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) 1996-2018, 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 <netdb.h>
78 #include <limits.h>
79
80 #ifdef HAVE_SYS_SELECT_H
81 #include <sys/select.h>
82 #endif
83
84 #ifdef USE_BONJOUR
85 #include <dns_sd.h>
86 #endif
87
88 #ifdef USE_SYSTEMD
89 #include <systemd/sd-daemon.h>
90 #endif
91
92 #ifdef HAVE_PTHREAD_IS_THREADED_NP
93 #include <pthread.h>
94 #endif
95
96 #include "access/transam.h"
97 #include "access/xlog.h"
98 #include "bootstrap/bootstrap.h"
99 #include "catalog/pg_control.h"
100 #include "common/file_perm.h"
101 #include "common/ip.h"
102 #include "lib/ilist.h"
103 #include "libpq/auth.h"
104 #include "libpq/libpq.h"
105 #include "libpq/pqformat.h"
106 #include "libpq/pqsignal.h"
107 #include "miscadmin.h"
108 #include "pg_getopt.h"
109 #include "pgstat.h"
110 #include "port/pg_bswap.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 MaxConnections so
205 * number of backend slots available to non-superusers is
206 * (MaxConnections - 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 checkControlFile(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 * We should not be creating any files or directories before we check the
605 * data directory (see checkDataDir()), but just in case set the umask to
606 * the most restrictive (owner-only) permissions.
607 *
608 * checkDataDir() will reset the umask based on the data directory
609 * permissions.
610 */
611 umask(PG_MODE_MASK_OWNER);
612
613 /*
614 * Initialize random(3) so we don't get the same values in every run.
615 *
616 * Note: the seed is pretty predictable from externally-visible facts such
617 * as postmaster start time, so avoid using random() for security-critical
618 * random values during postmaster startup. At the time of first
619 * connection, PostmasterRandom will select a hopefully-more-random seed.
620 */
621 srandom((unsigned int) (MyProcPid ^ MyStartTime));
622
623 /*
624 * By default, palloc() requests in the postmaster will be allocated in
625 * the PostmasterContext, which is space that can be recycled by backends.
626 * Allocated data that needs to be available to backends should be
627 * allocated in TopMemoryContext.
628 */
629 PostmasterContext = AllocSetContextCreate(TopMemoryContext,
630 "Postmaster",
631 ALLOCSET_DEFAULT_SIZES);
632 MemoryContextSwitchTo(PostmasterContext);
633
634 /* Initialize paths to installation files */
635 getInstallationPaths(argv[0]);
636
637 /*
638 * Set up signal handlers for the postmaster process.
639 *
640 * In the postmaster, we use pqsignal_pm() rather than pqsignal() (which
641 * is used by all child processes and client processes). That has a
642 * couple of special behaviors:
643 *
644 * 1. Except on Windows, we tell sigaction() to block all signals for the
645 * duration of the signal handler. This is faster than our old approach
646 * of blocking/unblocking explicitly in the signal handler, and it should
647 * also prevent excessive stack consumption if signals arrive quickly.
648 *
649 * 2. We do not set the SA_RESTART flag. This is because signals will be
650 * blocked at all times except when ServerLoop is waiting for something to
651 * happen, and during that window, we want signals to exit the select(2)
652 * wait so that ServerLoop can respond if anything interesting happened.
653 * On some platforms, signals marked SA_RESTART would not cause the
654 * select() wait to end.
655 *
656 * Child processes will generally want SA_RESTART, so pqsignal() sets that
657 * flag. We expect children to set up their own handlers before
658 * unblocking signals.
659 *
660 * CAUTION: when changing this list, check for side-effects on the signal
661 * handling setup of child processes. See tcop/postgres.c,
662 * bootstrap/bootstrap.c, postmaster/bgwriter.c, postmaster/walwriter.c,
663 * postmaster/autovacuum.c, postmaster/pgarch.c, postmaster/pgstat.c,
664 * postmaster/syslogger.c, postmaster/bgworker.c and
665 * postmaster/checkpointer.c.
666 */
667 pqinitmask();
668 PG_SETMASK(&BlockSig);
669
670 pqsignal_pm(SIGHUP, SIGHUP_handler); /* reread config file and have
671 * children do same */
672 pqsignal_pm(SIGINT, pmdie); /* send SIGTERM and shut down */
673 pqsignal_pm(SIGQUIT, pmdie); /* send SIGQUIT and die */
674 pqsignal_pm(SIGTERM, pmdie); /* wait for children and shut down */
675 pqsignal_pm(SIGALRM, SIG_IGN); /* ignored */
676 pqsignal_pm(SIGPIPE, SIG_IGN); /* ignored */
677 pqsignal_pm(SIGUSR1, sigusr1_handler); /* message from child process */
678 pqsignal_pm(SIGUSR2, dummy_handler); /* unused, reserve for children */
679 pqsignal_pm(SIGCHLD, reaper); /* handle child termination */
680 pqsignal_pm(SIGTTIN, SIG_IGN); /* ignored */
681 pqsignal_pm(SIGTTOU, SIG_IGN); /* ignored */
682 /* ignore SIGXFSZ, so that ulimit violations work like disk full */
683 #ifdef SIGXFSZ
684 pqsignal_pm(SIGXFSZ, SIG_IGN); /* ignored */
685 #endif
686
687 /*
688 * Options setup
689 */
690 InitializeGUCOptions();
691
692 opterr = 1;
693
694 /*
695 * Parse command-line options. CAUTION: keep this in sync with
696 * tcop/postgres.c (the option sets should not conflict) and with the
697 * common help() function in main/main.c.
698 */
699 while ((opt = getopt(argc, argv, "B:bc:C:D:d:EeFf:h:ijk:lN:nOo:Pp:r:S:sTt:W:-:")) != -1)
700 {
701 switch (opt)
702 {
703 case 'B':
704 SetConfigOption("shared_buffers", optarg, PGC_POSTMASTER, PGC_S_ARGV);
705 break;
706
707 case 'b':
708 /* Undocumented flag used for binary upgrades */
709 IsBinaryUpgrade = true;
710 break;
711
712 case 'C':
713 output_config_variable = strdup(optarg);
714 break;
715
716 case 'D':
717 userDoption = strdup(optarg);
718 break;
719
720 case 'd':
721 set_debug_options(atoi(optarg), PGC_POSTMASTER, PGC_S_ARGV);
722 break;
723
724 case 'E':
725 SetConfigOption("log_statement", "all", PGC_POSTMASTER, PGC_S_ARGV);
726 break;
727
728 case 'e':
729 SetConfigOption("datestyle", "euro", PGC_POSTMASTER, PGC_S_ARGV);
730 break;
731
732 case 'F':
733 SetConfigOption("fsync", "false", PGC_POSTMASTER, PGC_S_ARGV);
734 break;
735
736 case 'f':
737 if (!set_plan_disabling_options(optarg, PGC_POSTMASTER, PGC_S_ARGV))
738 {
739 write_stderr("%s: invalid argument for option -f: \"%s\"\n",
740 progname, optarg);
741 ExitPostmaster(1);
742 }
743 break;
744
745 case 'h':
746 SetConfigOption("listen_addresses", optarg, PGC_POSTMASTER, PGC_S_ARGV);
747 break;
748
749 case 'i':
750 SetConfigOption("listen_addresses", "*", PGC_POSTMASTER, PGC_S_ARGV);
751 break;
752
753 case 'j':
754 /* only used by interactive backend */
755 break;
756
757 case 'k':
758 SetConfigOption("unix_socket_directories", optarg, PGC_POSTMASTER, PGC_S_ARGV);
759 break;
760
761 case 'l':
762 SetConfigOption("ssl", "true", PGC_POSTMASTER, PGC_S_ARGV);
763 break;
764
765 case 'N':
766 SetConfigOption("max_connections", optarg, PGC_POSTMASTER, PGC_S_ARGV);
767 break;
768
769 case 'n':
770 /* Don't reinit shared mem after abnormal exit */
771 Reinit = false;
772 break;
773
774 case 'O':
775 SetConfigOption("allow_system_table_mods", "true", PGC_POSTMASTER, PGC_S_ARGV);
776 break;
777
778 case 'o':
779 /* Other options to pass to the backend on the command line */
780 snprintf(ExtraOptions + strlen(ExtraOptions),
781 sizeof(ExtraOptions) - strlen(ExtraOptions),
782 " %s", optarg);
783 break;
784
785 case 'P':
786 SetConfigOption("ignore_system_indexes", "true", PGC_POSTMASTER, PGC_S_ARGV);
787 break;
788
789 case 'p':
790 SetConfigOption("port", optarg, PGC_POSTMASTER, PGC_S_ARGV);
791 break;
792
793 case 'r':
794 /* only used by single-user backend */
795 break;
796
797 case 'S':
798 SetConfigOption("work_mem", optarg, PGC_POSTMASTER, PGC_S_ARGV);
799 break;
800
801 case 's':
802 SetConfigOption("log_statement_stats", "true", PGC_POSTMASTER, PGC_S_ARGV);
803 break;
804
805 case 'T':
806
807 /*
808 * In the event that some backend dumps core, send SIGSTOP,
809 * rather than SIGQUIT, to all its peers. This lets the wily
810 * post_hacker collect core dumps from everyone.
811 */
812 SendStop = true;
813 break;
814
815 case 't':
816 {
817 const char *tmp = get_stats_option_name(optarg);
818
819 if (tmp)
820 {
821 SetConfigOption(tmp, "true", PGC_POSTMASTER, PGC_S_ARGV);
822 }
823 else
824 {
825 write_stderr("%s: invalid argument for option -t: \"%s\"\n",
826 progname, optarg);
827 ExitPostmaster(1);
828 }
829 break;
830 }
831
832 case 'W':
833 SetConfigOption("post_auth_delay", optarg, PGC_POSTMASTER, PGC_S_ARGV);
834 break;
835
836 case 'c':
837 case '-':
838 {
839 char *name,
840 *value;
841
842 ParseLongOption(optarg, &name, &value);
843 if (!value)
844 {
845 if (opt == '-')
846 ereport(ERROR,
847 (errcode(ERRCODE_SYNTAX_ERROR),
848 errmsg("--%s requires a value",
849 optarg)));
850 else
851 ereport(ERROR,
852 (errcode(ERRCODE_SYNTAX_ERROR),
853 errmsg("-c %s requires a value",
854 optarg)));
855 }
856
857 SetConfigOption(name, value, PGC_POSTMASTER, PGC_S_ARGV);
858 free(name);
859 if (value)
860 free(value);
861 break;
862 }
863
864 default:
865 write_stderr("Try \"%s --help\" for more information.\n",
866 progname);
867 ExitPostmaster(1);
868 }
869 }
870
871 /*
872 * Postmaster accepts no non-option switch arguments.
873 */
874 if (optind < argc)
875 {
876 write_stderr("%s: invalid argument: \"%s\"\n",
877 progname, argv[optind]);
878 write_stderr("Try \"%s --help\" for more information.\n",
879 progname);
880 ExitPostmaster(1);
881 }
882
883 /*
884 * Locate the proper configuration files and data directory, and read
885 * postgresql.conf for the first time.
886 */
887 if (!SelectConfigFiles(userDoption, progname))
888 ExitPostmaster(2);
889
890 if (output_config_variable != NULL)
891 {
892 /*
893 * "-C guc" was specified, so print GUC's value and exit. No extra
894 * permission check is needed because the user is reading inside the
895 * data dir.
896 */
897 const char *config_val = GetConfigOption(output_config_variable,
898 false, false);
899
900 puts(config_val ? config_val : "");
901 ExitPostmaster(0);
902 }
903
904 /* Verify that DataDir looks reasonable */
905 checkDataDir();
906
907 /* Check that pg_control exists */
908 checkControlFile();
909
910 /* And switch working directory into it */
911 ChangeToDataDir();
912
913 /*
914 * Check for invalid combinations of GUC settings.
915 */
916 if (ReservedBackends + max_wal_senders >= MaxConnections)
917 {
918 write_stderr("%s: superuser_reserved_connections (%d) plus max_wal_senders (%d) must be less than max_connections (%d)\n",
919 progname,
920 ReservedBackends, max_wal_senders, MaxConnections);
921 ExitPostmaster(1);
922 }
923 if (XLogArchiveMode > ARCHIVE_MODE_OFF && wal_level == WAL_LEVEL_MINIMAL)
924 ereport(ERROR,
925 (errmsg("WAL archival cannot be enabled when wal_level is \"minimal\"")));
926 if (max_wal_senders > 0 && wal_level == WAL_LEVEL_MINIMAL)
927 ereport(ERROR,
928 (errmsg("WAL streaming (max_wal_senders > 0) requires wal_level \"replica\" or \"logical\"")));
929
930 /*
931 * Other one-time internal sanity checks can go here, if they are fast.
932 * (Put any slow processing further down, after postmaster.pid creation.)
933 */
934 if (!CheckDateTokenTables())
935 {
936 write_stderr("%s: invalid datetoken tables, please fix\n", progname);
937 ExitPostmaster(1);
938 }
939
940 /*
941 * Now that we are done processing the postmaster arguments, reset
942 * getopt(3) library so that it will work correctly in subprocesses.
943 */
944 optind = 1;
945 #ifdef HAVE_INT_OPTRESET
946 optreset = 1; /* some systems need this too */
947 #endif
948
949 /* For debugging: display postmaster environment */
950 {
951 extern char **environ;
952 char **p;
953
954 ereport(DEBUG3,
955 (errmsg_internal("%s: PostmasterMain: initial environment dump:",
956 progname)));
957 ereport(DEBUG3,
958 (errmsg_internal("-----------------------------------------")));
959 for (p = environ; *p; ++p)
960 ereport(DEBUG3,
961 (errmsg_internal("\t%s", *p)));
962 ereport(DEBUG3,
963 (errmsg_internal("-----------------------------------------")));
964 }
965
966 /*
967 * Create lockfile for data directory.
968 *
969 * We want to do this before we try to grab the input sockets, because the
970 * data directory interlock is more reliable than the socket-file
971 * interlock (thanks to whoever decided to put socket files in /tmp :-().
972 * For the same reason, it's best to grab the TCP socket(s) before the
973 * Unix socket(s).
974 *
975 * Also note that this internally sets up the on_proc_exit function that
976 * is responsible for removing both data directory and socket lockfiles;
977 * so it must happen before opening sockets so that at exit, the socket
978 * lockfiles go away after CloseServerPorts runs.
979 */
980 CreateDataDirLockFile(true);
981
982 /*
983 * Read the control file (for error checking and config info).
984 *
985 * Since we verify the control file's CRC, this has a useful side effect
986 * on machines where we need a run-time test for CRC support instructions.
987 * The postmaster will do the test once at startup, and then its child
988 * processes will inherit the correct function pointer and not need to
989 * repeat the test.
990 */
991 LocalProcessControlFile(false);
992
993 /*
994 * Initialize SSL library, if specified.
995 */
996 #ifdef USE_SSL
997 if (EnableSSL)
998 {
999 (void) secure_initialize(true);
1000 LoadedSSL = true;
1001 }
1002 #endif
1003
1004 /*
1005 * Register the apply launcher. Since it registers a background worker,
1006 * it needs to be called before InitializeMaxBackends(), and it's probably
1007 * a good idea to call it before any modules had chance to take the
1008 * background worker slots.
1009 */
1010 ApplyLauncherRegister();
1011
1012 /*
1013 * process any libraries that should be preloaded at postmaster start
1014 */
1015 process_shared_preload_libraries();
1016
1017 /*
1018 * Now that loadable modules have had their chance to register background
1019 * workers, calculate MaxBackends.
1020 */
1021 InitializeMaxBackends();
1022
1023 /*
1024 * Establish input sockets.
1025 *
1026 * First, mark them all closed, and set up an on_proc_exit function that's
1027 * charged with closing the sockets again at postmaster shutdown.
1028 */
1029 for (i = 0; i < MAXLISTEN; i++)
1030 ListenSocket[i] = PGINVALID_SOCKET;
1031
1032 on_proc_exit(CloseServerPorts, 0);
1033
1034 if (ListenAddresses)
1035 {
1036 char *rawstring;
1037 List *elemlist;
1038 ListCell *l;
1039 int success = 0;
1040
1041 /* Need a modifiable copy of ListenAddresses */
1042 rawstring = pstrdup(ListenAddresses);
1043
1044 /* Parse string into list of hostnames */
1045 if (!SplitIdentifierString(rawstring, ',', &elemlist))
1046 {
1047 /* syntax error in list */
1048 ereport(FATAL,
1049 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1050 errmsg("invalid list syntax in parameter \"%s\"",
1051 "listen_addresses")));
1052 }
1053
1054 foreach(l, elemlist)
1055 {
1056 char *curhost = (char *) lfirst(l);
1057
1058 if (strcmp(curhost, "*") == 0)
1059 status = StreamServerPort(AF_UNSPEC, NULL,
1060 (unsigned short) PostPortNumber,
1061 NULL,
1062 ListenSocket, MAXLISTEN);
1063 else
1064 status = StreamServerPort(AF_UNSPEC, curhost,
1065 (unsigned short) PostPortNumber,
1066 NULL,
1067 ListenSocket, MAXLISTEN);
1068
1069 if (status == STATUS_OK)
1070 {
1071 success++;
1072 /* record the first successful host addr in lockfile */
1073 if (!listen_addr_saved)
1074 {
1075 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, curhost);
1076 listen_addr_saved = true;
1077 }
1078 }
1079 else
1080 ereport(WARNING,
1081 (errmsg("could not create listen socket for \"%s\"",
1082 curhost)));
1083 }
1084
1085 if (!success && elemlist != NIL)
1086 ereport(FATAL,
1087 (errmsg("could not create any TCP/IP sockets")));
1088
1089 list_free(elemlist);
1090 pfree(rawstring);
1091 }
1092
1093 #ifdef USE_BONJOUR
1094 /* Register for Bonjour only if we opened TCP socket(s) */
1095 if (enable_bonjour && ListenSocket[0] != PGINVALID_SOCKET)
1096 {
1097 DNSServiceErrorType err;
1098
1099 /*
1100 * We pass 0 for interface_index, which will result in registering on
1101 * all "applicable" interfaces. It's not entirely clear from the
1102 * DNS-SD docs whether this would be appropriate if we have bound to
1103 * just a subset of the available network interfaces.
1104 */
1105 err = DNSServiceRegister(&bonjour_sdref,
1106 0,
1107 0,
1108 bonjour_name,
1109 "_postgresql._tcp.",
1110 NULL,
1111 NULL,
1112 pg_hton16(PostPortNumber),
1113 0,
1114 NULL,
1115 NULL,
1116 NULL);
1117 if (err != kDNSServiceErr_NoError)
1118 elog(LOG, "DNSServiceRegister() failed: error code %ld",
1119 (long) err);
1120
1121 /*
1122 * We don't bother to read the mDNS daemon's reply, and we expect that
1123 * it will automatically terminate our registration when the socket is
1124 * closed at postmaster termination. So there's nothing more to be
1125 * done here. However, the bonjour_sdref is kept around so that
1126 * forked children can close their copies of the socket.
1127 */
1128 }
1129 #endif
1130
1131 #ifdef HAVE_UNIX_SOCKETS
1132 if (Unix_socket_directories)
1133 {
1134 char *rawstring;
1135 List *elemlist;
1136 ListCell *l;
1137 int success = 0;
1138
1139 /* Need a modifiable copy of Unix_socket_directories */
1140 rawstring = pstrdup(Unix_socket_directories);
1141
1142 /* Parse string into list of directories */
1143 if (!SplitDirectoriesString(rawstring, ',', &elemlist))
1144 {
1145 /* syntax error in list */
1146 ereport(FATAL,
1147 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1148 errmsg("invalid list syntax in parameter \"%s\"",
1149 "unix_socket_directories")));
1150 }
1151
1152 foreach(l, elemlist)
1153 {
1154 char *socketdir = (char *) lfirst(l);
1155
1156 status = StreamServerPort(AF_UNIX, NULL,
1157 (unsigned short) PostPortNumber,
1158 socketdir,
1159 ListenSocket, MAXLISTEN);
1160
1161 if (status == STATUS_OK)
1162 {
1163 success++;
1164 /* record the first successful Unix socket in lockfile */
1165 if (success == 1)
1166 AddToDataDirLockFile(LOCK_FILE_LINE_SOCKET_DIR, socketdir);
1167 }
1168 else
1169 ereport(WARNING,
1170 (errmsg("could not create Unix-domain socket in directory \"%s\"",
1171 socketdir)));
1172 }
1173
1174 if (!success && elemlist != NIL)
1175 ereport(FATAL,
1176 (errmsg("could not create any Unix-domain sockets")));
1177
1178 list_free_deep(elemlist);
1179 pfree(rawstring);
1180 }
1181 #endif
1182
1183 /*
1184 * check that we have some socket to listen on
1185 */
1186 if (ListenSocket[0] == PGINVALID_SOCKET)
1187 ereport(FATAL,
1188 (errmsg("no socket created for listening")));
1189
1190 /*
1191 * If no valid TCP ports, write an empty line for listen address,
1192 * indicating the Unix socket must be used. Note that this line is not
1193 * added to the lock file until there is a socket backing it.
1194 */
1195 if (!listen_addr_saved)
1196 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, "");
1197
1198 /*
1199 * Set up shared memory and semaphores.
1200 */
1201 reset_shared(PostPortNumber);
1202
1203 /*
1204 * Estimate number of openable files. This must happen after setting up
1205 * semaphores, because on some platforms semaphores count as open files.
1206 */
1207 set_max_safe_fds();
1208
1209 /*
1210 * Set reference point for stack-depth checking.
1211 */
1212 set_stack_base();
1213
1214 /*
1215 * Initialize pipe (or process handle on Windows) that allows children to
1216 * wake up from sleep on postmaster death.
1217 */
1218 InitPostmasterDeathWatchHandle();
1219
1220 #ifdef WIN32
1221
1222 /*
1223 * Initialize I/O completion port used to deliver list of dead children.
1224 */
1225 win32ChildQueue = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 1);
1226 if (win32ChildQueue == NULL)
1227 ereport(FATAL,
1228 (errmsg("could not create I/O completion port for child queue")));
1229 #endif
1230
1231 /*
1232 * Record postmaster options. We delay this till now to avoid recording
1233 * bogus options (eg, NBuffers too high for available memory).
1234 */
1235 if (!CreateOptsFile(argc, argv, my_exec_path))
1236 ExitPostmaster(1);
1237
1238 #ifdef EXEC_BACKEND
1239 /* Write out nondefault GUC settings for child processes to use */
1240 write_nondefault_variables(PGC_POSTMASTER);
1241 #endif
1242
1243 /*
1244 * Write the external PID file if requested
1245 */
1246 if (external_pid_file)
1247 {
1248 FILE *fpidfile = fopen(external_pid_file, "w");
1249
1250 if (fpidfile)
1251 {
1252 fprintf(fpidfile, "%d\n", MyProcPid);
1253 fclose(fpidfile);
1254
1255 /* Make PID file world readable */
1256 if (chmod(external_pid_file, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH) != 0)
1257 write_stderr("%s: could not change permissions of external PID file \"%s\": %s\n",
1258 progname, external_pid_file, strerror(errno));
1259 }
1260 else
1261 write_stderr("%s: could not write external PID file \"%s\": %s\n",
1262 progname, external_pid_file, strerror(errno));
1263
1264 on_proc_exit(unlink_external_pid_file, 0);
1265 }
1266
1267 /*
1268 * Remove old temporary files. At this point there can be no other
1269 * Postgres processes running in this directory, so this should be safe.
1270 */
1271 RemovePgTempFiles();
1272
1273 /*
1274 * Forcibly remove the files signaling a standby promotion request.
1275 * Otherwise, the existence of those files triggers a promotion too early,
1276 * whether a user wants that or not.
1277 *
1278 * This removal of files is usually unnecessary because they can exist
1279 * only during a few moments during a standby promotion. However there is
1280 * a race condition: if pg_ctl promote is executed and creates the files
1281 * during a promotion, the files can stay around even after the server is
1282 * brought up to new master. Then, if new standby starts by using the
1283 * backup taken from that master, the files can exist at the server
1284 * startup and should be removed in order to avoid an unexpected
1285 * promotion.
1286 *
1287 * Note that promotion signal files need to be removed before the startup
1288 * process is invoked. Because, after that, they can be used by
1289 * postmaster's SIGUSR1 signal handler.
1290 */
1291 RemovePromoteSignalFiles();
1292
1293 /* Remove any outdated file holding the current log filenames. */
1294 if (unlink(LOG_METAINFO_DATAFILE) < 0 && errno != ENOENT)
1295 ereport(LOG,
1296 (errcode_for_file_access(),
1297 errmsg("could not remove file \"%s\": %m",
1298 LOG_METAINFO_DATAFILE)));
1299
1300 /*
1301 * If enabled, start up syslogger collection subprocess
1302 */
1303 SysLoggerPID = SysLogger_Start();
1304
1305 /*
1306 * Reset whereToSendOutput from DestDebug (its starting state) to
1307 * DestNone. This stops ereport from sending log messages to stderr unless
1308 * Log_destination permits. We don't do this until the postmaster is
1309 * fully launched, since startup failures may as well be reported to
1310 * stderr.
1311 *
1312 * If we are in fact disabling logging to stderr, first emit a log message
1313 * saying so, to provide a breadcrumb trail for users who may not remember
1314 * that their logging is configured to go somewhere else.
1315 */
1316 if (!(Log_destination & LOG_DESTINATION_STDERR))
1317 ereport(LOG,
1318 (errmsg("ending log output to stderr"),
1319 errhint("Future log output will go to log destination \"%s\".",
1320 Log_destination_string)));
1321
1322 whereToSendOutput = DestNone;
1323
1324 /*
1325 * Initialize stats collection subsystem (this does NOT start the
1326 * collector process!)
1327 */
1328 pgstat_init();
1329
1330 /*
1331 * Initialize the autovacuum subsystem (again, no process start yet)
1332 */
1333 autovac_init();
1334
1335 /*
1336 * Load configuration files for client authentication.
