 fprintf(fp, " \"%s\"", argv[i]);
5666 fputs("\n", fp);
5667
5668 if (fclose(fp))
5669 {
5670 elog(LOG, "could not write file \"%s\": %m", OPTS_FILE);
5671 return false;
5672 }
5673
5674 return true;
5675 }
5676
5677
5678 /*
5679 * MaxLivePostmasterChildren
5680 *
5681 * This reports the number of entries needed in per-child-process arrays
5682 * (the PMChildFlags array, and if EXEC_BACKEND the ShmemBackendArray).
5683 * These arrays include regular backends, autovac workers, walsenders
5684 * and background workers, but not special children nor dead_end children.
5685 * This allows the arrays to have a fixed maximum size, to wit the same
5686 * too-many-children limit enforced by canAcceptConnections(). The exact value
5687 * isn't too critical as long as it's more than MaxBackends.
5688 */
5689 int
5690 MaxLivePostmasterChildren(void)
5691 {
5692 return 2 * (MaxConnections + autovacuum_max_workers + 1 +
5693 max_worker_processes);
5694 }
5695
5696 /*
5697 * Connect background worker to a database.
5698 */
5699 void
5700 BackgroundWorkerInitializeConnection(char *dbname, char *username)
5701 {
5702 BackgroundWorker *worker = MyBgworkerEntry;
5703
5704 /* XXX is this the right errcode? */
5705 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5706 ereport(FATAL,
5707 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5708 errmsg("database connection requirement not indicated during registration")));
5709
5710 InitPostgres(dbname, InvalidOid, username, InvalidOid, NULL);
5711
5712 /* it had better not gotten out of "init" mode yet */
5713 if (!IsInitProcessingMode())
5714 ereport(ERROR,
5715 (errmsg("invalid processing mode in background worker")));
5716 SetProcessingMode(NormalProcessing);
5717 }
5718
5719 /*
5720 * Connect background worker to a database using OIDs.
5721 */
5722 void
5723 BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid)
5724 {
5725 BackgroundWorker *worker = MyBgworkerEntry;
5726
5727 /* XXX is this the right errcode? */
5728 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5729 ereport(FATAL,
5730 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5731 errmsg("database connection requirement not indicated during registration")));
5732
5733 InitPostgres(NULL, dboid, NULL, useroid, NULL);
5734
5735 /* it had better not gotten out of "init" mode yet */
5736 if (!IsInitProcessingMode())
5737 ereport(ERROR,
5738 (errmsg("invalid processing mode in background worker")));
5739 SetProcessingMode(NormalProcessing);
5740 }
5741
5742 /*
5743 * Block/unblock signals in a background worker
5744 */
5745 void
5746 BackgroundWorkerBlockSignals(void)
5747 {
5748 PG_SETMASK(&BlockSig);
5749 }
5750
5751 void
5752 BackgroundWorkerUnblockSignals(void)
5753 {
5754 PG_SETMASK(&UnBlockSig);
5755 }
5756
5757 #ifdef EXEC_BACKEND
5758 static pid_t
5759 bgworker_forkexec(int shmem_slot)
5760 {
5761 char *av[10];
5762 int ac = 0;
5763 char forkav[MAXPGPATH];
5764
5765 snprintf(forkav, MAXPGPATH, "--forkbgworker=%d", shmem_slot);
5766
5767 av[ac++] = "postgres";
5768 av[ac++] = forkav;
5769 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5770 av[ac] = NULL;
5771
5772 Assert(ac < lengthof(av));
5773
5774 return postmaster_forkexec(ac, av);
5775 }
5776 #endif
5777
5778 /*
5779 * Start a new bgworker.
5780 * Starting time conditions must have been checked already.
5781 *
5782 * Returns true on success, false on failure.
5783 * In either case, update the RegisteredBgWorker's state appropriately.
5784 *
5785 * This code is heavily based on autovacuum.c, q.v.
5786 */
5787 static bool
5788 do_start_bgworker(RegisteredBgWorker *rw)
5789 {
5790 pid_t worker_pid;
5791
5792 Assert(rw->rw_pid == 0);
5793
5794 /*
5795 * Allocate and assign the Backend element. Note we must do this before
5796 * forking, so that we can handle failures (out of memory or child-process
5797 * slots) cleanly.
