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) 1996-2019, 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 "common/string.h"
103 #include "lib/ilist.h"
104 #include "libpq/auth.h"
105 #include "libpq/libpq.h"
106 #include "libpq/pqformat.h"
107 #include "libpq/pqsignal.h"
108 #include "miscadmin.h"
109 #include "pg_getopt.h"
110 #include "pgstat.h"
111 #include "port/pg_bswap.h"
112 #include "postmaster/autovacuum.h"
113 #include "postmaster/bgworker_internals.h"
114 #include "postmaster/fork_process.h"
115 #include "postmaster/pgarch.h"
116 #include "postmaster/postmaster.h"
117 #include "postmaster/syslogger.h"
118 #include "replication/logicallauncher.h"
119 #include "replication/walsender.h"
120 #include "storage/fd.h"
121 #include "storage/ipc.h"
122 #include "storage/pg_shmem.h"
123 #include "storage/pmsignal.h"
124 #include "storage/proc.h"
125 #include "tcop/tcopprot.h"
126 #include "utils/builtins.h"
127 #include "utils/datetime.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/timestamp.h"
133 #include "utils/varlena.h"
134
135 #ifdef EXEC_BACKEND
136 #include "storage/spin.h"
137 #endif
138
139
140 /*
141 * Possible types of a backend. Beyond being the possible bkend_type values in
142 * struct bkend, these are OR-able request flag bits for SignalSomeChildren()
143 * and CountChildren().
144 */
145 #define BACKEND_TYPE_NORMAL 0x0001 /* normal backend */
146 #define BACKEND_TYPE_AUTOVAC 0x0002 /* autovacuum worker process */
147 #define BACKEND_TYPE_WALSND 0x0004 /* walsender process */
148 #define BACKEND_TYPE_BGWORKER 0x0008 /* bgworker process */
149 #define BACKEND_TYPE_ALL 0x000F /* OR of all the above */
150
151 /*
152 * List of active backends (or child processes anyway; we don't actually
153 * know whether a given child has become a backend or is still in the
154 * authorization phase). This is used mainly to keep track of how many
155 * children we have and send them appropriate signals when necessary.
156 *
157 * "Special" children such as the startup, bgwriter and autovacuum launcher
158 * tasks are not in this list. Autovacuum worker and walsender are in it.
159 * Also, "dead_end" children are in it: these are children launched just for
160 * the purpose of sending a friendly rejection message to a would-be client.
161 * We must track them because they are attached to shared memory, but we know
162 * they will never become live backends. dead_end children are not assigned a
163 * PMChildSlot.
164 *
165 * Background workers are in this list, too.
166 */
167 typedef struct bkend
168 {
169 pid_t pid; /* process id of backend */
170 int32 cancel_key; /* cancel key for cancels for this backend */
171 int child_slot; /* PMChildSlot for this backend, if any */
172
173 /*
174 * Flavor of backend or auxiliary process. Note that BACKEND_TYPE_WALSND
175 * backends initially announce themselves as BACKEND_TYPE_NORMAL, so if
176 * bkend_type is normal, you should check for a recent transition.
177 */
178 int bkend_type;
179 bool dead_end; /* is it going to send an error and quit? */
180 bool bgworker_notify; /* gets bgworker start/stop notifications */
181 dlist_node elem; /* list link in BackendList */
182 } Backend;
183
184 static dlist_head BackendList = DLIST_STATIC_INIT(BackendList);
185
186 #ifdef EXEC_BACKEND
187 static Backend *ShmemBackendArray;
188 #endif
189
190 BackgroundWorker *MyBgworkerEntry = NULL;
191
192
193
194 /* The socket number we are listening for connections on */
195 int PostPortNumber;
196
197 /* The directory names for Unix socket(s) */
198 char *Unix_socket_directories;
199
200 /* The TCP listen address(es) */
201 char *ListenAddresses;
202
203 /*
204 * ReservedBackends is the number of backends reserved for superuser use.
205 * This number is taken out of the pool size given by MaxConnections so
206 * number of backend slots available to non-superusers is
207 * (MaxConnections - ReservedBackends). Note what this really means is
208 * "if there are <= ReservedBackends connections available, only superusers
209 * can make new connections" --- pre-existing superuser connections don't
210 * count against the limit.
211 */
212 int ReservedBackends;
213
214 /* The socket(s) we're listening to. */
215 #define MAXLISTEN 64
216 static pgsocket ListenSocket[MAXLISTEN];
217
218 /*
219 * Set by the -o option
220 */
221 static char ExtraOptions[MAXPGPATH];
222
223 /*
224 * These globals control the behavior of the postmaster in case some
225 * backend dumps core. Normally, it kills all peers of the dead backend
226 * and reinitializes shared memory. By specifying -s or -n, we can have
227 * the postmaster stop (rather than kill) peers and not reinitialize
228 * shared data structures. (Reinit is currently dead code, though.)
229 */
230 static bool Reinit = true;
231 static int SendStop = false;
232
233 /* still more option variables */
234 bool EnableSSL = false;
235
236 int PreAuthDelay = 0;
237 int AuthenticationTimeout = 60;
238
239 bool log_hostname; /* for ps display and logging */
240 bool Log_connections = false;
241 bool Db_user_namespace = false;
242
243 bool enable_bonjour = false;
244 char *bonjour_name;
245 bool restart_after_crash = true;
246
247 /* PIDs of special child processes; 0 when not running */
248 static pid_t StartupPID = 0,
249 BgWriterPID = 0,
250 CheckpointerPID = 0,
251 WalWriterPID = 0,
252 WalReceiverPID = 0,
253 AutoVacPID = 0,
254 PgArchPID = 0,
255 PgStatPID = 0,
256 SysLoggerPID = 0;
257
258 /* Startup process's status */
259 typedef enum
260 {
261 STARTUP_NOT_RUNNING,
262 STARTUP_RUNNING,
263 STARTUP_SIGNALED, /* we sent it a SIGQUIT or SIGKILL */
264 STARTUP_CRASHED
265 } StartupStatusEnum;
266
267 static StartupStatusEnum StartupStatus = STARTUP_NOT_RUNNING;
268
269 /* Startup/shutdown state */
270 #define NoShutdown 0
271 #define SmartShutdown 1
272 #define FastShutdown 2
273 #define ImmediateShutdown 3
274
275 static int Shutdown = NoShutdown;
276
277 static bool FatalError = false; /* T if recovering from backend crash */
278
279 /*
280 * We use a simple state machine to control startup, shutdown, and
281 * crash recovery (which is rather like shutdown followed by startup).
282 *
283 * After doing all the postmaster initialization work, we enter PM_STARTUP
284 * state and the startup process is launched. The startup process begins by
285 * reading the control file and other preliminary initialization steps.
286 * In a normal startup, or after crash recovery, the startup process exits
287 * with exit code 0 and we switch to PM_RUN state. However, archive recovery
288 * is handled specially since it takes much longer and we would like to support
289 * hot standby during archive recovery.
290 *
291 * When the startup process is ready to start archive recovery, it signals the
292 * postmaster, and we switch to PM_RECOVERY state. The background writer and
293 * checkpointer are launched, while the startup process continues applying WAL.
294 * If Hot Standby is enabled, then, after reaching a consistent point in WAL
295 * redo, startup process signals us again, and we switch to PM_HOT_STANDBY
296 * state and begin accepting connections to perform read-only queries. When
297 * archive recovery is finished, the startup process exits with exit code 0
298 * and we switch to PM_RUN state.
299 *
300 * Normal child backends can only be launched when we are in PM_RUN or
301 * PM_HOT_STANDBY state. (connsAllowed can also restrict launching.)
302 * In other states we handle connection requests by launching "dead_end"
303 * child processes, which will simply send the client an error message and
304 * quit. (We track these in the BackendList so that we can know when they
305 * are all gone; this is important because they're still connected to shared
306 * memory, and would interfere with an attempt to destroy the shmem segment,
307 * possibly leading to SHMALL failure when we try to make a new one.)
308 * In PM_WAIT_DEAD_END state we are waiting for all the dead_end children
309 * to drain out of the system, and therefore stop accepting connection
310 * requests at all until the last existing child has quit (which hopefully
311 * will not be very long).
312 *
313 * Notice that this state variable does not distinguish *why* we entered
314 * states later than PM_RUN --- Shutdown and FatalError must be consulted
315 * to find that out. FatalError is never true in PM_RECOVERY, PM_HOT_STANDBY,
316 * or PM_RUN states, nor in PM_SHUTDOWN states (because we don't enter those
317 * states when trying to recover from a crash). It can be true in PM_STARTUP
318 * state, because we don't clear it until we've successfully started WAL redo.
319 */
320 typedef enum
321 {
322 PM_INIT, /* postmaster starting */
323 PM_STARTUP, /* waiting for startup subprocess */
324 PM_RECOVERY, /* in archive recovery mode */
325 PM_HOT_STANDBY, /* in hot standby mode */
326 PM_RUN, /* normal "database is alive" state */
327 PM_STOP_BACKENDS, /* need to stop remaining backends */
328 PM_WAIT_BACKENDS, /* waiting for live backends to exit */
329 PM_SHUTDOWN, /* waiting for checkpointer to do shutdown
330 * ckpt */
331 PM_SHUTDOWN_2, /* waiting for archiver and walsenders to
332 * finish */
333 PM_WAIT_DEAD_END, /* waiting for dead_end children to exit */
334 PM_NO_CHILDREN /* all important children have exited */
335 } PMState;
336
337 static PMState pmState = PM_INIT;
338
339 /*
340 * While performing a "smart shutdown", we restrict new connections but stay
341 * in PM_RUN or PM_HOT_STANDBY state until all the client backends are gone.
342 * connsAllowed is a sub-state indicator showing the active restriction.
343 * It is of no interest unless pmState is PM_RUN or PM_HOT_STANDBY.
344 */
345 typedef enum
346 {
347 ALLOW_ALL_CONNS, /* normal not-shutting-down state */
348 ALLOW_SUPERUSER_CONNS, /* only superusers can connect */
349 ALLOW_NO_CONNS /* no new connections allowed, period */
350 } ConnsAllowedState;
351
352 static ConnsAllowedState connsAllowed = ALLOW_ALL_CONNS;
353
354 /* Start time of SIGKILL timeout during immediate shutdown or child crash */
355 /* Zero means timeout is not running */
356 static time_t AbortStartTime = 0;
357
358 /* Length of said timeout */
359 #define SIGKILL_CHILDREN_AFTER_SECS 5
360
361 static bool ReachedNormalRunning = false; /* T if we've reached PM_RUN */
362
363 bool ClientAuthInProgress = false; /* T during new-client
364 * authentication */
365
366 bool redirection_done = false; /* stderr redirected for syslogger? */
367
368 /* received START_AUTOVAC_LAUNCHER signal */
369 static volatile sig_atomic_t start_autovac_launcher = false;
370
371 /* the launcher needs to be signalled to communicate some condition */
372 static volatile bool avlauncher_needs_signal = false;
373
374 /* received START_WALRECEIVER signal */
375 static volatile sig_atomic_t WalReceiverRequested = false;
376
377 /* set when there's a worker that needs to be started up */
378 static volatile bool StartWorkerNeeded = true;
379 static volatile bool HaveCrashedWorker = false;
380
381 #ifdef USE_SSL
382 /* Set when and if SSL has been initialized properly */
383 static bool LoadedSSL = false;
384 #endif
385
386 #ifdef USE_BONJOUR
387 static DNSServiceRef bonjour_sdref = NULL;
388 #endif
389
390 /*
391 * postmaster.c - function prototypes
392 */
393 static void CloseServerPorts(int status, Datum arg);
394 static void unlink_external_pid_file(int status, Datum arg);
395 static void getInstallationPaths(const char *argv0);
396 static void checkControlFile(void);
397 static Port *ConnCreate(int serverFd);
398 static void ConnFree(Port *port);
399 static void reset_shared(int port);
400 static void SIGHUP_handler(SIGNAL_ARGS);
401 static void pmdie(SIGNAL_ARGS);
402 static void reaper(SIGNAL_ARGS);
403 static void sigusr1_handler(SIGNAL_ARGS);
404 static void process_startup_packet_die(SIGNAL_ARGS);
405 static void process_startup_packet_quickdie(SIGNAL_ARGS);
406 static void dummy_handler(SIGNAL_ARGS);
407 static void StartupPacketTimeoutHandler(void);
408 static void CleanupBackend(int pid, int exitstatus);
409 static bool CleanupBackgroundWorker(int pid, int exitstatus);
410 static void HandleChildCrash(int pid, int exitstatus, const char *procname);
411 static void LogChildExit(int lev, const char *procname,
412 int pid, int exitstatus);
413 static void PostmasterStateMachine(void);
414 static void BackendInitialize(Port *port);
415 static void BackendRun(Port *port) pg_attribute_noreturn();
416 static void ExitPostmaster(int status) pg_attribute_noreturn();
417 static int ServerLoop(void);
418 static int BackendStartup(Port *port);
419 static int ProcessStartupPacket(Port *port, bool ssl_done, bool gss_done);
420 static void SendNegotiateProtocolVersion(List *unrecognized_protocol_options);
421 static void processCancelRequest(Port *port, void *pkt);
422 static int initMasks(fd_set *rmask);
423 static void report_fork_failure_to_client(Port *port, int errnum);
424 static CAC_state canAcceptConnections(int backend_type);
425 static bool RandomCancelKey(int32 *cancel_key);
426 static void signal_child(pid_t pid, int signal);
427 static bool SignalSomeChildren(int signal, int targets);
428 static void TerminateChildren(int signal);
429
430 #define SignalChildren(sig) SignalSomeChildren(sig, BACKEND_TYPE_ALL)
431
432 static int CountChildren(int target);
433 static bool assign_backendlist_entry(RegisteredBgWorker *rw);
434 static void maybe_start_bgworkers(void);
435 static bool CreateOptsFile(int argc, char *argv[], char *fullprogname);
436 static pid_t StartChildProcess(AuxProcType type);
437 static void StartAutovacuumWorker(void);
438 static void MaybeStartWalReceiver(void);
439 static void InitPostmasterDeathWatchHandle(void);
440
441 /*
442 * Archiver is allowed to start up at the current postmaster state?
443 *
444 * If WAL archiving is enabled always, we are allowed to start archiver
445 * even during recovery.
446 */
447 #define PgArchStartupAllowed() \
448 ((XLogArchivingActive() && pmState == PM_RUN) || \
449 (XLogArchivingAlways() && \
450 (pmState == PM_RECOVERY || pmState == PM_HOT_STANDBY)))
451
452 #ifdef EXEC_BACKEND
453
454 #ifdef WIN32
455 #define WNOHANG 0 /* ignored, so any integer value will do */
456
457 static pid_t waitpid(pid_t pid, int *exitstatus, int options);
458 static void WINAPI pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired);
459
460 static HANDLE win32ChildQueue;
461
462 typedef struct
463 {
464 HANDLE waitHandle;
465 HANDLE procHandle;
466 DWORD procId;
467 } win32_deadchild_waitinfo;
468 #endif /* WIN32 */
469
470 static pid_t backend_forkexec(Port *port);
471 static pid_t internal_forkexec(int argc, char *argv[], Port *port);
472
473 /* Type for a socket that can be inherited to a client process */
474 #ifdef WIN32
475 typedef struct
476 {
477 SOCKET origsocket; /* Original socket value, or PGINVALID_SOCKET
478 * if not a socket */
479 WSAPROTOCOL_INFO wsainfo;
480 } InheritableSocket;
481 #else
482 typedef int InheritableSocket;
483 #endif
484
485 /*
486 * Structure contains all variables passed to exec:ed backends
487 */
488 typedef struct
489 {
490 Port port;
491 InheritableSocket portsocket;
492 char DataDir[MAXPGPATH];
493 pgsocket ListenSocket[MAXLISTEN];
494 int32 MyCancelKey;
495 int MyPMChildSlot;
496 #ifndef WIN32
497 unsigned long UsedShmemSegID;
498 #else
499 void *ShmemProtectiveRegion;
500 HANDLE UsedShmemSegID;
501 #endif
502 void *UsedShmemSegAddr;
503 slock_t *ShmemLock;
504 VariableCache ShmemVariableCache;
505 Backend *ShmemBackendArray;
506 #ifndef HAVE_SPINLOCKS
507 PGSemaphore *SpinlockSemaArray;
508 #endif
509 int NamedLWLockTrancheRequests;
510 NamedLWLockTranche *NamedLWLockTrancheArray;
511 LWLockPadded *MainLWLockArray;
512 slock_t *ProcStructLock;
513 PROC_HDR *ProcGlobal;
514 PGPROC *AuxiliaryProcs;
515 PGPROC *PreparedXactProcs;
516 PMSignalData *PMSignalState;
517 InheritableSocket pgStatSock;
518 pid_t PostmasterPid;
519 TimestampTz PgStartTime;
520 TimestampTz PgReloadTime;
521 pg_time_t first_syslogger_file_time;
522 bool redirection_done;
523 bool IsBinaryUpgrade;
524 int max_safe_fds;
525 int MaxBackends;
526 #ifdef WIN32
527 HANDLE PostmasterHandle;
528 HANDLE initial_signal_pipe;
529 HANDLE syslogPipe[2];
530 #else
531 int postmaster_alive_fds[2];
532 int syslogPipe[2];
533 #endif
534 char my_exec_path[MAXPGPATH];
535 char pkglib_path[MAXPGPATH];
536 char ExtraOptions[MAXPGPATH];
537 } BackendParameters;
538
539 static void read_backend_variables(char *id, Port *port);
540 static void restore_backend_variables(BackendParameters *param, Port *port);
541
542 #ifndef WIN32
543 static bool save_backend_variables(BackendParameters *param, Port *port);
544 #else
545 static bool save_backend_variables(BackendParameters *param, Port *port,
546 HANDLE childProcess, pid_t childPid);
547 #endif
548
549 static void ShmemBackendArrayAdd(Backend *bn);
550 static void ShmemBackendArrayRemove(Backend *bn);
551 #endif /* EXEC_BACKEND */
552
553 #define StartupDataBase() StartChildProcess(StartupProcess)
554 #define StartBackgroundWriter() StartChildProcess(BgWriterProcess)
555 #define StartCheckpointer() StartChildProcess(CheckpointerProcess)
556 #define StartWalWriter() StartChildProcess(WalWriterProcess)
557 #define StartWalReceiver() StartChildProcess(WalReceiverProcess)
558
559 /* Macros to check exit status of a child process */
560 #define EXIT_STATUS_0(st) ((st) == 0)
561 #define EXIT_STATUS_1(st) (WIFEXITED(st) && WEXITSTATUS(st) == 1)
562 #define EXIT_STATUS_3(st) (WIFEXITED(st) && WEXITSTATUS(st) == 3)
563
564 #ifndef WIN32
565 /*
566 * File descriptors for pipe used to monitor if postmaster is alive.
567 * First is POSTMASTER_FD_WATCH, second is POSTMASTER_FD_OWN.
568 */
569 int postmaster_alive_fds[2] = {-1, -1};
570 #else
571 /* Process handle of postmaster used for the same purpose on Windows */
572 HANDLE PostmasterHandle;
573 #endif
574
575 /*
576 * Postmaster main entry point
577 */
578 void
PostmasterMain(int argc,char * argv[])579 PostmasterMain(int argc, char *argv[])
580 {
581 int opt;
582 int status;
583 char *userDoption = NULL;
584 bool listen_addr_saved = false;
585 int i;
586 char *output_config_variable = NULL;
587
588 InitProcessGlobals();
589
590 PostmasterPid = MyProcPid;
591
592 IsPostmasterEnvironment = true;
593
594 /*
595 * We should not be creating any files or directories before we check the
596 * data directory (see checkDataDir()), but just in case set the umask to
597 * the most restrictive (owner-only) permissions.
598 *
599 * checkDataDir() will reset the umask based on the data directory
600 * permissions.
601 */
602 umask(PG_MODE_MASK_OWNER);
603
604 /*
605 * By default, palloc() requests in the postmaster will be allocated in
606 * the PostmasterContext, which is space that can be recycled by backends.
607 * Allocated data that needs to be available to backends should be
608 * allocated in TopMemoryContext.
609 */
610 PostmasterContext = AllocSetContextCreate(TopMemoryContext,
611 "Postmaster",
612 ALLOCSET_DEFAULT_SIZES);
613 MemoryContextSwitchTo(PostmasterContext);
614
615 /* Initialize paths to installation files */
616 getInstallationPaths(argv[0]);
617
618 /*
619 * Set up signal handlers for the postmaster process.
620 *
621 * In the postmaster, we use pqsignal_pm() rather than pqsignal() (which
622 * is used by all child processes and client processes). That has a
623 * couple of special behaviors:
624 *
625 * 1. Except on Windows, we tell sigaction() to block all signals for the
626 * duration of the signal handler. This is faster than our old approach
627 * of blocking/unblocking explicitly in the signal handler, and it should
628 * also prevent excessive stack consumption if signals arrive quickly.
629 *
630 * 2. We do not set the SA_RESTART flag. This is because signals will be
631 * blocked at all times except when ServerLoop is waiting for something to
632 * happen, and during that window, we want signals to exit the select(2)
633 * wait so that ServerLoop can respond if anything interesting happened.
634 * On some platforms, signals marked SA_RESTART would not cause the
635 * select() wait to end.
636 *
637 * Child processes will generally want SA_RESTART, so pqsignal() sets that
638 * flag. We expect children to set up their own handlers before
639 * unblocking signals.
640 *
641 * CAUTION: when changing this list, check for side-effects on the signal
642 * handling setup of child processes. See tcop/postgres.c,
643 * bootstrap/bootstrap.c, postmaster/bgwriter.c, postmaster/walwriter.c,
644 * postmaster/autovacuum.c, postmaster/pgarch.c, postmaster/pgstat.c,
645 * postmaster/syslogger.c, postmaster/bgworker.c and
646 * postmaster/checkpointer.c.
647 */
648 pqinitmask();
649 PG_SETMASK(&BlockSig);
650
651 pqsignal_pm(SIGHUP, SIGHUP_handler); /* reread config file and have
652 * children do same */
653 pqsignal_pm(SIGINT, pmdie); /* send SIGTERM and shut down */
654 pqsignal_pm(SIGQUIT, pmdie); /* send SIGQUIT and die */
655 pqsignal_pm(SIGTERM, pmdie); /* wait for children and shut down */
656 pqsignal_pm(SIGALRM, SIG_IGN); /* ignored */
657 pqsignal_pm(SIGPIPE, SIG_IGN); /* ignored */
658 pqsignal_pm(SIGUSR1, sigusr1_handler); /* message from child process */
659 pqsignal_pm(SIGUSR2, dummy_handler); /* unused, reserve for children */
660 pqsignal_pm(SIGCHLD, reaper); /* handle child termination */
661
662 /*
663 * No other place in Postgres should touch SIGTTIN/SIGTTOU handling. We
664 * ignore those signals in a postmaster environment, so that there is no
665 * risk of a child process freezing up due to writing to stderr. But for
666 * a standalone backend, their default handling is reasonable. Hence, all
667 * child processes should just allow the inherited settings to stand.
668 */
669 #ifdef SIGTTIN
670 pqsignal_pm(SIGTTIN, SIG_IGN); /* ignored */
671 #endif
672 #ifdef SIGTTOU
673 pqsignal_pm(SIGTTOU, SIG_IGN); /* ignored */
674 #endif
675
676 /* ignore SIGXFSZ, so that ulimit violations work like disk full */
677 #ifdef SIGXFSZ
678 pqsignal_pm(SIGXFSZ, SIG_IGN); /* ignored */
679 #endif
680
681 /*
682 * Options setup
683 */
684 InitializeGUCOptions();
685
686 opterr = 1;
687
688 /*
689 * Parse command-line options. CAUTION: keep this in sync with
690 * tcop/postgres.c (the option sets should not conflict) and with the
691 * common help() function in main/main.c.
692 */
693 while ((opt = getopt(argc, argv, "B:bc:C:D:d:EeFf:h:ijk:lN:nOo:Pp:r:S:sTt:W:-:")) != -1)
694 {
695 switch (opt)
696 {
697 case 'B':
698 SetConfigOption("shared_buffers", optarg, PGC_POSTMASTER, PGC_S_ARGV);
699 break;
700
701 case 'b':
702 /* Undocumented flag used for binary upgrades */
703 IsBinaryUpgrade = true;
704 break;
705
706 case 'C':
707 output_config_variable = strdup(optarg);
708 break;
709
710 case 'D':
711 userDoption = strdup(optarg);
712 break;
713
714 case 'd':
715 set_debug_options(atoi(optarg), PGC_POSTMASTER, PGC_S_ARGV);
716 break;
717
718 case 'E':
719 SetConfigOption("log_statement", "all", PGC_POSTMASTER, PGC_S_ARGV);
720 break;
721
722 case 'e':
723 SetConfigOption("datestyle", "euro", PGC_POSTMASTER, PGC_S_ARGV);
724 break;
725
726 case 'F':
727 SetConfigOption("fsync", "false", PGC_POSTMASTER, PGC_S_ARGV);
728 break;
729
730 case 'f':
731 if (!set_plan_disabling_options(optarg, PGC_POSTMASTER, PGC_S_ARGV))
732 {
733 write_stderr("%s: invalid argument for option -f: \"%s\"\n",
734 progname, optarg);
735 ExitPostmaster(1);
736 }
737 break;
738
739 case 'h':
740 SetConfigOption("listen_addresses", optarg, PGC_POSTMASTER, PGC_S_ARGV);
741 break;
742
743 case 'i':
744 SetConfigOption("listen_addresses", "*", PGC_POSTMASTER, PGC_S_ARGV);
745 break;
746
747 case 'j':
748 /* only used by interactive backend */
749 break;
750
751 case 'k':
752 SetConfigOption("unix_socket_directories", optarg, PGC_POSTMASTER, PGC_S_ARGV);
753 break;
754
755 case 'l':
756 SetConfigOption("ssl", "true", PGC_POSTMASTER, PGC_S_ARGV);
757 break;
758
759 case 'N':
760 SetConfigOption("max_connections", optarg, PGC_POSTMASTER, PGC_S_ARGV);
761 break;
762
763 case 'n':
764 /* Don't reinit shared mem after abnormal exit */
765 Reinit = false;
766 break;
767
768 case 'O':
769 SetConfigOption("allow_system_table_mods", "true", PGC_POSTMASTER, PGC_S_ARGV);
770 break;
771
772 case 'o':
773 /* Other options to pass to the backend on the command line */
774 snprintf(ExtraOptions + strlen(ExtraOptions),
775 sizeof(ExtraOptions) - strlen(ExtraOptions),
776 " %s", optarg);
777 break;
778
779 case 'P':
780 SetConfigOption("ignore_system_indexes", "true", PGC_POSTMASTER, PGC_S_ARGV);
781 break;
782
783 case 'p':
784 SetConfigOption("port", optarg, PGC_POSTMASTER, PGC_S_ARGV);
785 break;
786
787 case 'r':
788 /* only used by single-user backend */
789 break;
790
791 case 'S':
792 SetConfigOption("work_mem", optarg, PGC_POSTMASTER, PGC_S_ARGV);
793 break;
794
795 case 's':
796 SetConfigOption("log_statement_stats", "true", PGC_POSTMASTER, PGC_S_ARGV);
797 break;
798
799 case 'T':
800
801 /*
802 * In the event that some backend dumps core, send SIGSTOP,
803 * rather than SIGQUIT, to all its peers. This lets the wily
804 * post_hacker collect core dumps from everyone.
805 */
806 SendStop = true;
807 break;
808
809 case 't':
810 {
811 const char *tmp = get_stats_option_name(optarg);
812
813 if (tmp)
814 {
815 SetConfigOption(tmp, "true", PGC_POSTMASTER, PGC_S_ARGV);
816 }
817 else
818 {
819 write_stderr("%s: invalid argument for option -t: \"%s\"\n",
820 progname, optarg);
821 ExitPostmaster(1);
822 }
823 break;
824 }
825
826 case 'W':
827 SetConfigOption("post_auth_delay", optarg, PGC_POSTMASTER, PGC_S_ARGV);
828 break;
829
830 case 'c':
831 case '-':
832 {
833 char *name,
834 *value;
835
836 ParseLongOption(optarg, &name, &value);
837 if (!value)
838 {
839 if (opt == '-')
840 ereport(ERROR,
841 (errcode(ERRCODE_SYNTAX_ERROR),
842 errmsg("--%s requires a value",
843 optarg)));
844 else
845 ereport(ERROR,
846 (errcode(ERRCODE_SYNTAX_ERROR),
847 errmsg("-c %s requires a value",
848 optarg)));
849 }
850
851 SetConfigOption(name, value, PGC_POSTMASTER, PGC_S_ARGV);
852 free(name);
853 if (value)
854 free(value);
855 break;
856 }
857
858 default:
859 write_stderr("Try \"%s --help\" for more information.\n",
860 progname);
861 ExitPostmaster(1);
862 }
863 }
864
865 /*
866 * Postmaster accepts no non-option switch arguments.