1337 */
1338 if (!load_hba())
1339 {
1340 /*
1341 * It makes no sense to continue if we fail to load the HBA file,
1342 * since there is no way to connect to the database in this case.
1343 */
1344 ereport(FATAL,
1345 (errmsg("could not load pg_hba.conf")));
1346 }
1347 if (!load_ident())
1348 {
1349 /*
1350 * We can start up without the IDENT file, although it means that you
1351 * cannot log in using any of the authentication methods that need a
1352 * user name mapping. load_ident() already logged the details of error
1353 * to the log.
1354 */
1355 }
1356
1357 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1358
1359 /*
1360 * On macOS, libintl replaces setlocale() with a version that calls
1361 * CFLocaleCopyCurrent() when its second argument is "" and every relevant
1362 * environment variable is unset or empty. CFLocaleCopyCurrent() makes
1363 * the process multithreaded. The postmaster calls sigprocmask() and
1364 * calls fork() without an immediate exec(), both of which have undefined
1365 * behavior in a multithreaded program. A multithreaded postmaster is the
1366 * normal case on Windows, which offers neither fork() nor sigprocmask().
1367 */
1368 if (pthread_is_threaded_np() != 0)
1369 ereport(FATAL,
1370 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1371 errmsg("postmaster became multithreaded during startup"),
1372 errhint("Set the LC_ALL environment variable to a valid locale.")));
1373 #endif
1374
1375 /*
1376 * Remember postmaster startup time
1377 */
1378 PgStartTime = GetCurrentTimestamp();
1379 #ifndef HAVE_STRONG_RANDOM
1380 /* RandomCancelKey wants its own copy */
1381 gettimeofday(&random_start_time, NULL);
1382 #endif
1383
1384 /*
1385 * Report postmaster status in the postmaster.pid file, to allow pg_ctl to
1386 * see what's happening.
1387 */
1388 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STARTING);
1389
1390 /*
1391 * We're ready to rock and roll...
1392 */
1393 StartupPID = StartupDataBase();
1394 Assert(StartupPID != 0);
1395 StartupStatus = STARTUP_RUNNING;
1396 pmState = PM_STARTUP;
1397
1398 /* Some workers may be scheduled to start now */
1399 maybe_start_bgworkers();
1400
1401 status = ServerLoop();
1402
1403 /*
1404 * ServerLoop probably shouldn't ever return, but if it does, close down.
1405 */
1406 ExitPostmaster(status != STATUS_OK);
1407
1408 abort(); /* not reached */
1409 }
1410
1411
1412 /*
1413 * on_proc_exit callback to close server's listen sockets
1414 */
1415 static void
CloseServerPorts(int status,Datum arg)1416 CloseServerPorts(int status, Datum arg)
1417 {
1418 int i;
1419
1420 /*
1421 * First, explicitly close all the socket FDs. We used to just let this
1422 * happen implicitly at postmaster exit, but it's better to close them
1423 * before we remove the postmaster.pid lockfile; otherwise there's a race
1424 * condition if a new postmaster wants to re-use the TCP port number.
1425 */
1426 for (i = 0; i < MAXLISTEN; i++)
1427 {
1428 if (ListenSocket[i] != PGINVALID_SOCKET)
1429 {
1430 StreamClose(ListenSocket[i]);
1431 ListenSocket[i] = PGINVALID_SOCKET;
1432 }
1433 }
1434
1435 /*
1436 * Next, remove any filesystem entries for Unix sockets. To avoid race
1437 * conditions against incoming postmasters, this must happen after closing
1438 * the sockets and before removing lock files.
1439 */
1440 RemoveSocketFiles();
1441
1442 /*
1443 * We don't do anything about socket lock files here; those will be
1444 * removed in a later on_proc_exit callback.
1445 */
1446 }
1447
1448 /*
1449 * on_proc_exit callback to delete external_pid_file
1450 */
1451 static void
unlink_external_pid_file(int status,Datum arg)1452 unlink_external_pid_file(int status, Datum arg)
1453 {
1454 if (external_pid_file)
1455 unlink(external_pid_file);
1456 }
1457
1458
1459 /*
1460 * Compute and check the directory paths to files that are part of the
1461 * installation (as deduced from the postgres executable's own location)
1462 */
1463 static void
getInstallationPaths(const char * argv0)1464 getInstallationPaths(const char *argv0)
1465 {
1466 DIR *pdir;
1467
1468 /* Locate the postgres executable itself */
1469 if (find_my_exec(argv0, my_exec_path) < 0)
1470 elog(FATAL, "%s: could not locate my own executable path", argv0);
1471
1472 #ifdef EXEC_BACKEND
1473 /* Locate executable backend before we change working directory */
1474 if (find_other_exec(argv0, "postgres", PG_BACKEND_VERSIONSTR,
1475 postgres_exec_path) < 0)
1476 ereport(FATAL,
1477 (errmsg("%s: could not locate matching postgres executable",
1478 argv0)));
1479 #endif
1480
1481 /*
1482 * Locate the pkglib directory --- this has to be set early in case we try
1483 * to load any modules from it in response to postgresql.conf entries.
1484 */
1485 get_pkglib_path(my_exec_path, pkglib_path);
1486
1487 /*
1488 * Verify that there's a readable directory there; otherwise the Postgres
1489 * installation is incomplete or corrupt. (A typical cause of this
1490 * failure is that the postgres executable has been moved or hardlinked to
1491 * some directory that's not a sibling of the installation lib/
1492 * directory.)
1493 */
1494 pdir = AllocateDir(pkglib_path);
1495 if (pdir == NULL)
1496 ereport(ERROR,
1497 (errcode_for_file_access(),
1498 errmsg("could not open directory \"%s\": %m",
1499 pkglib_path),
1500 errhint("This may indicate an incomplete PostgreSQL installation, or that the file \"%s\" has been moved away from its proper location.",
1501 my_exec_path)));
1502 FreeDir(pdir);
1503
1504 /*
1505 * XXX is it worth similarly checking the share/ directory? If the lib/
1506 * directory is there, then share/ probably is too.
1507 */
1508 }
1509
1510 /*
1511 * Check that pg_control exists in the correct location in the data directory.
1512 *
1513 * No attempt is made to validate the contents of pg_control here. This is
1514 * just a sanity check to see if we are looking at a real data directory.
1515 */
1516 static void
checkControlFile(void)1517 checkControlFile(void)
1518 {
1519 char path[MAXPGPATH];
1520 FILE *fp;
1521
1522 snprintf(path, sizeof(path), "%s/global/pg_control", DataDir);
1523
1524 fp = AllocateFile(path, PG_BINARY_R);
1525 if (fp == NULL)
1526 {
1527 write_stderr("%s: could not find the database system\n"
1528 "Expected to find it in the directory \"%s\",\n"
1529 "but could not open file \"%s\": %s\n",
1530 progname, DataDir, path, strerror(errno));
1531 ExitPostmaster(2);
1532 }
1533 FreeFile(fp);
1534 }
1535
1536 /*
1537 * Determine how long should we let ServerLoop sleep.
1538 *
1539 * In normal conditions we wait at most one minute, to ensure that the other
1540 * background tasks handled by ServerLoop get done even when no requests are
1541 * arriving. However, if there are background workers waiting to be started,
1542 * we don't actually sleep so that they are quickly serviced. Other exception
1543 * cases are as shown in the code.
1544 */
1545 static void
DetermineSleepTime(struct timeval * timeout)1546 DetermineSleepTime(struct timeval *timeout)
1547 {
1548 TimestampTz next_wakeup = 0;
1549
1550 /*
1551 * Normal case: either there are no background workers at all, or we're in
1552 * a shutdown sequence (during which we ignore bgworkers altogether).
1553 */
1554 if (Shutdown > NoShutdown ||
1555 (!StartWorkerNeeded && !HaveCrashedWorker))
1556 {
1557 if (AbortStartTime != 0)
1558 {
1559 /* time left to abort; clamp to 0 in case it already expired */
1560 timeout->tv_sec = SIGKILL_CHILDREN_AFTER_SECS -
1561 (time(NULL) - AbortStartTime);
1562 timeout->tv_sec = Max(timeout->tv_sec, 0);
1563 timeout->tv_usec = 0;
1564 }
1565 else
1566 {
1567 timeout->tv_sec = 60;
1568 timeout->tv_usec = 0;
1569 }
1570 return;
1571 }
1572
1573 if (StartWorkerNeeded)
1574 {
1575 timeout->tv_sec = 0;
1576 timeout->tv_usec = 0;
1577 return;
1578 }
1579
1580 if (HaveCrashedWorker)
1581 {
1582 slist_mutable_iter siter;
1583
1584 /*
1585 * When there are crashed bgworkers, we sleep just long enough that
1586 * they are restarted when they request to be. Scan the list to
1587 * determine the minimum of all wakeup times according to most recent
1588 * crash time and requested restart interval.
1589 */
1590 slist_foreach_modify(siter, &BackgroundWorkerList)
1591 {
1592 RegisteredBgWorker *rw;
1593 TimestampTz this_wakeup;
1594
1595 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
1596
1597 if (rw->rw_crashed_at == 0)
1598 continue;
1599
1600 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART
1601 || rw->rw_terminate)
1602 {
1603 ForgetBackgroundWorker(&siter);
1604 continue;
1605 }
1606
1607 this_wakeup = TimestampTzPlusMilliseconds(rw->rw_crashed_at,
1608 1000L * rw->rw_worker.bgw_restart_time);
1609 if (next_wakeup == 0 || this_wakeup < next_wakeup)
1610 next_wakeup = this_wakeup;
1611 }
1612 }
1613
1614 if (next_wakeup != 0)
1615 {
1616 long secs;
1617 int microsecs;
1618
1619 TimestampDifference(GetCurrentTimestamp(), next_wakeup,
1620 &secs, µsecs);
1621 timeout->tv_sec = secs;
1622 timeout->tv_usec = microsecs;
1623
1624 /* Ensure we don't exceed one minute */
1625 if (timeout->tv_sec > 60)
1626 {
1627 timeout->tv_sec = 60;
1628 timeout->tv_usec = 0;
1629 }
1630 }
1631 else
1632 {
1633 timeout->tv_sec = 60;
1634 timeout->tv_usec = 0;
1635 }
1636 }
1637
1638 /*
1639 * Main idle loop of postmaster
1640 *
1641 * NB: Needs to be called with signals blocked
1642 */
1643 static int
ServerLoop(void)1644 ServerLoop(void)
1645 {
1646 fd_set readmask;
1647 int nSockets;
1648 time_t last_lockfile_recheck_time,
1649 last_touch_time;
1650
1651 last_lockfile_recheck_time = last_touch_time = time(NULL);
1652
1653 nSockets = initMasks(&readmask);
1654
1655 for (;;)
1656 {
1657 fd_set rmask;
1658 int selres;
1659 time_t now;
1660
1661 /*
1662 * Wait for a connection request to arrive.
1663 *
1664 * We block all signals except while sleeping. That makes it safe for
1665 * signal handlers, which again block all signals while executing, to
1666 * do nontrivial work.
1667 *
1668 * If we are in PM_WAIT_DEAD_END state, then we don't want to accept
1669 * any new connections, so we don't call select(), and just sleep.
1670 */
1671 memcpy((char *) &rmask, (char *) &readmask, sizeof(fd_set));
1672
1673 if (pmState == PM_WAIT_DEAD_END)
1674 {
1675 PG_SETMASK(&UnBlockSig);
1676
1677 pg_usleep(100000L); /* 100 msec seems reasonable */
1678 selres = 0;
1679
1680 PG_SETMASK(&BlockSig);
1681 }
1682 else
1683 {
1684 /* must set timeout each time; some OSes change it! */
1685 struct timeval timeout;
1686
1687 /* Needs to run with blocked signals! */
1688 DetermineSleepTime(&timeout);
1689
1690 PG_SETMASK(&UnBlockSig);
1691
1692 selres = select(nSockets, &rmask, NULL, NULL, &timeout);
1693
1694 PG_SETMASK(&BlockSig);
1695 }
1696
1697 /* Now check the select() result */
1698 if (selres < 0)
1699 {
1700 if (errno != EINTR && errno != EWOULDBLOCK)
1701 {
1702 ereport(LOG,
1703 (errcode_for_socket_access(),
1704 errmsg("select() failed in postmaster: %m")));
1705 return STATUS_ERROR;
1706 }
1707 }
1708
1709 /*
1710 * New connection pending on any of our sockets? If so, fork a child
1711 * process to deal with it.
1712 */
1713 if (selres > 0)
1714 {
1715 int i;
1716
1717 for (i = 0; i < MAXLISTEN; i++)
1718 {
1719 if (ListenSocket[i] == PGINVALID_SOCKET)
1720 break;
1721 if (FD_ISSET(ListenSocket[i], &rmask))
1722 {
1723 Port *port;
1724
1725 port = ConnCreate(ListenSocket[i]);
1726 if (port)
1727 {
1728 BackendStartup(port);
1729
1730 /*
1731 * We no longer need the open socket or port structure
1732 * in this process
1733 */
1734 StreamClose(port->sock);
1735 ConnFree(port);
1736 }
1737 }
1738 }
1739 }
1740
1741 /* If we have lost the log collector, try to start a new one */
1742 if (SysLoggerPID == 0 && Logging_collector)
1743 SysLoggerPID = SysLogger_Start();
1744
1745 /*
1746 * If no background writer process is running, and we are not in a
1747 * state that prevents it, start one. It doesn't matter if this
1748 * fails, we'll just try again later. Likewise for the checkpointer.
1749 */
1750 if (pmState == PM_RUN || pmState == PM_RECOVERY ||
1751 pmState == PM_HOT_STANDBY)
1752 {
1753 if (CheckpointerPID == 0)
1754 CheckpointerPID = StartCheckpointer();
1755 if (BgWriterPID == 0)
1756 BgWriterPID = StartBackgroundWriter();
1757 }
1758
1759 /*
1760 * Likewise, if we have lost the walwriter process, try to start a new
1761 * one. But this is needed only in normal operation (else we cannot
1762 * be writing any new WAL).
1763 */
1764 if (WalWriterPID == 0 && pmState == PM_RUN)
1765 WalWriterPID = StartWalWriter();
1766
1767 /*
1768 * If we have lost the autovacuum launcher, try to start a new one. We
1769 * don't want autovacuum to run in binary upgrade mode because
1770 * autovacuum might update relfrozenxid for empty tables before the
1771 * physical files are put in place.
1772 */
1773 if (!IsBinaryUpgrade && AutoVacPID == 0 &&
1774 (AutoVacuumingActive() || start_autovac_launcher) &&
1775 pmState == PM_RUN)
1776 {
1777 AutoVacPID = StartAutoVacLauncher();
1778 if (AutoVacPID != 0)
1779 start_autovac_launcher = false; /* signal processed */
1780 }
1781
1782 /* If we have lost the stats collector, try to start a new one */
1783 if (PgStatPID == 0 &&
1784 (pmState == PM_RUN || pmState == PM_HOT_STANDBY))
1785 PgStatPID = pgstat_start();
1786
1787 /* If we have lost the archiver, try to start a new one. */
1788 if (PgArchPID == 0 && PgArchStartupAllowed())
1789 PgArchPID = pgarch_start();
1790
1791 /* If we need to signal the autovacuum launcher, do so now */
1792 if (avlauncher_needs_signal)
1793 {
1794 avlauncher_needs_signal = false;
1795 if (AutoVacPID != 0)
1796 kill(AutoVacPID, SIGUSR2);
1797 }
1798
1799 /* If we need to start a WAL receiver, try to do that now */
1800 if (WalReceiverRequested)
1801 MaybeStartWalReceiver();
1802
1803 /* Get other worker processes running, if needed */
1804 if (StartWorkerNeeded || HaveCrashedWorker)
1805 maybe_start_bgworkers();
1806
1807 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1808
1809 /*
1810 * With assertions enabled, check regularly for appearance of
1811 * additional threads. All builds check at start and exit.
1812 */
1813 Assert(pthread_is_threaded_np() == 0);
1814 #endif
1815
1816 /*
1817 * Lastly, check to see if it's time to do some things that we don't
1818 * want to do every single time through the loop, because they're a
1819 * bit expensive. Note that there's up to a minute of slop in when
1820 * these tasks will be performed, since DetermineSleepTime() will let
1821 * us sleep at most that long; except for SIGKILL timeout which has
1822 * special-case logic there.
1823 */
1824 now = time(NULL);
1825
1826 /*
1827 * If we already sent SIGQUIT to children and they are slow to shut
1828 * down, it's time to send them SIGKILL. This doesn't happen
1829 * normally, but under certain conditions backends can get stuck while
1830 * shutting down. This is a last measure to get them unwedged.
1831 *
1832 * Note we also do this during recovery from a process crash.
1833 */
1834 if ((Shutdown >= ImmediateShutdown || (FatalError && !SendStop)) &&
1835 AbortStartTime != 0 &&
1836 (now - AbortStartTime) >= SIGKILL_CHILDREN_AFTER_SECS)
1837 {
1838 /* We were gentle with them before. Not anymore */
1839 TerminateChildren(SIGKILL);
1840 /* reset flag so we don't SIGKILL again */
1841 AbortStartTime = 0;
1842 }
1843
1844 /*
1845 * Once a minute, verify that postmaster.pid hasn't been removed or
1846 * overwritten. If it has, we force a shutdown. This avoids having
1847 * postmasters and child processes hanging around after their database
1848 * is gone, and maybe causing problems if a new database cluster is
1849 * created in the same place. It also provides some protection
1850 * against a DBA foolishly removing postmaster.pid and manually
1851 * starting a new postmaster. Data corruption is likely to ensue from
1852 * that anyway, but we can minimize the damage by aborting ASAP.
1853 */
1854 if (now - last_lockfile_recheck_time >= 1 * SECS_PER_MINUTE)
1855 {
1856 if (!RecheckDataDirLockFile())
1857 {
1858 ereport(LOG,
1859 (errmsg("performing immediate shutdown because data directory lock file is invalid")));
1860 kill(MyProcPid, SIGQUIT);
1861 }
1862 last_lockfile_recheck_time = now;
1863 }
1864
1865 /*
1866 * Touch Unix socket and lock files every 58 minutes, to ensure that
1867 * they are not removed by overzealous /tmp-cleaning tasks. We assume
1868 * no one runs cleaners with cutoff times of less than an hour ...
1869 */
1870 if (now - last_touch_time >= 58 * SECS_PER_MINUTE)
1871 {
1872 TouchSocketFiles();
1873 TouchSocketLockFiles();
1874 last_touch_time = now;
1875 }
1876 }
1877 }
1878
1879 /*
1880 * Initialise the masks for select() for the ports we are listening on.
1881 * Return the number of sockets to listen on.
1882 */
1883 static int
initMasks(fd_set * rmask)1884 initMasks(fd_set *rmask)
1885 {
1886 int maxsock = -1;
1887 int i;
1888
1889 FD_ZERO(rmask);
1890
1891 for (i = 0; i < MAXLISTEN; i++)
1892 {
1893 int fd = ListenSocket[i];
1894
1895 if (fd == PGINVALID_SOCKET)
1896 break;
1897 FD_SET(fd, rmask);
1898
1899 if (fd > maxsock)
1900 maxsock = fd;
1901 }
1902
1903 return maxsock + 1;
1904 }
1905
1906
1907 /*
1908 * Read a client's startup packet and do something according to it.
1909 *
1910 * Returns STATUS_OK or STATUS_ERROR, or might call ereport(FATAL) and
1911 * not return at all.
1912 *
1913 * (Note that ereport(FATAL) stuff is sent to the client, so only use it
1914 * if that's what you want. Return STATUS_ERROR if you don't want to
1915 * send anything to the client, which would typically be appropriate
1916 * if we detect a communications failure.)
1917 */
1918 static int
ProcessStartupPacket(Port * port,bool SSLdone)1919 ProcessStartupPacket(Port *port, bool SSLdone)
1920 {
1921 int32 len;
1922 void *buf;
1923 ProtocolVersion proto;
1924 MemoryContext oldcontext;
1925
1926 pq_startmsgread();
1927 if (pq_getbytes((char *) &len, 4) == EOF)
1928 {
1929 /*
1930 * EOF after SSLdone probably means the client didn't like our
1931 * response to NEGOTIATE_SSL_CODE. That's not an error condition, so
1932 * don't clutter the log with a complaint.
1933 */
1934 if (!SSLdone)
1935 ereport(COMMERROR,
1936 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1937 errmsg("incomplete startup packet")));
1938 return STATUS_ERROR;
1939 }
1940
1941 len = pg_ntoh32(len);
1942 len -= 4;
1943
1944 if (len < (int32) sizeof(ProtocolVersion) ||
1945 len > MAX_STARTUP_PACKET_LENGTH)
1946 {
1947 ereport(COMMERROR,
1948 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1949 errmsg("invalid length of startup packet")));
1950 return STATUS_ERROR;
1951 }
1952
1953 /*
1954 * Allocate at least the size of an old-style startup packet, plus one
1955 * extra byte, and make sure all are zeroes. This ensures we will have
1956 * null termination of all strings, in both fixed- and variable-length
1957 * packet layouts.
1958 */
1959 if (len <= (int32) sizeof(StartupPacket))
1960 buf = palloc0(sizeof(StartupPacket) + 1);
1961 else
1962 buf = palloc0(len + 1);
1963
1964 if (pq_getbytes(buf, len) == EOF)
1965 {
1966 ereport(COMMERROR,
1967 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1968 errmsg("incomplete startup packet")));
1969 return STATUS_ERROR;
1970 }
1971 pq_endmsgread();
1972
1973 /*
1974 * The first field is either a protocol version number or a special
1975 * request code.
1976 */
1977 port->proto = proto = pg_ntoh32(*((ProtocolVersion *) buf));
1978
1979 if (proto == CANCEL_REQUEST_CODE)
1980 {
1981 processCancelRequest(port, buf);
1982 /* Not really an error, but we don't want to proceed further */
1983 return STATUS_ERROR;
1984 }
1985
1986 if (proto == NEGOTIATE_SSL_CODE && !SSLdone)
1987 {
1988 char SSLok;
1989
1990 #ifdef USE_SSL
1991 /* No SSL when disabled or on Unix sockets */
1992 if (!LoadedSSL || IS_AF_UNIX(port->laddr.addr.ss_family))
1993 SSLok = 'N';
1994 else
1995 SSLok = 'S'; /* Support for SSL */
1996 #else
1997 SSLok = 'N'; /* No support for SSL */
1998 #endif
1999
2000 retry1:
2001 if (send(port->sock, &SSLok, 1, 0) != 1)
2002 {
2003 if (errno == EINTR)
2004 goto retry1; /* if interrupted, just retry */
2005 ereport(COMMERROR,
2006 (errcode_for_socket_access(),
2007 errmsg("failed to send SSL negotiation response: %m")));
2008 return STATUS_ERROR; /* close the connection */
2009 }
2010
2011 #ifdef USE_SSL
2012 if (SSLok == 'S' && secure_open_server(port) == -1)
2013 return STATUS_ERROR;
2014 #endif
2015
2016 /*
2017 * At this point we should have no data already buffered. If we do,
2018 * it was received before we performed the SSL handshake, so it wasn't
2019 * encrypted and indeed may have been injected by a man-in-the-middle.
2020 * We report this case to the client.
2021 */
2022 if (pq_buffer_has_data())
2023 ereport(FATAL,
2024 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2025 errmsg("received unencrypted data after SSL request"),
2026 errdetail("This could be either a client-software bug or evidence of an attempted man-in-the-middle attack.")));
2027
2028 /* regular startup packet, cancel, etc packet should follow... */
2029 /* but not another SSL negotiation request */
2030 return ProcessStartupPacket(port, true);
2031 }
2032
2033 /* Could add additional special packet types here */
2034
2035 /*
2036 * Set FrontendProtocol now so that ereport() knows what format to send if
2037 * we fail during startup.
2038 */
2039 FrontendProtocol = proto;
2040
2041 /* Check that the major protocol version is in range. */
2042 if (PG_PROTOCOL_MAJOR(proto) < PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST) ||
2043 PG_PROTOCOL_MAJOR(proto) > PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST))
2044 ereport(FATAL,
2045 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2046 errmsg("unsupported frontend protocol %u.%u: server supports %u.0 to %u.%u",
2047 PG_PROTOCOL_MAJOR(proto), PG_PROTOCOL_MINOR(proto),
2048 PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST),
2049 PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST),
2050 PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST))));
2051
2052 /*
2053 * Now fetch parameters out of startup packet and save them into the Port
2054 * structure. All data structures attached to the Port struct must be
2055 * allocated in TopMemoryContext so that they will remain available in a
2056 * running backend (even after PostmasterContext is destroyed). We need
2057 * not worry about leaking this storage on failure, since we aren't in the
2058 * postmaster process anymore.
2059 */
2060 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
2061
2062 if (PG_PROTOCOL_MAJOR(proto) >= 3)
2063 {
2064 int32 offset = sizeof(ProtocolVersion);
2065 List *unrecognized_protocol_options = NIL;
2066
2067 /*
2068 * Scan packet body for name/option pairs. We can assume any string
2069 * beginning within the packet body is null-terminated, thanks to
2070 * zeroing extra byte above.
2071 */
2072 port->guc_options = NIL;
2073
2074 while (offset < len)
2075 {
2076 char *nameptr = ((char *) buf) + offset;
2077 int32 valoffset;
2078 char *valptr;
2079
2080 if (*nameptr == '\0')
2081 break; /* found packet terminator */
2082 valoffset = offset + strlen(nameptr) + 1;
2083 if (valoffset >= len)
2084 break; /* missing value, will complain below */
2085 valptr = ((char *) buf) + valoffset;
2086
2087 if (strcmp(nameptr, "database") == 0)
2088 port->database_name = pstrdup(valptr);
2089 else if (strcmp(nameptr, "user") == 0)
2090 port->user_name = pstrdup(valptr);
2091 else if (strcmp(nameptr, "options") == 0)
2092 port->cmdline_options = pstrdup(valptr);
2093 else if (strcmp(nameptr, "replication") == 0)
2094 {
2095 /*
2096 * Due to backward compatibility concerns the replication
2097 * parameter is a hybrid beast which allows the value to be
2098 * either boolean or the string 'database'. The latter
2099 * connects to a specific database which is e.g. required for
2100 * logical decoding while.