5798 *
5799 * Treat failure as though the worker had crashed. That way, the
5800 * postmaster will wait a bit before attempting to start it again; if we
5801 * tried again right away, most likely we'd find ourselves hitting the
5802 * same resource-exhaustion condition.
5803 */
5804 if (!assign_backendlist_entry(rw))
5805 {
5806 rw->rw_crashed_at = GetCurrentTimestamp();
5807 return false;
5808 }
5809
5810 ereport(DEBUG1,
5811 (errmsg("starting background worker process \"%s\"",
5812 rw->rw_worker.bgw_name)));
5813
5814 #ifdef EXEC_BACKEND
5815 switch ((worker_pid = bgworker_forkexec(rw->rw_shmem_slot)))
5816 #else
5817 switch ((worker_pid = fork_process()))
5818 #endif
5819 {
5820 case -1:
5821 /* in postmaster, fork failed ... */
5822 ereport(LOG,
5823 (errmsg("could not fork worker process: %m")));
5824 /* undo what assign_backendlist_entry did */
5825 ReleasePostmasterChildSlot(rw->rw_child_slot);
5826 rw->rw_child_slot = 0;
5827 free(rw->rw_backend);
5828 rw->rw_backend = NULL;
5829 /* mark entry as crashed, so we'll try again later */
5830 rw->rw_crashed_at = GetCurrentTimestamp();
5831 break;
5832
5833 #ifndef EXEC_BACKEND
5834 case 0:
5835 /* in postmaster child ... */
5836 InitPostmasterChild();
5837
5838 /* Close the postmaster's sockets */
5839 ClosePostmasterPorts(false);
5840
5841 /*
5842 * Before blowing away PostmasterContext, save this bgworker's
5843 * data where it can find it.
5844 */
5845 MyBgworkerEntry = (BackgroundWorker *)
5846 MemoryContextAlloc(TopMemoryContext, sizeof(BackgroundWorker));
5847 memcpy(MyBgworkerEntry, &rw->rw_worker, sizeof(BackgroundWorker));
5848
5849 /* Release postmaster's working memory context */
5850 MemoryContextSwitchTo(TopMemoryContext);
5851 MemoryContextDelete(PostmasterContext);
5852 PostmasterContext = NULL;
5853
5854 StartBackgroundWorker();
5855
5856 exit(1); /* should not get here */
5857 break;
5858 #endif
5859 default:
5860 /* in postmaster, fork successful ... */
5861 rw->rw_pid = worker_pid;
5862 rw->rw_backend->pid = rw->rw_pid;
5863 ReportBackgroundWorkerPID(rw);
5864 /* add new worker to lists of backends */
5865 dlist_push_head(&BackendList, &rw->rw_backend->elem);
5866 #ifdef EXEC_BACKEND
5867 ShmemBackendArrayAdd(rw->rw_backend);
5868 #endif
5869 return true;
5870 }
5871
5872 return false;
5873 }
5874
5875 /*
5876 * Does the current postmaster state require starting a worker with the
5877 * specified start_time?
5878 */
5879 static bool
5880 bgworker_should_start_now(BgWorkerStartTime start_time)
5881 {
5882 switch (pmState)
5883 {
5884 case PM_NO_CHILDREN:
5885 case PM_WAIT_DEAD_END:
5886 case PM_SHUTDOWN_2:
5887 case PM_SHUTDOWN:
5888 case PM_WAIT_BACKENDS:
5889 case PM_STOP_BACKENDS:
5890 break;
5891
5892 case PM_RUN:
5893 if (start_time == BgWorkerStart_RecoveryFinished)
5894 return true;
5895 /* fall through */
5896
5897 case PM_HOT_STANDBY:
5898 if (start_time == BgWorkerStart_ConsistentState)
5899 return true;
5900 /* fall through */
5901
5902 case PM_RECOVERY:
5903 case PM_STARTUP:
5904 case PM_INIT:
5905 if (start_time == BgWorkerStart_PostmasterStart)
5906 return true;
5907 /* fall through */
5908
5909 }
5910
5911 return false;
5912 }
5913
5914 /*
5915 * Allocate the Backend struct for a connected background worker, but don't
5916 * add it to the list of backends just yet.