867 */
868 if (optind < argc)
869 {
870 write_stderr("%s: invalid argument: \"%s\"\n",
871 progname, argv[optind]);
872 write_stderr("Try \"%s --help\" for more information.\n",
873 progname);
874 ExitPostmaster(1);
875 }
876
877 /*
878 * Locate the proper configuration files and data directory, and read
879 * postgresql.conf for the first time.
880 */
881 if (!SelectConfigFiles(userDoption, progname))
882 ExitPostmaster(2);
883
884 if (output_config_variable != NULL)
885 {
886 /*
887 * "-C guc" was specified, so print GUC's value and exit. No extra
888 * permission check is needed because the user is reading inside the
889 * data dir.
890 */
891 const char *config_val = GetConfigOption(output_config_variable,
892 false, false);
893
894 puts(config_val ? config_val : "");
895 ExitPostmaster(0);
896 }
897
898 /* Verify that DataDir looks reasonable */
899 checkDataDir();
900
901 /* Check that pg_control exists */
902 checkControlFile();
903
904 /* And switch working directory into it */
905 ChangeToDataDir();
906
907 /*
908 * Check for invalid combinations of GUC settings.
909 */
910 if (ReservedBackends >= MaxConnections)
911 {
912 write_stderr("%s: superuser_reserved_connections (%d) must be less than max_connections (%d)\n",
913 progname,
914 ReservedBackends, MaxConnections);
915 ExitPostmaster(1);
916 }
917 if (XLogArchiveMode > ARCHIVE_MODE_OFF && wal_level == WAL_LEVEL_MINIMAL)
918 ereport(ERROR,
919 (errmsg("WAL archival cannot be enabled when wal_level is \"minimal\"")));
920 if (max_wal_senders > 0 && wal_level == WAL_LEVEL_MINIMAL)
921 ereport(ERROR,
922 (errmsg("WAL streaming (max_wal_senders > 0) requires wal_level \"replica\" or \"logical\"")));
923
924 /*
925 * Other one-time internal sanity checks can go here, if they are fast.
926 * (Put any slow processing further down, after postmaster.pid creation.)
927 */
928 if (!CheckDateTokenTables())
929 {
930 write_stderr("%s: invalid datetoken tables, please fix\n", progname);
931 ExitPostmaster(1);
932 }
933
934 /*
935 * Now that we are done processing the postmaster arguments, reset
936 * getopt(3) library so that it will work correctly in subprocesses.
937 */
938 optind = 1;
939 #ifdef HAVE_INT_OPTRESET
940 optreset = 1; /* some systems need this too */
941 #endif
942
943 /* For debugging: display postmaster environment */
944 {
945 extern char **environ;
946 char **p;
947
948 ereport(DEBUG3,
949 (errmsg_internal("%s: PostmasterMain: initial environment dump:",
950 progname)));
951 ereport(DEBUG3,
952 (errmsg_internal("-----------------------------------------")));
953 for (p = environ; *p; ++p)
954 ereport(DEBUG3,
955 (errmsg_internal("\t%s", *p)));
956 ereport(DEBUG3,
957 (errmsg_internal("-----------------------------------------")));
958 }
959
960 /*
961 * Create lockfile for data directory.
962 *
963 * We want to do this before we try to grab the input sockets, because the
964 * data directory interlock is more reliable than the socket-file
965 * interlock (thanks to whoever decided to put socket files in /tmp :-().
966 * For the same reason, it's best to grab the TCP socket(s) before the
967 * Unix socket(s).
968 *
969 * Also note that this internally sets up the on_proc_exit function that
970 * is responsible for removing both data directory and socket lockfiles;
971 * so it must happen before opening sockets so that at exit, the socket
972 * lockfiles go away after CloseServerPorts runs.
973 */
974 CreateDataDirLockFile(true);
975
976 /*
977 * Read the control file (for error checking and config info).
978 *
979 * Since we verify the control file's CRC, this has a useful side effect
980 * on machines where we need a run-time test for CRC support instructions.
981 * The postmaster will do the test once at startup, and then its child
982 * processes will inherit the correct function pointer and not need to
983 * repeat the test.
984 */
985 LocalProcessControlFile(false);
986
987 /*
988 * Initialize SSL library, if specified.
989 */
990 #ifdef USE_SSL
991 if (EnableSSL)
992 {
993 (void) secure_initialize(true);
994 LoadedSSL = true;
995 }
996 #endif
997
998 /*
999 * Register the apply launcher. Since it registers a background worker,
1000 * it needs to be called before InitializeMaxBackends(), and it's probably
1001 * a good idea to call it before any modules had chance to take the
1002 * background worker slots.
1003 */
1004 ApplyLauncherRegister();
1005
1006 /*
1007 * process any libraries that should be preloaded at postmaster start
1008 */
1009 process_shared_preload_libraries();
1010
1011 /*
1012 * Now that loadable modules have had their chance to register background
1013 * workers, calculate MaxBackends.
1014 */
1015 InitializeMaxBackends();
1016
1017 /* Report server startup in log */
1018 ereport(LOG,
1019 (errmsg("starting %s", PG_VERSION_STR)));
1020
1021 /*
1022 * Establish input sockets.
1023 *
1024 * First, mark them all closed, and set up an on_proc_exit function that's
1025 * charged with closing the sockets again at postmaster shutdown.
1026 */
1027 for (i = 0; i < MAXLISTEN; i++)
1028 ListenSocket[i] = PGINVALID_SOCKET;
1029
1030 on_proc_exit(CloseServerPorts, 0);
1031
1032 if (ListenAddresses)
1033 {
1034 char *rawstring;
1035 List *elemlist;
1036 ListCell *l;
1037 int success = 0;
1038
1039 /* Need a modifiable copy of ListenAddresses */
1040 rawstring = pstrdup(ListenAddresses);
1041
1042 /* Parse string into list of hostnames */
1043 if (!SplitIdentifierString(rawstring, ',', &elemlist))
1044 {
1045 /* syntax error in list */
1046 ereport(FATAL,
1047 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1048 errmsg("invalid list syntax in parameter \"%s\"",
1049 "listen_addresses")));
1050 }
1051
1052 foreach(l, elemlist)
1053 {
1054 char *curhost = (char *) lfirst(l);
1055
1056 if (strcmp(curhost, "*") == 0)
1057 status = StreamServerPort(AF_UNSPEC, NULL,
1058 (unsigned short) PostPortNumber,
1059 NULL,
1060 ListenSocket, MAXLISTEN);
1061 else
1062 status = StreamServerPort(AF_UNSPEC, curhost,
1063 (unsigned short) PostPortNumber,
1064 NULL,
1065 ListenSocket, MAXLISTEN);
1066
1067 if (status == STATUS_OK)
1068 {
1069 success++;
1070 /* record the first successful host addr in lockfile */
1071 if (!listen_addr_saved)
1072 {
1073 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, curhost);
1074 listen_addr_saved = true;
1075 }
1076 }
1077 else
1078 ereport(WARNING,
1079 (errmsg("could not create listen socket for \"%s\"",
1080 curhost)));
1081 }
1082
1083 if (!success && elemlist != NIL)
1084 ereport(FATAL,
1085 (errmsg("could not create any TCP/IP sockets")));
1086
1087 list_free(elemlist);
1088 pfree(rawstring);
1089 }
1090
1091 #ifdef USE_BONJOUR
1092 /* Register for Bonjour only if we opened TCP socket(s) */
1093 if (enable_bonjour && ListenSocket[0] != PGINVALID_SOCKET)
1094 {
1095 DNSServiceErrorType err;
1096
1097 /*
1098 * We pass 0 for interface_index, which will result in registering on
1099 * all "applicable" interfaces. It's not entirely clear from the
1100 * DNS-SD docs whether this would be appropriate if we have bound to
1101 * just a subset of the available network interfaces.
1102 */
1103 err = DNSServiceRegister(&bonjour_sdref,
1104 0,
1105 0,
1106 bonjour_name,
1107 "_postgresql._tcp.",
1108 NULL,
1109 NULL,
1110 pg_hton16(PostPortNumber),
1111 0,
1112 NULL,
1113 NULL,
1114 NULL);
1115 if (err != kDNSServiceErr_NoError)
1116 elog(LOG, "DNSServiceRegister() failed: error code %ld",
1117 (long) err);
1118
1119 /*
1120 * We don't bother to read the mDNS daemon's reply, and we expect that
1121 * it will automatically terminate our registration when the socket is
1122 * closed at postmaster termination. So there's nothing more to be
1123 * done here. However, the bonjour_sdref is kept around so that
1124 * forked children can close their copies of the socket.
1125 */
1126 }
1127 #endif
1128
1129 #ifdef HAVE_UNIX_SOCKETS
1130 if (Unix_socket_directories)
1131 {
1132 char *rawstring;
1133 List *elemlist;
1134 ListCell *l;
1135 int success = 0;
1136
1137 /* Need a modifiable copy of Unix_socket_directories */
1138 rawstring = pstrdup(Unix_socket_directories);
1139
1140 /* Parse string into list of directories */
1141 if (!SplitDirectoriesString(rawstring, ',', &elemlist))
1142 {
1143 /* syntax error in list */
1144 ereport(FATAL,
1145 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1146 errmsg("invalid list syntax in parameter \"%s\"",
1147 "unix_socket_directories")));
1148 }
1149
1150 foreach(l, elemlist)
1151 {
1152 char *socketdir = (char *) lfirst(l);
1153
1154 status = StreamServerPort(AF_UNIX, NULL,
1155 (unsigned short) PostPortNumber,
1156 socketdir,
1157 ListenSocket, MAXLISTEN);
1158
1159 if (status == STATUS_OK)
1160 {
1161 success++;
1162 /* record the first successful Unix socket in lockfile */
1163 if (success == 1)
1164 AddToDataDirLockFile(LOCK_FILE_LINE_SOCKET_DIR, socketdir);
1165 }
1166 else
1167 ereport(WARNING,
1168 (errmsg("could not create Unix-domain socket in directory \"%s\"",
1169 socketdir)));
1170 }
1171
1172 if (!success && elemlist != NIL)
1173 ereport(FATAL,
1174 (errmsg("could not create any Unix-domain sockets")));
1175
1176 list_free_deep(elemlist);
1177 pfree(rawstring);
1178 }
1179 #endif
1180
1181 /*
1182 * check that we have some socket to listen on
1183 */
1184 if (ListenSocket[0] == PGINVALID_SOCKET)
1185 ereport(FATAL,
1186 (errmsg("no socket created for listening")));
1187
1188 /*
1189 * If no valid TCP ports, write an empty line for listen address,
1190 * indicating the Unix socket must be used. Note that this line is not
1191 * added to the lock file until there is a socket backing it.
1192 */
1193 if (!listen_addr_saved)
1194 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, "");
1195
1196 /*
1197 * Set up shared memory and semaphores.
1198 */
1199 reset_shared(PostPortNumber);
1200
1201 /*
1202 * Estimate number of openable files. This must happen after setting up
1203 * semaphores, because on some platforms semaphores count as open files.
1204 */
1205 set_max_safe_fds();
1206
1207 /*
1208 * Set reference point for stack-depth checking.
1209 */
1210 set_stack_base();
1211
1212 /*
1213 * Initialize pipe (or process handle on Windows) that allows children to
1214 * wake up from sleep on postmaster death.
1215 */
1216 InitPostmasterDeathWatchHandle();
1217
1218 #ifdef WIN32
1219
1220 /*
1221 * Initialize I/O completion port used to deliver list of dead children.
1222 */
1223 win32ChildQueue = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 1);
1224 if (win32ChildQueue == NULL)
1225 ereport(FATAL,
1226 (errmsg("could not create I/O completion port for child queue")));
1227 #endif
1228
1229 /*
1230 * Record postmaster options. We delay this till now to avoid recording
1231 * bogus options (eg, NBuffers too high for available memory).
1232 */
1233 if (!CreateOptsFile(argc, argv, my_exec_path))
1234 ExitPostmaster(1);
1235
1236 #ifdef EXEC_BACKEND
1237 /* Write out nondefault GUC settings for child processes to use */
1238 write_nondefault_variables(PGC_POSTMASTER);
1239 #endif
1240
1241 /*
1242 * Write the external PID file if requested
1243 */
1244 if (external_pid_file)
1245 {
1246 FILE *fpidfile = fopen(external_pid_file, "w");
1247
1248 if (fpidfile)
1249 {
1250 fprintf(fpidfile, "%d\n", MyProcPid);
1251 fclose(fpidfile);
1252
1253 /* Make PID file world readable */
1254 if (chmod(external_pid_file, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH) != 0)
1255 write_stderr("%s: could not change permissions of external PID file \"%s\": %s\n",
1256 progname, external_pid_file, strerror(errno));
1257 }
1258 else
1259 write_stderr("%s: could not write external PID file \"%s\": %s\n",
1260 progname, external_pid_file, strerror(errno));
1261
1262 on_proc_exit(unlink_external_pid_file, 0);
1263 }
1264
1265 /*
1266 * Remove old temporary files. At this point there can be no other
1267 * Postgres processes running in this directory, so this should be safe.
1268 */
1269 RemovePgTempFiles();
1270
1271 /*
1272 * Forcibly remove the files signaling a standby promotion request.
1273 * Otherwise, the existence of those files triggers a promotion too early,
1274 * whether a user wants that or not.
1275 *
1276 * This removal of files is usually unnecessary because they can exist
1277 * only during a few moments during a standby promotion. However there is
1278 * a race condition: if pg_ctl promote is executed and creates the files
1279 * during a promotion, the files can stay around even after the server is
1280 * brought up to new master. Then, if new standby starts by using the
1281 * backup taken from that master, the files can exist at the server
1282 * startup and should be removed in order to avoid an unexpected
1283 * promotion.
1284 *
1285 * Note that promotion signal files need to be removed before the startup
1286 * process is invoked. Because, after that, they can be used by
1287 * postmaster's SIGUSR1 signal handler.
1288 */
1289 RemovePromoteSignalFiles();
1290
1291 /* Do the same for logrotate signal file */
1292 RemoveLogrotateSignalFiles();
1293
1294 /* Remove any outdated file holding the current log filenames. */
1295 if (unlink(LOG_METAINFO_DATAFILE) < 0 && errno != ENOENT)
1296 ereport(LOG,
1297 (errcode_for_file_access(),
1298 errmsg("could not remove file \"%s\": %m",
1299 LOG_METAINFO_DATAFILE)));
1300
1301 /*
1302 * If enabled, start up syslogger collection subprocess
1303 */
1304 SysLoggerPID = SysLogger_Start();
1305
1306 /*
1307 * Reset whereToSendOutput from DestDebug (its starting state) to
1308 * DestNone. This stops ereport from sending log messages to stderr unless
1309 * Log_destination permits. We don't do this until the postmaster is
1310 * fully launched, since startup failures may as well be reported to
1311 * stderr.
1312 *
1313 * If we are in fact disabling logging to stderr, first emit a log message
1314 * saying so, to provide a breadcrumb trail for users who may not remember
1315 * that their logging is configured to go somewhere else.
1316 */
1317 if (!(Log_destination & LOG_DESTINATION_STDERR))
1318 ereport(LOG,
1319 (errmsg("ending log output to stderr"),
1320 errhint("Future log output will go to log destination \"%s\".",
1321 Log_destination_string)));
1322
1323 whereToSendOutput = DestNone;
1324
1325 /*
1326 * Initialize stats collection subsystem (this does NOT start the
1327 * collector process!)
1328 */
1329 pgstat_init();
1330
1331 /*
1332 * Initialize the autovacuum subsystem (again, no process start yet)
1333 */
1334 autovac_init();
1335
1336 /*
1337 * Load configuration files for client authentication.
1338 */
1339 if (!load_hba())
1340 {
1341 /*
1342 * It makes no sense to continue if we fail to load the HBA file,
1343 * since there is no way to connect to the database in this case.
1344 */
1345 ereport(FATAL,
1346 (errmsg("could not load pg_hba.conf")));
1347 }
1348 if (!load_ident())
1349 {
1350 /*
1351 * We can start up without the IDENT file, although it means that you
1352 * cannot log in using any of the authentication methods that need a
1353 * user name mapping. load_ident() already logged the details of error
1354 * to the log.
1355 */
1356 }
1357
1358 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1359
1360 /*
1361 * On macOS, libintl replaces setlocale() with a version that calls
1362 * CFLocaleCopyCurrent() when its second argument is "" and every relevant
1363 * environment variable is unset or empty. CFLocaleCopyCurrent() makes
1364 * the process multithreaded. The postmaster calls sigprocmask() and
1365 * calls fork() without an immediate exec(), both of which have undefined
1366 * behavior in a multithreaded program. A multithreaded postmaster is the
1367 * normal case on Windows, which offers neither fork() nor sigprocmask().
1368 */
1369 if (pthread_is_threaded_np() != 0)
1370 ereport(FATAL,
1371 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1372 errmsg("postmaster became multithreaded during startup"),
1373 errhint("Set the LC_ALL environment variable to a valid locale.")));
1374 #endif
1375
1376 /*
1377 * Remember postmaster startup time
1378 */
1379 PgStartTime = GetCurrentTimestamp();
1380
1381 /*
1382 * Report postmaster status in the postmaster.pid file, to allow pg_ctl to
1383 * see what's happening.
1384 */
1385 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STARTING);
1386
1387 /*
1388 * We're ready to rock and roll...
1389 */
1390 StartupPID = StartupDataBase();
1391 Assert(StartupPID != 0);
1392 StartupStatus = STARTUP_RUNNING;
1393 pmState = PM_STARTUP;
1394
1395 /* Some workers may be scheduled to start now */
1396 maybe_start_bgworkers();
1397
1398 status = ServerLoop();
1399
1400 /*
1401 * ServerLoop probably shouldn't ever return, but if it does, close down.
1402 */
1403 ExitPostmaster(status != STATUS_OK);
1404
1405 abort(); /* not reached */
1406 }
1407
1408
1409 /*
1410 * on_proc_exit callback to close server's listen sockets
1411 */
1412 static void
CloseServerPorts(int status,Datum arg)1413 CloseServerPorts(int status, Datum arg)
1414 {
1415 int i;
1416
1417 /*
1418 * First, explicitly close all the socket FDs. We used to just let this
1419 * happen implicitly at postmaster exit, but it's better to close them
1420 * before we remove the postmaster.pid lockfile; otherwise there's a race
1421 * condition if a new postmaster wants to re-use the TCP port number.
1422 */
1423 for (i = 0; i < MAXLISTEN; i++)
1424 {
1425 if (ListenSocket[i] != PGINVALID_SOCKET)
1426 {
1427 StreamClose(ListenSocket[i]);
1428 ListenSocket[i] = PGINVALID_SOCKET;
1429 }
1430 }
1431
1432 /*
1433 * Next, remove any filesystem entries for Unix sockets. To avoid race
1434 * conditions against incoming postmasters, this must happen after closing
1435 * the sockets and before removing lock files.
1436 */
1437 RemoveSocketFiles();
1438
1439 /*
1440 * We don't do anything about socket lock files here; those will be
1441 * removed in a later on_proc_exit callback.
1442 */
1443 }
1444
1445 /*
1446 * on_proc_exit callback to delete external_pid_file
1447 */
1448 static void
unlink_external_pid_file(int status,Datum arg)1449 unlink_external_pid_file(int status, Datum arg)
1450 {
1451 if (external_pid_file)
1452 unlink(external_pid_file);
1453 }
1454
1455
1456 /*
1457 * Compute and check the directory paths to files that are part of the
1458 * installation (as deduced from the postgres executable's own location)
1459 */
1460 static void
getInstallationPaths(const char * argv0)1461 getInstallationPaths(const char *argv0)
1462 {
1463 DIR *pdir;
1464
1465 /* Locate the postgres executable itself */
1466 if (find_my_exec(argv0, my_exec_path) < 0)
1467 elog(FATAL, "%s: could not locate my own executable path", argv0);
1468
1469 #ifdef EXEC_BACKEND
1470 /* Locate executable backend before we change working directory */
1471 if (find_other_exec(argv0, "postgres", PG_BACKEND_VERSIONSTR,
1472 postgres_exec_path) < 0)
1473 ereport(FATAL,
1474 (errmsg("%s: could not locate matching postgres executable",
1475 argv0)));
1476 #endif
1477
1478 /*
1479 * Locate the pkglib directory --- this has to be set early in case we try
1480 * to load any modules from it in response to postgresql.conf entries.
1481 */
1482 get_pkglib_path(my_exec_path, pkglib_path);
1483
1484 /*
1485 * Verify that there's a readable directory there; otherwise the Postgres
1486 * installation is incomplete or corrupt. (A typical cause of this
1487 * failure is that the postgres executable has been moved or hardlinked to
1488 * some directory that's not a sibling of the installation lib/
1489 * directory.)
1490 */
1491 pdir = AllocateDir(pkglib_path);
1492 if (pdir == NULL)
1493 ereport(ERROR,
1494 (errcode_for_file_access(),
1495 errmsg("could not open directory \"%s\": %m",
1496 pkglib_path),
1497 errhint("This may indicate an incomplete PostgreSQL installation, or that the file \"%s\" has been moved away from its proper location.",
1498 my_exec_path)));
1499 FreeDir(pdir);
1500
1501 /*
1502 * XXX is it worth similarly checking the share/ directory? If the lib/
1503 * directory is there, then share/ probably is too.
1504 */
1505 }
1506
1507 /*
1508 * Check that pg_control exists in the correct location in the data directory.
1509 *
1510 * No attempt is made to validate the contents of pg_control here. This is
1511 * just a sanity check to see if we are looking at a real data directory.
1512 */
1513 static void
checkControlFile(void)1514 checkControlFile(void)
1515 {
1516 char path[MAXPGPATH];
1517 FILE *fp;
1518
1519 snprintf(path, sizeof(path), "%s/global/pg_control", DataDir);
1520
1521 fp = AllocateFile(path, PG_BINARY_R);
1522 if (fp == NULL)
1523 {
1524 write_stderr("%s: could not find the database system\n"
1525 "Expected to find it in the directory \"%s\",\n"
1526 "but could not open file \"%s\": %s\n",
1527 progname, DataDir, path, strerror(errno));
1528 ExitPostmaster(2);
1529 }
1530 FreeFile(fp);
1531 }
1532
1533 /*
1534 * Determine how long should we let ServerLoop sleep.
1535 *
1536 * In normal conditions we wait at most one minute, to ensure that the other
1537 * background tasks handled by ServerLoop get done even when no requests are
1538 * arriving. However, if there are background workers waiting to be started,
1539 * we don't actually sleep so that they are quickly serviced. Other exception
1540 * cases are as shown in the code.
1541 */
1542 static void
DetermineSleepTime(struct timeval * timeout)1543 DetermineSleepTime(struct timeval *timeout)
1544 {
1545 TimestampTz next_wakeup = 0;
1546
1547 /*
1548 * Normal case: either there are no background workers at all, or we're in
1549 * a shutdown sequence (during which we ignore bgworkers altogether).
1550 */
1551 if (Shutdown > NoShutdown ||
1552 (!StartWorkerNeeded && !HaveCrashedWorker))
1553 {
1554 if (AbortStartTime != 0)
1555 {
1556 /* time left to abort; clamp to 0 in case it already expired */
1557 timeout->tv_sec = SIGKILL_CHILDREN_AFTER_SECS -
1558 (time(NULL) - AbortStartTime);
1559 timeout->tv_sec = Max(timeout->tv_sec, 0);
1560 timeout->tv_usec = 0;
1561 }
1562 else
1563 {
1564 timeout->tv_sec = 60;
1565 timeout->tv_usec = 0;
1566 }
1567 return;
1568 }
1569
1570 if (StartWorkerNeeded)
1571 {
1572 timeout->tv_sec = 0;
1573 timeout->tv_usec = 0;
1574 return;
1575 }
1576
1577 if (HaveCrashedWorker)
1578 {
1579 slist_mutable_iter siter;
1580
1581 /*
1582 * When there are crashed bgworkers, we sleep just long enough that
1583 * they are restarted when they request to be. Scan the list to
1584 * determine the minimum of all wakeup times according to most recent
1585 * crash time and requested restart interval.
1586 */
1587 slist_foreach_modify(siter, &BackgroundWorkerList)
1588 {
1589 RegisteredBgWorker *rw;
1590 TimestampTz this_wakeup;
1591
1592 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
1593
1594 if (rw->rw_crashed_at == 0)
1595 continue;
1596
1597 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART
1598 || rw->rw_terminate)
1599 {
1600 ForgetBackgroundWorker(&siter);
1601 continue;
1602 }
1603
1604 this_wakeup = TimestampTzPlusMilliseconds(rw->rw_crashed_at,
1605 1000L * rw->rw_worker.bgw_restart_time);
1606 if (next_wakeup == 0 || this_wakeup < next_wakeup)
1607 next_wakeup = this_wakeup;
1608 }
1609 }
1610
1611 if (next_wakeup != 0)
1612 {
1613 long secs;
1614 int microsecs;
1615
1616 TimestampDifference(GetCurrentTimestamp(), next_wakeup,
1617 &secs, µsecs);
1618 timeout->tv_sec = secs;
1619 timeout->tv_usec = microsecs;
1620
1621 /* Ensure we don't exceed one minute */
1622 if (timeout->tv_sec > 60)
1623 {
1624 timeout->tv_sec = 60;
1625 timeout->tv_usec = 0;
1626 }
1627 }
1628 else
1629 {
1630 timeout->tv_sec = 60;
1631 timeout->tv_usec = 0;
1632 }
1633 }
1634
1635 /*
1636 * Main idle loop of postmaster
1637 *
1638 * NB: Needs to be called with signals blocked
1639 */
1640 static int
ServerLoop(void)1641 ServerLoop(void)
1642 {
1643 fd_set readmask;
1644 int nSockets;
1645 time_t last_lockfile_recheck_time,
1646 last_touch_time;
1647
1648 last_lockfile_recheck_time = last_touch_time = time(NULL);
1649
1650 nSockets = initMasks(&readmask);
1651
1652 for (;;)
1653 {
1654 fd_set rmask;
1655 int selres;
1656 time_t now;
1657
1658 /*
1659 * Wait for a connection request to arrive.
1660 *
1661 * We block all signals except while sleeping. That makes it safe for
1662 * signal handlers, which again block all signals while executing, to
1663 * do nontrivial work.
1664 *
1665 * If we are in PM_WAIT_DEAD_END state, then we don't want to accept
1666 * any new connections, so we don't call select(), and just sleep.
1667 */
1668 memcpy((char *) &rmask, (char *) &readmask, sizeof(fd_set));
1669
1670 if (pmState == PM_WAIT_DEAD_END)
1671 {
1672 PG_SETMASK(&UnBlockSig);
1673
1674 pg_usleep(100000L); /* 100 msec seems reasonable */
1675 selres = 0;
1676
1677 PG_SETMASK(&BlockSig);
1678 }
1679 else
1680 {
1681 /* must set timeout each time; some OSes change it! */
1682 struct timeval timeout;
1683
1684 /* Needs to run with blocked signals! */
1685 DetermineSleepTime(&timeout);
1686
1687 PG_SETMASK(&UnBlockSig);
1688
1689 selres = select(nSockets, &rmask, NULL, NULL, &timeout);
1690
1691 PG_SETMASK(&BlockSig);
1692 }
1693
1694 /* Now check the select() result */
1695 if (selres < 0)
1696 {
1697 if (errno != EINTR && errno != EWOULDBLOCK)
1698 {
1699 ereport(LOG,
1700 (errcode_for_socket_access(),
1701 errmsg("select() failed in postmaster: %m")));
1702 return STATUS_ERROR;
1703 }
1704 }
1705
1706 /*
1707 * New connection pending on any of our sockets? If so, fork a child
1708 * process to deal with it.
1709 */
1710 if (selres > 0)
1711 {
1712 int i;
1713
1714 for (i = 0; i < MAXLISTEN; i++)
1715 {
1716 if (ListenSocket[i] == PGINVALID_SOCKET)
1717 break;
1718 if (FD_ISSET(ListenSocket[i], &rmask))
1719 {
1720 Port *port;
1721
1722 port = ConnCreate(ListenSocket[i]);
1723 if (port)
1724 {
1725 BackendStartup(port);
1726
1727 /*
1728 * We no longer need the open socket or port structure
1729 * in this process
1730 */
1731 StreamClose(port->sock);
1732 ConnFree(port);
1733 }
1734 }
1735 }
1736 }
1737
1738 /* If we have lost the log collector, try to start a new one */
1739 if (SysLoggerPID == 0 && Logging_collector)
1740 SysLoggerPID = SysLogger_Start();
1741
1742 /*
1743 * If no background writer process is running, and we are not in a
1744 * state that prevents it, start one. It doesn't matter if this
1745 * fails, we'll just try again later. Likewise for the checkpointer.