2101 */
2102 if (strcmp(valptr, "database") == 0)
2103 {
2104 am_walsender = true;
2105 am_db_walsender = true;
2106 }
2107 else if (!parse_bool(valptr, &am_walsender))
2108 ereport(FATAL,
2109 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
2110 errmsg("invalid value for parameter \"%s\": \"%s\"",
2111 "replication",
2112 valptr),
2113 errhint("Valid values are: \"false\", 0, \"true\", 1, \"database\".")));
2114 }
2115 else if (strncmp(nameptr, "_pq_.", 5) == 0)
2116 {
2117 /*
2118 * Any option beginning with _pq_. is reserved for use as a
2119 * protocol-level option, but at present no such options are
2120 * defined.
2121 */
2122 unrecognized_protocol_options =
2123 lappend(unrecognized_protocol_options, pstrdup(nameptr));
2124 }
2125 else
2126 {
2127 /* Assume it's a generic GUC option */
2128 port->guc_options = lappend(port->guc_options,
2129 pstrdup(nameptr));
2130 port->guc_options = lappend(port->guc_options,
2131 pstrdup(valptr));
2132 }
2133 offset = valoffset + strlen(valptr) + 1;
2134 }
2135
2136 /*
2137 * If we didn't find a packet terminator exactly at the end of the
2138 * given packet length, complain.
2139 */
2140 if (offset != len - 1)
2141 ereport(FATAL,
2142 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2143 errmsg("invalid startup packet layout: expected terminator as last byte")));
2144
2145 /*
2146 * If the client requested a newer protocol version or if the client
2147 * requested any protocol options we didn't recognize, let them know
2148 * the newest minor protocol version we do support and the names of
2149 * any unrecognized options.
2150 */
2151 if (PG_PROTOCOL_MINOR(proto) > PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST) ||
2152 unrecognized_protocol_options != NIL)
2153 SendNegotiateProtocolVersion(unrecognized_protocol_options);
2154 }
2155 else
2156 {
2157 /*
2158 * Get the parameters from the old-style, fixed-width-fields startup
2159 * packet as C strings. The packet destination was cleared first so a
2160 * short packet has zeros silently added. We have to be prepared to
2161 * truncate the pstrdup result for oversize fields, though.
2162 */
2163 StartupPacket *packet = (StartupPacket *) buf;
2164
2165 port->database_name = pstrdup(packet->database);
2166 if (strlen(port->database_name) > sizeof(packet->database))
2167 port->database_name[sizeof(packet->database)] = '\0';
2168 port->user_name = pstrdup(packet->user);
2169 if (strlen(port->user_name) > sizeof(packet->user))
2170 port->user_name[sizeof(packet->user)] = '\0';
2171 port->cmdline_options = pstrdup(packet->options);
2172 if (strlen(port->cmdline_options) > sizeof(packet->options))
2173 port->cmdline_options[sizeof(packet->options)] = '\0';
2174 port->guc_options = NIL;
2175 }
2176
2177 /* Check a user name was given. */
2178 if (port->user_name == NULL || port->user_name[0] == '\0')
2179 ereport(FATAL,
2180 (errcode(ERRCODE_INVALID_AUTHORIZATION_SPECIFICATION),
2181 errmsg("no PostgreSQL user name specified in startup packet")));
2182
2183 /* The database defaults to the user name. */
2184 if (port->database_name == NULL || port->database_name[0] == '\0')
2185 port->database_name = pstrdup(port->user_name);
2186
2187 if (Db_user_namespace)
2188 {
2189 /*
2190 * If user@, it is a global user, remove '@'. We only want to do this
2191 * if there is an '@' at the end and no earlier in the user string or
2192 * they may fake as a local user of another database attaching to this
2193 * database.
2194 */
2195 if (strchr(port->user_name, '@') ==
2196 port->user_name + strlen(port->user_name) - 1)
2197 *strchr(port->user_name, '@') = '\0';
2198 else
2199 {
2200 /* Append '@' and dbname */
2201 port->user_name = psprintf("%s@%s", port->user_name, port->database_name);
2202 }
2203 }
2204
2205 /*
2206 * Truncate given database and user names to length of a Postgres name.
2207 * This avoids lookup failures when overlength names are given.
2208 */
2209 if (strlen(port->database_name) >= NAMEDATALEN)
2210 port->database_name[NAMEDATALEN - 1] = '\0';
2211 if (strlen(port->user_name) >= NAMEDATALEN)
2212 port->user_name[NAMEDATALEN - 1] = '\0';
2213
2214 /*
2215 * Normal walsender backends, e.g. for streaming replication, are not
2216 * connected to a particular database. But walsenders used for logical
2217 * replication need to connect to a specific database. We allow streaming
2218 * replication commands to be issued even if connected to a database as it
2219 * can make sense to first make a basebackup and then stream changes
2220 * starting from that.
2221 */
2222 if (am_walsender && !am_db_walsender)
2223 port->database_name[0] = '\0';
2224
2225 /*
2226 * Done putting stuff in TopMemoryContext.
2227 */
2228 MemoryContextSwitchTo(oldcontext);
2229
2230 /*
2231 * If we're going to reject the connection due to database state, say so
2232 * now instead of wasting cycles on an authentication exchange. (This also
2233 * allows a pg_ping utility to be written.)
2234 */
2235 switch (port->canAcceptConnections)
2236 {
2237 case CAC_STARTUP:
2238 ereport(FATAL,
2239 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2240 errmsg("the database system is starting up")));
2241 break;
2242 case CAC_SHUTDOWN:
2243 ereport(FATAL,
2244 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2245 errmsg("the database system is shutting down")));
2246 break;
2247 case CAC_RECOVERY:
2248 ereport(FATAL,
2249 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2250 errmsg("the database system is in recovery mode")));
2251 break;
2252 case CAC_TOOMANY:
2253 ereport(FATAL,
2254 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
2255 errmsg("sorry, too many clients already")));
2256 break;
2257 case CAC_SUPERUSER:
2258 /* OK for now, will check in InitPostgres */
2259 break;
2260 case CAC_OK:
2261 break;
2262 }
2263
2264 return STATUS_OK;
2265 }
2266
2267 /*
2268 * Send a NegotiateProtocolVersion to the client. This lets the client know
2269 * that they have requested a newer minor protocol version than we are able
2270 * to speak. We'll speak the highest version we know about; the client can,
2271 * of course, abandon the connection if that's a problem.
2272 *
2273 * We also include in the response a list of protocol options we didn't
2274 * understand. This allows clients to include optional parameters that might
2275 * be present either in newer protocol versions or third-party protocol
2276 * extensions without fear of having to reconnect if those options are not
2277 * understood, while at the same time making certain that the client is aware
2278 * of which options were actually accepted.
2279 */
2280 static void
SendNegotiateProtocolVersion(List * unrecognized_protocol_options)2281 SendNegotiateProtocolVersion(List *unrecognized_protocol_options)
2282 {
2283 StringInfoData buf;
2284 ListCell *lc;
2285
2286 pq_beginmessage(&buf, 'v'); /* NegotiateProtocolVersion */
2287 pq_sendint32(&buf, PG_PROTOCOL_LATEST);
2288 pq_sendint32(&buf, list_length(unrecognized_protocol_options));
2289 foreach(lc, unrecognized_protocol_options)
2290 pq_sendstring(&buf, lfirst(lc));
2291 pq_endmessage(&buf);
2292
2293 /* no need to flush, some other message will follow */
2294 }
2295
2296 /*
2297 * The client has sent a cancel request packet, not a normal
2298 * start-a-new-connection packet. Perform the necessary processing.
2299 * Nothing is sent back to the client.
2300 */
2301 static void
processCancelRequest(Port * port,void * pkt)2302 processCancelRequest(Port *port, void *pkt)
2303 {
2304 CancelRequestPacket *canc = (CancelRequestPacket *) pkt;
2305 int backendPID;
2306 int32 cancelAuthCode;
2307 Backend *bp;
2308
2309 #ifndef EXEC_BACKEND
2310 dlist_iter iter;
2311 #else
2312 int i;
2313 #endif
2314
2315 backendPID = (int) pg_ntoh32(canc->backendPID);
2316 cancelAuthCode = (int32) pg_ntoh32(canc->cancelAuthCode);
2317
2318 /*
2319 * See if we have a matching backend. In the EXEC_BACKEND case, we can no
2320 * longer access the postmaster's own backend list, and must rely on the
2321 * duplicate array in shared memory.
2322 */
2323 #ifndef EXEC_BACKEND
2324 dlist_foreach(iter, &BackendList)
2325 {
2326 bp = dlist_container(Backend, elem, iter.cur);
2327 #else
2328 for (i = MaxLivePostmasterChildren() - 1; i >= 0; i--)
2329 {
2330 bp = (Backend *) &ShmemBackendArray[i];
2331 #endif
2332 if (bp->pid == backendPID)
2333 {
2334 if (bp->cancel_key == cancelAuthCode)
2335 {
2336 /* Found a match; signal that backend to cancel current op */
2337 ereport(DEBUG2,
2338 (errmsg_internal("processing cancel request: sending SIGINT to process %d",
2339 backendPID)));
2340 signal_child(bp->pid, SIGINT);
2341 }
2342 else
2343 /* Right PID, wrong key: no way, Jose */
2344 ereport(LOG,
2345 (errmsg("wrong key in cancel request for process %d",
2346 backendPID)));
2347 return;
2348 }
2349 }
2350
2351 /* No matching backend */
2352 ereport(LOG,
2353 (errmsg("PID %d in cancel request did not match any process",
2354 backendPID)));
2355 }
2356
2357 /*
2358 * canAcceptConnections --- check to see if database state allows connections
2359 * of the specified type. backend_type can be BACKEND_TYPE_NORMAL,
2360 * BACKEND_TYPE_AUTOVAC, or BACKEND_TYPE_BGWORKER. (Note that we don't yet
2361 * know whether a NORMAL connection might turn into a walsender.)
2362 */
2363 static CAC_state
2364 canAcceptConnections(int backend_type)
2365 {
2366 CAC_state result = CAC_OK;
2367
2368 /*
2369 * Can't start backends when in startup/shutdown/inconsistent recovery
2370 * state. We treat autovac workers the same as user backends for this
2371 * purpose. However, bgworkers are excluded from this test; we expect
2372 * bgworker_should_start_now() decided whether the DB state allows them.
2373 */
2374 if (pmState != PM_RUN && pmState != PM_HOT_STANDBY &&
2375 backend_type != BACKEND_TYPE_BGWORKER)
2376 {
2377 if (Shutdown > NoShutdown)
2378 return CAC_SHUTDOWN; /* shutdown is pending */
2379 else if (!FatalError &&
2380 (pmState == PM_STARTUP ||
2381 pmState == PM_RECOVERY))
2382 return CAC_STARTUP; /* normal startup */
2383 else
2384 return CAC_RECOVERY; /* else must be crash recovery */
2385 }
2386
2387 /*
2388 * "Smart shutdown" restrictions are applied only to normal connections,
2389 * not to autovac workers or bgworkers. When only superusers can connect,
2390 * we return CAC_SUPERUSER to indicate that superuserness must be checked
2391 * later. Note that neither CAC_OK nor CAC_SUPERUSER can safely be
2392 * returned until we have checked for too many children.
2393 */
2394 if (connsAllowed != ALLOW_ALL_CONNS &&
2395 backend_type == BACKEND_TYPE_NORMAL)
2396 {
2397 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
2398 result = CAC_SUPERUSER; /* allow superusers only */
2399 else
2400 return CAC_SHUTDOWN; /* shutdown is pending */
2401 }
2402
2403 /*
2404 * Don't start too many children.
2405 *
2406 * We allow more connections here than we can have backends because some
2407 * might still be authenticating; they might fail auth, or some existing
2408 * backend might exit before the auth cycle is completed. The exact
2409 * MaxBackends limit is enforced when a new backend tries to join the
2410 * shared-inval backend array.
2411 *
2412 * The limit here must match the sizes of the per-child-process arrays;
2413 * see comments for MaxLivePostmasterChildren().
2414 */
2415 if (CountChildren(BACKEND_TYPE_ALL) >= MaxLivePostmasterChildren())
2416 result = CAC_TOOMANY;
2417
2418 return result;
2419 }
2420
2421
2422 /*
2423 * ConnCreate -- create a local connection data structure
2424 *
2425 * Returns NULL on failure, other than out-of-memory which is fatal.
2426 */
2427 static Port *
2428 ConnCreate(int serverFd)
2429 {
2430 Port *port;
2431
2432 if (!(port = (Port *) calloc(1, sizeof(Port))))
2433 {
2434 ereport(LOG,
2435 (errcode(ERRCODE_OUT_OF_MEMORY),
2436 errmsg("out of memory")));
2437 ExitPostmaster(1);
2438 }
2439
2440 if (StreamConnection(serverFd, port) != STATUS_OK)
2441 {
2442 if (port->sock != PGINVALID_SOCKET)
2443 StreamClose(port->sock);
2444 ConnFree(port);
2445 return NULL;
2446 }
2447
2448 /*
2449 * Allocate GSSAPI specific state struct
2450 */
2451 #ifndef EXEC_BACKEND
2452 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
2453 port->gss = (pg_gssinfo *) calloc(1, sizeof(pg_gssinfo));
2454 if (!port->gss)
2455 {
2456 ereport(LOG,
2457 (errcode(ERRCODE_OUT_OF_MEMORY),
2458 errmsg("out of memory")));
2459 ExitPostmaster(1);
2460 }
2461 #endif
2462 #endif
2463
2464 return port;
2465 }
2466
2467
2468 /*
2469 * ConnFree -- free a local connection data structure
2470 */
2471 static void
2472 ConnFree(Port *conn)
2473 {
2474 #ifdef USE_SSL
2475 secure_close(conn);
2476 #endif
2477 if (conn->gss)
2478 free(conn->gss);
2479 free(conn);
2480 }
2481
2482
2483 /*
2484 * ClosePostmasterPorts -- close all the postmaster's open sockets
2485 *
2486 * This is called during child process startup to release file descriptors
2487 * that are not needed by that child process. The postmaster still has
2488 * them open, of course.
2489 *
2490 * Note: we pass am_syslogger as a boolean because we don't want to set
2491 * the global variable yet when this is called.
2492 */
2493 void
2494 ClosePostmasterPorts(bool am_syslogger)
2495 {
2496 int i;
2497
2498 #ifndef WIN32
2499
2500 /*
2501 * Close the write end of postmaster death watch pipe. It's important to
2502 * do this as early as possible, so that if postmaster dies, others won't
2503 * think that it's still running because we're holding the pipe open.
2504 */
2505 if (close(postmaster_alive_fds[POSTMASTER_FD_OWN]))
2506 ereport(FATAL,
2507 (errcode_for_file_access(),
2508 errmsg_internal("could not close postmaster death monitoring pipe in child process: %m")));
2509 postmaster_alive_fds[POSTMASTER_FD_OWN] = -1;
2510 #endif
2511
2512 /* Close the listen sockets */
2513 for (i = 0; i < MAXLISTEN; i++)
2514 {
2515 if (ListenSocket[i] != PGINVALID_SOCKET)
2516 {
2517 StreamClose(ListenSocket[i]);
2518 ListenSocket[i] = PGINVALID_SOCKET;
2519 }
2520 }
2521
2522 /* If using syslogger, close the read side of the pipe */
2523 if (!am_syslogger)
2524 {
2525 #ifndef WIN32
2526 if (syslogPipe[0] >= 0)
2527 close(syslogPipe[0]);
2528 syslogPipe[0] = -1;
2529 #else
2530 if (syslogPipe[0])
2531 CloseHandle(syslogPipe[0]);
2532 syslogPipe[0] = 0;
2533 #endif
2534 }
2535
2536 #ifdef USE_BONJOUR
2537 /* If using Bonjour, close the connection to the mDNS daemon */
2538 if (bonjour_sdref)
2539 close(DNSServiceRefSockFD(bonjour_sdref));
2540 #endif
2541 }
2542
2543
2544 /*
2545 * reset_shared -- reset shared memory and semaphores
2546 */
2547 static void
2548 reset_shared(int port)
2549 {
2550 /*
2551 * Create or re-create shared memory and semaphores.
2552 *
2553 * Note: in each "cycle of life" we will normally assign the same IPC keys
2554 * (if using SysV shmem and/or semas), since the port number is used to
2555 * determine IPC keys. This helps ensure that we will clean up dead IPC
2556 * objects if the postmaster crashes and is restarted.
2557 */
2558 CreateSharedMemoryAndSemaphores(port);
2559 }
2560
2561
2562 /*
2563 * SIGHUP -- reread config files, and tell children to do same
2564 */
2565 static void
2566 SIGHUP_handler(SIGNAL_ARGS)
2567 {
2568 int save_errno = errno;
2569
2570 /*
2571 * We rely on the signal mechanism to have blocked all signals ... except
2572 * on Windows, which lacks sigaction(), so we have to do it manually.
2573 */
2574 #ifdef WIN32
2575 PG_SETMASK(&BlockSig);
2576 #endif
2577
2578 if (Shutdown <= SmartShutdown)
2579 {
2580 ereport(LOG,
2581 (errmsg("received SIGHUP, reloading configuration files")));
2582 ProcessConfigFile(PGC_SIGHUP);
2583 SignalChildren(SIGHUP);
2584 if (StartupPID != 0)
2585 signal_child(StartupPID, SIGHUP);
2586 if (BgWriterPID != 0)
2587 signal_child(BgWriterPID, SIGHUP);
2588 if (CheckpointerPID != 0)
2589 signal_child(CheckpointerPID, SIGHUP);
2590 if (WalWriterPID != 0)
2591 signal_child(WalWriterPID, SIGHUP);
2592 if (WalReceiverPID != 0)
2593 signal_child(WalReceiverPID, SIGHUP);
2594 if (AutoVacPID != 0)
2595 signal_child(AutoVacPID, SIGHUP);
2596 if (PgArchPID != 0)
2597 signal_child(PgArchPID, SIGHUP);
2598 if (SysLoggerPID != 0)
2599 signal_child(SysLoggerPID, SIGHUP);
2600 if (PgStatPID != 0)
2601 signal_child(PgStatPID, SIGHUP);
2602
2603 /* Reload authentication config files too */
2604 if (!load_hba())
2605 ereport(LOG,
2606 (errmsg("pg_hba.conf was not reloaded")));
2607
2608 if (!load_ident())
2609 ereport(LOG,
2610 (errmsg("pg_ident.conf was not reloaded")));
2611
2612 #ifdef USE_SSL
2613 /* Reload SSL configuration as well */
2614 if (EnableSSL)
2615 {
2616 if (secure_initialize(false) == 0)
2617 LoadedSSL = true;
2618 else
2619 ereport(LOG,
2620 (errmsg("SSL configuration was not reloaded")));
2621 }
2622 else
2623 {
2624 secure_destroy();
2625 LoadedSSL = false;
2626 }
2627 #endif
2628
2629 #ifdef EXEC_BACKEND
2630 /* Update the starting-point file for future children */
2631 write_nondefault_variables(PGC_SIGHUP);
2632 #endif
2633 }
2634
2635 #ifdef WIN32
2636 PG_SETMASK(&UnBlockSig);
2637 #endif
2638
2639 errno = save_errno;
2640 }
2641
2642
2643 /*
2644 * pmdie -- signal handler for processing various postmaster signals.
2645 */
2646 static void
2647 pmdie(SIGNAL_ARGS)
2648 {
2649 int save_errno = errno;
2650
2651 /*
2652 * We rely on the signal mechanism to have blocked all signals ... except
2653 * on Windows, which lacks sigaction(), so we have to do it manually.
2654 */
2655 #ifdef WIN32
2656 PG_SETMASK(&BlockSig);
2657 #endif
2658
2659 ereport(DEBUG2,
2660 (errmsg_internal("postmaster received signal %d",
2661 postgres_signal_arg)));
2662
2663 switch (postgres_signal_arg)
2664 {
2665 case SIGTERM:
2666
2667 /*
2668 * Smart Shutdown:
2669 *
2670 * Wait for children to end their work, then shut down.
2671 */
2672 if (Shutdown >= SmartShutdown)
2673 break;
2674 Shutdown = SmartShutdown;
2675 ereport(LOG,
2676 (errmsg("received smart shutdown request")));
2677
2678 /* Report status */
2679 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2680 #ifdef USE_SYSTEMD
2681 sd_notify(0, "STOPPING=1");
2682 #endif
2683
2684 /*
2685 * If we reached normal running, we have to wait for any online
2686 * backup mode to end; otherwise go straight to waiting for client
2687 * backends to exit. (The difference is that in the former state,
2688 * we'll still let in new superuser clients, so that somebody can
2689 * end the online backup mode.) If already in PM_STOP_BACKENDS or
2690 * a later state, do not change it.
2691 */
2692 if (pmState == PM_RUN)
2693 connsAllowed = ALLOW_SUPERUSER_CONNS;
2694 else if (pmState == PM_HOT_STANDBY)
2695 connsAllowed = ALLOW_NO_CONNS;
2696 else if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2697 {
2698 /* There should be no clients, so proceed to stop children */
2699 pmState = PM_STOP_BACKENDS;
2700 }
2701
2702 /*
2703 * Now wait for online backup mode to end and backends to exit. If
2704 * that is already the case, PostmasterStateMachine will take the
2705 * next step.
2706 */
2707 PostmasterStateMachine();
2708 break;
2709
2710 case SIGINT:
2711
2712 /*
2713 * Fast Shutdown:
2714 *
2715 * Abort all children with SIGTERM (rollback active transactions
2716 * and exit) and shut down when they are gone.
2717 */
2718 if (Shutdown >= FastShutdown)
2719 break;
2720 Shutdown = FastShutdown;
2721 ereport(LOG,
2722 (errmsg("received fast shutdown request")));
2723
2724 /* Report status */
2725 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2726 #ifdef USE_SYSTEMD
2727 sd_notify(0, "STOPPING=1");
2728 #endif
2729
2730 if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2731 {
2732 /* Just shut down background processes silently */
2733 pmState = PM_STOP_BACKENDS;
2734 }
2735 else if (pmState == PM_RUN ||
2736 pmState == PM_HOT_STANDBY)
2737 {
2738 /* Report that we're about to zap live client sessions */
2739 ereport(LOG,
2740 (errmsg("aborting any active transactions")));
2741 pmState = PM_STOP_BACKENDS;
2742 }
2743
2744 /*
2745 * PostmasterStateMachine will issue any necessary signals, or
2746 * take the next step if no child processes need to be killed.
2747 */
2748 PostmasterStateMachine();
2749 break;
2750
2751 case SIGQUIT:
2752
2753 /*
2754 * Immediate Shutdown:
2755 *
2756 * abort all children with SIGQUIT, wait for them to exit,
2757 * terminate remaining ones with SIGKILL, then exit without
2758 * attempt to properly shut down the data base system.
2759 */
2760 if (Shutdown >= ImmediateShutdown)
2761 break;
2762 Shutdown = ImmediateShutdown;
2763 ereport(LOG,
2764 (errmsg("received immediate shutdown request")));
2765
2766 /* Report status */
2767 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2768 #ifdef USE_SYSTEMD
2769 sd_notify(0, "STOPPING=1");
2770 #endif
2771
2772 TerminateChildren(SIGQUIT);
2773 pmState = PM_WAIT_BACKENDS;
2774
2775 /* set stopwatch for them to die */
2776 AbortStartTime = time(NULL);
2777
2778 /*
2779 * Now wait for backends to exit. If there are none,
2780 * PostmasterStateMachine will take the next step.
2781 */
2782 PostmasterStateMachine();
2783 break;
2784 }
2785
2786 #ifdef WIN32
2787 PG_SETMASK(&UnBlockSig);
2788 #endif
2789
2790 errno = save_errno;
2791 }
2792
2793 /*
2794 * Reaper -- signal handler to cleanup after a child process dies.
2795 */
2796 static void
2797 reaper(SIGNAL_ARGS)
2798 {
2799 int save_errno = errno;
2800 int pid; /* process id of dead child process */
2801 int exitstatus; /* its exit status */
2802
2803 /*
2804 * We rely on the signal mechanism to have blocked all signals ... except
2805 * on Windows, which lacks sigaction(), so we have to do it manually.
2806 */
2807 #ifdef WIN32
2808 PG_SETMASK(&BlockSig);
2809 #endif
2810
2811 ereport(DEBUG4,
2812 (errmsg_internal("reaping dead processes")));
2813
2814 while ((pid = waitpid(-1, &exitstatus, WNOHANG)) > 0)
2815 {
2816 /*
2817 * Check if this child was a startup process.
2818 */
2819 if (pid == StartupPID)
2820 {
2821 StartupPID = 0;
2822
2823 /*
2824 * Startup process exited in response to a shutdown request (or it
2825 * completed normally regardless of the shutdown request).
2826 */
2827 if (Shutdown > NoShutdown &&
2828 (EXIT_STATUS_0(exitstatus) || EXIT_STATUS_1(exitstatus)))
2829 {
2830 StartupStatus = STARTUP_NOT_RUNNING;
2831 pmState = PM_WAIT_BACKENDS;
2832 /* PostmasterStateMachine logic does the rest */
2833 continue;
2834 }
2835
2836 if (EXIT_STATUS_3(exitstatus))
2837 {
2838 ereport(LOG,
2839 (errmsg("shutdown at recovery target")));
2840 StartupStatus = STARTUP_NOT_RUNNING;
2841 Shutdown = Max(Shutdown, SmartShutdown);
2842 TerminateChildren(SIGTERM);
2843 pmState = PM_WAIT_BACKENDS;
2844 /* PostmasterStateMachine logic does the rest */
2845 continue;
2846 }
2847
2848 /*
2849 * Unexpected exit of startup process (including FATAL exit)
2850 * during PM_STARTUP is treated as catastrophic. There are no
2851 * other processes running yet, so we can just exit.