5917 *
5918 * On failure, return false without changing any worker state.
5919 *
5920 * Some info from the Backend is copied into the passed rw.
5921 */
5922 static bool
5923 assign_backendlist_entry(RegisteredBgWorker *rw)
5924 {
5925 Backend *bn;
5926
5927 /*
5928 * Check that database state allows another connection. Currently the
5929 * only possible failure is CAC_TOOMANY, so we just log an error message
5930 * based on that rather than checking the error code precisely.
5931 */
5932 if (canAcceptConnections(BACKEND_TYPE_BGWORKER) != CAC_OK)
5933 {
5934 ereport(LOG,
5935 (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
5936 errmsg("no slot available for new worker process")));
5937 return false;
5938 }
5939
5940 /*
5941 * Compute the cancel key that will be assigned to this session. We
5942 * probably don't need cancel keys for background workers, but we'd better
5943 * have something random in the field to prevent unfriendly people from
5944 * sending cancels to them.
5945 */
5946 if (!RandomCancelKey(&MyCancelKey))
5947 {
5948 ereport(LOG,
5949 (errcode(ERRCODE_INTERNAL_ERROR),
5950 errmsg("could not generate random cancel key")));
5951 return false;
5952 }
5953
5954 bn = malloc(sizeof(Backend));
5955 if (bn == NULL)
5956 {
5957 ereport(LOG,
5958 (errcode(ERRCODE_OUT_OF_MEMORY),
5959 errmsg("out of memory")));
5960 return false;
5961 }
5962
5963 bn->cancel_key = MyCancelKey;
5964 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5965 bn->bkend_type = BACKEND_TYPE_BGWORKER;
5966 bn->dead_end = false;
5967 bn->bgworker_notify = false;
5968
5969 rw->rw_backend = bn;
5970 rw->rw_child_slot = bn->child_slot;
5971
5972 return true;
5973 }
5974
5975 /*
5976 * If the time is right, start background worker(s).
5977 *
5978 * As a side effect, the bgworker control variables are set or reset
5979 * depending on whether more workers may need to be started.
5980 *
5981 * We limit the number of workers started per call, to avoid consuming the
5982 * postmaster's attention for too long when many such requests are pending.
5983 * As long as StartWorkerNeeded is true, ServerLoop will not block and will
5984 * call this function again after dealing with any other issues.
5985 */
5986 static void
5987 maybe_start_bgworkers(void)
5988 {
5989 #define MAX_BGWORKERS_TO_LAUNCH 100
5990 int num_launched = 0;
5991 TimestampTz now = 0;
5992 slist_mutable_iter iter;
5993
5994 /*
5995 * During crash recovery, we have no need to be called until the state
5996 * transition out of recovery.
5997 */
5998 if (FatalError)
5999 {
6000 StartWorkerNeeded = false;
6001 HaveCrashedWorker = false;
6002 return;
6003 }
6004
6005 /* Don't need to be called again unless we find a reason for it below */
6006 StartWorkerNeeded = false;
6007 HaveCrashedWorker = false;
6008
6009 slist_foreach_modify(iter, &BackgroundWorkerList)
6010 {
6011 RegisteredBgWorker *rw;
6012
6013 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
6014
6015 /* ignore if already running */
6016 if (rw->rw_pid != 0)
6017 continue;
6018
6019 /* if marked for death, clean up and remove from list */
6020 if (rw->rw_terminate)
6021 {
6022 ForgetBackgroundWorker(&iter);
6023 continue;
6024 }
6025
6026 /*
6027 * If this worker has crashed previously, maybe it needs to be
6028 * restarted (unless on registration it specified it doesn't want to
6029 * be restarted at all). Check how long ago did a crash last happen.
6030 * If the last crash is too recent, don't start it right away; let it
6031 * be restarted once enough time has passed.