1746 */
1747 if (pmState == PM_RUN || pmState == PM_RECOVERY ||
1748 pmState == PM_HOT_STANDBY)
1749 {
1750 if (CheckpointerPID == 0)
1751 CheckpointerPID = StartCheckpointer();
1752 if (BgWriterPID == 0)
1753 BgWriterPID = StartBackgroundWriter();
1754 }
1755
1756 /*
1757 * Likewise, if we have lost the walwriter process, try to start a new
1758 * one. But this is needed only in normal operation (else we cannot
1759 * be writing any new WAL).
1760 */
1761 if (WalWriterPID == 0 && pmState == PM_RUN)
1762 WalWriterPID = StartWalWriter();
1763
1764 /*
1765 * If we have lost the autovacuum launcher, try to start a new one. We
1766 * don't want autovacuum to run in binary upgrade mode because
1767 * autovacuum might update relfrozenxid for empty tables before the
1768 * physical files are put in place.
1769 */
1770 if (!IsBinaryUpgrade && AutoVacPID == 0 &&
1771 (AutoVacuumingActive() || start_autovac_launcher) &&
1772 pmState == PM_RUN)
1773 {
1774 AutoVacPID = StartAutoVacLauncher();
1775 if (AutoVacPID != 0)
1776 start_autovac_launcher = false; /* signal processed */
1777 }
1778
1779 /* If we have lost the stats collector, try to start a new one */
1780 if (PgStatPID == 0 &&
1781 (pmState == PM_RUN || pmState == PM_HOT_STANDBY))
1782 PgStatPID = pgstat_start();
1783
1784 /* If we have lost the archiver, try to start a new one. */
1785 if (PgArchPID == 0 && PgArchStartupAllowed())
1786 PgArchPID = pgarch_start();
1787
1788 /* If we need to signal the autovacuum launcher, do so now */
1789 if (avlauncher_needs_signal)
1790 {
1791 avlauncher_needs_signal = false;
1792 if (AutoVacPID != 0)
1793 kill(AutoVacPID, SIGUSR2);
1794 }
1795
1796 /* If we need to start a WAL receiver, try to do that now */
1797 if (WalReceiverRequested)
1798 MaybeStartWalReceiver();
1799
1800 /* Get other worker processes running, if needed */
1801 if (StartWorkerNeeded || HaveCrashedWorker)
1802 maybe_start_bgworkers();
1803
1804 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1805
1806 /*
1807 * With assertions enabled, check regularly for appearance of
1808 * additional threads. All builds check at start and exit.
1809 */
1810 Assert(pthread_is_threaded_np() == 0);
1811 #endif
1812
1813 /*
1814 * Lastly, check to see if it's time to do some things that we don't
1815 * want to do every single time through the loop, because they're a
1816 * bit expensive. Note that there's up to a minute of slop in when
1817 * these tasks will be performed, since DetermineSleepTime() will let
1818 * us sleep at most that long; except for SIGKILL timeout which has
1819 * special-case logic there.
1820 */
1821 now = time(NULL);
1822
1823 /*
1824 * If we already sent SIGQUIT to children and they are slow to shut
1825 * down, it's time to send them SIGKILL. This doesn't happen
1826 * normally, but under certain conditions backends can get stuck while
1827 * shutting down. This is a last measure to get them unwedged.
1828 *
1829 * Note we also do this during recovery from a process crash.
1830 */
1831 if ((Shutdown >= ImmediateShutdown || (FatalError && !SendStop)) &&
1832 AbortStartTime != 0 &&
1833 (now - AbortStartTime) >= SIGKILL_CHILDREN_AFTER_SECS)
1834 {
1835 /* We were gentle with them before. Not anymore */
1836 TerminateChildren(SIGKILL);
1837 /* reset flag so we don't SIGKILL again */
1838 AbortStartTime = 0;
1839 }
1840
1841 /*
1842 * Once a minute, verify that postmaster.pid hasn't been removed or
1843 * overwritten. If it has, we force a shutdown. This avoids having
1844 * postmasters and child processes hanging around after their database
1845 * is gone, and maybe causing problems if a new database cluster is
1846 * created in the same place. It also provides some protection
1847 * against a DBA foolishly removing postmaster.pid and manually
1848 * starting a new postmaster. Data corruption is likely to ensue from
1849 * that anyway, but we can minimize the damage by aborting ASAP.
1850 */
1851 if (now - last_lockfile_recheck_time >= 1 * SECS_PER_MINUTE)
1852 {
1853 if (!RecheckDataDirLockFile())
1854 {
1855 ereport(LOG,
1856 (errmsg("performing immediate shutdown because data directory lock file is invalid")));
1857 kill(MyProcPid, SIGQUIT);
1858 }
1859 last_lockfile_recheck_time = now;
1860 }
1861
1862 /*
1863 * Touch Unix socket and lock files every 58 minutes, to ensure that
1864 * they are not removed by overzealous /tmp-cleaning tasks. We assume
1865 * no one runs cleaners with cutoff times of less than an hour ...
1866 */
1867 if (now - last_touch_time >= 58 * SECS_PER_MINUTE)
1868 {
1869 TouchSocketFiles();
1870 TouchSocketLockFiles();
1871 last_touch_time = now;
1872 }
1873 }
1874 }
1875
1876 /*
1877 * Initialise the masks for select() for the ports we are listening on.
1878 * Return the number of sockets to listen on.
1879 */
1880 static int
initMasks(fd_set * rmask)1881 initMasks(fd_set *rmask)
1882 {
1883 int maxsock = -1;
1884 int i;
1885
1886 FD_ZERO(rmask);
1887
1888 for (i = 0; i < MAXLISTEN; i++)
1889 {
1890 int fd = ListenSocket[i];
1891
1892 if (fd == PGINVALID_SOCKET)
1893 break;
1894 FD_SET(fd, rmask);
1895
1896 if (fd > maxsock)
1897 maxsock = fd;
1898 }
1899
1900 return maxsock + 1;
1901 }
1902
1903
1904 /*
1905 * Read a client's startup packet and do something according to it.
1906 *
1907 * Returns STATUS_OK or STATUS_ERROR, or might call ereport(FATAL) and
1908 * not return at all.
1909 *
1910 * (Note that ereport(FATAL) stuff is sent to the client, so only use it
1911 * if that's what you want. Return STATUS_ERROR if you don't want to
1912 * send anything to the client, which would typically be appropriate
1913 * if we detect a communications failure.)
1914 *
1915 * Set ssl_done and/or gss_done when negotiation of an encrypted layer
1916 * (currently, TLS or GSSAPI) is completed. A successful negotiation of either
1917 * encryption layer sets both flags, but a rejected negotiation sets only the
1918 * flag for that layer, since the client may wish to try the other one. We
1919 * should make no assumption here about the order in which the client may make
1920 * requests.
1921 */
1922 static int
ProcessStartupPacket(Port * port,bool ssl_done,bool gss_done)1923 ProcessStartupPacket(Port *port, bool ssl_done, bool gss_done)
1924 {
1925 int32 len;
1926 void *buf;
1927 ProtocolVersion proto;
1928 MemoryContext oldcontext;
1929
1930 pq_startmsgread();
1931
1932 /*
1933 * Grab the first byte of the length word separately, so that we can tell
1934 * whether we have no data at all or an incomplete packet. (This might
1935 * sound inefficient, but it's not really, because of buffering in
1936 * pqcomm.c.)
1937 */
1938 if (pq_getbytes((char *) &len, 1) == EOF)
1939 {
1940 /*
1941 * If we get no data at all, don't clutter the log with a complaint;
1942 * such cases often occur for legitimate reasons. An example is that
1943 * we might be here after responding to NEGOTIATE_SSL_CODE, and if the
1944 * client didn't like our response, it'll probably just drop the
1945 * connection. Service-monitoring software also often just opens and
1946 * closes a connection without sending anything. (So do port
1947 * scanners, which may be less benign, but it's not really our job to
1948 * notice those.)
1949 */
1950 return STATUS_ERROR;
1951 }
1952
1953 if (pq_getbytes(((char *) &len) + 1, 3) == EOF)
1954 {
1955 /* Got a partial length word, so bleat about that */
1956 if (!ssl_done && !gss_done)
1957 ereport(COMMERROR,
1958 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1959 errmsg("incomplete startup packet")));
1960 return STATUS_ERROR;
1961 }
1962
1963 len = pg_ntoh32(len);
1964 len -= 4;
1965
1966 if (len < (int32) sizeof(ProtocolVersion) ||
1967 len > MAX_STARTUP_PACKET_LENGTH)
1968 {
1969 ereport(COMMERROR,
1970 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1971 errmsg("invalid length of startup packet")));
1972 return STATUS_ERROR;
1973 }
1974
1975 /*
1976 * Allocate at least the size of an old-style startup packet, plus one
1977 * extra byte, and make sure all are zeroes. This ensures we will have
1978 * null termination of all strings, in both fixed- and variable-length
1979 * packet layouts.
1980 */
1981 if (len <= (int32) sizeof(StartupPacket))
1982 buf = palloc0(sizeof(StartupPacket) + 1);
1983 else
1984 buf = palloc0(len + 1);
1985
1986 if (pq_getbytes(buf, len) == EOF)
1987 {
1988 ereport(COMMERROR,
1989 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1990 errmsg("incomplete startup packet")));
1991 return STATUS_ERROR;
1992 }
1993 pq_endmsgread();
1994
1995 /*
1996 * The first field is either a protocol version number or a special
1997 * request code.
1998 */
1999 port->proto = proto = pg_ntoh32(*((ProtocolVersion *) buf));
2000
2001 if (proto == CANCEL_REQUEST_CODE)
2002 {
2003 processCancelRequest(port, buf);
2004 /* Not really an error, but we don't want to proceed further */
2005 return STATUS_ERROR;
2006 }
2007
2008 if (proto == NEGOTIATE_SSL_CODE && !ssl_done)
2009 {
2010 char SSLok;
2011
2012 #ifdef USE_SSL
2013 /* No SSL when disabled or on Unix sockets */
2014 if (!LoadedSSL || IS_AF_UNIX(port->laddr.addr.ss_family))
2015 SSLok = 'N';
2016 else
2017 SSLok = 'S'; /* Support for SSL */
2018 #else
2019 SSLok = 'N'; /* No support for SSL */
2020 #endif
2021
2022 retry1:
2023 if (send(port->sock, &SSLok, 1, 0) != 1)
2024 {
2025 if (errno == EINTR)
2026 goto retry1; /* if interrupted, just retry */
2027 ereport(COMMERROR,
2028 (errcode_for_socket_access(),
2029 errmsg("failed to send SSL negotiation response: %m")));
2030 return STATUS_ERROR; /* close the connection */
2031 }
2032
2033 #ifdef USE_SSL
2034 if (SSLok == 'S' && secure_open_server(port) == -1)
2035 return STATUS_ERROR;
2036 #endif
2037
2038 /*
2039 * At this point we should have no data already buffered. If we do,
2040 * it was received before we performed the SSL handshake, so it wasn't
2041 * encrypted and indeed may have been injected by a man-in-the-middle.
2042 * We report this case to the client.
2043 */
2044 if (pq_buffer_has_data())
2045 ereport(FATAL,
2046 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2047 errmsg("received unencrypted data after SSL request"),
2048 errdetail("This could be either a client-software bug or evidence of an attempted man-in-the-middle attack.")));
2049
2050 /*
2051 * regular startup packet, cancel, etc packet should follow, but not
2052 * another SSL negotiation request, and a GSS request should only
2053 * follow if SSL was rejected (client may negotiate in either order)
2054 */
2055 return ProcessStartupPacket(port, true, SSLok == 'S');
2056 }
2057 else if (proto == NEGOTIATE_GSS_CODE && !gss_done)
2058 {
2059 char GSSok = 'N';
2060
2061 #ifdef ENABLE_GSS
2062 /* No GSSAPI encryption when on Unix socket */
2063 if (!IS_AF_UNIX(port->laddr.addr.ss_family))
2064 GSSok = 'G';
2065 #endif
2066
2067 while (send(port->sock, &GSSok, 1, 0) != 1)
2068 {
2069 if (errno == EINTR)
2070 continue;
2071 ereport(COMMERROR,
2072 (errcode_for_socket_access(),
2073 errmsg("failed to send GSSAPI negotiation response: %m")));
2074 return STATUS_ERROR; /* close the connection */
2075 }
2076
2077 #ifdef ENABLE_GSS
2078 if (GSSok == 'G' && secure_open_gssapi(port) == -1)
2079 return STATUS_ERROR;
2080 #endif
2081
2082 /*
2083 * At this point we should have no data already buffered. If we do,
2084 * it was received before we performed the GSS handshake, so it wasn't
2085 * encrypted and indeed may have been injected by a man-in-the-middle.
2086 * We report this case to the client.
2087 */
2088 if (pq_buffer_has_data())
2089 ereport(FATAL,
2090 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2091 errmsg("received unencrypted data after GSSAPI encryption request"),
2092 errdetail("This could be either a client-software bug or evidence of an attempted man-in-the-middle attack.")));
2093
2094 /*
2095 * regular startup packet, cancel, etc packet should follow, but not
2096 * another GSS negotiation request, and an SSL request should only
2097 * follow if GSS was rejected (client may negotiate in either order)
2098 */
2099 return ProcessStartupPacket(port, GSSok == 'G', true);
2100 }
2101
2102 /* Could add additional special packet types here */
2103
2104 /*
2105 * Set FrontendProtocol now so that ereport() knows what format to send if
2106 * we fail during startup.
2107 */
2108 FrontendProtocol = proto;
2109
2110 /* Check that the major protocol version is in range. */
2111 if (PG_PROTOCOL_MAJOR(proto) < PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST) ||
2112 PG_PROTOCOL_MAJOR(proto) > PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST))
2113 ereport(FATAL,
2114 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2115 errmsg("unsupported frontend protocol %u.%u: server supports %u.0 to %u.%u",
2116 PG_PROTOCOL_MAJOR(proto), PG_PROTOCOL_MINOR(proto),
2117 PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST),
2118 PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST),
2119 PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST))));
2120
2121 /*
2122 * Now fetch parameters out of startup packet and save them into the Port
2123 * structure. All data structures attached to the Port struct must be
2124 * allocated in TopMemoryContext so that they will remain available in a
2125 * running backend (even after PostmasterContext is destroyed). We need
2126 * not worry about leaking this storage on failure, since we aren't in the
2127 * postmaster process anymore.
2128 */
2129 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
2130
2131 if (PG_PROTOCOL_MAJOR(proto) >= 3)
2132 {
2133 int32 offset = sizeof(ProtocolVersion);
2134 List *unrecognized_protocol_options = NIL;
2135
2136 /*
2137 * Scan packet body for name/option pairs. We can assume any string
2138 * beginning within the packet body is null-terminated, thanks to
2139 * zeroing extra byte above.
2140 */
2141 port->guc_options = NIL;
2142
2143 while (offset < len)
2144 {
2145 char *nameptr = ((char *) buf) + offset;
2146 int32 valoffset;
2147 char *valptr;
2148
2149 if (*nameptr == '\0')
2150 break; /* found packet terminator */
2151 valoffset = offset + strlen(nameptr) + 1;
2152 if (valoffset >= len)
2153 break; /* missing value, will complain below */
2154 valptr = ((char *) buf) + valoffset;
2155
2156 if (strcmp(nameptr, "database") == 0)
2157 port->database_name = pstrdup(valptr);
2158 else if (strcmp(nameptr, "user") == 0)
2159 port->user_name = pstrdup(valptr);
2160 else if (strcmp(nameptr, "options") == 0)
2161 port->cmdline_options = pstrdup(valptr);
2162 else if (strcmp(nameptr, "replication") == 0)
2163 {
2164 /*
2165 * Due to backward compatibility concerns the replication
2166 * parameter is a hybrid beast which allows the value to be
2167 * either boolean or the string 'database'. The latter
2168 * connects to a specific database which is e.g. required for
2169 * logical decoding while.
2170 */
2171 if (strcmp(valptr, "database") == 0)
2172 {
2173 am_walsender = true;
2174 am_db_walsender = true;
2175 }
2176 else if (!parse_bool(valptr, &am_walsender))
2177 ereport(FATAL,
2178 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
2179 errmsg("invalid value for parameter \"%s\": \"%s\"",
2180 "replication",
2181 valptr),
2182 errhint("Valid values are: \"false\", 0, \"true\", 1, \"database\".")));
2183 }
2184 else if (strncmp(nameptr, "_pq_.", 5) == 0)
2185 {
2186 /*
2187 * Any option beginning with _pq_. is reserved for use as a
2188 * protocol-level option, but at present no such options are
2189 * defined.
2190 */
2191 unrecognized_protocol_options =
2192 lappend(unrecognized_protocol_options, pstrdup(nameptr));
2193 }
2194 else
2195 {
2196 /* Assume it's a generic GUC option */
2197 port->guc_options = lappend(port->guc_options,
2198 pstrdup(nameptr));
2199 port->guc_options = lappend(port->guc_options,
2200 pstrdup(valptr));
2201
2202 /*
2203 * Copy application_name to port if we come across it. This
2204 * is done so we can log the application_name in the
2205 * connection authorization message. Note that the GUC would
2206 * be used but we haven't gone through GUC setup yet.
2207 */
2208 if (strcmp(nameptr, "application_name") == 0)
2209 {
2210 char *tmp_app_name = pstrdup(valptr);
2211
2212 pg_clean_ascii(tmp_app_name);
2213
2214 port->application_name = tmp_app_name;
2215 }
2216 }
2217 offset = valoffset + strlen(valptr) + 1;
2218 }
2219
2220 /*
2221 * If we didn't find a packet terminator exactly at the end of the
2222 * given packet length, complain.
2223 */
2224 if (offset != len - 1)
2225 ereport(FATAL,
2226 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2227 errmsg("invalid startup packet layout: expected terminator as last byte")));
2228
2229 /*
2230 * If the client requested a newer protocol version or if the client
2231 * requested any protocol options we didn't recognize, let them know
2232 * the newest minor protocol version we do support and the names of
2233 * any unrecognized options.
2234 */
2235 if (PG_PROTOCOL_MINOR(proto) > PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST) ||
2236 unrecognized_protocol_options != NIL)
2237 SendNegotiateProtocolVersion(unrecognized_protocol_options);
2238 }
2239 else
2240 {
2241 /*
2242 * Get the parameters from the old-style, fixed-width-fields startup
2243 * packet as C strings. The packet destination was cleared first so a
2244 * short packet has zeros silently added. We have to be prepared to
2245 * truncate the pstrdup result for oversize fields, though.
2246 */
2247 StartupPacket *packet = (StartupPacket *) buf;
2248
2249 port->database_name = pstrdup(packet->database);
2250 if (strlen(port->database_name) > sizeof(packet->database))
2251 port->database_name[sizeof(packet->database)] = '\0';
2252 port->user_name = pstrdup(packet->user);
2253 if (strlen(port->user_name) > sizeof(packet->user))
2254 port->user_name[sizeof(packet->user)] = '\0';
2255 port->cmdline_options = pstrdup(packet->options);
2256 if (strlen(port->cmdline_options) > sizeof(packet->options))
2257 port->cmdline_options[sizeof(packet->options)] = '\0';
2258 port->guc_options = NIL;
2259 }
2260
2261 /* Check a user name was given. */
2262 if (port->user_name == NULL || port->user_name[0] == '\0')
2263 ereport(FATAL,
2264 (errcode(ERRCODE_INVALID_AUTHORIZATION_SPECIFICATION),
2265 errmsg("no PostgreSQL user name specified in startup packet")));
2266
2267 /* The database defaults to the user name. */
2268 if (port->database_name == NULL || port->database_name[0] == '\0')
2269 port->database_name = pstrdup(port->user_name);
2270
2271 if (Db_user_namespace)
2272 {
2273 /*
2274 * If user@, it is a global user, remove '@'. We only want to do this
2275 * if there is an '@' at the end and no earlier in the user string or
2276 * they may fake as a local user of another database attaching to this
2277 * database.
2278 */
2279 if (strchr(port->user_name, '@') ==
2280 port->user_name + strlen(port->user_name) - 1)
2281 *strchr(port->user_name, '@') = '\0';
2282 else
2283 {
2284 /* Append '@' and dbname */
2285 port->user_name = psprintf("%s@%s", port->user_name, port->database_name);
2286 }
2287 }
2288
2289 /*
2290 * Truncate given database and user names to length of a Postgres name.
2291 * This avoids lookup failures when overlength names are given.
2292 */
2293 if (strlen(port->database_name) >= NAMEDATALEN)
2294 port->database_name[NAMEDATALEN - 1] = '\0';
2295 if (strlen(port->user_name) >= NAMEDATALEN)
2296 port->user_name[NAMEDATALEN - 1] = '\0';
2297
2298 /*
2299 * Normal walsender backends, e.g. for streaming replication, are not
2300 * connected to a particular database. But walsenders used for logical
2301 * replication need to connect to a specific database. We allow streaming
2302 * replication commands to be issued even if connected to a database as it
2303 * can make sense to first make a basebackup and then stream changes
2304 * starting from that.
2305 */
2306 if (am_walsender && !am_db_walsender)
2307 port->database_name[0] = '\0';
2308
2309 /*
2310 * Done putting stuff in TopMemoryContext.
2311 */
2312 MemoryContextSwitchTo(oldcontext);
2313
2314 /*
2315 * If we're going to reject the connection due to database state, say so
2316 * now instead of wasting cycles on an authentication exchange. (This also
2317 * allows a pg_ping utility to be written.)
2318 */
2319 switch (port->canAcceptConnections)
2320 {
2321 case CAC_STARTUP:
2322 ereport(FATAL,
2323 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2324 errmsg("the database system is starting up")));
2325 break;
2326 case CAC_SHUTDOWN:
2327 ereport(FATAL,
2328 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2329 errmsg("the database system is shutting down")));
2330 break;
2331 case CAC_RECOVERY:
2332 ereport(FATAL,
2333 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2334 errmsg("the database system is in recovery mode")));
2335 break;
2336 case CAC_TOOMANY:
2337 ereport(FATAL,
2338 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
2339 errmsg("sorry, too many clients already")));
2340 break;
2341 case CAC_SUPERUSER:
2342 /* OK for now, will check in InitPostgres */
2343 break;
2344 case CAC_OK:
2345 break;
2346 }
2347
2348 return STATUS_OK;
2349 }
2350
2351 /*
2352 * Send a NegotiateProtocolVersion to the client. This lets the client know
2353 * that they have requested a newer minor protocol version than we are able
2354 * to speak. We'll speak the highest version we know about; the client can,
2355 * of course, abandon the connection if that's a problem.
2356 *
2357 * We also include in the response a list of protocol options we didn't
2358 * understand. This allows clients to include optional parameters that might
2359 * be present either in newer protocol versions or third-party protocol
2360 * extensions without fear of having to reconnect if those options are not
2361 * understood, while at the same time making certain that the client is aware
2362 * of which options were actually accepted.
2363 */
2364 static void
SendNegotiateProtocolVersion(List * unrecognized_protocol_options)2365 SendNegotiateProtocolVersion(List *unrecognized_protocol_options)
2366 {
2367 StringInfoData buf;
2368 ListCell *lc;
2369
2370 pq_beginmessage(&buf, 'v'); /* NegotiateProtocolVersion */
2371 pq_sendint32(&buf, PG_PROTOCOL_LATEST);
2372 pq_sendint32(&buf, list_length(unrecognized_protocol_options));
2373 foreach(lc, unrecognized_protocol_options)
2374 pq_sendstring(&buf, lfirst(lc));
2375 pq_endmessage(&buf);
2376
2377 /* no need to flush, some other message will follow */
2378 }
2379
2380 /*
2381 * The client has sent a cancel request packet, not a normal
2382 * start-a-new-connection packet. Perform the necessary processing.
2383 * Nothing is sent back to the client.
2384 */
2385 static void
processCancelRequest(Port * port,void * pkt)2386 processCancelRequest(Port *port, void *pkt)
2387 {
2388 CancelRequestPacket *canc = (CancelRequestPacket *) pkt;
2389 int backendPID;
2390 int32 cancelAuthCode;
2391 Backend *bp;
2392
2393 #ifndef EXEC_BACKEND
2394 dlist_iter iter;
2395 #else
2396 int i;
2397 #endif
2398
2399 backendPID = (int) pg_ntoh32(canc->backendPID);
2400 cancelAuthCode = (int32) pg_ntoh32(canc->cancelAuthCode);
2401
2402 /*
2403 * See if we have a matching backend. In the EXEC_BACKEND case, we can no
2404 * longer access the postmaster's own backend list, and must rely on the
2405 * duplicate array in shared memory.
2406 */
2407 #ifndef EXEC_BACKEND
2408 dlist_foreach(iter, &BackendList)
2409 {
2410 bp = dlist_container(Backend, elem, iter.cur);
2411 #else
2412 for (i = MaxLivePostmasterChildren() - 1; i >= 0; i--)
2413 {
2414 bp = (Backend *) &ShmemBackendArray[i];
2415 #endif
2416 if (bp->pid == backendPID)
2417 {
2418 if (bp->cancel_key == cancelAuthCode)
2419 {
2420 /* Found a match; signal that backend to cancel current op */
2421 ereport(DEBUG2,
2422 (errmsg_internal("processing cancel request: sending SIGINT to process %d",
2423 backendPID)));
2424 signal_child(bp->pid, SIGINT);
2425 }
2426 else
2427 /* Right PID, wrong key: no way, Jose */
2428 ereport(LOG,
2429 (errmsg("wrong key in cancel request for process %d",
2430 backendPID)));
2431 return;
2432 }
2433 #ifndef EXEC_BACKEND /* make GNU Emacs 26.1 see brace balance */
2434 }
2435 #else
2436 }
2437 #endif
2438
2439 /* No matching backend */
2440 ereport(LOG,
2441 (errmsg("PID %d in cancel request did not match any process",
2442 backendPID)));
2443 }
2444
2445 /*
2446 * canAcceptConnections --- check to see if database state allows connections
2447 * of the specified type. backend_type can be BACKEND_TYPE_NORMAL,
2448 * BACKEND_TYPE_AUTOVAC, or BACKEND_TYPE_BGWORKER. (Note that we don't yet
2449 * know whether a NORMAL connection might turn into a walsender.)
2450 */
2451 static CAC_state
canAcceptConnections(int backend_type)2452 canAcceptConnections(int backend_type)
2453 {
2454 CAC_state result = CAC_OK;
2455
2456 /*
2457 * Can't start backends when in startup/shutdown/inconsistent recovery
2458 * state. We treat autovac workers the same as user backends for this
2459 * purpose. However, bgworkers are excluded from this test; we expect
2460 * bgworker_should_start_now() decided whether the DB state allows them.
2461 */
2462 if (pmState != PM_RUN && pmState != PM_HOT_STANDBY &&
2463 backend_type != BACKEND_TYPE_BGWORKER)
2464 {
2465 if (Shutdown > NoShutdown)
2466 return CAC_SHUTDOWN; /* shutdown is pending */
2467 else if (!FatalError &&
2468 (pmState == PM_STARTUP ||
2469 pmState == PM_RECOVERY))
2470 return CAC_STARTUP; /* normal startup */
2471 else
2472 return CAC_RECOVERY; /* else must be crash recovery */
2473 }
2474
2475 /*
2476 * "Smart shutdown" restrictions are applied only to normal connections,
2477 * not to autovac workers or bgworkers. When only superusers can connect,
2478 * we return CAC_SUPERUSER to indicate that superuserness must be checked
2479 * later. Note that neither CAC_OK nor CAC_SUPERUSER can safely be
2480 * returned until we have checked for too many children.
2481 */
2482 if (connsAllowed != ALLOW_ALL_CONNS &&
2483 backend_type == BACKEND_TYPE_NORMAL)
2484 {
2485 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
2486 result = CAC_SUPERUSER; /* allow superusers only */
2487 else
2488 return CAC_SHUTDOWN; /* shutdown is pending */
2489 }
2490
2491 /*
2492 * Don't start too many children.
2493 *
2494 * We allow more connections here than we can have backends because some
2495 * might still be authenticating; they might fail auth, or some existing
2496 * backend might exit before the auth cycle is completed. The exact
2497 * MaxBackends limit is enforced when a new backend tries to join the
2498 * shared-inval backend array.
2499 *
2500 * The limit here must match the sizes of the per-child-process arrays;
2501 * see comments for MaxLivePostmasterChildren().
2502 */
2503 if (CountChildren(BACKEND_TYPE_ALL) >= MaxLivePostmasterChildren())
2504 result = CAC_TOOMANY;
2505
2506 return result;
2507 }
2508
2509
2510 /*
2511 * ConnCreate -- create a local connection data structure
2512 *
2513 * Returns NULL on failure, other than out-of-memory which is fatal.
