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) 1996-2016, PostgreSQL Global Development Group
36 * Portions Copyright (c) 1994, Regents of the University of California
37 *
38 *
39 * IDENTIFICATION
40 * src/backend/postmaster/postmaster.c
41 *
42 * NOTES
43 *
44 * Initialization:
45 * The Postmaster sets up shared memory data structures
46 * for the backends.
47 *
48 * Synchronization:
49 * The Postmaster shares memory with the backends but should avoid
50 * touching shared memory, so as not to become stuck if a crashing
51 * backend screws up locks or shared memory. Likewise, the Postmaster
52 * should never block on messages from frontend clients.
53 *
54 * Garbage Collection:
55 * The Postmaster cleans up after backends if they have an emergency
56 * exit and/or core dump.
57 *
58 * Error Reporting:
59 * Use write_stderr() only for reporting "interactive" errors
60 * (essentially, bogus arguments on the command line). Once the
61 * postmaster is launched, use ereport().
62 *
63 *-------------------------------------------------------------------------
64 */
65
66 #include "postgres.h"
67
68 #include <unistd.h>
69 #include <signal.h>
70 #include <time.h>
71 #include <sys/wait.h>
72 #include <ctype.h>
73 #include <sys/stat.h>
74 #include <sys/socket.h>
75 #include <fcntl.h>
76 #include <sys/param.h>
77 #include <netinet/in.h>
78 #include <arpa/inet.h>
79 #include <netdb.h>
80 #include <limits.h>
81
82 #ifdef HAVE_SYS_SELECT_H
83 #include <sys/select.h>
84 #endif
85
86 #ifdef USE_BONJOUR
87 #include <dns_sd.h>
88 #endif
89
90 #ifdef USE_SYSTEMD
91 #include <systemd/sd-daemon.h>
92 #endif
93
94 #ifdef HAVE_PTHREAD_IS_THREADED_NP
95 #include <pthread.h>
96 #endif
97
98 #include "access/transam.h"
99 #include "access/xlog.h"
100 #include "bootstrap/bootstrap.h"
101 #include "catalog/pg_control.h"
102 #include "lib/ilist.h"
103 #include "libpq/auth.h"
104 #include "libpq/ip.h"
105 #include "libpq/libpq.h"
106 #include "libpq/pqformat.h"
107 #include "libpq/pqsignal.h"
108 #include "miscadmin.h"
109 #include "pg_getopt.h"
110 #include "pgstat.h"
111 #include "postmaster/autovacuum.h"
112 #include "postmaster/bgworker_internals.h"
113 #include "postmaster/fork_process.h"
114 #include "postmaster/pgarch.h"
115 #include "postmaster/postmaster.h"
116 #include "postmaster/syslogger.h"
117 #include "replication/walsender.h"
118 #include "storage/fd.h"
119 #include "storage/ipc.h"
120 #include "storage/pg_shmem.h"
121 #include "storage/pmsignal.h"
122 #include "storage/proc.h"
123 #include "tcop/tcopprot.h"
124 #include "utils/builtins.h"
125 #include "utils/datetime.h"
126 #include "utils/dynamic_loader.h"
127 #include "utils/memutils.h"
128 #include "utils/ps_status.h"
129 #include "utils/timeout.h"
130
131 #ifdef EXEC_BACKEND
132 #include "storage/spin.h"
133 #endif
134
135
136 /*
137 * Possible types of a backend. Beyond being the possible bkend_type values in
138 * struct bkend, these are OR-able request flag bits for SignalSomeChildren()
139 * and CountChildren().
140 */
141 #define BACKEND_TYPE_NORMAL 0x0001 /* normal backend */
142 #define BACKEND_TYPE_AUTOVAC 0x0002 /* autovacuum worker process */
143 #define BACKEND_TYPE_WALSND 0x0004 /* walsender process */
144 #define BACKEND_TYPE_BGWORKER 0x0008 /* bgworker process */
145 #define BACKEND_TYPE_ALL 0x000F /* OR of all the above */
146
147 /*
148 * List of active backends (or child processes anyway; we don't actually
149 * know whether a given child has become a backend or is still in the
150 * authorization phase). This is used mainly to keep track of how many
151 * children we have and send them appropriate signals when necessary.
152 *
153 * "Special" children such as the startup, bgwriter and autovacuum launcher
154 * tasks are not in this list. Autovacuum worker and walsender are in it.
155 * Also, "dead_end" children are in it: these are children launched just for
156 * the purpose of sending a friendly rejection message to a would-be client.
157 * We must track them because they are attached to shared memory, but we know
158 * they will never become live backends. dead_end children are not assigned a
159 * PMChildSlot.
160 *
161 * Background workers are in this list, too.
162 */
163 typedef struct bkend
164 {
165 pid_t pid; /* process id of backend */
166 long cancel_key; /* cancel key for cancels for this backend */
167 int child_slot; /* PMChildSlot for this backend, if any */
168
169 /*
170 * Flavor of backend or auxiliary process. Note that BACKEND_TYPE_WALSND
171 * backends initially announce themselves as BACKEND_TYPE_NORMAL, so if
172 * bkend_type is normal, you should check for a recent transition.