2852 */
2853 if (pmState == PM_STARTUP && !EXIT_STATUS_0(exitstatus))
2854 {
2855 LogChildExit(LOG, _("startup process"),
2856 pid, exitstatus);
2857 ereport(LOG,
2858 (errmsg("aborting startup due to startup process failure")));
2859 ExitPostmaster(1);
2860 }
2861
2862 /*
2863 * After PM_STARTUP, any unexpected exit (including FATAL exit) of
2864 * the startup process is catastrophic, so kill other children,
2865 * and set StartupStatus so we don't try to reinitialize after
2866 * they're gone. Exception: if StartupStatus is STARTUP_SIGNALED,
2867 * then we previously sent the startup process a SIGQUIT; so
2868 * that's probably the reason it died, and we do want to try to
2869 * restart in that case.
2870 */
2871 if (!EXIT_STATUS_0(exitstatus))
2872 {
2873 if (StartupStatus == STARTUP_SIGNALED)
2874 StartupStatus = STARTUP_NOT_RUNNING;
2875 else
2876 StartupStatus = STARTUP_CRASHED;
2877 HandleChildCrash(pid, exitstatus,
2878 _("startup process"));
2879 continue;
2880 }
2881
2882 /*
2883 * Startup succeeded, commence normal operations
2884 */
2885 StartupStatus = STARTUP_NOT_RUNNING;
2886 FatalError = false;
2887 Assert(AbortStartTime == 0);
2888 ReachedNormalRunning = true;
2889 pmState = PM_RUN;
2890 connsAllowed = ALLOW_ALL_CONNS;
2891
2892 /*
2893 * Crank up the background tasks, if we didn't do that already
2894 * when we entered consistent recovery state. It doesn't matter
2895 * if this fails, we'll just try again later.
2896 */
2897 if (CheckpointerPID == 0)
2898 CheckpointerPID = StartCheckpointer();
2899 if (BgWriterPID == 0)
2900 BgWriterPID = StartBackgroundWriter();
2901 if (WalWriterPID == 0)
2902 WalWriterPID = StartWalWriter();
2903
2904 /*
2905 * Likewise, start other special children as needed. In a restart
2906 * situation, some of them may be alive already.
2907 */
2908 if (!IsBinaryUpgrade && AutoVacuumingActive() && AutoVacPID == 0)
2909 AutoVacPID = StartAutoVacLauncher();
2910 if (PgArchStartupAllowed() && PgArchPID == 0)
2911 PgArchPID = pgarch_start();
2912 if (PgStatPID == 0)
2913 PgStatPID = pgstat_start();
2914
2915 /* workers may be scheduled to start now */
2916 maybe_start_bgworkers();
2917
2918 /* at this point we are really open for business */
2919 ereport(LOG,
2920 (errmsg("database system is ready to accept connections")));
2921
2922 /* Report status */
2923 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
2924 #ifdef USE_SYSTEMD
2925 sd_notify(0, "READY=1");
2926 #endif
2927
2928 continue;
2929 }
2930
2931 /*
2932 * Was it the bgwriter? Normal exit can be ignored; we'll start a new
2933 * one at the next iteration of the postmaster's main loop, if
2934 * necessary. Any other exit condition is treated as a crash.
2935 */
2936 if (pid == BgWriterPID)
2937 {
2938 BgWriterPID = 0;
2939 if (!EXIT_STATUS_0(exitstatus))
2940 HandleChildCrash(pid, exitstatus,
2941 _("background writer process"));
2942 continue;
2943 }
2944
2945 /*
2946 * Was it the checkpointer?
2947 */
2948 if (pid == CheckpointerPID)
2949 {
2950 CheckpointerPID = 0;
2951 if (EXIT_STATUS_0(exitstatus) && pmState == PM_SHUTDOWN)
2952 {
2953 /*
2954 * OK, we saw normal exit of the checkpointer after it's been
2955 * told to shut down. We expect that it wrote a shutdown
2956 * checkpoint. (If for some reason it didn't, recovery will
2957 * occur on next postmaster start.)
2958 *
2959 * At this point we should have no normal backend children
2960 * left (else we'd not be in PM_SHUTDOWN state) but we might
2961 * have dead_end children to wait for.
2962 *
2963 * If we have an archiver subprocess, tell it to do a last
2964 * archive cycle and quit. Likewise, if we have walsender
2965 * processes, tell them to send any remaining WAL and quit.
2966 */
2967 Assert(Shutdown > NoShutdown);
2968
2969 /* Waken archiver for the last time */
2970 if (PgArchPID != 0)
2971 signal_child(PgArchPID, SIGUSR2);
2972
2973 /*
2974 * Waken walsenders for the last time. No regular backends
2975 * should be around anymore.
2976 */
2977 SignalChildren(SIGUSR2);
2978
2979 pmState = PM_SHUTDOWN_2;
2980
2981 /*
2982 * We can also shut down the stats collector now; there's
2983 * nothing left for it to do.
2984 */
2985 if (PgStatPID != 0)
2986 signal_child(PgStatPID, SIGQUIT);
2987 }
2988 else
2989 {
2990 /*
2991 * Any unexpected exit of the checkpointer (including FATAL
2992 * exit) is treated as a crash.
2993 */
2994 HandleChildCrash(pid, exitstatus,
2995 _("checkpointer process"));
2996 }
2997
2998 continue;
2999 }
3000
3001 /*
3002 * Was it the wal writer? Normal exit can be ignored; we'll start a
3003 * new one at the next iteration of the postmaster's main loop, if
3004 * necessary. Any other exit condition is treated as a crash.
3005 */
3006 if (pid == WalWriterPID)
3007 {
3008 WalWriterPID = 0;
3009 if (!EXIT_STATUS_0(exitstatus))
3010 HandleChildCrash(pid, exitstatus,
3011 _("WAL writer process"));
3012 continue;
3013 }
3014
3015 /*
3016 * Was it the wal receiver? If exit status is zero (normal) or one
3017 * (FATAL exit), we assume everything is all right just like normal
3018 * backends. (If we need a new wal receiver, we'll start one at the
3019 * next iteration of the postmaster's main loop.)
3020 */
3021 if (pid == WalReceiverPID)
3022 {
3023 WalReceiverPID = 0;
3024 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3025 HandleChildCrash(pid, exitstatus,
3026 _("WAL receiver process"));
3027 continue;
3028 }
3029
3030 /*
3031 * Was it the autovacuum launcher? Normal exit can be ignored; we'll
3032 * start a new one at the next iteration of the postmaster's main
3033 * loop, if necessary. Any other exit condition is treated as a
3034 * crash.
3035 */
3036 if (pid == AutoVacPID)
3037 {
3038 AutoVacPID = 0;
3039 if (!EXIT_STATUS_0(exitstatus))
3040 HandleChildCrash(pid, exitstatus,
3041 _("autovacuum launcher process"));
3042 continue;
3043 }
3044
3045 /*
3046 * Was it the archiver? If so, just try to start a new one; no need
3047 * to force reset of the rest of the system. (If fail, we'll try
3048 * again in future cycles of the main loop.). Unless we were waiting
3049 * for it to shut down; don't restart it in that case, and
3050 * PostmasterStateMachine() will advance to the next shutdown step.
3051 */
3052 if (pid == PgArchPID)
3053 {
3054 PgArchPID = 0;
3055 if (!EXIT_STATUS_0(exitstatus))
3056 LogChildExit(LOG, _("archiver process"),
3057 pid, exitstatus);
3058 if (PgArchStartupAllowed())
3059 PgArchPID = pgarch_start();
3060 continue;
3061 }
3062
3063 /*
3064 * Was it the statistics collector? If so, just try to start a new
3065 * one; no need to force reset of the rest of the system. (If fail,
3066 * we'll try again in future cycles of the main loop.)
3067 */
3068 if (pid == PgStatPID)
3069 {
3070 PgStatPID = 0;
3071 if (!EXIT_STATUS_0(exitstatus))
3072 LogChildExit(LOG, _("statistics collector process"),
3073 pid, exitstatus);
3074 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3075 PgStatPID = pgstat_start();
3076 continue;
3077 }
3078
3079 /* Was it the system logger? If so, try to start a new one */
3080 if (pid == SysLoggerPID)
3081 {
3082 SysLoggerPID = 0;
3083 /* for safety's sake, launch new logger *first* */
3084 SysLoggerPID = SysLogger_Start();
3085 if (!EXIT_STATUS_0(exitstatus))
3086 LogChildExit(LOG, _("system logger process"),
3087 pid, exitstatus);
3088 continue;
3089 }
3090
3091 /* Was it one of our background workers? */
3092 if (CleanupBackgroundWorker(pid, exitstatus))
3093 {
3094 /* have it be restarted */
3095 HaveCrashedWorker = true;
3096 continue;
3097 }
3098
3099 /*
3100 * Else do standard backend child cleanup.
3101 */
3102 CleanupBackend(pid, exitstatus);
3103 } /* loop over pending child-death reports */
3104
3105 /*
3106 * After cleaning out the SIGCHLD queue, see if we have any state changes
3107 * or actions to make.
3108 */
3109 PostmasterStateMachine();
3110
3111 /* Done with signal handler */
3112 #ifdef WIN32
3113 PG_SETMASK(&UnBlockSig);
3114 #endif
3115
3116 errno = save_errno;
3117 }
3118
3119 /*
3120 * Scan the bgworkers list and see if the given PID (which has just stopped
3121 * or crashed) is in it. Handle its shutdown if so, and return true. If not a
3122 * bgworker, return false.
3123 *
3124 * This is heavily based on CleanupBackend. One important difference is that
3125 * we don't know yet that the dying process is a bgworker, so we must be silent
3126 * until we're sure it is.
3127 */
3128 static bool
3129 CleanupBackgroundWorker(int pid,
3130 int exitstatus) /* child's exit status */
3131 {
3132 char namebuf[MAXPGPATH];
3133 slist_mutable_iter iter;
3134
3135 slist_foreach_modify(iter, &BackgroundWorkerList)
3136 {
3137 RegisteredBgWorker *rw;
3138
3139 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
3140
3141 if (rw->rw_pid != pid)
3142 continue;
3143
3144 #ifdef WIN32
3145 /* see CleanupBackend */
3146 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3147 exitstatus = 0;
3148 #endif
3149
3150 snprintf(namebuf, MAXPGPATH, _("background worker \"%s\""),
3151 rw->rw_worker.bgw_type);
3152
3153
3154 if (!EXIT_STATUS_0(exitstatus))
3155 {
3156 /* Record timestamp, so we know when to restart the worker. */
3157 rw->rw_crashed_at = GetCurrentTimestamp();
3158 }
3159 else
3160 {
3161 /* Zero exit status means terminate */
3162 rw->rw_crashed_at = 0;
3163 rw->rw_terminate = true;
3164 }
3165
3166 /*
3167 * Additionally, for shared-memory-connected workers, just like a
3168 * backend, any exit status other than 0 or 1 is considered a crash
3169 * and causes a system-wide restart.
3170 */
3171 if ((rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3172 {
3173 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3174 {
3175 HandleChildCrash(pid, exitstatus, namebuf);
3176 return true;
3177 }
3178 }
3179
3180 /*
3181 * We must release the postmaster child slot whether this worker is
3182 * connected to shared memory or not, but we only treat it as a crash
3183 * if it is in fact connected.
3184 */
3185 if (!ReleasePostmasterChildSlot(rw->rw_child_slot) &&
3186 (rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3187 {
3188 HandleChildCrash(pid, exitstatus, namebuf);
3189 return true;
3190 }
3191
3192 /* Get it out of the BackendList and clear out remaining data */
3193 dlist_delete(&rw->rw_backend->elem);
3194 #ifdef EXEC_BACKEND
3195 ShmemBackendArrayRemove(rw->rw_backend);
3196 #endif
3197
3198 /*
3199 * It's possible that this background worker started some OTHER
3200 * background worker and asked to be notified when that worker started
3201 * or stopped. If so, cancel any notifications destined for the
3202 * now-dead backend.
3203 */
3204 if (rw->rw_backend->bgworker_notify)
3205 BackgroundWorkerStopNotifications(rw->rw_pid);
3206 free(rw->rw_backend);
3207 rw->rw_backend = NULL;
3208 rw->rw_pid = 0;
3209 rw->rw_child_slot = 0;
3210 ReportBackgroundWorkerExit(&iter); /* report child death */
3211
3212 LogChildExit(EXIT_STATUS_0(exitstatus) ? DEBUG1 : LOG,
3213 namebuf, pid, exitstatus);
3214
3215 return true;
3216 }
3217
3218 return false;
3219 }
3220
3221 /*
3222 * CleanupBackend -- cleanup after terminated backend.
3223 *
3224 * Remove all local state associated with backend.
3225 *
3226 * If you change this, see also CleanupBackgroundWorker.
3227 */
3228 static void
3229 CleanupBackend(int pid,
3230 int exitstatus) /* child's exit status. */
3231 {
3232 dlist_mutable_iter iter;
3233
3234 LogChildExit(DEBUG2, _("server process"), pid, exitstatus);
3235
3236 /*
3237 * If a backend dies in an ugly way then we must signal all other backends
3238 * to quickdie. If exit status is zero (normal) or one (FATAL exit), we
3239 * assume everything is all right and proceed to remove the backend from
3240 * the active backend list.
3241 */
3242
3243 #ifdef WIN32
3244
3245 /*
3246 * On win32, also treat ERROR_WAIT_NO_CHILDREN (128) as nonfatal case,
3247 * since that sometimes happens under load when the process fails to start
3248 * properly (long before it starts using shared memory). Microsoft reports
3249 * it is related to mutex failure:
3250 * http://archives.postgresql.org/pgsql-hackers/2010-09/msg00790.php
3251 */
3252 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3253 {
3254 LogChildExit(LOG, _("server process"), pid, exitstatus);
3255 exitstatus = 0;
3256 }
3257 #endif
3258
3259 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3260 {
3261 HandleChildCrash(pid, exitstatus, _("server process"));
3262 return;
3263 }
3264
3265 dlist_foreach_modify(iter, &BackendList)
3266 {
3267 Backend *bp = dlist_container(Backend, elem, iter.cur);
3268
3269 if (bp->pid == pid)
3270 {
3271 if (!bp->dead_end)
3272 {
3273 if (!ReleasePostmasterChildSlot(bp->child_slot))
3274 {
3275 /*
3276 * Uh-oh, the child failed to clean itself up. Treat as a
3277 * crash after all.
3278 */
3279 HandleChildCrash(pid, exitstatus, _("server process"));
3280 return;
3281 }
3282 #ifdef EXEC_BACKEND
3283 ShmemBackendArrayRemove(bp);
3284 #endif
3285 }
3286 if (bp->bgworker_notify)
3287 {
3288 /*
3289 * This backend may have been slated to receive SIGUSR1 when
3290 * some background worker started or stopped. Cancel those
3291 * notifications, as we don't want to signal PIDs that are not
3292 * PostgreSQL backends. This gets skipped in the (probably
3293 * very common) case where the backend has never requested any
3294 * such notifications.
3295 */
3296 BackgroundWorkerStopNotifications(bp->pid);
3297 }
3298 dlist_delete(iter.cur);
3299 free(bp);
3300 break;
3301 }
3302 }
3303 }
3304
3305 /*
3306 * HandleChildCrash -- cleanup after failed backend, bgwriter, checkpointer,
3307 * walwriter, autovacuum, or background worker.
3308 *
3309 * The objectives here are to clean up our local state about the child
3310 * process, and to signal all other remaining children to quickdie.
3311 */
3312 static void
3313 HandleChildCrash(int pid, int exitstatus, const char *procname)
3314 {
3315 dlist_mutable_iter iter;
3316 slist_iter siter;
3317 Backend *bp;
3318 bool take_action;
3319
3320 /*
3321 * We only log messages and send signals if this is the first process
3322 * crash and we're not doing an immediate shutdown; otherwise, we're only
3323 * here to update postmaster's idea of live processes. If we have already
3324 * signalled children, nonzero exit status is to be expected, so don't
3325 * clutter log.
3326 */
3327 take_action = !FatalError && Shutdown != ImmediateShutdown;
3328
3329 if (take_action)
3330 {
3331 LogChildExit(LOG, procname, pid, exitstatus);
3332 ereport(LOG,
3333 (errmsg("terminating any other active server processes")));
3334 }
3335
3336 /* Process background workers. */
3337 slist_foreach(siter, &BackgroundWorkerList)
3338 {
3339 RegisteredBgWorker *rw;
3340
3341 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
3342 if (rw->rw_pid == 0)
3343 continue; /* not running */
3344 if (rw->rw_pid == pid)
3345 {
3346 /*
3347 * Found entry for freshly-dead worker, so remove it.
3348 */
3349 (void) ReleasePostmasterChildSlot(rw->rw_child_slot);
3350 dlist_delete(&rw->rw_backend->elem);
3351 #ifdef EXEC_BACKEND
3352 ShmemBackendArrayRemove(rw->rw_backend);
3353 #endif
3354 free(rw->rw_backend);
3355 rw->rw_backend = NULL;
3356 rw->rw_pid = 0;
3357 rw->rw_child_slot = 0;
3358 /* don't reset crashed_at */
3359 /* don't report child stop, either */
3360 /* Keep looping so we can signal remaining workers */
3361 }
3362 else
3363 {
3364 /*
3365 * This worker is still alive. Unless we did so already, tell it
3366 * to commit hara-kiri.
3367 *
3368 * SIGQUIT is the special signal that says exit without proc_exit
3369 * and let the user know what's going on. But if SendStop is set
3370 * (-s on command line), then we send SIGSTOP instead, so that we
3371 * can get core dumps from all backends by hand.
3372 */
3373 if (take_action)
3374 {
3375 ereport(DEBUG2,
3376 (errmsg_internal("sending %s to process %d",
3377 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3378 (int) rw->rw_pid)));
3379 signal_child(rw->rw_pid, (SendStop ? SIGSTOP : SIGQUIT));
3380 }
3381 }
3382 }
3383
3384 /* Process regular backends */
3385 dlist_foreach_modify(iter, &BackendList)
3386 {
3387 bp = dlist_container(Backend, elem, iter.cur);
3388
3389 if (bp->pid == pid)
3390 {
3391 /*
3392 * Found entry for freshly-dead backend, so remove it.
3393 */
3394 if (!bp->dead_end)
3395 {
3396 (void) ReleasePostmasterChildSlot(bp->child_slot);
3397 #ifdef EXEC_BACKEND
3398 ShmemBackendArrayRemove(bp);
3399 #endif
3400 }
3401 dlist_delete(iter.cur);
3402 free(bp);
3403 /* Keep looping so we can signal remaining backends */
3404 }
3405 else
3406 {
3407 /*
3408 * This backend is still alive. Unless we did so already, tell it
3409 * to commit hara-kiri.
3410 *
3411 * SIGQUIT is the special signal that says exit without proc_exit
3412 * and let the user know what's going on. But if SendStop is set
3413 * (-s on command line), then we send SIGSTOP instead, so that we
3414 * can get core dumps from all backends by hand.
3415 *
3416 * We could exclude dead_end children here, but at least in the
3417 * SIGSTOP case it seems better to include them.
3418 *
3419 * Background workers were already processed above; ignore them
3420 * here.
3421 */
3422 if (bp->bkend_type == BACKEND_TYPE_BGWORKER)
3423 continue;
3424
3425 if (take_action)
3426 {
3427 ereport(DEBUG2,
3428 (errmsg_internal("sending %s to process %d",
3429 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3430 (int) bp->pid)));
3431 signal_child(bp->pid, (SendStop ? SIGSTOP : SIGQUIT));
3432 }
3433 }
3434 }
3435
3436 /* Take care of the startup process too */
3437 if (pid == StartupPID)
3438 {
3439 StartupPID = 0;
3440 StartupStatus = STARTUP_CRASHED;
3441 }
3442 else if (StartupPID != 0 && take_action)
3443 {
3444 ereport(DEBUG2,
3445 (errmsg_internal("sending %s to process %d",
3446 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3447 (int) StartupPID)));
3448 signal_child(StartupPID, (SendStop ? SIGSTOP : SIGQUIT));
3449 StartupStatus = STARTUP_SIGNALED;
3450 }
3451
3452 /* Take care of the bgwriter too */
3453 if (pid == BgWriterPID)
3454 BgWriterPID = 0;
3455 else if (BgWriterPID != 0 && take_action)
3456 {
3457 ereport(DEBUG2,
3458 (errmsg_internal("sending %s to process %d",
3459 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3460 (int) BgWriterPID)));
3461 signal_child(BgWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3462 }
3463
3464 /* Take care of the checkpointer too */
3465 if (pid == CheckpointerPID)
3466 CheckpointerPID = 0;
3467 else if (CheckpointerPID != 0 && take_action)
3468 {
3469 ereport(DEBUG2,
3470 (errmsg_internal("sending %s to process %d",
3471 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3472 (int) CheckpointerPID)));
3473 signal_child(CheckpointerPID, (SendStop ? SIGSTOP : SIGQUIT));
3474 }
3475
3476 /* Take care of the walwriter too */
3477 if (pid == WalWriterPID)
3478 WalWriterPID = 0;
3479 else if (WalWriterPID != 0 && take_action)
3480 {
3481 ereport(DEBUG2,
3482 (errmsg_internal("sending %s to process %d",
3483 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3484 (int) WalWriterPID)));
3485 signal_child(WalWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3486 }
3487
3488 /* Take care of the walreceiver too */
3489 if (pid == WalReceiverPID)
3490 WalReceiverPID = 0;
3491 else if (WalReceiverPID != 0 && take_action)
3492 {
3493 ereport(DEBUG2,
3494 (errmsg_internal("sending %s to process %d",
3495 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3496 (int) WalReceiverPID)));
3497 signal_child(WalReceiverPID, (SendStop ? SIGSTOP : SIGQUIT));
3498 }
3499
3500 /* Take care of the autovacuum launcher too */
3501 if (pid == AutoVacPID)
3502 AutoVacPID = 0;
3503 else if (AutoVacPID != 0 && take_action)
3504 {
3505 ereport(DEBUG2,
3506 (errmsg_internal("sending %s to process %d",
3507 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3508 (int) AutoVacPID)));
3509 signal_child(AutoVacPID, (SendStop ? SIGSTOP : SIGQUIT));
3510 }
3511
3512 /*
3513 * Force a power-cycle of the pgarch process too. (This isn't absolutely
3514 * necessary, but it seems like a good idea for robustness, and it
3515 * simplifies the state-machine logic in the case where a shutdown request
3516 * arrives during crash processing.)
3517 */
3518 if (PgArchPID != 0 && take_action)
3519 {
3520 ereport(DEBUG2,
3521 (errmsg_internal("sending %s to process %d",
3522 "SIGQUIT",
3523 (int) PgArchPID)));
3524 signal_child(PgArchPID, SIGQUIT);
3525 }
3526
3527 /*
3528 * Force a power-cycle of the pgstat process too. (This isn't absolutely
3529 * necessary, but it seems like a good idea for robustness, and it
3530 * simplifies the state-machine logic in the case where a shutdown request
3531 * arrives during crash processing.)
3532 */
3533 if (PgStatPID != 0 && take_action)
3534 {
3535 ereport(DEBUG2,
3536 (errmsg_internal("sending %s to process %d",
3537 "SIGQUIT",
3538 (int) PgStatPID)));
3539 signal_child(PgStatPID, SIGQUIT);
3540 allow_immediate_pgstat_restart();
3541 }
3542
3543 /* We do NOT restart the syslogger */
3544
3545 if (Shutdown != ImmediateShutdown)
3546 FatalError = true;
3547
3548 /* We now transit into a state of waiting for children to die */
3549 if (pmState == PM_RECOVERY ||
3550 pmState == PM_HOT_STANDBY ||
3551 pmState == PM_RUN ||
3552 pmState == PM_STOP_BACKENDS ||
3553 pmState == PM_SHUTDOWN)
3554 pmState = PM_WAIT_BACKENDS;
3555
3556 /*
3557 * .. and if this doesn't happen quickly enough, now the clock is ticking
3558 * for us to kill them without mercy.
3559 */
3560 if (AbortStartTime == 0)
3561 AbortStartTime = time(NULL);
3562 }
3563
3564 /*
3565 * Log the death of a child process.
3566 */
3567 static void
3568 LogChildExit(int lev, const char *procname, int pid, int exitstatus)
3569 {
3570 /*
3571 * size of activity_buffer is arbitrary, but set equal to default
3572 * track_activity_query_size
3573 */
3574 char activity_buffer[1024];
3575 const char *activity = NULL;
3576
3577 if (!EXIT_STATUS_0(exitstatus))
3578 activity = pgstat_get_crashed_backend_activity(pid,
3579 activity_buffer,
3580 sizeof(activity_buffer));
3581
3582 if (WIFEXITED(exitstatus))
3583 ereport(lev,
3584
3585 /*------
3586 translator: %s is a noun phrase describing a child process, such as
3587 "server process" */
3588 (errmsg("%s (PID %d) exited with exit code %d",
3589 procname, pid, WEXITSTATUS(exitstatus)),
3590 activity ? errdetail("Failed process was running: %s", activity) : 0));
3591 else if (WIFSIGNALED(exitstatus))
3592 {
3593 #if defined(WIN32)
3594 ereport(lev,
3595
3596 /*------
3597 translator: %s is a noun phrase describing a child process, such as
3598 "server process" */
3599 (errmsg("%s (PID %d) was terminated by exception 0x%X",
3600 procname, pid, WTERMSIG(exitstatus)),
3601 errhint("See C include file \"ntstatus.h\" for a description of the hexadecimal value."),
3602 activity ? errdetail("Failed process was running: %s", activity) : 0));
3603 #else
3604 ereport(lev,
3605
3606 /*------
3607 translator: %s is a noun phrase describing a child process, such as
3608 "server process" */
3609 (errmsg("%s (PID %d) was terminated by signal %d: %s",
3610 procname, pid, WTERMSIG(exitstatus),
3611 pg_strsignal(WTERMSIG(exitstatus))),
3612 activity ? errdetail("Failed process was running: %s", activity) : 0));
3613 #endif
3614 }
3615 else
3616 ereport(lev,
3617
3618 /*------
3619 translator: %s is a noun phrase describing a child process, such as
3620 "server process" */
3621 (errmsg("%s (PID %d) exited with unrecognized status %d",
3622 procname, pid, exitstatus),
3623 activity ? errdetail("Failed process was running: %s", activity) : 0));
3624 }
3625
3626 /*
3627 * Advance the postmaster's state machine and take actions as appropriate
3628 *
3629 * This is common code for pmdie(), reaper() and sigusr1_handler(), which
3630 * receive the signals that might mean we need to change state.