2514 */
2515 static Port *
ConnCreate(int serverFd)2516 ConnCreate(int serverFd)
2517 {
2518 Port *port;
2519
2520 if (!(port = (Port *) calloc(1, sizeof(Port))))
2521 {
2522 ereport(LOG,
2523 (errcode(ERRCODE_OUT_OF_MEMORY),
2524 errmsg("out of memory")));
2525 ExitPostmaster(1);
2526 }
2527
2528 if (StreamConnection(serverFd, port) != STATUS_OK)
2529 {
2530 if (port->sock != PGINVALID_SOCKET)
2531 StreamClose(port->sock);
2532 ConnFree(port);
2533 return NULL;
2534 }
2535
2536 return port;
2537 }
2538
2539
2540 /*
2541 * ConnFree -- free a local connection data structure
2542 *
2543 * Caller has already closed the socket if any, so there's not much
2544 * to do here.
2545 */
2546 static void
ConnFree(Port * conn)2547 ConnFree(Port *conn)
2548 {
2549 free(conn);
2550 }
2551
2552
2553 /*
2554 * ClosePostmasterPorts -- close all the postmaster's open sockets
2555 *
2556 * This is called during child process startup to release file descriptors
2557 * that are not needed by that child process. The postmaster still has
2558 * them open, of course.
2559 *
2560 * Note: we pass am_syslogger as a boolean because we don't want to set
2561 * the global variable yet when this is called.
2562 */
2563 void
ClosePostmasterPorts(bool am_syslogger)2564 ClosePostmasterPorts(bool am_syslogger)
2565 {
2566 int i;
2567
2568 #ifndef WIN32
2569
2570 /*
2571 * Close the write end of postmaster death watch pipe. It's important to
2572 * do this as early as possible, so that if postmaster dies, others won't
2573 * think that it's still running because we're holding the pipe open.
2574 */
2575 if (close(postmaster_alive_fds[POSTMASTER_FD_OWN]))
2576 ereport(FATAL,
2577 (errcode_for_file_access(),
2578 errmsg_internal("could not close postmaster death monitoring pipe in child process: %m")));
2579 postmaster_alive_fds[POSTMASTER_FD_OWN] = -1;
2580 #endif
2581
2582 /* Close the listen sockets */
2583 for (i = 0; i < MAXLISTEN; i++)
2584 {
2585 if (ListenSocket[i] != PGINVALID_SOCKET)
2586 {
2587 StreamClose(ListenSocket[i]);
2588 ListenSocket[i] = PGINVALID_SOCKET;
2589 }
2590 }
2591
2592 /* If using syslogger, close the read side of the pipe */
2593 if (!am_syslogger)
2594 {
2595 #ifndef WIN32
2596 if (syslogPipe[0] >= 0)
2597 close(syslogPipe[0]);
2598 syslogPipe[0] = -1;
2599 #else
2600 if (syslogPipe[0])
2601 CloseHandle(syslogPipe[0]);
2602 syslogPipe[0] = 0;
2603 #endif
2604 }
2605
2606 #ifdef USE_BONJOUR
2607 /* If using Bonjour, close the connection to the mDNS daemon */
2608 if (bonjour_sdref)
2609 close(DNSServiceRefSockFD(bonjour_sdref));
2610 #endif
2611 }
2612
2613
2614 /*
2615 * InitProcessGlobals -- set MyProcPid, MyStartTime[stamp], random seeds
2616 *
2617 * Called early in the postmaster and every backend.
2618 */
2619 void
InitProcessGlobals(void)2620 InitProcessGlobals(void)
2621 {
2622 unsigned int rseed;
2623
2624 MyProcPid = getpid();
2625 MyStartTimestamp = GetCurrentTimestamp();
2626 MyStartTime = timestamptz_to_time_t(MyStartTimestamp);
2627
2628 /*
2629 * Set a different seed for random() in every process. We want something
2630 * unpredictable, so if possible, use high-quality random bits for the
2631 * seed. Otherwise, fall back to a seed based on timestamp and PID.
2632 */
2633 if (!pg_strong_random(&rseed, sizeof(rseed)))
2634 {
2635 /*
2636 * Since PIDs and timestamps tend to change more frequently in their
2637 * least significant bits, shift the timestamp left to allow a larger
2638 * total number of seeds in a given time period. Since that would
2639 * leave only 20 bits of the timestamp that cycle every ~1 second,
2640 * also mix in some higher bits.
2641 */
2642 rseed = ((uint64) MyProcPid) ^
2643 ((uint64) MyStartTimestamp << 12) ^
2644 ((uint64) MyStartTimestamp >> 20);
2645 }
2646 srandom(rseed);
2647 }
2648
2649
2650 /*
2651 * reset_shared -- reset shared memory and semaphores
2652 */
2653 static void
reset_shared(int port)2654 reset_shared(int port)
2655 {
2656 /*
2657 * Create or re-create shared memory and semaphores.
2658 *
2659 * Note: in each "cycle of life" we will normally assign the same IPC keys
2660 * (if using SysV shmem and/or semas), since the port number is used to
2661 * determine IPC keys. This helps ensure that we will clean up dead IPC
2662 * objects if the postmaster crashes and is restarted.
2663 */
2664 CreateSharedMemoryAndSemaphores(port);
2665 }
2666
2667
2668 /*
2669 * SIGHUP -- reread config files, and tell children to do same
2670 */
2671 static void
SIGHUP_handler(SIGNAL_ARGS)2672 SIGHUP_handler(SIGNAL_ARGS)
2673 {
2674 int save_errno = errno;
2675
2676 /*
2677 * We rely on the signal mechanism to have blocked all signals ... except
2678 * on Windows, which lacks sigaction(), so we have to do it manually.
2679 */
2680 #ifdef WIN32
2681 PG_SETMASK(&BlockSig);
2682 #endif
2683
2684 if (Shutdown <= SmartShutdown)
2685 {
2686 ereport(LOG,
2687 (errmsg("received SIGHUP, reloading configuration files")));
2688 ProcessConfigFile(PGC_SIGHUP);
2689 SignalChildren(SIGHUP);
2690 if (StartupPID != 0)
2691 signal_child(StartupPID, SIGHUP);
2692 if (BgWriterPID != 0)
2693 signal_child(BgWriterPID, SIGHUP);
2694 if (CheckpointerPID != 0)
2695 signal_child(CheckpointerPID, SIGHUP);
2696 if (WalWriterPID != 0)
2697 signal_child(WalWriterPID, SIGHUP);
2698 if (WalReceiverPID != 0)
2699 signal_child(WalReceiverPID, SIGHUP);
2700 if (AutoVacPID != 0)
2701 signal_child(AutoVacPID, SIGHUP);
2702 if (PgArchPID != 0)
2703 signal_child(PgArchPID, SIGHUP);
2704 if (SysLoggerPID != 0)
2705 signal_child(SysLoggerPID, SIGHUP);
2706 if (PgStatPID != 0)
2707 signal_child(PgStatPID, SIGHUP);
2708
2709 /* Reload authentication config files too */
2710 if (!load_hba())
2711 ereport(LOG,
2712 /* translator: %s is a configuration file */
2713 (errmsg("%s was not reloaded", "pg_hba.conf")));
2714
2715 if (!load_ident())
2716 ereport(LOG,
2717 (errmsg("%s was not reloaded", "pg_ident.conf")));
2718
2719 #ifdef USE_SSL
2720 /* Reload SSL configuration as well */
2721 if (EnableSSL)
2722 {
2723 if (secure_initialize(false) == 0)
2724 LoadedSSL = true;
2725 else
2726 ereport(LOG,
2727 (errmsg("SSL configuration was not reloaded")));
2728 }
2729 else
2730 {
2731 secure_destroy();
2732 LoadedSSL = false;
2733 }
2734 #endif
2735
2736 #ifdef EXEC_BACKEND
2737 /* Update the starting-point file for future children */
2738 write_nondefault_variables(PGC_SIGHUP);
2739 #endif
2740 }
2741
2742 #ifdef WIN32
2743 PG_SETMASK(&UnBlockSig);
2744 #endif
2745
2746 errno = save_errno;
2747 }
2748
2749
2750 /*
2751 * pmdie -- signal handler for processing various postmaster signals.
2752 */
2753 static void
pmdie(SIGNAL_ARGS)2754 pmdie(SIGNAL_ARGS)
2755 {
2756 int save_errno = errno;
2757
2758 /*
2759 * We rely on the signal mechanism to have blocked all signals ... except
2760 * on Windows, which lacks sigaction(), so we have to do it manually.
2761 */
2762 #ifdef WIN32
2763 PG_SETMASK(&BlockSig);
2764 #endif
2765
2766 ereport(DEBUG2,
2767 (errmsg_internal("postmaster received signal %d",
2768 postgres_signal_arg)));
2769
2770 switch (postgres_signal_arg)
2771 {
2772 case SIGTERM:
2773
2774 /*
2775 * Smart Shutdown:
2776 *
2777 * Wait for children to end their work, then shut down.
2778 */
2779 if (Shutdown >= SmartShutdown)
2780 break;
2781 Shutdown = SmartShutdown;
2782 ereport(LOG,
2783 (errmsg("received smart shutdown request")));
2784
2785 /* Report status */
2786 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2787 #ifdef USE_SYSTEMD
2788 sd_notify(0, "STOPPING=1");
2789 #endif
2790
2791 /*
2792 * If we reached normal running, we have to wait for any online
2793 * backup mode to end; otherwise go straight to waiting for client
2794 * backends to exit. (The difference is that in the former state,
2795 * we'll still let in new superuser clients, so that somebody can
2796 * end the online backup mode.) If already in PM_STOP_BACKENDS or
2797 * a later state, do not change it.
2798 */
2799 if (pmState == PM_RUN)
2800 connsAllowed = ALLOW_SUPERUSER_CONNS;
2801 else if (pmState == PM_HOT_STANDBY)
2802 connsAllowed = ALLOW_NO_CONNS;
2803 else if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2804 {
2805 /* There should be no clients, so proceed to stop children */
2806 pmState = PM_STOP_BACKENDS;
2807 }
2808
2809 /*
2810 * Now wait for online backup mode to end and backends to exit. If
2811 * that is already the case, PostmasterStateMachine will take the
2812 * next step.
2813 */
2814 PostmasterStateMachine();
2815 break;
2816
2817 case SIGINT:
2818
2819 /*
2820 * Fast Shutdown:
2821 *
2822 * Abort all children with SIGTERM (rollback active transactions
2823 * and exit) and shut down when they are gone.
2824 */
2825 if (Shutdown >= FastShutdown)
2826 break;
2827 Shutdown = FastShutdown;
2828 ereport(LOG,
2829 (errmsg("received fast shutdown request")));
2830
2831 /* Report status */
2832 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2833 #ifdef USE_SYSTEMD
2834 sd_notify(0, "STOPPING=1");
2835 #endif
2836
2837 if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2838 {
2839 /* Just shut down background processes silently */
2840 pmState = PM_STOP_BACKENDS;
2841 }
2842 else if (pmState == PM_RUN ||
2843 pmState == PM_HOT_STANDBY)
2844 {
2845 /* Report that we're about to zap live client sessions */
2846 ereport(LOG,
2847 (errmsg("aborting any active transactions")));
2848 pmState = PM_STOP_BACKENDS;
2849 }
2850
2851 /*
2852 * PostmasterStateMachine will issue any necessary signals, or
2853 * take the next step if no child processes need to be killed.
2854 */
2855 PostmasterStateMachine();
2856 break;
2857
2858 case SIGQUIT:
2859
2860 /*
2861 * Immediate Shutdown:
2862 *
2863 * abort all children with SIGQUIT, wait for them to exit,
2864 * terminate remaining ones with SIGKILL, then exit without
2865 * attempt to properly shut down the data base system.
2866 */
2867 if (Shutdown >= ImmediateShutdown)
2868 break;
2869 Shutdown = ImmediateShutdown;
2870 ereport(LOG,
2871 (errmsg("received immediate shutdown request")));
2872
2873 /* Report status */
2874 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2875 #ifdef USE_SYSTEMD
2876 sd_notify(0, "STOPPING=1");
2877 #endif
2878
2879 TerminateChildren(SIGQUIT);
2880 pmState = PM_WAIT_BACKENDS;
2881
2882 /* set stopwatch for them to die */
2883 AbortStartTime = time(NULL);
2884
2885 /*
2886 * Now wait for backends to exit. If there are none,
2887 * PostmasterStateMachine will take the next step.
2888 */
2889 PostmasterStateMachine();
2890 break;
2891 }
2892
2893 #ifdef WIN32
2894 PG_SETMASK(&UnBlockSig);
2895 #endif
2896
2897 errno = save_errno;
2898 }
2899
2900 /*
2901 * Reaper -- signal handler to cleanup after a child process dies.
2902 */
2903 static void
reaper(SIGNAL_ARGS)2904 reaper(SIGNAL_ARGS)
2905 {
2906 int save_errno = errno;
2907 int pid; /* process id of dead child process */
2908 int exitstatus; /* its exit status */
2909
2910 /*
2911 * We rely on the signal mechanism to have blocked all signals ... except
2912 * on Windows, which lacks sigaction(), so we have to do it manually.
2913 */
2914 #ifdef WIN32
2915 PG_SETMASK(&BlockSig);
2916 #endif
2917
2918 ereport(DEBUG4,
2919 (errmsg_internal("reaping dead processes")));
2920
2921 while ((pid = waitpid(-1, &exitstatus, WNOHANG)) > 0)
2922 {
2923 /*
2924 * Check if this child was a startup process.
2925 */
2926 if (pid == StartupPID)
2927 {
2928 StartupPID = 0;
2929
2930 /*
2931 * Startup process exited in response to a shutdown request (or it
2932 * completed normally regardless of the shutdown request).
2933 */
2934 if (Shutdown > NoShutdown &&
2935 (EXIT_STATUS_0(exitstatus) || EXIT_STATUS_1(exitstatus)))
2936 {
2937 StartupStatus = STARTUP_NOT_RUNNING;
2938 pmState = PM_WAIT_BACKENDS;
2939 /* PostmasterStateMachine logic does the rest */
2940 continue;
2941 }
2942
2943 if (EXIT_STATUS_3(exitstatus))
2944 {
2945 ereport(LOG,
2946 (errmsg("shutdown at recovery target")));
2947 StartupStatus = STARTUP_NOT_RUNNING;
2948 Shutdown = Max(Shutdown, SmartShutdown);
2949 TerminateChildren(SIGTERM);
2950 pmState = PM_WAIT_BACKENDS;
2951 /* PostmasterStateMachine logic does the rest */
2952 continue;
2953 }
2954
2955 /*
2956 * Unexpected exit of startup process (including FATAL exit)
2957 * during PM_STARTUP is treated as catastrophic. There are no
2958 * other processes running yet, so we can just exit.
2959 */
2960 if (pmState == PM_STARTUP && !EXIT_STATUS_0(exitstatus))
2961 {
2962 LogChildExit(LOG, _("startup process"),
2963 pid, exitstatus);
2964 ereport(LOG,
2965 (errmsg("aborting startup due to startup process failure")));
2966 ExitPostmaster(1);
2967 }
2968
2969 /*
2970 * After PM_STARTUP, any unexpected exit (including FATAL exit) of
2971 * the startup process is catastrophic, so kill other children,
2972 * and set StartupStatus so we don't try to reinitialize after
2973 * they're gone. Exception: if StartupStatus is STARTUP_SIGNALED,
2974 * then we previously sent the startup process a SIGQUIT; so
2975 * that's probably the reason it died, and we do want to try to
2976 * restart in that case.
2977 */
2978 if (!EXIT_STATUS_0(exitstatus))
2979 {
2980 if (StartupStatus == STARTUP_SIGNALED)
2981 StartupStatus = STARTUP_NOT_RUNNING;
2982 else
2983 StartupStatus = STARTUP_CRASHED;
2984 HandleChildCrash(pid, exitstatus,
2985 _("startup process"));
2986 continue;
2987 }
2988
2989 /*
2990 * Startup succeeded, commence normal operations
2991 */
2992 StartupStatus = STARTUP_NOT_RUNNING;
2993 FatalError = false;
2994 Assert(AbortStartTime == 0);
2995 ReachedNormalRunning = true;
2996 pmState = PM_RUN;
2997 connsAllowed = ALLOW_ALL_CONNS;
2998
2999 /*
3000 * Crank up the background tasks, if we didn't do that already
3001 * when we entered consistent recovery state. It doesn't matter
3002 * if this fails, we'll just try again later.
3003 */
3004 if (CheckpointerPID == 0)
3005 CheckpointerPID = StartCheckpointer();
3006 if (BgWriterPID == 0)
3007 BgWriterPID = StartBackgroundWriter();
3008 if (WalWriterPID == 0)
3009 WalWriterPID = StartWalWriter();
3010
3011 /*
3012 * Likewise, start other special children as needed. In a restart
3013 * situation, some of them may be alive already.
3014 */
3015 if (!IsBinaryUpgrade && AutoVacuumingActive() && AutoVacPID == 0)
3016 AutoVacPID = StartAutoVacLauncher();
3017 if (PgArchStartupAllowed() && PgArchPID == 0)
3018 PgArchPID = pgarch_start();
3019 if (PgStatPID == 0)
3020 PgStatPID = pgstat_start();
3021
3022 /* workers may be scheduled to start now */
3023 maybe_start_bgworkers();
3024
3025 /* at this point we are really open for business */
3026 ereport(LOG,
3027 (errmsg("database system is ready to accept connections")));
3028
3029 /* Report status */
3030 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
3031 #ifdef USE_SYSTEMD
3032 sd_notify(0, "READY=1");
3033 #endif
3034
3035 continue;
3036 }
3037
3038 /*
3039 * Was it the bgwriter? Normal exit can be ignored; we'll start a new
3040 * one at the next iteration of the postmaster's main loop, if
3041 * necessary. Any other exit condition is treated as a crash.
3042 */
3043 if (pid == BgWriterPID)
3044 {
3045 BgWriterPID = 0;
3046 if (!EXIT_STATUS_0(exitstatus))
3047 HandleChildCrash(pid, exitstatus,
3048 _("background writer process"));
3049 continue;
3050 }
3051
3052 /*
3053 * Was it the checkpointer?
3054 */
3055 if (pid == CheckpointerPID)
3056 {
3057 CheckpointerPID = 0;
3058 if (EXIT_STATUS_0(exitstatus) && pmState == PM_SHUTDOWN)
3059 {
3060 /*
3061 * OK, we saw normal exit of the checkpointer after it's been
3062 * told to shut down. We expect that it wrote a shutdown
3063 * checkpoint. (If for some reason it didn't, recovery will
3064 * occur on next postmaster start.)
3065 *
3066 * At this point we should have no normal backend children
3067 * left (else we'd not be in PM_SHUTDOWN state) but we might
3068 * have dead_end children to wait for.
3069 *
3070 * If we have an archiver subprocess, tell it to do a last
3071 * archive cycle and quit. Likewise, if we have walsender
3072 * processes, tell them to send any remaining WAL and quit.
3073 */
3074 Assert(Shutdown > NoShutdown);
3075
3076 /* Waken archiver for the last time */
3077 if (PgArchPID != 0)
3078 signal_child(PgArchPID, SIGUSR2);
3079
3080 /*
3081 * Waken walsenders for the last time. No regular backends
3082 * should be around anymore.
3083 */
3084 SignalChildren(SIGUSR2);
3085
3086 pmState = PM_SHUTDOWN_2;
3087
3088 /*
3089 * We can also shut down the stats collector now; there's
3090 * nothing left for it to do.
3091 */
3092 if (PgStatPID != 0)
3093 signal_child(PgStatPID, SIGQUIT);
3094 }
3095 else
3096 {
3097 /*
3098 * Any unexpected exit of the checkpointer (including FATAL
3099 * exit) is treated as a crash.
3100 */
3101 HandleChildCrash(pid, exitstatus,
3102 _("checkpointer process"));
3103 }
3104
3105 continue;
3106 }
3107
3108 /*
3109 * Was it the wal writer? Normal exit can be ignored; we'll start a
3110 * new one at the next iteration of the postmaster's main loop, if
3111 * necessary. Any other exit condition is treated as a crash.
3112 */
3113 if (pid == WalWriterPID)
3114 {
3115 WalWriterPID = 0;
3116 if (!EXIT_STATUS_0(exitstatus))
3117 HandleChildCrash(pid, exitstatus,
3118 _("WAL writer process"));
3119 continue;
3120 }
3121
3122 /*
3123 * Was it the wal receiver? If exit status is zero (normal) or one
3124 * (FATAL exit), we assume everything is all right just like normal
3125 * backends. (If we need a new wal receiver, we'll start one at the
3126 * next iteration of the postmaster's main loop.)
3127 */
3128 if (pid == WalReceiverPID)
3129 {
3130 WalReceiverPID = 0;
3131 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3132 HandleChildCrash(pid, exitstatus,
3133 _("WAL receiver process"));
3134 continue;
3135 }
3136
3137 /*
3138 * Was it the autovacuum launcher? Normal exit can be ignored; we'll
3139 * start a new one at the next iteration of the postmaster's main
3140 * loop, if necessary. Any other exit condition is treated as a
3141 * crash.
3142 */
3143 if (pid == AutoVacPID)
3144 {
3145 AutoVacPID = 0;
3146 if (!EXIT_STATUS_0(exitstatus))
3147 HandleChildCrash(pid, exitstatus,
3148 _("autovacuum launcher process"));
3149 continue;
3150 }
3151
3152 /*
3153 * Was it the archiver? If so, just try to start a new one; no need
3154 * to force reset of the rest of the system. (If fail, we'll try
3155 * again in future cycles of the main loop.). Unless we were waiting
3156 * for it to shut down; don't restart it in that case, and
3157 * PostmasterStateMachine() will advance to the next shutdown step.
3158 */
3159 if (pid == PgArchPID)
3160 {
3161 PgArchPID = 0;
3162 if (!EXIT_STATUS_0(exitstatus))
3163 LogChildExit(LOG, _("archiver process"),
3164 pid, exitstatus);
3165 if (PgArchStartupAllowed())
3166 PgArchPID = pgarch_start();
3167 continue;
3168 }
3169
3170 /*
3171 * Was it the statistics collector? If so, just try to start a new
3172 * one; no need to force reset of the rest of the system. (If fail,
3173 * we'll try again in future cycles of the main loop.)
3174 */
3175 if (pid == PgStatPID)
3176 {
3177 PgStatPID = 0;
3178 if (!EXIT_STATUS_0(exitstatus))
3179 LogChildExit(LOG, _("statistics collector process"),
3180 pid, exitstatus);
3181 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3182 PgStatPID = pgstat_start();
3183 continue;
3184 }
3185
3186 /* Was it the system logger? If so, try to start a new one */
3187 if (pid == SysLoggerPID)
3188 {
3189 SysLoggerPID = 0;
3190 /* for safety's sake, launch new logger *first* */
3191 SysLoggerPID = SysLogger_Start();
3192 if (!EXIT_STATUS_0(exitstatus))
3193 LogChildExit(LOG, _("system logger process"),
3194 pid, exitstatus);
3195 continue;
3196 }
3197
3198 /* Was it one of our background workers? */
3199 if (CleanupBackgroundWorker(pid, exitstatus))
3200 {
3201 /* have it be restarted */
3202 HaveCrashedWorker = true;
3203 continue;
3204 }
3205
3206 /*
3207 * Else do standard backend child cleanup.
3208 */
3209 CleanupBackend(pid, exitstatus);
3210 } /* loop over pending child-death reports */
3211
3212 /*
3213 * After cleaning out the SIGCHLD queue, see if we have any state changes
3214 * or actions to make.
3215 */
3216 PostmasterStateMachine();
3217
3218 /* Done with signal handler */
3219 #ifdef WIN32
3220 PG_SETMASK(&UnBlockSig);
3221 #endif
3222
3223 errno = save_errno;
3224 }
3225
3226 /*
3227 * Scan the bgworkers list and see if the given PID (which has just stopped
3228 * or crashed) is in it. Handle its shutdown if so, and return true. If not a
3229 * bgworker, return false.
3230 *
3231 * This is heavily based on CleanupBackend. One important difference is that
3232 * we don't know yet that the dying process is a bgworker, so we must be silent
3233 * until we're sure it is.
3234 */
3235 static bool
CleanupBackgroundWorker(int pid,int exitstatus)3236 CleanupBackgroundWorker(int pid,
3237 int exitstatus) /* child's exit status */
3238 {
3239 char namebuf[MAXPGPATH];
3240 slist_mutable_iter iter;
3241
3242 slist_foreach_modify(iter, &BackgroundWorkerList)
3243 {
3244 RegisteredBgWorker *rw;
3245
3246 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
3247
3248 if (rw->rw_pid != pid)
3249 continue;
3250
3251 #ifdef WIN32
3252 /* see CleanupBackend */
3253 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3254 exitstatus = 0;
3255 #endif
3256
3257 snprintf(namebuf, MAXPGPATH, _("background worker \"%s\""),
3258 rw->rw_worker.bgw_type);
3259
3260
3261 if (!EXIT_STATUS_0(exitstatus))
3262 {
3263 /* Record timestamp, so we know when to restart the worker. */
3264 rw->rw_crashed_at = GetCurrentTimestamp();
3265 }
3266 else
3267 {
3268 /* Zero exit status means terminate */
3269 rw->rw_crashed_at = 0;
3270 rw->rw_terminate = true;
3271 }
3272
3273 /*
3274 * Additionally, for shared-memory-connected workers, just like a
3275 * backend, any exit status other than 0 or 1 is considered a crash
3276 * and causes a system-wide restart.
3277 */
3278 if ((rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3279 {
3280 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3281 {
3282 HandleChildCrash(pid, exitstatus, namebuf);
3283 return true;
3284 }
3285 }
3286
3287 /*
3288 * We must release the postmaster child slot whether this worker is
3289 * connected to shared memory or not, but we only treat it as a crash
3290 * if it is in fact connected.
3291 */
3292 if (!ReleasePostmasterChildSlot(rw->rw_child_slot) &&
3293 (rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3294 {
3295 HandleChildCrash(pid, exitstatus, namebuf);
3296 return true;
3297 }
3298
3299 /* Get it out of the BackendList and clear out remaining data */
3300 dlist_delete(&rw->rw_backend->elem);
3301 #ifdef EXEC_BACKEND
3302 ShmemBackendArrayRemove(rw->rw_backend);
3303 #endif
3304
3305 /*
3306 * It's possible that this background worker started some OTHER
3307 * background worker and asked to be notified when that worker started
3308 * or stopped. If so, cancel any notifications destined for the
3309 * now-dead backend.
3310 */
3311 if (rw->rw_backend->bgworker_notify)
3312 BackgroundWorkerStopNotifications(rw->rw_pid);
3313 free(rw->rw_backend);
3314 rw->rw_backend = NULL;
3315 rw->rw_pid = 0;
3316 rw->rw_child_slot = 0;
3317 ReportBackgroundWorkerExit(&iter); /* report child death */
3318
3319 LogChildExit(EXIT_STATUS_0(exitstatus) ? DEBUG1 : LOG,
3320 namebuf, pid, exitstatus);
3321
3322 return true;
3323 }
3324
3325 return false;
3326 }
3327
3328 /*
3329 * CleanupBackend -- cleanup after terminated backend.
3330 *
3331 * Remove all local state associated with backend.
3332 *
3333 * If you change this, see also CleanupBackgroundWorker.
3334 */
3335 static void
CleanupBackend(int pid,int exitstatus)3336 CleanupBackend(int pid,
3337 int exitstatus) /* child's exit status. */
3338 {
3339 dlist_mutable_iter iter;
3340
3341 LogChildExit(DEBUG2, _("server process"), pid, exitstatus);
3342
3343 /*
3344 * If a backend dies in an ugly way then we must signal all other backends
3345 * to quickdie. If exit status is zero (normal) or one (FATAL exit), we
3346 * assume everything is all right and proceed to remove the backend from
3347 * the active backend list.
3348 */
3349
3350 #ifdef WIN32
3351
3352 /*
3353 * On win32, also treat ERROR_WAIT_NO_CHILDREN (128) as nonfatal case,
3354 * since that sometimes happens under load when the process fails to start
3355 * properly (long before it starts using shared memory). Microsoft reports
3356 * it is related to mutex failure:
3357 * http://archives.postgresql.org/pgsql-hackers/2010-09/msg00790.php
3358 */
3359 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3360 {
3361 LogChildExit(LOG, _("server process"), pid, exitstatus);
3362 exitstatus = 0;
3363 }
3364 #endif
3365
3366 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3367 {
3368 HandleChildCrash(pid, exitstatus, _("server process"));
3369 return;
3370 }
3371
3372 dlist_foreach_modify(iter, &BackendList)
3373 {
3374 Backend *bp = dlist_container(Backend, elem, iter.cur);
3375
3376 if (bp->pid == pid)
3377 {
3378 if (!bp->dead_end)
3379 {
3380 if (!ReleasePostmasterChildSlot(bp->child_slot))
3381 {
3382 /*
3383 * Uh-oh, the child failed to clean itself up. Treat as a
3384 * crash after all.
3385 */
3386 HandleChildCrash(pid, exitstatus, _("server process"));
3387 return;
3388 }
3389 #ifdef EXEC_BACKEND
3390 ShmemBackendArrayRemove(bp);
3391 #endif
3392 }
3393 if (bp->bgworker_notify)
3394 {
3395 /*
3396 * This backend may have been slated to receive SIGUSR1 when
3397 * some background worker started or stopped. Cancel those
3398 * notifications, as we don't want to signal PIDs that are not
3399 * PostgreSQL backends. This gets skipped in the (probably
3400 * very common) case where the backend has never requested any
3401 * such notifications.