173 */
174 int bkend_type;
175 bool dead_end; /* is it going to send an error and quit? */
176 bool bgworker_notify; /* gets bgworker start/stop notifications */
177 dlist_node elem; /* list link in BackendList */
178 } Backend;
179
180 static dlist_head BackendList = DLIST_STATIC_INIT(BackendList);
181
182 #ifdef EXEC_BACKEND
183 static Backend *ShmemBackendArray;
184 #endif
185
186 BackgroundWorker *MyBgworkerEntry = NULL;
187
188
189
190 /* The socket number we are listening for connections on */
191 int PostPortNumber;
192
193 /* The directory names for Unix socket(s) */
194 char *Unix_socket_directories;
195
196 /* The TCP listen address(es) */
197 char *ListenAddresses;
198
199 /*
200 * ReservedBackends is the number of backends reserved for superuser use.
201 * This number is taken out of the pool size given by MaxBackends so
202 * number of backend slots available to non-superusers is
203 * (MaxBackends - ReservedBackends). Note what this really means is
204 * "if there are <= ReservedBackends connections available, only superusers
205 * can make new connections" --- pre-existing superuser connections don't
206 * count against the limit.
207 */
208 int ReservedBackends;
209
210 /* The socket(s) we're listening to. */
211 #define MAXLISTEN 64
212 static pgsocket ListenSocket[MAXLISTEN];
213
214 /*
215 * Set by the -o option
216 */
217 static char ExtraOptions[MAXPGPATH];
218
219 /*
220 * These globals control the behavior of the postmaster in case some
221 * backend dumps core. Normally, it kills all peers of the dead backend
222 * and reinitializes shared memory. By specifying -s or -n, we can have
223 * the postmaster stop (rather than kill) peers and not reinitialize
224 * shared data structures. (Reinit is currently dead code, though.)
225 */
226 static bool Reinit = true;
227 static int SendStop = false;
228
229 /* still more option variables */
230 bool EnableSSL = false;
231
232 int PreAuthDelay = 0;
233 int AuthenticationTimeout = 60;
234
235 bool log_hostname; /* for ps display and logging */
236 bool Log_connections = false;
237 bool Db_user_namespace = false;
238
239 bool enable_bonjour = false;
240 char *bonjour_name;
241 bool restart_after_crash = true;
242
243 /* PIDs of special child processes; 0 when not running */
244 static pid_t StartupPID = 0,
245 BgWriterPID = 0,
246 CheckpointerPID = 0,
247 WalWriterPID = 0,
248 WalReceiverPID = 0,
249 AutoVacPID = 0,
250 PgArchPID = 0,
251 PgStatPID = 0,
252 SysLoggerPID = 0;
253
254 /* Startup process's status */
255 typedef enum
256 {
257 STARTUP_NOT_RUNNING,
258 STARTUP_RUNNING,
259 STARTUP_SIGNALED, /* we sent it a SIGQUIT or SIGKILL */
260 STARTUP_CRASHED
261 } StartupStatusEnum;
262
263 static StartupStatusEnum StartupStatus = STARTUP_NOT_RUNNING;
264
265 /* Startup/shutdown state */
266 #define NoShutdown 0
267 #define SmartShutdown 1
268 #define FastShutdown 2
269 #define ImmediateShutdown 3
270
271 static int Shutdown = NoShutdown;
272
273 static bool FatalError = false; /* T if recovering from backend crash */
274
275 /*
276 * We use a simple state machine to control startup, shutdown, and
277 * crash recovery (which is rather like shutdown followed by startup).
278 *
279 * After doing all the postmaster initialization work, we enter PM_STARTUP
280 * state and the startup process is launched. The startup process begins by
281 * reading the control file and other preliminary initialization steps.
282 * In a normal startup, or after crash recovery, the startup process exits
283 * with exit code 0 and we switch to PM_RUN state. However, archive recovery
284 * is handled specially since it takes much longer and we would like to support
285 * hot standby during archive recovery.
286 *
287 * When the startup process is ready to start archive recovery, it signals the
288 * postmaster, and we switch to PM_RECOVERY state. The background writer and
289 * checkpointer are launched, while the startup process continues applying WAL.
290 * If Hot Standby is enabled, then, after reaching a consistent point in WAL
291 * redo, startup process signals us again, and we switch to PM_HOT_STANDBY
292 * state and begin accepting connections to perform read-only queries. When
293 * archive recovery is finished, the startup process exits with exit code 0
294 * and we switch to PM_RUN state.
295 *
296 * Normal child backends can only be launched when we are in PM_RUN or
297 * PM_HOT_STANDBY state. (connsAllowed can also restrict launching.)
298 * In other states we handle connection requests by launching "dead_end"
299 * child processes, which will simply send the client an error message and
300 * quit. (We track these in the BackendList so that we can know when they
301 * are all gone; this is important because they're still connected to shared
302 * memory, and would interfere with an attempt to destroy the shmem segment,
303 * possibly leading to SHMALL failure when we try to make a new one.)
304 * In PM_WAIT_DEAD_END state we are waiting for all the dead_end children
305 * to drain out of the system, and therefore stop accepting connection
306 * requests at all until the last existing child has quit (which hopefully
307 * will not be very long).
308 *
309 * Notice that this state variable does not distinguish *why* we entered
310 * states later than PM_RUN --- Shutdown and FatalError must be consulted
311 * to find that out. FatalError is never true in PM_RECOVERY, PM_HOT_STANDBY,
312 * or PM_RUN states, nor in PM_SHUTDOWN states (because we don't enter those
313 * states when trying to recover from a crash). It can be true in PM_STARTUP
314 * state, because we don't clear it until we've successfully started WAL redo.
