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) 1996-2020, 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/interrupt.h"
116 #include "postmaster/pgarch.h"
117 #include "postmaster/postmaster.h"
118 #include "postmaster/syslogger.h"
119 #include "replication/logicallauncher.h"
120 #include "replication/walsender.h"
121 #include "storage/fd.h"
122 #include "storage/ipc.h"
123 #include "storage/pg_shmem.h"
124 #include "storage/pmsignal.h"
125 #include "storage/proc.h"
126 #include "tcop/tcopprot.h"
127 #include "utils/builtins.h"
128 #include "utils/datetime.h"
129 #include "utils/memutils.h"
130 #include "utils/pidfile.h"
131 #include "utils/ps_status.h"
132 #include "utils/timeout.h"
133 #include "utils/timestamp.h"
134 #include "utils/varlena.h"
135
136 #ifdef EXEC_BACKEND
137 #include "storage/spin.h"
138 #endif
139
140
141 /*
142 * Possible types of a backend. Beyond being the possible bkend_type values in
143 * struct bkend, these are OR-able request flag bits for SignalSomeChildren()
144 * and CountChildren().
145 */
146 #define BACKEND_TYPE_NORMAL 0x0001 /* normal backend */
147 #define BACKEND_TYPE_AUTOVAC 0x0002 /* autovacuum worker process */
148 #define BACKEND_TYPE_WALSND 0x0004 /* walsender process */
149 #define BACKEND_TYPE_BGWORKER 0x0008 /* bgworker process */
150 #define BACKEND_TYPE_ALL 0x000F /* OR of all the above */
151
152 /*
153 * List of active backends (or child processes anyway; we don't actually
154 * know whether a given child has become a backend or is still in the
155 * authorization phase). This is used mainly to keep track of how many
156 * children we have and send them appropriate signals when necessary.
157 *
158 * "Special" children such as the startup, bgwriter and autovacuum launcher
159 * tasks are not in this list. Autovacuum worker and walsender are in it.
160 * Also, "dead_end" children are in it: these are children launched just for
161 * the purpose of sending a friendly rejection message to a would-be client.
162 * We must track them because they are attached to shared memory, but we know
163 * they will never become live backends. dead_end children are not assigned a
164 * PMChildSlot.
165 *
166 * Background workers are in this list, too.
167 */
168 typedef struct bkend
169 {
170 pid_t pid; /* process id of backend */
171 int32 cancel_key; /* cancel key for cancels for this backend */
172 int child_slot; /* PMChildSlot for this backend, if any */
173
174 /*
175 * Flavor of backend or auxiliary process. Note that BACKEND_TYPE_WALSND
176 * backends initially announce themselves as BACKEND_TYPE_NORMAL, so if
177 * bkend_type is normal, you should check for a recent transition.
178 */
179 int bkend_type;
180 bool dead_end; /* is it going to send an error and quit? */
181 bool bgworker_notify; /* gets bgworker start/stop notifications */
182 dlist_node elem; /* list link in BackendList */
183 } Backend;
184
185 static dlist_head BackendList = DLIST_STATIC_INIT(BackendList);
186
187 #ifdef EXEC_BACKEND
188 static Backend *ShmemBackendArray;
189 #endif
190
191 BackgroundWorker *MyBgworkerEntry = NULL;
192
193
194
195 /* The socket number we are listening for connections on */
196 int PostPortNumber;
197
198 /* The directory names for Unix socket(s) */
199 char *Unix_socket_directories;
200
201 /* The TCP listen address(es) */
202 char *ListenAddresses;
203
204 /*
205 * ReservedBackends is the number of backends reserved for superuser use.
206 * This number is taken out of the pool size given by MaxConnections so
207 * number of backend slots available to non-superusers is
208 * (MaxConnections - ReservedBackends). Note what this really means is
209 * "if there are <= ReservedBackends connections available, only superusers
210 * can make new connections" --- pre-existing superuser connections don't
211 * count against the limit.
212 */
213 int ReservedBackends;
214
215 /* The socket(s) we're listening to. */
216 #define MAXLISTEN 64
217 static pgsocket ListenSocket[MAXLISTEN];
218
219 /*
220 * Set by the -o option
221 */
222 static char ExtraOptions[MAXPGPATH];
223
224 /*
225 * These globals control the behavior of the postmaster in case some
226 * backend dumps core. Normally, it kills all peers of the dead backend
227 * and reinitializes shared memory. By specifying -s or -n, we can have
228 * the postmaster stop (rather than kill) peers and not reinitialize
229 * shared data structures. (Reinit is currently dead code, though.)
230 */
231 static bool Reinit = true;
232 static int SendStop = false;
233
234 /* still more option variables */
235 bool EnableSSL = false;
236
237 int PreAuthDelay = 0;
238 int AuthenticationTimeout = 60;
239
240 bool log_hostname; /* for ps display and logging */
241 bool Log_connections = false;
242 bool Db_user_namespace = false;
243
244 bool enable_bonjour = false;
245 char *bonjour_name;
246 bool restart_after_crash = true;
247
248 /* PIDs of special child processes; 0 when not running */
249 static pid_t StartupPID = 0,
250 BgWriterPID = 0,
251 CheckpointerPID = 0,
252 WalWriterPID = 0,
253 WalReceiverPID = 0,
254 AutoVacPID = 0,
255 PgArchPID = 0,
256 PgStatPID = 0,
257 SysLoggerPID = 0;
258
259 /* Startup process's status */
260 typedef enum
261 {
262 STARTUP_NOT_RUNNING,
263 STARTUP_RUNNING,
264 STARTUP_SIGNALED, /* we sent it a SIGQUIT or SIGKILL */
265 STARTUP_CRASHED
266 } StartupStatusEnum;
267
268 static StartupStatusEnum StartupStatus = STARTUP_NOT_RUNNING;
269
270 /* Startup/shutdown state */
271 #define NoShutdown 0
272 #define SmartShutdown 1
273 #define FastShutdown 2
274 #define ImmediateShutdown 3
275
276 static int Shutdown = NoShutdown;
277
278 static bool FatalError = false; /* T if recovering from backend crash */
279
280 /*
281 * We use a simple state machine to control startup, shutdown, and
282 * crash recovery (which is rather like shutdown followed by startup).
283 *
284 * After doing all the postmaster initialization work, we enter PM_STARTUP
285 * state and the startup process is launched. The startup process begins by
286 * reading the control file and other preliminary initialization steps.
287 * In a normal startup, or after crash recovery, the startup process exits
288 * with exit code 0 and we switch to PM_RUN state. However, archive recovery
289 * is handled specially since it takes much longer and we would like to support
290 * hot standby during archive recovery.
291 *
292 * When the startup process is ready to start archive recovery, it signals the
293 * postmaster, and we switch to PM_RECOVERY state. The background writer and
294 * checkpointer are launched, while the startup process continues applying WAL.
295 * If Hot Standby is enabled, then, after reaching a consistent point in WAL
296 * redo, startup process signals us again, and we switch to PM_HOT_STANDBY
297 * state and begin accepting connections to perform read-only queries. When
298 * archive recovery is finished, the startup process exits with exit code 0
299 * and we switch to PM_RUN state.
300 *
301 * Normal child backends can only be launched when we are in PM_RUN or
302 * PM_HOT_STANDBY state. (connsAllowed can also restrict launching.)
303 * In other states we handle connection requests by launching "dead_end"
304 * child processes, which will simply send the client an error message and
305 * quit. (We track these in the BackendList so that we can know when they
306 * are all gone; this is important because they're still connected to shared
307 * memory, and would interfere with an attempt to destroy the shmem segment,
308 * possibly leading to SHMALL failure when we try to make a new one.)
309 * In PM_WAIT_DEAD_END state we are waiting for all the dead_end children
310 * to drain out of the system, and therefore stop accepting connection
311 * requests at all until the last existing child has quit (which hopefully
312 * will not be very long).
313 *
314 * Notice that this state variable does not distinguish *why* we entered
315 * states later than PM_RUN --- Shutdown and FatalError must be consulted
316 * to find that out. FatalError is never true in PM_RECOVERY, PM_HOT_STANDBY,
317 * or PM_RUN states, nor in PM_SHUTDOWN states (because we don't enter those
318 * states when trying to recover from a crash). It can be true in PM_STARTUP
319 * state, because we don't clear it until we've successfully started WAL redo.
320 */
321 typedef enum
322 {
323 PM_INIT, /* postmaster starting */
324 PM_STARTUP, /* waiting for startup subprocess */
325 PM_RECOVERY, /* in archive recovery mode */
326 PM_HOT_STANDBY, /* in hot standby mode */
327 PM_RUN, /* normal "database is alive" state */
328 PM_STOP_BACKENDS, /* need to stop remaining backends */
329 PM_WAIT_BACKENDS, /* waiting for live backends to exit */
330 PM_SHUTDOWN, /* waiting for checkpointer to do shutdown
331 * ckpt */
332 PM_SHUTDOWN_2, /* waiting for archiver and walsenders to
333 * finish */
334 PM_WAIT_DEAD_END, /* waiting for dead_end children to exit */
335 PM_NO_CHILDREN /* all important children have exited */
336 } PMState;
337
338 static PMState pmState = PM_INIT;
339
340 /*
341 * While performing a "smart shutdown", we restrict new connections but stay
342 * in PM_RUN or PM_HOT_STANDBY state until all the client backends are gone.
343 * connsAllowed is a sub-state indicator showing the active restriction.
344 * It is of no interest unless pmState is PM_RUN or PM_HOT_STANDBY.
345 */
346 typedef enum
347 {
348 ALLOW_ALL_CONNS, /* normal not-shutting-down state */
349 ALLOW_SUPERUSER_CONNS, /* only superusers can connect */
350 ALLOW_NO_CONNS /* no new connections allowed, period */
351 } ConnsAllowedState;
352
353 static ConnsAllowedState connsAllowed = ALLOW_ALL_CONNS;
354
355 /* Start time of SIGKILL timeout during immediate shutdown or child crash */
356 /* Zero means timeout is not running */
357 static time_t AbortStartTime = 0;
358
359 /* Length of said timeout */
360 #define SIGKILL_CHILDREN_AFTER_SECS 5
361
362 static bool ReachedNormalRunning = false; /* T if we've reached PM_RUN */
363
364 bool ClientAuthInProgress = false; /* T during new-client
365 * authentication */
366
367 bool redirection_done = false; /* stderr redirected for syslogger? */
368
369 /* received START_AUTOVAC_LAUNCHER signal */
370 static volatile sig_atomic_t start_autovac_launcher = false;
371
372 /* the launcher needs to be signaled to communicate some condition */
373 static volatile bool avlauncher_needs_signal = false;
374
375 /* received START_WALRECEIVER signal */
376 static volatile sig_atomic_t WalReceiverRequested = false;
377
378 /* set when there's a worker that needs to be started up */
379 static volatile bool StartWorkerNeeded = true;
380 static volatile bool HaveCrashedWorker = false;
381
382 #ifdef USE_SSL
383 /* Set when and if SSL has been initialized properly */
384 static bool LoadedSSL = false;
385 #endif
386
387 #ifdef USE_BONJOUR
388 static DNSServiceRef bonjour_sdref = NULL;
389 #endif
390
391 /*
392 * postmaster.c - function prototypes
393 */
394 static void CloseServerPorts(int status, Datum arg);
395 static void unlink_external_pid_file(int status, Datum arg);
396 static void getInstallationPaths(const char *argv0);
397 static void checkControlFile(void);
398 static Port *ConnCreate(int serverFd);
399 static void ConnFree(Port *port);
400 static void reset_shared(void);
401 static void SIGHUP_handler(SIGNAL_ARGS);
402 static void pmdie(SIGNAL_ARGS);
403 static void reaper(SIGNAL_ARGS);
404 static void sigusr1_handler(SIGNAL_ARGS);
405 static void process_startup_packet_die(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 * Register the apply launcher. Since it registers a background worker,
989 * it needs to be called before InitializeMaxBackends(), and it's probably
990 * a good idea to call it before any modules had chance to take the
991 * background worker slots.
992 */
993 ApplyLauncherRegister();
994
995 /*
996 * process any libraries that should be preloaded at postmaster start
997 */
998 process_shared_preload_libraries();
999
1000 /*
1001 * Initialize SSL library, if specified.
1002 */
1003 #ifdef USE_SSL
1004 if (EnableSSL)
1005 {
1006 (void) secure_initialize(true);
1007 LoadedSSL = true;
1008 }
1009 #endif
1010
1011 /*
1012 * Now that loadable modules have had their chance to register background
1013 * workers, calculate MaxBackends.
1014 */
1015 InitializeMaxBackends();
1016
1017 /*
1018 * Set up shared memory and semaphores.
1019 */
1020 reset_shared();
1021
1022 /*
1023 * Estimate number of openable files. This must happen after setting up
1024 * semaphores, because on some platforms semaphores count as open files.
1025 */
1026 set_max_safe_fds();
1027
1028 /*
1029 * Set reference point for stack-depth checking.
1030 */
1031 set_stack_base();
1032
1033 /*
1034 * Initialize pipe (or process handle on Windows) that allows children to
1035 * wake up from sleep on postmaster death.
1036 */
1037 InitPostmasterDeathWatchHandle();
1038
1039 #ifdef WIN32
1040
1041 /*
1042 * Initialize I/O completion port used to deliver list of dead children.
1043 */
1044 win32ChildQueue = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 1);
1045 if (win32ChildQueue == NULL)
1046 ereport(FATAL,
1047 (errmsg("could not create I/O completion port for child queue")));
1048 #endif
1049
1050 #ifdef EXEC_BACKEND
1051 /* Write out nondefault GUC settings for child processes to use */
1052 write_nondefault_variables(PGC_POSTMASTER);
1053
1054 /*
1055 * Clean out the temp directory used to transmit parameters to child
1056 * processes (see internal_forkexec, below). We must do this before
1057 * launching any child processes, else we have a race condition: we could
1058 * remove a parameter file before the child can read it. It should be
1059 * safe to do so now, because we verified earlier that there are no
1060 * conflicting Postgres processes in this data directory.
1061 */
1062 RemovePgTempFilesInDir(PG_TEMP_FILES_DIR, true, false);
1063 #endif
1064
1065 /*
1066 * Forcibly remove the files signaling a standby promotion request.
1067 * Otherwise, the existence of those files triggers a promotion too early,
1068 * whether a user wants that or not.
1069 *
1070 * This removal of files is usually unnecessary because they can exist
1071 * only during a few moments during a standby promotion. However there is
1072 * a race condition: if pg_ctl promote is executed and creates the files
1073 * during a promotion, the files can stay around even after the server is
1074 * brought up to new master. Then, if new standby starts by using the
1075 * backup taken from that master, the files can exist at the server
1076 * startup and should be removed in order to avoid an unexpected
1077 * promotion.
1078 *
1079 * Note that promotion signal files need to be removed before the startup
1080 * process is invoked. Because, after that, they can be used by
1081 * postmaster's SIGUSR1 signal handler.
1082 */
1083 RemovePromoteSignalFiles();
1084
1085 /* Do the same for logrotate signal file */
1086 RemoveLogrotateSignalFiles();
1087
1088 /* Remove any outdated file holding the current log filenames. */
1089 if (unlink(LOG_METAINFO_DATAFILE) < 0 && errno != ENOENT)
1090 ereport(LOG,
1091 (errcode_for_file_access(),
1092 errmsg("could not remove file \"%s\": %m",
1093 LOG_METAINFO_DATAFILE)));
1094
1095 /*
1096 * If enabled, start up syslogger collection subprocess
1097 */
1098 SysLoggerPID = SysLogger_Start();
1099
1100 /*
1101 * Reset whereToSendOutput from DestDebug (its starting state) to
1102 * DestNone. This stops ereport from sending log messages to stderr unless
1103 * Log_destination permits. We don't do this until the postmaster is
1104 * fully launched, since startup failures may as well be reported to
1105 * stderr.
1106 *
1107 * If we are in fact disabling logging to stderr, first emit a log message
1108 * saying so, to provide a breadcrumb trail for users who may not remember
1109 * that their logging is configured to go somewhere else.
1110 */
1111 if (!(Log_destination & LOG_DESTINATION_STDERR))
1112 ereport(LOG,
1113 (errmsg("ending log output to stderr"),
1114 errhint("Future log output will go to log destination \"%s\".",
1115 Log_destination_string)));
1116
1117 whereToSendOutput = DestNone;
1118
1119 /*
1120 * Report server startup in log. While we could emit this much earlier,
1121 * it seems best to do so after starting the log collector, if we intend
1122 * to use one.
1123 */
1124 ereport(LOG,
1125 (errmsg("starting %s", PG_VERSION_STR)));
1126
1127 /*
1128 * Establish input sockets.
1129 *
1130 * First, mark them all closed, and set up an on_proc_exit function that's
1131 * charged with closing the sockets again at postmaster shutdown.
1132 */
1133 for (i = 0; i < MAXLISTEN; i++)
1134 ListenSocket[i] = PGINVALID_SOCKET;
1135
1136 on_proc_exit(CloseServerPorts, 0);
1137
1138 if (ListenAddresses)
1139 {
1140 char *rawstring;
1141 List *elemlist;
1142 ListCell *l;
1143 int success = 0;
1144
1145 /* Need a modifiable copy of ListenAddresses */
1146 rawstring = pstrdup(ListenAddresses);
1147
1148 /* Parse string into list of hostnames */
1149 if (!SplitGUCList(rawstring, ',', &elemlist))
1150 {
1151 /* syntax error in list */
1152 ereport(FATAL,
1153 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1154 errmsg("invalid list syntax in parameter \"%s\"",
1155 "listen_addresses")));
1156 }
1157
1158 foreach(l, elemlist)
1159 {
1160 char *curhost = (char *) lfirst(l);
1161
1162 if (strcmp(curhost, "*") == 0)
1163 status = StreamServerPort(AF_UNSPEC, NULL,
1164 (unsigned short) PostPortNumber,
1165 NULL,
1166 ListenSocket, MAXLISTEN);
1167 else
1168 status = StreamServerPort(AF_UNSPEC, curhost,
1169 (unsigned short) PostPortNumber,
1170 NULL,
1171 ListenSocket, MAXLISTEN);
1172
1173 if (status == STATUS_OK)
1174 {
1175 success++;
1176 /* record the first successful host addr in lockfile */
1177 if (!listen_addr_saved)
1178 {
1179 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, curhost);
1180 listen_addr_saved = true;
1181 }
1182 }
1183 else
1184 ereport(WARNING,
1185 (errmsg("could not create listen socket for \"%s\"",
1186 curhost)));
1187 }
1188
1189 if (!success && elemlist != NIL)
1190 ereport(FATAL,
1191 (errmsg("could not create any TCP/IP sockets")));
1192
1193 list_free(elemlist);
1194 pfree(rawstring);
1195 }
1196
1197 #ifdef USE_BONJOUR
1198 /* Register for Bonjour only if we opened TCP socket(s) */
1199 if (enable_bonjour && ListenSocket[0] != PGINVALID_SOCKET)
1200 {
1201 DNSServiceErrorType err;
1202
1203 /*
1204 * We pass 0 for interface_index, which will result in registering on
1205 * all "applicable" interfaces. It's not entirely clear from the
1206 * DNS-SD docs whether this would be appropriate if we have bound to
1207 * just a subset of the available network interfaces.
1208 */
1209 err = DNSServiceRegister(&bonjour_sdref,
1210 0,
1211 0,
1212 bonjour_name,
1213 "_postgresql._tcp.",
1214 NULL,
1215 NULL,
1216 pg_hton16(PostPortNumber),
1217 0,
1218 NULL,
1219 NULL,
1220 NULL);
1221 if (err != kDNSServiceErr_NoError)
1222 elog(LOG, "DNSServiceRegister() failed: error code %ld",
1223 (long) err);
1224
1225 /*
1226 * We don't bother to read the mDNS daemon's reply, and we expect that
1227 * it will automatically terminate our registration when the socket is
1228 * closed at postmaster termination. So there's nothing more to be
1229 * done here. However, the bonjour_sdref is kept around so that
1230 * forked children can close their copies of the socket.
1231 */
1232 }
1233 #endif
1234
1235 #ifdef HAVE_UNIX_SOCKETS
1236 if (Unix_socket_directories)
1237 {
1238 char *rawstring;
1239 List *elemlist;
1240 ListCell *l;
1241 int success = 0;
1242
1243 /* Need a modifiable copy of Unix_socket_directories */
1244 rawstring = pstrdup(Unix_socket_directories);
1245
1246 /* Parse string into list of directories */
1247 if (!SplitDirectoriesString(rawstring, ',', &elemlist))
1248 {
1249 /* syntax error in list */
1250 ereport(FATAL,
1251 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1252 errmsg("invalid list syntax in parameter \"%s\"",
1253 "unix_socket_directories")));
1254 }
1255
1256 foreach(l, elemlist)
1257 {
1258 char *socketdir = (char *) lfirst(l);
1259
1260 status = StreamServerPort(AF_UNIX, NULL,
1261 (unsigned short) PostPortNumber,
1262 socketdir,
1263 ListenSocket, MAXLISTEN);
1264
1265 if (status == STATUS_OK)
1266 {
1267 success++;
1268 /* record the first successful Unix socket in lockfile */
1269 if (success == 1)
1270 AddToDataDirLockFile(LOCK_FILE_LINE_SOCKET_DIR, socketdir);
1271 }
1272 else
1273 ereport(WARNING,
1274 (errmsg("could not create Unix-domain socket in directory \"%s\"",
1275 socketdir)));
1276 }
1277
1278 if (!success && elemlist != NIL)
1279 ereport(FATAL,
1280 (errmsg("could not create any Unix-domain sockets")));
1281
1282 list_free_deep(elemlist);
1283 pfree(rawstring);
1284 }
1285 #endif
1286
1287 /*
1288 * check that we have some socket to listen on
1289 */
1290 if (ListenSocket[0] == PGINVALID_SOCKET)
1291 ereport(FATAL,
1292 (errmsg("no socket created for listening")));
1293
1294 /*
1295 * If no valid TCP ports, write an empty line for listen address,
1296 * indicating the Unix socket must be used. Note that this line is not
1297 * added to the lock file until there is a socket backing it.
1298 */
1299 if (!listen_addr_saved)
1300 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, "");
1301
1302 /*
1303 * Record postmaster options. We delay this till now to avoid recording
1304 * bogus options (eg, unusable port number).
1305 */
1306 if (!CreateOptsFile(argc, argv, my_exec_path))
1307 ExitPostmaster(1);
1308
1309 /*
1310 * Write the external PID file if requested
1311 */
1312 if (external_pid_file)
1313 {
1314 FILE *fpidfile = fopen(external_pid_file, "w");
1315
1316 if (fpidfile)
1317 {
1318 fprintf(fpidfile, "%d\n", MyProcPid);
1319 fclose(fpidfile);
1320
1321 /* Make PID file world readable */
1322 if (chmod(external_pid_file, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH) != 0)
1323 write_stderr("%s: could not change permissions of external PID file \"%s\": %s\n",
1324 progname, external_pid_file, strerror(errno));
1325 }
1326 else
1327 write_stderr("%s: could not write external PID file \"%s\": %s\n",
1328 progname, external_pid_file, strerror(errno));
1329
1330 on_proc_exit(unlink_external_pid_file, 0);
1331 }
1332
1333 /*
1334 * Remove old temporary files. At this point there can be no other
1335 * Postgres processes running in this directory, so this should be safe.
1336 */
1337 RemovePgTempFiles();
1338
1339 /*
1340 * Initialize stats collection subsystem (this does NOT start the
1341 * collector process!)
1342 */
1343 pgstat_init();
1344
1345 /*
1346 * Initialize the autovacuum subsystem (again, no process start yet)
1347 */
1348 autovac_init();
1349
1350 /*
1351 * Load configuration files for client authentication.
1352 */
1353 if (!load_hba())
1354 {
1355 /*
1356 * It makes no sense to continue if we fail to load the HBA file,
1357 * since there is no way to connect to the database in this case.
1358 */
1359 ereport(FATAL,
1360 (errmsg("could not load pg_hba.conf")));
1361 }
1362 if (!load_ident())
1363 {
1364 /*
1365 * We can start up without the IDENT file, although it means that you
1366 * cannot log in using any of the authentication methods that need a
1367 * user name mapping. load_ident() already logged the details of error
1368 * to the log.
1369 */
1370 }
1371
1372 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1373
1374 /*
1375 * On macOS, libintl replaces setlocale() with a version that calls
1376 * CFLocaleCopyCurrent() when its second argument is "" and every relevant
1377 * environment variable is unset or empty. CFLocaleCopyCurrent() makes
1378 * the process multithreaded. The postmaster calls sigprocmask() and
1379 * calls fork() without an immediate exec(), both of which have undefined
1380 * behavior in a multithreaded program. A multithreaded postmaster is the
1381 * normal case on Windows, which offers neither fork() nor sigprocmask().
1382 */
1383 if (pthread_is_threaded_np() != 0)
1384 ereport(FATAL,
1385 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1386 errmsg("postmaster became multithreaded during startup"),
1387 errhint("Set the LC_ALL environment variable to a valid locale.")));
1388 #endif
1389
1390 /*
1391 * Remember postmaster startup time
1392 */
1393 PgStartTime = GetCurrentTimestamp();
1394
1395 /*
1396 * Report postmaster status in the postmaster.pid file, to allow pg_ctl to
1397 * see what's happening.
1398 */
1399 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STARTING);
1400
1401 /*
1402 * We're ready to rock and roll...
1403 */
1404 StartupPID = StartupDataBase();
1405 Assert(StartupPID != 0);
1406 StartupStatus = STARTUP_RUNNING;
1407 pmState = PM_STARTUP;
1408
1409 /* Some workers may be scheduled to start now */
1410 maybe_start_bgworkers();
1411
1412 status = ServerLoop();
1413
1414 /*
1415 * ServerLoop probably shouldn't ever return, but if it does, close down.
1416 */
1417 ExitPostmaster(status != STATUS_OK);
1418
1419 abort(); /* not reached */
1420 }
1421
1422
1423 /*
1424 * on_proc_exit callback to close server's listen sockets
1425 */
1426 static void
CloseServerPorts(int status,Datum arg)1427 CloseServerPorts(int status, Datum arg)
1428 {
1429 int i;
1430
1431 /*
1432 * First, explicitly close all the socket FDs. We used to just let this
1433 * happen implicitly at postmaster exit, but it's better to close them
1434 * before we remove the postmaster.pid lockfile; otherwise there's a race
1435 * condition if a new postmaster wants to re-use the TCP port number.