3402 */
3403 BackgroundWorkerStopNotifications(bp->pid);
3404 }
3405 dlist_delete(iter.cur);
3406 free(bp);
3407 break;
3408 }
3409 }
3410 }
3411
3412 /*
3413 * HandleChildCrash -- cleanup after failed backend, bgwriter, checkpointer,
3414 * walwriter, autovacuum, or background worker.
3415 *
3416 * The objectives here are to clean up our local state about the child
3417 * process, and to signal all other remaining children to quickdie.
3418 */
3419 static void
HandleChildCrash(int pid,int exitstatus,const char * procname)3420 HandleChildCrash(int pid, int exitstatus, const char *procname)
3421 {
3422 dlist_mutable_iter iter;
3423 slist_iter siter;
3424 Backend *bp;
3425 bool take_action;
3426
3427 /*
3428 * We only log messages and send signals if this is the first process
3429 * crash and we're not doing an immediate shutdown; otherwise, we're only
3430 * here to update postmaster's idea of live processes. If we have already
3431 * signalled children, nonzero exit status is to be expected, so don't
3432 * clutter log.
3433 */
3434 take_action = !FatalError && Shutdown != ImmediateShutdown;
3435
3436 if (take_action)
3437 {
3438 LogChildExit(LOG, procname, pid, exitstatus);
3439 ereport(LOG,
3440 (errmsg("terminating any other active server processes")));
3441 }
3442
3443 /* Process background workers. */
3444 slist_foreach(siter, &BackgroundWorkerList)
3445 {
3446 RegisteredBgWorker *rw;
3447
3448 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
3449 if (rw->rw_pid == 0)
3450 continue; /* not running */
3451 if (rw->rw_pid == pid)
3452 {
3453 /*
3454 * Found entry for freshly-dead worker, so remove it.
3455 */
3456 (void) ReleasePostmasterChildSlot(rw->rw_child_slot);
3457 dlist_delete(&rw->rw_backend->elem);
3458 #ifdef EXEC_BACKEND
3459 ShmemBackendArrayRemove(rw->rw_backend);
3460 #endif
3461 free(rw->rw_backend);
3462 rw->rw_backend = NULL;
3463 rw->rw_pid = 0;
3464 rw->rw_child_slot = 0;
3465 /* don't reset crashed_at */
3466 /* don't report child stop, either */
3467 /* Keep looping so we can signal remaining workers */
3468 }
3469 else
3470 {
3471 /*
3472 * This worker is still alive. Unless we did so already, tell it
3473 * to commit hara-kiri.
3474 *
3475 * SIGQUIT is the special signal that says exit without proc_exit
3476 * and let the user know what's going on. But if SendStop is set
3477 * (-s on command line), then we send SIGSTOP instead, so that we
3478 * can get core dumps from all backends by hand.
3479 */
3480 if (take_action)
3481 {
3482 ereport(DEBUG2,
3483 (errmsg_internal("sending %s to process %d",
3484 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3485 (int) rw->rw_pid)));
3486 signal_child(rw->rw_pid, (SendStop ? SIGSTOP : SIGQUIT));
3487 }
3488 }
3489 }
3490
3491 /* Process regular backends */
3492 dlist_foreach_modify(iter, &BackendList)
3493 {
3494 bp = dlist_container(Backend, elem, iter.cur);
3495
3496 if (bp->pid == pid)
3497 {
3498 /*
3499 * Found entry for freshly-dead backend, so remove it.
3500 */
3501 if (!bp->dead_end)
3502 {
3503 (void) ReleasePostmasterChildSlot(bp->child_slot);
3504 #ifdef EXEC_BACKEND
3505 ShmemBackendArrayRemove(bp);
3506 #endif
3507 }
3508 dlist_delete(iter.cur);
3509 free(bp);
3510 /* Keep looping so we can signal remaining backends */
3511 }
3512 else
3513 {
3514 /*
3515 * This backend is still alive. Unless we did so already, tell it
3516 * to commit hara-kiri.
3517 *
3518 * SIGQUIT is the special signal that says exit without proc_exit
3519 * and let the user know what's going on. But if SendStop is set
3520 * (-s on command line), then we send SIGSTOP instead, so that we
3521 * can get core dumps from all backends by hand.
3522 *
3523 * We could exclude dead_end children here, but at least in the
3524 * SIGSTOP case it seems better to include them.
3525 *
3526 * Background workers were already processed above; ignore them
3527 * here.
3528 */
3529 if (bp->bkend_type == BACKEND_TYPE_BGWORKER)
3530 continue;
3531
3532 if (take_action)
3533 {
3534 ereport(DEBUG2,
3535 (errmsg_internal("sending %s to process %d",
3536 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3537 (int) bp->pid)));
3538 signal_child(bp->pid, (SendStop ? SIGSTOP : SIGQUIT));
3539 }
3540 }
3541 }
3542
3543 /* Take care of the startup process too */
3544 if (pid == StartupPID)
3545 {
3546 StartupPID = 0;
3547 StartupStatus = STARTUP_CRASHED;
3548 }
3549 else if (StartupPID != 0 && take_action)
3550 {
3551 ereport(DEBUG2,
3552 (errmsg_internal("sending %s to process %d",
3553 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3554 (int) StartupPID)));
3555 signal_child(StartupPID, (SendStop ? SIGSTOP : SIGQUIT));
3556 StartupStatus = STARTUP_SIGNALED;
3557 }
3558
3559 /* Take care of the bgwriter too */
3560 if (pid == BgWriterPID)
3561 BgWriterPID = 0;
3562 else if (BgWriterPID != 0 && take_action)
3563 {
3564 ereport(DEBUG2,
3565 (errmsg_internal("sending %s to process %d",
3566 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3567 (int) BgWriterPID)));
3568 signal_child(BgWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3569 }
3570
3571 /* Take care of the checkpointer too */
3572 if (pid == CheckpointerPID)
3573 CheckpointerPID = 0;
3574 else if (CheckpointerPID != 0 && take_action)
3575 {
3576 ereport(DEBUG2,
3577 (errmsg_internal("sending %s to process %d",
3578 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3579 (int) CheckpointerPID)));
3580 signal_child(CheckpointerPID, (SendStop ? SIGSTOP : SIGQUIT));
3581 }
3582
3583 /* Take care of the walwriter too */
3584 if (pid == WalWriterPID)
3585 WalWriterPID = 0;
3586 else if (WalWriterPID != 0 && take_action)
3587 {
3588 ereport(DEBUG2,
3589 (errmsg_internal("sending %s to process %d",
3590 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3591 (int) WalWriterPID)));
3592 signal_child(WalWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3593 }
3594
3595 /* Take care of the walreceiver too */
3596 if (pid == WalReceiverPID)
3597 WalReceiverPID = 0;
3598 else if (WalReceiverPID != 0 && take_action)
3599 {
3600 ereport(DEBUG2,
3601 (errmsg_internal("sending %s to process %d",
3602 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3603 (int) WalReceiverPID)));
3604 signal_child(WalReceiverPID, (SendStop ? SIGSTOP : SIGQUIT));
3605 }
3606
3607 /* Take care of the autovacuum launcher too */
3608 if (pid == AutoVacPID)
3609 AutoVacPID = 0;
3610 else if (AutoVacPID != 0 && take_action)
3611 {
3612 ereport(DEBUG2,
3613 (errmsg_internal("sending %s to process %d",
3614 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3615 (int) AutoVacPID)));
3616 signal_child(AutoVacPID, (SendStop ? SIGSTOP : SIGQUIT));
3617 }
3618
3619 /*
3620 * Force a power-cycle of the pgarch process too. (This isn't absolutely
3621 * necessary, but it seems like a good idea for robustness, and it
3622 * simplifies the state-machine logic in the case where a shutdown request
3623 * arrives during crash processing.)
3624 */
3625 if (PgArchPID != 0 && take_action)
3626 {
3627 ereport(DEBUG2,
3628 (errmsg_internal("sending %s to process %d",
3629 "SIGQUIT",
3630 (int) PgArchPID)));
3631 signal_child(PgArchPID, SIGQUIT);
3632 }
3633
3634 /*
3635 * Force a power-cycle of the pgstat process too. (This isn't absolutely
3636 * necessary, but it seems like a good idea for robustness, and it
3637 * simplifies the state-machine logic in the case where a shutdown request
3638 * arrives during crash processing.)
3639 */
3640 if (PgStatPID != 0 && take_action)
3641 {
3642 ereport(DEBUG2,
3643 (errmsg_internal("sending %s to process %d",
3644 "SIGQUIT",
3645 (int) PgStatPID)));
3646 signal_child(PgStatPID, SIGQUIT);
3647 allow_immediate_pgstat_restart();
3648 }
3649
3650 /* We do NOT restart the syslogger */
3651
3652 if (Shutdown != ImmediateShutdown)
3653 FatalError = true;
3654
3655 /* We now transit into a state of waiting for children to die */
3656 if (pmState == PM_RECOVERY ||
3657 pmState == PM_HOT_STANDBY ||
3658 pmState == PM_RUN ||
3659 pmState == PM_STOP_BACKENDS ||
3660 pmState == PM_SHUTDOWN)
3661 pmState = PM_WAIT_BACKENDS;
3662
3663 /*
3664 * .. and if this doesn't happen quickly enough, now the clock is ticking
3665 * for us to kill them without mercy.
3666 */
3667 if (AbortStartTime == 0)
3668 AbortStartTime = time(NULL);
3669 }
3670
3671 /*
3672 * Log the death of a child process.
3673 */
3674 static void
LogChildExit(int lev,const char * procname,int pid,int exitstatus)3675 LogChildExit(int lev, const char *procname, int pid, int exitstatus)
3676 {
3677 /*
3678 * size of activity_buffer is arbitrary, but set equal to default
3679 * track_activity_query_size
3680 */
3681 char activity_buffer[1024];
3682 const char *activity = NULL;
3683
3684 if (!EXIT_STATUS_0(exitstatus))
3685 activity = pgstat_get_crashed_backend_activity(pid,
3686 activity_buffer,
3687 sizeof(activity_buffer));
3688
3689 if (WIFEXITED(exitstatus))
3690 ereport(lev,
3691
3692 /*------
3693 translator: %s is a noun phrase describing a child process, such as
3694 "server process" */
3695 (errmsg("%s (PID %d) exited with exit code %d",
3696 procname, pid, WEXITSTATUS(exitstatus)),
3697 activity ? errdetail("Failed process was running: %s", activity) : 0));
3698 else if (WIFSIGNALED(exitstatus))
3699 {
3700 #if defined(WIN32)
3701 ereport(lev,
3702
3703 /*------
3704 translator: %s is a noun phrase describing a child process, such as
3705 "server process" */
3706 (errmsg("%s (PID %d) was terminated by exception 0x%X",
3707 procname, pid, WTERMSIG(exitstatus)),
3708 errhint("See C include file \"ntstatus.h\" for a description of the hexadecimal value."),
3709 activity ? errdetail("Failed process was running: %s", activity) : 0));
3710 #else
3711 ereport(lev,
3712
3713 /*------
3714 translator: %s is a noun phrase describing a child process, such as
3715 "server process" */
3716 (errmsg("%s (PID %d) was terminated by signal %d: %s",
3717 procname, pid, WTERMSIG(exitstatus),
3718 pg_strsignal(WTERMSIG(exitstatus))),
3719 activity ? errdetail("Failed process was running: %s", activity) : 0));
3720 #endif
3721 }
3722 else
3723 ereport(lev,
3724
3725 /*------
3726 translator: %s is a noun phrase describing a child process, such as
3727 "server process" */
3728 (errmsg("%s (PID %d) exited with unrecognized status %d",
3729 procname, pid, exitstatus),
3730 activity ? errdetail("Failed process was running: %s", activity) : 0));
3731 }
3732
3733 /*
3734 * Advance the postmaster's state machine and take actions as appropriate
3735 *
3736 * This is common code for pmdie(), reaper() and sigusr1_handler(), which
3737 * receive the signals that might mean we need to change state.
3738 */
3739 static void
PostmasterStateMachine(void)3740 PostmasterStateMachine(void)
3741 {
3742 /* If we're doing a smart shutdown, try to advance that state. */
3743 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3744 {
3745 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
3746 {
3747 /*
3748 * ALLOW_SUPERUSER_CONNS state ends as soon as online backup mode
3749 * is not active.
3750 */
3751 if (!BackupInProgress())
3752 connsAllowed = ALLOW_NO_CONNS;
3753 }
3754
3755 if (connsAllowed == ALLOW_NO_CONNS)
3756 {
3757 /*
3758 * ALLOW_NO_CONNS state ends when we have no normal client
3759 * backends running. Then we're ready to stop other children.
3760 */
3761 if (CountChildren(BACKEND_TYPE_NORMAL) == 0)
3762 pmState = PM_STOP_BACKENDS;
3763 }
3764 }
3765
3766 /*
3767 * If we're ready to do so, signal child processes to shut down. (This
3768 * isn't a persistent state, but treating it as a distinct pmState allows
3769 * us to share this code across multiple shutdown code paths.)
3770 */
3771 if (pmState == PM_STOP_BACKENDS)
3772 {
3773 /*
3774 * Forget any pending requests for background workers, since we're no
3775 * longer willing to launch any new workers. (If additional requests
3776 * arrive, BackgroundWorkerStateChange will reject them.)
3777 */
3778 ForgetUnstartedBackgroundWorkers();
3779
3780 /* Signal all backend children except walsenders */
3781 SignalSomeChildren(SIGTERM,
3782 BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND);
3783 /* and the autovac launcher too */
3784 if (AutoVacPID != 0)
3785 signal_child(AutoVacPID, SIGTERM);
3786 /* and the bgwriter too */
3787 if (BgWriterPID != 0)
3788 signal_child(BgWriterPID, SIGTERM);
3789 /* and the walwriter too */
3790 if (WalWriterPID != 0)
3791 signal_child(WalWriterPID, SIGTERM);
3792 /* If we're in recovery, also stop startup and walreceiver procs */
3793 if (StartupPID != 0)
3794 signal_child(StartupPID, SIGTERM);
3795 if (WalReceiverPID != 0)
3796 signal_child(WalReceiverPID, SIGTERM);
3797 /* checkpointer, archiver, stats, and syslogger may continue for now */
3798
3799 /* Now transition to PM_WAIT_BACKENDS state to wait for them to die */
3800 pmState = PM_WAIT_BACKENDS;
3801 }
3802
3803 /*
3804 * If we are in a state-machine state that implies waiting for backends to
3805 * exit, see if they're all gone, and change state if so.
3806 */
3807 if (pmState == PM_WAIT_BACKENDS)
3808 {
3809 /*
3810 * PM_WAIT_BACKENDS state ends when we have no regular backends
3811 * (including autovac workers), no bgworkers (including unconnected
3812 * ones), and no walwriter, autovac launcher or bgwriter. If we are
3813 * doing crash recovery or an immediate shutdown then we expect the
3814 * checkpointer to exit as well, otherwise not. The archiver, stats,
3815 * and syslogger processes are disregarded since they are not
3816 * connected to shared memory; we also disregard dead_end children
3817 * here. Walsenders are also disregarded, they will be terminated
3818 * later after writing the checkpoint record, like the archiver
3819 * process.
3820 */
3821 if (CountChildren(BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND) == 0 &&
3822 StartupPID == 0 &&
3823 WalReceiverPID == 0 &&
3824 BgWriterPID == 0 &&
3825 (CheckpointerPID == 0 ||
3826 (!FatalError && Shutdown < ImmediateShutdown)) &&
3827 WalWriterPID == 0 &&
3828 AutoVacPID == 0)
3829 {
3830 if (Shutdown >= ImmediateShutdown || FatalError)
3831 {
3832 /*
3833 * Start waiting for dead_end children to die. This state
3834 * change causes ServerLoop to stop creating new ones.
3835 */
3836 pmState = PM_WAIT_DEAD_END;
3837
3838 /*
3839 * We already SIGQUIT'd the archiver and stats processes, if
3840 * any, when we started immediate shutdown or entered
3841 * FatalError state.
3842 */
3843 }
3844 else
3845 {
3846 /*
3847 * If we get here, we are proceeding with normal shutdown. All
3848 * the regular children are gone, and it's time to tell the
3849 * checkpointer to do a shutdown checkpoint.
3850 */
3851 Assert(Shutdown > NoShutdown);
3852 /* Start the checkpointer if not running */
3853 if (CheckpointerPID == 0)
3854 CheckpointerPID = StartCheckpointer();
3855 /* And tell it to shut down */
3856 if (CheckpointerPID != 0)
3857 {
3858 signal_child(CheckpointerPID, SIGUSR2);
3859 pmState = PM_SHUTDOWN;
3860 }
3861 else
3862 {
3863 /*
3864 * If we failed to fork a checkpointer, just shut down.
3865 * Any required cleanup will happen at next restart. We
3866 * set FatalError so that an "abnormal shutdown" message
3867 * gets logged when we exit.
3868 */
3869 FatalError = true;
3870 pmState = PM_WAIT_DEAD_END;
3871
3872 /* Kill the walsenders, archiver and stats collector too */
3873 SignalChildren(SIGQUIT);
3874 if (PgArchPID != 0)
3875 signal_child(PgArchPID, SIGQUIT);
3876 if (PgStatPID != 0)
3877 signal_child(PgStatPID, SIGQUIT);
3878 }
3879 }
3880 }
3881 }
3882
3883 if (pmState == PM_SHUTDOWN_2)
3884 {
3885 /*
3886 * PM_SHUTDOWN_2 state ends when there's no other children than
3887 * dead_end children left. There shouldn't be any regular backends
3888 * left by now anyway; what we're really waiting for is walsenders and
3889 * archiver.
3890 */
3891 if (PgArchPID == 0 && CountChildren(BACKEND_TYPE_ALL) == 0)
3892 {
3893 pmState = PM_WAIT_DEAD_END;
3894 }
3895 }
3896
3897 if (pmState == PM_WAIT_DEAD_END)
3898 {
3899 /*
3900 * PM_WAIT_DEAD_END state ends when the BackendList is entirely empty
3901 * (ie, no dead_end children remain), and the archiver and stats
3902 * collector are gone too.
3903 *
3904 * The reason we wait for those two is to protect them against a new
3905 * postmaster starting conflicting subprocesses; this isn't an
3906 * ironclad protection, but it at least helps in the
3907 * shutdown-and-immediately-restart scenario. Note that they have
3908 * already been sent appropriate shutdown signals, either during a
3909 * normal state transition leading up to PM_WAIT_DEAD_END, or during
3910 * FatalError processing.
3911 */
3912 if (dlist_is_empty(&BackendList) &&
3913 PgArchPID == 0 && PgStatPID == 0)
3914 {
3915 /* These other guys should be dead already */
3916 Assert(StartupPID == 0);
3917 Assert(WalReceiverPID == 0);
3918 Assert(BgWriterPID == 0);
3919 Assert(CheckpointerPID == 0);
3920 Assert(WalWriterPID == 0);
3921 Assert(AutoVacPID == 0);
3922 /* syslogger is not considered here */
3923 pmState = PM_NO_CHILDREN;
3924 }
3925 }
3926
3927 /*
3928 * If we've been told to shut down, we exit as soon as there are no
3929 * remaining children. If there was a crash, cleanup will occur at the
3930 * next startup. (Before PostgreSQL 8.3, we tried to recover from the
3931 * crash before exiting, but that seems unwise if we are quitting because
3932 * we got SIGTERM from init --- there may well not be time for recovery
3933 * before init decides to SIGKILL us.)
3934 *
3935 * Note that the syslogger continues to run. It will exit when it sees
3936 * EOF on its input pipe, which happens when there are no more upstream
3937 * processes.
3938 */
3939 if (Shutdown > NoShutdown && pmState == PM_NO_CHILDREN)
3940 {
3941 if (FatalError)
3942 {
3943 ereport(LOG, (errmsg("abnormal database system shutdown")));
3944 ExitPostmaster(1);
3945 }
3946 else
3947 {
3948 /*
3949 * Terminate exclusive backup mode to avoid recovery after a clean
3950 * fast shutdown. Since an exclusive backup can only be taken
3951 * during normal running (and not, for example, while running
3952 * under Hot Standby) it only makes sense to do this if we reached
3953 * normal running. If we're still in recovery, the backup file is
3954 * one we're recovering *from*, and we must keep it around so that
3955 * recovery restarts from the right place.
3956 */
3957 if (ReachedNormalRunning)
3958 CancelBackup();
3959
3960 /* Normal exit from the postmaster is here */
3961 ExitPostmaster(0);
3962 }
3963 }
3964
3965 /*
3966 * If the startup process failed, or the user does not want an automatic
3967 * restart after backend crashes, wait for all non-syslogger children to
3968 * exit, and then exit postmaster. We don't try to reinitialize when the
3969 * startup process fails, because more than likely it will just fail again
3970 * and we will keep trying forever.
3971 */
3972 if (pmState == PM_NO_CHILDREN &&
3973 (StartupStatus == STARTUP_CRASHED || !restart_after_crash))
3974 ExitPostmaster(1);
3975
3976 /*
3977 * If we need to recover from a crash, wait for all non-syslogger children
3978 * to exit, then reset shmem and StartupDataBase.
3979 */
3980 if (FatalError && pmState == PM_NO_CHILDREN)
3981 {
3982 ereport(LOG,
3983 (errmsg("all server processes terminated; reinitializing")));
3984
3985 /* allow background workers to immediately restart */
3986 ResetBackgroundWorkerCrashTimes();
3987
3988 shmem_exit(1);
3989
3990 /* re-read control file into local memory */
3991 LocalProcessControlFile(true);
3992
3993 reset_shared(PostPortNumber);
3994
3995 StartupPID = StartupDataBase();
3996 Assert(StartupPID != 0);
3997 StartupStatus = STARTUP_RUNNING;
3998 pmState = PM_STARTUP;
3999 /* crash recovery started, reset SIGKILL flag */
4000 AbortStartTime = 0;
4001 }
4002 }
4003
4004
4005 /*
4006 * Send a signal to a postmaster child process
4007 *
4008 * On systems that have setsid(), each child process sets itself up as a
4009 * process group leader. For signals that are generally interpreted in the
4010 * appropriate fashion, we signal the entire process group not just the
4011 * direct child process. This allows us to, for example, SIGQUIT a blocked
4012 * archive_recovery script, or SIGINT a script being run by a backend via
4013 * system().
4014 *
4015 * There is a race condition for recently-forked children: they might not
4016 * have executed setsid() yet. So we signal the child directly as well as
4017 * the group. We assume such a child will handle the signal before trying
4018 * to spawn any grandchild processes. We also assume that signaling the
4019 * child twice will not cause any problems.
4020 */
4021 static void
signal_child(pid_t pid,int signal)4022 signal_child(pid_t pid, int signal)
4023 {
4024 if (kill(pid, signal) < 0)
4025 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) pid, signal);
4026 #ifdef HAVE_SETSID
4027 switch (signal)
4028 {
4029 case SIGINT:
4030 case SIGTERM:
4031 case SIGQUIT:
4032 case SIGSTOP:
4033 case SIGKILL:
4034 if (kill(-pid, signal) < 0)
4035 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) (-pid), signal);
4036 break;
4037 default:
4038 break;
4039 }
4040 #endif
4041 }
4042
4043 /*
4044 * Send a signal to the targeted children (but NOT special children;
4045 * dead_end children are never signaled, either).
4046 */
4047 static bool
SignalSomeChildren(int signal,int target)4048 SignalSomeChildren(int signal, int target)
4049 {
4050 dlist_iter iter;
4051 bool signaled = false;
4052
4053 dlist_foreach(iter, &BackendList)
4054 {
4055 Backend *bp = dlist_container(Backend, elem, iter.cur);
4056
4057 if (bp->dead_end)
4058 continue;
4059
4060 /*
4061 * Since target == BACKEND_TYPE_ALL is the most common case, we test
4062 * it first and avoid touching shared memory for every child.
4063 */
4064 if (target != BACKEND_TYPE_ALL)
4065 {
4066 /*
4067 * Assign bkend_type for any recently announced WAL Sender
4068 * processes.
4069 */
4070 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
4071 IsPostmasterChildWalSender(bp->child_slot))
4072 bp->bkend_type = BACKEND_TYPE_WALSND;
4073
4074 if (!(target & bp->bkend_type))
4075 continue;
4076 }
4077
4078 ereport(DEBUG4,
4079 (errmsg_internal("sending signal %d to process %d",
4080 signal, (int) bp->pid)));
4081 signal_child(bp->pid, signal);
4082 signaled = true;
4083 }
4084 return signaled;
4085 }
4086
4087 /*
4088 * Send a termination signal to children. This considers all of our children
4089 * processes, except syslogger and dead_end backends.
4090 */
4091 static void
TerminateChildren(int signal)4092 TerminateChildren(int signal)
4093 {
4094 SignalChildren(signal);
4095 if (StartupPID != 0)
4096 {
4097 signal_child(StartupPID, signal);
4098 if (signal == SIGQUIT || signal == SIGKILL)
4099 StartupStatus = STARTUP_SIGNALED;
4100 }
4101 if (BgWriterPID != 0)
4102 signal_child(BgWriterPID, signal);
4103 if (CheckpointerPID != 0)
4104 signal_child(CheckpointerPID, signal);
4105 if (WalWriterPID != 0)
4106 signal_child(WalWriterPID, signal);
4107 if (WalReceiverPID != 0)
4108 signal_child(WalReceiverPID, signal);
4109 if (AutoVacPID != 0)
4110 signal_child(AutoVacPID, signal);
4111 if (PgArchPID != 0)
4112 signal_child(PgArchPID, signal);
4113 if (PgStatPID != 0)
4114 signal_child(PgStatPID, signal);
4115 }
4116
4117 /*
4118 * BackendStartup -- start backend process
4119 *
4120 * returns: STATUS_ERROR if the fork failed, STATUS_OK otherwise.
4121 *
4122 * Note: if you change this code, also consider StartAutovacuumWorker.
4123 */
4124 static int
BackendStartup(Port * port)4125 BackendStartup(Port *port)
4126 {
4127 Backend *bn; /* for backend cleanup */
4128 pid_t pid;
4129
4130 /*
4131 * Create backend data structure. Better before the fork() so we can
4132 * handle failure cleanly.
4133 */
4134 bn = (Backend *) malloc(sizeof(Backend));
4135 if (!bn)
4136 {
4137 ereport(LOG,
4138 (errcode(ERRCODE_OUT_OF_MEMORY),
4139 errmsg("out of memory")));
4140 return STATUS_ERROR;
4141 }
4142
4143 /*
4144 * Compute the cancel key that will be assigned to this backend. The
4145 * backend will have its own copy in the forked-off process' value of
4146 * MyCancelKey, so that it can transmit the key to the frontend.
4147 */
4148 if (!RandomCancelKey(&MyCancelKey))
4149 {
4150 free(bn);
4151 ereport(LOG,
4152 (errcode(ERRCODE_INTERNAL_ERROR),
4153 errmsg("could not generate random cancel key")));
4154 return STATUS_ERROR;
4155 }
4156
4157 bn->cancel_key = MyCancelKey;
4158
4159 /* Pass down canAcceptConnections state */
4160 port->canAcceptConnections = canAcceptConnections(BACKEND_TYPE_NORMAL);
4161 bn->dead_end = (port->canAcceptConnections != CAC_OK &&
4162 port->canAcceptConnections != CAC_SUPERUSER);
4163
4164 /*
4165 * Unless it's a dead_end child, assign it a child slot number
4166 */
4167 if (!bn->dead_end)
4168 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
4169 else
4170 bn->child_slot = 0;
4171
4172 /* Hasn't asked to be notified about any bgworkers yet */
4173 bn->bgworker_notify = false;
4174
4175 #ifdef EXEC_BACKEND
4176 pid = backend_forkexec(port);
4177 #else /* !EXEC_BACKEND */
4178 pid = fork_process();
4179 if (pid == 0) /* child */
4180 {
4181 free(bn);
4182
4183 /* Detangle from postmaster */
4184 InitPostmasterChild();
4185
4186 /* Close the postmaster's sockets */
4187 ClosePostmasterPorts(false);
4188
4189 /* Perform additional initialization and collect startup packet */
4190 BackendInitialize(port);
4191
4192 /* And run the backend */
4193 BackendRun(port);
4194 }
4195 #endif /* EXEC_BACKEND */
4196
4197 if (pid < 0)
4198 {
4199 /* in parent, fork failed */
4200 int save_errno = errno;
4201
4202 if (!bn->dead_end)
4203 (void) ReleasePostmasterChildSlot(bn->child_slot);
4204 free(bn);
4205 errno = save_errno;
4206 ereport(LOG,
4207 (errmsg("could not fork new process for connection: %m")));
4208 report_fork_failure_to_client(port, save_errno);
4209 return STATUS_ERROR;
4210 }
4211
4212 /* in parent, successful fork */
4213 ereport(DEBUG2,
4214 (errmsg_internal("forked new backend, pid=%d socket=%d",
4215 (int) pid, (int) port->sock)));
4216
4217 /*
4218 * Everything's been successful, it's safe to add this backend to our list
4219 * of backends.