315 */
316 typedef enum
317 {
318 PM_INIT, /* postmaster starting */
319 PM_STARTUP, /* waiting for startup subprocess */
320 PM_RECOVERY, /* in archive recovery mode */
321 PM_HOT_STANDBY, /* in hot standby mode */
322 PM_RUN, /* normal "database is alive" state */
323 PM_STOP_BACKENDS, /* need to stop remaining backends */
324 PM_WAIT_BACKENDS, /* waiting for live backends to exit */
325 PM_SHUTDOWN, /* waiting for checkpointer to do shutdown
326 * ckpt */
327 PM_SHUTDOWN_2, /* waiting for archiver and walsenders to
328 * finish */
329 PM_WAIT_DEAD_END, /* waiting for dead_end children to exit */
330 PM_NO_CHILDREN /* all important children have exited */
331 } PMState;
332
333 static PMState pmState = PM_INIT;
334
335 /*
336 * While performing a "smart shutdown", we restrict new connections but stay
337 * in PM_RUN or PM_HOT_STANDBY state until all the client backends are gone.
338 * connsAllowed is a sub-state indicator showing the active restriction.
339 * It is of no interest unless pmState is PM_RUN or PM_HOT_STANDBY.
340 */
341 typedef enum
342 {
343 ALLOW_ALL_CONNS, /* normal not-shutting-down state */
344 ALLOW_SUPERUSER_CONNS, /* only superusers can connect */
345 ALLOW_NO_CONNS /* no new connections allowed, period */
346 } ConnsAllowedState;
347
348 static ConnsAllowedState connsAllowed = ALLOW_ALL_CONNS;
349
350 /* Start time of SIGKILL timeout during immediate shutdown or child crash */
351 /* Zero means timeout is not running */
352 static time_t AbortStartTime = 0;
353 /* Length of said timeout */
354 #define SIGKILL_CHILDREN_AFTER_SECS 5
355
356 static bool ReachedNormalRunning = false; /* T if we've reached PM_RUN */
357
358 bool ClientAuthInProgress = false; /* T during new-client
359 * authentication */
360
361 bool redirection_done = false; /* stderr redirected for syslogger? */
362
363 /* received START_AUTOVAC_LAUNCHER signal */
364 static volatile sig_atomic_t start_autovac_launcher = false;
365
366 /* the launcher needs to be signalled to communicate some condition */
367 static volatile bool avlauncher_needs_signal = false;
368
369 /* received START_WALRECEIVER signal */
370 static volatile sig_atomic_t WalReceiverRequested = false;
371
372 /* set when there's a worker that needs to be started up */
373 static volatile bool StartWorkerNeeded = true;
374 static volatile bool HaveCrashedWorker = false;
375
376 /*
377 * State for assigning random salts and cancel keys.
378 * Also, the global MyCancelKey passes the cancel key assigned to a given
379 * backend from the postmaster to that backend (via fork).
380 */
381 static unsigned int random_seed = 0;
382 static struct timeval random_start_time;
383
384 #ifdef USE_BONJOUR
385 static DNSServiceRef bonjour_sdref = NULL;
386 #endif
387
388 /*
389 * postmaster.c - function prototypes
390 */
391 static void CloseServerPorts(int status, Datum arg);
392 static void unlink_external_pid_file(int status, Datum arg);
393 static void getInstallationPaths(const char *argv0);
394 static void checkDataDir(void);
395 static Port *ConnCreate(int serverFd);
396 static void ConnFree(Port *port);
397 static void reset_shared(int port);
398 static void SIGHUP_handler(SIGNAL_ARGS);
399 static void pmdie(SIGNAL_ARGS);
400 static void reaper(SIGNAL_ARGS);
401 static void sigusr1_handler(SIGNAL_ARGS);
402 static void process_startup_packet_die(SIGNAL_ARGS);
403 static void process_startup_packet_quickdie(SIGNAL_ARGS);
404 static void dummy_handler(SIGNAL_ARGS);
405 static void StartupPacketTimeoutHandler(void);
406 static void CleanupBackend(int pid, int exitstatus);
407 static bool CleanupBackgroundWorker(int pid, int exitstatus);
408 static void HandleChildCrash(int pid, int exitstatus, const char *procname);
409 static void LogChildExit(int lev, const char *procname,
410 int pid, int exitstatus);
411 static void PostmasterStateMachine(void);
412 static void BackendInitialize(Port *port);
413 static void BackendRun(Port *port) pg_attribute_noreturn();
414 static void ExitPostmaster(int status) pg_attribute_noreturn();
415 static int ServerLoop(void);
416 static int BackendStartup(Port *port);
417 static int ProcessStartupPacket(Port *port, bool SSLdone);
418 static void SendNegotiateProtocolVersion(List *unrecognized_protocol_options);
419 static void processCancelRequest(Port *port, void *pkt);
420 static int initMasks(fd_set *rmask);
421 static void report_fork_failure_to_client(Port *port, int errnum);
422 static CAC_state canAcceptConnections(int backend_type);
423 static long PostmasterRandom(void);
424 static void RandomSalt(char *md5Salt);
425 static void signal_child(pid_t pid, int signal);
426 static bool SignalSomeChildren(int signal, int targets);
427 static void TerminateChildren(int signal);
428
429 #define SignalChildren(sig) SignalSomeChildren(sig, BACKEND_TYPE_ALL)
430
431 static int CountChildren(int target);
432 static bool assign_backendlist_entry(RegisteredBgWorker *rw);
433 static void maybe_start_bgworkers(void);
434 static bool CreateOptsFile(int argc, char *argv[], char *fullprogname);
435 static pid_t StartChildProcess(AuxProcType type);
436 static void StartAutovacuumWorker(void);
437 static void MaybeStartWalReceiver(void);
438 static void InitPostmasterDeathWatchHandle(void);
439
440 /*
441 * Archiver is allowed to start up at the current postmaster state?