1436 */
1437 for (i = 0; i < MAXLISTEN; i++)
1438 {
1439 if (ListenSocket[i] != PGINVALID_SOCKET)
1440 {
1441 StreamClose(ListenSocket[i]);
1442 ListenSocket[i] = PGINVALID_SOCKET;
1443 }
1444 }
1445
1446 /*
1447 * Next, remove any filesystem entries for Unix sockets. To avoid race
1448 * conditions against incoming postmasters, this must happen after closing
1449 * the sockets and before removing lock files.
1450 */
1451 RemoveSocketFiles();
1452
1453 /*
1454 * We don't do anything about socket lock files here; those will be
1455 * removed in a later on_proc_exit callback.
1456 */
1457 }
1458
1459 /*
1460 * on_proc_exit callback to delete external_pid_file
1461 */
1462 static void
unlink_external_pid_file(int status,Datum arg)1463 unlink_external_pid_file(int status, Datum arg)
1464 {
1465 if (external_pid_file)
1466 unlink(external_pid_file);
1467 }
1468
1469
1470 /*
1471 * Compute and check the directory paths to files that are part of the
1472 * installation (as deduced from the postgres executable's own location)
1473 */
1474 static void
getInstallationPaths(const char * argv0)1475 getInstallationPaths(const char *argv0)
1476 {
1477 DIR *pdir;
1478
1479 /* Locate the postgres executable itself */
1480 if (find_my_exec(argv0, my_exec_path) < 0)
1481 elog(FATAL, "%s: could not locate my own executable path", argv0);
1482
1483 #ifdef EXEC_BACKEND
1484 /* Locate executable backend before we change working directory */
1485 if (find_other_exec(argv0, "postgres", PG_BACKEND_VERSIONSTR,
1486 postgres_exec_path) < 0)
1487 ereport(FATAL,
1488 (errmsg("%s: could not locate matching postgres executable",
1489 argv0)));
1490 #endif
1491
1492 /*
1493 * Locate the pkglib directory --- this has to be set early in case we try
1494 * to load any modules from it in response to postgresql.conf entries.
1495 */
1496 get_pkglib_path(my_exec_path, pkglib_path);
1497
1498 /*
1499 * Verify that there's a readable directory there; otherwise the Postgres
1500 * installation is incomplete or corrupt. (A typical cause of this
1501 * failure is that the postgres executable has been moved or hardlinked to
1502 * some directory that's not a sibling of the installation lib/
1503 * directory.)
1504 */
1505 pdir = AllocateDir(pkglib_path);
1506 if (pdir == NULL)
1507 ereport(ERROR,
1508 (errcode_for_file_access(),
1509 errmsg("could not open directory \"%s\": %m",
1510 pkglib_path),
1511 errhint("This may indicate an incomplete PostgreSQL installation, or that the file \"%s\" has been moved away from its proper location.",
1512 my_exec_path)));
1513 FreeDir(pdir);
1514
1515 /*
1516 * XXX is it worth similarly checking the share/ directory? If the lib/
1517 * directory is there, then share/ probably is too.
1518 */
1519 }
1520
1521 /*
1522 * Check that pg_control exists in the correct location in the data directory.
1523 *
1524 * No attempt is made to validate the contents of pg_control here. This is
1525 * just a sanity check to see if we are looking at a real data directory.
1526 */
1527 static void
checkControlFile(void)1528 checkControlFile(void)
1529 {
1530 char path[MAXPGPATH];
1531 FILE *fp;
1532
1533 snprintf(path, sizeof(path), "%s/global/pg_control", DataDir);
1534
1535 fp = AllocateFile(path, PG_BINARY_R);
1536 if (fp == NULL)
1537 {
1538 write_stderr("%s: could not find the database system\n"
1539 "Expected to find it in the directory \"%s\",\n"
1540 "but could not open file \"%s\": %s\n",
1541 progname, DataDir, path, strerror(errno));
1542 ExitPostmaster(2);
1543 }
1544 FreeFile(fp);
1545 }
1546
1547 /*
1548 * Determine how long should we let ServerLoop sleep.
1549 *
1550 * In normal conditions we wait at most one minute, to ensure that the other
1551 * background tasks handled by ServerLoop get done even when no requests are
1552 * arriving. However, if there are background workers waiting to be started,
1553 * we don't actually sleep so that they are quickly serviced. Other exception
1554 * cases are as shown in the code.
1555 */
1556 static void
DetermineSleepTime(struct timeval * timeout)1557 DetermineSleepTime(struct timeval *timeout)
1558 {
1559 TimestampTz next_wakeup = 0;
1560
1561 /*
1562 * Normal case: either there are no background workers at all, or we're in
1563 * a shutdown sequence (during which we ignore bgworkers altogether).
1564 */
1565 if (Shutdown > NoShutdown ||
1566 (!StartWorkerNeeded && !HaveCrashedWorker))
1567 {
1568 if (AbortStartTime != 0)
1569 {
1570 /* time left to abort; clamp to 0 in case it already expired */
1571 timeout->tv_sec = SIGKILL_CHILDREN_AFTER_SECS -
1572 (time(NULL) - AbortStartTime);
1573 timeout->tv_sec = Max(timeout->tv_sec, 0);
1574 timeout->tv_usec = 0;
1575 }
1576 else
1577 {
1578 timeout->tv_sec = 60;
1579 timeout->tv_usec = 0;
1580 }
1581 return;
1582 }
1583
1584 if (StartWorkerNeeded)
1585 {
1586 timeout->tv_sec = 0;
1587 timeout->tv_usec = 0;
1588 return;
1589 }
1590
1591 if (HaveCrashedWorker)
1592 {
1593 slist_mutable_iter siter;
1594
1595 /*
1596 * When there are crashed bgworkers, we sleep just long enough that
1597 * they are restarted when they request to be. Scan the list to
1598 * determine the minimum of all wakeup times according to most recent
1599 * crash time and requested restart interval.
1600 */
1601 slist_foreach_modify(siter, &BackgroundWorkerList)
1602 {
1603 RegisteredBgWorker *rw;
1604 TimestampTz this_wakeup;
1605
1606 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
1607
1608 if (rw->rw_crashed_at == 0)
1609 continue;
1610
1611 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART
1612 || rw->rw_terminate)
1613 {
1614 ForgetBackgroundWorker(&siter);
1615 continue;
1616 }
1617
1618 this_wakeup = TimestampTzPlusMilliseconds(rw->rw_crashed_at,
1619 1000L * rw->rw_worker.bgw_restart_time);
1620 if (next_wakeup == 0 || this_wakeup < next_wakeup)
1621 next_wakeup = this_wakeup;
1622 }
1623 }
1624
1625 if (next_wakeup != 0)
1626 {
1627 long secs;
1628 int microsecs;
1629
1630 TimestampDifference(GetCurrentTimestamp(), next_wakeup,
1631 &secs, µsecs);
1632 timeout->tv_sec = secs;
1633 timeout->tv_usec = microsecs;
1634
1635 /* Ensure we don't exceed one minute */
1636 if (timeout->tv_sec > 60)
1637 {
1638 timeout->tv_sec = 60;
1639 timeout->tv_usec = 0;
1640 }
1641 }
1642 else
1643 {
1644 timeout->tv_sec = 60;
1645 timeout->tv_usec = 0;
1646 }
1647 }
1648
1649 /*
1650 * Main idle loop of postmaster
1651 *
1652 * NB: Needs to be called with signals blocked
1653 */
1654 static int
ServerLoop(void)1655 ServerLoop(void)
1656 {
1657 fd_set readmask;
1658 int nSockets;
1659 time_t last_lockfile_recheck_time,
1660 last_touch_time;
1661
1662 last_lockfile_recheck_time = last_touch_time = time(NULL);
1663
1664 nSockets = initMasks(&readmask);
1665
1666 for (;;)
1667 {
1668 fd_set rmask;
1669 int selres;
1670 time_t now;
1671
1672 /*
1673 * Wait for a connection request to arrive.
1674 *
1675 * We block all signals except while sleeping. That makes it safe for
1676 * signal handlers, which again block all signals while executing, to
1677 * do nontrivial work.
1678 *
1679 * If we are in PM_WAIT_DEAD_END state, then we don't want to accept
1680 * any new connections, so we don't call select(), and just sleep.
1681 */
1682 memcpy((char *) &rmask, (char *) &readmask, sizeof(fd_set));
1683
1684 if (pmState == PM_WAIT_DEAD_END)
1685 {
1686 PG_SETMASK(&UnBlockSig);
1687
1688 pg_usleep(100000L); /* 100 msec seems reasonable */
1689 selres = 0;
1690
1691 PG_SETMASK(&BlockSig);
1692 }
1693 else
1694 {
1695 /* must set timeout each time; some OSes change it! */
1696 struct timeval timeout;
1697
1698 /* Needs to run with blocked signals! */
1699 DetermineSleepTime(&timeout);
1700
1701 PG_SETMASK(&UnBlockSig);
1702
1703 selres = select(nSockets, &rmask, NULL, NULL, &timeout);
1704
1705 PG_SETMASK(&BlockSig);
1706 }
1707
1708 /* Now check the select() result */
1709 if (selres < 0)
1710 {
1711 if (errno != EINTR && errno != EWOULDBLOCK)
1712 {
1713 ereport(LOG,
1714 (errcode_for_socket_access(),
1715 errmsg("select() failed in postmaster: %m")));
1716 return STATUS_ERROR;
1717 }
1718 }
1719
1720 /*
1721 * New connection pending on any of our sockets? If so, fork a child
1722 * process to deal with it.
1723 */
1724 if (selres > 0)
1725 {
1726 int i;
1727
1728 for (i = 0; i < MAXLISTEN; i++)
1729 {
1730 if (ListenSocket[i] == PGINVALID_SOCKET)
1731 break;
1732 if (FD_ISSET(ListenSocket[i], &rmask))
1733 {
1734 Port *port;
1735
1736 port = ConnCreate(ListenSocket[i]);
1737 if (port)
1738 {
1739 BackendStartup(port);
1740
1741 /*
1742 * We no longer need the open socket or port structure
1743 * in this process
1744 */
1745 StreamClose(port->sock);
1746 ConnFree(port);
1747 }
1748 }
1749 }
1750 }
1751
1752 /* If we have lost the log collector, try to start a new one */
1753 if (SysLoggerPID == 0 && Logging_collector)
1754 SysLoggerPID = SysLogger_Start();
1755
1756 /*
1757 * If no background writer process is running, and we are not in a
1758 * state that prevents it, start one. It doesn't matter if this
1759 * fails, we'll just try again later. Likewise for the checkpointer.
1760 */
1761 if (pmState == PM_RUN || pmState == PM_RECOVERY ||
1762 pmState == PM_HOT_STANDBY)
1763 {
1764 if (CheckpointerPID == 0)
1765 CheckpointerPID = StartCheckpointer();
1766 if (BgWriterPID == 0)
1767 BgWriterPID = StartBackgroundWriter();
1768 }
1769
1770 /*
1771 * Likewise, if we have lost the walwriter process, try to start a new
1772 * one. But this is needed only in normal operation (else we cannot
1773 * be writing any new WAL).
1774 */
1775 if (WalWriterPID == 0 && pmState == PM_RUN)
1776 WalWriterPID = StartWalWriter();
1777
1778 /*
1779 * If we have lost the autovacuum launcher, try to start a new one. We
1780 * don't want autovacuum to run in binary upgrade mode because
1781 * autovacuum might update relfrozenxid for empty tables before the
1782 * physical files are put in place.
1783 */
1784 if (!IsBinaryUpgrade && AutoVacPID == 0 &&
1785 (AutoVacuumingActive() || start_autovac_launcher) &&
1786 pmState == PM_RUN)
1787 {
1788 AutoVacPID = StartAutoVacLauncher();
1789 if (AutoVacPID != 0)
1790 start_autovac_launcher = false; /* signal processed */
1791 }
1792
1793 /* If we have lost the stats collector, try to start a new one */
1794 if (PgStatPID == 0 &&
1795 (pmState == PM_RUN || pmState == PM_HOT_STANDBY))
1796 PgStatPID = pgstat_start();
1797
1798 /* If we have lost the archiver, try to start a new one. */
1799 if (PgArchPID == 0 && PgArchStartupAllowed())
1800 PgArchPID = pgarch_start();
1801
1802 /* If we need to signal the autovacuum launcher, do so now */
1803 if (avlauncher_needs_signal)
1804 {
1805 avlauncher_needs_signal = false;
1806 if (AutoVacPID != 0)
1807 kill(AutoVacPID, SIGUSR2);
1808 }
1809
1810 /* If we need to start a WAL receiver, try to do that now */
1811 if (WalReceiverRequested)
1812 MaybeStartWalReceiver();
1813
1814 /* Get other worker processes running, if needed */
1815 if (StartWorkerNeeded || HaveCrashedWorker)
1816 maybe_start_bgworkers();
1817
1818 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1819
1820 /*
1821 * With assertions enabled, check regularly for appearance of
1822 * additional threads. All builds check at start and exit.
1823 */
1824 Assert(pthread_is_threaded_np() == 0);
1825 #endif
1826
1827 /*
1828 * Lastly, check to see if it's time to do some things that we don't
1829 * want to do every single time through the loop, because they're a
1830 * bit expensive. Note that there's up to a minute of slop in when
1831 * these tasks will be performed, since DetermineSleepTime() will let
1832 * us sleep at most that long; except for SIGKILL timeout which has
1833 * special-case logic there.
1834 */
1835 now = time(NULL);
1836
1837 /*
1838 * If we already sent SIGQUIT to children and they are slow to shut
1839 * down, it's time to send them SIGKILL. This doesn't happen
1840 * normally, but under certain conditions backends can get stuck while
1841 * shutting down. This is a last measure to get them unwedged.
1842 *
1843 * Note we also do this during recovery from a process crash.
1844 */
1845 if ((Shutdown >= ImmediateShutdown || (FatalError && !SendStop)) &&
1846 AbortStartTime != 0 &&
1847 (now - AbortStartTime) >= SIGKILL_CHILDREN_AFTER_SECS)
1848 {
1849 /* We were gentle with them before. Not anymore */
1850 TerminateChildren(SIGKILL);
1851 /* reset flag so we don't SIGKILL again */
1852 AbortStartTime = 0;
1853 }
1854
1855 /*
1856 * Once a minute, verify that postmaster.pid hasn't been removed or
1857 * overwritten. If it has, we force a shutdown. This avoids having
1858 * postmasters and child processes hanging around after their database
1859 * is gone, and maybe causing problems if a new database cluster is
1860 * created in the same place. It also provides some protection
1861 * against a DBA foolishly removing postmaster.pid and manually
1862 * starting a new postmaster. Data corruption is likely to ensue from
1863 * that anyway, but we can minimize the damage by aborting ASAP.
1864 */
1865 if (now - last_lockfile_recheck_time >= 1 * SECS_PER_MINUTE)
1866 {
1867 if (!RecheckDataDirLockFile())
1868 {
1869 ereport(LOG,
1870 (errmsg("performing immediate shutdown because data directory lock file is invalid")));
1871 kill(MyProcPid, SIGQUIT);
1872 }
1873 last_lockfile_recheck_time = now;
1874 }
1875
1876 /*
1877 * Touch Unix socket and lock files every 58 minutes, to ensure that
1878 * they are not removed by overzealous /tmp-cleaning tasks. We assume
1879 * no one runs cleaners with cutoff times of less than an hour ...
1880 */
1881 if (now - last_touch_time >= 58 * SECS_PER_MINUTE)
1882 {
1883 TouchSocketFiles();
1884 TouchSocketLockFiles();
1885 last_touch_time = now;
1886 }
1887 }
1888 }
1889
1890 /*
1891 * Initialise the masks for select() for the ports we are listening on.
1892 * Return the number of sockets to listen on.
1893 */
1894 static int
initMasks(fd_set * rmask)1895 initMasks(fd_set *rmask)
1896 {
1897 int maxsock = -1;
1898 int i;
1899
1900 FD_ZERO(rmask);
1901
1902 for (i = 0; i < MAXLISTEN; i++)
1903 {
1904 int fd = ListenSocket[i];
1905
1906 if (fd == PGINVALID_SOCKET)
1907 break;
1908 FD_SET(fd, rmask);
1909
1910 if (fd > maxsock)
1911 maxsock = fd;
1912 }
1913
1914 return maxsock + 1;
1915 }
1916
1917
1918 /*
1919 * Read a client's startup packet and do something according to it.
1920 *
1921 * Returns STATUS_OK or STATUS_ERROR, or might call ereport(FATAL) and
1922 * not return at all.
1923 *
1924 * (Note that ereport(FATAL) stuff is sent to the client, so only use it
1925 * if that's what you want. Return STATUS_ERROR if you don't want to
1926 * send anything to the client, which would typically be appropriate
1927 * if we detect a communications failure.)
1928 *
1929 * Set ssl_done and/or gss_done when negotiation of an encrypted layer
1930 * (currently, TLS or GSSAPI) is completed. A successful negotiation of either
1931 * encryption layer sets both flags, but a rejected negotiation sets only the
1932 * flag for that layer, since the client may wish to try the other one. We
1933 * should make no assumption here about the order in which the client may make
1934 * requests.
1935 */
1936 static int
ProcessStartupPacket(Port * port,bool ssl_done,bool gss_done)1937 ProcessStartupPacket(Port *port, bool ssl_done, bool gss_done)
1938 {
1939 int32 len;
1940 void *buf;
1941 ProtocolVersion proto;
1942 MemoryContext oldcontext;
1943
1944 pq_startmsgread();
1945
1946 /*
1947 * Grab the first byte of the length word separately, so that we can tell
1948 * whether we have no data at all or an incomplete packet. (This might
1949 * sound inefficient, but it's not really, because of buffering in
1950 * pqcomm.c.)
1951 */
1952 if (pq_getbytes((char *) &len, 1) == EOF)
1953 {
1954 /*
1955 * If we get no data at all, don't clutter the log with a complaint;
1956 * such cases often occur for legitimate reasons. An example is that
1957 * we might be here after responding to NEGOTIATE_SSL_CODE, and if the
1958 * client didn't like our response, it'll probably just drop the
1959 * connection. Service-monitoring software also often just opens and
1960 * closes a connection without sending anything. (So do port
1961 * scanners, which may be less benign, but it's not really our job to
1962 * notice those.)
1963 */
1964 return STATUS_ERROR;
1965 }
1966
1967 if (pq_getbytes(((char *) &len) + 1, 3) == EOF)
1968 {
1969 /* Got a partial length word, so bleat about that */
1970 if (!ssl_done && !gss_done)
1971 ereport(COMMERROR,
1972 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1973 errmsg("incomplete startup packet")));
1974 return STATUS_ERROR;
1975 }
1976
1977 len = pg_ntoh32(len);
1978 len -= 4;
1979
1980 if (len < (int32) sizeof(ProtocolVersion) ||
1981 len > MAX_STARTUP_PACKET_LENGTH)
1982 {
1983 ereport(COMMERROR,
1984 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1985 errmsg("invalid length of startup packet")));
1986 return STATUS_ERROR;
1987 }
1988
1989 /*
1990 * Allocate at least the size of an old-style startup packet, plus one
1991 * extra byte, and make sure all are zeroes. This ensures we will have
1992 * null termination of all strings, in both fixed- and variable-length
1993 * packet layouts.
1994 */
1995 if (len <= (int32) sizeof(StartupPacket))
1996 buf = palloc0(sizeof(StartupPacket) + 1);
1997 else
1998 buf = palloc0(len + 1);
1999
2000 if (pq_getbytes(buf, len) == EOF)
2001 {
2002 ereport(COMMERROR,
2003 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2004 errmsg("incomplete startup packet")));
2005 return STATUS_ERROR;
2006 }
2007 pq_endmsgread();
2008
2009 /*
2010 * The first field is either a protocol version number or a special
2011 * request code.
2012 */
2013 port->proto = proto = pg_ntoh32(*((ProtocolVersion *) buf));
2014
2015 if (proto == CANCEL_REQUEST_CODE)
2016 {
2017 processCancelRequest(port, buf);
2018 /* Not really an error, but we don't want to proceed further */
2019 return STATUS_ERROR;
2020 }
2021
2022 if (proto == NEGOTIATE_SSL_CODE && !ssl_done)
2023 {
2024 char SSLok;
2025
2026 #ifdef USE_SSL
2027 /* No SSL when disabled or on Unix sockets */
2028 if (!LoadedSSL || IS_AF_UNIX(port->laddr.addr.ss_family))
2029 SSLok = 'N';
2030 else
2031 SSLok = 'S'; /* Support for SSL */
2032 #else
2033 SSLok = 'N'; /* No support for SSL */
2034 #endif
2035
2036 retry1:
2037 if (send(port->sock, &SSLok, 1, 0) != 1)
2038 {
2039 if (errno == EINTR)
2040 goto retry1; /* if interrupted, just retry */
2041 ereport(COMMERROR,
2042 (errcode_for_socket_access(),
2043 errmsg("failed to send SSL negotiation response: %m")));
2044 return STATUS_ERROR; /* close the connection */
2045 }
2046
2047 #ifdef USE_SSL
2048 if (SSLok == 'S' && secure_open_server(port) == -1)
2049 return STATUS_ERROR;
2050 #endif
2051
2052 /*
2053 * At this point we should have no data already buffered. If we do,
2054 * it was received before we performed the SSL handshake, so it wasn't
2055 * encrypted and indeed may have been injected by a man-in-the-middle.
2056 * We report this case to the client.
2057 */
2058 if (pq_buffer_has_data())
2059 ereport(FATAL,
2060 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2061 errmsg("received unencrypted data after SSL request"),
2062 errdetail("This could be either a client-software bug or evidence of an attempted man-in-the-middle attack.")));
2063
2064 /*
2065 * regular startup packet, cancel, etc packet should follow, but not
2066 * another SSL negotiation request, and a GSS request should only
2067 * follow if SSL was rejected (client may negotiate in either order)
2068 */
2069 return ProcessStartupPacket(port, true, SSLok == 'S');
2070 }
2071 else if (proto == NEGOTIATE_GSS_CODE && !gss_done)
2072 {
2073 char GSSok = 'N';
2074
2075 #ifdef ENABLE_GSS
2076 /* No GSSAPI encryption when on Unix socket */
2077 if (!IS_AF_UNIX(port->laddr.addr.ss_family))
2078 GSSok = 'G';
2079 #endif
2080
2081 while (send(port->sock, &GSSok, 1, 0) != 1)
2082 {
2083 if (errno == EINTR)
2084 continue;
2085 ereport(COMMERROR,
2086 (errcode_for_socket_access(),
2087 errmsg("failed to send GSSAPI negotiation response: %m")));
2088 return STATUS_ERROR; /* close the connection */
2089 }
2090
2091 #ifdef ENABLE_GSS
2092 if (GSSok == 'G' && secure_open_gssapi(port) == -1)
2093 return STATUS_ERROR;
2094 #endif
2095
2096 /*
2097 * At this point we should have no data already buffered. If we do,
2098 * it was received before we performed the GSS handshake, so it wasn't
2099 * encrypted and indeed may have been injected by a man-in-the-middle.
2100 * We report this case to the client.
2101 */
2102 if (pq_buffer_has_data())
2103 ereport(FATAL,
2104 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2105 errmsg("received unencrypted data after GSSAPI encryption request"),
2106 errdetail("This could be either a client-software bug or evidence of an attempted man-in-the-middle attack.")));
2107
2108 /*
2109 * regular startup packet, cancel, etc packet should follow, but not
2110 * another GSS negotiation request, and an SSL request should only
2111 * follow if GSS was rejected (client may negotiate in either order)
2112 */
2113 return ProcessStartupPacket(port, GSSok == 'G', true);
2114 }
2115
2116 /* Could add additional special packet types here */
2117
2118 /*
2119 * Set FrontendProtocol now so that ereport() knows what format to send if
2120 * we fail during startup.
2121 */
2122 FrontendProtocol = proto;
2123
2124 /* Check that the major protocol version is in range. */
2125 if (PG_PROTOCOL_MAJOR(proto) < PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST) ||
2126 PG_PROTOCOL_MAJOR(proto) > PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST))
2127 ereport(FATAL,
2128 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2129 errmsg("unsupported frontend protocol %u.%u: server supports %u.0 to %u.%u",
2130 PG_PROTOCOL_MAJOR(proto), PG_PROTOCOL_MINOR(proto),
2131 PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST),
2132 PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST),
2133 PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST))));
2134
2135 /*
2136 * Now fetch parameters out of startup packet and save them into the Port
2137 * structure. All data structures attached to the Port struct must be
2138 * allocated in TopMemoryContext so that they will remain available in a
2139 * running backend (even after PostmasterContext is destroyed). We need
2140 * not worry about leaking this storage on failure, since we aren't in the
2141 * postmaster process anymore.
2142 */
2143 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
2144
2145 if (PG_PROTOCOL_MAJOR(proto) >= 3)
2146 {
2147 int32 offset = sizeof(ProtocolVersion);
2148 List *unrecognized_protocol_options = NIL;
2149
2150 /*
2151 * Scan packet body for name/option pairs. We can assume any string
2152 * beginning within the packet body is null-terminated, thanks to
2153 * zeroing extra byte above.
2154 */
2155 port->guc_options = NIL;
2156
2157 while (offset < len)
2158 {
2159 char *nameptr = ((char *) buf) + offset;
2160 int32 valoffset;
2161 char *valptr;
2162
2163 if (*nameptr == '\0')
2164 break; /* found packet terminator */
2165 valoffset = offset + strlen(nameptr) + 1;
2166 if (valoffset >= len)
2167 break; /* missing value, will complain below */
2168 valptr = ((char *) buf) + valoffset;
2169
2170 if (strcmp(nameptr, "database") == 0)
2171 port->database_name = pstrdup(valptr);
2172 else if (strcmp(nameptr, "user") == 0)
2173 port->user_name = pstrdup(valptr);
2174 else if (strcmp(nameptr, "options") == 0)
2175 port->cmdline_options = pstrdup(valptr);
2176 else if (strcmp(nameptr, "replication") == 0)
2177 {
2178 /*
2179 * Due to backward compatibility concerns the replication
2180 * parameter is a hybrid beast which allows the value to be
2181 * either boolean or the string 'database'. The latter
2182 * connects to a specific database which is e.g. required for
2183 * logical decoding while.