6032 */
6033 if (rw->rw_crashed_at != 0)
6034 {
6035 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
6036 {
6037 int notify_pid;
6038
6039 notify_pid = rw->rw_worker.bgw_notify_pid;
6040
6041 ForgetBackgroundWorker(&iter);
6042
6043 /* Report worker is gone now. */
6044 if (notify_pid != 0)
6045 kill(notify_pid, SIGUSR1);
6046
6047 continue;
6048 }
6049
6050 /* read system time only when needed */
6051 if (now == 0)
6052 now = GetCurrentTimestamp();
6053
6054 if (!TimestampDifferenceExceeds(rw->rw_crashed_at, now,
6055 rw->rw_worker.bgw_restart_time * 1000))
6056 {
6057 /* Set flag to remember that we have workers to start later */
6058 HaveCrashedWorker = true;
6059 continue;
6060 }
6061 }
6062
6063 if (bgworker_should_start_now(rw->rw_worker.bgw_start_time))
6064 {
6065 /* reset crash time before trying to start worker */
6066 rw->rw_crashed_at = 0;
6067
6068 /*
6069 * Try to start the worker.
6070 *
6071 * On failure, give up processing workers for now, but set
6072 * StartWorkerNeeded so we'll come back here on the next iteration
6073 * of ServerLoop to try again. (We don't want to wait, because
6074 * there might be additional ready-to-run workers.) We could set
6075 * HaveCrashedWorker as well, since this worker is now marked
6076 * crashed, but there's no need because the next run of this
6077 * function will do that.
6078 */
6079 if (!do_start_bgworker(rw))
6080 {
6081 StartWorkerNeeded = true;
6082 return;
6083 }
6084
6085 /*
6086 * If we've launched as many workers as allowed, quit, but have
6087 * ServerLoop call us again to look for additional ready-to-run
6088 * workers. There might not be any, but we'll find out the next
6089 * time we run.
6090 */
6091 if (++num_launched >= MAX_BGWORKERS_TO_LAUNCH)
6092 {
6093 StartWorkerNeeded = true;
6094 return;
6095 }
6096 }
6097 }
6098 }
6099
6100 /*
6101 * When a backend asks to be notified about worker state changes, we
6102 * set a flag in its backend entry. The background worker machinery needs
6103 * to know when such backends exit.
6104 */
6105 bool
6106 PostmasterMarkPIDForWorkerNotify(int pid)
6107 {
6108 dlist_iter iter;
6109 Backend *bp;
6110
6111 dlist_foreach(iter, &BackendList)
6112 {
6113 bp = dlist_container(Backend, elem, iter.cur);
6114 if (bp->pid == pid)
6115 {
6116 bp->bgworker_notify = true;
6117 return true;
6118 }
6119 }
6120 return false;
6121 }
6122
6123 #ifdef EXEC_BACKEND
6124
6125 /*
6126 * The following need to be available to the save/restore_backend_variables
6127 * functions. They are marked NON_EXEC_STATIC in their home modules.
6128 */
6129 extern slock_t *ShmemLock;
6130 extern slock_t *ProcStructLock;
6131 extern PGPROC *AuxiliaryProcs;
6132 extern PMSignalData *PMSignalState;
6133 extern pgsocket pgStatSock;
6134 extern pg_time_t first_syslogger_file_time;
6135
6136 #ifndef WIN32
6137 #define write_inheritable_socket(dest, src, childpid) ((*(dest) = (src)), true)
6138 #define read_inheritable_socket(dest, src) (*(dest) = *(src))
6139 #else
6140 static bool write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE child);
6141 static bool write_inheritable_socket(InheritableSocket *dest, SOCKET src,
6142 pid_t childPid);
6143 static void read_inheritable_socket(SOCKET *dest, InheritableSocket *src);
6144 #endif
6145
6146
6147 /* Save critical backend variables into the BackendParameters struct */
6148 #ifndef WIN32
6149 static bool
6150 save_backend_variables(BackendParameters *param, Port *port)
6151 #else
6152 static bool
6153 save_backend_variables(BackendParameters *param, Port *port,
6154 HANDLE childProcess, pid_t childPid)
6155 #endif
6156 {
6157 memcpy(¶m->port, port, sizeof(Port));
6158 if (!write_inheritable_socket(¶m->portsocket, port->sock, childPid))
6159 return false;
6160