442 *
443 * If WAL archiving is enabled always, we are allowed to start archiver
444 * even during recovery.
445 */
446 #define PgArchStartupAllowed() \
447 ((XLogArchivingActive() && pmState == PM_RUN) || \
448 (XLogArchivingAlways() && \
449 (pmState == PM_RECOVERY || pmState == PM_HOT_STANDBY)))
450
451 #ifdef EXEC_BACKEND
452
453 #ifdef WIN32
454 #define WNOHANG 0 /* ignored, so any integer value will do */
455
456 static pid_t waitpid(pid_t pid, int *exitstatus, int options);
457 static void WINAPI pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired);
458
459 static HANDLE win32ChildQueue;
460
461 typedef struct
462 {
463 HANDLE waitHandle;
464 HANDLE procHandle;
465 DWORD procId;
466 } win32_deadchild_waitinfo;
467 #endif /* WIN32 */
468
469 static pid_t backend_forkexec(Port *port);
470 static pid_t internal_forkexec(int argc, char *argv[], Port *port);
471
472 /* Type for a socket that can be inherited to a client process */
473 #ifdef WIN32
474 typedef struct
475 {
476 SOCKET origsocket; /* Original socket value, or PGINVALID_SOCKET
477 * if not a socket */
478 WSAPROTOCOL_INFO wsainfo;
479 } InheritableSocket;
480 #else
481 typedef int InheritableSocket;
482 #endif
483
484 /*
485 * Structure contains all variables passed to exec:ed backends
486 */
487 typedef struct
488 {
489 Port port;
490 InheritableSocket portsocket;
491 char DataDir[MAXPGPATH];
492 pgsocket ListenSocket[MAXLISTEN];
493 long MyCancelKey;
494 int MyPMChildSlot;
495 #ifndef WIN32
496 unsigned long UsedShmemSegID;
497 #else
498 void *ShmemProtectiveRegion;
499 HANDLE UsedShmemSegID;
500 #endif
501 void *UsedShmemSegAddr;
502 slock_t *ShmemLock;
503 VariableCache ShmemVariableCache;
504 Backend *ShmemBackendArray;
505 #ifndef HAVE_SPINLOCKS
506 PGSemaphore SpinlockSemaArray;
507 #endif
508 int NamedLWLockTrancheRequests;
509 NamedLWLockTranche *NamedLWLockTrancheArray;
510 LWLockPadded *MainLWLockArray;
511 slock_t *ProcStructLock;
512 PROC_HDR *ProcGlobal;
513 PGPROC *AuxiliaryProcs;
514 PGPROC *PreparedXactProcs;
515 PMSignalData *PMSignalState;
516 InheritableSocket pgStatSock;
517 pid_t PostmasterPid;
518 TimestampTz PgStartTime;
519 TimestampTz PgReloadTime;
520 pg_time_t first_syslogger_file_time;
521 bool redirection_done;
522 bool IsBinaryUpgrade;
523 int max_safe_fds;
524 int MaxBackends;
525 #ifdef WIN32
526 HANDLE PostmasterHandle;
527 HANDLE initial_signal_pipe;
528 HANDLE syslogPipe[2];
529 #else
530 int postmaster_alive_fds[2];
531 int syslogPipe[2];
532 #endif
533 char my_exec_path[MAXPGPATH];
534 char pkglib_path[MAXPGPATH];
535 char ExtraOptions[MAXPGPATH];
536 } BackendParameters;
537
538 static void read_backend_variables(char *id, Port *port);
539 static void restore_backend_variables(BackendParameters *param, Port *port);
540
541 #ifndef WIN32
542 static bool save_backend_variables(BackendParameters *param, Port *port);
543 #else
544 static bool save_backend_variables(BackendParameters *param, Port *port,
545 HANDLE childProcess, pid_t childPid);
546 #endif
547
548 static void ShmemBackendArrayAdd(Backend *bn);
549 static void ShmemBackendArrayRemove(Backend *bn);
550 #endif /* EXEC_BACKEND */
551
552 #define StartupDataBase() StartChildProcess(StartupProcess)
553 #define StartBackgroundWriter() StartChildProcess(BgWriterProcess)
554 #define StartCheckpointer() StartChildProcess(CheckpointerProcess)
555 #define StartWalWriter() StartChildProcess(WalWriterProcess)
556 #define StartWalReceiver() StartChildProcess(WalReceiverProcess)
557
558 /* Macros to check exit status of a child process */
559 #define EXIT_STATUS_0(st) ((st) == 0)
560 #define EXIT_STATUS_1(st) (WIFEXITED(st) && WEXITSTATUS(st) == 1)
561 #define EXIT_STATUS_3(st) (WIFEXITED(st) && WEXITSTATUS(st) == 3)
562
563 #ifndef WIN32
564 /*
565 * File descriptors for pipe used to monitor if postmaster is alive.
566 * First is POSTMASTER_FD_WATCH, second is POSTMASTER_FD_OWN.