2184 */
2185 if (strcmp(valptr, "database") == 0)
2186 {
2187 am_walsender = true;
2188 am_db_walsender = true;
2189 }
2190 else if (!parse_bool(valptr, &am_walsender))
2191 ereport(FATAL,
2192 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
2193 errmsg("invalid value for parameter \"%s\": \"%s\"",
2194 "replication",
2195 valptr),
2196 errhint("Valid values are: \"false\", 0, \"true\", 1, \"database\".")));
2197 }
2198 else if (strncmp(nameptr, "_pq_.", 5) == 0)
2199 {
2200 /*
2201 * Any option beginning with _pq_. is reserved for use as a
2202 * protocol-level option, but at present no such options are
2203 * defined.
2204 */
2205 unrecognized_protocol_options =
2206 lappend(unrecognized_protocol_options, pstrdup(nameptr));
2207 }
2208 else
2209 {
2210 /* Assume it's a generic GUC option */
2211 port->guc_options = lappend(port->guc_options,
2212 pstrdup(nameptr));
2213 port->guc_options = lappend(port->guc_options,
2214 pstrdup(valptr));
2215
2216 /*
2217 * Copy application_name to port if we come across it. This
2218 * is done so we can log the application_name in the
2219 * connection authorization message. Note that the GUC would
2220 * be used but we haven't gone through GUC setup yet.
2221 */
2222 if (strcmp(nameptr, "application_name") == 0)
2223 {
2224 char *tmp_app_name = pstrdup(valptr);
2225
2226 pg_clean_ascii(tmp_app_name);
2227
2228 port->application_name = tmp_app_name;
2229 }
2230 }
2231 offset = valoffset + strlen(valptr) + 1;
2232 }
2233
2234 /*
2235 * If we didn't find a packet terminator exactly at the end of the
2236 * given packet length, complain.
2237 */
2238 if (offset != len - 1)
2239 ereport(FATAL,
2240 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2241 errmsg("invalid startup packet layout: expected terminator as last byte")));
2242
2243 /*
2244 * If the client requested a newer protocol version or if the client
2245 * requested any protocol options we didn't recognize, let them know
2246 * the newest minor protocol version we do support and the names of
2247 * any unrecognized options.
2248 */
2249 if (PG_PROTOCOL_MINOR(proto) > PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST) ||
2250 unrecognized_protocol_options != NIL)
2251 SendNegotiateProtocolVersion(unrecognized_protocol_options);
2252 }
2253 else
2254 {
2255 /*
2256 * Get the parameters from the old-style, fixed-width-fields startup
2257 * packet as C strings. The packet destination was cleared first so a
2258 * short packet has zeros silently added. We have to be prepared to
2259 * truncate the pstrdup result for oversize fields, though.
2260 */
2261 StartupPacket *packet = (StartupPacket *) buf;
2262
2263 port->database_name = pstrdup(packet->database);
2264 if (strlen(port->database_name) > sizeof(packet->database))
2265 port->database_name[sizeof(packet->database)] = '\0';
2266 port->user_name = pstrdup(packet->user);
2267 if (strlen(port->user_name) > sizeof(packet->user))
2268 port->user_name[sizeof(packet->user)] = '\0';
2269 port->cmdline_options = pstrdup(packet->options);
2270 if (strlen(port->cmdline_options) > sizeof(packet->options))
2271 port->cmdline_options[sizeof(packet->options)] = '\0';
2272 port->guc_options = NIL;
2273 }
2274
2275 /* Check a user name was given. */
2276 if (port->user_name == NULL || port->user_name[0] == '\0')
2277 ereport(FATAL,
2278 (errcode(ERRCODE_INVALID_AUTHORIZATION_SPECIFICATION),
2279 errmsg("no PostgreSQL user name specified in startup packet")));
2280
2281 /* The database defaults to the user name. */
2282 if (port->database_name == NULL || port->database_name[0] == '\0')
2283 port->database_name = pstrdup(port->user_name);
2284
2285 if (Db_user_namespace)
2286 {
2287 /*
2288 * If user@, it is a global user, remove '@'. We only want to do this
2289 * if there is an '@' at the end and no earlier in the user string or
2290 * they may fake as a local user of another database attaching to this
2291 * database.
2292 */
2293 if (strchr(port->user_name, '@') ==
2294 port->user_name + strlen(port->user_name) - 1)
2295 *strchr(port->user_name, '@') = '\0';
2296 else
2297 {
2298 /* Append '@' and dbname */
2299 port->user_name = psprintf("%s@%s", port->user_name, port->database_name);
2300 }
2301 }
2302
2303 /*
2304 * Truncate given database and user names to length of a Postgres name.
2305 * This avoids lookup failures when overlength names are given.
2306 */
2307 if (strlen(port->database_name) >= NAMEDATALEN)
2308 port->database_name[NAMEDATALEN - 1] = '\0';
2309 if (strlen(port->user_name) >= NAMEDATALEN)
2310 port->user_name[NAMEDATALEN - 1] = '\0';
2311
2312 if (am_walsender)
2313 MyBackendType = B_WAL_SENDER;
2314 else
2315 MyBackendType = B_BACKEND;
2316
2317 /*
2318 * Normal walsender backends, e.g. for streaming replication, are not
2319 * connected to a particular database. But walsenders used for logical
2320 * replication need to connect to a specific database. We allow streaming
2321 * replication commands to be issued even if connected to a database as it
2322 * can make sense to first make a basebackup and then stream changes
2323 * starting from that.
2324 */
2325 if (am_walsender && !am_db_walsender)
2326 port->database_name[0] = '\0';
2327
2328 /*
2329 * Done putting stuff in TopMemoryContext.
2330 */
2331 MemoryContextSwitchTo(oldcontext);
2332
2333 /*
2334 * If we're going to reject the connection due to database state, say so
2335 * now instead of wasting cycles on an authentication exchange. (This also
2336 * allows a pg_ping utility to be written.)
2337 */
2338 switch (port->canAcceptConnections)
2339 {
2340 case CAC_STARTUP:
2341 ereport(FATAL,
2342 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2343 errmsg("the database system is starting up")));
2344 break;
2345 case CAC_SHUTDOWN:
2346 ereport(FATAL,
2347 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2348 errmsg("the database system is shutting down")));
2349 break;
2350 case CAC_RECOVERY:
2351 ereport(FATAL,
2352 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2353 errmsg("the database system is in recovery mode")));
2354 break;
2355 case CAC_TOOMANY:
2356 ereport(FATAL,
2357 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
2358 errmsg("sorry, too many clients already")));
2359 break;
2360 case CAC_SUPERUSER:
2361 /* OK for now, will check in InitPostgres */
2362 break;
2363 case CAC_OK:
2364 break;
2365 }
2366
2367 return STATUS_OK;
2368 }
2369
2370 /*
2371 * Send a NegotiateProtocolVersion to the client. This lets the client know
2372 * that they have requested a newer minor protocol version than we are able
2373 * to speak. We'll speak the highest version we know about; the client can,
2374 * of course, abandon the connection if that's a problem.
2375 *
2376 * We also include in the response a list of protocol options we didn't
2377 * understand. This allows clients to include optional parameters that might
2378 * be present either in newer protocol versions or third-party protocol
2379 * extensions without fear of having to reconnect if those options are not
2380 * understood, while at the same time making certain that the client is aware
2381 * of which options were actually accepted.
2382 */
2383 static void
SendNegotiateProtocolVersion(List * unrecognized_protocol_options)2384 SendNegotiateProtocolVersion(List *unrecognized_protocol_options)
2385 {
2386 StringInfoData buf;
2387 ListCell *lc;
2388
2389 pq_beginmessage(&buf, 'v'); /* NegotiateProtocolVersion */
2390 pq_sendint32(&buf, PG_PROTOCOL_LATEST);
2391 pq_sendint32(&buf, list_length(unrecognized_protocol_options));
2392 foreach(lc, unrecognized_protocol_options)
2393 pq_sendstring(&buf, lfirst(lc));
2394 pq_endmessage(&buf);
2395
2396 /* no need to flush, some other message will follow */
2397 }
2398
2399 /*
2400 * The client has sent a cancel request packet, not a normal
2401 * start-a-new-connection packet. Perform the necessary processing.
2402 * Nothing is sent back to the client.
2403 */
2404 static void
processCancelRequest(Port * port,void * pkt)2405 processCancelRequest(Port *port, void *pkt)
2406 {
2407 CancelRequestPacket *canc = (CancelRequestPacket *) pkt;
2408 int backendPID;
2409 int32 cancelAuthCode;
2410 Backend *bp;
2411
2412 #ifndef EXEC_BACKEND
2413 dlist_iter iter;
2414 #else
2415 int i;
2416 #endif
2417
2418 backendPID = (int) pg_ntoh32(canc->backendPID);
2419 cancelAuthCode = (int32) pg_ntoh32(canc->cancelAuthCode);
2420
2421 /*
2422 * See if we have a matching backend. In the EXEC_BACKEND case, we can no
2423 * longer access the postmaster's own backend list, and must rely on the
2424 * duplicate array in shared memory.
2425 */
2426 #ifndef EXEC_BACKEND
2427 dlist_foreach(iter, &BackendList)
2428 {
2429 bp = dlist_container(Backend, elem, iter.cur);
2430 #else
2431 for (i = MaxLivePostmasterChildren() - 1; i >= 0; i--)
2432 {
2433 bp = (Backend *) &ShmemBackendArray[i];
2434 #endif
2435 if (bp->pid == backendPID)
2436 {
2437 if (bp->cancel_key == cancelAuthCode)
2438 {
2439 /* Found a match; signal that backend to cancel current op */
2440 ereport(DEBUG2,
2441 (errmsg_internal("processing cancel request: sending SIGINT to process %d",
2442 backendPID)));
2443 signal_child(bp->pid, SIGINT);
2444 }
2445 else
2446 /* Right PID, wrong key: no way, Jose */
2447 ereport(LOG,
2448 (errmsg("wrong key in cancel request for process %d",
2449 backendPID)));
2450 return;
2451 }
2452 #ifndef EXEC_BACKEND /* make GNU Emacs 26.1 see brace balance */
2453 }
2454 #else
2455 }
2456 #endif
2457
2458 /* No matching backend */
2459 ereport(LOG,
2460 (errmsg("PID %d in cancel request did not match any process",
2461 backendPID)));
2462 }
2463
2464 /*
2465 * canAcceptConnections --- check to see if database state allows connections
2466 * of the specified type. backend_type can be BACKEND_TYPE_NORMAL,
2467 * BACKEND_TYPE_AUTOVAC, or BACKEND_TYPE_BGWORKER. (Note that we don't yet
2468 * know whether a NORMAL connection might turn into a walsender.)
2469 */
2470 static CAC_state
canAcceptConnections(int backend_type)2471 canAcceptConnections(int backend_type)
2472 {
2473 CAC_state result = CAC_OK;
2474
2475 /*
2476 * Can't start backends when in startup/shutdown/inconsistent recovery
2477 * state. We treat autovac workers the same as user backends for this
2478 * purpose. However, bgworkers are excluded from this test; we expect
2479 * bgworker_should_start_now() decided whether the DB state allows them.
2480 */
2481 if (pmState != PM_RUN && pmState != PM_HOT_STANDBY &&
2482 backend_type != BACKEND_TYPE_BGWORKER)
2483 {
2484 if (Shutdown > NoShutdown)
2485 return CAC_SHUTDOWN; /* shutdown is pending */
2486 else if (!FatalError &&
2487 (pmState == PM_STARTUP ||
2488 pmState == PM_RECOVERY))
2489 return CAC_STARTUP; /* normal startup */
2490 else
2491 return CAC_RECOVERY; /* else must be crash recovery */
2492 }
2493
2494 /*
2495 * "Smart shutdown" restrictions are applied only to normal connections,
2496 * not to autovac workers or bgworkers. When only superusers can connect,
2497 * we return CAC_SUPERUSER to indicate that superuserness must be checked
2498 * later. Note that neither CAC_OK nor CAC_SUPERUSER can safely be
2499 * returned until we have checked for too many children.
2500 */
2501 if (connsAllowed != ALLOW_ALL_CONNS &&
2502 backend_type == BACKEND_TYPE_NORMAL)
2503 {
2504 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
2505 result = CAC_SUPERUSER; /* allow superusers only */
2506 else
2507 return CAC_SHUTDOWN; /* shutdown is pending */
2508 }
2509
2510 /*
2511 * Don't start too many children.
2512 *
2513 * We allow more connections here than we can have backends because some
2514 * might still be authenticating; they might fail auth, or some existing
2515 * backend might exit before the auth cycle is completed. The exact
2516 * MaxBackends limit is enforced when a new backend tries to join the
2517 * shared-inval backend array.
2518 *
2519 * The limit here must match the sizes of the per-child-process arrays;
2520 * see comments for MaxLivePostmasterChildren().
2521 */
2522 if (CountChildren(BACKEND_TYPE_ALL) >= MaxLivePostmasterChildren())
2523 result = CAC_TOOMANY;
2524
2525 return result;
2526 }
2527
2528
2529 /*
2530 * ConnCreate -- create a local connection data structure
2531 *
2532 * Returns NULL on failure, other than out-of-memory which is fatal.
2533 */
2534 static Port *
ConnCreate(int serverFd)2535 ConnCreate(int serverFd)
2536 {
2537 Port *port;
2538
2539 if (!(port = (Port *) calloc(1, sizeof(Port))))
2540 {
2541 ereport(LOG,
2542 (errcode(ERRCODE_OUT_OF_MEMORY),
2543 errmsg("out of memory")));
2544 ExitPostmaster(1);
2545 }
2546
2547 if (StreamConnection(serverFd, port) != STATUS_OK)
2548 {
2549 if (port->sock != PGINVALID_SOCKET)
2550 StreamClose(port->sock);
2551 ConnFree(port);
2552 return NULL;
2553 }
2554
2555 return port;
2556 }
2557
2558
2559 /*
2560 * ConnFree -- free a local connection data structure
2561 *
2562 * Caller has already closed the socket if any, so there's not much
2563 * to do here.
2564 */
2565 static void
ConnFree(Port * conn)2566 ConnFree(Port *conn)
2567 {
2568 free(conn);
2569 }
2570
2571
2572 /*
2573 * ClosePostmasterPorts -- close all the postmaster's open sockets
2574 *
2575 * This is called during child process startup to release file descriptors
2576 * that are not needed by that child process. The postmaster still has
2577 * them open, of course.
2578 *
2579 * Note: we pass am_syslogger as a boolean because we don't want to set
2580 * the global variable yet when this is called.
2581 */
2582 void
ClosePostmasterPorts(bool am_syslogger)2583 ClosePostmasterPorts(bool am_syslogger)
2584 {
2585 int i;
2586
2587 #ifndef WIN32
2588
2589 /*
2590 * Close the write end of postmaster death watch pipe. It's important to
2591 * do this as early as possible, so that if postmaster dies, others won't
2592 * think that it's still running because we're holding the pipe open.
2593 */
2594 if (close(postmaster_alive_fds[POSTMASTER_FD_OWN]) != 0)
2595 ereport(FATAL,
2596 (errcode_for_file_access(),
2597 errmsg_internal("could not close postmaster death monitoring pipe in child process: %m")));
2598 postmaster_alive_fds[POSTMASTER_FD_OWN] = -1;
2599 /* Notify fd.c that we released one pipe FD. */
2600 ReleaseExternalFD();
2601 #endif
2602
2603 /*
2604 * Close the postmaster's listen sockets. These aren't tracked by fd.c,
2605 * so we don't call ReleaseExternalFD() here.
2606 */
2607 for (i = 0; i < MAXLISTEN; i++)
2608 {
2609 if (ListenSocket[i] != PGINVALID_SOCKET)
2610 {
2611 StreamClose(ListenSocket[i]);
2612 ListenSocket[i] = PGINVALID_SOCKET;
2613 }
2614 }
2615
2616 /*
2617 * If using syslogger, close the read side of the pipe. We don't bother
2618 * tracking this in fd.c, either.
2619 */
2620 if (!am_syslogger)
2621 {
2622 #ifndef WIN32
2623 if (syslogPipe[0] >= 0)
2624 close(syslogPipe[0]);
2625 syslogPipe[0] = -1;
2626 #else
2627 if (syslogPipe[0])
2628 CloseHandle(syslogPipe[0]);
2629 syslogPipe[0] = 0;
2630 #endif
2631 }
2632
2633 #ifdef USE_BONJOUR
2634 /* If using Bonjour, close the connection to the mDNS daemon */
2635 if (bonjour_sdref)
2636 close(DNSServiceRefSockFD(bonjour_sdref));
2637 #endif
2638 }
2639
2640
2641 /*
2642 * InitProcessGlobals -- set MyProcPid, MyStartTime[stamp], random seeds
2643 *
2644 * Called early in the postmaster and every backend.
2645 */
2646 void
InitProcessGlobals(void)2647 InitProcessGlobals(void)
2648 {
2649 unsigned int rseed;
2650
2651 MyProcPid = getpid();
2652 MyStartTimestamp = GetCurrentTimestamp();
2653 MyStartTime = timestamptz_to_time_t(MyStartTimestamp);
2654
2655 /*
2656 * Set a different seed for random() in every process. We want something
2657 * unpredictable, so if possible, use high-quality random bits for the
2658 * seed. Otherwise, fall back to a seed based on timestamp and PID.
2659 */
2660 if (!pg_strong_random(&rseed, sizeof(rseed)))
2661 {
2662 /*
2663 * Since PIDs and timestamps tend to change more frequently in their
2664 * least significant bits, shift the timestamp left to allow a larger
2665 * total number of seeds in a given time period. Since that would
2666 * leave only 20 bits of the timestamp that cycle every ~1 second,
2667 * also mix in some higher bits.
2668 */
2669 rseed = ((uint64) MyProcPid) ^
2670 ((uint64) MyStartTimestamp << 12) ^
2671 ((uint64) MyStartTimestamp >> 20);
2672 }
2673 srandom(rseed);
2674 }
2675
2676
2677 /*
2678 * reset_shared -- reset shared memory and semaphores
2679 */
2680 static void
reset_shared(void)2681 reset_shared(void)
2682 {
2683 /*
2684 * Create or re-create shared memory and semaphores.
2685 *
2686 * Note: in each "cycle of life" we will normally assign the same IPC keys
2687 * (if using SysV shmem and/or semas). This helps ensure that we will
2688 * clean up dead IPC objects if the postmaster crashes and is restarted.
2689 */
2690 CreateSharedMemoryAndSemaphores();
2691 }
2692
2693
2694 /*
2695 * SIGHUP -- reread config files, and tell children to do same
2696 */
2697 static void
SIGHUP_handler(SIGNAL_ARGS)2698 SIGHUP_handler(SIGNAL_ARGS)
2699 {
2700 int save_errno = errno;
2701
2702 /*
2703 * We rely on the signal mechanism to have blocked all signals ... except
2704 * on Windows, which lacks sigaction(), so we have to do it manually.
2705 */
2706 #ifdef WIN32
2707 PG_SETMASK(&BlockSig);
2708 #endif
2709
2710 if (Shutdown <= SmartShutdown)
2711 {
2712 ereport(LOG,
2713 (errmsg("received SIGHUP, reloading configuration files")));
2714 ProcessConfigFile(PGC_SIGHUP);
2715 SignalChildren(SIGHUP);
2716 if (StartupPID != 0)
2717 signal_child(StartupPID, SIGHUP);
2718 if (BgWriterPID != 0)
2719 signal_child(BgWriterPID, SIGHUP);
2720 if (CheckpointerPID != 0)
2721 signal_child(CheckpointerPID, SIGHUP);
2722 if (WalWriterPID != 0)
2723 signal_child(WalWriterPID, SIGHUP);
2724 if (WalReceiverPID != 0)
2725 signal_child(WalReceiverPID, SIGHUP);
2726 if (AutoVacPID != 0)
2727 signal_child(AutoVacPID, SIGHUP);
2728 if (PgArchPID != 0)
2729 signal_child(PgArchPID, SIGHUP);
2730 if (SysLoggerPID != 0)
2731 signal_child(SysLoggerPID, SIGHUP);
2732 if (PgStatPID != 0)
2733 signal_child(PgStatPID, SIGHUP);
2734
2735 /* Reload authentication config files too */
2736 if (!load_hba())
2737 ereport(LOG,
2738 /* translator: %s is a configuration file */
2739 (errmsg("%s was not reloaded", "pg_hba.conf")));
2740
2741 if (!load_ident())
2742 ereport(LOG,
2743 (errmsg("%s was not reloaded", "pg_ident.conf")));
2744
2745 #ifdef USE_SSL
2746 /* Reload SSL configuration as well */
2747 if (EnableSSL)
2748 {
2749 if (secure_initialize(false) == 0)
2750 LoadedSSL = true;
2751 else
2752 ereport(LOG,
2753 (errmsg("SSL configuration was not reloaded")));
2754 }
2755 else
2756 {
2757 secure_destroy();
2758 LoadedSSL = false;
2759 }
2760 #endif
2761
2762 #ifdef EXEC_BACKEND
2763 /* Update the starting-point file for future children */
2764 write_nondefault_variables(PGC_SIGHUP);
2765 #endif
2766 }
2767
2768 #ifdef WIN32
2769 PG_SETMASK(&UnBlockSig);
2770 #endif
2771
2772 errno = save_errno;
2773 }
2774
2775
2776 /*
2777 * pmdie -- signal handler for processing various postmaster signals.
2778 */
2779 static void
pmdie(SIGNAL_ARGS)2780 pmdie(SIGNAL_ARGS)
2781 {
2782 int save_errno = errno;
2783
2784 /*
2785 * We rely on the signal mechanism to have blocked all signals ... except
2786 * on Windows, which lacks sigaction(), so we have to do it manually.
2787 */
2788 #ifdef WIN32
2789 PG_SETMASK(&BlockSig);
2790 #endif
2791
2792 ereport(DEBUG2,
2793 (errmsg_internal("postmaster received signal %d",
2794 postgres_signal_arg)));
2795
2796 switch (postgres_signal_arg)
2797 {
2798 case SIGTERM:
2799
2800 /*
2801 * Smart Shutdown:
2802 *
2803 * Wait for children to end their work, then shut down.
2804 */
2805 if (Shutdown >= SmartShutdown)
2806 break;
2807 Shutdown = SmartShutdown;
2808 ereport(LOG,
2809 (errmsg("received smart shutdown request")));
2810
2811 /* Report status */
2812 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2813 #ifdef USE_SYSTEMD
2814 sd_notify(0, "STOPPING=1");
2815 #endif
2816
2817 /*
2818 * If we reached normal running, we have to wait for any online
2819 * backup mode to end; otherwise go straight to waiting for client
2820 * backends to exit. (The difference is that in the former state,
2821 * we'll still let in new superuser clients, so that somebody can
2822 * end the online backup mode.) If already in PM_STOP_BACKENDS or
2823 * a later state, do not change it.
2824 */
2825 if (pmState == PM_RUN)
2826 connsAllowed = ALLOW_SUPERUSER_CONNS;
2827 else if (pmState == PM_HOT_STANDBY)
2828 connsAllowed = ALLOW_NO_CONNS;
2829 else if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2830 {
2831 /* There should be no clients, so proceed to stop children */
2832 pmState = PM_STOP_BACKENDS;
2833 }
2834
2835 /*
2836 * Now wait for online backup mode to end and backends to exit. If
2837 * that is already the case, PostmasterStateMachine will take the
2838 * next step.
2839 */
2840 PostmasterStateMachine();
2841 break;
2842
2843 case SIGINT:
2844
2845 /*
2846 * Fast Shutdown:
2847 *
2848 * Abort all children with SIGTERM (rollback active transactions
2849 * and exit) and shut down when they are gone.
2850 */
2851 if (Shutdown >= FastShutdown)
2852 break;
2853 Shutdown = FastShutdown;
2854 ereport(LOG,
2855 (errmsg("received fast shutdown request")));
2856
2857 /* Report status */
2858 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2859 #ifdef USE_SYSTEMD
2860 sd_notify(0, "STOPPING=1");
2861 #endif
2862
2863 if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2864 {
2865 /* Just shut down background processes silently */
2866 pmState = PM_STOP_BACKENDS;
2867 }
2868 else if (pmState == PM_RUN ||
2869 pmState == PM_HOT_STANDBY)
2870 {
2871 /* Report that we're about to zap live client sessions */
2872 ereport(LOG,
2873 (errmsg("aborting any active transactions")));
2874 pmState = PM_STOP_BACKENDS;
2875 }
2876
2877 /*
2878 * PostmasterStateMachine will issue any necessary signals, or
2879 * take the next step if no child processes need to be killed.
2880 */
2881 PostmasterStateMachine();
2882 break;
2883
2884 case SIGQUIT:
2885
2886 /*
2887 * Immediate Shutdown:
2888 *
2889 * abort all children with SIGQUIT, wait for them to exit,
2890 * terminate remaining ones with SIGKILL, then exit without
2891 * attempt to properly shut down the data base system.
2892 */
2893 if (Shutdown >= ImmediateShutdown)
2894 break;
2895 Shutdown = ImmediateShutdown;
2896 ereport(LOG,
2897 (errmsg("received immediate shutdown request")));
2898
2899 /* Report status */
2900 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2901 #ifdef USE_SYSTEMD
2902 sd_notify(0, "STOPPING=1");
2903 #endif
2904
2905 TerminateChildren(SIGQUIT);
2906 pmState = PM_WAIT_BACKENDS;
2907
2908 /* set stopwatch for them to die */
2909 AbortStartTime = time(NULL);
2910
2911 /*
2912 * Now wait for backends to exit. If there are none,
2913 * PostmasterStateMachine will take the next step.
2914 */
2915 PostmasterStateMachine();
2916 break;
2917 }
2918
2919 #ifdef WIN32
2920 PG_SETMASK(&UnBlockSig);
2921 #endif
2922
2923 errno = save_errno;
2924 }
2925
2926 /*
2927 * Reaper -- signal handler to cleanup after a child process dies.
2928 */
2929 static void
reaper(SIGNAL_ARGS)2930 reaper(SIGNAL_ARGS)
2931 {
2932 int save_errno = errno;
2933 int pid; /* process id of dead child process */
2934 int exitstatus; /* its exit status */
2935
2936 /*
2937 * We rely on the signal mechanism to have blocked all signals ... except
2938 * on Windows, which lacks sigaction(), so we have to do it manually.
2939 */
2940 #ifdef WIN32
2941 PG_SETMASK(&BlockSig);
2942 #endif
2943
2944 ereport(DEBUG4,
2945 (errmsg_internal("reaping dead processes")));
2946
2947 while ((pid = waitpid(-1, &exitstatus, WNOHANG)) > 0)
2948 {
2949 /*
2950 * Check if this child was a startup process.