6161 strlcpy(param->DataDir, DataDir, MAXPGPATH);
6162
6163 memcpy(¶m->ListenSocket, &ListenSocket, sizeof(ListenSocket));
6164
6165 param->MyCancelKey = MyCancelKey;
6166 param->MyPMChildSlot = MyPMChildSlot;
6167
6168 #ifdef WIN32
6169 param->ShmemProtectiveRegion = ShmemProtectiveRegion;
6170 #endif
6171 param->UsedShmemSegID = UsedShmemSegID;
6172 param->UsedShmemSegAddr = UsedShmemSegAddr;
6173
6174 param->ShmemLock = ShmemLock;
6175 param->ShmemVariableCache = ShmemVariableCache;
6176 param->ShmemBackendArray = ShmemBackendArray;
6177
6178 #ifndef HAVE_SPINLOCKS
6179 param->SpinlockSemaArray = SpinlockSemaArray;
6180 #endif
6181 param->NamedLWLockTrancheRequests = NamedLWLockTrancheRequests;
6182 param->NamedLWLockTrancheArray = NamedLWLockTrancheArray;
6183 param->MainLWLockArray = MainLWLockArray;
6184 param->ProcStructLock = ProcStructLock;
6185 param->ProcGlobal = ProcGlobal;
6186 param->AuxiliaryProcs = AuxiliaryProcs;
6187 param->PreparedXactProcs = PreparedXactProcs;
6188 param->PMSignalState = PMSignalState;
6189 if (!write_inheritable_socket(¶m->pgStatSock, pgStatSock, childPid))
6190 return false;
6191
6192 param->PostmasterPid = PostmasterPid;
6193 param->PgStartTime = PgStartTime;
6194 param->PgReloadTime = PgReloadTime;
6195 param->first_syslogger_file_time = first_syslogger_file_time;
6196
6197 param->redirection_done = redirection_done;
6198 param->IsBinaryUpgrade = IsBinaryUpgrade;
6199 param->max_safe_fds = max_safe_fds;
6200
6201 param->MaxBackends = MaxBackends;
6202
6203 #ifdef WIN32
6204 param->PostmasterHandle = PostmasterHandle;
6205 if (!write_duplicated_handle(¶m->initial_signal_pipe,
6206 pgwin32_create_signal_listener(childPid),
6207 childProcess))
6208 return false;
6209 #else
6210 memcpy(¶m->postmaster_alive_fds, &postmaster_alive_fds,
6211 sizeof(postmaster_alive_fds));
6212 #endif
6213
6214 memcpy(¶m->syslogPipe, &syslogPipe, sizeof(syslogPipe));
6215
6216 strlcpy(param->my_exec_path, my_exec_path, MAXPGPATH);
6217
6218 strlcpy(param->pkglib_path, pkglib_path, MAXPGPATH);
6219
6220 strlcpy(param->ExtraOptions, ExtraOptions, MAXPGPATH);
6221
6222 return true;
6223 }
6224
6225
6226 #ifdef WIN32
6227 /*
6228 * Duplicate a handle for usage in a child process, and write the child
6229 * process instance of the handle to the parameter file.
6230 */
6231 static bool
6232 write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE childProcess)
6233 {
6234 HANDLE hChild = INVALID_HANDLE_VALUE;
6235
6236 if (!DuplicateHandle(GetCurrentProcess(),
6237 src,
6238 childProcess,
6239 &hChild,
6240 0,
6241 TRUE,
6242 DUPLICATE_CLOSE_SOURCE | DUPLICATE_SAME_ACCESS))
6243 {
6244 ereport(LOG,
6245 (errmsg_internal("could not duplicate handle to be written to backend parameter file: error code %lu",
6246 GetLastError())));
6247 return false;
6248 }
6249
6250 *dest = hChild;
6251 return true;
6252 }
6253
6254 /*
6255 * Duplicate a socket for usage in a child process, and write the resulting
6256 * structure to the parameter file.
6257 * This is required because a number of LSPs (Layered Service Providers) very
6258 * common on Windows (antivirus, firewalls, download managers etc) break
6259 * straight socket inheritance.
6260 */
6261 static bool
6262 write_inheritable_socket(InheritableSocket *dest, SOCKET src, pid_t childpid)
6263 {
6264 dest->origsocket = src;
6265 if (src != 0 && src != PGINVALID_SOCKET)
6266 {
6267 /* Actual socket */
6268 if (WSADuplicateSocket(src, childpid, &dest->wsainfo) != 0)
6269 {
6270 ereport(LOG,
6271 (errmsg("could not duplicate socket %d for use in backend: error code %d",
6272 (int) src, WSAGetLastError())));
6273 return false;
6274 }
6275 }
6276 return true;
6277 }
6278
6279 /*
6280 * Read a duplicate socket structure back, and get the socket descriptor.