567 */
568 int postmaster_alive_fds[2] = {-1, -1};
569 #else
570 /* Process handle of postmaster used for the same purpose on Windows */
571 HANDLE PostmasterHandle;
572 #endif
573
574 /*
575 * Postmaster main entry point
576 */
577 void
PostmasterMain(int argc,char * argv[])578 PostmasterMain(int argc, char *argv[])
579 {
580 int opt;
581 int status;
582 char *userDoption = NULL;
583 bool listen_addr_saved = false;
584 int i;
585 char *output_config_variable = NULL;
586
587 MyProcPid = PostmasterPid = getpid();
588
589 MyStartTime = time(NULL);
590
591 IsPostmasterEnvironment = true;
592
593 /*
594 * for security, no dir or file created can be group or other accessible
595 */
596 umask(S_IRWXG | S_IRWXO);
597
598 /*
599 * Initialize random(3) so we don't get the same values in every run.
600 *
601 * Note: the seed is pretty predictable from externally-visible facts such
602 * as postmaster start time, so avoid using random() for security-critical
603 * random values during postmaster startup. At the time of first
604 * connection, PostmasterRandom will select a hopefully-more-random seed.
605 */
606 srandom((unsigned int) (MyProcPid ^ MyStartTime));
607
608 /*
609 * By default, palloc() requests in the postmaster will be allocated in
610 * the PostmasterContext, which is space that can be recycled by backends.
611 * Allocated data that needs to be available to backends should be
612 * allocated in TopMemoryContext.
613 */
614 PostmasterContext = AllocSetContextCreate(TopMemoryContext,
615 "Postmaster",
616 ALLOCSET_DEFAULT_SIZES);
617 MemoryContextSwitchTo(PostmasterContext);
618
619 /* Initialize paths to installation files */
620 getInstallationPaths(argv[0]);
621
622 /*
623 * Set up signal handlers for the postmaster process.
624 *
625 * In the postmaster, we use pqsignal_pm() rather than pqsignal() (which
626 * is used by all child processes and client processes). That has a
627 * couple of special behaviors:
628 *
629 * 1. Except on Windows, we tell sigaction() to block all signals for the
630 * duration of the signal handler. This is faster than our old approach
631 * of blocking/unblocking explicitly in the signal handler, and it should
632 * also prevent excessive stack consumption if signals arrive quickly.
633 *
634 * 2. We do not set the SA_RESTART flag. This is because signals will be
635 * blocked at all times except when ServerLoop is waiting for something to
636 * happen, and during that window, we want signals to exit the select(2)
637 * wait so that ServerLoop can respond if anything interesting happened.
638 * On some platforms, signals marked SA_RESTART would not cause the
639 * select() wait to end.
640 *
641 * Child processes will generally want SA_RESTART, so pqsignal() sets that
642 * flag. We expect children to set up their own handlers before
643 * unblocking signals.
644 *
645 * CAUTION: when changing this list, check for side-effects on the signal
646 * handling setup of child processes. See tcop/postgres.c,
647 * bootstrap/bootstrap.c, postmaster/bgwriter.c, postmaster/walwriter.c,
648 * postmaster/autovacuum.c, postmaster/pgarch.c, postmaster/pgstat.c,
649 * postmaster/syslogger.c, postmaster/bgworker.c and
650 * postmaster/checkpointer.c.
651 */
652 pqinitmask();
653 PG_SETMASK(&BlockSig);
654
655 pqsignal_pm(SIGHUP, SIGHUP_handler); /* reread config file and have
656 * children do same */
657 pqsignal_pm(SIGINT, pmdie); /* send SIGTERM and shut down */
658 pqsignal_pm(SIGQUIT, pmdie); /* send SIGQUIT and die */
659 pqsignal_pm(SIGTERM, pmdie); /* wait for children and shut down */
660 pqsignal_pm(SIGALRM, SIG_IGN); /* ignored */
661 pqsignal_pm(SIGPIPE, SIG_IGN); /* ignored */
662 pqsignal_pm(SIGUSR1, sigusr1_handler); /* message from child process */
663 pqsignal_pm(SIGUSR2, dummy_handler); /* unused, reserve for children */
664 pqsignal_pm(SIGCHLD, reaper); /* handle child termination */
665 pqsignal_pm(SIGTTIN, SIG_IGN); /* ignored */
666 pqsignal_pm(SIGTTOU, SIG_IGN); /* ignored */
667 /* ignore SIGXFSZ, so that ulimit violations work like disk full */
668 #ifdef SIGXFSZ
669 pqsignal_pm(SIGXFSZ, SIG_IGN); /* ignored */
670 #endif
671
672 /*
673 * Options setup
674 */
675 InitializeGUCOptions();
676
677 opterr = 1;
678
679 /*
680 * Parse command-line options. CAUTION: keep this in sync with
681 * tcop/postgres.c (the option sets should not conflict) and with the
682 * common help() function in main/main.c.