3631 */
3632 static void
3633 PostmasterStateMachine(void)
3634 {
3635 /* If we're doing a smart shutdown, try to advance that state. */
3636 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3637 {
3638 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
3639 {
3640 /*
3641 * ALLOW_SUPERUSER_CONNS state ends as soon as online backup mode
3642 * is not active.
3643 */
3644 if (!BackupInProgress())
3645 connsAllowed = ALLOW_NO_CONNS;
3646 }
3647
3648 if (connsAllowed == ALLOW_NO_CONNS)
3649 {
3650 /*
3651 * ALLOW_NO_CONNS state ends when we have no normal client
3652 * backends running. Then we're ready to stop other children.
3653 */
3654 if (CountChildren(BACKEND_TYPE_NORMAL) == 0)
3655 pmState = PM_STOP_BACKENDS;
3656 }
3657 }
3658
3659 /*
3660 * If we're ready to do so, signal child processes to shut down. (This
3661 * isn't a persistent state, but treating it as a distinct pmState allows
3662 * us to share this code across multiple shutdown code paths.)
3663 */
3664 if (pmState == PM_STOP_BACKENDS)
3665 {
3666 /*
3667 * Forget any pending requests for background workers, since we're no
3668 * longer willing to launch any new workers. (If additional requests
3669 * arrive, BackgroundWorkerStateChange will reject them.)
3670 */
3671 ForgetUnstartedBackgroundWorkers();
3672
3673 /* Signal all backend children except walsenders */
3674 SignalSomeChildren(SIGTERM,
3675 BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND);
3676 /* and the autovac launcher too */
3677 if (AutoVacPID != 0)
3678 signal_child(AutoVacPID, SIGTERM);
3679 /* and the bgwriter too */
3680 if (BgWriterPID != 0)
3681 signal_child(BgWriterPID, SIGTERM);
3682 /* and the walwriter too */
3683 if (WalWriterPID != 0)
3684 signal_child(WalWriterPID, SIGTERM);
3685 /* If we're in recovery, also stop startup and walreceiver procs */
3686 if (StartupPID != 0)
3687 signal_child(StartupPID, SIGTERM);
3688 if (WalReceiverPID != 0)
3689 signal_child(WalReceiverPID, SIGTERM);
3690 /* checkpointer, archiver, stats, and syslogger may continue for now */
3691
3692 /* Now transition to PM_WAIT_BACKENDS state to wait for them to die */
3693 pmState = PM_WAIT_BACKENDS;
3694 }
3695
3696 /*
3697 * If we are in a state-machine state that implies waiting for backends to
3698 * exit, see if they're all gone, and change state if so.
3699 */
3700 if (pmState == PM_WAIT_BACKENDS)
3701 {
3702 /*
3703 * PM_WAIT_BACKENDS state ends when we have no regular backends
3704 * (including autovac workers), no bgworkers (including unconnected
3705 * ones), and no walwriter, autovac launcher or bgwriter. If we are
3706 * doing crash recovery or an immediate shutdown then we expect the
3707 * checkpointer to exit as well, otherwise not. The archiver, stats,
3708 * and syslogger processes are disregarded since they are not
3709 * connected to shared memory; we also disregard dead_end children
3710 * here. Walsenders are also disregarded, they will be terminated
3711 * later after writing the checkpoint record, like the archiver
3712 * process.
3713 */
3714 if (CountChildren(BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND) == 0 &&
3715 StartupPID == 0 &&
3716 WalReceiverPID == 0 &&
3717 BgWriterPID == 0 &&
3718 (CheckpointerPID == 0 ||
3719 (!FatalError && Shutdown < ImmediateShutdown)) &&
3720 WalWriterPID == 0 &&
3721 AutoVacPID == 0)
3722 {
3723 if (Shutdown >= ImmediateShutdown || FatalError)
3724 {
3725 /*
3726 * Start waiting for dead_end children to die. This state
3727 * change causes ServerLoop to stop creating new ones.
3728 */
3729 pmState = PM_WAIT_DEAD_END;
3730
3731 /*
3732 * We already SIGQUIT'd the archiver and stats processes, if
3733 * any, when we started immediate shutdown or entered
3734 * FatalError state.
3735 */
3736 }
3737 else
3738 {
3739 /*
3740 * If we get here, we are proceeding with normal shutdown. All
3741 * the regular children are gone, and it's time to tell the
3742 * checkpointer to do a shutdown checkpoint.
3743 */
3744 Assert(Shutdown > NoShutdown);
3745 /* Start the checkpointer if not running */
3746 if (CheckpointerPID == 0)
3747 CheckpointerPID = StartCheckpointer();
3748 /* And tell it to shut down */
3749 if (CheckpointerPID != 0)
3750 {
3751 signal_child(CheckpointerPID, SIGUSR2);
3752 pmState = PM_SHUTDOWN;
3753 }
3754 else
3755 {
3756 /*
3757 * If we failed to fork a checkpointer, just shut down.
3758 * Any required cleanup will happen at next restart. We
3759 * set FatalError so that an "abnormal shutdown" message
3760 * gets logged when we exit.
3761 */
3762 FatalError = true;
3763 pmState = PM_WAIT_DEAD_END;
3764
3765 /* Kill the walsenders, archiver and stats collector too */
3766 SignalChildren(SIGQUIT);
3767 if (PgArchPID != 0)
3768 signal_child(PgArchPID, SIGQUIT);
3769 if (PgStatPID != 0)
3770 signal_child(PgStatPID, SIGQUIT);
3771 }
3772 }
3773 }
3774 }
3775
3776 if (pmState == PM_SHUTDOWN_2)
3777 {
3778 /*
3779 * PM_SHUTDOWN_2 state ends when there's no other children than
3780 * dead_end children left. There shouldn't be any regular backends
3781 * left by now anyway; what we're really waiting for is walsenders and
3782 * archiver.
3783 */
3784 if (PgArchPID == 0 && CountChildren(BACKEND_TYPE_ALL) == 0)
3785 {
3786 pmState = PM_WAIT_DEAD_END;
3787 }
3788 }
3789
3790 if (pmState == PM_WAIT_DEAD_END)
3791 {
3792 /*
3793 * PM_WAIT_DEAD_END state ends when the BackendList is entirely empty
3794 * (ie, no dead_end children remain), and the archiver and stats
3795 * collector are gone too.
3796 *
3797 * The reason we wait for those two is to protect them against a new
3798 * postmaster starting conflicting subprocesses; this isn't an
3799 * ironclad protection, but it at least helps in the
3800 * shutdown-and-immediately-restart scenario. Note that they have
3801 * already been sent appropriate shutdown signals, either during a
3802 * normal state transition leading up to PM_WAIT_DEAD_END, or during
3803 * FatalError processing.
3804 */
3805 if (dlist_is_empty(&BackendList) &&
3806 PgArchPID == 0 && PgStatPID == 0)
3807 {
3808 /* These other guys should be dead already */
3809 Assert(StartupPID == 0);
3810 Assert(WalReceiverPID == 0);
3811 Assert(BgWriterPID == 0);
3812 Assert(CheckpointerPID == 0);
3813 Assert(WalWriterPID == 0);
3814 Assert(AutoVacPID == 0);
3815 /* syslogger is not considered here */
3816 pmState = PM_NO_CHILDREN;
3817 }
3818 }
3819
3820 /*
3821 * If we've been told to shut down, we exit as soon as there are no
3822 * remaining children. If there was a crash, cleanup will occur at the
3823 * next startup. (Before PostgreSQL 8.3, we tried to recover from the
3824 * crash before exiting, but that seems unwise if we are quitting because
3825 * we got SIGTERM from init --- there may well not be time for recovery
3826 * before init decides to SIGKILL us.)
3827 *
3828 * Note that the syslogger continues to run. It will exit when it sees
3829 * EOF on its input pipe, which happens when there are no more upstream
3830 * processes.
3831 */
3832 if (Shutdown > NoShutdown && pmState == PM_NO_CHILDREN)
3833 {
3834 if (FatalError)
3835 {
3836 ereport(LOG, (errmsg("abnormal database system shutdown")));
3837 ExitPostmaster(1);
3838 }
3839 else
3840 {
3841 /*
3842 * Terminate exclusive backup mode to avoid recovery after a clean
3843 * fast shutdown. Since an exclusive backup can only be taken
3844 * during normal running (and not, for example, while running
3845 * under Hot Standby) it only makes sense to do this if we reached
3846 * normal running. If we're still in recovery, the backup file is
3847 * one we're recovering *from*, and we must keep it around so that
3848 * recovery restarts from the right place.
3849 */
3850 if (ReachedNormalRunning)
3851 CancelBackup();
3852
3853 /* Normal exit from the postmaster is here */
3854 ExitPostmaster(0);
3855 }
3856 }
3857
3858 /*
3859 * If the startup process failed, or the user does not want an automatic
3860 * restart after backend crashes, wait for all non-syslogger children to
3861 * exit, and then exit postmaster. We don't try to reinitialize when the
3862 * startup process fails, because more than likely it will just fail again
3863 * and we will keep trying forever.
3864 */
3865 if (pmState == PM_NO_CHILDREN &&
3866 (StartupStatus == STARTUP_CRASHED || !restart_after_crash))
3867 ExitPostmaster(1);
3868
3869 /*
3870 * If we need to recover from a crash, wait for all non-syslogger children
3871 * to exit, then reset shmem and StartupDataBase.
3872 */
3873 if (FatalError && pmState == PM_NO_CHILDREN)
3874 {
3875 ereport(LOG,
3876 (errmsg("all server processes terminated; reinitializing")));
3877
3878 /* allow background workers to immediately restart */
3879 ResetBackgroundWorkerCrashTimes();
3880
3881 shmem_exit(1);
3882
3883 /* re-read control file into local memory */
3884 LocalProcessControlFile(true);
3885
3886 reset_shared(PostPortNumber);
3887
3888 StartupPID = StartupDataBase();
3889 Assert(StartupPID != 0);
3890 StartupStatus = STARTUP_RUNNING;
3891 pmState = PM_STARTUP;
3892 /* crash recovery started, reset SIGKILL flag */
3893 AbortStartTime = 0;
3894 }
3895 }
3896
3897
3898 /*
3899 * Send a signal to a postmaster child process
3900 *
3901 * On systems that have setsid(), each child process sets itself up as a
3902 * process group leader. For signals that are generally interpreted in the
3903 * appropriate fashion, we signal the entire process group not just the
3904 * direct child process. This allows us to, for example, SIGQUIT a blocked
3905 * archive_recovery script, or SIGINT a script being run by a backend via
3906 * system().
3907 *
3908 * There is a race condition for recently-forked children: they might not
3909 * have executed setsid() yet. So we signal the child directly as well as
3910 * the group. We assume such a child will handle the signal before trying
3911 * to spawn any grandchild processes. We also assume that signaling the
3912 * child twice will not cause any problems.
3913 */
3914 static void
3915 signal_child(pid_t pid, int signal)
3916 {
3917 if (kill(pid, signal) < 0)
3918 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) pid, signal);
3919 #ifdef HAVE_SETSID
3920 switch (signal)
3921 {
3922 case SIGINT:
3923 case SIGTERM:
3924 case SIGQUIT:
3925 case SIGSTOP:
3926 case SIGKILL:
3927 if (kill(-pid, signal) < 0)
3928 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) (-pid), signal);
3929 break;
3930 default:
3931 break;
3932 }
3933 #endif
3934 }
3935
3936 /*
3937 * Send a signal to the targeted children (but NOT special children;
3938 * dead_end children are never signaled, either).
3939 */
3940 static bool
3941 SignalSomeChildren(int signal, int target)
3942 {
3943 dlist_iter iter;
3944 bool signaled = false;
3945
3946 dlist_foreach(iter, &BackendList)
3947 {
3948 Backend *bp = dlist_container(Backend, elem, iter.cur);
3949
3950 if (bp->dead_end)
3951 continue;
3952
3953 /*
3954 * Since target == BACKEND_TYPE_ALL is the most common case, we test
3955 * it first and avoid touching shared memory for every child.
3956 */
3957 if (target != BACKEND_TYPE_ALL)
3958 {
3959 /*
3960 * Assign bkend_type for any recently announced WAL Sender
3961 * processes.
3962 */
3963 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
3964 IsPostmasterChildWalSender(bp->child_slot))
3965 bp->bkend_type = BACKEND_TYPE_WALSND;
3966
3967 if (!(target & bp->bkend_type))
3968 continue;
3969 }
3970
3971 ereport(DEBUG4,
3972 (errmsg_internal("sending signal %d to process %d",
3973 signal, (int) bp->pid)));
3974 signal_child(bp->pid, signal);
3975 signaled = true;
3976 }
3977 return signaled;
3978 }
3979
3980 /*
3981 * Send a termination signal to children. This considers all of our children
3982 * processes, except syslogger and dead_end backends.
3983 */
3984 static void
3985 TerminateChildren(int signal)
3986 {
3987 SignalChildren(signal);
3988 if (StartupPID != 0)
3989 {
3990 signal_child(StartupPID, signal);
3991 if (signal == SIGQUIT || signal == SIGKILL)
3992 StartupStatus = STARTUP_SIGNALED;
3993 }
3994 if (BgWriterPID != 0)
3995 signal_child(BgWriterPID, signal);
3996 if (CheckpointerPID != 0)
3997 signal_child(CheckpointerPID, signal);
3998 if (WalWriterPID != 0)
3999 signal_child(WalWriterPID, signal);
4000 if (WalReceiverPID != 0)
4001 signal_child(WalReceiverPID, signal);
4002 if (AutoVacPID != 0)
4003 signal_child(AutoVacPID, signal);
4004 if (PgArchPID != 0)
4005 signal_child(PgArchPID, signal);
4006 if (PgStatPID != 0)
4007 signal_child(PgStatPID, signal);
4008 }
4009
4010 /*
4011 * BackendStartup -- start backend process
4012 *
4013 * returns: STATUS_ERROR if the fork failed, STATUS_OK otherwise.
4014 *
4015 * Note: if you change this code, also consider StartAutovacuumWorker.
4016 */
4017 static int
4018 BackendStartup(Port *port)
4019 {
4020 Backend *bn; /* for backend cleanup */
4021 pid_t pid;
4022
4023 /*
4024 * Create backend data structure. Better before the fork() so we can
4025 * handle failure cleanly.
4026 */
4027 bn = (Backend *) malloc(sizeof(Backend));
4028 if (!bn)
4029 {
4030 ereport(LOG,
4031 (errcode(ERRCODE_OUT_OF_MEMORY),
4032 errmsg("out of memory")));
4033 return STATUS_ERROR;
4034 }
4035
4036 /*
4037 * Compute the cancel key that will be assigned to this backend. The
4038 * backend will have its own copy in the forked-off process' value of
4039 * MyCancelKey, so that it can transmit the key to the frontend.
4040 */
4041 if (!RandomCancelKey(&MyCancelKey))
4042 {
4043 free(bn);
4044 ereport(LOG,
4045 (errcode(ERRCODE_INTERNAL_ERROR),
4046 errmsg("could not generate random cancel key")));
4047 return STATUS_ERROR;
4048 }
4049
4050 bn->cancel_key = MyCancelKey;
4051
4052 /* Pass down canAcceptConnections state */
4053 port->canAcceptConnections = canAcceptConnections(BACKEND_TYPE_NORMAL);
4054 bn->dead_end = (port->canAcceptConnections != CAC_OK &&
4055 port->canAcceptConnections != CAC_SUPERUSER);
4056
4057 /*
4058 * Unless it's a dead_end child, assign it a child slot number
4059 */
4060 if (!bn->dead_end)
4061 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
4062 else
4063 bn->child_slot = 0;
4064
4065 /* Hasn't asked to be notified about any bgworkers yet */
4066 bn->bgworker_notify = false;
4067
4068 #ifdef EXEC_BACKEND
4069 pid = backend_forkexec(port);
4070 #else /* !EXEC_BACKEND */
4071 pid = fork_process();
4072 if (pid == 0) /* child */
4073 {
4074 free(bn);
4075
4076 /* Detangle from postmaster */
4077 InitPostmasterChild();
4078
4079 /* Close the postmaster's sockets */
4080 ClosePostmasterPorts(false);
4081
4082 /* Perform additional initialization and collect startup packet */
4083 BackendInitialize(port);
4084
4085 /* And run the backend */
4086 BackendRun(port);
4087 }
4088 #endif /* EXEC_BACKEND */
4089
4090 if (pid < 0)
4091 {
4092 /* in parent, fork failed */
4093 int save_errno = errno;
4094
4095 if (!bn->dead_end)
4096 (void) ReleasePostmasterChildSlot(bn->child_slot);
4097 free(bn);
4098 errno = save_errno;
4099 ereport(LOG,
4100 (errmsg("could not fork new process for connection: %m")));
4101 report_fork_failure_to_client(port, save_errno);
4102 return STATUS_ERROR;
4103 }
4104
4105 /* in parent, successful fork */
4106 ereport(DEBUG2,
4107 (errmsg_internal("forked new backend, pid=%d socket=%d",
4108 (int) pid, (int) port->sock)));
4109
4110 /*
4111 * Everything's been successful, it's safe to add this backend to our list
4112 * of backends.
4113 */
4114 bn->pid = pid;
4115 bn->bkend_type = BACKEND_TYPE_NORMAL; /* Can change later to WALSND */
4116 dlist_push_head(&BackendList, &bn->elem);
4117
4118 #ifdef EXEC_BACKEND
4119 if (!bn->dead_end)
4120 ShmemBackendArrayAdd(bn);
4121 #endif
4122
4123 return STATUS_OK;
4124 }
4125
4126 /*
4127 * Try to report backend fork() failure to client before we close the
4128 * connection. Since we do not care to risk blocking the postmaster on
4129 * this connection, we set the connection to non-blocking and try only once.
4130 *
4131 * This is grungy special-purpose code; we cannot use backend libpq since
4132 * it's not up and running.
4133 */
4134 static void
4135 report_fork_failure_to_client(Port *port, int errnum)
4136 {
4137 char buffer[1000];
4138 int rc;
4139
4140 /* Format the error message packet (always V2 protocol) */
4141 snprintf(buffer, sizeof(buffer), "E%s%s\n",
4142 _("could not fork new process for connection: "),
4143 strerror(errnum));
4144
4145 /* Set port to non-blocking. Don't do send() if this fails */
4146 if (!pg_set_noblock(port->sock))
4147 return;
4148
4149 /* We'll retry after EINTR, but ignore all other failures */
4150 do
4151 {
4152 rc = send(port->sock, buffer, strlen(buffer) + 1, 0);
4153 } while (rc < 0 && errno == EINTR);
4154 }
4155
4156
4157 /*
4158 * BackendInitialize -- initialize an interactive (postmaster-child)
4159 * backend process, and collect the client's startup packet.
4160 *
4161 * returns: nothing. Will not return at all if there's any failure.
4162 *
4163 * Note: this code does not depend on having any access to shared memory.
4164 * In the EXEC_BACKEND case, we are physically attached to shared memory
4165 * but have not yet set up most of our local pointers to shmem structures.
4166 */
4167 static void
4168 BackendInitialize(Port *port)
4169 {
4170 int status;
4171 int ret;
4172 char remote_host[NI_MAXHOST];
4173 char remote_port[NI_MAXSERV];
4174 char remote_ps_data[NI_MAXHOST];
4175
4176 /* Save port etc. for ps status */
4177 MyProcPort = port;
4178
4179 /*
4180 * PreAuthDelay is a debugging aid for investigating problems in the
4181 * authentication cycle: it can be set in postgresql.conf to allow time to
4182 * attach to the newly-forked backend with a debugger. (See also
4183 * PostAuthDelay, which we allow clients to pass through PGOPTIONS, but it
4184 * is not honored until after authentication.)
4185 */
4186 if (PreAuthDelay > 0)
4187 pg_usleep(PreAuthDelay * 1000000L);
4188
4189 /* This flag will remain set until InitPostgres finishes authentication */
4190 ClientAuthInProgress = true; /* limit visibility of log messages */
4191
4192 /* save process start time */
4193 port->SessionStartTime = GetCurrentTimestamp();
4194 MyStartTime = timestamptz_to_time_t(port->SessionStartTime);
4195
4196 /* set these to empty in case they are needed before we set them up */
4197 port->remote_host = "";
4198 port->remote_port = "";
4199
4200 /*
4201 * Initialize libpq and enable reporting of ereport errors to the client.
4202 * Must do this now because authentication uses libpq to send messages.
4203 */
4204 pq_init(); /* initialize libpq to talk to client */
4205 whereToSendOutput = DestRemote; /* now safe to ereport to client */
4206
4207 /*
4208 * We arrange to do proc_exit(1) if we receive SIGTERM or timeout while
4209 * trying to collect the startup packet; while SIGQUIT results in
4210 * _exit(2). Otherwise the postmaster cannot shutdown the database FAST
4211 * or IMMED cleanly if a buggy client fails to send the packet promptly.
4212 *
4213 * XXX this is pretty dangerous; signal handlers should not call anything
4214 * as complex as proc_exit() directly. We minimize the hazard by not
4215 * keeping these handlers active for longer than we must. However, it
4216 * seems necessary to be able to escape out of DNS lookups as well as the
4217 * startup packet reception proper, so we can't narrow the scope further
4218 * than is done here.
4219 *
4220 * XXX it follows that the remainder of this function must tolerate losing
4221 * control at any instant. Likewise, any pg_on_exit_callback registered
4222 * before or during this function must be prepared to execute at any
4223 * instant between here and the end of this function. Furthermore,
4224 * affected callbacks execute partially or not at all when a second
4225 * exit-inducing signal arrives after proc_exit_prepare() decrements
4226 * on_proc_exit_index. (Thanks to that mechanic, callbacks need not
4227 * anticipate more than one call.) This is fragile; it ought to instead
4228 * follow the norm of handling interrupts at selected, safe opportunities.
4229 */
4230 pqsignal(SIGTERM, process_startup_packet_die);
4231 pqsignal(SIGQUIT, process_startup_packet_quickdie);
4232 InitializeTimeouts(); /* establishes SIGALRM handler */
4233 PG_SETMASK(&StartupBlockSig);
4234
4235 /*
4236 * Get the remote host name and port for logging and status display.
4237 */
4238 remote_host[0] = '\0';
4239 remote_port[0] = '\0';
4240 if ((ret = pg_getnameinfo_all(&port->raddr.addr, port->raddr.salen,
4241 remote_host, sizeof(remote_host),
4242 remote_port, sizeof(remote_port),
4243 (log_hostname ? 0 : NI_NUMERICHOST) | NI_NUMERICSERV)) != 0)
4244 ereport(WARNING,
4245 (errmsg_internal("pg_getnameinfo_all() failed: %s",
4246 gai_strerror(ret))));
4247 if (remote_port[0] == '\0')
4248 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s", remote_host);
4249 else
4250 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s(%s)", remote_host, remote_port);
4251
4252 /*
4253 * Save remote_host and remote_port in port structure (after this, they
4254 * will appear in log_line_prefix data for log messages).
4255 */
4256 port->remote_host = strdup(remote_host);
4257 port->remote_port = strdup(remote_port);
4258
4259 /* And now we can issue the Log_connections message, if wanted */
4260 if (Log_connections)
4261 {
4262 if (remote_port[0])
4263 ereport(LOG,
4264 (errmsg("connection received: host=%s port=%s",
4265 remote_host,
4266 remote_port)));
4267 else
4268 ereport(LOG,
4269 (errmsg("connection received: host=%s",
4270 remote_host)));
4271 }
4272
4273 /*
4274 * If we did a reverse lookup to name, we might as well save the results
4275 * rather than possibly repeating the lookup during authentication.
4276 *
4277 * Note that we don't want to specify NI_NAMEREQD above, because then we'd
4278 * get nothing useful for a client without an rDNS entry. Therefore, we
4279 * must check whether we got a numeric IPv4 or IPv6 address, and not save
4280 * it into remote_hostname if so. (This test is conservative and might
4281 * sometimes classify a hostname as numeric, but an error in that
4282 * direction is safe; it only results in a possible extra lookup.)
4283 */
4284 if (log_hostname &&
4285 ret == 0 &&
4286 strspn(remote_host, "0123456789.") < strlen(remote_host) &&
4287 strspn(remote_host, "0123456789ABCDEFabcdef:") < strlen(remote_host))
4288 port->remote_hostname = strdup(remote_host);
4289
4290 /*
4291 * Ready to begin client interaction. We will give up and proc_exit(1)
4292 * after a time delay, so that a broken client can't hog a connection
4293 * indefinitely. PreAuthDelay and any DNS interactions above don't count
4294 * against the time limit.
4295 *
4296 * Note: AuthenticationTimeout is applied here while waiting for the
4297 * startup packet, and then again in InitPostgres for the duration of any
4298 * authentication operations. So a hostile client could tie up the
4299 * process for nearly twice AuthenticationTimeout before we kick him off.
4300 *
4301 * Note: because PostgresMain will call InitializeTimeouts again, the
4302 * registration of STARTUP_PACKET_TIMEOUT will be lost. This is okay
4303 * since we never use it again after this function.
4304 */
4305 RegisterTimeout(STARTUP_PACKET_TIMEOUT, StartupPacketTimeoutHandler);
4306 enable_timeout_after(STARTUP_PACKET_TIMEOUT, AuthenticationTimeout * 1000);
4307
4308 /*
4309 * Receive the startup packet (which might turn out to be a cancel request
4310 * packet).