2951 */
2952 if (pid == StartupPID)
2953 {
2954 StartupPID = 0;
2955
2956 /*
2957 * Startup process exited in response to a shutdown request (or it
2958 * completed normally regardless of the shutdown request).
2959 */
2960 if (Shutdown > NoShutdown &&
2961 (EXIT_STATUS_0(exitstatus) || EXIT_STATUS_1(exitstatus)))
2962 {
2963 StartupStatus = STARTUP_NOT_RUNNING;
2964 pmState = PM_WAIT_BACKENDS;
2965 /* PostmasterStateMachine logic does the rest */
2966 continue;
2967 }
2968
2969 if (EXIT_STATUS_3(exitstatus))
2970 {
2971 ereport(LOG,
2972 (errmsg("shutdown at recovery target")));
2973 StartupStatus = STARTUP_NOT_RUNNING;
2974 Shutdown = Max(Shutdown, SmartShutdown);
2975 TerminateChildren(SIGTERM);
2976 pmState = PM_WAIT_BACKENDS;
2977 /* PostmasterStateMachine logic does the rest */
2978 continue;
2979 }
2980
2981 /*
2982 * Unexpected exit of startup process (including FATAL exit)
2983 * during PM_STARTUP is treated as catastrophic. There are no
2984 * other processes running yet, so we can just exit.
2985 */
2986 if (pmState == PM_STARTUP &&
2987 StartupStatus != STARTUP_SIGNALED &&
2988 !EXIT_STATUS_0(exitstatus))
2989 {
2990 LogChildExit(LOG, _("startup process"),
2991 pid, exitstatus);
2992 ereport(LOG,
2993 (errmsg("aborting startup due to startup process failure")));
2994 ExitPostmaster(1);
2995 }
2996
2997 /*
2998 * After PM_STARTUP, any unexpected exit (including FATAL exit) of
2999 * the startup process is catastrophic, so kill other children,
3000 * and set StartupStatus so we don't try to reinitialize after
3001 * they're gone. Exception: if StartupStatus is STARTUP_SIGNALED,
3002 * then we previously sent the startup process a SIGQUIT; so
3003 * that's probably the reason it died, and we do want to try to
3004 * restart in that case.
3005 *
3006 * This stanza also handles the case where we sent a SIGQUIT
3007 * during PM_STARTUP due to some dead_end child crashing: in that
3008 * situation, if the startup process dies on the SIGQUIT, we need
3009 * to transition to PM_WAIT_BACKENDS state which will allow
3010 * PostmasterStateMachine to restart the startup process. (On the
3011 * other hand, the startup process might complete normally, if we
3012 * were too late with the SIGQUIT. In that case we'll fall
3013 * through and commence normal operations.)
3014 */
3015 if (!EXIT_STATUS_0(exitstatus))
3016 {
3017 if (StartupStatus == STARTUP_SIGNALED)
3018 {
3019 StartupStatus = STARTUP_NOT_RUNNING;
3020 if (pmState == PM_STARTUP)
3021 pmState = PM_WAIT_BACKENDS;
3022 }
3023 else
3024 StartupStatus = STARTUP_CRASHED;
3025 HandleChildCrash(pid, exitstatus,
3026 _("startup process"));
3027 continue;
3028 }
3029
3030 /*
3031 * Startup succeeded, commence normal operations
3032 */
3033 StartupStatus = STARTUP_NOT_RUNNING;
3034 FatalError = false;
3035 AbortStartTime = 0;
3036 ReachedNormalRunning = true;
3037 pmState = PM_RUN;
3038 connsAllowed = ALLOW_ALL_CONNS;
3039
3040 /*
3041 * Crank up the background tasks, if we didn't do that already
3042 * when we entered consistent recovery state. It doesn't matter
3043 * if this fails, we'll just try again later.
3044 */
3045 if (CheckpointerPID == 0)
3046 CheckpointerPID = StartCheckpointer();
3047 if (BgWriterPID == 0)
3048 BgWriterPID = StartBackgroundWriter();
3049 if (WalWriterPID == 0)
3050 WalWriterPID = StartWalWriter();
3051
3052 /*
3053 * Likewise, start other special children as needed. In a restart
3054 * situation, some of them may be alive already.
3055 */
3056 if (!IsBinaryUpgrade && AutoVacuumingActive() && AutoVacPID == 0)
3057 AutoVacPID = StartAutoVacLauncher();
3058 if (PgArchStartupAllowed() && PgArchPID == 0)
3059 PgArchPID = pgarch_start();
3060 if (PgStatPID == 0)
3061 PgStatPID = pgstat_start();
3062
3063 /* workers may be scheduled to start now */
3064 maybe_start_bgworkers();
3065
3066 /* at this point we are really open for business */
3067 ereport(LOG,
3068 (errmsg("database system is ready to accept connections")));
3069
3070 /* Report status */
3071 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
3072 #ifdef USE_SYSTEMD
3073 sd_notify(0, "READY=1");
3074 #endif
3075
3076 continue;
3077 }
3078
3079 /*
3080 * Was it the bgwriter? Normal exit can be ignored; we'll start a new
3081 * one at the next iteration of the postmaster's main loop, if
3082 * necessary. Any other exit condition is treated as a crash.
3083 */
3084 if (pid == BgWriterPID)
3085 {
3086 BgWriterPID = 0;
3087 if (!EXIT_STATUS_0(exitstatus))
3088 HandleChildCrash(pid, exitstatus,
3089 _("background writer process"));
3090 continue;
3091 }
3092
3093 /*
3094 * Was it the checkpointer?
3095 */
3096 if (pid == CheckpointerPID)
3097 {
3098 CheckpointerPID = 0;
3099 if (EXIT_STATUS_0(exitstatus) && pmState == PM_SHUTDOWN)
3100 {
3101 /*
3102 * OK, we saw normal exit of the checkpointer after it's been
3103 * told to shut down. We expect that it wrote a shutdown
3104 * checkpoint. (If for some reason it didn't, recovery will
3105 * occur on next postmaster start.)
3106 *
3107 * At this point we should have no normal backend children
3108 * left (else we'd not be in PM_SHUTDOWN state) but we might
3109 * have dead_end children to wait for.
3110 *
3111 * If we have an archiver subprocess, tell it to do a last
3112 * archive cycle and quit. Likewise, if we have walsender
3113 * processes, tell them to send any remaining WAL and quit.
3114 */
3115 Assert(Shutdown > NoShutdown);
3116
3117 /* Waken archiver for the last time */
3118 if (PgArchPID != 0)
3119 signal_child(PgArchPID, SIGUSR2);
3120
3121 /*
3122 * Waken walsenders for the last time. No regular backends
3123 * should be around anymore.
3124 */
3125 SignalChildren(SIGUSR2);
3126
3127 pmState = PM_SHUTDOWN_2;
3128
3129 /*
3130 * We can also shut down the stats collector now; there's
3131 * nothing left for it to do.
3132 */
3133 if (PgStatPID != 0)
3134 signal_child(PgStatPID, SIGQUIT);
3135 }
3136 else
3137 {
3138 /*
3139 * Any unexpected exit of the checkpointer (including FATAL
3140 * exit) is treated as a crash.
3141 */
3142 HandleChildCrash(pid, exitstatus,
3143 _("checkpointer process"));
3144 }
3145
3146 continue;
3147 }
3148
3149 /*
3150 * Was it the wal writer? Normal exit can be ignored; we'll start a
3151 * new one at the next iteration of the postmaster's main loop, if
3152 * necessary. Any other exit condition is treated as a crash.
3153 */
3154 if (pid == WalWriterPID)
3155 {
3156 WalWriterPID = 0;
3157 if (!EXIT_STATUS_0(exitstatus))
3158 HandleChildCrash(pid, exitstatus,
3159 _("WAL writer process"));
3160 continue;
3161 }
3162
3163 /*
3164 * Was it the wal receiver? If exit status is zero (normal) or one
3165 * (FATAL exit), we assume everything is all right just like normal
3166 * backends. (If we need a new wal receiver, we'll start one at the
3167 * next iteration of the postmaster's main loop.)
3168 */
3169 if (pid == WalReceiverPID)
3170 {
3171 WalReceiverPID = 0;
3172 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3173 HandleChildCrash(pid, exitstatus,
3174 _("WAL receiver process"));
3175 continue;
3176 }
3177
3178 /*
3179 * Was it the autovacuum launcher? Normal exit can be ignored; we'll
3180 * start a new one at the next iteration of the postmaster's main
3181 * loop, if necessary. Any other exit condition is treated as a
3182 * crash.
3183 */
3184 if (pid == AutoVacPID)
3185 {
3186 AutoVacPID = 0;
3187 if (!EXIT_STATUS_0(exitstatus))
3188 HandleChildCrash(pid, exitstatus,
3189 _("autovacuum launcher process"));
3190 continue;
3191 }
3192
3193 /*
3194 * Was it the archiver? If so, just try to start a new one; no need
3195 * to force reset of the rest of the system. (If fail, we'll try
3196 * again in future cycles of the main loop.). Unless we were waiting
3197 * for it to shut down; don't restart it in that case, and
3198 * PostmasterStateMachine() will advance to the next shutdown step.
3199 */
3200 if (pid == PgArchPID)
3201 {
3202 PgArchPID = 0;
3203 if (!EXIT_STATUS_0(exitstatus))
3204 LogChildExit(LOG, _("archiver process"),
3205 pid, exitstatus);
3206 if (PgArchStartupAllowed())
3207 PgArchPID = pgarch_start();
3208 continue;
3209 }
3210
3211 /*
3212 * Was it the statistics collector? If so, just try to start a new
3213 * one; no need to force reset of the rest of the system. (If fail,
3214 * we'll try again in future cycles of the main loop.)
3215 */
3216 if (pid == PgStatPID)
3217 {
3218 PgStatPID = 0;
3219 if (!EXIT_STATUS_0(exitstatus))
3220 LogChildExit(LOG, _("statistics collector process"),
3221 pid, exitstatus);
3222 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3223 PgStatPID = pgstat_start();
3224 continue;
3225 }
3226
3227 /* Was it the system logger? If so, try to start a new one */
3228 if (pid == SysLoggerPID)
3229 {
3230 SysLoggerPID = 0;
3231 /* for safety's sake, launch new logger *first* */
3232 SysLoggerPID = SysLogger_Start();
3233 if (!EXIT_STATUS_0(exitstatus))
3234 LogChildExit(LOG, _("system logger process"),
3235 pid, exitstatus);
3236 continue;
3237 }
3238
3239 /* Was it one of our background workers? */
3240 if (CleanupBackgroundWorker(pid, exitstatus))
3241 {
3242 /* have it be restarted */
3243 HaveCrashedWorker = true;
3244 continue;
3245 }
3246
3247 /*
3248 * Else do standard backend child cleanup.
3249 */
3250 CleanupBackend(pid, exitstatus);
3251 } /* loop over pending child-death reports */
3252
3253 /*
3254 * After cleaning out the SIGCHLD queue, see if we have any state changes
3255 * or actions to make.
3256 */
3257 PostmasterStateMachine();
3258
3259 /* Done with signal handler */
3260 #ifdef WIN32
3261 PG_SETMASK(&UnBlockSig);
3262 #endif
3263
3264 errno = save_errno;
3265 }
3266
3267 /*
3268 * Scan the bgworkers list and see if the given PID (which has just stopped
3269 * or crashed) is in it. Handle its shutdown if so, and return true. If not a
3270 * bgworker, return false.
3271 *
3272 * This is heavily based on CleanupBackend. One important difference is that
3273 * we don't know yet that the dying process is a bgworker, so we must be silent
3274 * until we're sure it is.
3275 */
3276 static bool
CleanupBackgroundWorker(int pid,int exitstatus)3277 CleanupBackgroundWorker(int pid,
3278 int exitstatus) /* child's exit status */
3279 {
3280 char namebuf[MAXPGPATH];
3281 slist_mutable_iter iter;
3282
3283 slist_foreach_modify(iter, &BackgroundWorkerList)
3284 {
3285 RegisteredBgWorker *rw;
3286
3287 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
3288
3289 if (rw->rw_pid != pid)
3290 continue;
3291
3292 #ifdef WIN32
3293 /* see CleanupBackend */
3294 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3295 exitstatus = 0;
3296 #endif
3297
3298 snprintf(namebuf, MAXPGPATH, _("background worker \"%s\""),
3299 rw->rw_worker.bgw_type);
3300
3301
3302 if (!EXIT_STATUS_0(exitstatus))
3303 {
3304 /* Record timestamp, so we know when to restart the worker. */
3305 rw->rw_crashed_at = GetCurrentTimestamp();
3306 }
3307 else
3308 {
3309 /* Zero exit status means terminate */
3310 rw->rw_crashed_at = 0;
3311 rw->rw_terminate = true;
3312 }
3313
3314 /*
3315 * Additionally, for shared-memory-connected workers, just like a
3316 * backend, any exit status other than 0 or 1 is considered a crash
3317 * and causes a system-wide restart.
3318 */
3319 if ((rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3320 {
3321 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3322 {
3323 HandleChildCrash(pid, exitstatus, namebuf);
3324 return true;
3325 }
3326 }
3327
3328 /*
3329 * We must release the postmaster child slot whether this worker is
3330 * connected to shared memory or not, but we only treat it as a crash
3331 * if it is in fact connected.
3332 */
3333 if (!ReleasePostmasterChildSlot(rw->rw_child_slot) &&
3334 (rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3335 {
3336 HandleChildCrash(pid, exitstatus, namebuf);
3337 return true;
3338 }
3339
3340 /* Get it out of the BackendList and clear out remaining data */
3341 dlist_delete(&rw->rw_backend->elem);
3342 #ifdef EXEC_BACKEND
3343 ShmemBackendArrayRemove(rw->rw_backend);
3344 #endif
3345
3346 /*
3347 * It's possible that this background worker started some OTHER
3348 * background worker and asked to be notified when that worker started
3349 * or stopped. If so, cancel any notifications destined for the
3350 * now-dead backend.
3351 */
3352 if (rw->rw_backend->bgworker_notify)
3353 BackgroundWorkerStopNotifications(rw->rw_pid);
3354 free(rw->rw_backend);
3355 rw->rw_backend = NULL;
3356 rw->rw_pid = 0;
3357 rw->rw_child_slot = 0;
3358 ReportBackgroundWorkerExit(&iter); /* report child death */
3359
3360 LogChildExit(EXIT_STATUS_0(exitstatus) ? DEBUG1 : LOG,
3361 namebuf, pid, exitstatus);
3362
3363 return true;
3364 }
3365
3366 return false;
3367 }
3368
3369 /*
3370 * CleanupBackend -- cleanup after terminated backend.
3371 *
3372 * Remove all local state associated with backend.
3373 *
3374 * If you change this, see also CleanupBackgroundWorker.
3375 */
3376 static void
CleanupBackend(int pid,int exitstatus)3377 CleanupBackend(int pid,
3378 int exitstatus) /* child's exit status. */
3379 {
3380 dlist_mutable_iter iter;
3381
3382 LogChildExit(DEBUG2, _("server process"), pid, exitstatus);
3383
3384 /*
3385 * If a backend dies in an ugly way then we must signal all other backends
3386 * to quickdie. If exit status is zero (normal) or one (FATAL exit), we
3387 * assume everything is all right and proceed to remove the backend from
3388 * the active backend list.
3389 */
3390
3391 #ifdef WIN32
3392
3393 /*
3394 * On win32, also treat ERROR_WAIT_NO_CHILDREN (128) as nonfatal case,
3395 * since that sometimes happens under load when the process fails to start
3396 * properly (long before it starts using shared memory). Microsoft reports
3397 * it is related to mutex failure:
3398 * http://archives.postgresql.org/pgsql-hackers/2010-09/msg00790.php
3399 */
3400 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3401 {
3402 LogChildExit(LOG, _("server process"), pid, exitstatus);
3403 exitstatus = 0;
3404 }
3405 #endif
3406
3407 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3408 {
3409 HandleChildCrash(pid, exitstatus, _("server process"));
3410 return;
3411 }
3412
3413 dlist_foreach_modify(iter, &BackendList)
3414 {
3415 Backend *bp = dlist_container(Backend, elem, iter.cur);
3416
3417 if (bp->pid == pid)
3418 {
3419 if (!bp->dead_end)
3420 {
3421 if (!ReleasePostmasterChildSlot(bp->child_slot))
3422 {
3423 /*
3424 * Uh-oh, the child failed to clean itself up. Treat as a
3425 * crash after all.
3426 */
3427 HandleChildCrash(pid, exitstatus, _("server process"));
3428 return;
3429 }
3430 #ifdef EXEC_BACKEND
3431 ShmemBackendArrayRemove(bp);
3432 #endif
3433 }
3434 if (bp->bgworker_notify)
3435 {
3436 /*
3437 * This backend may have been slated to receive SIGUSR1 when
3438 * some background worker started or stopped. Cancel those
3439 * notifications, as we don't want to signal PIDs that are not
3440 * PostgreSQL backends. This gets skipped in the (probably
3441 * very common) case where the backend has never requested any
3442 * such notifications.
3443 */
3444 BackgroundWorkerStopNotifications(bp->pid);
3445 }
3446 dlist_delete(iter.cur);
3447 free(bp);
3448 break;
3449 }
3450 }
3451 }
3452
3453 /*
3454 * HandleChildCrash -- cleanup after failed backend, bgwriter, checkpointer,
3455 * walwriter, autovacuum, or background worker.
3456 *
3457 * The objectives here are to clean up our local state about the child
3458 * process, and to signal all other remaining children to quickdie.
3459 */
3460 static void
HandleChildCrash(int pid,int exitstatus,const char * procname)3461 HandleChildCrash(int pid, int exitstatus, const char *procname)
3462 {
3463 dlist_mutable_iter iter;
3464 slist_iter siter;
3465 Backend *bp;
3466 bool take_action;
3467
3468 /*
3469 * We only log messages and send signals if this is the first process
3470 * crash and we're not doing an immediate shutdown; otherwise, we're only
3471 * here to update postmaster's idea of live processes. If we have already
3472 * signaled children, nonzero exit status is to be expected, so don't
3473 * clutter log.
3474 */
3475 take_action = !FatalError && Shutdown != ImmediateShutdown;
3476
3477 if (take_action)
3478 {
3479 LogChildExit(LOG, procname, pid, exitstatus);
3480 ereport(LOG,
3481 (errmsg("terminating any other active server processes")));
3482 }
3483
3484 /* Process background workers. */
3485 slist_foreach(siter, &BackgroundWorkerList)
3486 {
3487 RegisteredBgWorker *rw;
3488
3489 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
3490 if (rw->rw_pid == 0)
3491 continue; /* not running */
3492 if (rw->rw_pid == pid)
3493 {
3494 /*
3495 * Found entry for freshly-dead worker, so remove it.
3496 */
3497 (void) ReleasePostmasterChildSlot(rw->rw_child_slot);
3498 dlist_delete(&rw->rw_backend->elem);
3499 #ifdef EXEC_BACKEND
3500 ShmemBackendArrayRemove(rw->rw_backend);
3501 #endif
3502 free(rw->rw_backend);
3503 rw->rw_backend = NULL;
3504 rw->rw_pid = 0;
3505 rw->rw_child_slot = 0;
3506 /* don't reset crashed_at */
3507 /* don't report child stop, either */
3508 /* Keep looping so we can signal remaining workers */
3509 }
3510 else
3511 {
3512 /*
3513 * This worker is still alive. Unless we did so already, tell it
3514 * to commit hara-kiri.
3515 *
3516 * SIGQUIT is the special signal that says exit without proc_exit
3517 * and let the user know what's going on. But if SendStop is set
3518 * (-s on command line), then we send SIGSTOP instead, so that we
3519 * can get core dumps from all backends by hand.
3520 */
3521 if (take_action)
3522 {
3523 ereport(DEBUG2,
3524 (errmsg_internal("sending %s to process %d",
3525 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3526 (int) rw->rw_pid)));
3527 signal_child(rw->rw_pid, (SendStop ? SIGSTOP : SIGQUIT));
3528 }
3529 }
3530 }
3531
3532 /* Process regular backends */
3533 dlist_foreach_modify(iter, &BackendList)
3534 {
3535 bp = dlist_container(Backend, elem, iter.cur);
3536
3537 if (bp->pid == pid)
3538 {
3539 /*
3540 * Found entry for freshly-dead backend, so remove it.
3541 */
3542 if (!bp->dead_end)
3543 {
3544 (void) ReleasePostmasterChildSlot(bp->child_slot);
3545 #ifdef EXEC_BACKEND
3546 ShmemBackendArrayRemove(bp);
3547 #endif
3548 }
3549 dlist_delete(iter.cur);
3550 free(bp);
3551 /* Keep looping so we can signal remaining backends */
3552 }
3553 else
3554 {
3555 /*
3556 * This backend is still alive. Unless we did so already, tell it
3557 * to commit hara-kiri.
3558 *
3559 * SIGQUIT is the special signal that says exit without proc_exit
3560 * and let the user know what's going on. But if SendStop is set
3561 * (-s on command line), then we send SIGSTOP instead, so that we
3562 * can get core dumps from all backends by hand.
3563 *
3564 * We could exclude dead_end children here, but at least in the
3565 * SIGSTOP case it seems better to include them.
3566 *
3567 * Background workers were already processed above; ignore them
3568 * here.
3569 */
3570 if (bp->bkend_type == BACKEND_TYPE_BGWORKER)
3571 continue;
3572
3573 if (take_action)
3574 {
3575 ereport(DEBUG2,
3576 (errmsg_internal("sending %s to process %d",
3577 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3578 (int) bp->pid)));
3579 signal_child(bp->pid, (SendStop ? SIGSTOP : SIGQUIT));
3580 }
3581 }
3582 }
3583
3584 /* Take care of the startup process too */
3585 if (pid == StartupPID)
3586 {
3587 StartupPID = 0;
3588 /* Caller adjusts StartupStatus, so don't touch it here */
3589 }
3590 else if (StartupPID != 0 && take_action)
3591 {
3592 ereport(DEBUG2,
3593 (errmsg_internal("sending %s to process %d",
3594 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3595 (int) StartupPID)));
3596 signal_child(StartupPID, (SendStop ? SIGSTOP : SIGQUIT));
3597 StartupStatus = STARTUP_SIGNALED;
3598 }
3599
3600 /* Take care of the bgwriter too */
3601 if (pid == BgWriterPID)
3602 BgWriterPID = 0;
3603 else if (BgWriterPID != 0 && take_action)
3604 {
3605 ereport(DEBUG2,
3606 (errmsg_internal("sending %s to process %d",
3607 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3608 (int) BgWriterPID)));
3609 signal_child(BgWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3610 }
3611
3612 /* Take care of the checkpointer too */
3613 if (pid == CheckpointerPID)
3614 CheckpointerPID = 0;
3615 else if (CheckpointerPID != 0 && take_action)
3616 {
3617 ereport(DEBUG2,
3618 (errmsg_internal("sending %s to process %d",
3619 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3620 (int) CheckpointerPID)));
3621 signal_child(CheckpointerPID, (SendStop ? SIGSTOP : SIGQUIT));
3622 }
3623
3624 /* Take care of the walwriter too */
3625 if (pid == WalWriterPID)
3626 WalWriterPID = 0;
3627 else if (WalWriterPID != 0 && take_action)
3628 {
3629 ereport(DEBUG2,
3630 (errmsg_internal("sending %s to process %d",
3631 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3632 (int) WalWriterPID)));
3633 signal_child(WalWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3634 }
3635
3636 /* Take care of the walreceiver too */
3637 if (pid == WalReceiverPID)
3638 WalReceiverPID = 0;
3639 else if (WalReceiverPID != 0 && take_action)
3640 {
3641 ereport(DEBUG2,
3642 (errmsg_internal("sending %s to process %d",
3643 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3644 (int) WalReceiverPID)));
3645 signal_child(WalReceiverPID, (SendStop ? SIGSTOP : SIGQUIT));
3646 }
3647
3648 /* Take care of the autovacuum launcher too */
3649 if (pid == AutoVacPID)
3650 AutoVacPID = 0;
3651 else if (AutoVacPID != 0 && take_action)
3652 {
3653 ereport(DEBUG2,
3654 (errmsg_internal("sending %s to process %d",
3655 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3656 (int) AutoVacPID)));
3657 signal_child(AutoVacPID, (SendStop ? SIGSTOP : SIGQUIT));
3658 }
3659
3660 /*
3661 * Force a power-cycle of the pgarch process too. (This isn't absolutely
3662 * necessary, but it seems like a good idea for robustness, and it
3663 * simplifies the state-machine logic in the case where a shutdown request
3664 * arrives during crash processing.)
3665 */
3666 if (PgArchPID != 0 && take_action)
3667 {
3668 ereport(DEBUG2,
3669 (errmsg_internal("sending %s to process %d",
3670 "SIGQUIT",
3671 (int) PgArchPID)));
3672 signal_child(PgArchPID, SIGQUIT);
3673 }
3674
3675 /*
3676 * Force a power-cycle of the pgstat process too. (This isn't absolutely
3677 * necessary, but it seems like a good idea for robustness, and it
3678 * simplifies the state-machine logic in the case where a shutdown request
3679 * arrives during crash processing.)
3680 */
3681 if (PgStatPID != 0 && take_action)
3682 {
3683 ereport(DEBUG2,
3684 (errmsg_internal("sending %s to process %d",
3685 "SIGQUIT",
3686 (int) PgStatPID)));
3687 signal_child(PgStatPID, SIGQUIT);
3688 allow_immediate_pgstat_restart();
3689 }
3690
3691 /* We do NOT restart the syslogger */
3692
3693 if (Shutdown != ImmediateShutdown)
3694 FatalError = true;
3695
3696 /* We now transit into a state of waiting for children to die */
3697 if (pmState == PM_RECOVERY ||
3698 pmState == PM_HOT_STANDBY ||
3699 pmState == PM_RUN ||
3700 pmState == PM_STOP_BACKENDS ||
3701 pmState == PM_SHUTDOWN)
3702 pmState = PM_WAIT_BACKENDS;
3703
3704 /*
3705 * .. and if this doesn't happen quickly enough, now the clock is ticking
3706 * for us to kill them without mercy.
3707 */
3708 if (AbortStartTime == 0)
3709 AbortStartTime = time(NULL);
3710 }
3711
3712 /*
3713 * Log the death of a child process.