6281 */
6282 static void
6283 read_inheritable_socket(SOCKET *dest, InheritableSocket *src)
6284 {
6285 SOCKET s;
6286
6287 if (src->origsocket == PGINVALID_SOCKET || src->origsocket == 0)
6288 {
6289 /* Not a real socket! */
6290 *dest = src->origsocket;
6291 }
6292 else
6293 {
6294 /* Actual socket, so create from structure */
6295 s = WSASocket(FROM_PROTOCOL_INFO,
6296 FROM_PROTOCOL_INFO,
6297 FROM_PROTOCOL_INFO,
6298 &src->wsainfo,
6299 0,
6300 0);
6301 if (s == INVALID_SOCKET)
6302 {
6303 write_stderr("could not create inherited socket: error code %d\n",
6304 WSAGetLastError());
6305 exit(1);
6306 }
6307 *dest = s;
6308
6309 /*
6310 * To make sure we don't get two references to the same socket, close
6311 * the original one. (This would happen when inheritance actually
6312 * works..
6313 */
6314 closesocket(src->origsocket);
6315 }
6316 }
6317 #endif
6318
6319 static void
6320 read_backend_variables(char *id, Port *port)
6321 {
6322 BackendParameters param;
6323
6324 #ifndef WIN32
6325 /* Non-win32 implementation reads from file */
6326 FILE *fp;
6327
6328 /* Open file */
6329 fp = AllocateFile(id, PG_BINARY_R);
6330 if (!fp)
6331 {
6332 write_stderr("could not open backend variables file \"%s\": %s\n",
6333 id, strerror(errno));
6334 exit(1);
6335 }
6336
6337 if (fread(¶m, sizeof(param), 1, fp) != 1)
6338 {
6339 write_stderr("could not read from backend variables file \"%s\": %s\n",
6340 id, strerror(errno));
6341 exit(1);
6342 }
6343
6344 /* Release file */
6345 FreeFile(fp);
6346 if (unlink(id) != 0)
6347 {
6348 write_stderr("could not remove file \"%s\": %s\n",
6349 id, strerror(errno));
6350 exit(1);
6351 }
6352 #else
6353 /* Win32 version uses mapped file */
6354 HANDLE paramHandle;
6355 BackendParameters *paramp;
6356
6357 #ifdef _WIN64
6358 paramHandle = (HANDLE) _atoi64(id);
6359 #else
6360 paramHandle = (HANDLE) atol(id);
6361 #endif
6362 paramp = MapViewOfFile(paramHandle, FILE_MAP_READ, 0, 0, 0);
6363 if (!paramp)
6364 {
6365 write_stderr("could not map view of backend variables: error code %lu\n",
6366 GetLastError());
6367 exit(1);
6368 }
6369
6370 memcpy(¶m, paramp, sizeof(BackendParameters));
6371
6372 if (!UnmapViewOfFile(paramp))
6373 {
6374 write_stderr("could not unmap view of backend variables: error code %lu\n",
6375 GetLastError());
6376 exit(1);
6377 }
6378
6379 if (!CloseHandle(paramHandle))
6380 {
6381 write_stderr("could not close handle to backend parameter variables: error code %lu\n",
6382 GetLastError());
6383 exit(1);
6384 }
6385 #endif
6386
6387 restore_backend_variables(¶m, port);
6388 }
6389
6390 /* Restore critical backend variables from the BackendParameters struct */
6391 static void
6392 restore_backend_variables(BackendParameters *param, Port *port)
6393 {
6394 memcpy(port, ¶m->port, sizeof(Port));
6395 read_inheritable_socket(&port->sock, ¶m->portsocket);
6396
6397 SetDataDir(param->DataDir);
6398
6399 memcpy(&ListenSocket, ¶m->ListenSocket, sizeof(ListenSocket));
6400
6401 MyCancelKey = param->MyCancelKey;
6402 MyPMChildSlot = param->MyPMChildSlot;
6403
6404 #ifdef WIN32
6405 ShmemProtectiveRegion = param->ShmemProtectiveRegion;