4311 */
4312 status = ProcessStartupPacket(port, false);
4313
4314 /*
4315 * Disable the timeout, and prevent SIGTERM/SIGQUIT again.
4316 */
4317 disable_timeout(STARTUP_PACKET_TIMEOUT, false);
4318 PG_SETMASK(&BlockSig);
4319
4320 /*
4321 * Stop here if it was bad or a cancel packet. ProcessStartupPacket
4322 * already did any appropriate error reporting.
4323 */
4324 if (status != STATUS_OK)
4325 proc_exit(0);
4326
4327 /*
4328 * Now that we have the user and database name, we can set the process
4329 * title for ps. It's good to do this as early as possible in startup.
4330 *
4331 * For a walsender, the ps display is set in the following form:
4332 *
4333 * postgres: walsender <user> <host> <activity>
4334 *
4335 * To achieve that, we pass "walsender" as username and username as dbname
4336 * to init_ps_display(). XXX: should add a new variant of
4337 * init_ps_display() to avoid abusing the parameters like this.
4338 */
4339 if (am_walsender)
4340 init_ps_display(pgstat_get_backend_desc(B_WAL_SENDER), port->user_name, remote_ps_data,
4341 update_process_title ? "authentication" : "");
4342 else
4343 init_ps_display(port->user_name, port->database_name, remote_ps_data,
4344 update_process_title ? "authentication" : "");
4345 }
4346
4347
4348 /*
4349 * BackendRun -- set up the backend's argument list and invoke PostgresMain()
4350 *
4351 * returns:
4352 * Shouldn't return at all.
4353 * If PostgresMain() fails, return status.
4354 */
4355 static void
4356 BackendRun(Port *port)
4357 {
4358 char **av;
4359 int maxac;
4360 int ac;
4361 long secs;
4362 int usecs;
4363 int i;
4364
4365 /*
4366 * Don't want backend to be able to see the postmaster random number
4367 * generator state. We have to clobber the static random_seed *and* start
4368 * a new random sequence in the random() library function.
4369 */
4370 #ifndef HAVE_STRONG_RANDOM
4371 random_seed = 0;
4372 random_start_time.tv_usec = 0;
4373 #endif
4374 /* slightly hacky way to convert timestamptz into integers */
4375 TimestampDifference(0, port->SessionStartTime, &secs, &usecs);
4376 srandom((unsigned int) (MyProcPid ^ (usecs << 12) ^ secs));
4377
4378 /*
4379 * Now, build the argv vector that will be given to PostgresMain.
4380 *
4381 * The maximum possible number of commandline arguments that could come
4382 * from ExtraOptions is (strlen(ExtraOptions) + 1) / 2; see
4383 * pg_split_opts().
4384 */
4385 maxac = 2; /* for fixed args supplied below */
4386 maxac += (strlen(ExtraOptions) + 1) / 2;
4387
4388 av = (char **) MemoryContextAlloc(TopMemoryContext,
4389 maxac * sizeof(char *));
4390 ac = 0;
4391
4392 av[ac++] = "postgres";
4393
4394 /*
4395 * Pass any backend switches specified with -o on the postmaster's own
4396 * command line. We assume these are secure.
4397 */
4398 pg_split_opts(av, &ac, ExtraOptions);
4399
4400 av[ac] = NULL;
4401
4402 Assert(ac < maxac);
4403
4404 /*
4405 * Debug: print arguments being passed to backend
4406 */
4407 ereport(DEBUG3,
4408 (errmsg_internal("%s child[%d]: starting with (",
4409 progname, (int) getpid())));
4410 for (i = 0; i < ac; ++i)
4411 ereport(DEBUG3,
4412 (errmsg_internal("\t%s", av[i])));
4413 ereport(DEBUG3,
4414 (errmsg_internal(")")));
4415
4416 /*
4417 * Make sure we aren't in PostmasterContext anymore. (We can't delete it
4418 * just yet, though, because InitPostgres will need the HBA data.)
4419 */
4420 MemoryContextSwitchTo(TopMemoryContext);
4421
4422 PostgresMain(ac, av, port->database_name, port->user_name);
4423 }
4424
4425
4426 #ifdef EXEC_BACKEND
4427
4428 /*
4429 * postmaster_forkexec -- fork and exec a postmaster subprocess
4430 *
4431 * The caller must have set up the argv array already, except for argv[2]
4432 * which will be filled with the name of the temp variable file.
4433 *
4434 * Returns the child process PID, or -1 on fork failure (a suitable error
4435 * message has been logged on failure).
4436 *
4437 * All uses of this routine will dispatch to SubPostmasterMain in the
4438 * child process.
4439 */
4440 pid_t
4441 postmaster_forkexec(int argc, char *argv[])
4442 {
4443 Port port;
4444
4445 /* This entry point passes dummy values for the Port variables */
4446 memset(&port, 0, sizeof(port));
4447 return internal_forkexec(argc, argv, &port);
4448 }
4449
4450 /*
4451 * backend_forkexec -- fork/exec off a backend process
4452 *
4453 * Some operating systems (WIN32) don't have fork() so we have to simulate
4454 * it by storing parameters that need to be passed to the child and
4455 * then create a new child process.
4456 *
4457 * returns the pid of the fork/exec'd process, or -1 on failure
4458 */
4459 static pid_t
4460 backend_forkexec(Port *port)
4461 {
4462 char *av[4];
4463 int ac = 0;
4464
4465 av[ac++] = "postgres";
4466 av[ac++] = "--forkbackend";
4467 av[ac++] = NULL; /* filled in by internal_forkexec */
4468
4469 av[ac] = NULL;
4470 Assert(ac < lengthof(av));
4471
4472 return internal_forkexec(ac, av, port);
4473 }
4474
4475 #ifndef WIN32
4476
4477 /*
4478 * internal_forkexec non-win32 implementation
4479 *
4480 * - writes out backend variables to the parameter file
4481 * - fork():s, and then exec():s the child process
4482 */
4483 static pid_t
4484 internal_forkexec(int argc, char *argv[], Port *port)
4485 {
4486 static unsigned long tmpBackendFileNum = 0;
4487 pid_t pid;
4488 char tmpfilename[MAXPGPATH];
4489 BackendParameters param;
4490 FILE *fp;
4491
4492 if (!save_backend_variables(¶m, port))
4493 return -1; /* log made by save_backend_variables */
4494
4495 /* Calculate name for temp file */
4496 snprintf(tmpfilename, MAXPGPATH, "%s/%s.backend_var.%d.%lu",
4497 PG_TEMP_FILES_DIR, PG_TEMP_FILE_PREFIX,
4498 MyProcPid, ++tmpBackendFileNum);
4499
4500 /* Open file */
4501 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4502 if (!fp)
4503 {
4504 /*
4505 * As in OpenTemporaryFileInTablespace, try to make the temp-file
4506 * directory, ignoring errors.
4507 */
4508 (void) MakePGDirectory(PG_TEMP_FILES_DIR);
4509
4510 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4511 if (!fp)
4512 {
4513 ereport(LOG,
4514 (errcode_for_file_access(),
4515 errmsg("could not create file \"%s\": %m",
4516 tmpfilename)));
4517 return -1;
4518 }
4519 }
4520
4521 if (fwrite(¶m, sizeof(param), 1, fp) != 1)
4522 {
4523 ereport(LOG,
4524 (errcode_for_file_access(),
4525 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4526 FreeFile(fp);
4527 return -1;
4528 }
4529
4530 /* Release file */
4531 if (FreeFile(fp))
4532 {
4533 ereport(LOG,
4534 (errcode_for_file_access(),
4535 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4536 return -1;
4537 }
4538
4539 /* Make sure caller set up argv properly */
4540 Assert(argc >= 3);
4541 Assert(argv[argc] == NULL);
4542 Assert(strncmp(argv[1], "--fork", 6) == 0);
4543 Assert(argv[2] == NULL);
4544
4545 /* Insert temp file name after --fork argument */
4546 argv[2] = tmpfilename;
4547
4548 /* Fire off execv in child */
4549 if ((pid = fork_process()) == 0)
4550 {
4551 if (execv(postgres_exec_path, argv) < 0)
4552 {
4553 ereport(LOG,
4554 (errmsg("could not execute server process \"%s\": %m",
4555 postgres_exec_path)));
4556 /* We're already in the child process here, can't return */
4557 exit(1);
4558 }
4559 }
4560
4561 return pid; /* Parent returns pid, or -1 on fork failure */
4562 }
4563 #else /* WIN32 */
4564
4565 /*
4566 * internal_forkexec win32 implementation
4567 *
4568 * - starts backend using CreateProcess(), in suspended state
4569 * - writes out backend variables to the parameter file
4570 * - during this, duplicates handles and sockets required for
4571 * inheritance into the new process
4572 * - resumes execution of the new process once the backend parameter
4573 * file is complete.
4574 */
4575 static pid_t
4576 internal_forkexec(int argc, char *argv[], Port *port)
4577 {
4578 int retry_count = 0;
4579 STARTUPINFO si;
4580 PROCESS_INFORMATION pi;
4581 int i;
4582 int j;
4583 char cmdLine[MAXPGPATH * 2];
4584 HANDLE paramHandle;
4585 BackendParameters *param;
4586 SECURITY_ATTRIBUTES sa;
4587 char paramHandleStr[32];
4588 win32_deadchild_waitinfo *childinfo;
4589
4590 /* Make sure caller set up argv properly */
4591 Assert(argc >= 3);
4592 Assert(argv[argc] == NULL);
4593 Assert(strncmp(argv[1], "--fork", 6) == 0);
4594 Assert(argv[2] == NULL);
4595
4596 /* Resume here if we need to retry */
4597 retry:
4598
4599 /* Set up shared memory for parameter passing */
4600 ZeroMemory(&sa, sizeof(sa));
4601 sa.nLength = sizeof(sa);
4602 sa.bInheritHandle = TRUE;
4603 paramHandle = CreateFileMapping(INVALID_HANDLE_VALUE,
4604 &sa,
4605 PAGE_READWRITE,
4606 0,
4607 sizeof(BackendParameters),
4608 NULL);
4609 if (paramHandle == INVALID_HANDLE_VALUE)
4610 {
4611 elog(LOG, "could not create backend parameter file mapping: error code %lu",
4612 GetLastError());
4613 return -1;
4614 }
4615
4616 param = MapViewOfFile(paramHandle, FILE_MAP_WRITE, 0, 0, sizeof(BackendParameters));
4617 if (!param)
4618 {
4619 elog(LOG, "could not map backend parameter memory: error code %lu",
4620 GetLastError());
4621 CloseHandle(paramHandle);
4622 return -1;
4623 }
4624
4625 /* Insert temp file name after --fork argument */
4626 #ifdef _WIN64
4627 sprintf(paramHandleStr, "%llu", (LONG_PTR) paramHandle);
4628 #else
4629 sprintf(paramHandleStr, "%lu", (DWORD) paramHandle);
4630 #endif
4631 argv[2] = paramHandleStr;
4632
4633 /* Format the cmd line */
4634 cmdLine[sizeof(cmdLine) - 1] = '\0';
4635 cmdLine[sizeof(cmdLine) - 2] = '\0';
4636 snprintf(cmdLine, sizeof(cmdLine) - 1, "\"%s\"", postgres_exec_path);
4637 i = 0;
4638 while (argv[++i] != NULL)
4639 {
4640 j = strlen(cmdLine);
4641 snprintf(cmdLine + j, sizeof(cmdLine) - 1 - j, " \"%s\"", argv[i]);
4642 }
4643 if (cmdLine[sizeof(cmdLine) - 2] != '\0')
4644 {
4645 elog(LOG, "subprocess command line too long");
4646 return -1;
4647 }
4648
4649 memset(&pi, 0, sizeof(pi));
4650 memset(&si, 0, sizeof(si));
4651 si.cb = sizeof(si);
4652
4653 /*
4654 * Create the subprocess in a suspended state. This will be resumed later,
4655 * once we have written out the parameter file.
4656 */
4657 if (!CreateProcess(NULL, cmdLine, NULL, NULL, TRUE, CREATE_SUSPENDED,
4658 NULL, NULL, &si, &pi))
4659 {
4660 elog(LOG, "CreateProcess call failed: %m (error code %lu)",
4661 GetLastError());
4662 return -1;
4663 }
4664
4665 if (!save_backend_variables(param, port, pi.hProcess, pi.dwProcessId))
4666 {
4667 /*
4668 * log made by save_backend_variables, but we have to clean up the
4669 * mess with the half-started process
4670 */
4671 if (!TerminateProcess(pi.hProcess, 255))
4672 ereport(LOG,
4673 (errmsg_internal("could not terminate unstarted process: error code %lu",
4674 GetLastError())));
4675 CloseHandle(pi.hProcess);
4676 CloseHandle(pi.hThread);
4677 return -1; /* log made by save_backend_variables */
4678 }
4679
4680 /* Drop the parameter shared memory that is now inherited to the backend */
4681 if (!UnmapViewOfFile(param))
4682 elog(LOG, "could not unmap view of backend parameter file: error code %lu",
4683 GetLastError());
4684 if (!CloseHandle(paramHandle))
4685 elog(LOG, "could not close handle to backend parameter file: error code %lu",
4686 GetLastError());
4687
4688 /*
4689 * Reserve the memory region used by our main shared memory segment before
4690 * we resume the child process. Normally this should succeed, but if ASLR
4691 * is active then it might sometimes fail due to the stack or heap having
4692 * gotten mapped into that range. In that case, just terminate the
4693 * process and retry.
4694 */
4695 if (!pgwin32_ReserveSharedMemoryRegion(pi.hProcess))
4696 {
4697 /* pgwin32_ReserveSharedMemoryRegion already made a log entry */
4698 if (!TerminateProcess(pi.hProcess, 255))
4699 ereport(LOG,
4700 (errmsg_internal("could not terminate process that failed to reserve memory: error code %lu",
4701 GetLastError())));
4702 CloseHandle(pi.hProcess);
4703 CloseHandle(pi.hThread);
4704 if (++retry_count < 100)
4705 goto retry;
4706 ereport(LOG,
4707 (errmsg("giving up after too many tries to reserve shared memory"),
4708 errhint("This might be caused by ASLR or antivirus software.")));
4709 return -1;
4710 }
4711
4712 /*
4713 * Now that the backend variables are written out, we start the child
4714 * thread so it can start initializing while we set up the rest of the
4715 * parent state.
4716 */
4717 if (ResumeThread(pi.hThread) == -1)
4718 {
4719 if (!TerminateProcess(pi.hProcess, 255))
4720 {
4721 ereport(LOG,
4722 (errmsg_internal("could not terminate unstartable process: error code %lu",
4723 GetLastError())));
4724 CloseHandle(pi.hProcess);
4725 CloseHandle(pi.hThread);
4726 return -1;
4727 }
4728 CloseHandle(pi.hProcess);
4729 CloseHandle(pi.hThread);
4730 ereport(LOG,
4731 (errmsg_internal("could not resume thread of unstarted process: error code %lu",
4732 GetLastError())));
4733 return -1;
4734 }
4735
4736 /*
4737 * Queue a waiter for to signal when this child dies. The wait will be
4738 * handled automatically by an operating system thread pool.
4739 *
4740 * Note: use malloc instead of palloc, since it needs to be thread-safe.
4741 * Struct will be free():d from the callback function that runs on a
4742 * different thread.
4743 */
4744 childinfo = malloc(sizeof(win32_deadchild_waitinfo));
4745 if (!childinfo)
4746 ereport(FATAL,
4747 (errcode(ERRCODE_OUT_OF_MEMORY),
4748 errmsg("out of memory")));
4749
4750 childinfo->procHandle = pi.hProcess;
4751 childinfo->procId = pi.dwProcessId;
4752
4753 if (!RegisterWaitForSingleObject(&childinfo->waitHandle,
4754 pi.hProcess,
4755 pgwin32_deadchild_callback,
4756 childinfo,
4757 INFINITE,
4758 WT_EXECUTEONLYONCE | WT_EXECUTEINWAITTHREAD))
4759 ereport(FATAL,
4760 (errmsg_internal("could not register process for wait: error code %lu",
4761 GetLastError())));
4762
4763 /* Don't close pi.hProcess here - the wait thread needs access to it */
4764
4765 CloseHandle(pi.hThread);
4766
4767 return pi.dwProcessId;
4768 }
4769 #endif /* WIN32 */
4770
4771
4772 /*
4773 * SubPostmasterMain -- Get the fork/exec'd process into a state equivalent
4774 * to what it would be if we'd simply forked on Unix, and then
4775 * dispatch to the appropriate place.
4776 *
4777 * The first two command line arguments are expected to be "--forkFOO"
4778 * (where FOO indicates which postmaster child we are to become), and
4779 * the name of a variables file that we can read to load data that would
4780 * have been inherited by fork() on Unix. Remaining arguments go to the
4781 * subprocess FooMain() routine.
4782 */
4783 void
4784 SubPostmasterMain(int argc, char *argv[])
4785 {
4786 Port port;
4787
4788 /* In EXEC_BACKEND case we will not have inherited these settings */
4789 IsPostmasterEnvironment = true;
4790 whereToSendOutput = DestNone;
4791
4792 /* Setup as postmaster child */
4793 InitPostmasterChild();
4794
4795 /* Setup essential subsystems (to ensure elog() behaves sanely) */
4796 InitializeGUCOptions();
4797
4798 /* Check we got appropriate args */
4799 if (argc < 3)
4800 elog(FATAL, "invalid subpostmaster invocation");
4801
4802 /* Read in the variables file */
4803 memset(&port, 0, sizeof(Port));
4804 read_backend_variables(argv[2], &port);
4805
4806 /* Close the postmaster's sockets (as soon as we know them) */
4807 ClosePostmasterPorts(strcmp(argv[1], "--forklog") == 0);
4808
4809 /*
4810 * Set reference point for stack-depth checking
4811 */
4812 set_stack_base();
4813
4814 /*
4815 * Set up memory area for GSS information. Mirrors the code in ConnCreate
4816 * for the non-exec case.
4817 */
4818 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
4819 port.gss = (pg_gssinfo *) calloc(1, sizeof(pg_gssinfo));
4820 if (!port.gss)
4821 ereport(FATAL,
4822 (errcode(ERRCODE_OUT_OF_MEMORY),
4823 errmsg("out of memory")));
4824 #endif
4825
4826 /*
4827 * If appropriate, physically re-attach to shared memory segment. We want
4828 * to do this before going any further to ensure that we can attach at the
4829 * same address the postmaster used. On the other hand, if we choose not
4830 * to re-attach, we may have other cleanup to do.
4831 *
4832 * If testing EXEC_BACKEND on Linux, you should run this as root before
4833 * starting the postmaster:
4834 *
4835 * echo 0 >/proc/sys/kernel/randomize_va_space
4836 *
4837 * This prevents using randomized stack and code addresses that cause the
4838 * child process's memory map to be different from the parent's, making it
4839 * sometimes impossible to attach to shared memory at the desired address.
4840 * Return the setting to its old value (usually '1' or '2') when finished.
4841 */
4842 if (strcmp(argv[1], "--forkbackend") == 0 ||
4843 strcmp(argv[1], "--forkavlauncher") == 0 ||
4844 strcmp(argv[1], "--forkavworker") == 0 ||
4845 strcmp(argv[1], "--forkboot") == 0 ||
4846 strncmp(argv[1], "--forkbgworker=", 15) == 0)
4847 PGSharedMemoryReAttach();
4848 else
4849 PGSharedMemoryNoReAttach();
4850
4851 /* autovacuum needs this set before calling InitProcess */
4852 if (strcmp(argv[1], "--forkavlauncher") == 0)
4853 AutovacuumLauncherIAm();
4854 if (strcmp(argv[1], "--forkavworker") == 0)
4855 AutovacuumWorkerIAm();
4856
4857 /*
4858 * Start our win32 signal implementation. This has to be done after we
4859 * read the backend variables, because we need to pick up the signal pipe
4860 * from the parent process.
4861 */
4862 #ifdef WIN32
4863 pgwin32_signal_initialize();
4864 #endif
4865
4866 /* In EXEC_BACKEND case we will not have inherited these settings */
4867 pqinitmask();
4868 PG_SETMASK(&BlockSig);
4869
4870 /* Read in remaining GUC variables */
4871 read_nondefault_variables();
4872
4873 /*
4874 * Check that the data directory looks valid, which will also check the
4875 * privileges on the data directory and update our umask and file/group
4876 * variables for creating files later. Note: this should really be done
4877 * before we create any files or directories.
4878 */
4879 checkDataDir();
4880
4881 /*
4882 * (re-)read control file, as it contains config. The postmaster will
4883 * already have read this, but this process doesn't know about that.
4884 */
4885 LocalProcessControlFile(false);
4886
4887 /*
4888 * Reload any libraries that were preloaded by the postmaster. Since we
4889 * exec'd this process, those libraries didn't come along with us; but we
4890 * should load them into all child processes to be consistent with the
4891 * non-EXEC_BACKEND behavior.
4892 */
4893 process_shared_preload_libraries();
4894
4895 /* Run backend or appropriate child */
4896 if (strcmp(argv[1], "--forkbackend") == 0)
4897 {
4898 Assert(argc == 3); /* shouldn't be any more args */
4899
4900 /*
4901 * Need to reinitialize the SSL library in the backend, since the
4902 * context structures contain function pointers and cannot be passed
4903 * through the parameter file.
4904 *
4905 * If for some reason reload fails (maybe the user installed broken
4906 * key files), soldier on without SSL; that's better than all
4907 * connections becoming impossible.
4908 *
4909 * XXX should we do this in all child processes? For the moment it's
4910 * enough to do it in backend children.
4911 */
4912 #ifdef USE_SSL
4913 if (EnableSSL)
4914 {
4915 if (secure_initialize(false) == 0)
4916 LoadedSSL = true;
4917 else
4918 ereport(LOG,
4919 (errmsg("SSL configuration could not be loaded in child process")));
4920 }
4921 #endif
4922
4923 /*
4924 * Perform additional initialization and collect startup packet.
4925 *
4926 * We want to do this before InitProcess() for a couple of reasons: 1.
4927 * so that we aren't eating up a PGPROC slot while waiting on the
4928 * client. 2. so that if InitProcess() fails due to being out of
4929 * PGPROC slots, we have already initialized libpq and are able to
4930 * report the error to the client.
4931 */
4932 BackendInitialize(&port);
4933
4934 /* Restore basic shared memory pointers */
4935 InitShmemAccess(UsedShmemSegAddr);
4936
4937 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4938 InitProcess();
4939
4940 /* Attach process to shared data structures */
4941 CreateSharedMemoryAndSemaphores(0);
4942
4943 /* And run the backend */
4944 BackendRun(&port); /* does not return */
4945 }
4946 if (strcmp(argv[1], "--forkboot") == 0)
4947 {
4948 /* Restore basic shared memory pointers */
4949 InitShmemAccess(UsedShmemSegAddr);
4950
4951 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4952 InitAuxiliaryProcess();
4953
4954 /* Attach process to shared data structures */
4955 CreateSharedMemoryAndSemaphores(0);
4956
4957 AuxiliaryProcessMain(argc - 2, argv + 2); /* does not return */
4958 }
4959 if (strcmp(argv[1], "--forkavlauncher") == 0)
4960 {
4961 /* Restore basic shared memory pointers */
4962 InitShmemAccess(UsedShmemSegAddr);
4963
4964 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4965 InitProcess();
4966
4967 /* Attach process to shared data structures */
4968 CreateSharedMemoryAndSemaphores(0);
4969
4970 AutoVacLauncherMain(argc - 2, argv + 2); /* does not return */
4971 }
4972 if (strcmp(argv[1], "--forkavworker") == 0)
4973 {
4974 /* Restore basic shared memory pointers */
4975 InitShmemAccess(UsedShmemSegAddr);
4976
4977 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4978 InitProcess();
4979
4980 /* Attach process to shared data structures */
4981 CreateSharedMemoryAndSemaphores(0);
4982
4983 AutoVacWorkerMain(argc - 2, argv + 2); /* does not return */
4984 }
4985 if (strncmp(argv[1], "--forkbgworker=", 15) == 0)
4986 {
4987 int shmem_slot;
4988
4989 /* do this as early as possible; in particular, before InitProcess() */
4990 IsBackgroundWorker = true;
4991
4992 /* Restore basic shared memory pointers */
4993 InitShmemAccess(UsedShmemSegAddr);
4994
4995 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4996 InitProcess();
4997
4998 /* Attach process to shared data structures */
4999 CreateSharedMemoryAndSemaphores(0);
5000
5001 /* Fetch MyBgworkerEntry from shared memory */
5002 shmem_slot = atoi(argv[1] + 15);
5003 MyBgworkerEntry = BackgroundWorkerEntry(shmem_slot);
5004
5005 StartBackgroundWorker();
5006 }
5007 if (strcmp(argv[1], "--forkarch") == 0)
5008 {
5009 /* Do not want to attach to shared memory */
5010
5011 PgArchiverMain(argc, argv); /* does not return */
5012 }
5013 if (strcmp(argv[1], "--forkcol") == 0)
5014 {
5015 /* Do not want to attach to shared memory */
5016
5017 PgstatCollectorMain(argc, argv); /* does not return */
5018 }
5019 if (strcmp(argv[1], "--forklog") == 0)
5020 {
5021 /* Do not want to attach to shared memory */
5022
5023 SysLoggerMain(argc, argv); /* does not return */
5024 }
5025
5026 abort(); /* shouldn't get here */
5027 }
5028 #endif /* EXEC_BACKEND */
5029
5030
5031 /*
5032 * ExitPostmaster -- cleanup
5033 *
5034 * Do NOT call exit() directly --- always go through here!
5035 */
5036 static void
5037 ExitPostmaster(int status)
5038 {
5039 #ifdef HAVE_PTHREAD_IS_THREADED_NP
5040
5041 /*
5042 * There is no known cause for a postmaster to become multithreaded after
5043 * startup. Recheck to account for the possibility of unknown causes.
5044 * This message uses LOG level, because an unclean shutdown at this point
5045 * would usually not look much different from a clean shutdown.