3714 */
3715 static void
LogChildExit(int lev,const char * procname,int pid,int exitstatus)3716 LogChildExit(int lev, const char *procname, int pid, int exitstatus)
3717 {
3718 /*
3719 * size of activity_buffer is arbitrary, but set equal to default
3720 * track_activity_query_size
3721 */
3722 char activity_buffer[1024];
3723 const char *activity = NULL;
3724
3725 if (!EXIT_STATUS_0(exitstatus))
3726 activity = pgstat_get_crashed_backend_activity(pid,
3727 activity_buffer,
3728 sizeof(activity_buffer));
3729
3730 if (WIFEXITED(exitstatus))
3731 ereport(lev,
3732
3733 /*------
3734 translator: %s is a noun phrase describing a child process, such as
3735 "server process" */
3736 (errmsg("%s (PID %d) exited with exit code %d",
3737 procname, pid, WEXITSTATUS(exitstatus)),
3738 activity ? errdetail("Failed process was running: %s", activity) : 0));
3739 else if (WIFSIGNALED(exitstatus))
3740 {
3741 #if defined(WIN32)
3742 ereport(lev,
3743
3744 /*------
3745 translator: %s is a noun phrase describing a child process, such as
3746 "server process" */
3747 (errmsg("%s (PID %d) was terminated by exception 0x%X",
3748 procname, pid, WTERMSIG(exitstatus)),
3749 errhint("See C include file \"ntstatus.h\" for a description of the hexadecimal value."),
3750 activity ? errdetail("Failed process was running: %s", activity) : 0));
3751 #else
3752 ereport(lev,
3753
3754 /*------
3755 translator: %s is a noun phrase describing a child process, such as
3756 "server process" */
3757 (errmsg("%s (PID %d) was terminated by signal %d: %s",
3758 procname, pid, WTERMSIG(exitstatus),
3759 pg_strsignal(WTERMSIG(exitstatus))),
3760 activity ? errdetail("Failed process was running: %s", activity) : 0));
3761 #endif
3762 }
3763 else
3764 ereport(lev,
3765
3766 /*------
3767 translator: %s is a noun phrase describing a child process, such as
3768 "server process" */
3769 (errmsg("%s (PID %d) exited with unrecognized status %d",
3770 procname, pid, exitstatus),
3771 activity ? errdetail("Failed process was running: %s", activity) : 0));
3772 }
3773
3774 /*
3775 * Advance the postmaster's state machine and take actions as appropriate
3776 *
3777 * This is common code for pmdie(), reaper() and sigusr1_handler(), which
3778 * receive the signals that might mean we need to change state.
3779 */
3780 static void
PostmasterStateMachine(void)3781 PostmasterStateMachine(void)
3782 {
3783 /* If we're doing a smart shutdown, try to advance that state. */
3784 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3785 {
3786 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
3787 {
3788 /*
3789 * ALLOW_SUPERUSER_CONNS state ends as soon as online backup mode
3790 * is not active.
3791 */
3792 if (!BackupInProgress())
3793 connsAllowed = ALLOW_NO_CONNS;
3794 }
3795
3796 if (connsAllowed == ALLOW_NO_CONNS)
3797 {
3798 /*
3799 * ALLOW_NO_CONNS state ends when we have no normal client
3800 * backends running. Then we're ready to stop other children.
3801 */
3802 if (CountChildren(BACKEND_TYPE_NORMAL) == 0)
3803 pmState = PM_STOP_BACKENDS;
3804 }
3805 }
3806
3807 /*
3808 * If we're ready to do so, signal child processes to shut down. (This
3809 * isn't a persistent state, but treating it as a distinct pmState allows
3810 * us to share this code across multiple shutdown code paths.)
3811 */
3812 if (pmState == PM_STOP_BACKENDS)
3813 {
3814 /*
3815 * Forget any pending requests for background workers, since we're no
3816 * longer willing to launch any new workers. (If additional requests
3817 * arrive, BackgroundWorkerStateChange will reject them.)
3818 */
3819 ForgetUnstartedBackgroundWorkers();
3820
3821 /* Signal all backend children except walsenders */
3822 SignalSomeChildren(SIGTERM,
3823 BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND);
3824 /* and the autovac launcher too */
3825 if (AutoVacPID != 0)
3826 signal_child(AutoVacPID, SIGTERM);
3827 /* and the bgwriter too */
3828 if (BgWriterPID != 0)
3829 signal_child(BgWriterPID, SIGTERM);
3830 /* and the walwriter too */
3831 if (WalWriterPID != 0)
3832 signal_child(WalWriterPID, SIGTERM);
3833 /* If we're in recovery, also stop startup and walreceiver procs */
3834 if (StartupPID != 0)
3835 signal_child(StartupPID, SIGTERM);
3836 if (WalReceiverPID != 0)
3837 signal_child(WalReceiverPID, SIGTERM);
3838 /* checkpointer, archiver, stats, and syslogger may continue for now */
3839
3840 /* Now transition to PM_WAIT_BACKENDS state to wait for them to die */
3841 pmState = PM_WAIT_BACKENDS;
3842 }
3843
3844 /*
3845 * If we are in a state-machine state that implies waiting for backends to
3846 * exit, see if they're all gone, and change state if so.
3847 */
3848 if (pmState == PM_WAIT_BACKENDS)
3849 {
3850 /*
3851 * PM_WAIT_BACKENDS state ends when we have no regular backends
3852 * (including autovac workers), no bgworkers (including unconnected
3853 * ones), and no walwriter, autovac launcher or bgwriter. If we are
3854 * doing crash recovery or an immediate shutdown then we expect the
3855 * checkpointer to exit as well, otherwise not. The archiver, stats,
3856 * and syslogger processes are disregarded since they are not
3857 * connected to shared memory; we also disregard dead_end children
3858 * here. Walsenders are also disregarded, they will be terminated
3859 * later after writing the checkpoint record, like the archiver
3860 * process.
3861 */
3862 if (CountChildren(BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND) == 0 &&
3863 StartupPID == 0 &&
3864 WalReceiverPID == 0 &&
3865 BgWriterPID == 0 &&
3866 (CheckpointerPID == 0 ||
3867 (!FatalError && Shutdown < ImmediateShutdown)) &&
3868 WalWriterPID == 0 &&
3869 AutoVacPID == 0)
3870 {
3871 if (Shutdown >= ImmediateShutdown || FatalError)
3872 {
3873 /*
3874 * Start waiting for dead_end children to die. This state
3875 * change causes ServerLoop to stop creating new ones.
3876 */
3877 pmState = PM_WAIT_DEAD_END;
3878
3879 /*
3880 * We already SIGQUIT'd the archiver and stats processes, if
3881 * any, when we started immediate shutdown or entered
3882 * FatalError state.
3883 */
3884 }
3885 else
3886 {
3887 /*
3888 * If we get here, we are proceeding with normal shutdown. All
3889 * the regular children are gone, and it's time to tell the
3890 * checkpointer to do a shutdown checkpoint.
3891 */
3892 Assert(Shutdown > NoShutdown);
3893 /* Start the checkpointer if not running */
3894 if (CheckpointerPID == 0)
3895 CheckpointerPID = StartCheckpointer();
3896 /* And tell it to shut down */
3897 if (CheckpointerPID != 0)
3898 {
3899 signal_child(CheckpointerPID, SIGUSR2);
3900 pmState = PM_SHUTDOWN;
3901 }
3902 else
3903 {
3904 /*
3905 * If we failed to fork a checkpointer, just shut down.
3906 * Any required cleanup will happen at next restart. We
3907 * set FatalError so that an "abnormal shutdown" message
3908 * gets logged when we exit.
3909 */
3910 FatalError = true;
3911 pmState = PM_WAIT_DEAD_END;
3912
3913 /* Kill the walsenders, archiver and stats collector too */
3914 SignalChildren(SIGQUIT);
3915 if (PgArchPID != 0)
3916 signal_child(PgArchPID, SIGQUIT);
3917 if (PgStatPID != 0)
3918 signal_child(PgStatPID, SIGQUIT);
3919 }
3920 }
3921 }
3922 }
3923
3924 if (pmState == PM_SHUTDOWN_2)
3925 {
3926 /*
3927 * PM_SHUTDOWN_2 state ends when there's no other children than
3928 * dead_end children left. There shouldn't be any regular backends
3929 * left by now anyway; what we're really waiting for is walsenders and
3930 * archiver.
3931 */
3932 if (PgArchPID == 0 && CountChildren(BACKEND_TYPE_ALL) == 0)
3933 {
3934 pmState = PM_WAIT_DEAD_END;
3935 }
3936 }
3937
3938 if (pmState == PM_WAIT_DEAD_END)
3939 {
3940 /*
3941 * PM_WAIT_DEAD_END state ends when the BackendList is entirely empty
3942 * (ie, no dead_end children remain), and the archiver and stats
3943 * collector are gone too.
3944 *
3945 * The reason we wait for those two is to protect them against a new
3946 * postmaster starting conflicting subprocesses; this isn't an
3947 * ironclad protection, but it at least helps in the
3948 * shutdown-and-immediately-restart scenario. Note that they have
3949 * already been sent appropriate shutdown signals, either during a
3950 * normal state transition leading up to PM_WAIT_DEAD_END, or during
3951 * FatalError processing.
3952 */
3953 if (dlist_is_empty(&BackendList) &&
3954 PgArchPID == 0 && PgStatPID == 0)
3955 {
3956 /* These other guys should be dead already */
3957 Assert(StartupPID == 0);
3958 Assert(WalReceiverPID == 0);
3959 Assert(BgWriterPID == 0);
3960 Assert(CheckpointerPID == 0);
3961 Assert(WalWriterPID == 0);
3962 Assert(AutoVacPID == 0);
3963 /* syslogger is not considered here */
3964 pmState = PM_NO_CHILDREN;
3965 }
3966 }
3967
3968 /*
3969 * If we've been told to shut down, we exit as soon as there are no
3970 * remaining children. If there was a crash, cleanup will occur at the
3971 * next startup. (Before PostgreSQL 8.3, we tried to recover from the
3972 * crash before exiting, but that seems unwise if we are quitting because
3973 * we got SIGTERM from init --- there may well not be time for recovery
3974 * before init decides to SIGKILL us.)
3975 *
3976 * Note that the syslogger continues to run. It will exit when it sees
3977 * EOF on its input pipe, which happens when there are no more upstream
3978 * processes.
3979 */
3980 if (Shutdown > NoShutdown && pmState == PM_NO_CHILDREN)
3981 {
3982 if (FatalError)
3983 {
3984 ereport(LOG, (errmsg("abnormal database system shutdown")));
3985 ExitPostmaster(1);
3986 }
3987 else
3988 {
3989 /*
3990 * Terminate exclusive backup mode to avoid recovery after a clean
3991 * fast shutdown. Since an exclusive backup can only be taken
3992 * during normal running (and not, for example, while running
3993 * under Hot Standby) it only makes sense to do this if we reached
3994 * normal running. If we're still in recovery, the backup file is
3995 * one we're recovering *from*, and we must keep it around so that
3996 * recovery restarts from the right place.
3997 */
3998 if (ReachedNormalRunning)
3999 CancelBackup();
4000
4001 /* Normal exit from the postmaster is here */
4002 ExitPostmaster(0);
4003 }
4004 }
4005
4006 /*
4007 * If the startup process failed, or the user does not want an automatic
4008 * restart after backend crashes, wait for all non-syslogger children to
4009 * exit, and then exit postmaster. We don't try to reinitialize when the
4010 * startup process fails, because more than likely it will just fail again
4011 * and we will keep trying forever.
4012 */
4013 if (pmState == PM_NO_CHILDREN &&
4014 (StartupStatus == STARTUP_CRASHED || !restart_after_crash))
4015 ExitPostmaster(1);
4016
4017 /*
4018 * If we need to recover from a crash, wait for all non-syslogger children
4019 * to exit, then reset shmem and StartupDataBase.
4020 */
4021 if (FatalError && pmState == PM_NO_CHILDREN)
4022 {
4023 ereport(LOG,
4024 (errmsg("all server processes terminated; reinitializing")));
4025
4026 /* allow background workers to immediately restart */
4027 ResetBackgroundWorkerCrashTimes();
4028
4029 shmem_exit(1);
4030
4031 /* re-read control file into local memory */
4032 LocalProcessControlFile(true);
4033
4034 reset_shared();
4035
4036 StartupPID = StartupDataBase();
4037 Assert(StartupPID != 0);
4038 StartupStatus = STARTUP_RUNNING;
4039 pmState = PM_STARTUP;
4040 /* crash recovery started, reset SIGKILL flag */
4041 AbortStartTime = 0;
4042 }
4043 }
4044
4045
4046 /*
4047 * Send a signal to a postmaster child process
4048 *
4049 * On systems that have setsid(), each child process sets itself up as a
4050 * process group leader. For signals that are generally interpreted in the
4051 * appropriate fashion, we signal the entire process group not just the
4052 * direct child process. This allows us to, for example, SIGQUIT a blocked
4053 * archive_recovery script, or SIGINT a script being run by a backend via
4054 * system().
4055 *
4056 * There is a race condition for recently-forked children: they might not
4057 * have executed setsid() yet. So we signal the child directly as well as
4058 * the group. We assume such a child will handle the signal before trying
4059 * to spawn any grandchild processes. We also assume that signaling the
4060 * child twice will not cause any problems.
4061 */
4062 static void
signal_child(pid_t pid,int signal)4063 signal_child(pid_t pid, int signal)
4064 {
4065 if (kill(pid, signal) < 0)
4066 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) pid, signal);
4067 #ifdef HAVE_SETSID
4068 switch (signal)
4069 {
4070 case SIGINT:
4071 case SIGTERM:
4072 case SIGQUIT:
4073 case SIGSTOP:
4074 case SIGKILL:
4075 if (kill(-pid, signal) < 0)
4076 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) (-pid), signal);
4077 break;
4078 default:
4079 break;
4080 }
4081 #endif
4082 }
4083
4084 /*
4085 * Send a signal to the targeted children (but NOT special children;
4086 * dead_end children are never signaled, either).
4087 */
4088 static bool
SignalSomeChildren(int signal,int target)4089 SignalSomeChildren(int signal, int target)
4090 {
4091 dlist_iter iter;
4092 bool signaled = false;
4093
4094 dlist_foreach(iter, &BackendList)
4095 {
4096 Backend *bp = dlist_container(Backend, elem, iter.cur);
4097
4098 if (bp->dead_end)
4099 continue;
4100
4101 /*
4102 * Since target == BACKEND_TYPE_ALL is the most common case, we test
4103 * it first and avoid touching shared memory for every child.
4104 */
4105 if (target != BACKEND_TYPE_ALL)
4106 {
4107 /*
4108 * Assign bkend_type for any recently announced WAL Sender
4109 * processes.
4110 */
4111 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
4112 IsPostmasterChildWalSender(bp->child_slot))
4113 bp->bkend_type = BACKEND_TYPE_WALSND;
4114
4115 if (!(target & bp->bkend_type))
4116 continue;
4117 }
4118
4119 ereport(DEBUG4,
4120 (errmsg_internal("sending signal %d to process %d",
4121 signal, (int) bp->pid)));
4122 signal_child(bp->pid, signal);
4123 signaled = true;
4124 }
4125 return signaled;
4126 }
4127
4128 /*
4129 * Send a termination signal to children. This considers all of our children
4130 * processes, except syslogger and dead_end backends.
4131 */
4132 static void
TerminateChildren(int signal)4133 TerminateChildren(int signal)
4134 {
4135 SignalChildren(signal);
4136 if (StartupPID != 0)
4137 {
4138 signal_child(StartupPID, signal);
4139 if (signal == SIGQUIT || signal == SIGKILL)
4140 StartupStatus = STARTUP_SIGNALED;
4141 }
4142 if (BgWriterPID != 0)
4143 signal_child(BgWriterPID, signal);
4144 if (CheckpointerPID != 0)
4145 signal_child(CheckpointerPID, signal);
4146 if (WalWriterPID != 0)
4147 signal_child(WalWriterPID, signal);
4148 if (WalReceiverPID != 0)
4149 signal_child(WalReceiverPID, signal);
4150 if (AutoVacPID != 0)
4151 signal_child(AutoVacPID, signal);
4152 if (PgArchPID != 0)
4153 signal_child(PgArchPID, signal);
4154 if (PgStatPID != 0)
4155 signal_child(PgStatPID, signal);
4156 }
4157
4158 /*
4159 * BackendStartup -- start backend process
4160 *
4161 * returns: STATUS_ERROR if the fork failed, STATUS_OK otherwise.
4162 *
4163 * Note: if you change this code, also consider StartAutovacuumWorker.
4164 */
4165 static int
BackendStartup(Port * port)4166 BackendStartup(Port *port)
4167 {
4168 Backend *bn; /* for backend cleanup */
4169 pid_t pid;
4170
4171 /*
4172 * Create backend data structure. Better before the fork() so we can
4173 * handle failure cleanly.
4174 */
4175 bn = (Backend *) malloc(sizeof(Backend));
4176 if (!bn)
4177 {
4178 ereport(LOG,
4179 (errcode(ERRCODE_OUT_OF_MEMORY),
4180 errmsg("out of memory")));
4181 return STATUS_ERROR;
4182 }
4183
4184 /*
4185 * Compute the cancel key that will be assigned to this backend. The
4186 * backend will have its own copy in the forked-off process' value of
4187 * MyCancelKey, so that it can transmit the key to the frontend.
4188 */
4189 if (!RandomCancelKey(&MyCancelKey))
4190 {
4191 free(bn);
4192 ereport(LOG,
4193 (errcode(ERRCODE_INTERNAL_ERROR),
4194 errmsg("could not generate random cancel key")));
4195 return STATUS_ERROR;
4196 }
4197
4198 bn->cancel_key = MyCancelKey;
4199
4200 /* Pass down canAcceptConnections state */
4201 port->canAcceptConnections = canAcceptConnections(BACKEND_TYPE_NORMAL);
4202 bn->dead_end = (port->canAcceptConnections != CAC_OK &&
4203 port->canAcceptConnections != CAC_SUPERUSER);
4204
4205 /*
4206 * Unless it's a dead_end child, assign it a child slot number
4207 */
4208 if (!bn->dead_end)
4209 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
4210 else
4211 bn->child_slot = 0;
4212
4213 /* Hasn't asked to be notified about any bgworkers yet */
4214 bn->bgworker_notify = false;
4215
4216 #ifdef EXEC_BACKEND
4217 pid = backend_forkexec(port);
4218 #else /* !EXEC_BACKEND */
4219 pid = fork_process();
4220 if (pid == 0) /* child */
4221 {
4222 free(bn);
4223
4224 /* Detangle from postmaster */
4225 InitPostmasterChild();
4226
4227 /* Close the postmaster's sockets */
4228 ClosePostmasterPorts(false);
4229
4230 /* Perform additional initialization and collect startup packet */
4231 BackendInitialize(port);
4232
4233 /* And run the backend */
4234 BackendRun(port);
4235 }
4236 #endif /* EXEC_BACKEND */
4237
4238 if (pid < 0)
4239 {
4240 /* in parent, fork failed */
4241 int save_errno = errno;
4242
4243 if (!bn->dead_end)
4244 (void) ReleasePostmasterChildSlot(bn->child_slot);
4245 free(bn);
4246 errno = save_errno;
4247 ereport(LOG,
4248 (errmsg("could not fork new process for connection: %m")));
4249 report_fork_failure_to_client(port, save_errno);
4250 return STATUS_ERROR;
4251 }
4252
4253 /* in parent, successful fork */
4254 ereport(DEBUG2,
4255 (errmsg_internal("forked new backend, pid=%d socket=%d",
4256 (int) pid, (int) port->sock)));
4257
4258 /*
4259 * Everything's been successful, it's safe to add this backend to our list
4260 * of backends.
4261 */
4262 bn->pid = pid;
4263 bn->bkend_type = BACKEND_TYPE_NORMAL; /* Can change later to WALSND */
4264 dlist_push_head(&BackendList, &bn->elem);
4265
4266 #ifdef EXEC_BACKEND
4267 if (!bn->dead_end)
4268 ShmemBackendArrayAdd(bn);
4269 #endif
4270
4271 return STATUS_OK;
4272 }
4273
4274 /*
4275 * Try to report backend fork() failure to client before we close the
4276 * connection. Since we do not care to risk blocking the postmaster on
4277 * this connection, we set the connection to non-blocking and try only once.
4278 *
4279 * This is grungy special-purpose code; we cannot use backend libpq since
4280 * it's not up and running.
4281 */
4282 static void
report_fork_failure_to_client(Port * port,int errnum)4283 report_fork_failure_to_client(Port *port, int errnum)
4284 {
4285 char buffer[1000];
4286 int rc;
4287
4288 /* Format the error message packet (always V2 protocol) */
4289 snprintf(buffer, sizeof(buffer), "E%s%s\n",
4290 _("could not fork new process for connection: "),
4291 strerror(errnum));
4292
4293 /* Set port to non-blocking. Don't do send() if this fails */
4294 if (!pg_set_noblock(port->sock))
4295 return;
4296
4297 /* We'll retry after EINTR, but ignore all other failures */
4298 do
4299 {
4300 rc = send(port->sock, buffer, strlen(buffer) + 1, 0);
4301 } while (rc < 0 && errno == EINTR);
4302 }
4303
4304
4305 /*
4306 * BackendInitialize -- initialize an interactive (postmaster-child)
4307 * backend process, and collect the client's startup packet.
4308 *
4309 * returns: nothing. Will not return at all if there's any failure.
4310 *
4311 * Note: this code does not depend on having any access to shared memory.
4312 * In the EXEC_BACKEND case, we are physically attached to shared memory
4313 * but have not yet set up most of our local pointers to shmem structures.
4314 */
4315 static void
BackendInitialize(Port * port)4316 BackendInitialize(Port *port)
4317 {
4318 int status;
4319 int ret;
4320 char remote_host[NI_MAXHOST];
4321 char remote_port[NI_MAXSERV];
4322 StringInfoData ps_data;
4323
4324 /* Save port etc. for ps status */
4325 MyProcPort = port;
4326
4327 /* Tell fd.c about the long-lived FD associated with the port */
4328 ReserveExternalFD();
4329
4330 /*
4331 * PreAuthDelay is a debugging aid for investigating problems in the
4332 * authentication cycle: it can be set in postgresql.conf to allow time to
4333 * attach to the newly-forked backend with a debugger. (See also
4334 * PostAuthDelay, which we allow clients to pass through PGOPTIONS, but it
4335 * is not honored until after authentication.)
4336 */
4337 if (PreAuthDelay > 0)
4338 pg_usleep(PreAuthDelay * 1000000L);
4339
4340 /* This flag will remain set until InitPostgres finishes authentication */
4341 ClientAuthInProgress = true; /* limit visibility of log messages */
4342
4343 /* set these to empty in case they are needed before we set them up */
4344 port->remote_host = "";
4345 port->remote_port = "";
4346
4347 /*
4348 * Initialize libpq and enable reporting of ereport errors to the client.
4349 * Must do this now because authentication uses libpq to send messages.
4350 */
4351 pq_init(); /* initialize libpq to talk to client */
4352 whereToSendOutput = DestRemote; /* now safe to ereport to client */
4353
4354 /*
4355 * We arrange to do proc_exit(1) if we receive SIGTERM or timeout while
4356 * trying to collect the startup packet; while SIGQUIT results in
4357 * _exit(2). Otherwise the postmaster cannot shutdown the database FAST
4358 * or IMMED cleanly if a buggy client fails to send the packet promptly.
4359 *
4360 * XXX this is pretty dangerous; signal handlers should not call anything
4361 * as complex as proc_exit() directly. We minimize the hazard by not
4362 * keeping these handlers active for longer than we must. However, it
4363 * seems necessary to be able to escape out of DNS lookups as well as the
4364 * startup packet reception proper, so we can't narrow the scope further
4365 * than is done here.
4366 *
4367 * XXX it follows that the remainder of this function must tolerate losing
4368 * control at any instant. Likewise, any pg_on_exit_callback registered
4369 * before or during this function must be prepared to execute at any
4370 * instant between here and the end of this function. Furthermore,
4371 * affected callbacks execute partially or not at all when a second
4372 * exit-inducing signal arrives after proc_exit_prepare() decrements
4373 * on_proc_exit_index. (Thanks to that mechanic, callbacks need not
4374 * anticipate more than one call.) This is fragile; it ought to instead
4375 * follow the norm of handling interrupts at selected, safe opportunities.
4376 */
4377 pqsignal(SIGTERM, process_startup_packet_die);
4378 pqsignal(SIGQUIT, SignalHandlerForCrashExit);
4379 InitializeTimeouts(); /* establishes SIGALRM handler */
4380 PG_SETMASK(&StartupBlockSig);
4381
4382 /*
4383 * Get the remote host name and port for logging and status display.
4384 */
4385 remote_host[0] = '\0';
4386 remote_port[0] = '\0';
4387 if ((ret = pg_getnameinfo_all(&port->raddr.addr, port->raddr.salen,
4388 remote_host, sizeof(remote_host),
4389 remote_port, sizeof(remote_port),
4390 (log_hostname ? 0 : NI_NUMERICHOST) | NI_NUMERICSERV)) != 0)
4391 ereport(WARNING,
4392 (errmsg_internal("pg_getnameinfo_all() failed: %s",
4393 gai_strerror(ret))));
4394
4395 /*
4396 * Save remote_host and remote_port in port structure (after this, they
4397 * will appear in log_line_prefix data for log messages).
4398 */
4399 port->remote_host = strdup(remote_host);
4400 port->remote_port = strdup(remote_port);
4401
4402 /* And now we can issue the Log_connections message, if wanted */
4403 if (Log_connections)
4404 {
4405 if (remote_port[0])
4406 ereport(LOG,
4407 (errmsg("connection received: host=%s port=%s",
4408 remote_host,
4409 remote_port)));
4410 else
4411 ereport(LOG,
4412 (errmsg("connection received: host=%s",
4413 remote_host)));
4414 }
4415
4416 /*
4417 * If we did a reverse lookup to name, we might as well save the results
4418 * rather than possibly repeating the lookup during authentication.
4419 *
4420 * Note that we don't want to specify NI_NAMEREQD above, because then we'd
4421 * get nothing useful for a client without an rDNS entry. Therefore, we
4422 * must check whether we got a numeric IPv4 or IPv6 address, and not save
4423 * it into remote_hostname if so. (This test is conservative and might
4424 * sometimes classify a hostname as numeric, but an error in that
4425 * direction is safe; it only results in a possible extra lookup.)