6406 #endif
6407 UsedShmemSegID = param->UsedShmemSegID;
6408 UsedShmemSegAddr = param->UsedShmemSegAddr;
6409
6410 ShmemLock = param->ShmemLock;
6411 ShmemVariableCache = param->ShmemVariableCache;
6412 ShmemBackendArray = param->ShmemBackendArray;
6413
6414 #ifndef HAVE_SPINLOCKS
6415 SpinlockSemaArray = param->SpinlockSemaArray;
6416 #endif
6417 NamedLWLockTrancheRequests = param->NamedLWLockTrancheRequests;
6418 NamedLWLockTrancheArray = param->NamedLWLockTrancheArray;
6419 MainLWLockArray = param->MainLWLockArray;
6420 ProcStructLock = param->ProcStructLock;
6421 ProcGlobal = param->ProcGlobal;
6422 AuxiliaryProcs = param->AuxiliaryProcs;
6423 PreparedXactProcs = param->PreparedXactProcs;
6424 PMSignalState = param->PMSignalState;
6425 read_inheritable_socket(&pgStatSock, ¶m->pgStatSock);
6426
6427 PostmasterPid = param->PostmasterPid;
6428 PgStartTime = param->PgStartTime;
6429 PgReloadTime = param->PgReloadTime;
6430 first_syslogger_file_time = param->first_syslogger_file_time;
6431
6432 redirection_done = param->redirection_done;
6433 IsBinaryUpgrade = param->IsBinaryUpgrade;
6434 max_safe_fds = param->max_safe_fds;
6435
6436 MaxBackends = param->MaxBackends;
6437
6438 #ifdef WIN32
6439 PostmasterHandle = param->PostmasterHandle;
6440 pgwin32_initial_signal_pipe = param->initial_signal_pipe;
6441 #else
6442 memcpy(&postmaster_alive_fds, ¶m->postmaster_alive_fds,
6443 sizeof(postmaster_alive_fds));
6444 #endif
6445
6446 memcpy(&syslogPipe, ¶m->syslogPipe, sizeof(syslogPipe));
6447
6448 strlcpy(my_exec_path, param->my_exec_path, MAXPGPATH);
6449
6450 strlcpy(pkglib_path, param->pkglib_path, MAXPGPATH);
6451
6452 strlcpy(ExtraOptions, param->ExtraOptions, MAXPGPATH);
6453 }
6454
6455
6456 Size
6457 ShmemBackendArraySize(void)
6458 {
6459 return mul_size(MaxLivePostmasterChildren(), sizeof(Backend));
6460 }
6461
6462 void
6463 ShmemBackendArrayAllocation(void)
6464 {
6465 Size size = ShmemBackendArraySize();
6466
6467 ShmemBackendArray = (Backend *) ShmemAlloc(size);
6468 /* Mark all slots as empty */
6469 memset(ShmemBackendArray, 0, size);
6470 }
6471
6472 static void
6473 ShmemBackendArrayAdd(Backend *bn)
6474 {
6475 /* The array slot corresponding to my PMChildSlot should be free */
6476 int i = bn->child_slot - 1;
6477
6478 Assert(ShmemBackendArray[i].pid == 0);
6479 ShmemBackendArray[i] = *bn;
6480 }
6481
6482 static void
6483 ShmemBackendArrayRemove(Backend *bn)
6484 {
6485 int i = bn->child_slot - 1;
6486
6487 Assert(ShmemBackendArray[i].pid == bn->pid);
6488 /* Mark the slot as empty */
6489 ShmemBackendArray[i].pid = 0;
6490 }
6491 #endif /* EXEC_BACKEND */
6492
6493
6494 #ifdef WIN32
6495
6496 /*
6497 * Subset implementation of waitpid() for Windows. We assume pid is -1
6498 * (that is, check all child processes) and options is WNOHANG (don't wait).
6499 */
6500 static pid_t
6501 waitpid(pid_t pid, int *exitstatus, int options)
6502 {
6503 DWORD dwd;
6504 ULONG_PTR key;
6505 OVERLAPPED *ovl;
6506
6507 /*
6508 * Check if there are any dead children. If there are, return the pid of
6509 * the first one that died.