5046 */
5047 if (pthread_is_threaded_np() != 0)
5048 ereport(LOG,
5049 (errcode(ERRCODE_INTERNAL_ERROR),
5050 errmsg_internal("postmaster became multithreaded"),
5051 errdetail("Please report this to <pgsql-bugs@postgresql.org>.")));
5052 #endif
5053
5054 /* should cleanup shared memory and kill all backends */
5055
5056 /*
5057 * Not sure of the semantics here. When the Postmaster dies, should the
5058 * backends all be killed? probably not.
5059 *
5060 * MUST -- vadim 05-10-1999
5061 */
5062
5063 proc_exit(status);
5064 }
5065
5066 /*
5067 * sigusr1_handler - handle signal conditions from child processes
5068 */
5069 static void
5070 sigusr1_handler(SIGNAL_ARGS)
5071 {
5072 int save_errno = errno;
5073
5074 /*
5075 * We rely on the signal mechanism to have blocked all signals ... except
5076 * on Windows, which lacks sigaction(), so we have to do it manually.
5077 */
5078 #ifdef WIN32
5079 PG_SETMASK(&BlockSig);
5080 #endif
5081
5082 /*
5083 * RECOVERY_STARTED and BEGIN_HOT_STANDBY signals are ignored in
5084 * unexpected states. If the startup process quickly starts up, completes
5085 * recovery, exits, we might process the death of the startup process
5086 * first. We don't want to go back to recovery in that case.
5087 */
5088 if (CheckPostmasterSignal(PMSIGNAL_RECOVERY_STARTED) &&
5089 pmState == PM_STARTUP && Shutdown == NoShutdown)
5090 {
5091 /* WAL redo has started. We're out of reinitialization. */
5092 FatalError = false;
5093 Assert(AbortStartTime == 0);
5094
5095 /*
5096 * Crank up the background tasks. It doesn't matter if this fails,
5097 * we'll just try again later.
5098 */
5099 Assert(CheckpointerPID == 0);
5100 CheckpointerPID = StartCheckpointer();
5101 Assert(BgWriterPID == 0);
5102 BgWriterPID = StartBackgroundWriter();
5103
5104 /*
5105 * Start the archiver if we're responsible for (re-)archiving received
5106 * files.
5107 */
5108 Assert(PgArchPID == 0);
5109 if (XLogArchivingAlways())
5110 PgArchPID = pgarch_start();
5111
5112 /*
5113 * If we aren't planning to enter hot standby mode later, treat
5114 * RECOVERY_STARTED as meaning we're out of startup, and report status
5115 * accordingly.
5116 */
5117 if (!EnableHotStandby)
5118 {
5119 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STANDBY);
5120 #ifdef USE_SYSTEMD
5121 sd_notify(0, "READY=1");
5122 #endif
5123 }
5124
5125 pmState = PM_RECOVERY;
5126 }
5127
5128 if (CheckPostmasterSignal(PMSIGNAL_BEGIN_HOT_STANDBY) &&
5129 pmState == PM_RECOVERY && Shutdown == NoShutdown)
5130 {
5131 /*
5132 * Likewise, start other special children as needed.
5133 */
5134 Assert(PgStatPID == 0);
5135 PgStatPID = pgstat_start();
5136
5137 ereport(LOG,
5138 (errmsg("database system is ready to accept read only connections")));
5139
5140 /* Report status */
5141 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
5142 #ifdef USE_SYSTEMD
5143 sd_notify(0, "READY=1");
5144 #endif
5145
5146 pmState = PM_HOT_STANDBY;
5147 connsAllowed = ALLOW_ALL_CONNS;
5148
5149 /* Some workers may be scheduled to start now */
5150 StartWorkerNeeded = true;
5151 }
5152
5153 /* Process background worker state changes. */
5154 if (CheckPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE))
5155 {
5156 /* Accept new worker requests only if not stopping. */
5157 BackgroundWorkerStateChange(pmState < PM_STOP_BACKENDS);
5158 StartWorkerNeeded = true;
5159 }
5160
5161 if (StartWorkerNeeded || HaveCrashedWorker)
5162 maybe_start_bgworkers();
5163
5164 if (CheckPostmasterSignal(PMSIGNAL_WAKEN_ARCHIVER) &&
5165 PgArchPID != 0)
5166 {
5167 /*
5168 * Send SIGUSR1 to archiver process, to wake it up and begin archiving
5169 * next WAL file.
5170 */
5171 signal_child(PgArchPID, SIGUSR1);
5172 }
5173
5174 if (CheckPostmasterSignal(PMSIGNAL_ROTATE_LOGFILE) &&
5175 SysLoggerPID != 0)
5176 {
5177 /* Tell syslogger to rotate logfile */
5178 signal_child(SysLoggerPID, SIGUSR1);
5179 }
5180
5181 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER) &&
5182 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5183 {
5184 /*
5185 * Start one iteration of the autovacuum daemon, even if autovacuuming
5186 * is nominally not enabled. This is so we can have an active defense
5187 * against transaction ID wraparound. We set a flag for the main loop
5188 * to do it rather than trying to do it here --- this is because the
5189 * autovac process itself may send the signal, and we want to handle
5190 * that by launching another iteration as soon as the current one
5191 * completes.
5192 */
5193 start_autovac_launcher = true;
5194 }
5195
5196 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER) &&
5197 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5198 {
5199 /* The autovacuum launcher wants us to start a worker process. */
5200 StartAutovacuumWorker();
5201 }
5202
5203 if (CheckPostmasterSignal(PMSIGNAL_START_WALRECEIVER))
5204 {
5205 /* Startup Process wants us to start the walreceiver process. */
5206 /* Start immediately if possible, else remember request for later. */
5207 WalReceiverRequested = true;
5208 MaybeStartWalReceiver();
5209 }
5210
5211 /*
5212 * Try to advance postmaster's state machine, if a child requests it.
5213 *
5214 * Be careful about the order of this action relative to sigusr1_handler's
5215 * other actions. Generally, this should be after other actions, in case
5216 * they have effects PostmasterStateMachine would need to know about.
5217 * However, we should do it before the CheckPromoteSignal step, which
5218 * cannot have any (immediate) effect on the state machine, but does
5219 * depend on what state we're in now.
5220 */
5221 if (CheckPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE))
5222 {
5223 PostmasterStateMachine();
5224 }
5225
5226 if (StartupPID != 0 &&
5227 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5228 pmState == PM_HOT_STANDBY) &&
5229 CheckPromoteSignal())
5230 {
5231 /* Tell startup process to finish recovery */
5232 signal_child(StartupPID, SIGUSR2);
5233 }
5234
5235 #ifdef WIN32
5236 PG_SETMASK(&UnBlockSig);
5237 #endif
5238
5239 errno = save_errno;
5240 }
5241
5242 /*
5243 * SIGTERM while processing startup packet.
5244 * Clean up and exit(1).
5245 *
5246 * Running proc_exit() from a signal handler is pretty unsafe, since we
5247 * can't know what code we've interrupted. But the alternative of using
5248 * _exit(2) is also unpalatable, since it'd mean that a "fast shutdown"
5249 * would cause a database crash cycle (forcing WAL replay at restart)
5250 * if any sessions are in authentication. So we live with it for now.
5251 *
5252 * One might be tempted to try to send a message indicating why we are
5253 * disconnecting. However, that would make this even more unsafe. Also,
5254 * it seems undesirable to provide clues about the database's state to
5255 * a client that has not yet completed authentication.
5256 */
5257 static void
5258 process_startup_packet_die(SIGNAL_ARGS)
5259 {
5260 proc_exit(1);
5261 }
5262
5263 /*
5264 * SIGQUIT while processing startup packet.
5265 *
5266 * Some backend has bought the farm,
5267 * so we need to stop what we're doing and exit.
5268 */
5269 static void
5270 process_startup_packet_quickdie(SIGNAL_ARGS)
5271 {
5272 /*
5273 * We DO NOT want to run proc_exit() or atexit() callbacks; they wouldn't
5274 * be safe to run from a signal handler. Just nail the windows shut and
5275 * get out of town.
5276 *
5277 * Note we do _exit(2) not _exit(1). This is to force the postmaster into
5278 * a system reset cycle if someone sends a manual SIGQUIT to a random
5279 * backend. (While it might be safe to do _exit(1), since this session
5280 * shouldn't have touched shared memory yet, there seems little point in
5281 * taking any risks.)
5282 */
5283 _exit(2);
5284 }
5285
5286 /*
5287 * Dummy signal handler
5288 *
5289 * We use this for signals that we don't actually use in the postmaster,
5290 * but we do use in backends. If we were to SIG_IGN such signals in the
5291 * postmaster, then a newly started backend might drop a signal that arrives
5292 * before it's able to reconfigure its signal processing. (See notes in
5293 * tcop/postgres.c.)
5294 */
5295 static void
5296 dummy_handler(SIGNAL_ARGS)
5297 {
5298 }
5299
5300 /*
5301 * Timeout while processing startup packet.
5302 * As for process_startup_packet_die(), we clean up and exit(1).
5303 *
5304 * This is theoretically just as hazardous as in process_startup_packet_die(),
5305 * although in practice we're almost certainly waiting for client input,
5306 * which greatly reduces the risk.
5307 */
5308 static void
5309 StartupPacketTimeoutHandler(void)
5310 {
5311 proc_exit(1);
5312 }
5313
5314
5315 /*
5316 * Generate a random cancel key.
5317 */
5318 static bool
5319 RandomCancelKey(int32 *cancel_key)
5320 {
5321 #ifdef HAVE_STRONG_RANDOM
5322 return pg_strong_random((char *) cancel_key, sizeof(int32));
5323 #else
5324
5325 /*
5326 * If built with --disable-strong-random, use plain old erand48.
5327 *
5328 * We cannot use pg_backend_random() in postmaster, because it stores its
5329 * state in shared memory.
5330 */
5331 static unsigned short seed[3];
5332
5333 /*
5334 * Select a random seed at the time of first receiving a request.
5335 */
5336 if (random_seed == 0)
5337 {
5338 struct timeval random_stop_time;
5339
5340 gettimeofday(&random_stop_time, NULL);
5341
5342 seed[0] = (unsigned short) random_start_time.tv_usec;
5343 seed[1] = (unsigned short) (random_stop_time.tv_usec) ^ (random_start_time.tv_usec >> 16);
5344 seed[2] = (unsigned short) (random_stop_time.tv_usec >> 16);
5345
5346 random_seed = 1;
5347 }
5348
5349 *cancel_key = pg_jrand48(seed);
5350
5351 return true;
5352 #endif
5353 }
5354
5355 /*
5356 * Count up number of child processes of specified types (dead_end children
5357 * are always excluded).
5358 */
5359 static int
5360 CountChildren(int target)
5361 {
5362 dlist_iter iter;
5363 int cnt = 0;
5364
5365 dlist_foreach(iter, &BackendList)
5366 {
5367 Backend *bp = dlist_container(Backend, elem, iter.cur);
5368
5369 if (bp->dead_end)
5370 continue;
5371
5372 /*
5373 * Since target == BACKEND_TYPE_ALL is the most common case, we test
5374 * it first and avoid touching shared memory for every child.
5375 */
5376 if (target != BACKEND_TYPE_ALL)
5377 {
5378 /*
5379 * Assign bkend_type for any recently announced WAL Sender
5380 * processes.
5381 */
5382 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
5383 IsPostmasterChildWalSender(bp->child_slot))
5384 bp->bkend_type = BACKEND_TYPE_WALSND;
5385
5386 if (!(target & bp->bkend_type))
5387 continue;
5388 }
5389
5390 cnt++;
5391 }
5392 return cnt;
5393 }
5394
5395
5396 /*
5397 * StartChildProcess -- start an auxiliary process for the postmaster
5398 *
5399 * "type" determines what kind of child will be started. All child types
5400 * initially go to AuxiliaryProcessMain, which will handle common setup.
5401 *
5402 * Return value of StartChildProcess is subprocess' PID, or 0 if failed
5403 * to start subprocess.
5404 */
5405 static pid_t
5406 StartChildProcess(AuxProcType type)
5407 {
5408 pid_t pid;
5409 char *av[10];
5410 int ac = 0;
5411 char typebuf[32];
5412
5413 /*
5414 * Set up command-line arguments for subprocess
5415 */
5416 av[ac++] = "postgres";
5417
5418 #ifdef EXEC_BACKEND
5419 av[ac++] = "--forkboot";
5420 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5421 #endif
5422
5423 snprintf(typebuf, sizeof(typebuf), "-x%d", type);
5424 av[ac++] = typebuf;
5425
5426 av[ac] = NULL;
5427 Assert(ac < lengthof(av));
5428
5429 #ifdef EXEC_BACKEND
5430 pid = postmaster_forkexec(ac, av);
5431 #else /* !EXEC_BACKEND */
5432 pid = fork_process();
5433
5434 if (pid == 0) /* child */
5435 {
5436 InitPostmasterChild();
5437
5438 /* Close the postmaster's sockets */
5439 ClosePostmasterPorts(false);
5440
5441 /* Release postmaster's working memory context */
5442 MemoryContextSwitchTo(TopMemoryContext);
5443 MemoryContextDelete(PostmasterContext);
5444 PostmasterContext = NULL;
5445
5446 AuxiliaryProcessMain(ac, av);
5447 ExitPostmaster(0);
5448 }
5449 #endif /* EXEC_BACKEND */
5450
5451 if (pid < 0)
5452 {
5453 /* in parent, fork failed */
5454 int save_errno = errno;
5455
5456 errno = save_errno;
5457 switch (type)
5458 {
5459 case StartupProcess:
5460 ereport(LOG,
5461 (errmsg("could not fork startup process: %m")));
5462 break;
5463 case BgWriterProcess:
5464 ereport(LOG,
5465 (errmsg("could not fork background writer process: %m")));
5466 break;
5467 case CheckpointerProcess:
5468 ereport(LOG,
5469 (errmsg("could not fork checkpointer process: %m")));
5470 break;
5471 case WalWriterProcess:
5472 ereport(LOG,
5473 (errmsg("could not fork WAL writer process: %m")));
5474 break;
5475 case WalReceiverProcess:
5476 ereport(LOG,
5477 (errmsg("could not fork WAL receiver process: %m")));
5478 break;
5479 default:
5480 ereport(LOG,
5481 (errmsg("could not fork process: %m")));
5482 break;
5483 }
5484
5485 /*
5486 * fork failure is fatal during startup, but there's no need to choke
5487 * immediately if starting other child types fails.
5488 */
5489 if (type == StartupProcess)
5490 ExitPostmaster(1);
5491 return 0;
5492 }
5493
5494 /*
5495 * in parent, successful fork
5496 */
5497 return pid;
5498 }
5499
5500 /*
5501 * StartAutovacuumWorker
5502 * Start an autovac worker process.
5503 *
5504 * This function is here because it enters the resulting PID into the
5505 * postmaster's private backends list.
5506 *
5507 * NB -- this code very roughly matches BackendStartup.
5508 */
5509 static void
5510 StartAutovacuumWorker(void)
5511 {
5512 Backend *bn;
5513
5514 /*
5515 * If not in condition to run a process, don't try, but handle it like a
5516 * fork failure. This does not normally happen, since the signal is only
5517 * supposed to be sent by autovacuum launcher when it's OK to do it, but
5518 * we have to check to avoid race-condition problems during DB state
5519 * changes.
5520 */
5521 if (canAcceptConnections(BACKEND_TYPE_AUTOVAC) == CAC_OK)
5522 {
5523 /*
5524 * Compute the cancel key that will be assigned to this session. We
5525 * probably don't need cancel keys for autovac workers, but we'd
5526 * better have something random in the field to prevent unfriendly
5527 * people from sending cancels to them.
5528 */
5529 if (!RandomCancelKey(&MyCancelKey))
5530 {
5531 ereport(LOG,
5532 (errcode(ERRCODE_INTERNAL_ERROR),
5533 errmsg("could not generate random cancel key")));
5534 return;
5535 }
5536
5537 bn = (Backend *) malloc(sizeof(Backend));
5538 if (bn)
5539 {
5540 bn->cancel_key = MyCancelKey;
5541
5542 /* Autovac workers are not dead_end and need a child slot */
5543 bn->dead_end = false;
5544 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5545 bn->bgworker_notify = false;
5546
5547 bn->pid = StartAutoVacWorker();
5548 if (bn->pid > 0)
5549 {
5550 bn->bkend_type = BACKEND_TYPE_AUTOVAC;
5551 dlist_push_head(&BackendList, &bn->elem);
5552 #ifdef EXEC_BACKEND
5553 ShmemBackendArrayAdd(bn);
5554 #endif
5555 /* all OK */
5556 return;
5557 }
5558
5559 /*
5560 * fork failed, fall through to report -- actual error message was
5561 * logged by StartAutoVacWorker
5562 */
5563 (void) ReleasePostmasterChildSlot(bn->child_slot);
5564 free(bn);
5565 }
5566 else
5567 ereport(LOG,
5568 (errcode(ERRCODE_OUT_OF_MEMORY),
5569 errmsg("out of memory")));
5570 }
5571
5572 /*
5573 * Report the failure to the launcher, if it's running. (If it's not, we
5574 * might not even be connected to shared memory, so don't try to call
5575 * AutoVacWorkerFailed.) Note that we also need to signal it so that it
5576 * responds to the condition, but we don't do that here, instead waiting
5577 * for ServerLoop to do it. This way we avoid a ping-pong signalling in
5578 * quick succession between the autovac launcher and postmaster in case
5579 * things get ugly.
5580 */
5581 if (AutoVacPID != 0)
5582 {
5583 AutoVacWorkerFailed();
5584 avlauncher_needs_signal = true;
5585 }
5586 }
5587
5588 /*
5589 * MaybeStartWalReceiver
5590 * Start the WAL receiver process, if not running and our state allows.
5591 *
5592 * Note: if WalReceiverPID is already nonzero, it might seem that we should
5593 * clear WalReceiverRequested. However, there's a race condition if the
5594 * walreceiver terminates and the startup process immediately requests a new
5595 * one: it's quite possible to get the signal for the request before reaping
5596 * the dead walreceiver process. Better to risk launching an extra
5597 * walreceiver than to miss launching one we need. (The walreceiver code
5598 * has logic to recognize that it should go away if not needed.)
5599 */
5600 static void
5601 MaybeStartWalReceiver(void)
5602 {
5603 if (WalReceiverPID == 0 &&
5604 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5605 pmState == PM_HOT_STANDBY) &&
5606 Shutdown <= SmartShutdown)
5607 {
5608 WalReceiverPID = StartWalReceiver();
5609 if (WalReceiverPID != 0)
5610 WalReceiverRequested = false;
5611 /* else leave the flag set, so we'll try again later */
5612 }
5613 }
5614
5615
5616 /*
5617 * Create the opts file
5618 */
5619 static bool
5620 CreateOptsFile(int argc, char *argv[], char *fullprogname)
5621 {
5622 FILE *fp;
5623 int i;
5624
5625 #define OPTS_FILE "postmaster.opts"
5626
5627 if ((fp = fopen(OPTS_FILE, "w")) == NULL)
5628 {
5629 elog(LOG, "could not create file \"%s\": %m", OPTS_FILE);
5630 return false;
5631 }
5632
5633 fprintf(fp, "%s", fullprogname);
5634 for (i = 1; i < argc; i++)
5635 fprintf(fp, " \"%s\"", argv[i]);
5636 fputs("\n", fp);
5637
5638 if (fclose(fp))
5639 {
5640 elog(LOG, "could not write file \"%s\": %m", OPTS_FILE);
5641 return false;
5642 }
5643
5644 return true;
5645 }
5646
5647
5648 /*
5649 * MaxLivePostmasterChildren
5650 *
5651 * This reports the number of entries needed in per-child-process arrays
5652 * (the PMChildFlags array, and if EXEC_BACKEND the ShmemBackendArray).
5653 * These arrays include regular backends, autovac workers, walsenders
5654 * and background workers, but not special children nor dead_end children.
5655 * This allows the arrays to have a fixed maximum size, to wit the same
5656 * too-many-children limit enforced by canAcceptConnections(). The exact value
5657 * isn't too critical as long as it's more than MaxBackends.
5658 */
5659 int
5660 MaxLivePostmasterChildren(void)
5661 {
5662 return 2 * (MaxConnections + autovacuum_max_workers + 1 +
5663 max_worker_processes);
5664 }
5665
5666 /*
5667 * Connect background worker to a database.
5668 */
5669 void
5670 BackgroundWorkerInitializeConnection(const char *dbname, const char *username, uint32 flags)
5671 {
5672 BackgroundWorker *worker = MyBgworkerEntry;
5673
5674 /* XXX is this the right errcode? */
5675 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5676 ereport(FATAL,
5677 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5678 errmsg("database connection requirement not indicated during registration")));
5679
5680 InitPostgres(dbname, InvalidOid, username, InvalidOid, NULL, (flags & BGWORKER_BYPASS_ALLOWCONN) != 0);
5681
5682 /* it had better not gotten out of "init" mode yet */
5683 if (!IsInitProcessingMode())
5684 ereport(ERROR,
5685 (errmsg("invalid processing mode in background worker")));
5686 SetProcessingMode(NormalProcessing);
5687 }
5688
5689 /*
5690 * Connect background worker to a database using OIDs.
5691 */
5692 void
5693 BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid, uint32 flags)
5694 {
5695 BackgroundWorker *worker = MyBgworkerEntry;
5696
5697 /* XXX is this the right errcode? */
5698 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5699 ereport(FATAL,
5700 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5701 errmsg("database connection requirement not indicated during registration")));
5702
5703 InitPostgres(NULL, dboid, NULL, useroid, NULL, (flags & BGWORKER_BYPASS_ALLOWCONN) != 0);
5704
5705 /* it had better not gotten out of "init" mode yet */
5706 if (!IsInitProcessingMode())
5707 ereport(ERROR,
5708 (errmsg("invalid processing mode in background worker")));
5709 SetProcessingMode(NormalProcessing);
5710 }
5711
5712 /*
5713 * Block/unblock signals in a background worker
5714 */
5715 void
5716 BackgroundWorkerBlockSignals(void)
5717 {
5718 PG_SETMASK(&BlockSig);
5719 }
5720
5721 void
5722 BackgroundWorkerUnblockSignals(void)
5723 {
5724 PG_SETMASK(&UnBlockSig);
5725 }
5726
5727 #ifdef EXEC_BACKEND
5728 static pid_t
5729 bgworker_forkexec(int shmem_slot)
5730 {
5731 char *av[10];
5732 int ac = 0;
5733 char forkav[MAXPGPATH];
5734
5735 snprintf(forkav, MAXPGPATH, "--forkbgworker=%d", shmem_slot);
5736
5737 av[ac++] = "postgres";
5738 av[ac++] = forkav;
5739 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5740 av[ac] = NULL;
5741
5742 Assert(ac < lengthof(av));
5743
5744 return postmaster_forkexec(ac, av);
5745 }
5746 #endif
5747
5748 /*
5749 * Start a new bgworker.
5750 * Starting time conditions must have been checked already.
5751 *
5752 * Returns true on success, false on failure.
5753 * In either case, update the RegisteredBgWorker's state appropriately.
5754 *
5755 * This code is heavily based on autovacuum.c, q.v.
5756 */
5757 static bool
5758 do_start_bgworker(RegisteredBgWorker *rw)
5759 {
5760 pid_t worker_pid;
5761
5762 Assert(rw->rw_pid == 0);
5763
5764 /*
5765 * Allocate and assign the Backend element. Note we must do this before
5766 * forking, so that we can handle failures (out of memory or child-process
5767 * slots) cleanly.
5768 *
5769 * Treat failure as though the worker had crashed. That way, the
5770 * postmaster will wait a bit before attempting to start it again; if we
5771 * tried again right away, most likely we'd find ourselves hitting the
5772 * same resource-exhaustion condition.
5773 */
5774 if (!assign_backendlist_entry(rw))
5775 {
5776 rw->rw_crashed_at = GetCurrentTimestamp();
5777 return false;
5778 }
5779
5780 ereport(DEBUG1,
5781 (errmsg("starting background worker process \"%s\"",
5782 rw->rw_worker.bgw_name)));
5783
5784 #ifdef EXEC_BACKEND
5785 switch ((worker_pid = bgworker_forkexec(rw->rw_shmem_slot)))
5786 #else
5787 switch ((worker_pid = fork_process()))
5788 #endif
5789 {
5790 case -1:
5791 /* in postmaster, fork failed ... */
5792 ereport(LOG,
5793 (errmsg("could not fork worker process: %m")));
5794 /* undo what assign_backendlist_entry did */
5795 ReleasePostmasterChildSlot(rw->rw_child_slot);
5796 rw->rw_child_slot = 0;
5797 free(rw->rw_backend);
5798 rw->rw_backend = NULL;
5799 /* mark entry as crashed, so we'll try again later */
5800 rw->rw_crashed_at = GetCurrentTimestamp();
5801 break;
5802
5803 #ifndef EXEC_BACKEND
5804 case 0:
5805 /* in postmaster child ... */
5806 InitPostmasterChild();
5807
5808 /* Close the postmaster's sockets */
5809 ClosePostmasterPorts(false);
5810
5811 /*
5812 * Before blowing away PostmasterContext, save this bgworker's
5813 * data where it can find it.
5814 */
5815 MyBgworkerEntry = (BackgroundWorker *)
5816 MemoryContextAlloc(TopMemoryContext, sizeof(BackgroundWorker));
5817 memcpy(MyBgworkerEntry, &rw->rw_worker, sizeof(BackgroundWorker));
5818
5819 /* Release postmaster's working memory context */
5820 MemoryContextSwitchTo(TopMemoryContext);
5821 MemoryContextDelete(PostmasterContext);
5822 PostmasterContext = NULL;
5823
5824 StartBackgroundWorker();
5825
5826 exit(1); /* should not get here */
5827 break;
5828 #endif
5829 default:
5830 /* in postmaster, fork successful ... */
5831 rw->rw_pid = worker_pid;
5832 rw->rw_backend->pid = rw->rw_pid;
5833 ReportBackgroundWorkerPID(rw);
5834 /* add new worker to lists of backends */
5835 dlist_push_head(&BackendList, &rw->rw_backend->elem);
5836 #ifdef EXEC_BACKEND
5837 ShmemBackendArrayAdd(rw->rw_backend);
5838 #endif
5839 return true;
5840 }
5841
5842 return false;
5843 }
5844
5845 /*
5846 * Does the current postmaster state require starting a worker with the
5847 * specified start_time?