4426 */
4427 if (log_hostname &&
4428 ret == 0 &&
4429 strspn(remote_host, "0123456789.") < strlen(remote_host) &&
4430 strspn(remote_host, "0123456789ABCDEFabcdef:") < strlen(remote_host))
4431 port->remote_hostname = strdup(remote_host);
4432
4433 /*
4434 * Ready to begin client interaction. We will give up and proc_exit(1)
4435 * after a time delay, so that a broken client can't hog a connection
4436 * indefinitely. PreAuthDelay and any DNS interactions above don't count
4437 * against the time limit.
4438 *
4439 * Note: AuthenticationTimeout is applied here while waiting for the
4440 * startup packet, and then again in InitPostgres for the duration of any
4441 * authentication operations. So a hostile client could tie up the
4442 * process for nearly twice AuthenticationTimeout before we kick him off.
4443 *
4444 * Note: because PostgresMain will call InitializeTimeouts again, the
4445 * registration of STARTUP_PACKET_TIMEOUT will be lost. This is okay
4446 * since we never use it again after this function.
4447 */
4448 RegisterTimeout(STARTUP_PACKET_TIMEOUT, StartupPacketTimeoutHandler);
4449 enable_timeout_after(STARTUP_PACKET_TIMEOUT, AuthenticationTimeout * 1000);
4450
4451 /*
4452 * Receive the startup packet (which might turn out to be a cancel request
4453 * packet).
4454 */
4455 status = ProcessStartupPacket(port, false, false);
4456
4457 /*
4458 * Disable the timeout, and prevent SIGTERM/SIGQUIT again.
4459 */
4460 disable_timeout(STARTUP_PACKET_TIMEOUT, false);
4461 PG_SETMASK(&BlockSig);
4462
4463 /*
4464 * Stop here if it was bad or a cancel packet. ProcessStartupPacket
4465 * already did any appropriate error reporting.
4466 */
4467 if (status != STATUS_OK)
4468 proc_exit(0);
4469
4470 /*
4471 * Now that we have the user and database name, we can set the process
4472 * title for ps. It's good to do this as early as possible in startup.
4473 */
4474 initStringInfo(&ps_data);
4475 if (am_walsender)
4476 appendStringInfo(&ps_data, "%s ", GetBackendTypeDesc(B_WAL_SENDER));
4477 appendStringInfo(&ps_data, "%s ", port->user_name);
4478 if (!am_walsender)
4479 appendStringInfo(&ps_data, "%s ", port->database_name);
4480 appendStringInfo(&ps_data, "%s", port->remote_host);
4481 if (port->remote_port[0] != '\0')
4482 appendStringInfo(&ps_data, "(%s)", port->remote_port);
4483
4484 init_ps_display(ps_data.data);
4485 pfree(ps_data.data);
4486
4487 set_ps_display("initializing");
4488 }
4489
4490
4491 /*
4492 * BackendRun -- set up the backend's argument list and invoke PostgresMain()
4493 *
4494 * returns:
4495 * Shouldn't return at all.
4496 * If PostgresMain() fails, return status.
4497 */
4498 static void
BackendRun(Port * port)4499 BackendRun(Port *port)
4500 {
4501 char **av;
4502 int maxac;
4503 int ac;
4504 int i;
4505
4506 /*
4507 * Now, build the argv vector that will be given to PostgresMain.
4508 *
4509 * The maximum possible number of commandline arguments that could come
4510 * from ExtraOptions is (strlen(ExtraOptions) + 1) / 2; see
4511 * pg_split_opts().
4512 */
4513 maxac = 2; /* for fixed args supplied below */
4514 maxac += (strlen(ExtraOptions) + 1) / 2;
4515
4516 av = (char **) MemoryContextAlloc(TopMemoryContext,
4517 maxac * sizeof(char *));
4518 ac = 0;
4519
4520 av[ac++] = "postgres";
4521
4522 /*
4523 * Pass any backend switches specified with -o on the postmaster's own
4524 * command line. We assume these are secure.
4525 */
4526 pg_split_opts(av, &ac, ExtraOptions);
4527
4528 av[ac] = NULL;
4529
4530 Assert(ac < maxac);
4531
4532 /*
4533 * Debug: print arguments being passed to backend
4534 */
4535 ereport(DEBUG3,
4536 (errmsg_internal("%s child[%d]: starting with (",
4537 progname, (int) getpid())));
4538 for (i = 0; i < ac; ++i)
4539 ereport(DEBUG3,
4540 (errmsg_internal("\t%s", av[i])));
4541 ereport(DEBUG3,
4542 (errmsg_internal(")")));
4543
4544 /*
4545 * Make sure we aren't in PostmasterContext anymore. (We can't delete it
4546 * just yet, though, because InitPostgres will need the HBA data.)
4547 */
4548 MemoryContextSwitchTo(TopMemoryContext);
4549
4550 PostgresMain(ac, av, port->database_name, port->user_name);
4551 }
4552
4553
4554 #ifdef EXEC_BACKEND
4555
4556 /*
4557 * postmaster_forkexec -- fork and exec a postmaster subprocess
4558 *
4559 * The caller must have set up the argv array already, except for argv[2]
4560 * which will be filled with the name of the temp variable file.
4561 *
4562 * Returns the child process PID, or -1 on fork failure (a suitable error
4563 * message has been logged on failure).
4564 *
4565 * All uses of this routine will dispatch to SubPostmasterMain in the
4566 * child process.
4567 */
4568 pid_t
postmaster_forkexec(int argc,char * argv[])4569 postmaster_forkexec(int argc, char *argv[])
4570 {
4571 Port port;
4572
4573 /* This entry point passes dummy values for the Port variables */
4574 memset(&port, 0, sizeof(port));
4575 return internal_forkexec(argc, argv, &port);
4576 }
4577
4578 /*
4579 * backend_forkexec -- fork/exec off a backend process
4580 *
4581 * Some operating systems (WIN32) don't have fork() so we have to simulate
4582 * it by storing parameters that need to be passed to the child and
4583 * then create a new child process.
4584 *
4585 * returns the pid of the fork/exec'd process, or -1 on failure
4586 */
4587 static pid_t
backend_forkexec(Port * port)4588 backend_forkexec(Port *port)
4589 {
4590 char *av[4];
4591 int ac = 0;
4592
4593 av[ac++] = "postgres";
4594 av[ac++] = "--forkbackend";
4595 av[ac++] = NULL; /* filled in by internal_forkexec */
4596
4597 av[ac] = NULL;
4598 Assert(ac < lengthof(av));
4599
4600 return internal_forkexec(ac, av, port);
4601 }
4602
4603 #ifndef WIN32
4604
4605 /*
4606 * internal_forkexec non-win32 implementation
4607 *
4608 * - writes out backend variables to the parameter file
4609 * - fork():s, and then exec():s the child process
4610 */
4611 static pid_t
internal_forkexec(int argc,char * argv[],Port * port)4612 internal_forkexec(int argc, char *argv[], Port *port)
4613 {
4614 static unsigned long tmpBackendFileNum = 0;
4615 pid_t pid;
4616 char tmpfilename[MAXPGPATH];
4617 BackendParameters param;
4618 FILE *fp;
4619
4620 if (!save_backend_variables(¶m, port))
4621 return -1; /* log made by save_backend_variables */
4622
4623 /* Calculate name for temp file */
4624 snprintf(tmpfilename, MAXPGPATH, "%s/%s.backend_var.%d.%lu",
4625 PG_TEMP_FILES_DIR, PG_TEMP_FILE_PREFIX,
4626 MyProcPid, ++tmpBackendFileNum);
4627
4628 /* Open file */
4629 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4630 if (!fp)
4631 {
4632 /*
4633 * As in OpenTemporaryFileInTablespace, try to make the temp-file
4634 * directory, ignoring errors.
4635 */
4636 (void) MakePGDirectory(PG_TEMP_FILES_DIR);
4637
4638 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4639 if (!fp)
4640 {
4641 ereport(LOG,
4642 (errcode_for_file_access(),
4643 errmsg("could not create file \"%s\": %m",
4644 tmpfilename)));
4645 return -1;
4646 }
4647 }
4648
4649 if (fwrite(¶m, sizeof(param), 1, fp) != 1)
4650 {
4651 ereport(LOG,
4652 (errcode_for_file_access(),
4653 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4654 FreeFile(fp);
4655 return -1;
4656 }
4657
4658 /* Release file */
4659 if (FreeFile(fp))
4660 {
4661 ereport(LOG,
4662 (errcode_for_file_access(),
4663 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4664 return -1;
4665 }
4666
4667 /* Make sure caller set up argv properly */
4668 Assert(argc >= 3);
4669 Assert(argv[argc] == NULL);
4670 Assert(strncmp(argv[1], "--fork", 6) == 0);
4671 Assert(argv[2] == NULL);
4672
4673 /* Insert temp file name after --fork argument */
4674 argv[2] = tmpfilename;
4675
4676 /* Fire off execv in child */
4677 if ((pid = fork_process()) == 0)
4678 {
4679 if (execv(postgres_exec_path, argv) < 0)
4680 {
4681 ereport(LOG,
4682 (errmsg("could not execute server process \"%s\": %m",
4683 postgres_exec_path)));
4684 /* We're already in the child process here, can't return */
4685 exit(1);
4686 }
4687 }
4688
4689 return pid; /* Parent returns pid, or -1 on fork failure */
4690 }
4691 #else /* WIN32 */
4692
4693 /*
4694 * internal_forkexec win32 implementation
4695 *
4696 * - starts backend using CreateProcess(), in suspended state
4697 * - writes out backend variables to the parameter file
4698 * - during this, duplicates handles and sockets required for
4699 * inheritance into the new process
4700 * - resumes execution of the new process once the backend parameter
4701 * file is complete.
4702 */
4703 static pid_t
internal_forkexec(int argc,char * argv[],Port * port)4704 internal_forkexec(int argc, char *argv[], Port *port)
4705 {
4706 int retry_count = 0;
4707 STARTUPINFO si;
4708 PROCESS_INFORMATION pi;
4709 int i;
4710 int j;
4711 char cmdLine[MAXPGPATH * 2];
4712 HANDLE paramHandle;
4713 BackendParameters *param;
4714 SECURITY_ATTRIBUTES sa;
4715 char paramHandleStr[32];
4716 win32_deadchild_waitinfo *childinfo;
4717
4718 /* Make sure caller set up argv properly */
4719 Assert(argc >= 3);
4720 Assert(argv[argc] == NULL);
4721 Assert(strncmp(argv[1], "--fork", 6) == 0);
4722 Assert(argv[2] == NULL);
4723
4724 /* Resume here if we need to retry */
4725 retry:
4726
4727 /* Set up shared memory for parameter passing */
4728 ZeroMemory(&sa, sizeof(sa));
4729 sa.nLength = sizeof(sa);
4730 sa.bInheritHandle = TRUE;
4731 paramHandle = CreateFileMapping(INVALID_HANDLE_VALUE,
4732 &sa,
4733 PAGE_READWRITE,
4734 0,
4735 sizeof(BackendParameters),
4736 NULL);
4737 if (paramHandle == INVALID_HANDLE_VALUE)
4738 {
4739 elog(LOG, "could not create backend parameter file mapping: error code %lu",
4740 GetLastError());
4741 return -1;
4742 }
4743
4744 param = MapViewOfFile(paramHandle, FILE_MAP_WRITE, 0, 0, sizeof(BackendParameters));
4745 if (!param)
4746 {
4747 elog(LOG, "could not map backend parameter memory: error code %lu",
4748 GetLastError());
4749 CloseHandle(paramHandle);
4750 return -1;
4751 }
4752
4753 /* Insert temp file name after --fork argument */
4754 #ifdef _WIN64
4755 sprintf(paramHandleStr, "%llu", (LONG_PTR) paramHandle);
4756 #else
4757 sprintf(paramHandleStr, "%lu", (DWORD) paramHandle);
4758 #endif
4759 argv[2] = paramHandleStr;
4760
4761 /* Format the cmd line */
4762 cmdLine[sizeof(cmdLine) - 1] = '\0';
4763 cmdLine[sizeof(cmdLine) - 2] = '\0';
4764 snprintf(cmdLine, sizeof(cmdLine) - 1, "\"%s\"", postgres_exec_path);
4765 i = 0;
4766 while (argv[++i] != NULL)
4767 {
4768 j = strlen(cmdLine);
4769 snprintf(cmdLine + j, sizeof(cmdLine) - 1 - j, " \"%s\"", argv[i]);
4770 }
4771 if (cmdLine[sizeof(cmdLine) - 2] != '\0')
4772 {
4773 elog(LOG, "subprocess command line too long");
4774 UnmapViewOfFile(param);
4775 CloseHandle(paramHandle);
4776 return -1;
4777 }
4778
4779 memset(&pi, 0, sizeof(pi));
4780 memset(&si, 0, sizeof(si));
4781 si.cb = sizeof(si);
4782
4783 /*
4784 * Create the subprocess in a suspended state. This will be resumed later,
4785 * once we have written out the parameter file.
4786 */
4787 if (!CreateProcess(NULL, cmdLine, NULL, NULL, TRUE, CREATE_SUSPENDED,
4788 NULL, NULL, &si, &pi))
4789 {
4790 elog(LOG, "CreateProcess call failed: %m (error code %lu)",
4791 GetLastError());
4792 UnmapViewOfFile(param);
4793 CloseHandle(paramHandle);
4794 return -1;
4795 }
4796
4797 if (!save_backend_variables(param, port, pi.hProcess, pi.dwProcessId))
4798 {
4799 /*
4800 * log made by save_backend_variables, but we have to clean up the
4801 * mess with the half-started process
4802 */
4803 if (!TerminateProcess(pi.hProcess, 255))
4804 ereport(LOG,
4805 (errmsg_internal("could not terminate unstarted process: error code %lu",
4806 GetLastError())));
4807 CloseHandle(pi.hProcess);
4808 CloseHandle(pi.hThread);
4809 UnmapViewOfFile(param);
4810 CloseHandle(paramHandle);
4811 return -1; /* log made by save_backend_variables */
4812 }
4813
4814 /* Drop the parameter shared memory that is now inherited to the backend */
4815 if (!UnmapViewOfFile(param))
4816 elog(LOG, "could not unmap view of backend parameter file: error code %lu",
4817 GetLastError());
4818 if (!CloseHandle(paramHandle))
4819 elog(LOG, "could not close handle to backend parameter file: error code %lu",
4820 GetLastError());
4821
4822 /*
4823 * Reserve the memory region used by our main shared memory segment before
4824 * we resume the child process. Normally this should succeed, but if ASLR
4825 * is active then it might sometimes fail due to the stack or heap having
4826 * gotten mapped into that range. In that case, just terminate the
4827 * process and retry.
4828 */
4829 if (!pgwin32_ReserveSharedMemoryRegion(pi.hProcess))
4830 {
4831 /* pgwin32_ReserveSharedMemoryRegion already made a log entry */
4832 if (!TerminateProcess(pi.hProcess, 255))
4833 ereport(LOG,
4834 (errmsg_internal("could not terminate process that failed to reserve memory: error code %lu",
4835 GetLastError())));
4836 CloseHandle(pi.hProcess);
4837 CloseHandle(pi.hThread);
4838 if (++retry_count < 100)
4839 goto retry;
4840 ereport(LOG,
4841 (errmsg("giving up after too many tries to reserve shared memory"),
4842 errhint("This might be caused by ASLR or antivirus software.")));
4843 return -1;
4844 }
4845
4846 /*
4847 * Now that the backend variables are written out, we start the child
4848 * thread so it can start initializing while we set up the rest of the
4849 * parent state.
4850 */
4851 if (ResumeThread(pi.hThread) == -1)
4852 {
4853 if (!TerminateProcess(pi.hProcess, 255))
4854 {
4855 ereport(LOG,
4856 (errmsg_internal("could not terminate unstartable process: error code %lu",
4857 GetLastError())));
4858 CloseHandle(pi.hProcess);
4859 CloseHandle(pi.hThread);
4860 return -1;
4861 }
4862 CloseHandle(pi.hProcess);
4863 CloseHandle(pi.hThread);
4864 ereport(LOG,
4865 (errmsg_internal("could not resume thread of unstarted process: error code %lu",
4866 GetLastError())));
4867 return -1;
4868 }
4869
4870 /*
4871 * Queue a waiter to signal when this child dies. The wait will be handled
4872 * automatically by an operating system thread pool.
4873 *
4874 * Note: use malloc instead of palloc, since it needs to be thread-safe.
4875 * Struct will be free():d from the callback function that runs on a
4876 * different thread.
4877 */
4878 childinfo = malloc(sizeof(win32_deadchild_waitinfo));
4879 if (!childinfo)
4880 ereport(FATAL,
4881 (errcode(ERRCODE_OUT_OF_MEMORY),
4882 errmsg("out of memory")));
4883
4884 childinfo->procHandle = pi.hProcess;
4885 childinfo->procId = pi.dwProcessId;
4886
4887 if (!RegisterWaitForSingleObject(&childinfo->waitHandle,
4888 pi.hProcess,
4889 pgwin32_deadchild_callback,
4890 childinfo,
4891 INFINITE,
4892 WT_EXECUTEONLYONCE | WT_EXECUTEINWAITTHREAD))
4893 ereport(FATAL,
4894 (errmsg_internal("could not register process for wait: error code %lu",
4895 GetLastError())));
4896
4897 /* Don't close pi.hProcess here - the wait thread needs access to it */
4898
4899 CloseHandle(pi.hThread);
4900
4901 return pi.dwProcessId;
4902 }
4903 #endif /* WIN32 */
4904
4905
4906 /*
4907 * SubPostmasterMain -- Get the fork/exec'd process into a state equivalent
4908 * to what it would be if we'd simply forked on Unix, and then
4909 * dispatch to the appropriate place.
4910 *
4911 * The first two command line arguments are expected to be "--forkFOO"
4912 * (where FOO indicates which postmaster child we are to become), and
4913 * the name of a variables file that we can read to load data that would
4914 * have been inherited by fork() on Unix. Remaining arguments go to the
4915 * subprocess FooMain() routine.
4916 */
4917 void
SubPostmasterMain(int argc,char * argv[])4918 SubPostmasterMain(int argc, char *argv[])
4919 {
4920 Port port;
4921
4922 /* In EXEC_BACKEND case we will not have inherited these settings */
4923 IsPostmasterEnvironment = true;
4924 whereToSendOutput = DestNone;
4925
4926 /* Setup as postmaster child */
4927 InitPostmasterChild();
4928
4929 /* Setup essential subsystems (to ensure elog() behaves sanely) */
4930 InitializeGUCOptions();
4931
4932 /* Check we got appropriate args */
4933 if (argc < 3)
4934 elog(FATAL, "invalid subpostmaster invocation");
4935
4936 /* Read in the variables file */
4937 memset(&port, 0, sizeof(Port));
4938 read_backend_variables(argv[2], &port);
4939
4940 /* Close the postmaster's sockets (as soon as we know them) */
4941 ClosePostmasterPorts(strcmp(argv[1], "--forklog") == 0);
4942
4943 /*
4944 * If appropriate, physically re-attach to shared memory segment. We want
4945 * to do this before going any further to ensure that we can attach at the
4946 * same address the postmaster used. On the other hand, if we choose not
4947 * to re-attach, we may have other cleanup to do.
4948 *
4949 * If testing EXEC_BACKEND on Linux, you should run this as root before
4950 * starting the postmaster:
4951 *
4952 * echo 0 >/proc/sys/kernel/randomize_va_space
4953 *
4954 * This prevents using randomized stack and code addresses that cause the
4955 * child process's memory map to be different from the parent's, making it
4956 * sometimes impossible to attach to shared memory at the desired address.
4957 * Return the setting to its old value (usually '1' or '2') when finished.
4958 */
4959 if (strcmp(argv[1], "--forkbackend") == 0 ||
4960 strcmp(argv[1], "--forkavlauncher") == 0 ||
4961 strcmp(argv[1], "--forkavworker") == 0 ||
4962 strcmp(argv[1], "--forkboot") == 0 ||
4963 strncmp(argv[1], "--forkbgworker=", 15) == 0)
4964 PGSharedMemoryReAttach();
4965 else
4966 PGSharedMemoryNoReAttach();
4967
4968 /* autovacuum needs this set before calling InitProcess */
4969 if (strcmp(argv[1], "--forkavlauncher") == 0)
4970 AutovacuumLauncherIAm();
4971 if (strcmp(argv[1], "--forkavworker") == 0)
4972 AutovacuumWorkerIAm();
4973
4974 /*
4975 * Start our win32 signal implementation. This has to be done after we
4976 * read the backend variables, because we need to pick up the signal pipe
4977 * from the parent process.
4978 */
4979 #ifdef WIN32
4980 pgwin32_signal_initialize();
4981 #endif
4982
4983 /* In EXEC_BACKEND case we will not have inherited these settings */
4984 pqinitmask();
4985 PG_SETMASK(&BlockSig);
4986
4987 /* Read in remaining GUC variables */
4988 read_nondefault_variables();
4989
4990 /*
4991 * Check that the data directory looks valid, which will also check the
4992 * privileges on the data directory and update our umask and file/group
4993 * variables for creating files later. Note: this should really be done
4994 * before we create any files or directories.
4995 */
4996 checkDataDir();
4997
4998 /*
4999 * (re-)read control file, as it contains config. The postmaster will
5000 * already have read this, but this process doesn't know about that.
5001 */
5002 LocalProcessControlFile(false);
5003
5004 /*
5005 * Reload any libraries that were preloaded by the postmaster. Since we
5006 * exec'd this process, those libraries didn't come along with us; but we
5007 * should load them into all child processes to be consistent with the
5008 * non-EXEC_BACKEND behavior.
5009 */
5010 process_shared_preload_libraries();
5011
5012 /* Run backend or appropriate child */
5013 if (strcmp(argv[1], "--forkbackend") == 0)
5014 {
5015 Assert(argc == 3); /* shouldn't be any more args */
5016
5017 /*
5018 * Need to reinitialize the SSL library in the backend, since the
5019 * context structures contain function pointers and cannot be passed
5020 * through the parameter file.
5021 *
5022 * If for some reason reload fails (maybe the user installed broken
5023 * key files), soldier on without SSL; that's better than all
5024 * connections becoming impossible.
5025 *
5026 * XXX should we do this in all child processes? For the moment it's
5027 * enough to do it in backend children.
5028 */
5029 #ifdef USE_SSL
5030 if (EnableSSL)
5031 {
5032 if (secure_initialize(false) == 0)
5033 LoadedSSL = true;
5034 else
5035 ereport(LOG,
5036 (errmsg("SSL configuration could not be loaded in child process")));
5037 }
5038 #endif
5039
5040 /*
5041 * Perform additional initialization and collect startup packet.
5042 *
5043 * We want to do this before InitProcess() for a couple of reasons: 1.
5044 * so that we aren't eating up a PGPROC slot while waiting on the
5045 * client. 2. so that if InitProcess() fails due to being out of
5046 * PGPROC slots, we have already initialized libpq and are able to
5047 * report the error to the client.
5048 */
5049 BackendInitialize(&port);
5050
5051 /* Restore basic shared memory pointers */
5052 InitShmemAccess(UsedShmemSegAddr);
5053
5054 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5055 InitProcess();
5056
5057 /* Attach process to shared data structures */
5058 CreateSharedMemoryAndSemaphores();
5059
5060 /* And run the backend */
5061 BackendRun(&port); /* does not return */
5062 }
5063 if (strcmp(argv[1], "--forkboot") == 0)
5064 {
5065 /* Restore basic shared memory pointers */
5066 InitShmemAccess(UsedShmemSegAddr);
5067
5068 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5069 InitAuxiliaryProcess();
5070
5071 /* Attach process to shared data structures */
5072 CreateSharedMemoryAndSemaphores();
5073
5074 AuxiliaryProcessMain(argc - 2, argv + 2); /* does not return */
5075 }
5076 if (strcmp(argv[1], "--forkavlauncher") == 0)
5077 {
5078 /* Restore basic shared memory pointers */
5079 InitShmemAccess(UsedShmemSegAddr);
5080
5081 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5082 InitProcess();
5083
5084 /* Attach process to shared data structures */
5085 CreateSharedMemoryAndSemaphores();
5086
5087 AutoVacLauncherMain(argc - 2, argv + 2); /* does not return */
5088 }
5089 if (strcmp(argv[1], "--forkavworker") == 0)
5090 {
5091 /* Restore basic shared memory pointers */
5092 InitShmemAccess(UsedShmemSegAddr);
5093
5094 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5095 InitProcess();
5096
5097 /* Attach process to shared data structures */
5098 CreateSharedMemoryAndSemaphores();
5099
5100 AutoVacWorkerMain(argc - 2, argv + 2); /* does not return */
5101 }
5102 if (strncmp(argv[1], "--forkbgworker=", 15) == 0)
5103 {
5104 int shmem_slot;
5105
5106 /* do this as early as possible; in particular, before InitProcess() */
5107 IsBackgroundWorker = true;
5108
5109 /* Restore basic shared memory pointers */
5110 InitShmemAccess(UsedShmemSegAddr);
5111
5112 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
5113 InitProcess();
5114
5115 /* Attach process to shared data structures */
5116 CreateSharedMemoryAndSemaphores();
5117
5118 /* Fetch MyBgworkerEntry from shared memory */
5119 shmem_slot = atoi(argv[1] + 15);
5120 MyBgworkerEntry = BackgroundWorkerEntry(shmem_slot);
5121
5122 StartBackgroundWorker();
5123 }
5124 if (strcmp(argv[1], "--forkarch") == 0)
5125 {
5126 /* Do not want to attach to shared memory */
5127
5128 PgArchiverMain(argc, argv); /* does not return */
5129 }
5130 if (strcmp(argv[1], "--forkcol") == 0)
5131 {
5132 /* Do not want to attach to shared memory */
5133
5134 PgstatCollectorMain(argc, argv); /* does not return */
5135 }
5136 if (strcmp(argv[1], "--forklog") == 0)
5137 {
5138 /* Do not want to attach to shared memory */
5139
5140 SysLoggerMain(argc, argv); /* does not return */
5141 }
5142
5143 abort(); /* shouldn't get here */
5144 }
5145 #endif /* EXEC_BACKEND */
5146
5147
5148 /*
5149 * ExitPostmaster -- cleanup
5150 *
5151 * Do NOT call exit() directly --- always go through here!
5152 */
5153 static void
ExitPostmaster(int status)5154 ExitPostmaster(int status)
5155 {
5156 #ifdef HAVE_PTHREAD_IS_THREADED_NP
5157
5158 /*
5159 * There is no known cause for a postmaster to become multithreaded after
5160 * startup. Recheck to account for the possibility of unknown causes.