6510 */
6511 if (GetQueuedCompletionStatus(win32ChildQueue, &dwd, &key, &ovl, 0))
6512 {
6513 *exitstatus = (int) key;
6514 return dwd;
6515 }
6516
6517 return -1;
6518 }
6519
6520 /*
6521 * Note! Code below executes on a thread pool! All operations must
6522 * be thread safe! Note that elog() and friends must *not* be used.
6523 */
6524 static void WINAPI
6525 pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired)
6526 {
6527 win32_deadchild_waitinfo *childinfo = (win32_deadchild_waitinfo *) lpParameter;
6528 DWORD exitcode;
6529
6530 if (TimerOrWaitFired)
6531 return; /* timeout. Should never happen, since we use
6532 * INFINITE as timeout value. */
6533
6534 /*
6535 * Remove handle from wait - required even though it's set to wait only
6536 * once
6537 */
6538 UnregisterWaitEx(childinfo->waitHandle, NULL);
6539
6540 if (!GetExitCodeProcess(childinfo->procHandle, &exitcode))
6541 {
6542 /*
6543 * Should never happen. Inform user and set a fixed exitcode.
6544 */
6545 write_stderr("could not read exit code for process\n");
6546 exitcode = 255;
6547 }
6548
6549 if (!PostQueuedCompletionStatus(win32ChildQueue, childinfo->procId, (ULONG_PTR) exitcode, NULL))
6550 write_stderr("could not post child completion status\n");
6551
6552 /*
6553 * Handle is per-process, so we close it here instead of in the
6554 * originating thread
6555 */
6556 CloseHandle(childinfo->procHandle);
6557
6558 /*
6559 * Free struct that was allocated before the call to
6560 * RegisterWaitForSingleObject()
6561 */
6562 free(childinfo);
6563
6564 /* Queue SIGCHLD signal */
6565 pg_queue_signal(SIGCHLD);
6566 }
6567 #endif /* WIN32 */
6568
6569 /*
6570 * Initialize one and only handle for monitoring postmaster death.
6571 *
6572 * Called once in the postmaster, so that child processes can subsequently
6573 * monitor if their parent is dead.
6574 */
6575 static void
6576 InitPostmasterDeathWatchHandle(void)
6577 {
6578 #ifndef WIN32
6579
6580 /*
6581 * Create a pipe. Postmaster holds the write end of the pipe open
6582 * (POSTMASTER_FD_OWN), and children hold the read end. Children can pass
6583 * the read file descriptor to select() to wake up in case postmaster
6584 * dies, or check for postmaster death with a (read() == 0). Children must
6585 * close the write end as soon as possible after forking, because EOF
6586 * won't be signaled in the read end until all processes have closed the
6587 * write fd. That is taken care of in ClosePostmasterPorts().
6588 */
6589 Assert(MyProcPid == PostmasterPid);
6590 if (pipe(postmaster_alive_fds) < 0)
6591 ereport(FATAL,
6592 (errcode_for_file_access(),
6593 errmsg_internal("could not create pipe to monitor postmaster death: %m")));
6594
6595 /*
6596 * Set O_NONBLOCK to allow testing for the fd's presence with a read()
6597 * call.
6598 */
6599 if (fcntl(postmaster_alive_fds[POSTMASTER_FD_WATCH], F_SETFL, O_NONBLOCK) == -1)
6600 ereport(FATAL,
6601 (errcode_for_socket_access(),
6602 errmsg_internal("could not set postmaster death monitoring pipe to nonblocking mode: %m")));
6603 #else
6604
6605 /*
6606 * On Windows, we use a process handle for the same purpose.
6607 */
6608 if (DuplicateHandle(GetCurrentProcess(),
6609 GetCurrentProcess(),
6610 GetCurrentProcess(),
6611 &PostmasterHandle,
6612 0,
6613 TRUE,
6614 DUPLICATE_SAME_ACCESS) == 0)
6615 ereport(FATAL,
6616 (errmsg_internal("could not duplicate postmaster handle: error code %lu",
6617 GetLastError())));
6618 #endif /* WIN32 */
6619 }
6620