683 */
684 while ((opt = getopt(argc, argv, "B:bc:C:D:d:EeFf:h:ijk:lN:nOo:Pp:r:S:sTt:W:-:")) != -1)
685 {
686 switch (opt)
687 {
688 case 'B':
689 SetConfigOption("shared_buffers", optarg, PGC_POSTMASTER, PGC_S_ARGV);
690 break;
691
692 case 'b':
693 /* Undocumented flag used for binary upgrades */
694 IsBinaryUpgrade = true;
695 break;
696
697 case 'C':
698 output_config_variable = strdup(optarg);
699 break;
700
701 case 'D':
702 userDoption = strdup(optarg);
703 break;
704
705 case 'd':
706 set_debug_options(atoi(optarg), PGC_POSTMASTER, PGC_S_ARGV);
707 break;
708
709 case 'E':
710 SetConfigOption("log_statement", "all", PGC_POSTMASTER, PGC_S_ARGV);
711 break;
712
713 case 'e':
714 SetConfigOption("datestyle", "euro", PGC_POSTMASTER, PGC_S_ARGV);
715 break;
716
717 case 'F':
718 SetConfigOption("fsync", "false", PGC_POSTMASTER, PGC_S_ARGV);
719 break;
720
721 case 'f':
722 if (!set_plan_disabling_options(optarg, PGC_POSTMASTER, PGC_S_ARGV))
723 {
724 write_stderr("%s: invalid argument for option -f: \"%s\"\n",
725 progname, optarg);
726 ExitPostmaster(1);
727 }
728 break;
729
730 case 'h':
731 SetConfigOption("listen_addresses", optarg, PGC_POSTMASTER, PGC_S_ARGV);
732 break;
733
734 case 'i':
735 SetConfigOption("listen_addresses", "*", PGC_POSTMASTER, PGC_S_ARGV);
736 break;
737
738 case 'j':
739 /* only used by interactive backend */
740 break;
741
742 case 'k':
743 SetConfigOption("unix_socket_directories", optarg, PGC_POSTMASTER, PGC_S_ARGV);
744 break;
745
746 case 'l':
747 SetConfigOption("ssl", "true", PGC_POSTMASTER, PGC_S_ARGV);
748 break;
749
750 case 'N':
751 SetConfigOption("max_connections", optarg, PGC_POSTMASTER, PGC_S_ARGV);
752 break;
753
754 case 'n':
755 /* Don't reinit shared mem after abnormal exit */
756 Reinit = false;
757 break;
758
759 case 'O':
760 SetConfigOption("allow_system_table_mods", "true", PGC_POSTMASTER, PGC_S_ARGV);
761 break;
762
763 case 'o':
764 /* Other options to pass to the backend on the command line */
765 snprintf(ExtraOptions + strlen(ExtraOptions),
766 sizeof(ExtraOptions) - strlen(ExtraOptions),
767 " %s", optarg);
768 break;
769
770 case 'P':
771 SetConfigOption("ignore_system_indexes", "true", PGC_POSTMASTER, PGC_S_ARGV);
772 break;
773
774 case 'p':
775 SetConfigOption("port", optarg, PGC_POSTMASTER, PGC_S_ARGV);
776 break;
777
778 case 'r':
779 /* only used by single-user backend */
780 break;
781
782 case 'S':
783 SetConfigOption("work_mem", optarg, PGC_POSTMASTER, PGC_S_ARGV);
784 break;
785
786 case 's':
787 SetConfigOption("log_statement_stats", "true", PGC_POSTMASTER, PGC_S_ARGV);
788 break;
789
790 case 'T':
791
792 /*
793 * In the event that some backend dumps core, send SIGSTOP,
794 * rather than SIGQUIT, to all its peers. This lets the wily
795 * post_hacker collect core dumps from everyone.
796 */
797 SendStop = true;
798 break;
799
800 case 't':
801 {
802 const char *tmp = get_stats_option_name(optarg);
803
804 if (tmp)
805 {
806 SetConfigOption(tmp, "true", PGC_POSTMASTER, PGC_S_ARGV);
807 }
808 else
809 {
810 write_stderr("%s: invalid argument for option -t: \"%s\"\n",
811 progname, optarg);
812 ExitPostmaster(1);
813 }
814 break;
815 }
816
817 case 'W':
818 SetConfigOption("post_auth_delay", optarg, PGC_POSTMASTER, PGC_S_ARGV);
819 break;
820
821 case 'c':
822 case '-':
823 {
824 char *name,
825 *value;
826
827 ParseLongOption(optarg, &name, &value);
828 if (!value)
829 {
830 if (opt == '-')
831 ereport(ERROR,
832 (errcode(ERRCODE_SYNTAX_ERROR),
833 errmsg("--%s requires a value",
834 optarg)));
835 else
836 ereport(ERROR,
837 (errcode(ERRCODE_SYNTAX_ERROR),
838 errmsg("-c %s requires a value",
839 optarg)));
840 }
841
842 SetConfigOption(name, value, PGC_POSTMASTER, PGC_S_ARGV);
843 free(name);
844 if (value)
845 free(value);
846 break;
847 }
848
849 default:
850 write_stderr("Try \"%s --help\" for more information.\n",
851 progname);
852 ExitPostmaster(1);
853 }
854 }
855
856 /*
857 * Postmaster accepts no non-option switch arguments.
858 */
859 if (optind < argc)
860 {
861 write_stderr("%s: invalid argument: \"%s\"\n",
862 progname, argv[optind]);
863 write_stderr("Try \"%s --help\" for more information.\n",
864 progname);
865 ExitPostmaster(1);
866 }
867
868 /*
869 * Locate the proper configuration files and data directory, and read
870 * postgresql.conf for the first time.
871 */
872 if (!SelectConfigFiles(userDoption, progname))
873 ExitPostmaster(2);
874
875 if (output_config_variable != NULL)
876 {
877 /*
878 * "-C guc" was specified, so print GUC's value and exit. No extra
879 * permission check is needed because the user is reading inside the
880 * data dir.