5161 * This message uses LOG level, because an unclean shutdown at this point
5162 * would usually not look much different from a clean shutdown.
5163 */
5164 if (pthread_is_threaded_np() != 0)
5165 ereport(LOG,
5166 (errcode(ERRCODE_INTERNAL_ERROR),
5167 errmsg_internal("postmaster became multithreaded"),
5168 errdetail("Please report this to <%s>.", PACKAGE_BUGREPORT)));
5169 #endif
5170
5171 /* should cleanup shared memory and kill all backends */
5172
5173 /*
5174 * Not sure of the semantics here. When the Postmaster dies, should the
5175 * backends all be killed? probably not.
5176 *
5177 * MUST -- vadim 05-10-1999
5178 */
5179
5180 proc_exit(status);
5181 }
5182
5183 /*
5184 * sigusr1_handler - handle signal conditions from child processes
5185 */
5186 static void
sigusr1_handler(SIGNAL_ARGS)5187 sigusr1_handler(SIGNAL_ARGS)
5188 {
5189 int save_errno = errno;
5190
5191 /*
5192 * We rely on the signal mechanism to have blocked all signals ... except
5193 * on Windows, which lacks sigaction(), so we have to do it manually.
5194 */
5195 #ifdef WIN32
5196 PG_SETMASK(&BlockSig);
5197 #endif
5198
5199 /*
5200 * RECOVERY_STARTED and BEGIN_HOT_STANDBY signals are ignored in
5201 * unexpected states. If the startup process quickly starts up, completes
5202 * recovery, exits, we might process the death of the startup process
5203 * first. We don't want to go back to recovery in that case.
5204 */
5205 if (CheckPostmasterSignal(PMSIGNAL_RECOVERY_STARTED) &&
5206 pmState == PM_STARTUP && Shutdown == NoShutdown)
5207 {
5208 /* WAL redo has started. We're out of reinitialization. */
5209 FatalError = false;
5210 AbortStartTime = 0;
5211
5212 /*
5213 * Crank up the background tasks. It doesn't matter if this fails,
5214 * we'll just try again later.
5215 */
5216 Assert(CheckpointerPID == 0);
5217 CheckpointerPID = StartCheckpointer();
5218 Assert(BgWriterPID == 0);
5219 BgWriterPID = StartBackgroundWriter();
5220
5221 /*
5222 * Start the archiver if we're responsible for (re-)archiving received
5223 * files.
5224 */
5225 Assert(PgArchPID == 0);
5226 if (XLogArchivingAlways())
5227 PgArchPID = pgarch_start();
5228
5229 /*
5230 * If we aren't planning to enter hot standby mode later, treat
5231 * RECOVERY_STARTED as meaning we're out of startup, and report status
5232 * accordingly.
5233 */
5234 if (!EnableHotStandby)
5235 {
5236 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STANDBY);
5237 #ifdef USE_SYSTEMD
5238 sd_notify(0, "READY=1");
5239 #endif
5240 }
5241
5242 pmState = PM_RECOVERY;
5243 }
5244
5245 if (CheckPostmasterSignal(PMSIGNAL_BEGIN_HOT_STANDBY) &&
5246 pmState == PM_RECOVERY && Shutdown == NoShutdown)
5247 {
5248 /*
5249 * Likewise, start other special children as needed.
5250 */
5251 Assert(PgStatPID == 0);
5252 PgStatPID = pgstat_start();
5253
5254 ereport(LOG,
5255 (errmsg("database system is ready to accept read only connections")));
5256
5257 /* Report status */
5258 AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
5259 #ifdef USE_SYSTEMD
5260 sd_notify(0, "READY=1");
5261 #endif
5262
5263 pmState = PM_HOT_STANDBY;
5264 connsAllowed = ALLOW_ALL_CONNS;
5265
5266 /* Some workers may be scheduled to start now */
5267 StartWorkerNeeded = true;
5268 }
5269
5270 /* Process background worker state changes. */
5271 if (CheckPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE))
5272 {
5273 /* Accept new worker requests only if not stopping. */
5274 BackgroundWorkerStateChange(pmState < PM_STOP_BACKENDS);
5275 StartWorkerNeeded = true;
5276 }
5277
5278 if (StartWorkerNeeded || HaveCrashedWorker)
5279 maybe_start_bgworkers();
5280
5281 if (CheckPostmasterSignal(PMSIGNAL_WAKEN_ARCHIVER) &&
5282 PgArchPID != 0)
5283 {
5284 /*
5285 * Send SIGUSR1 to archiver process, to wake it up and begin archiving
5286 * next WAL file.
5287 */
5288 signal_child(PgArchPID, SIGUSR1);
5289 }
5290
5291 /* Tell syslogger to rotate logfile if requested */
5292 if (SysLoggerPID != 0)
5293 {
5294 if (CheckLogrotateSignal())
5295 {
5296 signal_child(SysLoggerPID, SIGUSR1);
5297 RemoveLogrotateSignalFiles();
5298 }
5299 else if (CheckPostmasterSignal(PMSIGNAL_ROTATE_LOGFILE))
5300 {
5301 signal_child(SysLoggerPID, SIGUSR1);
5302 }
5303 }
5304
5305 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER) &&
5306 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5307 {
5308 /*
5309 * Start one iteration of the autovacuum daemon, even if autovacuuming
5310 * is nominally not enabled. This is so we can have an active defense
5311 * against transaction ID wraparound. We set a flag for the main loop
5312 * to do it rather than trying to do it here --- this is because the
5313 * autovac process itself may send the signal, and we want to handle
5314 * that by launching another iteration as soon as the current one
5315 * completes.
5316 */
5317 start_autovac_launcher = true;
5318 }
5319
5320 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER) &&
5321 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5322 {
5323 /* The autovacuum launcher wants us to start a worker process. */
5324 StartAutovacuumWorker();
5325 }
5326
5327 if (CheckPostmasterSignal(PMSIGNAL_START_WALRECEIVER))
5328 {
5329 /* Startup Process wants us to start the walreceiver process. */
5330 /* Start immediately if possible, else remember request for later. */
5331 WalReceiverRequested = true;
5332 MaybeStartWalReceiver();
5333 }
5334
5335 /*
5336 * Try to advance postmaster's state machine, if a child requests it.
5337 *
5338 * Be careful about the order of this action relative to sigusr1_handler's
5339 * other actions. Generally, this should be after other actions, in case
5340 * they have effects PostmasterStateMachine would need to know about.
5341 * However, we should do it before the CheckPromoteSignal step, which
5342 * cannot have any (immediate) effect on the state machine, but does
5343 * depend on what state we're in now.
5344 */
5345 if (CheckPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE))
5346 {
5347 PostmasterStateMachine();
5348 }
5349
5350 if (StartupPID != 0 &&
5351 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5352 pmState == PM_HOT_STANDBY) &&
5353 CheckPromoteSignal())
5354 {
5355 /* Tell startup process to finish recovery */
5356 signal_child(StartupPID, SIGUSR2);
5357 }
5358
5359 #ifdef WIN32
5360 PG_SETMASK(&UnBlockSig);
5361 #endif
5362
5363 errno = save_errno;
5364 }
5365
5366 /*
5367 * SIGTERM while processing startup packet.
5368 * Clean up and exit(1).
5369 *
5370 * Running proc_exit() from a signal handler is pretty unsafe, since we
5371 * can't know what code we've interrupted. But the alternative of using
5372 * _exit(2) is also unpalatable, since it'd mean that a "fast shutdown"
5373 * would cause a database crash cycle (forcing WAL replay at restart)
5374 * if any sessions are in authentication. So we live with it for now.
5375 *
5376 * One might be tempted to try to send a message indicating why we are
5377 * disconnecting. However, that would make this even more unsafe. Also,
5378 * it seems undesirable to provide clues about the database's state to
5379 * a client that has not yet completed authentication.
5380 */
5381 static void
process_startup_packet_die(SIGNAL_ARGS)5382 process_startup_packet_die(SIGNAL_ARGS)
5383 {
5384 proc_exit(1);
5385 }
5386
5387 /*
5388 * Dummy signal handler
5389 *
5390 * We use this for signals that we don't actually use in the postmaster,
5391 * but we do use in backends. If we were to SIG_IGN such signals in the
5392 * postmaster, then a newly started backend might drop a signal that arrives
5393 * before it's able to reconfigure its signal processing. (See notes in
5394 * tcop/postgres.c.)
5395 */
5396 static void
dummy_handler(SIGNAL_ARGS)5397 dummy_handler(SIGNAL_ARGS)
5398 {
5399 }
5400
5401 /*
5402 * Timeout while processing startup packet.
5403 * As for process_startup_packet_die(), we clean up and exit(1).
5404 *
5405 * This is theoretically just as hazardous as in process_startup_packet_die(),
5406 * although in practice we're almost certainly waiting for client input,
5407 * which greatly reduces the risk.
5408 */
5409 static void
StartupPacketTimeoutHandler(void)5410 StartupPacketTimeoutHandler(void)
5411 {
5412 proc_exit(1);
5413 }
5414
5415
5416 /*
5417 * Generate a random cancel key.
5418 */
5419 static bool
RandomCancelKey(int32 * cancel_key)5420 RandomCancelKey(int32 *cancel_key)
5421 {
5422 return pg_strong_random(cancel_key, sizeof(int32));
5423 }
5424
5425 /*
5426 * Count up number of child processes of specified types (dead_end children
5427 * are always excluded).
5428 */
5429 static int
CountChildren(int target)5430 CountChildren(int target)
5431 {
5432 dlist_iter iter;
5433 int cnt = 0;
5434
5435 dlist_foreach(iter, &BackendList)
5436 {
5437 Backend *bp = dlist_container(Backend, elem, iter.cur);
5438
5439 if (bp->dead_end)
5440 continue;
5441
5442 /*
5443 * Since target == BACKEND_TYPE_ALL is the most common case, we test
5444 * it first and avoid touching shared memory for every child.
5445 */
5446 if (target != BACKEND_TYPE_ALL)
5447 {
5448 /*
5449 * Assign bkend_type for any recently announced WAL Sender
5450 * processes.
5451 */
5452 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
5453 IsPostmasterChildWalSender(bp->child_slot))
5454 bp->bkend_type = BACKEND_TYPE_WALSND;
5455
5456 if (!(target & bp->bkend_type))
5457 continue;
5458 }
5459
5460 cnt++;
5461 }
5462 return cnt;
5463 }
5464
5465
5466 /*
5467 * StartChildProcess -- start an auxiliary process for the postmaster
5468 *
5469 * "type" determines what kind of child will be started. All child types
5470 * initially go to AuxiliaryProcessMain, which will handle common setup.
5471 *
5472 * Return value of StartChildProcess is subprocess' PID, or 0 if failed
5473 * to start subprocess.
5474 */
5475 static pid_t
StartChildProcess(AuxProcType type)5476 StartChildProcess(AuxProcType type)
5477 {
5478 pid_t pid;
5479 char *av[10];
5480 int ac = 0;
5481 char typebuf[32];
5482
5483 /*
5484 * Set up command-line arguments for subprocess
5485 */
5486 av[ac++] = "postgres";
5487
5488 #ifdef EXEC_BACKEND
5489 av[ac++] = "--forkboot";
5490 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5491 #endif
5492
5493 snprintf(typebuf, sizeof(typebuf), "-x%d", type);
5494 av[ac++] = typebuf;
5495
5496 av[ac] = NULL;
5497 Assert(ac < lengthof(av));
5498
5499 #ifdef EXEC_BACKEND
5500 pid = postmaster_forkexec(ac, av);
5501 #else /* !EXEC_BACKEND */
5502 pid = fork_process();
5503
5504 if (pid == 0) /* child */
5505 {
5506 InitPostmasterChild();
5507
5508 /* Close the postmaster's sockets */
5509 ClosePostmasterPorts(false);
5510
5511 /* Release postmaster's working memory context */
5512 MemoryContextSwitchTo(TopMemoryContext);
5513 MemoryContextDelete(PostmasterContext);
5514 PostmasterContext = NULL;
5515
5516 AuxiliaryProcessMain(ac, av);
5517 ExitPostmaster(0);
5518 }
5519 #endif /* EXEC_BACKEND */
5520
5521 if (pid < 0)
5522 {
5523 /* in parent, fork failed */
5524 int save_errno = errno;
5525
5526 errno = save_errno;
5527 switch (type)
5528 {
5529 case StartupProcess:
5530 ereport(LOG,
5531 (errmsg("could not fork startup process: %m")));
5532 break;
5533 case BgWriterProcess:
5534 ereport(LOG,
5535 (errmsg("could not fork background writer process: %m")));
5536 break;
5537 case CheckpointerProcess:
5538 ereport(LOG,
5539 (errmsg("could not fork checkpointer process: %m")));
5540 break;
5541 case WalWriterProcess:
5542 ereport(LOG,
5543 (errmsg("could not fork WAL writer process: %m")));
5544 break;
5545 case WalReceiverProcess:
5546 ereport(LOG,
5547 (errmsg("could not fork WAL receiver process: %m")));
5548 break;
5549 default:
5550 ereport(LOG,
5551 (errmsg("could not fork process: %m")));
5552 break;
5553 }
5554
5555 /*
5556 * fork failure is fatal during startup, but there's no need to choke
5557 * immediately if starting other child types fails.
5558 */
5559 if (type == StartupProcess)
5560 ExitPostmaster(1);
5561 return 0;
5562 }
5563
5564 /*
5565 * in parent, successful fork
5566 */
5567 return pid;
5568 }
5569
5570 /*
5571 * StartAutovacuumWorker
5572 * Start an autovac worker process.
5573 *
5574 * This function is here because it enters the resulting PID into the
5575 * postmaster's private backends list.
5576 *
5577 * NB -- this code very roughly matches BackendStartup.
5578 */
5579 static void
StartAutovacuumWorker(void)5580 StartAutovacuumWorker(void)
5581 {
5582 Backend *bn;
5583
5584 /*
5585 * If not in condition to run a process, don't try, but handle it like a
5586 * fork failure. This does not normally happen, since the signal is only
5587 * supposed to be sent by autovacuum launcher when it's OK to do it, but
5588 * we have to check to avoid race-condition problems during DB state
5589 * changes.
5590 */
5591 if (canAcceptConnections(BACKEND_TYPE_AUTOVAC) == CAC_OK)
5592 {
5593 /*
5594 * Compute the cancel key that will be assigned to this session. We
5595 * probably don't need cancel keys for autovac workers, but we'd
5596 * better have something random in the field to prevent unfriendly
5597 * people from sending cancels to them.
5598 */
5599 if (!RandomCancelKey(&MyCancelKey))
5600 {
5601 ereport(LOG,
5602 (errcode(ERRCODE_INTERNAL_ERROR),
5603 errmsg("could not generate random cancel key")));
5604 return;
5605 }
5606
5607 bn = (Backend *) malloc(sizeof(Backend));
5608 if (bn)
5609 {
5610 bn->cancel_key = MyCancelKey;
5611
5612 /* Autovac workers are not dead_end and need a child slot */
5613 bn->dead_end = false;
5614 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5615 bn->bgworker_notify = false;
5616
5617 bn->pid = StartAutoVacWorker();
5618 if (bn->pid > 0)
5619 {
5620 bn->bkend_type = BACKEND_TYPE_AUTOVAC;
5621 dlist_push_head(&BackendList, &bn->elem);
5622 #ifdef EXEC_BACKEND
5623 ShmemBackendArrayAdd(bn);
5624 #endif
5625 /* all OK */
5626 return;
5627 }
5628
5629 /*
5630 * fork failed, fall through to report -- actual error message was
5631 * logged by StartAutoVacWorker
5632 */
5633 (void) ReleasePostmasterChildSlot(bn->child_slot);
5634 free(bn);
5635 }
5636 else
5637 ereport(LOG,
5638 (errcode(ERRCODE_OUT_OF_MEMORY),
5639 errmsg("out of memory")));
5640 }
5641
5642 /*
5643 * Report the failure to the launcher, if it's running. (If it's not, we
5644 * might not even be connected to shared memory, so don't try to call
5645 * AutoVacWorkerFailed.) Note that we also need to signal it so that it
5646 * responds to the condition, but we don't do that here, instead waiting
5647 * for ServerLoop to do it. This way we avoid a ping-pong signaling in
5648 * quick succession between the autovac launcher and postmaster in case
5649 * things get ugly.
5650 */
5651 if (AutoVacPID != 0)
5652 {
5653 AutoVacWorkerFailed();
5654 avlauncher_needs_signal = true;
5655 }
5656 }
5657
5658 /*
5659 * MaybeStartWalReceiver
5660 * Start the WAL receiver process, if not running and our state allows.
5661 *
5662 * Note: if WalReceiverPID is already nonzero, it might seem that we should
5663 * clear WalReceiverRequested. However, there's a race condition if the
5664 * walreceiver terminates and the startup process immediately requests a new
5665 * one: it's quite possible to get the signal for the request before reaping
5666 * the dead walreceiver process. Better to risk launching an extra
5667 * walreceiver than to miss launching one we need. (The walreceiver code
5668 * has logic to recognize that it should go away if not needed.)
5669 */
5670 static void
MaybeStartWalReceiver(void)5671 MaybeStartWalReceiver(void)
5672 {
5673 if (WalReceiverPID == 0 &&
5674 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5675 pmState == PM_HOT_STANDBY) &&
5676 Shutdown <= SmartShutdown)
5677 {
5678 WalReceiverPID = StartWalReceiver();
5679 if (WalReceiverPID != 0)
5680 WalReceiverRequested = false;
5681 /* else leave the flag set, so we'll try again later */
5682 }
5683 }
5684
5685
5686 /*
5687 * Create the opts file
5688 */
5689 static bool
CreateOptsFile(int argc,char * argv[],char * fullprogname)5690 CreateOptsFile(int argc, char *argv[], char *fullprogname)
5691 {
5692 FILE *fp;
5693 int i;
5694
5695 #define OPTS_FILE "postmaster.opts"
5696
5697 if ((fp = fopen(OPTS_FILE, "w")) == NULL)
5698 {
5699 elog(LOG, "could not create file \"%s\": %m", OPTS_FILE);
5700 return false;
5701 }
5702
5703 fprintf(fp, "%s", fullprogname);
5704 for (i = 1; i < argc; i++)
5705 fprintf(fp, " \"%s\"", argv[i]);
5706 fputs("\n", fp);
5707
5708 if (fclose(fp))
5709 {
5710 elog(LOG, "could not write file \"%s\": %m", OPTS_FILE);
5711 return false;
5712 }
5713
5714 return true;
5715 }
5716
5717
5718 /*
5719 * MaxLivePostmasterChildren
5720 *
5721 * This reports the number of entries needed in per-child-process arrays
5722 * (the PMChildFlags array, and if EXEC_BACKEND the ShmemBackendArray).
5723 * These arrays include regular backends, autovac workers, walsenders
5724 * and background workers, but not special children nor dead_end children.
5725 * This allows the arrays to have a fixed maximum size, to wit the same
5726 * too-many-children limit enforced by canAcceptConnections(). The exact value
5727 * isn't too critical as long as it's more than MaxBackends.
5728 */
5729 int
MaxLivePostmasterChildren(void)5730 MaxLivePostmasterChildren(void)
5731 {
5732 return 2 * (MaxConnections + autovacuum_max_workers + 1 +
5733 max_wal_senders + max_worker_processes);
5734 }
5735
5736 /*
5737 * Connect background worker to a database.
5738 */
5739 void
BackgroundWorkerInitializeConnection(const char * dbname,const char * username,uint32 flags)5740 BackgroundWorkerInitializeConnection(const char *dbname, const char *username, uint32 flags)
5741 {
5742 BackgroundWorker *worker = MyBgworkerEntry;
5743
5744 /* XXX is this the right errcode? */
5745 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5746 ereport(FATAL,
5747 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5748 errmsg("database connection requirement not indicated during registration")));
5749
5750 InitPostgres(dbname, InvalidOid, username, InvalidOid, NULL, (flags & BGWORKER_BYPASS_ALLOWCONN) != 0);
5751
5752 /* it had better not gotten out of "init" mode yet */
5753 if (!IsInitProcessingMode())
5754 ereport(ERROR,
5755 (errmsg("invalid processing mode in background worker")));
5756 SetProcessingMode(NormalProcessing);
5757 }
5758
5759 /*
5760 * Connect background worker to a database using OIDs.
5761 */
5762 void
BackgroundWorkerInitializeConnectionByOid(Oid dboid,Oid useroid,uint32 flags)5763 BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid, uint32 flags)
5764 {
5765 BackgroundWorker *worker = MyBgworkerEntry;
5766
5767 /* XXX is this the right errcode? */
5768 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5769 ereport(FATAL,
5770 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5771 errmsg("database connection requirement not indicated during registration")));
5772
5773 InitPostgres(NULL, dboid, NULL, useroid, NULL, (flags & BGWORKER_BYPASS_ALLOWCONN) != 0);
5774
5775 /* it had better not gotten out of "init" mode yet */
5776 if (!IsInitProcessingMode())
5777 ereport(ERROR,
5778 (errmsg("invalid processing mode in background worker")));
5779 SetProcessingMode(NormalProcessing);
5780 }
5781
5782 /*
5783 * Block/unblock signals in a background worker
5784 */
5785 void
BackgroundWorkerBlockSignals(void)5786 BackgroundWorkerBlockSignals(void)
5787 {
5788 PG_SETMASK(&BlockSig);
5789 }
5790
5791 void
BackgroundWorkerUnblockSignals(void)5792 BackgroundWorkerUnblockSignals(void)
5793 {
5794 PG_SETMASK(&UnBlockSig);
5795 }
5796
5797 #ifdef EXEC_BACKEND
5798 static pid_t
bgworker_forkexec(int shmem_slot)5799 bgworker_forkexec(int shmem_slot)
5800 {
5801 char *av[10];
5802 int ac = 0;
5803 char forkav[MAXPGPATH];
5804
5805 snprintf(forkav, MAXPGPATH, "--forkbgworker=%d", shmem_slot);
5806
5807 av[ac++] = "postgres";
5808 av[ac++] = forkav;
5809 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5810 av[ac] = NULL;
5811
5812 Assert(ac < lengthof(av));
5813
5814 return postmaster_forkexec(ac, av);
5815 }
5816 #endif
5817
5818 /*
5819 * Start a new bgworker.
5820 * Starting time conditions must have been checked already.
5821 *
5822 * Returns true on success, false on failure.
5823 * In either case, update the RegisteredBgWorker's state appropriately.
5824 *
5825 * This code is heavily based on autovacuum.c, q.v.
5826 */
5827 static bool
do_start_bgworker(RegisteredBgWorker * rw)5828 do_start_bgworker(RegisteredBgWorker *rw)
5829 {
5830 pid_t worker_pid;
5831
5832 Assert(rw->rw_pid == 0);
5833
5834 /*
5835 * Allocate and assign the Backend element. Note we must do this before
5836 * forking, so that we can handle failures (out of memory or child-process
5837 * slots) cleanly.
5838 *
5839 * Treat failure as though the worker had crashed. That way, the
5840 * postmaster will wait a bit before attempting to start it again; if we
5841 * tried again right away, most likely we'd find ourselves hitting the
5842 * same resource-exhaustion condition.
5843 */
5844 if (!assign_backendlist_entry(rw))
5845 {
5846 rw->rw_crashed_at = GetCurrentTimestamp();
5847 return false;
5848 }
5849
5850 ereport(DEBUG1,
5851 (errmsg("starting background worker process \"%s\"",
5852 rw->rw_worker.bgw_name)));
5853
5854 #ifdef EXEC_BACKEND
5855 switch ((worker_pid = bgworker_forkexec(rw->rw_shmem_slot)))
5856 #else
5857 switch ((worker_pid = fork_process()))
5858 #endif
5859 {
5860 case -1:
5861 /* in postmaster, fork failed ... */
5862 ereport(LOG,
5863 (errmsg("could not fork worker process: %m")));
5864 /* undo what assign_backendlist_entry did */
5865 ReleasePostmasterChildSlot(rw->rw_child_slot);
5866 rw->rw_child_slot = 0;
5867 free(rw->rw_backend);
5868 rw->rw_backend = NULL;
5869 /* mark entry as crashed, so we'll try again later */
5870 rw->rw_crashed_at = GetCurrentTimestamp();
5871 break;
5872
5873 #ifndef EXEC_BACKEND
5874 case 0:
5875 /* in postmaster child ... */
5876 InitPostmasterChild();
5877
5878 /* Close the postmaster's sockets */
5879 ClosePostmasterPorts(false);
5880
5881 /*
5882 * Before blowing away PostmasterContext, save this bgworker's
5883 * data where it can find it.
5884 */
5885 MyBgworkerEntry = (BackgroundWorker *)
5886 MemoryContextAlloc(TopMemoryContext, sizeof(BackgroundWorker));
5887 memcpy(MyBgworkerEntry, &rw->rw_worker, sizeof(BackgroundWorker));
5888
5889 /* Release postmaster's working memory context */
5890 MemoryContextSwitchTo(TopMemoryContext);
5891 MemoryContextDelete(PostmasterContext);
5892 PostmasterContext = NULL;
5893
5894 StartBackgroundWorker();
5895
5896 exit(1); /* should not get here */
5897 break;
5898 #endif
5899 default:
5900 /* in postmaster, fork successful ... */
5901 rw->rw_pid = worker_pid;
5902 rw->rw_backend->pid = rw->rw_pid;
5903 ReportBackgroundWorkerPID(rw);
5904 /* add new worker to lists of backends */
5905 dlist_push_head(&BackendList, &rw->rw_backend->elem);
5906 #ifdef EXEC_BACKEND
5907 ShmemBackendArrayAdd(rw->rw_backend);
5908 #endif
5909 return true;
5910 }
5911
5912 return false;
5913 }
5914
5915 /*
5916 * Does the current postmaster state require starting a worker with the
5917 * specified start_time?
5918 */
5919 static bool
bgworker_should_start_now(BgWorkerStartTime start_time)5920 bgworker_should_start_now(BgWorkerStartTime start_time)
5921 {
5922 switch (pmState)
5923 {
5924 case PM_NO_CHILDREN:
5925 case PM_WAIT_DEAD_END:
5926 case PM_SHUTDOWN_2:
5927 case PM_SHUTDOWN:
5928 case PM_WAIT_BACKENDS:
5929 case PM_STOP_BACKENDS:
5930 break;
5931
5932 case PM_RUN:
5933 if (start_time == BgWorkerStart_RecoveryFinished)
5934 return true;
5935 /* fall through */
5936
5937 case PM_HOT_STANDBY:
5938 if (start_time == BgWorkerStart_ConsistentState)
5939 return true;
5940 /* fall through */
5941
5942 case PM_RECOVERY:
5943 case PM_STARTUP:
5944 case PM_INIT:
5945 if (start_time == BgWorkerStart_PostmasterStart)
5946 return true;
5947 /* fall through */
5948
5949 }
5950
5951 return false;
5952 }
5953
5954 /*
5955 * Allocate the Backend struct for a connected background worker, but don't
5956 * add it to the list of backends just yet.