4220 */
4221 bn->pid = pid;
4222 bn->bkend_type = BACKEND_TYPE_NORMAL; /* Can change later to WALSND */
4223 dlist_push_head(&BackendList, &bn->elem);
4224
4225 #ifdef EXEC_BACKEND
4226 if (!bn->dead_end)
4227 ShmemBackendArrayAdd(bn);
4228 #endif
4229
4230 return STATUS_OK;
4231 }
4232
4233 /*
4234 * Try to report backend fork() failure to client before we close the
4235 * connection. Since we do not care to risk blocking the postmaster on
4236 * this connection, we set the connection to non-blocking and try only once.
4237 *
4238 * This is grungy special-purpose code; we cannot use backend libpq since
4239 * it's not up and running.
4240 */
4241 static void
report_fork_failure_to_client(Port * port,int errnum)4242 report_fork_failure_to_client(Port *port, int errnum)
4243 {
4244 char buffer[1000];
4245 int rc;
4246
4247 /* Format the error message packet (always V2 protocol) */
4248 snprintf(buffer, sizeof(buffer), "E%s%s\n",
4249 _("could not fork new process for connection: "),
4250 strerror(errnum));
4251
4252 /* Set port to non-blocking. Don't do send() if this fails */
4253 if (!pg_set_noblock(port->sock))
4254 return;
4255
4256 /* We'll retry after EINTR, but ignore all other failures */
4257 do
4258 {
4259 rc = send(port->sock, buffer, strlen(buffer) + 1, 0);
4260 } while (rc < 0 && errno == EINTR);
4261 }
4262
4263
4264 /*
4265 * BackendInitialize -- initialize an interactive (postmaster-child)
4266 * backend process, and collect the client's startup packet.
4267 *
4268 * returns: nothing. Will not return at all if there's any failure.
4269 *
4270 * Note: this code does not depend on having any access to shared memory.
4271 * In the EXEC_BACKEND case, we are physically attached to shared memory
4272 * but have not yet set up most of our local pointers to shmem structures.
4273 */
4274 static void
BackendInitialize(Port * port)4275 BackendInitialize(Port *port)
4276 {
4277 int status;
4278 int ret;
4279 char remote_host[NI_MAXHOST];
4280 char remote_port[NI_MAXSERV];
4281 char remote_ps_data[NI_MAXHOST];
4282
4283 /* Save port etc. for ps status */
4284 MyProcPort = port;
4285
4286 /*
4287 * PreAuthDelay is a debugging aid for investigating problems in the
4288 * authentication cycle: it can be set in postgresql.conf to allow time to
4289 * attach to the newly-forked backend with a debugger. (See also
4290 * PostAuthDelay, which we allow clients to pass through PGOPTIONS, but it
4291 * is not honored until after authentication.)
4292 */
4293 if (PreAuthDelay > 0)
4294 pg_usleep(PreAuthDelay * 1000000L);
4295
4296 /* This flag will remain set until InitPostgres finishes authentication */
4297 ClientAuthInProgress = true; /* limit visibility of log messages */
4298
4299 /* set these to empty in case they are needed before we set them up */
4300 port->remote_host = "";
4301 port->remote_port = "";
4302
4303 /*
4304 * Initialize libpq and enable reporting of ereport errors to the client.
4305 * Must do this now because authentication uses libpq to send messages.
4306 */
4307 pq_init(); /* initialize libpq to talk to client */
4308 whereToSendOutput = DestRemote; /* now safe to ereport to client */
4309
4310 /*
4311 * We arrange to do proc_exit(1) if we receive SIGTERM or timeout while
4312 * trying to collect the startup packet; while SIGQUIT results in
4313 * _exit(2). Otherwise the postmaster cannot shutdown the database FAST
4314 * or IMMED cleanly if a buggy client fails to send the packet promptly.
4315 *
4316 * XXX this is pretty dangerous; signal handlers should not call anything
4317 * as complex as proc_exit() directly. We minimize the hazard by not
4318 * keeping these handlers active for longer than we must. However, it
4319 * seems necessary to be able to escape out of DNS lookups as well as the
4320 * startup packet reception proper, so we can't narrow the scope further
4321 * than is done here.
4322 *
4323 * XXX it follows that the remainder of this function must tolerate losing
4324 * control at any instant. Likewise, any pg_on_exit_callback registered
4325 * before or during this function must be prepared to execute at any
4326 * instant between here and the end of this function. Furthermore,
4327 * affected callbacks execute partially or not at all when a second
4328 * exit-inducing signal arrives after proc_exit_prepare() decrements
4329 * on_proc_exit_index. (Thanks to that mechanic, callbacks need not
4330 * anticipate more than one call.) This is fragile; it ought to instead
4331 * follow the norm of handling interrupts at selected, safe opportunities.
4332 */
4333 pqsignal(SIGTERM, process_startup_packet_die);
4334 pqsignal(SIGQUIT, process_startup_packet_quickdie);
4335 InitializeTimeouts(); /* establishes SIGALRM handler */
4336 PG_SETMASK(&StartupBlockSig);
4337
4338 /*
4339 * Get the remote host name and port for logging and status display.
4340 */
4341 remote_host[0] = '\0';
4342 remote_port[0] = '\0';
4343 if ((ret = pg_getnameinfo_all(&port->raddr.addr, port->raddr.salen,
4344 remote_host, sizeof(remote_host),
4345 remote_port, sizeof(remote_port),
4346 (log_hostname ? 0 : NI_NUMERICHOST) | NI_NUMERICSERV)) != 0)
4347 ereport(WARNING,
4348 (errmsg_internal("pg_getnameinfo_all() failed: %s",
4349 gai_strerror(ret))));
4350 if (remote_port[0] == '\0')
4351 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s", remote_host);
4352 else
4353 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s(%s)", remote_host, remote_port);
4354
4355 /*
4356 * Save remote_host and remote_port in port structure (after this, they
4357 * will appear in log_line_prefix data for log messages).
4358 */
4359 port->remote_host = strdup(remote_host);
4360 port->remote_port = strdup(remote_port);
4361
4362 /* And now we can issue the Log_connections message, if wanted */
4363 if (Log_connections)
4364 {
4365 if (remote_port[0])
4366 ereport(LOG,
4367 (errmsg("connection received: host=%s port=%s",
4368 remote_host,
4369 remote_port)));
4370 else
4371 ereport(LOG,
4372 (errmsg("connection received: host=%s",
4373 remote_host)));
4374 }
4375
4376 /*
4377 * If we did a reverse lookup to name, we might as well save the results
4378 * rather than possibly repeating the lookup during authentication.
4379 *
4380 * Note that we don't want to specify NI_NAMEREQD above, because then we'd
4381 * get nothing useful for a client without an rDNS entry. Therefore, we
4382 * must check whether we got a numeric IPv4 or IPv6 address, and not save
4383 * it into remote_hostname if so. (This test is conservative and might
4384 * sometimes classify a hostname as numeric, but an error in that
4385 * direction is safe; it only results in a possible extra lookup.)
4386 */
4387 if (log_hostname &&
4388 ret == 0 &&
4389 strspn(remote_host, "0123456789.") < strlen(remote_host) &&
4390 strspn(remote_host, "0123456789ABCDEFabcdef:") < strlen(remote_host))
4391 port->remote_hostname = strdup(remote_host);
4392
4393 /*
4394 * Ready to begin client interaction. We will give up and proc_exit(1)
4395 * after a time delay, so that a broken client can't hog a connection
4396 * indefinitely. PreAuthDelay and any DNS interactions above don't count
4397 * against the time limit.
4398 *
4399 * Note: AuthenticationTimeout is applied here while waiting for the
4400 * startup packet, and then again in InitPostgres for the duration of any
4401 * authentication operations. So a hostile client could tie up the
4402 * process for nearly twice AuthenticationTimeout before we kick him off.
4403 *
4404 * Note: because PostgresMain will call InitializeTimeouts again, the
4405 * registration of STARTUP_PACKET_TIMEOUT will be lost. This is okay
4406 * since we never use it again after this function.
4407 */
4408 RegisterTimeout(STARTUP_PACKET_TIMEOUT, StartupPacketTimeoutHandler);
4409 enable_timeout_after(STARTUP_PACKET_TIMEOUT, AuthenticationTimeout * 1000);
4410
4411 /*
4412 * Receive the startup packet (which might turn out to be a cancel request
4413 * packet).
4414 */
4415 status = ProcessStartupPacket(port, false, false);
4416
4417 /*
4418 * Disable the timeout, and prevent SIGTERM/SIGQUIT again.
4419 */
4420 disable_timeout(STARTUP_PACKET_TIMEOUT, false);
4421 PG_SETMASK(&BlockSig);
4422
4423 /*
4424 * Stop here if it was bad or a cancel packet. ProcessStartupPacket
4425 * already did any appropriate error reporting.
4426 */
4427 if (status != STATUS_OK)
4428 proc_exit(0);
4429
4430 /*
4431 * Now that we have the user and database name, we can set the process
4432 * title for ps. It's good to do this as early as possible in startup.
4433 *
4434 * For a walsender, the ps display is set in the following form:
4435 *
4436 * postgres: walsender <user> <host> <activity>
4437 *
4438 * To achieve that, we pass "walsender" as username and username as dbname
4439 * to init_ps_display(). XXX: should add a new variant of
4440 * init_ps_display() to avoid abusing the parameters like this.
4441 */
4442 if (am_walsender)
4443 init_ps_display(pgstat_get_backend_desc(B_WAL_SENDER), port->user_name, remote_ps_data,
4444 update_process_title ? "authentication" : "");
4445 else
4446 init_ps_display(port->user_name, port->database_name, remote_ps_data,
4447 update_process_title ? "authentication" : "");
4448 }
4449
4450
4451 /*
4452 * BackendRun -- set up the backend's argument list and invoke PostgresMain()
4453 *
4454 * returns:
4455 * Shouldn't return at all.
4456 * If PostgresMain() fails, return status.
4457 */
4458 static void
BackendRun(Port * port)4459 BackendRun(Port *port)
4460 {
4461 char **av;
4462 int maxac;
4463 int ac;
4464 int i;
4465
4466 /*
4467 * Now, build the argv vector that will be given to PostgresMain.
4468 *
4469 * The maximum possible number of commandline arguments that could come
4470 * from ExtraOptions is (strlen(ExtraOptions) + 1) / 2; see
4471 * pg_split_opts().
4472 */
4473 maxac = 2; /* for fixed args supplied below */
4474 maxac += (strlen(ExtraOptions) + 1) / 2;
4475
4476 av = (char **) MemoryContextAlloc(TopMemoryContext,
4477 maxac * sizeof(char *));
4478 ac = 0;
4479
4480 av[ac++] = "postgres";
4481
4482 /*
4483 * Pass any backend switches specified with -o on the postmaster's own
4484 * command line. We assume these are secure.
4485 */
4486 pg_split_opts(av, &ac, ExtraOptions);
4487
4488 av[ac] = NULL;
4489
4490 Assert(ac < maxac);
4491
4492 /*
4493 * Debug: print arguments being passed to backend
4494 */
4495 ereport(DEBUG3,
4496 (errmsg_internal("%s child[%d]: starting with (",
4497 progname, (int) getpid())));
4498 for (i = 0; i < ac; ++i)
4499 ereport(DEBUG3,
4500 (errmsg_internal("\t%s", av[i])));
4501 ereport(DEBUG3,
4502 (errmsg_internal(")")));
4503
4504 /*
4505 * Make sure we aren't in PostmasterContext anymore. (We can't delete it
4506 * just yet, though, because InitPostgres will need the HBA data.)
4507 */
4508 MemoryContextSwitchTo(TopMemoryContext);
4509
4510 PostgresMain(ac, av, port->database_name, port->user_name);
4511 }
4512
4513
4514 #ifdef EXEC_BACKEND
4515
4516 /*
4517 * postmaster_forkexec -- fork and exec a postmaster subprocess
4518 *
4519 * The caller must have set up the argv array already, except for argv[2]
4520 * which will be filled with the name of the temp variable file.
4521 *
4522 * Returns the child process PID, or -1 on fork failure (a suitable error
4523 * message has been logged on failure).
4524 *
4525 * All uses of this routine will dispatch to SubPostmasterMain in the
4526 * child process.
4527 */
4528 pid_t
postmaster_forkexec(int argc,char * argv[])4529 postmaster_forkexec(int argc, char *argv[])
4530 {
4531 Port port;
4532
4533 /* This entry point passes dummy values for the Port variables */
4534 memset(&port, 0, sizeof(port));
4535 return internal_forkexec(argc, argv, &port);
4536 }
4537
4538 /*
4539 * backend_forkexec -- fork/exec off a backend process
4540 *
4541 * Some operating systems (WIN32) don't have fork() so we have to simulate
4542 * it by storing parameters that need to be passed to the child and
4543 * then create a new child process.
4544 *
4545 * returns the pid of the fork/exec'd process, or -1 on failure
4546 */
4547 static pid_t
backend_forkexec(Port * port)4548 backend_forkexec(Port *port)
4549 {
4550 char *av[4];
4551 int ac = 0;
4552
4553 av[ac++] = "postgres";
4554 av[ac++] = "--forkbackend";
4555 av[ac++] = NULL; /* filled in by internal_forkexec */
4556
4557 av[ac] = NULL;
4558 Assert(ac < lengthof(av));
4559
4560 return internal_forkexec(ac, av, port);
4561 }
4562
4563 #ifndef WIN32
4564
4565 /*
4566 * internal_forkexec non-win32 implementation
4567 *
4568 * - writes out backend variables to the parameter file
4569 * - fork():s, and then exec():s the child process
4570 */
4571 static pid_t
internal_forkexec(int argc,char * argv[],Port * port)4572 internal_forkexec(int argc, char *argv[], Port *port)
4573 {
4574 static unsigned long tmpBackendFileNum = 0;
4575 pid_t pid;
4576 char tmpfilename[MAXPGPATH];
4577 BackendParameters param;
4578 FILE *fp;
4579
4580 if (!save_backend_variables(¶m, port))
4581 return -1; /* log made by save_backend_variables */
4582
4583 /* Calculate name for temp file */
4584 snprintf(tmpfilename, MAXPGPATH, "%s/%s.backend_var.%d.%lu",
4585 PG_TEMP_FILES_DIR, PG_TEMP_FILE_PREFIX,
4586 MyProcPid, ++tmpBackendFileNum);
4587
4588 /* Open file */
4589 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4590 if (!fp)
4591 {
4592 /*
4593 * As in OpenTemporaryFileInTablespace, try to make the temp-file
4594 * directory, ignoring errors.
4595 */
4596 (void) MakePGDirectory(PG_TEMP_FILES_DIR);
4597
4598 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4599 if (!fp)
4600 {
4601 ereport(LOG,
4602 (errcode_for_file_access(),
4603 errmsg("could not create file \"%s\": %m",
4604 tmpfilename)));
4605 return -1;
4606 }
4607 }
4608
4609 if (fwrite(¶m, sizeof(param), 1, fp) != 1)
4610 {
4611 ereport(LOG,
4612 (errcode_for_file_access(),
4613 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4614 FreeFile(fp);
4615 return -1;
4616 }
4617
4618 /* Release file */
4619 if (FreeFile(fp))
4620 {
4621 ereport(LOG,
4622 (errcode_for_file_access(),
4623 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4624 return -1;
4625 }
4626
4627 /* Make sure caller set up argv properly */
4628 Assert(argc >= 3);
4629 Assert(argv[argc] == NULL);
4630 Assert(strncmp(argv[1], "--fork", 6) == 0);
4631 Assert(argv[2] == NULL);
4632
4633 /* Insert temp file name after --fork argument */
4634 argv[2] = tmpfilename;
4635
4636 /* Fire off execv in child */
4637 if ((pid = fork_process()) == 0)
4638 {
4639 if (execv(postgres_exec_path, argv) < 0)
4640 {
4641 ereport(LOG,
4642 (errmsg("could not execute server process \"%s\": %m",
4643 postgres_exec_path)));
4644 /* We're already in the child process here, can't return */
4645 exit(1);
4646 }
4647 }
4648
4649 return pid; /* Parent returns pid, or -1 on fork failure */
4650 }
4651 #else /* WIN32 */
4652
4653 /*
4654 * internal_forkexec win32 implementation
4655 *
4656 * - starts backend using CreateProcess(), in suspended state
4657 * - writes out backend variables to the parameter file
4658 * - during this, duplicates handles and sockets required for
4659 * inheritance into the new process
4660 * - resumes execution of the new process once the backend parameter
4661 * file is complete.
4662 */
4663 static pid_t
internal_forkexec(int argc,char * argv[],Port * port)4664 internal_forkexec(int argc, char *argv[], Port *port)
4665 {
4666 int retry_count = 0;
4667 STARTUPINFO si;
4668 PROCESS_INFORMATION pi;
4669 int i;
4670 int j;
4671 char cmdLine[MAXPGPATH * 2];
4672 HANDLE paramHandle;
4673 BackendParameters *param;
4674 SECURITY_ATTRIBUTES sa;
4675 char paramHandleStr[32];
4676 win32_deadchild_waitinfo *childinfo;
4677
4678 /* Make sure caller set up argv properly */
4679 Assert(argc >= 3);
4680 Assert(argv[argc] == NULL);
4681 Assert(strncmp(argv[1], "--fork", 6) == 0);
4682 Assert(argv[2] == NULL);
4683
4684 /* Resume here if we need to retry */
4685 retry:
4686
4687 /* Set up shared memory for parameter passing */
4688 ZeroMemory(&sa, sizeof(sa));
4689 sa.nLength = sizeof(sa);
4690 sa.bInheritHandle = TRUE;
4691 paramHandle = CreateFileMapping(INVALID_HANDLE_VALUE,
4692 &sa,
4693 PAGE_READWRITE,
4694 0,
4695 sizeof(BackendParameters),
4696 NULL);
4697 if (paramHandle == INVALID_HANDLE_VALUE)
4698 {
4699 elog(LOG, "could not create backend parameter file mapping: error code %lu",
4700 GetLastError());
4701 return -1;
4702 }
4703
4704 param = MapViewOfFile(paramHandle, FILE_MAP_WRITE, 0, 0, sizeof(BackendParameters));
4705 if (!param)
4706 {
4707 elog(LOG, "could not map backend parameter memory: error code %lu",
4708 GetLastError());
4709 CloseHandle(paramHandle);
4710 return -1;
4711 }
4712
4713 /* Insert temp file name after --fork argument */
4714 #ifdef _WIN64
4715 sprintf(paramHandleStr, "%llu", (LONG_PTR) paramHandle);
4716 #else
4717 sprintf(paramHandleStr, "%lu", (DWORD) paramHandle);
4718 #endif
4719 argv[2] = paramHandleStr;
4720
4721 /* Format the cmd line */
4722 cmdLine[sizeof(cmdLine) - 1] = '\0';
4723 cmdLine[sizeof(cmdLine) - 2] = '\0';
4724 snprintf(cmdLine, sizeof(cmdLine) - 1, "\"%s\"", postgres_exec_path);
4725 i = 0;
4726 while (argv[++i] != NULL)
4727 {
4728 j = strlen(cmdLine);
4729 snprintf(cmdLine + j, sizeof(cmdLine) - 1 - j, " \"%s\"", argv[i]);
4730 }
4731 if (cmdLine[sizeof(cmdLine) - 2] != '\0')
4732 {
4733 elog(LOG, "subprocess command line too long");
4734 return -1;
4735 }
4736
4737 memset(&pi, 0, sizeof(pi));
4738 memset(&si, 0, sizeof(si));
4739 si.cb = sizeof(si);
4740
4741 /*
4742 * Create the subprocess in a suspended state. This will be resumed later,
4743 * once we have written out the parameter file.
4744 */
4745 if (!CreateProcess(NULL, cmdLine, NULL, NULL, TRUE, CREATE_SUSPENDED,
4746 NULL, NULL, &si, &pi))
4747 {
4748 elog(LOG, "CreateProcess call failed: %m (error code %lu)",
4749 GetLastError());
4750 return -1;
4751 }
4752
4753 if (!save_backend_variables(param, port, pi.hProcess, pi.dwProcessId))
4754 {
4755 /*
4756 * log made by save_backend_variables, but we have to clean up the
4757 * mess with the half-started process
4758 */
4759 if (!TerminateProcess(pi.hProcess, 255))
4760 ereport(LOG,
4761 (errmsg_internal("could not terminate unstarted process: error code %lu",
4762 GetLastError())));
4763 CloseHandle(pi.hProcess);
4764 CloseHandle(pi.hThread);
4765 return -1; /* log made by save_backend_variables */
4766 }
4767
4768 /* Drop the parameter shared memory that is now inherited to the backend */
4769 if (!UnmapViewOfFile(param))
4770 elog(LOG, "could not unmap view of backend parameter file: error code %lu",
4771 GetLastError());
4772 if (!CloseHandle(paramHandle))
4773 elog(LOG, "could not close handle to backend parameter file: error code %lu",
4774 GetLastError());
4775
4776 /*
4777 * Reserve the memory region used by our main shared memory segment before
4778 * we resume the child process. Normally this should succeed, but if ASLR
4779 * is active then it might sometimes fail due to the stack or heap having
4780 * gotten mapped into that range. In that case, just terminate the
4781 * process and retry.
4782 */
4783 if (!pgwin32_ReserveSharedMemoryRegion(pi.hProcess))
4784 {
4785 /* pgwin32_ReserveSharedMemoryRegion already made a log entry */
4786 if (!TerminateProcess(pi.hProcess, 255))
4787 ereport(LOG,
4788 (errmsg_internal("could not terminate process that failed to reserve memory: error code %lu",
4789 GetLastError())));
4790 CloseHandle(pi.hProcess);
4791 CloseHandle(pi.hThread);
4792 if (++retry_count < 100)
4793 goto retry;
4794 ereport(LOG,
4795 (errmsg("giving up after too many tries to reserve shared memory"),
4796 errhint("This might be caused by ASLR or antivirus software.")));
4797 return -1;
4798 }
4799
4800 /*
4801 * Now that the backend variables are written out, we start the child
4802 * thread so it can start initializing while we set up the rest of the
4803 * parent state.
4804 */
4805 if (ResumeThread(pi.hThread) == -1)
4806 {
4807 if (!TerminateProcess(pi.hProcess, 255))
4808 {
4809 ereport(LOG,
4810 (errmsg_internal("could not terminate unstartable process: error code %lu",
4811 GetLastError())));
4812 CloseHandle(pi.hProcess);
4813 CloseHandle(pi.hThread);
4814 return -1;
4815 }
4816 CloseHandle(pi.hProcess);
4817 CloseHandle(pi.hThread);
4818 ereport(LOG,
4819 (errmsg_internal("could not resume thread of unstarted process: error code %lu",
4820 GetLastError())));
4821 return -1;
4822 }
4823
4824 /*
4825 * Queue a waiter to signal when this child dies. The wait will be handled
4826 * automatically by an operating system thread pool.
4827 *
4828 * Note: use malloc instead of palloc, since it needs to be thread-safe.
4829 * Struct will be free():d from the callback function that runs on a
4830 * different thread.
4831 */
4832 childinfo = malloc(sizeof(win32_deadchild_waitinfo));
4833 if (!childinfo)
4834 ereport(FATAL,
4835 (errcode(ERRCODE_OUT_OF_MEMORY),
4836 errmsg("out of memory")));
4837
4838 childinfo->procHandle = pi.hProcess;
4839 childinfo->procId = pi.dwProcessId;
4840
4841 if (!RegisterWaitForSingleObject(&childinfo->waitHandle,
4842 pi.hProcess,
4843 pgwin32_deadchild_callback,
4844 childinfo,
4845 INFINITE,
4846 WT_EXECUTEONLYONCE | WT_EXECUTEINWAITTHREAD))
4847 ereport(FATAL,
4848 (errmsg_internal("could not register process for wait: error code %lu",
4849 GetLastError())));
4850
4851 /* Don't close pi.hProcess here - the wait thread needs access to it */
4852
4853 CloseHandle(pi.hThread);
4854
4855 return pi.dwProcessId;
4856 }
4857 #endif /* WIN32 */
4858
4859
4860 /*
4861 * SubPostmasterMain -- Get the fork/exec'd process into a state equivalent
4862 * to what it would be if we'd simply forked on Unix, and then
4863 * dispatch to the appropriate place.
4864 *
4865 * The first two command line arguments are expected to be "--forkFOO"
4866 * (where FOO indicates which postmaster child we are to become), and
4867 * the name of a variables file that we can read to load data that would
4868 * have been inherited by fork() on Unix. Remaining arguments go to the
4869 * subprocess FooMain() routine.
4870 */
4871 void
SubPostmasterMain(int argc,char * argv[])4872 SubPostmasterMain(int argc, char *argv[])
4873 {
4874 Port port;
4875
4876 /* In EXEC_BACKEND case we will not have inherited these settings */
4877 IsPostmasterEnvironment = true;
4878 whereToSendOutput = DestNone;
4879
4880 /* Setup as postmaster child */
4881 InitPostmasterChild();
4882
4883 /* Setup essential subsystems (to ensure elog() behaves sanely) */
4884 InitializeGUCOptions();
4885
4886 /* Check we got appropriate args */
4887 if (argc < 3)
4888 elog(FATAL, "invalid subpostmaster invocation");
4889
4890 /* Read in the variables file */
4891 memset(&port, 0, sizeof(Port));
4892 read_backend_variables(argv[2], &port);
4893
4894 /* Close the postmaster's sockets (as soon as we know them) */
4895 ClosePostmasterPorts(strcmp(argv[1], "--forklog") == 0);
4896
4897 /*
4898 * Set reference point for stack-depth checking
4899 */
4900 set_stack_base();
4901
4902 /*
4903 * If appropriate, physically re-attach to shared memory segment. We want
4904 * to do this before going any further to ensure that we can attach at the
4905 * same address the postmaster used. On the other hand, if we choose not
4906 * to re-attach, we may have other cleanup to do.
4907 *
4908 * If testing EXEC_BACKEND on Linux, you should run this as root before
4909 * starting the postmaster:
4910 *
4911 * echo 0 >/proc/sys/kernel/randomize_va_space
4912 *
4913 * This prevents using randomized stack and code addresses that cause the
4914 * child process's memory map to be different from the parent's, making it
4915 * sometimes impossible to attach to shared memory at the desired address.
4916 * Return the setting to its old value (usually '1' or '2') when finished.
4917 */
4918 if (strcmp(argv[1], "--forkbackend") == 0 ||
4919 strcmp(argv[1], "--forkavlauncher") == 0 ||
4920 strcmp(argv[1], "--forkavworker") == 0 ||
4921 strcmp(argv[1], "--forkboot") == 0 ||
4922 strncmp(argv[1], "--forkbgworker=", 15) == 0)
4923 PGSharedMemoryReAttach();
4924 else
4925 PGSharedMemoryNoReAttach();
4926
4927 /* autovacuum needs this set before calling InitProcess */
4928 if (strcmp(argv[1], "--forkavlauncher") == 0)
4929 AutovacuumLauncherIAm();
4930 if (strcmp(argv[1], "--forkavworker") == 0)
4931 AutovacuumWorkerIAm();
4932
4933 /*
4934 * Start our win32 signal implementation. This has to be done after we
4935 * read the backend variables, because we need to pick up the signal pipe
4936 * from the parent process.
4937 */
4938 #ifdef WIN32
4939 pgwin32_signal_initialize();
4940 #endif
4941
4942 /* In EXEC_BACKEND case we will not have inherited these settings */
4943 pqinitmask();
4944 PG_SETMASK(&BlockSig);
4945
4946 /* Read in remaining GUC variables */
4947 read_nondefault_variables();
4948
4949 /*
4950 * Check that the data directory looks valid, which will also check the
4951 * privileges on the data directory and update our umask and file/group
4952 * variables for creating files later. Note: this should really be done
4953 * before we create any files or directories.
4954 */
4955 checkDataDir();
4956
4957 /*
4958 * (re-)read control file, as it contains config. The postmaster will
4959 * already have read this, but this process doesn't know about that.
4960 */
4961 LocalProcessControlFile(false);
4962
4963 /*
4964 * Reload any libraries that were preloaded by the postmaster. Since we
4965 * exec'd this process, those libraries didn't come along with us; but we
4966 * should load them into all child processes to be consistent with the
4967 * non-EXEC_BACKEND behavior.
4968 */
4969 process_shared_preload_libraries();
4970
4971 /* Run backend or appropriate child */
4972 if (strcmp(argv[1], "--forkbackend") == 0)
4973 {
4974 Assert(argc == 3); /* shouldn't be any more args */
4975
4976 /*
4977 * Need to reinitialize the SSL library in the backend, since the
4978 * context structures contain function pointers and cannot be passed
4979 * through the parameter file.