881 */
882 const char *config_val = GetConfigOption(output_config_variable,
883 false, false);
884
885 puts(config_val ? config_val : "");
886 ExitPostmaster(0);
887 }
888
889 /* Verify that DataDir looks reasonable */
890 checkDataDir();
891
892 /* And switch working directory into it */
893 ChangeToDataDir();
894
895 /*
896 * Check for invalid combinations of GUC settings.
897 */
898 if (ReservedBackends >= MaxConnections)
899 {
900 write_stderr("%s: superuser_reserved_connections must be less than max_connections\n", progname);
901 ExitPostmaster(1);
902 }
903 if (max_wal_senders >= MaxConnections)
904 {
905 write_stderr("%s: max_wal_senders must be less than max_connections\n", progname);
906 ExitPostmaster(1);
907 }
908 if (XLogArchiveMode > ARCHIVE_MODE_OFF && wal_level == WAL_LEVEL_MINIMAL)
909 ereport(ERROR,
910 (errmsg("WAL archival cannot be enabled when wal_level is \"minimal\"")));
911 if (max_wal_senders > 0 && wal_level == WAL_LEVEL_MINIMAL)
912 ereport(ERROR,
913 (errmsg("WAL streaming (max_wal_senders > 0) requires wal_level \"replica\" or \"logical\"")));
914
915 /*
916 * Other one-time internal sanity checks can go here, if they are fast.
917 * (Put any slow processing further down, after postmaster.pid creation.)
918 */
919 if (!CheckDateTokenTables())
920 {
921 write_stderr("%s: invalid datetoken tables, please fix\n", progname);
922 ExitPostmaster(1);
923 }
924
925 /*
926 * Now that we are done processing the postmaster arguments, reset
927 * getopt(3) library so that it will work correctly in subprocesses.
928 */
929 optind = 1;
930 #ifdef HAVE_INT_OPTRESET
931 optreset = 1; /* some systems need this too */
932 #endif
933
934 /* For debugging: display postmaster environment */
935 {
936 extern char **environ;
937 char **p;
938
939 ereport(DEBUG3,
940 (errmsg_internal("%s: PostmasterMain: initial environment dump:",
941 progname)));
942 ereport(DEBUG3,
943 (errmsg_internal("-----------------------------------------")));
944 for (p = environ; *p; ++p)
945 ereport(DEBUG3,
946 (errmsg_internal("\t%s", *p)));
947 ereport(DEBUG3,
948 (errmsg_internal("-----------------------------------------")));
949 }
950
951 /*
952 * Create lockfile for data directory.
953 *
954 * We want to do this before we try to grab the input sockets, because the
955 * data directory interlock is more reliable than the socket-file
956 * interlock (thanks to whoever decided to put socket files in /tmp :-().
957 * For the same reason, it's best to grab the TCP socket(s) before the
958 * Unix socket(s).
959 *
960 * Also note that this internally sets up the on_proc_exit function that
961 * is responsible for removing both data directory and socket lockfiles;
962 * so it must happen before opening sockets so that at exit, the socket
963 * lockfiles go away after CloseServerPorts runs.
964 */
965 CreateDataDirLockFile(true);
966
967 /*
968 * Initialize SSL library, if specified.
969 */
970 #ifdef USE_SSL
971 if (EnableSSL)
972 secure_initialize();
973 #endif
974
975 /*
976 * process any libraries that should be preloaded at postmaster start
977 */
978 process_shared_preload_libraries();
979
980 /*
981 * Now that loadable modules have had their chance to register background
982 * workers, calculate MaxBackends.
983 */
984 InitializeMaxBackends();
985
986 /*
987 * Establish input sockets.
988 *
989 * First, mark them all closed, and set up an on_proc_exit function that's
990 * charged with closing the sockets again at postmaster shutdown.
991 */
992 for (i = 0; i < MAXLISTEN; i++)
993 ListenSocket[i] = PGINVALID_SOCKET;
994
995 on_proc_exit(CloseServerPorts, 0);
996
997 if (ListenAddresses)
998 {
999 char *rawstring;
1000 List *elemlist;
1001 ListCell *l;
1002 int success = 0;
1003
1004 /* Need a modifiable copy of ListenAddresses */
1005 rawstring = pstrdup(ListenAddresses);
1006
1007 /* Parse string into list of hostnames */
1008 if (!SplitIdentifierString(rawstring, ',', &elemlist))
1009 {
1010 /* syntax error in list */
1011 ereport(FATAL,
1012 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1013 errmsg("invalid list syntax in parameter \"%s\"",
1014 "listen_addresses")));
1015 }
1016
1017 foreach(l, elemlist)
1018 {
1019 char *curhost = (char *) lfirst(l);
1020
1021 if (strcmp(curhost, "*") == 0)
1022 status = StreamServerPort(AF_UNSPEC, NULL,
1023 (unsigned short) PostPortNumber,
1024 NULL,
1025 ListenSocket, MAXLISTEN);
1026 else
1027 status = StreamServerPort(AF_UNSPEC, curhost,
1028 (unsigned short) PostPortNumber,
1029 NULL,
1030 ListenSocket, MAXLISTEN);
1031
1032 if (status == STATUS_OK)
1033 {
1034 success++;
1035 /* record the first successful host addr in lockfile */
1036 if (!listen_addr_saved)
1037 {
1038 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, curhost);
1039 listen_addr_saved = true;
1040 }
1041 }
1042 else
1043 ereport(WARNING,
1044 (errmsg("could not create listen socket for \"%s\"",
1045 curhost)));
1046 }
1047
1048 if (!success && elemlist != NIL)
1049 ereport(FATAL,
1050 (errmsg("could not create any TCP/IP sockets")));
1051
1052 list_free(elemlist);
1053 pfree(rawstring);
1054 }
1055
1056 #ifdef USE_BONJOUR
1057 /* Register for Bonjour only if we opened TCP socket(s) */
1058 if (enable_bonjour && ListenSocket[0] != PGINVALID_SOCKET)
1059 {
1060 DNSServiceErrorType err;
1061
1062 /*
1063 * We pass 0 for interface_index, which will result in registering on
1064 * all "applicable" interfaces. It's not entirely clear from the
1065 * DNS-SD docs whether this would be appropriate if we have bound to
1066 * just a subset of the available network interfaces.