5957 *
5958 * On failure, return false without changing any worker state.
5959 *
5960 * Some info from the Backend is copied into the passed rw.
5961 */
5962 static bool
assign_backendlist_entry(RegisteredBgWorker * rw)5963 assign_backendlist_entry(RegisteredBgWorker *rw)
5964 {
5965 Backend *bn;
5966
5967 /*
5968 * Check that database state allows another connection. Currently the
5969 * only possible failure is CAC_TOOMANY, so we just log an error message
5970 * based on that rather than checking the error code precisely.
5971 */
5972 if (canAcceptConnections(BACKEND_TYPE_BGWORKER) != CAC_OK)
5973 {
5974 ereport(LOG,
5975 (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
5976 errmsg("no slot available for new worker process")));
5977 return false;
5978 }
5979
5980 /*
5981 * Compute the cancel key that will be assigned to this session. We
5982 * probably don't need cancel keys for background workers, but we'd better
5983 * have something random in the field to prevent unfriendly people from
5984 * sending cancels to them.
5985 */
5986 if (!RandomCancelKey(&MyCancelKey))
5987 {
5988 ereport(LOG,
5989 (errcode(ERRCODE_INTERNAL_ERROR),
5990 errmsg("could not generate random cancel key")));
5991 return false;
5992 }
5993
5994 bn = malloc(sizeof(Backend));
5995 if (bn == NULL)
5996 {
5997 ereport(LOG,
5998 (errcode(ERRCODE_OUT_OF_MEMORY),
5999 errmsg("out of memory")));
6000 return false;
6001 }
6002
6003 bn->cancel_key = MyCancelKey;
6004 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
6005 bn->bkend_type = BACKEND_TYPE_BGWORKER;
6006 bn->dead_end = false;
6007 bn->bgworker_notify = false;
6008
6009 rw->rw_backend = bn;
6010 rw->rw_child_slot = bn->child_slot;
6011
6012 return true;
6013 }
6014
6015 /*
6016 * If the time is right, start background worker(s).
6017 *
6018 * As a side effect, the bgworker control variables are set or reset
6019 * depending on whether more workers may need to be started.
6020 *
6021 * We limit the number of workers started per call, to avoid consuming the
6022 * postmaster's attention for too long when many such requests are pending.
6023 * As long as StartWorkerNeeded is true, ServerLoop will not block and will
6024 * call this function again after dealing with any other issues.
6025 */
6026 static void
maybe_start_bgworkers(void)6027 maybe_start_bgworkers(void)
6028 {
6029 #define MAX_BGWORKERS_TO_LAUNCH 100
6030 int num_launched = 0;
6031 TimestampTz now = 0;
6032 slist_mutable_iter iter;
6033
6034 /*
6035 * During crash recovery, we have no need to be called until the state
6036 * transition out of recovery.
6037 */
6038 if (FatalError)
6039 {
6040 StartWorkerNeeded = false;
6041 HaveCrashedWorker = false;
6042 return;
6043 }
6044
6045 /* Don't need to be called again unless we find a reason for it below */
6046 StartWorkerNeeded = false;
6047 HaveCrashedWorker = false;
6048
6049 slist_foreach_modify(iter, &BackgroundWorkerList)
6050 {
6051 RegisteredBgWorker *rw;
6052
6053 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
6054
6055 /* ignore if already running */
6056 if (rw->rw_pid != 0)
6057 continue;
6058
6059 /* if marked for death, clean up and remove from list */
6060 if (rw->rw_terminate)
6061 {
6062 ForgetBackgroundWorker(&iter);
6063 continue;
6064 }
6065
6066 /*
6067 * If this worker has crashed previously, maybe it needs to be
6068 * restarted (unless on registration it specified it doesn't want to
6069 * be restarted at all). Check how long ago did a crash last happen.
6070 * If the last crash is too recent, don't start it right away; let it
6071 * be restarted once enough time has passed.
6072 */
6073 if (rw->rw_crashed_at != 0)
6074 {
6075 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
6076 {
6077 int notify_pid;
6078
6079 notify_pid = rw->rw_worker.bgw_notify_pid;
6080
6081 ForgetBackgroundWorker(&iter);
6082
6083 /* Report worker is gone now. */
6084 if (notify_pid != 0)
6085 kill(notify_pid, SIGUSR1);
6086
6087 continue;
6088 }
6089
6090 /* read system time only when needed */
6091 if (now == 0)
6092 now = GetCurrentTimestamp();
6093
6094 if (!TimestampDifferenceExceeds(rw->rw_crashed_at, now,
6095 rw->rw_worker.bgw_restart_time * 1000))
6096 {
6097 /* Set flag to remember that we have workers to start later */
6098 HaveCrashedWorker = true;
6099 continue;
6100 }
6101 }
6102
6103 if (bgworker_should_start_now(rw->rw_worker.bgw_start_time))
6104 {
6105 /* reset crash time before trying to start worker */
6106 rw->rw_crashed_at = 0;
6107
6108 /*
6109 * Try to start the worker.
6110 *
6111 * On failure, give up processing workers for now, but set
6112 * StartWorkerNeeded so we'll come back here on the next iteration
6113 * of ServerLoop to try again. (We don't want to wait, because
6114 * there might be additional ready-to-run workers.) We could set
6115 * HaveCrashedWorker as well, since this worker is now marked
6116 * crashed, but there's no need because the next run of this
6117 * function will do that.
6118 */
6119 if (!do_start_bgworker(rw))
6120 {
6121 StartWorkerNeeded = true;
6122 return;
6123 }
6124
6125 /*
6126 * If we've launched as many workers as allowed, quit, but have
6127 * ServerLoop call us again to look for additional ready-to-run
6128 * workers. There might not be any, but we'll find out the next
6129 * time we run.
6130 */
6131 if (++num_launched >= MAX_BGWORKERS_TO_LAUNCH)
6132 {
6133 StartWorkerNeeded = true;
6134 return;
6135 }
6136 }
6137 }
6138 }
6139
6140 /*
6141 * When a backend asks to be notified about worker state changes, we
6142 * set a flag in its backend entry. The background worker machinery needs
6143 * to know when such backends exit.
6144 */
6145 bool
PostmasterMarkPIDForWorkerNotify(int pid)6146 PostmasterMarkPIDForWorkerNotify(int pid)
6147 {
6148 dlist_iter iter;
6149 Backend *bp;
6150
6151 dlist_foreach(iter, &BackendList)
6152 {
6153 bp = dlist_container(Backend, elem, iter.cur);
6154 if (bp->pid == pid)
6155 {
6156 bp->bgworker_notify = true;
6157 return true;
6158 }
6159 }
6160 return false;
6161 }
6162
6163 #ifdef EXEC_BACKEND
6164
6165 /*
6166 * The following need to be available to the save/restore_backend_variables
6167 * functions. They are marked NON_EXEC_STATIC in their home modules.
6168 */
6169 extern slock_t *ShmemLock;
6170 extern slock_t *ProcStructLock;
6171 extern PGPROC *AuxiliaryProcs;
6172 extern PMSignalData *PMSignalState;
6173 extern pgsocket pgStatSock;
6174 extern pg_time_t first_syslogger_file_time;
6175
6176 #ifndef WIN32
6177 #define write_inheritable_socket(dest, src, childpid) ((*(dest) = (src)), true)
6178 #define read_inheritable_socket(dest, src) (*(dest) = *(src))
6179 #else
6180 static bool write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE child);
6181 static bool write_inheritable_socket(InheritableSocket *dest, SOCKET src,
6182 pid_t childPid);
6183 static void read_inheritable_socket(SOCKET *dest, InheritableSocket *src);
6184 #endif
6185
6186
6187 /* Save critical backend variables into the BackendParameters struct */
6188 #ifndef WIN32
6189 static bool
save_backend_variables(BackendParameters * param,Port * port)6190 save_backend_variables(BackendParameters *param, Port *port)
6191 #else
6192 static bool
6193 save_backend_variables(BackendParameters *param, Port *port,
6194 HANDLE childProcess, pid_t childPid)
6195 #endif
6196 {
6197 memcpy(¶m->port, port, sizeof(Port));
6198 if (!write_inheritable_socket(¶m->portsocket, port->sock, childPid))
6199 return false;
6200
6201 strlcpy(param->DataDir, DataDir, MAXPGPATH);
6202
6203 memcpy(¶m->ListenSocket, &ListenSocket, sizeof(ListenSocket));
6204
6205 param->MyCancelKey = MyCancelKey;
6206 param->MyPMChildSlot = MyPMChildSlot;
6207
6208 #ifdef WIN32
6209 param->ShmemProtectiveRegion = ShmemProtectiveRegion;
6210 #endif
6211 param->UsedShmemSegID = UsedShmemSegID;
6212 param->UsedShmemSegAddr = UsedShmemSegAddr;
6213
6214 param->ShmemLock = ShmemLock;
6215 param->ShmemVariableCache = ShmemVariableCache;
6216 param->ShmemBackendArray = ShmemBackendArray;
6217
6218 #ifndef HAVE_SPINLOCKS
6219 param->SpinlockSemaArray = SpinlockSemaArray;
6220 #endif
6221 param->NamedLWLockTrancheRequests = NamedLWLockTrancheRequests;
6222 param->NamedLWLockTrancheArray = NamedLWLockTrancheArray;
6223 param->MainLWLockArray = MainLWLockArray;
6224 param->ProcStructLock = ProcStructLock;
6225 param->ProcGlobal = ProcGlobal;
6226 param->AuxiliaryProcs = AuxiliaryProcs;
6227 param->PreparedXactProcs = PreparedXactProcs;
6228 param->PMSignalState = PMSignalState;
6229 if (!write_inheritable_socket(¶m->pgStatSock, pgStatSock, childPid))
6230 return false;
6231
6232 param->PostmasterPid = PostmasterPid;
6233 param->PgStartTime = PgStartTime;
6234 param->PgReloadTime = PgReloadTime;
6235 param->first_syslogger_file_time = first_syslogger_file_time;
6236
6237 param->redirection_done = redirection_done;
6238 param->IsBinaryUpgrade = IsBinaryUpgrade;
6239 param->max_safe_fds = max_safe_fds;
6240
6241 param->MaxBackends = MaxBackends;
6242
6243 #ifdef WIN32
6244 param->PostmasterHandle = PostmasterHandle;
6245 if (!write_duplicated_handle(¶m->initial_signal_pipe,
6246 pgwin32_create_signal_listener(childPid),
6247 childProcess))
6248 return false;
6249 #else
6250 memcpy(¶m->postmaster_alive_fds, &postmaster_alive_fds,
6251 sizeof(postmaster_alive_fds));
6252 #endif
6253
6254 memcpy(¶m->syslogPipe, &syslogPipe, sizeof(syslogPipe));
6255
6256 strlcpy(param->my_exec_path, my_exec_path, MAXPGPATH);
6257
6258 strlcpy(param->pkglib_path, pkglib_path, MAXPGPATH);
6259
6260 strlcpy(param->ExtraOptions, ExtraOptions, MAXPGPATH);
6261
6262 return true;
6263 }
6264
6265
6266 #ifdef WIN32
6267 /*
6268 * Duplicate a handle for usage in a child process, and write the child
6269 * process instance of the handle to the parameter file.
6270 */
6271 static bool
write_duplicated_handle(HANDLE * dest,HANDLE src,HANDLE childProcess)6272 write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE childProcess)
6273 {
6274 HANDLE hChild = INVALID_HANDLE_VALUE;
6275
6276 if (!DuplicateHandle(GetCurrentProcess(),
6277 src,
6278 childProcess,
6279 &hChild,
6280 0,
6281 TRUE,
6282 DUPLICATE_CLOSE_SOURCE | DUPLICATE_SAME_ACCESS))
6283 {
6284 ereport(LOG,
6285 (errmsg_internal("could not duplicate handle to be written to backend parameter file: error code %lu",
6286 GetLastError())));
6287 return false;
6288 }
6289
6290 *dest = hChild;
6291 return true;
6292 }
6293
6294 /*
6295 * Duplicate a socket for usage in a child process, and write the resulting
6296 * structure to the parameter file.
6297 * This is required because a number of LSPs (Layered Service Providers) very
6298 * common on Windows (antivirus, firewalls, download managers etc) break
6299 * straight socket inheritance.
6300 */
6301 static bool
write_inheritable_socket(InheritableSocket * dest,SOCKET src,pid_t childpid)6302 write_inheritable_socket(InheritableSocket *dest, SOCKET src, pid_t childpid)
6303 {
6304 dest->origsocket = src;
6305 if (src != 0 && src != PGINVALID_SOCKET)
6306 {
6307 /* Actual socket */
6308 if (WSADuplicateSocket(src, childpid, &dest->wsainfo) != 0)
6309 {
6310 ereport(LOG,
6311 (errmsg("could not duplicate socket %d for use in backend: error code %d",
6312 (int) src, WSAGetLastError())));
6313 return false;
6314 }
6315 }
6316 return true;
6317 }
6318
6319 /*
6320 * Read a duplicate socket structure back, and get the socket descriptor.
6321 */
6322 static void
read_inheritable_socket(SOCKET * dest,InheritableSocket * src)6323 read_inheritable_socket(SOCKET *dest, InheritableSocket *src)
6324 {
6325 SOCKET s;
6326
6327 if (src->origsocket == PGINVALID_SOCKET || src->origsocket == 0)
6328 {
6329 /* Not a real socket! */
6330 *dest = src->origsocket;
6331 }
6332 else
6333 {
6334 /* Actual socket, so create from structure */
6335 s = WSASocket(FROM_PROTOCOL_INFO,
6336 FROM_PROTOCOL_INFO,
6337 FROM_PROTOCOL_INFO,
6338 &src->wsainfo,
6339 0,
6340 0);
6341 if (s == INVALID_SOCKET)
6342 {
6343 write_stderr("could not create inherited socket: error code %d\n",
6344 WSAGetLastError());
6345 exit(1);
6346 }
6347 *dest = s;
6348
6349 /*
6350 * To make sure we don't get two references to the same socket, close
6351 * the original one. (This would happen when inheritance actually
6352 * works..
6353 */
6354 closesocket(src->origsocket);
6355 }
6356 }
6357 #endif
6358
6359 static void
read_backend_variables(char * id,Port * port)6360 read_backend_variables(char *id, Port *port)
6361 {
6362 BackendParameters param;
6363
6364 #ifndef WIN32
6365 /* Non-win32 implementation reads from file */
6366 FILE *fp;
6367
6368 /* Open file */
6369 fp = AllocateFile(id, PG_BINARY_R);
6370 if (!fp)
6371 {
6372 write_stderr("could not open backend variables file \"%s\": %s\n",
6373 id, strerror(errno));
6374 exit(1);
6375 }
6376
6377 if (fread(¶m, sizeof(param), 1, fp) != 1)
6378 {
6379 write_stderr("could not read from backend variables file \"%s\": %s\n",
6380 id, strerror(errno));
6381 exit(1);
6382 }
6383
6384 /* Release file */
6385 FreeFile(fp);
6386 if (unlink(id) != 0)
6387 {
6388 write_stderr("could not remove file \"%s\": %s\n",
6389 id, strerror(errno));
6390 exit(1);
6391 }
6392 #else
6393 /* Win32 version uses mapped file */
6394 HANDLE paramHandle;
6395 BackendParameters *paramp;
6396
6397 #ifdef _WIN64
6398 paramHandle = (HANDLE) _atoi64(id);
6399 #else
6400 paramHandle = (HANDLE) atol(id);
6401 #endif
6402 paramp = MapViewOfFile(paramHandle, FILE_MAP_READ, 0, 0, 0);
6403 if (!paramp)
6404 {
6405 write_stderr("could not map view of backend variables: error code %lu\n",
6406 GetLastError());
6407 exit(1);
6408 }
6409
6410 memcpy(¶m, paramp, sizeof(BackendParameters));
6411
6412 if (!UnmapViewOfFile(paramp))
6413 {
6414 write_stderr("could not unmap view of backend variables: error code %lu\n",
6415 GetLastError());
6416 exit(1);
6417 }
6418
6419 if (!CloseHandle(paramHandle))
6420 {
6421 write_stderr("could not close handle to backend parameter variables: error code %lu\n",
6422 GetLastError());
6423 exit(1);
6424 }
6425 #endif
6426
6427 restore_backend_variables(¶m, port);
6428 }
6429
6430 /* Restore critical backend variables from the BackendParameters struct */
6431 static void
restore_backend_variables(BackendParameters * param,Port * port)6432 restore_backend_variables(BackendParameters *param, Port *port)
6433 {
6434 memcpy(port, ¶m->port, sizeof(Port));
6435 read_inheritable_socket(&port->sock, ¶m->portsocket);
6436
6437 SetDataDir(param->DataDir);
6438
6439 memcpy(&ListenSocket, ¶m->ListenSocket, sizeof(ListenSocket));
6440
6441 MyCancelKey = param->MyCancelKey;
6442 MyPMChildSlot = param->MyPMChildSlot;
6443
6444 #ifdef WIN32
6445 ShmemProtectiveRegion = param->ShmemProtectiveRegion;
6446 #endif
6447 UsedShmemSegID = param->UsedShmemSegID;
6448 UsedShmemSegAddr = param->UsedShmemSegAddr;
6449
6450 ShmemLock = param->ShmemLock;
6451 ShmemVariableCache = param->ShmemVariableCache;
6452 ShmemBackendArray = param->ShmemBackendArray;
6453
6454 #ifndef HAVE_SPINLOCKS
6455 SpinlockSemaArray = param->SpinlockSemaArray;
6456 #endif
6457 NamedLWLockTrancheRequests = param->NamedLWLockTrancheRequests;
6458 NamedLWLockTrancheArray = param->NamedLWLockTrancheArray;
6459 MainLWLockArray = param->MainLWLockArray;
6460 ProcStructLock = param->ProcStructLock;
6461 ProcGlobal = param->ProcGlobal;
6462 AuxiliaryProcs = param->AuxiliaryProcs;
6463 PreparedXactProcs = param->PreparedXactProcs;
6464 PMSignalState = param->PMSignalState;
6465 read_inheritable_socket(&pgStatSock, ¶m->pgStatSock);
6466
6467 PostmasterPid = param->PostmasterPid;
6468 PgStartTime = param->PgStartTime;
6469 PgReloadTime = param->PgReloadTime;
6470 first_syslogger_file_time = param->first_syslogger_file_time;
6471
6472 redirection_done = param->redirection_done;
6473 IsBinaryUpgrade = param->IsBinaryUpgrade;
6474 max_safe_fds = param->max_safe_fds;
6475
6476 MaxBackends = param->MaxBackends;
6477
6478 #ifdef WIN32
6479 PostmasterHandle = param->PostmasterHandle;
6480 pgwin32_initial_signal_pipe = param->initial_signal_pipe;
6481 #else
6482 memcpy(&postmaster_alive_fds, ¶m->postmaster_alive_fds,
6483 sizeof(postmaster_alive_fds));
6484 #endif
6485
6486 memcpy(&syslogPipe, ¶m->syslogPipe, sizeof(syslogPipe));
6487
6488 strlcpy(my_exec_path, param->my_exec_path, MAXPGPATH);
6489
6490 strlcpy(pkglib_path, param->pkglib_path, MAXPGPATH);
6491
6492 strlcpy(ExtraOptions, param->ExtraOptions, MAXPGPATH);
6493
6494 /*
6495 * We need to restore fd.c's counts of externally-opened FDs; to avoid
6496 * confusion, be sure to do this after restoring max_safe_fds. (Note:
6497 * BackendInitialize will handle this for port->sock.)
6498 */
6499 #ifndef WIN32
6500 if (postmaster_alive_fds[0] >= 0)
6501 ReserveExternalFD();
6502 if (postmaster_alive_fds[1] >= 0)
6503 ReserveExternalFD();
6504 #endif
6505 if (pgStatSock != PGINVALID_SOCKET)
6506 ReserveExternalFD();
6507 }
6508
6509
6510 Size
ShmemBackendArraySize(void)6511 ShmemBackendArraySize(void)
6512 {
6513 return mul_size(MaxLivePostmasterChildren(), sizeof(Backend));
6514 }
6515
6516 void
ShmemBackendArrayAllocation(void)6517 ShmemBackendArrayAllocation(void)
6518 {
6519 Size size = ShmemBackendArraySize();
6520
6521 ShmemBackendArray = (Backend *) ShmemAlloc(size);
6522 /* Mark all slots as empty */
6523 memset(ShmemBackendArray, 0, size);
6524 }
6525
6526 static void
ShmemBackendArrayAdd(Backend * bn)6527 ShmemBackendArrayAdd(Backend *bn)
6528 {
6529 /* The array slot corresponding to my PMChildSlot should be free */
6530 int i = bn->child_slot - 1;
6531
6532 Assert(ShmemBackendArray[i].pid == 0);
6533 ShmemBackendArray[i] = *bn;
6534 }
6535
6536 static void
ShmemBackendArrayRemove(Backend * bn)6537 ShmemBackendArrayRemove(Backend *bn)
6538 {
6539 int i = bn->child_slot - 1;
6540
6541 Assert(ShmemBackendArray[i].pid == bn->pid);
6542 /* Mark the slot as empty */
6543 ShmemBackendArray[i].pid = 0;
6544 }
6545 #endif /* EXEC_BACKEND */
6546
6547
6548 #ifdef WIN32
6549
6550 /*
6551 * Subset implementation of waitpid() for Windows. We assume pid is -1
6552 * (that is, check all child processes) and options is WNOHANG (don't wait).
6553 */
6554 static pid_t
waitpid(pid_t pid,int * exitstatus,int options)6555 waitpid(pid_t pid, int *exitstatus, int options)
6556 {
6557 DWORD dwd;
6558 ULONG_PTR key;
6559 OVERLAPPED *ovl;
6560
6561 /*
6562 * Check if there are any dead children. If there are, return the pid of
6563 * the first one that died.
6564 */
6565 if (GetQueuedCompletionStatus(win32ChildQueue, &dwd, &key, &ovl, 0))
6566 {
6567 *exitstatus = (int) key;
6568 return dwd;
6569 }
6570
6571 return -1;
6572 }
6573
6574 /*
6575 * Note! Code below executes on a thread pool! All operations must
6576 * be thread safe! Note that elog() and friends must *not* be used.
6577 */
6578 static void WINAPI
pgwin32_deadchild_callback(PVOID lpParameter,BOOLEAN TimerOrWaitFired)6579 pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired)
6580 {
6581 win32_deadchild_waitinfo *childinfo = (win32_deadchild_waitinfo *) lpParameter;
6582 DWORD exitcode;
6583
6584 if (TimerOrWaitFired)
6585 return; /* timeout. Should never happen, since we use
6586 * INFINITE as timeout value. */
6587
6588 /*
6589 * Remove handle from wait - required even though it's set to wait only
6590 * once
6591 */
6592 UnregisterWaitEx(childinfo->waitHandle, NULL);
6593
6594 if (!GetExitCodeProcess(childinfo->procHandle, &exitcode))
6595 {
6596 /*
6597 * Should never happen. Inform user and set a fixed exitcode.
6598 */
6599 write_stderr("could not read exit code for process\n");
6600 exitcode = 255;
6601 }
6602
6603 if (!PostQueuedCompletionStatus(win32ChildQueue, childinfo->procId, (ULONG_PTR) exitcode, NULL))
6604 write_stderr("could not post child completion status\n");
6605
6606 /*
6607 * Handle is per-process, so we close it here instead of in the
6608 * originating thread
6609 */
6610 CloseHandle(childinfo->procHandle);
6611
6612 /*
6613 * Free struct that was allocated before the call to
6614 * RegisterWaitForSingleObject()
6615 */
6616 free(childinfo);
6617
6618 /* Queue SIGCHLD signal */
6619 pg_queue_signal(SIGCHLD);
6620 }
6621 #endif /* WIN32 */
6622
6623 /*
6624 * Initialize one and only handle for monitoring postmaster death.
6625 *
6626 * Called once in the postmaster, so that child processes can subsequently
6627 * monitor if their parent is dead.
6628 */
6629 static void
InitPostmasterDeathWatchHandle(void)6630 InitPostmasterDeathWatchHandle(void)
6631 {
6632 #ifndef WIN32
6633
6634 /*
6635 * Create a pipe. Postmaster holds the write end of the pipe open
6636 * (POSTMASTER_FD_OWN), and children hold the read end. Children can pass
6637 * the read file descriptor to select() to wake up in case postmaster
6638 * dies, or check for postmaster death with a (read() == 0). Children must
6639 * close the write end as soon as possible after forking, because EOF
6640 * won't be signaled in the read end until all processes have closed the
6641 * write fd. That is taken care of in ClosePostmasterPorts().
6642 */
6643 Assert(MyProcPid == PostmasterPid);
6644 if (pipe(postmaster_alive_fds) < 0)
6645 ereport(FATAL,
6646 (errcode_for_file_access(),
6647 errmsg_internal("could not create pipe to monitor postmaster death: %m")));
6648
6649 /* Notify fd.c that we've eaten two FDs for the pipe. */
6650 ReserveExternalFD();
6651 ReserveExternalFD();
6652
6653 /*
6654 * Set O_NONBLOCK to allow testing for the fd's presence with a read()
6655 * call.
6656 */
6657 if (fcntl(postmaster_alive_fds[POSTMASTER_FD_WATCH], F_SETFL, O_NONBLOCK) == -1)
6658 ereport(FATAL,
6659 (errcode_for_socket_access(),
6660 errmsg_internal("could not set postmaster death monitoring pipe to nonblocking mode: %m")));
6661 #else
6662
6663 /*
6664 * On Windows, we use a process handle for the same purpose.
6665 */
6666 if (DuplicateHandle(GetCurrentProcess(),
6667 GetCurrentProcess(),
6668 GetCurrentProcess(),
6669 &PostmasterHandle,
6670 0,
6671 TRUE,
6672 DUPLICATE_SAME_ACCESS) == 0)
6673 ereport(FATAL,
6674 (errmsg_internal("could not duplicate postmaster handle: error code %lu",
6675 GetLastError())));
6676 #endif /* WIN32 */
6677 }
6678