4980 *
4981 * If for some reason reload fails (maybe the user installed broken
4982 * key files), soldier on without SSL; that's better than all
4983 * connections becoming impossible.
4984 *
4985 * XXX should we do this in all child processes? For the moment it's
4986 * enough to do it in backend children.
4987 */
4988 #ifdef USE_SSL
4989 if (EnableSSL)
4990 {
4991 if (secure_initialize(false) == 0)
4992 LoadedSSL = true;
4993 else
4994 ereport(LOG,
4995 (errmsg("SSL configuration could not be loaded in child process")));
4996 }
4997 #endif
4998
4999 /*
5000 * Perform additional initialization and collect startup packet.
5001 *
5002 * We want to do this before InitProcess() for a couple of reasons: 1.
5003 * so that we aren't eating up a PGPROC slot while waiting on the
5004 * client. 2. so that if InitProcess() fails due to being out of
5005 * PGPROC slots, we have already initialized libpq and are able to
5006 * report the error to the client.
5007 */
5008 BackendInitialize(&port);
5009
5010 /* Restore basic shared memory pointers */
5011 InitShmemAccess(UsedShmemSegAddr);
5012
5013 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5014 InitProcess();
5015
5016 /* Attach process to shared data structures */
5017 CreateSharedMemoryAndSemaphores(0);
5018
5019 /* And run the backend */
5020 BackendRun(&port); /* does not return */
5021 }
5022 if (strcmp(argv[1], "--forkboot") == 0)
5023 {
5024 /* Restore basic shared memory pointers */
5025 InitShmemAccess(UsedShmemSegAddr);
5026
5027 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5028 InitAuxiliaryProcess();
5029
5030 /* Attach process to shared data structures */
5031 CreateSharedMemoryAndSemaphores(0);
5032
5033 AuxiliaryProcessMain(argc - 2, argv + 2); /* does not return */
5034 }
5035 if (strcmp(argv[1], "--forkavlauncher") == 0)
5036 {
5037 /* Restore basic shared memory pointers */
5038 InitShmemAccess(UsedShmemSegAddr);
5039
5040 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5041 InitProcess();
5042
5043 /* Attach process to shared data structures */
5044 CreateSharedMemoryAndSemaphores(0);
5045
5046 AutoVacLauncherMain(argc - 2, argv + 2); /* does not return */
5047 }
5048 if (strcmp(argv[1], "--forkavworker") == 0)
5049 {
5050 /* Restore basic shared memory pointers */
5051 InitShmemAccess(UsedShmemSegAddr);
5052
5053 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5054 InitProcess();
5055
5056 /* Attach process to shared data structures */
5057 CreateSharedMemoryAndSemaphores(0);
5058
5059 AutoVacWorkerMain(argc - 2, argv + 2); /* does not return */
5060 }
5061 if (strncmp(argv[1], "--forkbgworker=", 15) == 0)
5062 {
5063 int shmem_slot;
5064
5065 /* do this as early as possible; in particular, before InitProcess() */
5066 IsBackgroundWorker = true;
5067
5068 /* Restore basic shared memory pointers */
5069 InitShmemAccess(UsedShmemSegAddr);
5070
5071 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5072 InitProcess();
5073
5074 /* Attach process to shared data structures */
5075 CreateSharedMemoryAndSemaphores(0);
5076
5077 /* Fetch MyBgworkerEntry from shared memory */
5078 shmem_slot = atoi(argv[1] + 15);
5079 MyBgworkerEntry = BackgroundWorkerEntry(shmem_slot);
5080
5081 StartBackgroundWorker();
5082 }
5083 if (strcmp(argv[1], "--forkarch") == 0)
5084 {
5085 /* Do not want to attach to shared memory */
5086
5087 PgArchiverMain(argc, argv); /* does not return */
5088 }
5089 if (strcmp(argv[1], "--forkcol") == 0)
5090 {
5091 /* Do not want to attach to shared memory */
5092
5093 PgstatCollectorMain(argc, argv); /* does not return */
5094 }
5095 if (strcmp(argv[1], "--forklog") == 0)
5096 {
5097 /* Do not want to attach to shared memory */
5098
5099 SysLoggerMain(argc, argv); /* does not return */
5100 }
5101
5102 abort(); /* shouldn't get here */
5103 }
5104 #endif /* EXEC_BACKEND */
5105
5106
5107 /*
5108 * ExitPostmaster -- cleanup
5109 *
5110 * Do NOT call exit() directly --- always go through here!
5111 */
5112 static void
ExitPostmaster(int status)5113 ExitPostmaster(int status)
5114 {
5115 #ifdef HAVE_PTHREAD_IS_THREADED_NP
5116
5117 /*
5118 * There is no known cause for a postmaster to become multithreaded after
5119 * startup. Recheck to account for the possibility of unknown causes.
5120 * This message uses LOG level, because an unclean shutdown at this point
5121 * would usually not look much different from a clean shutdown.
5122 */
5123 if (pthread_is_threaded_np() != 0)
5124 ereport(LOG,
5125 (errcode(ERRCODE_INTERNAL_ERROR),
5126 errmsg_internal("postmaster became multithreaded"),
5127 errdetail("Please report this to <pgsql-bugs@lists.postgresql.org>.")));
5128 #endif
5129
5130 /* should cleanup shared memory and kill all backends */
5131
5132 /*
5133 * Not sure of the semantics here. When the Postmaster dies, should the
5134 * backends all be killed? probably not.
5135 *
5136 * MUST -- vadim 05-10-1999
5137 */
5138
5139 proc_exit(status);
5140 }
5141
5142 /*
5143 * sigusr1_handler - handle signal conditions from child processes
5144 */
5145 static void
sigusr1_handler(SIGNAL_ARGS)5146 sigusr1_handler(SIGNAL_ARGS)
5147 {
5148 int save_errno = errno;
5149
5150 /*
5151 * We rely on the signal mechanism to have blocked all signals ... except
5152 * on Windows, which lacks sigaction(), so we have to do it manually.
5153 */
5154 #ifdef WIN32
5155 PG_SETMASK(&BlockSig);
5156 #endif
5157
5158 /*
5159 * RECOVERY_STARTED and BEGIN_HOT_STANDBY signals are ignored in
5160 * unexpected states. If the startup process quickly starts up, completes
5161 * recovery, exits, we might process the death of the startup process
5162 * first. We don't want to go back to recovery in that case.
5163 */
5164 if (CheckPostmasterSignal(PMSIGNAL_RECOVERY_STARTED) &&
5165 pmState == PM_STARTUP && Shutdown == NoShutdown)
5166 {
5167 /* WAL redo has started. We're out of reinitialization. */
5168 FatalError = false;
5169 Assert(AbortStartTime == 0);
5170
5171 /*
5172 * Crank up the background tasks. It doesn't matter if this fails,
5173 * we'll just try again later.
5174 */
5175 Assert(CheckpointerPID == 0);
5176 CheckpointerPID = StartCheckpointer();
5177 Assert(BgWriterPID == 0);
5178 BgWriterPID = StartBackgroundWriter();
5179
5180 /*
5181 * Start the archiver if we're responsible for (re-)archiving received
5182 * files.
5183 */
5184 Assert(PgArchPID == 0);
5185 if (XLogArchivingAlways())
5186 PgArchPID = pgarch_start();
5187
5188 /*
5189 * If we aren't planning to enter hot standby mode later, treat
5190 * RECOVERY_STARTED as meaning we're out of startup, and report status
5191 * accordingly.
5192 */
5193 if (!EnableHotStandby)
5194 {
5195 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STANDBY);
5196 #ifdef USE_SYSTEMD
5197 sd_notify(0, "READY=1");
5198 #endif
5199 }
5200
5201 pmState = PM_RECOVERY;
5202 }
5203
5204 if (CheckPostmasterSignal(PMSIGNAL_BEGIN_HOT_STANDBY) &&
5205 pmState == PM_RECOVERY && Shutdown == NoShutdown)
5206 {
5207 /*
5208 * Likewise, start other special children as needed.
5209 */
5210 Assert(PgStatPID == 0);
5211 PgStatPID = pgstat_start();
5212
5213 ereport(LOG,
5214 (errmsg("database system is ready to accept read only connections")));
5215
5216 /* Report status */
5217 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
5218 #ifdef USE_SYSTEMD
5219 sd_notify(0, "READY=1");
5220 #endif
5221
5222 pmState = PM_HOT_STANDBY;
5223 connsAllowed = ALLOW_ALL_CONNS;
5224
5225 /* Some workers may be scheduled to start now */
5226 StartWorkerNeeded = true;
5227 }
5228
5229 /* Process background worker state changes. */
5230 if (CheckPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE))
5231 {
5232 /* Accept new worker requests only if not stopping. */
5233 BackgroundWorkerStateChange(pmState < PM_STOP_BACKENDS);
5234 StartWorkerNeeded = true;
5235 }
5236
5237 if (StartWorkerNeeded || HaveCrashedWorker)
5238 maybe_start_bgworkers();
5239
5240 if (CheckPostmasterSignal(PMSIGNAL_WAKEN_ARCHIVER) &&
5241 PgArchPID != 0)
5242 {
5243 /*
5244 * Send SIGUSR1 to archiver process, to wake it up and begin archiving
5245 * next WAL file.
5246 */
5247 signal_child(PgArchPID, SIGUSR1);
5248 }
5249
5250 /* Tell syslogger to rotate logfile if requested */
5251 if (SysLoggerPID != 0)
5252 {
5253 if (CheckLogrotateSignal())
5254 {
5255 signal_child(SysLoggerPID, SIGUSR1);
5256 RemoveLogrotateSignalFiles();
5257 }
5258 else if (CheckPostmasterSignal(PMSIGNAL_ROTATE_LOGFILE))
5259 {
5260 signal_child(SysLoggerPID, SIGUSR1);
5261 }
5262 }
5263
5264 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER) &&
5265 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5266 {
5267 /*
5268 * Start one iteration of the autovacuum daemon, even if autovacuuming
5269 * is nominally not enabled. This is so we can have an active defense
5270 * against transaction ID wraparound. We set a flag for the main loop
5271 * to do it rather than trying to do it here --- this is because the
5272 * autovac process itself may send the signal, and we want to handle
5273 * that by launching another iteration as soon as the current one
5274 * completes.
5275 */
5276 start_autovac_launcher = true;
5277 }
5278
5279 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER) &&
5280 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5281 {
5282 /* The autovacuum launcher wants us to start a worker process. */
5283 StartAutovacuumWorker();
5284 }
5285
5286 if (CheckPostmasterSignal(PMSIGNAL_START_WALRECEIVER))
5287 {
5288 /* Startup Process wants us to start the walreceiver process. */
5289 /* Start immediately if possible, else remember request for later. */
5290 WalReceiverRequested = true;
5291 MaybeStartWalReceiver();
5292 }
5293
5294 /*
5295 * Try to advance postmaster's state machine, if a child requests it.
5296 *
5297 * Be careful about the order of this action relative to sigusr1_handler's
5298 * other actions. Generally, this should be after other actions, in case
5299 * they have effects PostmasterStateMachine would need to know about.
5300 * However, we should do it before the CheckPromoteSignal step, which
5301 * cannot have any (immediate) effect on the state machine, but does
5302 * depend on what state we're in now.
5303 */
5304 if (CheckPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE))
5305 {
5306 PostmasterStateMachine();
5307 }
5308
5309 if (StartupPID != 0 &&
5310 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5311 pmState == PM_HOT_STANDBY) &&
5312 CheckPromoteSignal())
5313 {
5314 /* Tell startup process to finish recovery */
5315 signal_child(StartupPID, SIGUSR2);
5316 }
5317
5318 #ifdef WIN32
5319 PG_SETMASK(&UnBlockSig);
5320 #endif
5321
5322 errno = save_errno;
5323 }
5324
5325 /*
5326 * SIGTERM while processing startup packet.
5327 * Clean up and exit(1).
5328 *
5329 * Running proc_exit() from a signal handler is pretty unsafe, since we
5330 * can't know what code we've interrupted. But the alternative of using
5331 * _exit(2) is also unpalatable, since it'd mean that a "fast shutdown"
5332 * would cause a database crash cycle (forcing WAL replay at restart)
5333 * if any sessions are in authentication. So we live with it for now.
5334 *
5335 * One might be tempted to try to send a message indicating why we are
5336 * disconnecting. However, that would make this even more unsafe. Also,
5337 * it seems undesirable to provide clues about the database's state to
5338 * a client that has not yet completed authentication.
5339 */
5340 static void
process_startup_packet_die(SIGNAL_ARGS)5341 process_startup_packet_die(SIGNAL_ARGS)
5342 {
5343 proc_exit(1);
5344 }
5345
5346 /*
5347 * SIGQUIT while processing startup packet.
5348 *
5349 * Some backend has bought the farm,
5350 * so we need to stop what we're doing and exit.
5351 */
5352 static void
process_startup_packet_quickdie(SIGNAL_ARGS)5353 process_startup_packet_quickdie(SIGNAL_ARGS)
5354 {
5355 /*
5356 * We DO NOT want to run proc_exit() or atexit() callbacks; they wouldn't
5357 * be safe to run from a signal handler. Just nail the windows shut and
5358 * get out of town.
5359 *
5360 * Note we do _exit(2) not _exit(1). This is to force the postmaster into
5361 * a system reset cycle if someone sends a manual SIGQUIT to a random
5362 * backend. (While it might be safe to do _exit(1), since this session
5363 * shouldn't have touched shared memory yet, there seems little point in
5364 * taking any risks.)
5365 */
5366 _exit(2);
5367 }
5368
5369 /*
5370 * Dummy signal handler
5371 *
5372 * We use this for signals that we don't actually use in the postmaster,
5373 * but we do use in backends. If we were to SIG_IGN such signals in the
5374 * postmaster, then a newly started backend might drop a signal that arrives
5375 * before it's able to reconfigure its signal processing. (See notes in
5376 * tcop/postgres.c.)
5377 */
5378 static void
dummy_handler(SIGNAL_ARGS)5379 dummy_handler(SIGNAL_ARGS)
5380 {
5381 }
5382
5383 /*
5384 * Timeout while processing startup packet.
5385 * As for process_startup_packet_die(), we clean up and exit(1).
5386 *
5387 * This is theoretically just as hazardous as in process_startup_packet_die(),
5388 * although in practice we're almost certainly waiting for client input,
5389 * which greatly reduces the risk.
5390 */
5391 static void
StartupPacketTimeoutHandler(void)5392 StartupPacketTimeoutHandler(void)
5393 {
5394 proc_exit(1);
5395 }
5396
5397
5398 /*
5399 * Generate a random cancel key.
5400 */
5401 static bool
RandomCancelKey(int32 * cancel_key)5402 RandomCancelKey(int32 *cancel_key)
5403 {
5404 return pg_strong_random(cancel_key, sizeof(int32));
5405 }
5406
5407 /*
5408 * Count up number of child processes of specified types (dead_end children
5409 * are always excluded).
5410 */
5411 static int
CountChildren(int target)5412 CountChildren(int target)
5413 {
5414 dlist_iter iter;
5415 int cnt = 0;
5416
5417 dlist_foreach(iter, &BackendList)
5418 {
5419 Backend *bp = dlist_container(Backend, elem, iter.cur);
5420
5421 if (bp->dead_end)
5422 continue;
5423
5424 /*
5425 * Since target == BACKEND_TYPE_ALL is the most common case, we test
5426 * it first and avoid touching shared memory for every child.
5427 */
5428 if (target != BACKEND_TYPE_ALL)
5429 {
5430 /*
5431 * Assign bkend_type for any recently announced WAL Sender
5432 * processes.
5433 */
5434 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
5435 IsPostmasterChildWalSender(bp->child_slot))
5436 bp->bkend_type = BACKEND_TYPE_WALSND;
5437
5438 if (!(target & bp->bkend_type))
5439 continue;
5440 }
5441
5442 cnt++;
5443 }
5444 return cnt;
5445 }
5446
5447
5448 /*
5449 * StartChildProcess -- start an auxiliary process for the postmaster
5450 *
5451 * "type" determines what kind of child will be started. All child types
5452 * initially go to AuxiliaryProcessMain, which will handle common setup.
5453 *
5454 * Return value of StartChildProcess is subprocess' PID, or 0 if failed
5455 * to start subprocess.
5456 */
5457 static pid_t
StartChildProcess(AuxProcType type)5458 StartChildProcess(AuxProcType type)
5459 {
5460 pid_t pid;
5461 char *av[10];
5462 int ac = 0;
5463 char typebuf[32];
5464
5465 /*
5466 * Set up command-line arguments for subprocess
5467 */
5468 av[ac++] = "postgres";
5469
5470 #ifdef EXEC_BACKEND
5471 av[ac++] = "--forkboot";
5472 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5473 #endif
5474
5475 snprintf(typebuf, sizeof(typebuf), "-x%d", type);
5476 av[ac++] = typebuf;
5477
5478 av[ac] = NULL;
5479 Assert(ac < lengthof(av));
5480
5481 #ifdef EXEC_BACKEND
5482 pid = postmaster_forkexec(ac, av);
5483 #else /* !EXEC_BACKEND */
5484 pid = fork_process();
5485
5486 if (pid == 0) /* child */
5487 {
5488 InitPostmasterChild();
5489
5490 /* Close the postmaster's sockets */
5491 ClosePostmasterPorts(false);
5492
5493 /* Release postmaster's working memory context */
5494 MemoryContextSwitchTo(TopMemoryContext);
5495 MemoryContextDelete(PostmasterContext);
5496 PostmasterContext = NULL;
5497
5498 AuxiliaryProcessMain(ac, av);
5499 ExitPostmaster(0);
5500 }
5501 #endif /* EXEC_BACKEND */
5502
5503 if (pid < 0)
5504 {
5505 /* in parent, fork failed */
5506 int save_errno = errno;
5507
5508 errno = save_errno;
5509 switch (type)
5510 {
5511 case StartupProcess:
5512 ereport(LOG,
5513 (errmsg("could not fork startup process: %m")));
5514 break;
5515 case BgWriterProcess:
5516 ereport(LOG,
5517 (errmsg("could not fork background writer process: %m")));
5518 break;
5519 case CheckpointerProcess:
5520 ereport(LOG,
5521 (errmsg("could not fork checkpointer process: %m")));
5522 break;
5523 case WalWriterProcess:
5524 ereport(LOG,
5525 (errmsg("could not fork WAL writer process: %m")));
5526 break;
5527 case WalReceiverProcess:
5528 ereport(LOG,
5529 (errmsg("could not fork WAL receiver process: %m")));
5530 break;
5531 default:
5532 ereport(LOG,
5533 (errmsg("could not fork process: %m")));
5534 break;
5535 }
5536
5537 /*
5538 * fork failure is fatal during startup, but there's no need to choke
5539 * immediately if starting other child types fails.
5540 */
5541 if (type == StartupProcess)
5542 ExitPostmaster(1);
5543 return 0;
5544 }
5545
5546 /*
5547 * in parent, successful fork
5548 */
5549 return pid;
5550 }
5551
5552 /*
5553 * StartAutovacuumWorker
5554 * Start an autovac worker process.
5555 *
5556 * This function is here because it enters the resulting PID into the
5557 * postmaster's private backends list.
5558 *
5559 * NB -- this code very roughly matches BackendStartup.
5560 */
5561 static void
StartAutovacuumWorker(void)5562 StartAutovacuumWorker(void)
5563 {
5564 Backend *bn;
5565
5566 /*
5567 * If not in condition to run a process, don't try, but handle it like a
5568 * fork failure. This does not normally happen, since the signal is only
5569 * supposed to be sent by autovacuum launcher when it's OK to do it, but
5570 * we have to check to avoid race-condition problems during DB state
5571 * changes.
5572 */
5573 if (canAcceptConnections(BACKEND_TYPE_AUTOVAC) == CAC_OK)
5574 {
5575 /*
5576 * Compute the cancel key that will be assigned to this session. We
5577 * probably don't need cancel keys for autovac workers, but we'd
5578 * better have something random in the field to prevent unfriendly
5579 * people from sending cancels to them.
5580 */
5581 if (!RandomCancelKey(&MyCancelKey))
5582 {
5583 ereport(LOG,
5584 (errcode(ERRCODE_INTERNAL_ERROR),
5585 errmsg("could not generate random cancel key")));
5586 return;
5587 }
5588
5589 bn = (Backend *) malloc(sizeof(Backend));
5590 if (bn)
5591 {
5592 bn->cancel_key = MyCancelKey;
5593
5594 /* Autovac workers are not dead_end and need a child slot */
5595 bn->dead_end = false;
5596 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5597 bn->bgworker_notify = false;
5598
5599 bn->pid = StartAutoVacWorker();
5600 if (bn->pid > 0)
5601 {
5602 bn->bkend_type = BACKEND_TYPE_AUTOVAC;
5603 dlist_push_head(&BackendList, &bn->elem);
5604 #ifdef EXEC_BACKEND
5605 ShmemBackendArrayAdd(bn);
5606 #endif
5607 /* all OK */
5608 return;
5609 }
5610
5611 /*
5612 * fork failed, fall through to report -- actual error message was
5613 * logged by StartAutoVacWorker
5614 */
5615 (void) ReleasePostmasterChildSlot(bn->child_slot);
5616 free(bn);
5617 }
5618 else
5619 ereport(LOG,
5620 (errcode(ERRCODE_OUT_OF_MEMORY),
5621 errmsg("out of memory")));
5622 }
5623
5624 /*
5625 * Report the failure to the launcher, if it's running. (If it's not, we
5626 * might not even be connected to shared memory, so don't try to call
5627 * AutoVacWorkerFailed.) Note that we also need to signal it so that it
5628 * responds to the condition, but we don't do that here, instead waiting
5629 * for ServerLoop to do it. This way we avoid a ping-pong signalling in
5630 * quick succession between the autovac launcher and postmaster in case
5631 * things get ugly.
5632 */
5633 if (AutoVacPID != 0)
5634 {
5635 AutoVacWorkerFailed();
5636 avlauncher_needs_signal = true;
5637 }
5638 }
5639
5640 /*
5641 * MaybeStartWalReceiver
5642 * Start the WAL receiver process, if not running and our state allows.
5643 *
5644 * Note: if WalReceiverPID is already nonzero, it might seem that we should
5645 * clear WalReceiverRequested. However, there's a race condition if the
5646 * walreceiver terminates and the startup process immediately requests a new
5647 * one: it's quite possible to get the signal for the request before reaping
5648 * the dead walreceiver process. Better to risk launching an extra
5649 * walreceiver than to miss launching one we need. (The walreceiver code
5650 * has logic to recognize that it should go away if not needed.)
5651 */
5652 static void
MaybeStartWalReceiver(void)5653 MaybeStartWalReceiver(void)
5654 {
5655 if (WalReceiverPID == 0 &&
5656 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5657 pmState == PM_HOT_STANDBY) &&
5658 Shutdown <= SmartShutdown)
5659 {
5660 WalReceiverPID = StartWalReceiver();
5661 if (WalReceiverPID != 0)
5662 WalReceiverRequested = false;
5663 /* else leave the flag set, so we'll try again later */
5664 }
5665 }
5666
5667
5668 /*
5669 * Create the opts file
5670 */
5671 static bool
CreateOptsFile(int argc,char * argv[],char * fullprogname)5672 CreateOptsFile(int argc, char *argv[], char *fullprogname)
5673 {
5674 FILE *fp;
5675 int i;
5676
5677 #define OPTS_FILE "postmaster.opts"
5678
5679 if ((fp = fopen(OPTS_FILE, "w")) == NULL)
5680 {
5681 elog(LOG, "could not create file \"%s\": %m", OPTS_FILE);
5682 return false;
5683 }
5684
5685 fprintf(fp, "%s", fullprogname);
5686 for (i = 1; i < argc; i++)
5687 fprintf(fp, " \"%s\"", argv[i]);
5688 fputs("\n", fp);
5689
5690 if (fclose(fp))
5691 {
5692 elog(LOG, "could not write file \"%s\": %m", OPTS_FILE);
5693 return false;
5694 }
5695
5696 return true;
5697 }
5698
5699
5700 /*
5701 * MaxLivePostmasterChildren
5702 *
5703 * This reports the number of entries needed in per-child-process arrays
5704 * (the PMChildFlags array, and if EXEC_BACKEND the ShmemBackendArray).
5705 * These arrays include regular backends, autovac workers, walsenders
5706 * and background workers, but not special children nor dead_end children.
5707 * This allows the arrays to have a fixed maximum size, to wit the same
5708 * too-many-children limit enforced by canAcceptConnections(). The exact value
5709 * isn't too critical as long as it's more than MaxBackends.
5710 */
5711 int
MaxLivePostmasterChildren(void)5712 MaxLivePostmasterChildren(void)
5713 {
5714 return 2 * (MaxConnections + autovacuum_max_workers + 1 +
5715 max_wal_senders + max_worker_processes);
5716 }
5717
5718 /*
5719 * Connect background worker to a database.
5720 */
5721 void
BackgroundWorkerInitializeConnection(const char * dbname,const char * username,uint32 flags)5722 BackgroundWorkerInitializeConnection(const char *dbname, const char *username, uint32 flags)
5723 {
5724 BackgroundWorker *worker = MyBgworkerEntry;
5725
5726 /* XXX is this the right errcode? */
5727 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5728 ereport(FATAL,
5729 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5730 errmsg("database connection requirement not indicated during registration")));
5731
5732 InitPostgres(dbname, InvalidOid, username, InvalidOid, NULL, (flags & BGWORKER_BYPASS_ALLOWCONN) != 0);
5733
5734 /* it had better not gotten out of "init" mode yet */
5735 if (!IsInitProcessingMode())
5736 ereport(ERROR,
5737 (errmsg("invalid processing mode in background worker")));
5738 SetProcessingMode(NormalProcessing);
5739 }
5740
5741 /*
5742 * Connect background worker to a database using OIDs.
5743 */
5744 void
BackgroundWorkerInitializeConnectionByOid(Oid dboid,Oid useroid,uint32 flags)5745 BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid, uint32 flags)
5746 {
5747 BackgroundWorker *worker = MyBgworkerEntry;
5748
5749 /* XXX is this the right errcode? */
5750 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5751 ereport(FATAL,
5752 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5753 errmsg("database connection requirement not indicated during registration")));
5754
5755 InitPostgres(NULL, dboid, NULL, useroid, NULL, (flags & BGWORKER_BYPASS_ALLOWCONN) != 0);
5756
5757 /* it had better not gotten out of "init" mode yet */
5758 if (!IsInitProcessingMode())
5759 ereport(ERROR,
5760 (errmsg("invalid processing mode in background worker")));
5761 SetProcessingMode(NormalProcessing);
5762 }
5763
5764 /*
5765 * Block/unblock signals in a background worker
5766 */
5767 void
BackgroundWorkerBlockSignals(void)5768 BackgroundWorkerBlockSignals(void)
5769 {
5770 PG_SETMASK(&BlockSig);
5771 }
5772
5773 void
BackgroundWorkerUnblockSignals(void)5774 BackgroundWorkerUnblockSignals(void)
5775 {
5776 PG_SETMASK(&UnBlockSig);
5777 }
5778
5779 #ifdef EXEC_BACKEND
5780 static pid_t
bgworker_forkexec(int shmem_slot)5781 bgworker_forkexec(int shmem_slot)
5782 {
5783 char *av[10];
5784 int ac = 0;
5785 char forkav[MAXPGPATH];
5786
5787 snprintf(forkav, MAXPGPATH, "--forkbgworker=%d", shmem_slot);
5788
5789 av[ac++] = "postgres";
5790 av[ac++] = forkav;
5791 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5792 av[ac] = NULL;
5793
5794 Assert(ac < lengthof(av));
5795
5796 return postmaster_forkexec(ac, av);
5797 }
5798 #endif
5799
5800 /*
5801 * Start a new bgworker.
5802 * Starting time conditions must have been checked already.
5803 *
5804 * Returns true on success, false on failure.
5805 * In either case, update the RegisteredBgWorker's state appropriately.
5806 *
5807 * This code is heavily based on autovacuum.c, q.v.
5808 */
5809 static bool
do_start_bgworker(RegisteredBgWorker * rw)5810 do_start_bgworker(RegisteredBgWorker *rw)
5811 {
5812 pid_t worker_pid;
5813
5814 Assert(rw->rw_pid == 0);
5815
5816 /*
5817 * Allocate and assign the Backend element. Note we must do this before
5818 * forking, so that we can handle failures (out of memory or child-process
5819 * slots) cleanly.
5820 *
5821 * Treat failure as though the worker had crashed. That way, the
5822 * postmaster will wait a bit before attempting to start it again; if we
5823 * tried again right away, most likely we'd find ourselves hitting the
5824 * same resource-exhaustion condition.