5848 */
5849 static bool
5850 bgworker_should_start_now(BgWorkerStartTime start_time)
5851 {
5852 switch (pmState)
5853 {
5854 case PM_NO_CHILDREN:
5855 case PM_WAIT_DEAD_END:
5856 case PM_SHUTDOWN_2:
5857 case PM_SHUTDOWN:
5858 case PM_WAIT_BACKENDS:
5859 case PM_STOP_BACKENDS:
5860 break;
5861
5862 case PM_RUN:
5863 if (start_time == BgWorkerStart_RecoveryFinished)
5864 return true;
5865 /* fall through */
5866
5867 case PM_HOT_STANDBY:
5868 if (start_time == BgWorkerStart_ConsistentState)
5869 return true;
5870 /* fall through */
5871
5872 case PM_RECOVERY:
5873 case PM_STARTUP:
5874 case PM_INIT:
5875 if (start_time == BgWorkerStart_PostmasterStart)
5876 return true;
5877 /* fall through */
5878
5879 }
5880
5881 return false;
5882 }
5883
5884 /*
5885 * Allocate the Backend struct for a connected background worker, but don't
5886 * add it to the list of backends just yet.
5887 *
5888 * On failure, return false without changing any worker state.
5889 *
5890 * Some info from the Backend is copied into the passed rw.
5891 */
5892 static bool
5893 assign_backendlist_entry(RegisteredBgWorker *rw)
5894 {
5895 Backend *bn;
5896
5897 /*
5898 * Check that database state allows another connection. Currently the
5899 * only possible failure is CAC_TOOMANY, so we just log an error message
5900 * based on that rather than checking the error code precisely.
5901 */
5902 if (canAcceptConnections(BACKEND_TYPE_BGWORKER) != CAC_OK)
5903 {
5904 ereport(LOG,
5905 (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
5906 errmsg("no slot available for new worker process")));
5907 return false;
5908 }
5909
5910 /*
5911 * Compute the cancel key that will be assigned to this session. We
5912 * probably don't need cancel keys for background workers, but we'd better
5913 * have something random in the field to prevent unfriendly people from
5914 * sending cancels to them.
5915 */
5916 if (!RandomCancelKey(&MyCancelKey))
5917 {
5918 ereport(LOG,
5919 (errcode(ERRCODE_INTERNAL_ERROR),
5920 errmsg("could not generate random cancel key")));
5921 return false;
5922 }
5923
5924 bn = malloc(sizeof(Backend));
5925 if (bn == NULL)
5926 {
5927 ereport(LOG,
5928 (errcode(ERRCODE_OUT_OF_MEMORY),
5929 errmsg("out of memory")));
5930 return false;
5931 }
5932
5933 bn->cancel_key = MyCancelKey;
5934 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5935 bn->bkend_type = BACKEND_TYPE_BGWORKER;
5936 bn->dead_end = false;
5937 bn->bgworker_notify = false;
5938
5939 rw->rw_backend = bn;
5940 rw->rw_child_slot = bn->child_slot;
5941
5942 return true;
5943 }
5944
5945 /*
5946 * If the time is right, start background worker(s).
5947 *
5948 * As a side effect, the bgworker control variables are set or reset
5949 * depending on whether more workers may need to be started.
5950 *
5951 * We limit the number of workers started per call, to avoid consuming the
5952 * postmaster's attention for too long when many such requests are pending.
5953 * As long as StartWorkerNeeded is true, ServerLoop will not block and will
5954 * call this function again after dealing with any other issues.
5955 */
5956 static void
5957 maybe_start_bgworkers(void)
5958 {
5959 #define MAX_BGWORKERS_TO_LAUNCH 100
5960 int num_launched = 0;
5961 TimestampTz now = 0;
5962 slist_mutable_iter iter;
5963
5964 /*
5965 * During crash recovery, we have no need to be called until the state
5966 * transition out of recovery.
5967 */
5968 if (FatalError)
5969 {
5970 StartWorkerNeeded = false;
5971 HaveCrashedWorker = false;
5972 return;
5973 }
5974
5975 /* Don't need to be called again unless we find a reason for it below */
5976 StartWorkerNeeded = false;
5977 HaveCrashedWorker = false;
5978
5979 slist_foreach_modify(iter, &BackgroundWorkerList)
5980 {
5981 RegisteredBgWorker *rw;
5982
5983 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
5984
5985 /* ignore if already running */
5986 if (rw->rw_pid != 0)
5987 continue;
5988
5989 /* if marked for death, clean up and remove from list */
5990 if (rw->rw_terminate)
5991 {
5992 ForgetBackgroundWorker(&iter);
5993 continue;
5994 }
5995
5996 /*
5997 * If this worker has crashed previously, maybe it needs to be
5998 * restarted (unless on registration it specified it doesn't want to
5999 * be restarted at all). Check how long ago did a crash last happen.
6000 * If the last crash is too recent, don't start it right away; let it
6001 * be restarted once enough time has passed.
6002 */
6003 if (rw->rw_crashed_at != 0)
6004 {
6005 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
6006 {
6007 int notify_pid;
6008
6009 notify_pid = rw->rw_worker.bgw_notify_pid;
6010
6011 ForgetBackgroundWorker(&iter);
6012
6013 /* Report worker is gone now. */
6014 if (notify_pid != 0)
6015 kill(notify_pid, SIGUSR1);
6016
6017 continue;
6018 }
6019
6020 /* read system time only when needed */
6021 if (now == 0)
6022 now = GetCurrentTimestamp();
6023
6024 if (!TimestampDifferenceExceeds(rw->rw_crashed_at, now,
6025 rw->rw_worker.bgw_restart_time * 1000))
6026 {
6027 /* Set flag to remember that we have workers to start later */
6028 HaveCrashedWorker = true;
6029 continue;
6030 }
6031 }
6032
6033 if (bgworker_should_start_now(rw->rw_worker.bgw_start_time))
6034 {
6035 /* reset crash time before trying to start worker */
6036 rw->rw_crashed_at = 0;
6037
6038 /*
6039 * Try to start the worker.
6040 *
6041 * On failure, give up processing workers for now, but set
6042 * StartWorkerNeeded so we'll come back here on the next iteration
6043 * of ServerLoop to try again. (We don't want to wait, because
6044 * there might be additional ready-to-run workers.) We could set
6045 * HaveCrashedWorker as well, since this worker is now marked
6046 * crashed, but there's no need because the next run of this
6047 * function will do that.
6048 */
6049 if (!do_start_bgworker(rw))
6050 {
6051 StartWorkerNeeded = true;
6052 return;
6053 }
6054
6055 /*
6056 * If we've launched as many workers as allowed, quit, but have
6057 * ServerLoop call us again to look for additional ready-to-run
6058 * workers. There might not be any, but we'll find out the next
6059 * time we run.
6060 */
6061 if (++num_launched >= MAX_BGWORKERS_TO_LAUNCH)
6062 {
6063 StartWorkerNeeded = true;
6064 return;
6065 }
6066 }
6067 }
6068 }
6069
6070 /*
6071 * When a backend asks to be notified about worker state changes, we
6072 * set a flag in its backend entry. The background worker machinery needs
6073 * to know when such backends exit.
6074 */
6075 bool
6076 PostmasterMarkPIDForWorkerNotify(int pid)
6077 {
6078 dlist_iter iter;
6079 Backend *bp;
6080
6081 dlist_foreach(iter, &BackendList)
6082 {
6083 bp = dlist_container(Backend, elem, iter.cur);
6084 if (bp->pid == pid)
6085 {
6086 bp->bgworker_notify = true;
6087 return true;
6088 }
6089 }
6090 return false;
6091 }
6092
6093 #ifdef EXEC_BACKEND
6094
6095 /*
6096 * The following need to be available to the save/restore_backend_variables
6097 * functions. They are marked NON_EXEC_STATIC in their home modules.
6098 */
6099 extern slock_t *ShmemLock;
6100 extern slock_t *ProcStructLock;
6101 extern PGPROC *AuxiliaryProcs;
6102 extern PMSignalData *PMSignalState;
6103 extern pgsocket pgStatSock;
6104 extern pg_time_t first_syslogger_file_time;
6105
6106 #ifndef WIN32
6107 #define write_inheritable_socket(dest, src, childpid) ((*(dest) = (src)), true)
6108 #define read_inheritable_socket(dest, src) (*(dest) = *(src))
6109 #else
6110 static bool write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE child);
6111 static bool write_inheritable_socket(InheritableSocket *dest, SOCKET src,
6112 pid_t childPid);
6113 static void read_inheritable_socket(SOCKET *dest, InheritableSocket *src);
6114 #endif
6115
6116
6117 /* Save critical backend variables into the BackendParameters struct */
6118 #ifndef WIN32
6119 static bool
6120 save_backend_variables(BackendParameters *param, Port *port)
6121 #else
6122 static bool
6123 save_backend_variables(BackendParameters *param, Port *port,
6124 HANDLE childProcess, pid_t childPid)
6125 #endif
6126 {
6127 memcpy(¶m->port, port, sizeof(Port));
6128 if (!write_inheritable_socket(¶m->portsocket, port->sock, childPid))
6129 return false;
6130
6131 strlcpy(param->DataDir, DataDir, MAXPGPATH);
6132
6133 memcpy(¶m->ListenSocket, &ListenSocket, sizeof(ListenSocket));
6134
6135 param->MyCancelKey = MyCancelKey;
6136 param->MyPMChildSlot = MyPMChildSlot;
6137
6138 #ifdef WIN32
6139 param->ShmemProtectiveRegion = ShmemProtectiveRegion;
6140 #endif
6141 param->UsedShmemSegID = UsedShmemSegID;
6142 param->UsedShmemSegAddr = UsedShmemSegAddr;
6143
6144 param->ShmemLock = ShmemLock;
6145 param->ShmemVariableCache = ShmemVariableCache;
6146 param->ShmemBackendArray = ShmemBackendArray;
6147
6148 #ifndef HAVE_SPINLOCKS
6149 param->SpinlockSemaArray = SpinlockSemaArray;
6150 #endif
6151 param->NamedLWLockTrancheRequests = NamedLWLockTrancheRequests;
6152 param->NamedLWLockTrancheArray = NamedLWLockTrancheArray;
6153 param->MainLWLockArray = MainLWLockArray;
6154 param->ProcStructLock = ProcStructLock;
6155 param->ProcGlobal = ProcGlobal;
6156 param->AuxiliaryProcs = AuxiliaryProcs;
6157 param->PreparedXactProcs = PreparedXactProcs;
6158 param->PMSignalState = PMSignalState;
6159 if (!write_inheritable_socket(¶m->pgStatSock, pgStatSock, childPid))
6160 return false;
6161
6162 param->PostmasterPid = PostmasterPid;
6163 param->PgStartTime = PgStartTime;
6164 param->PgReloadTime = PgReloadTime;
6165 param->first_syslogger_file_time = first_syslogger_file_time;
6166
6167 param->redirection_done = redirection_done;
6168 param->IsBinaryUpgrade = IsBinaryUpgrade;
6169 param->max_safe_fds = max_safe_fds;
6170
6171 param->MaxBackends = MaxBackends;
6172
6173 #ifdef WIN32
6174 param->PostmasterHandle = PostmasterHandle;
6175 if (!write_duplicated_handle(¶m->initial_signal_pipe,
6176 pgwin32_create_signal_listener(childPid),
6177 childProcess))
6178 return false;
6179 #else
6180 memcpy(¶m->postmaster_alive_fds, &postmaster_alive_fds,
6181 sizeof(postmaster_alive_fds));
6182 #endif
6183
6184 memcpy(¶m->syslogPipe, &syslogPipe, sizeof(syslogPipe));
6185
6186 strlcpy(param->my_exec_path, my_exec_path, MAXPGPATH);
6187
6188 strlcpy(param->pkglib_path, pkglib_path, MAXPGPATH);
6189
6190 strlcpy(param->ExtraOptions, ExtraOptions, MAXPGPATH);
6191
6192 return true;
6193 }
6194
6195
6196 #ifdef WIN32
6197 /*
6198 * Duplicate a handle for usage in a child process, and write the child
6199 * process instance of the handle to the parameter file.
6200 */
6201 static bool
6202 write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE childProcess)
6203 {
6204 HANDLE hChild = INVALID_HANDLE_VALUE;
6205
6206 if (!DuplicateHandle(GetCurrentProcess(),
6207 src,
6208 childProcess,
6209 &hChild,
6210 0,
6211 TRUE,
6212 DUPLICATE_CLOSE_SOURCE | DUPLICATE_SAME_ACCESS))
6213 {
6214 ereport(LOG,
6215 (errmsg_internal("could not duplicate handle to be written to backend parameter file: error code %lu",
6216 GetLastError())));
6217 return false;
6218 }
6219
6220 *dest = hChild;
6221 return true;
6222 }
6223
6224 /*
6225 * Duplicate a socket for usage in a child process, and write the resulting
6226 * structure to the parameter file.
6227 * This is required because a number of LSPs (Layered Service Providers) very
6228 * common on Windows (antivirus, firewalls, download managers etc) break
6229 * straight socket inheritance.
6230 */
6231 static bool
6232 write_inheritable_socket(InheritableSocket *dest, SOCKET src, pid_t childpid)
6233 {
6234 dest->origsocket = src;
6235 if (src != 0 && src != PGINVALID_SOCKET)
6236 {
6237 /* Actual socket */
6238 if (WSADuplicateSocket(src, childpid, &dest->wsainfo) != 0)
6239 {
6240 ereport(LOG,
6241 (errmsg("could not duplicate socket %d for use in backend: error code %d",
6242 (int) src, WSAGetLastError())));
6243 return false;
6244 }
6245 }
6246 return true;
6247 }
6248
6249 /*
6250 * Read a duplicate socket structure back, and get the socket descriptor.
6251 */
6252 static void
6253 read_inheritable_socket(SOCKET *dest, InheritableSocket *src)
6254 {
6255 SOCKET s;
6256
6257 if (src->origsocket == PGINVALID_SOCKET || src->origsocket == 0)
6258 {
6259 /* Not a real socket! */
6260 *dest = src->origsocket;
6261 }
6262 else
6263 {
6264 /* Actual socket, so create from structure */
6265 s = WSASocket(FROM_PROTOCOL_INFO,
6266 FROM_PROTOCOL_INFO,
6267 FROM_PROTOCOL_INFO,
6268 &src->wsainfo,
6269 0,
6270 0);
6271 if (s == INVALID_SOCKET)
6272 {
6273 write_stderr("could not create inherited socket: error code %d\n",
6274 WSAGetLastError());
6275 exit(1);
6276 }
6277 *dest = s;
6278
6279 /*
6280 * To make sure we don't get two references to the same socket, close
6281 * the original one. (This would happen when inheritance actually
6282 * works..
6283 */
6284 closesocket(src->origsocket);
6285 }
6286 }
6287 #endif
6288
6289 static void
6290 read_backend_variables(char *id, Port *port)
6291 {
6292 BackendParameters param;
6293
6294 #ifndef WIN32
6295 /* Non-win32 implementation reads from file */
6296 FILE *fp;
6297
6298 /* Open file */
6299 fp = AllocateFile(id, PG_BINARY_R);
6300 if (!fp)
6301 {
6302 write_stderr("could not open backend variables file \"%s\": %s\n",
6303 id, strerror(errno));
6304 exit(1);
6305 }
6306
6307 if (fread(¶m, sizeof(param), 1, fp) != 1)
6308 {
6309 write_stderr("could not read from backend variables file \"%s\": %s\n",
6310 id, strerror(errno));
6311 exit(1);
6312 }
6313
6314 /* Release file */
6315 FreeFile(fp);
6316 if (unlink(id) != 0)
6317 {
6318 write_stderr("could not remove file \"%s\": %s\n",
6319 id, strerror(errno));
6320 exit(1);
6321 }
6322 #else
6323 /* Win32 version uses mapped file */
6324 HANDLE paramHandle;
6325 BackendParameters *paramp;
6326
6327 #ifdef _WIN64
6328 paramHandle = (HANDLE) _atoi64(id);
6329 #else
6330 paramHandle = (HANDLE) atol(id);
6331 #endif
6332 paramp = MapViewOfFile(paramHandle, FILE_MAP_READ, 0, 0, 0);
6333 if (!paramp)
6334 {
6335 write_stderr("could not map view of backend variables: error code %lu\n",
6336 GetLastError());
6337 exit(1);
6338 }
6339
6340 memcpy(¶m, paramp, sizeof(BackendParameters));
6341
6342 if (!UnmapViewOfFile(paramp))
6343 {
6344 write_stderr("could not unmap view of backend variables: error code %lu\n",
6345 GetLastError());
6346 exit(1);
6347 }
6348
6349 if (!CloseHandle(paramHandle))
6350 {
6351 write_stderr("could not close handle to backend parameter variables: error code %lu\n",
6352 GetLastError());
6353 exit(1);
6354 }
6355 #endif
6356
6357 restore_backend_variables(¶m, port);
6358 }
6359
6360 /* Restore critical backend variables from the BackendParameters struct */
6361 static void
6362 restore_backend_variables(BackendParameters *param, Port *port)
6363 {
6364 memcpy(port, ¶m->port, sizeof(Port));
6365 read_inheritable_socket(&port->sock, ¶m->portsocket);
6366
6367 SetDataDir(param->DataDir);
6368
6369 memcpy(&ListenSocket, ¶m->ListenSocket, sizeof(ListenSocket));
6370
6371 MyCancelKey = param->MyCancelKey;
6372 MyPMChildSlot = param->MyPMChildSlot;
6373
6374 #ifdef WIN32
6375 ShmemProtectiveRegion = param->ShmemProtectiveRegion;
6376 #endif
6377 UsedShmemSegID = param->UsedShmemSegID;
6378 UsedShmemSegAddr = param->UsedShmemSegAddr;
6379
6380 ShmemLock = param->ShmemLock;
6381 ShmemVariableCache = param->ShmemVariableCache;
6382 ShmemBackendArray = param->ShmemBackendArray;
6383
6384 #ifndef HAVE_SPINLOCKS
6385 SpinlockSemaArray = param->SpinlockSemaArray;
6386 #endif
6387 NamedLWLockTrancheRequests = param->NamedLWLockTrancheRequests;
6388 NamedLWLockTrancheArray = param->NamedLWLockTrancheArray;
6389 MainLWLockArray = param->MainLWLockArray;
6390 ProcStructLock = param->ProcStructLock;
6391 ProcGlobal = param->ProcGlobal;
6392 AuxiliaryProcs = param->AuxiliaryProcs;
6393 PreparedXactProcs = param->PreparedXactProcs;
6394 PMSignalState = param->PMSignalState;
6395 read_inheritable_socket(&pgStatSock, ¶m->pgStatSock);
6396
6397 PostmasterPid = param->PostmasterPid;
6398 PgStartTime = param->PgStartTime;
6399 PgReloadTime = param->PgReloadTime;
6400 first_syslogger_file_time = param->first_syslogger_file_time;
6401
6402 redirection_done = param->redirection_done;
6403 IsBinaryUpgrade = param->IsBinaryUpgrade;
6404 max_safe_fds = param->max_safe_fds;
6405
6406 MaxBackends = param->MaxBackends;
6407
6408 #ifdef WIN32
6409 PostmasterHandle = param->PostmasterHandle;
6410 pgwin32_initial_signal_pipe = param->initial_signal_pipe;
6411 #else
6412 memcpy(&postmaster_alive_fds, ¶m->postmaster_alive_fds,
6413 sizeof(postmaster_alive_fds));
6414 #endif
6415
6416 memcpy(&syslogPipe, ¶m->syslogPipe, sizeof(syslogPipe));
6417
6418 strlcpy(my_exec_path, param->my_exec_path, MAXPGPATH);
6419
6420 strlcpy(pkglib_path, param->pkglib_path, MAXPGPATH);
6421
6422 strlcpy(ExtraOptions, param->ExtraOptions, MAXPGPATH);
6423 }
6424
6425
6426 Size
6427 ShmemBackendArraySize(void)
6428 {
6429 return mul_size(MaxLivePostmasterChildren(), sizeof(Backend));
6430 }
6431
6432 void
6433 ShmemBackendArrayAllocation(void)
6434 {
6435 Size size = ShmemBackendArraySize();
6436
6437 ShmemBackendArray = (Backend *) ShmemAlloc(size);
6438 /* Mark all slots as empty */
6439 memset(ShmemBackendArray, 0, size);
6440 }
6441
6442 static void
6443 ShmemBackendArrayAdd(Backend *bn)
6444 {
6445 /* The array slot corresponding to my PMChildSlot should be free */
6446 int i = bn->child_slot - 1;
6447
6448 Assert(ShmemBackendArray[i].pid == 0);
6449 ShmemBackendArray[i] = *bn;
6450 }
6451
6452 static void
6453 ShmemBackendArrayRemove(Backend *bn)
6454 {
6455 int i = bn->child_slot - 1;
6456
6457 Assert(ShmemBackendArray[i].pid == bn->pid);
6458 /* Mark the slot as empty */
6459 ShmemBackendArray[i].pid = 0;
6460 }
6461 #endif /* EXEC_BACKEND */
6462
6463
6464 #ifdef WIN32
6465
6466 /*
6467 * Subset implementation of waitpid() for Windows. We assume pid is -1
6468 * (that is, check all child processes) and options is WNOHANG (don't wait).
6469 */
6470 static pid_t
6471 waitpid(pid_t pid, int *exitstatus, int options)
6472 {
6473 DWORD dwd;
6474 ULONG_PTR key;
6475 OVERLAPPED *ovl;
6476
6477 /*
6478 * Check if there are any dead children. If there are, return the pid of
6479 * the first one that died.
6480 */
6481 if (GetQueuedCompletionStatus(win32ChildQueue, &dwd, &key, &ovl, 0))
6482 {
6483 *exitstatus = (int) key;
6484 return dwd;
6485 }
6486
6487 return -1;
6488 }
6489
6490 /*
6491 * Note! Code below executes on a thread pool! All operations must
6492 * be thread safe! Note that elog() and friends must *not* be used.
6493 */
6494 static void WINAPI
6495 pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired)
6496 {
6497 win32_deadchild_waitinfo *childinfo = (win32_deadchild_waitinfo *) lpParameter;
6498 DWORD exitcode;
6499
6500 if (TimerOrWaitFired)
6501 return; /* timeout. Should never happen, since we use
6502 * INFINITE as timeout value. */
6503
6504 /*
6505 * Remove handle from wait - required even though it's set to wait only
6506 * once
6507 */
6508 UnregisterWaitEx(childinfo->waitHandle, NULL);
6509
6510 if (!GetExitCodeProcess(childinfo->procHandle, &exitcode))
6511 {
6512 /*
6513 * Should never happen. Inform user and set a fixed exitcode.
6514 */
6515 write_stderr("could not read exit code for process\n");
6516 exitcode = 255;
6517 }
6518
6519 if (!PostQueuedCompletionStatus(win32ChildQueue, childinfo->procId, (ULONG_PTR) exitcode, NULL))
6520 write_stderr("could not post child completion status\n");
6521
6522 /*
6523 * Handle is per-process, so we close it here instead of in the
6524 * originating thread
6525 */
6526 CloseHandle(childinfo->procHandle);
6527
6528 /*
6529 * Free struct that was allocated before the call to
6530 * RegisterWaitForSingleObject()
6531 */
6532 free(childinfo);
6533
6534 /* Queue SIGCHLD signal */
6535 pg_queue_signal(SIGCHLD);
6536 }
6537 #endif /* WIN32 */
6538
6539 /*
6540 * Initialize one and only handle for monitoring postmaster death.
6541 *
6542 * Called once in the postmaster, so that child processes can subsequently
6543 * monitor if their parent is dead.
6544 */
6545 static void
6546 InitPostmasterDeathWatchHandle(void)
6547 {
6548 #ifndef WIN32
6549
6550 /*
6551 * Create a pipe. Postmaster holds the write end of the pipe open
6552 * (POSTMASTER_FD_OWN), and children hold the read end. Children can pass
6553 * the read file descriptor to select() to wake up in case postmaster
6554 * dies, or check for postmaster death with a (read() == 0). Children must
6555 * close the write end as soon as possible after forking, because EOF
6556 * won't be signaled in the read end until all processes have closed the
6557 * write fd. That is taken care of in ClosePostmasterPorts().
6558 */
6559 Assert(MyProcPid == PostmasterPid);
6560 if (pipe(postmaster_alive_fds) < 0)
6561 ereport(FATAL,
6562 (errcode_for_file_access(),
6563 errmsg_internal("could not create pipe to monitor postmaster death: %m")));
6564
6565 /*
6566 * Set O_NONBLOCK to allow testing for the fd's presence with a read()
6567 * call.
6568 */
6569 if (fcntl(postmaster_alive_fds[POSTMASTER_FD_WATCH], F_SETFL, O_NONBLOCK) == -1)
6570 ereport(FATAL,
6571 (errcode_for_socket_access(),
6572 errmsg_internal("could not set postmaster death monitoring pipe to nonblocking mode: %m")));
6573 #else
6574
6575 /*
6576 * On Windows, we use a process handle for the same purpose.
6577 */
6578 if (DuplicateHandle(GetCurrentProcess(),
6579 GetCurrentProcess(),
6580 GetCurrentProcess(),
6581 &PostmasterHandle,
6582 0,
6583 TRUE,
6584 DUPLICATE_SAME_ACCESS) == 0)
6585 ereport(FATAL,
6586 (errmsg_internal("could not duplicate postmaster handle: error code %lu",
6587 GetLastError())));
6588 #endif /* WIN32 */
6589 }
6590