5825 */
5826 if (!assign_backendlist_entry(rw))
5827 {
5828 rw->rw_crashed_at = GetCurrentTimestamp();
5829 return false;
5830 }
5831
5832 ereport(DEBUG1,
5833 (errmsg("starting background worker process \"%s\"",
5834 rw->rw_worker.bgw_name)));
5835
5836 #ifdef EXEC_BACKEND
5837 switch ((worker_pid = bgworker_forkexec(rw->rw_shmem_slot)))
5838 #else
5839 switch ((worker_pid = fork_process()))
5840 #endif
5841 {
5842 case -1:
5843 /* in postmaster, fork failed ... */
5844 ereport(LOG,
5845 (errmsg("could not fork worker process: %m")));
5846 /* undo what assign_backendlist_entry did */
5847 ReleasePostmasterChildSlot(rw->rw_child_slot);
5848 rw->rw_child_slot = 0;
5849 free(rw->rw_backend);
5850 rw->rw_backend = NULL;
5851 /* mark entry as crashed, so we'll try again later */
5852 rw->rw_crashed_at = GetCurrentTimestamp();
5853 break;
5854
5855 #ifndef EXEC_BACKEND
5856 case 0:
5857 /* in postmaster child ... */
5858 InitPostmasterChild();
5859
5860 /* Close the postmaster's sockets */
5861 ClosePostmasterPorts(false);
5862
5863 /*
5864 * Before blowing away PostmasterContext, save this bgworker's
5865 * data where it can find it.
5866 */
5867 MyBgworkerEntry = (BackgroundWorker *)
5868 MemoryContextAlloc(TopMemoryContext, sizeof(BackgroundWorker));
5869 memcpy(MyBgworkerEntry, &rw->rw_worker, sizeof(BackgroundWorker));
5870
5871 /* Release postmaster's working memory context */
5872 MemoryContextSwitchTo(TopMemoryContext);
5873 MemoryContextDelete(PostmasterContext);
5874 PostmasterContext = NULL;
5875
5876 StartBackgroundWorker();
5877
5878 exit(1); /* should not get here */
5879 break;
5880 #endif
5881 default:
5882 /* in postmaster, fork successful ... */
5883 rw->rw_pid = worker_pid;
5884 rw->rw_backend->pid = rw->rw_pid;
5885 ReportBackgroundWorkerPID(rw);
5886 /* add new worker to lists of backends */
5887 dlist_push_head(&BackendList, &rw->rw_backend->elem);
5888 #ifdef EXEC_BACKEND
5889 ShmemBackendArrayAdd(rw->rw_backend);
5890 #endif
5891 return true;
5892 }
5893
5894 return false;
5895 }
5896
5897 /*
5898 * Does the current postmaster state require starting a worker with the
5899 * specified start_time?
5900 */
5901 static bool
bgworker_should_start_now(BgWorkerStartTime start_time)5902 bgworker_should_start_now(BgWorkerStartTime start_time)
5903 {
5904 switch (pmState)
5905 {
5906 case PM_NO_CHILDREN:
5907 case PM_WAIT_DEAD_END:
5908 case PM_SHUTDOWN_2:
5909 case PM_SHUTDOWN:
5910 case PM_WAIT_BACKENDS:
5911 case PM_STOP_BACKENDS:
5912 break;
5913
5914 case PM_RUN:
5915 if (start_time == BgWorkerStart_RecoveryFinished)
5916 return true;
5917 /* fall through */
5918
5919 case PM_HOT_STANDBY:
5920 if (start_time == BgWorkerStart_ConsistentState)
5921 return true;
5922 /* fall through */
5923
5924 case PM_RECOVERY:
5925 case PM_STARTUP:
5926 case PM_INIT:
5927 if (start_time == BgWorkerStart_PostmasterStart)
5928 return true;
5929 /* fall through */
5930
5931 }
5932
5933 return false;
5934 }
5935
5936 /*
5937 * Allocate the Backend struct for a connected background worker, but don't
5938 * add it to the list of backends just yet.
5939 *
5940 * On failure, return false without changing any worker state.
5941 *
5942 * Some info from the Backend is copied into the passed rw.
5943 */
5944 static bool
assign_backendlist_entry(RegisteredBgWorker * rw)5945 assign_backendlist_entry(RegisteredBgWorker *rw)
5946 {
5947 Backend *bn;
5948
5949 /*
5950 * Check that database state allows another connection. Currently the
5951 * only possible failure is CAC_TOOMANY, so we just log an error message
5952 * based on that rather than checking the error code precisely.
5953 */
5954 if (canAcceptConnections(BACKEND_TYPE_BGWORKER) != CAC_OK)
5955 {
5956 ereport(LOG,
5957 (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
5958 errmsg("no slot available for new worker process")));
5959 return false;
5960 }
5961
5962 /*
5963 * Compute the cancel key that will be assigned to this session. We
5964 * probably don't need cancel keys for background workers, but we'd better
5965 * have something random in the field to prevent unfriendly people from
5966 * sending cancels to them.
5967 */
5968 if (!RandomCancelKey(&MyCancelKey))
5969 {
5970 ereport(LOG,
5971 (errcode(ERRCODE_INTERNAL_ERROR),
5972 errmsg("could not generate random cancel key")));
5973 return false;
5974 }
5975
5976 bn = malloc(sizeof(Backend));
5977 if (bn == NULL)
5978 {
5979 ereport(LOG,
5980 (errcode(ERRCODE_OUT_OF_MEMORY),
5981 errmsg("out of memory")));
5982 return false;
5983 }
5984
5985 bn->cancel_key = MyCancelKey;
5986 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5987 bn->bkend_type = BACKEND_TYPE_BGWORKER;
5988 bn->dead_end = false;
5989 bn->bgworker_notify = false;
5990
5991 rw->rw_backend = bn;
5992 rw->rw_child_slot = bn->child_slot;
5993
5994 return true;
5995 }
5996
5997 /*
5998 * If the time is right, start background worker(s).
5999 *
6000 * As a side effect, the bgworker control variables are set or reset
6001 * depending on whether more workers may need to be started.
6002 *
6003 * We limit the number of workers started per call, to avoid consuming the
6004 * postmaster's attention for too long when many such requests are pending.
6005 * As long as StartWorkerNeeded is true, ServerLoop will not block and will
6006 * call this function again after dealing with any other issues.
6007 */
6008 static void
maybe_start_bgworkers(void)6009 maybe_start_bgworkers(void)
6010 {
6011 #define MAX_BGWORKERS_TO_LAUNCH 100
6012 int num_launched = 0;
6013 TimestampTz now = 0;
6014 slist_mutable_iter iter;
6015
6016 /*
6017 * During crash recovery, we have no need to be called until the state
6018 * transition out of recovery.
6019 */
6020 if (FatalError)
6021 {
6022 StartWorkerNeeded = false;
6023 HaveCrashedWorker = false;
6024 return;
6025 }
6026
6027 /* Don't need to be called again unless we find a reason for it below */
6028 StartWorkerNeeded = false;
6029 HaveCrashedWorker = false;
6030
6031 slist_foreach_modify(iter, &BackgroundWorkerList)
6032 {
6033 RegisteredBgWorker *rw;
6034
6035 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
6036
6037 /* ignore if already running */
6038 if (rw->rw_pid != 0)
6039 continue;
6040
6041 /* if marked for death, clean up and remove from list */
6042 if (rw->rw_terminate)
6043 {
6044 ForgetBackgroundWorker(&iter);
6045 continue;
6046 }
6047
6048 /*
6049 * If this worker has crashed previously, maybe it needs to be
6050 * restarted (unless on registration it specified it doesn't want to
6051 * be restarted at all). Check how long ago did a crash last happen.
6052 * If the last crash is too recent, don't start it right away; let it
6053 * be restarted once enough time has passed.
6054 */
6055 if (rw->rw_crashed_at != 0)
6056 {
6057 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
6058 {
6059 int notify_pid;
6060
6061 notify_pid = rw->rw_worker.bgw_notify_pid;
6062
6063 ForgetBackgroundWorker(&iter);
6064
6065 /* Report worker is gone now. */
6066 if (notify_pid != 0)
6067 kill(notify_pid, SIGUSR1);
6068
6069 continue;
6070 }
6071
6072 /* read system time only when needed */
6073 if (now == 0)
6074 now = GetCurrentTimestamp();
6075
6076 if (!TimestampDifferenceExceeds(rw->rw_crashed_at, now,
6077 rw->rw_worker.bgw_restart_time * 1000))
6078 {
6079 /* Set flag to remember that we have workers to start later */
6080 HaveCrashedWorker = true;
6081 continue;
6082 }
6083 }
6084
6085 if (bgworker_should_start_now(rw->rw_worker.bgw_start_time))
6086 {
6087 /* reset crash time before trying to start worker */
6088 rw->rw_crashed_at = 0;
6089
6090 /*
6091 * Try to start the worker.
6092 *
6093 * On failure, give up processing workers for now, but set
6094 * StartWorkerNeeded so we'll come back here on the next iteration
6095 * of ServerLoop to try again. (We don't want to wait, because
6096 * there might be additional ready-to-run workers.) We could set
6097 * HaveCrashedWorker as well, since this worker is now marked
6098 * crashed, but there's no need because the next run of this
6099 * function will do that.
6100 */
6101 if (!do_start_bgworker(rw))
6102 {
6103 StartWorkerNeeded = true;
6104 return;
6105 }
6106
6107 /*
6108 * If we've launched as many workers as allowed, quit, but have
6109 * ServerLoop call us again to look for additional ready-to-run
6110 * workers. There might not be any, but we'll find out the next
6111 * time we run.
6112 */
6113 if (++num_launched >= MAX_BGWORKERS_TO_LAUNCH)
6114 {
6115 StartWorkerNeeded = true;
6116 return;
6117 }
6118 }
6119 }
6120 }
6121
6122 /*
6123 * When a backend asks to be notified about worker state changes, we
6124 * set a flag in its backend entry. The background worker machinery needs
6125 * to know when such backends exit.
6126 */
6127 bool
PostmasterMarkPIDForWorkerNotify(int pid)6128 PostmasterMarkPIDForWorkerNotify(int pid)
6129 {
6130 dlist_iter iter;
6131 Backend *bp;
6132
6133 dlist_foreach(iter, &BackendList)
6134 {
6135 bp = dlist_container(Backend, elem, iter.cur);
6136 if (bp->pid == pid)
6137 {
6138 bp->bgworker_notify = true;
6139 return true;
6140 }
6141 }
6142 return false;
6143 }
6144
6145 #ifdef EXEC_BACKEND
6146
6147 /*
6148 * The following need to be available to the save/restore_backend_variables
6149 * functions. They are marked NON_EXEC_STATIC in their home modules.
6150 */
6151 extern slock_t *ShmemLock;
6152 extern slock_t *ProcStructLock;
6153 extern PGPROC *AuxiliaryProcs;
6154 extern PMSignalData *PMSignalState;
6155 extern pgsocket pgStatSock;
6156 extern pg_time_t first_syslogger_file_time;
6157
6158 #ifndef WIN32
6159 #define write_inheritable_socket(dest, src, childpid) ((*(dest) = (src)), true)
6160 #define read_inheritable_socket(dest, src) (*(dest) = *(src))
6161 #else
6162 static bool write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE child);
6163 static bool write_inheritable_socket(InheritableSocket *dest, SOCKET src,
6164 pid_t childPid);
6165 static void read_inheritable_socket(SOCKET *dest, InheritableSocket *src);
6166 #endif
6167
6168
6169 /* Save critical backend variables into the BackendParameters struct */
6170 #ifndef WIN32
6171 static bool
save_backend_variables(BackendParameters * param,Port * port)6172 save_backend_variables(BackendParameters *param, Port *port)
6173 #else
6174 static bool
6175 save_backend_variables(BackendParameters *param, Port *port,
6176 HANDLE childProcess, pid_t childPid)
6177 #endif
6178 {
6179 memcpy(¶m->port, port, sizeof(Port));
6180 if (!write_inheritable_socket(¶m->portsocket, port->sock, childPid))
6181 return false;
6182
6183 strlcpy(param->DataDir, DataDir, MAXPGPATH);
6184
6185 memcpy(¶m->ListenSocket, &ListenSocket, sizeof(ListenSocket));
6186
6187 param->MyCancelKey = MyCancelKey;
6188 param->MyPMChildSlot = MyPMChildSlot;
6189
6190 #ifdef WIN32
6191 param->ShmemProtectiveRegion = ShmemProtectiveRegion;
6192 #endif
6193 param->UsedShmemSegID = UsedShmemSegID;
6194 param->UsedShmemSegAddr = UsedShmemSegAddr;
6195
6196 param->ShmemLock = ShmemLock;
6197 param->ShmemVariableCache = ShmemVariableCache;
6198 param->ShmemBackendArray = ShmemBackendArray;
6199
6200 #ifndef HAVE_SPINLOCKS
6201 param->SpinlockSemaArray = SpinlockSemaArray;
6202 #endif
6203 param->NamedLWLockTrancheRequests = NamedLWLockTrancheRequests;
6204 param->NamedLWLockTrancheArray = NamedLWLockTrancheArray;
6205 param->MainLWLockArray = MainLWLockArray;
6206 param->ProcStructLock = ProcStructLock;
6207 param->ProcGlobal = ProcGlobal;
6208 param->AuxiliaryProcs = AuxiliaryProcs;
6209 param->PreparedXactProcs = PreparedXactProcs;
6210 param->PMSignalState = PMSignalState;
6211 if (!write_inheritable_socket(¶m->pgStatSock, pgStatSock, childPid))
6212 return false;
6213
6214 param->PostmasterPid = PostmasterPid;
6215 param->PgStartTime = PgStartTime;
6216 param->PgReloadTime = PgReloadTime;
6217 param->first_syslogger_file_time = first_syslogger_file_time;
6218
6219 param->redirection_done = redirection_done;
6220 param->IsBinaryUpgrade = IsBinaryUpgrade;
6221 param->max_safe_fds = max_safe_fds;
6222
6223 param->MaxBackends = MaxBackends;
6224
6225 #ifdef WIN32
6226 param->PostmasterHandle = PostmasterHandle;
6227 if (!write_duplicated_handle(¶m->initial_signal_pipe,
6228 pgwin32_create_signal_listener(childPid),
6229 childProcess))
6230 return false;
6231 #else
6232 memcpy(¶m->postmaster_alive_fds, &postmaster_alive_fds,
6233 sizeof(postmaster_alive_fds));
6234 #endif
6235
6236 memcpy(¶m->syslogPipe, &syslogPipe, sizeof(syslogPipe));
6237
6238 strlcpy(param->my_exec_path, my_exec_path, MAXPGPATH);
6239
6240 strlcpy(param->pkglib_path, pkglib_path, MAXPGPATH);
6241
6242 strlcpy(param->ExtraOptions, ExtraOptions, MAXPGPATH);
6243
6244 return true;
6245 }
6246
6247
6248 #ifdef WIN32
6249 /*
6250 * Duplicate a handle for usage in a child process, and write the child
6251 * process instance of the handle to the parameter file.
6252 */
6253 static bool
write_duplicated_handle(HANDLE * dest,HANDLE src,HANDLE childProcess)6254 write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE childProcess)
6255 {
6256 HANDLE hChild = INVALID_HANDLE_VALUE;
6257
6258 if (!DuplicateHandle(GetCurrentProcess(),
6259 src,
6260 childProcess,
6261 &hChild,
6262 0,
6263 TRUE,
6264 DUPLICATE_CLOSE_SOURCE | DUPLICATE_SAME_ACCESS))
6265 {
6266 ereport(LOG,
6267 (errmsg_internal("could not duplicate handle to be written to backend parameter file: error code %lu",
6268 GetLastError())));
6269 return false;
6270 }
6271
6272 *dest = hChild;
6273 return true;
6274 }
6275
6276 /*
6277 * Duplicate a socket for usage in a child process, and write the resulting
6278 * structure to the parameter file.
6279 * This is required because a number of LSPs (Layered Service Providers) very
6280 * common on Windows (antivirus, firewalls, download managers etc) break
6281 * straight socket inheritance.
6282 */
6283 static bool
write_inheritable_socket(InheritableSocket * dest,SOCKET src,pid_t childpid)6284 write_inheritable_socket(InheritableSocket *dest, SOCKET src, pid_t childpid)
6285 {
6286 dest->origsocket = src;
6287 if (src != 0 && src != PGINVALID_SOCKET)
6288 {
6289 /* Actual socket */
6290 if (WSADuplicateSocket(src, childpid, &dest->wsainfo) != 0)
6291 {
6292 ereport(LOG,
6293 (errmsg("could not duplicate socket %d for use in backend: error code %d",
6294 (int) src, WSAGetLastError())));
6295 return false;
6296 }
6297 }
6298 return true;
6299 }
6300
6301 /*
6302 * Read a duplicate socket structure back, and get the socket descriptor.
6303 */
6304 static void
read_inheritable_socket(SOCKET * dest,InheritableSocket * src)6305 read_inheritable_socket(SOCKET *dest, InheritableSocket *src)
6306 {
6307 SOCKET s;
6308
6309 if (src->origsocket == PGINVALID_SOCKET || src->origsocket == 0)
6310 {
6311 /* Not a real socket! */
6312 *dest = src->origsocket;
6313 }
6314 else
6315 {
6316 /* Actual socket, so create from structure */
6317 s = WSASocket(FROM_PROTOCOL_INFO,
6318 FROM_PROTOCOL_INFO,
6319 FROM_PROTOCOL_INFO,
6320 &src->wsainfo,
6321 0,
6322 0);
6323 if (s == INVALID_SOCKET)
6324 {
6325 write_stderr("could not create inherited socket: error code %d\n",
6326 WSAGetLastError());
6327 exit(1);
6328 }
6329 *dest = s;
6330
6331 /*
6332 * To make sure we don't get two references to the same socket, close
6333 * the original one. (This would happen when inheritance actually
6334 * works..
6335 */
6336 closesocket(src->origsocket);
6337 }
6338 }
6339 #endif
6340
6341 static void
read_backend_variables(char * id,Port * port)6342 read_backend_variables(char *id, Port *port)
6343 {
6344 BackendParameters param;
6345
6346 #ifndef WIN32
6347 /* Non-win32 implementation reads from file */
6348 FILE *fp;
6349
6350 /* Open file */
6351 fp = AllocateFile(id, PG_BINARY_R);
6352 if (!fp)
6353 {
6354 write_stderr("could not open backend variables file \"%s\": %s\n",
6355 id, strerror(errno));
6356 exit(1);
6357 }
6358
6359 if (fread(¶m, sizeof(param), 1, fp) != 1)
6360 {
6361 write_stderr("could not read from backend variables file \"%s\": %s\n",
6362 id, strerror(errno));
6363 exit(1);
6364 }
6365
6366 /* Release file */
6367 FreeFile(fp);
6368 if (unlink(id) != 0)
6369 {
6370 write_stderr("could not remove file \"%s\": %s\n",
6371 id, strerror(errno));
6372 exit(1);
6373 }
6374 #else
6375 /* Win32 version uses mapped file */
6376 HANDLE paramHandle;
6377 BackendParameters *paramp;
6378
6379 #ifdef _WIN64
6380 paramHandle = (HANDLE) _atoi64(id);
6381 #else
6382 paramHandle = (HANDLE) atol(id);
6383 #endif
6384 paramp = MapViewOfFile(paramHandle, FILE_MAP_READ, 0, 0, 0);
6385 if (!paramp)
6386 {
6387 write_stderr("could not map view of backend variables: error code %lu\n",
6388 GetLastError());
6389 exit(1);
6390 }
6391
6392 memcpy(¶m, paramp, sizeof(BackendParameters));
6393
6394 if (!UnmapViewOfFile(paramp))
6395 {
6396 write_stderr("could not unmap view of backend variables: error code %lu\n",
6397 GetLastError());
6398 exit(1);
6399 }
6400
6401 if (!CloseHandle(paramHandle))
6402 {
6403 write_stderr("could not close handle to backend parameter variables: error code %lu\n",
6404 GetLastError());
6405 exit(1);
6406 }
6407 #endif
6408
6409 restore_backend_variables(¶m, port);
6410 }
6411
6412 /* Restore critical backend variables from the BackendParameters struct */
6413 static void
restore_backend_variables(BackendParameters * param,Port * port)6414 restore_backend_variables(BackendParameters *param, Port *port)
6415 {
6416 memcpy(port, ¶m->port, sizeof(Port));
6417 read_inheritable_socket(&port->sock, ¶m->portsocket);
6418
6419 SetDataDir(param->DataDir);
6420
6421 memcpy(&ListenSocket, ¶m->ListenSocket, sizeof(ListenSocket));
6422
6423 MyCancelKey = param->MyCancelKey;
6424 MyPMChildSlot = param->MyPMChildSlot;
6425
6426 #ifdef WIN32
6427 ShmemProtectiveRegion = param->ShmemProtectiveRegion;
6428 #endif
6429 UsedShmemSegID = param->UsedShmemSegID;
6430 UsedShmemSegAddr = param->UsedShmemSegAddr;
6431
6432 ShmemLock = param->ShmemLock;
6433 ShmemVariableCache = param->ShmemVariableCache;
6434 ShmemBackendArray = param->ShmemBackendArray;
6435
6436 #ifndef HAVE_SPINLOCKS
6437 SpinlockSemaArray = param->SpinlockSemaArray;
6438 #endif
6439 NamedLWLockTrancheRequests = param->NamedLWLockTrancheRequests;
6440 NamedLWLockTrancheArray = param->NamedLWLockTrancheArray;
6441 MainLWLockArray = param->MainLWLockArray;
6442 ProcStructLock = param->ProcStructLock;
6443 ProcGlobal = param->ProcGlobal;
6444 AuxiliaryProcs = param->AuxiliaryProcs;
6445 PreparedXactProcs = param->PreparedXactProcs;
6446 PMSignalState = param->PMSignalState;
6447 read_inheritable_socket(&pgStatSock, ¶m->pgStatSock);
6448
6449 PostmasterPid = param->PostmasterPid;
6450 PgStartTime = param->PgStartTime;
6451 PgReloadTime = param->PgReloadTime;
6452 first_syslogger_file_time = param->first_syslogger_file_time;
6453
6454 redirection_done = param->redirection_done;
6455 IsBinaryUpgrade = param->IsBinaryUpgrade;
6456 max_safe_fds = param->max_safe_fds;
6457
6458 MaxBackends = param->MaxBackends;
6459
6460 #ifdef WIN32
6461 PostmasterHandle = param->PostmasterHandle;
6462 pgwin32_initial_signal_pipe = param->initial_signal_pipe;
6463 #else
6464 memcpy(&postmaster_alive_fds, ¶m->postmaster_alive_fds,
6465 sizeof(postmaster_alive_fds));
6466 #endif
6467
6468 memcpy(&syslogPipe, ¶m->syslogPipe, sizeof(syslogPipe));
6469
6470 strlcpy(my_exec_path, param->my_exec_path, MAXPGPATH);
6471
6472 strlcpy(pkglib_path, param->pkglib_path, MAXPGPATH);
6473
6474 strlcpy(ExtraOptions, param->ExtraOptions, MAXPGPATH);
6475 }
6476
6477
6478 Size
ShmemBackendArraySize(void)6479 ShmemBackendArraySize(void)
6480 {
6481 return mul_size(MaxLivePostmasterChildren(), sizeof(Backend));
6482 }
6483
6484 void
ShmemBackendArrayAllocation(void)6485 ShmemBackendArrayAllocation(void)
6486 {
6487 Size size = ShmemBackendArraySize();
6488
6489 ShmemBackendArray = (Backend *) ShmemAlloc(size);
6490 /* Mark all slots as empty */
6491 memset(ShmemBackendArray, 0, size);
6492 }
6493
6494 static void
ShmemBackendArrayAdd(Backend * bn)6495 ShmemBackendArrayAdd(Backend *bn)
6496 {
6497 /* The array slot corresponding to my PMChildSlot should be free */
6498 int i = bn->child_slot - 1;
6499
6500 Assert(ShmemBackendArray[i].pid == 0);
6501 ShmemBackendArray[i] = *bn;
6502 }
6503
6504 static void
ShmemBackendArrayRemove(Backend * bn)6505 ShmemBackendArrayRemove(Backend *bn)
6506 {
6507 int i = bn->child_slot - 1;
6508
6509 Assert(ShmemBackendArray[i].pid == bn->pid);
6510 /* Mark the slot as empty */
6511 ShmemBackendArray[i].pid = 0;
6512 }
6513 #endif /* EXEC_BACKEND */
6514
6515
6516 #ifdef WIN32
6517
6518 /*
6519 * Subset implementation of waitpid() for Windows. We assume pid is -1
6520 * (that is, check all child processes) and options is WNOHANG (don't wait).
6521 */
6522 static pid_t
waitpid(pid_t pid,int * exitstatus,int options)6523 waitpid(pid_t pid, int *exitstatus, int options)
6524 {
6525 DWORD dwd;
6526 ULONG_PTR key;
6527 OVERLAPPED *ovl;
6528
6529 /*
6530 * Check if there are any dead children. If there are, return the pid of
6531 * the first one that died.
6532 */
6533 if (GetQueuedCompletionStatus(win32ChildQueue, &dwd, &key, &ovl, 0))
6534 {
6535 *exitstatus = (int) key;
6536 return dwd;
6537 }
6538
6539 return -1;
6540 }
6541
6542 /*
6543 * Note! Code below executes on a thread pool! All operations must
6544 * be thread safe! Note that elog() and friends must *not* be used.
6545 */
6546 static void WINAPI
pgwin32_deadchild_callback(PVOID lpParameter,BOOLEAN TimerOrWaitFired)6547 pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired)
6548 {
6549 win32_deadchild_waitinfo *childinfo = (win32_deadchild_waitinfo *) lpParameter;
6550 DWORD exitcode;
6551
6552 if (TimerOrWaitFired)
6553 return; /* timeout. Should never happen, since we use
6554 * INFINITE as timeout value. */
6555
6556 /*
6557 * Remove handle from wait - required even though it's set to wait only
6558 * once
6559 */
6560 UnregisterWaitEx(childinfo->waitHandle, NULL);
6561
6562 if (!GetExitCodeProcess(childinfo->procHandle, &exitcode))
6563 {
6564 /*
6565 * Should never happen. Inform user and set a fixed exitcode.
6566 */
6567 write_stderr("could not read exit code for process\n");
6568 exitcode = 255;
6569 }
6570
6571 if (!PostQueuedCompletionStatus(win32ChildQueue, childinfo->procId, (ULONG_PTR) exitcode, NULL))
6572 write_stderr("could not post child completion status\n");
6573
6574 /*
6575 * Handle is per-process, so we close it here instead of in the
6576 * originating thread
6577 */
6578 CloseHandle(childinfo->procHandle);
6579
6580 /*
6581 * Free struct that was allocated before the call to
6582 * RegisterWaitForSingleObject()
6583 */
6584 free(childinfo);
6585
6586 /* Queue SIGCHLD signal */
6587 pg_queue_signal(SIGCHLD);
6588 }
6589 #endif /* WIN32 */
6590
6591 /*
6592 * Initialize one and only handle for monitoring postmaster death.
6593 *
6594 * Called once in the postmaster, so that child processes can subsequently
6595 * monitor if their parent is dead.
6596 */
6597 static void
InitPostmasterDeathWatchHandle(void)6598 InitPostmasterDeathWatchHandle(void)
6599 {
6600 #ifndef WIN32
6601
6602 /*
6603 * Create a pipe. Postmaster holds the write end of the pipe open
6604 * (POSTMASTER_FD_OWN), and children hold the read end. Children can pass
6605 * the read file descriptor to select() to wake up in case postmaster
6606 * dies, or check for postmaster death with a (read() == 0). Children must
6607 * close the write end as soon as possible after forking, because EOF
6608 * won't be signaled in the read end until all processes have closed the
6609 * write fd. That is taken care of in ClosePostmasterPorts().
6610 */
6611 Assert(MyProcPid == PostmasterPid);
6612 if (pipe(postmaster_alive_fds) < 0)
6613 ereport(FATAL,
6614 (errcode_for_file_access(),
6615 errmsg_internal("could not create pipe to monitor postmaster death: %m")));
6616
6617 /*
6618 * Set O_NONBLOCK to allow testing for the fd's presence with a read()
6619 * call.
6620 */
6621 if (fcntl(postmaster_alive_fds[POSTMASTER_FD_WATCH], F_SETFL, O_NONBLOCK) == -1)
6622 ereport(FATAL,
6623 (errcode_for_socket_access(),
6624 errmsg_internal("could not set postmaster death monitoring pipe to nonblocking mode: %m")));
6625 #else
6626
6627 /*
6628 * On Windows, we use a process handle for the same purpose.
6629 */
6630 if (DuplicateHandle(GetCurrentProcess(),
6631 GetCurrentProcess(),
6632 GetCurrentProcess(),
6633 &PostmasterHandle,
6634 0,
6635 TRUE,
6636 DUPLICATE_SAME_ACCESS) == 0)
6637 ereport(FATAL,
6638 (errmsg_internal("could not duplicate postmaster handle: error code %lu",
6639 GetLastError())));
6640 #endif /* WIN32 */
6641 }
6642