1067 */
1068 err = DNSServiceRegister(&bonjour_sdref,
1069 0,
1070 0,
1071 bonjour_name,
1072 "_postgresql._tcp.",
1073 NULL,
1074 NULL,
1075 htons(PostPortNumber),
1076 0,
1077 NULL,
1078 NULL,
1079 NULL);
1080 if (err != kDNSServiceErr_NoError)
1081 elog(LOG, "DNSServiceRegister() failed: error code %ld",
1082 (long) err);
1083
1084 /*
1085 * We don't bother to read the mDNS daemon's reply, and we expect that
1086 * it will automatically terminate our registration when the socket is
1087 * closed at postmaster termination. So there's nothing more to be
1088 * done here. However, the bonjour_sdref is kept around so that
1089 * forked children can close their copies of the socket.
1090 */
1091 }
1092 #endif
1093
1094 #ifdef HAVE_UNIX_SOCKETS
1095 if (Unix_socket_directories)
1096 {
1097 char *rawstring;
1098 List *elemlist;
1099 ListCell *l;
1100 int success = 0;
1101
1102 /* Need a modifiable copy of Unix_socket_directories */
1103 rawstring = pstrdup(Unix_socket_directories);
1104
1105 /* Parse string into list of directories */
1106 if (!SplitDirectoriesString(rawstring, ',', &elemlist))
1107 {
1108 /* syntax error in list */
1109 ereport(FATAL,
1110 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1111 errmsg("invalid list syntax in parameter \"%s\"",
1112 "unix_socket_directories")));
1113 }
1114
1115 foreach(l, elemlist)
1116 {
1117 char *socketdir = (char *) lfirst(l);
1118
1119 status = StreamServerPort(AF_UNIX, NULL,
1120 (unsigned short) PostPortNumber,
1121 socketdir,
1122 ListenSocket, MAXLISTEN);
1123
1124 if (status == STATUS_OK)
1125 {
1126 success++;
1127 /* record the first successful Unix socket in lockfile */
1128 if (success == 1)
1129 AddToDataDirLockFile(LOCK_FILE_LINE_SOCKET_DIR, socketdir);
1130 }
1131 else
1132 ereport(WARNING,
1133 (errmsg("could not create Unix-domain socket in directory \"%s\"",
1134 socketdir)));
1135 }
1136
1137 if (!success && elemlist != NIL)
1138 ereport(FATAL,
1139 (errmsg("could not create any Unix-domain sockets")));
1140
1141 list_free_deep(elemlist);
1142 pfree(rawstring);
1143 }
1144 #endif
1145
1146 /*
1147 * check that we have some socket to listen on
1148 */
1149 if (ListenSocket[0] == PGINVALID_SOCKET)
1150 ereport(FATAL,
1151 (errmsg("no socket created for listening")));
1152
1153 /*
1154 * If no valid TCP ports, write an empty line for listen address,
1155 * indicating the Unix socket must be used. Note that this line is not
1156 * added to the lock file until there is a socket backing it.
1157 */
1158 if (!listen_addr_saved)
1159 AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, "");
1160
1161 /*
1162 * Set up shared memory and semaphores.
1163 */
1164 reset_shared(PostPortNumber);
1165
1166 /*
1167 * Estimate number of openable files. This must happen after setting up
1168 * semaphores, because on some platforms semaphores count as open files.
1169 */
1170 set_max_safe_fds();
1171
1172 /*
1173 * Set reference point for stack-depth checking.
1174 */
1175 set_stack_base();
1176
1177 /*
1178 * Initialize pipe (or process handle on Windows) that allows children to
1179 * wake up from sleep on postmaster death.
1180 */
1181 InitPostmasterDeathWatchHandle();
1182
1183 #ifdef WIN32
1184
1185 /*
1186 * Initialize I/O completion port used to deliver list of dead children.
1187 */
1188 win32ChildQueue = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 1);
1189 if (win32ChildQueue == NULL)
1190 ereport(FATAL,
1191 (errmsg("could not create I/O completion port for child queue")));
1192 #endif
1193
1194 /*
1195 * Record postmaster options. We delay this till now to avoid recording
1196 * bogus options (eg, NBuffers too high for available memory).
1197 */
1198 if (!CreateOptsFile(argc, argv, my_exec_path))
1199 ExitPostmaster(1);
1200
1201 #ifdef EXEC_BACKEND
1202 /* Write out nondefault GUC settings for child processes to use */
1203 write_nondefault_variables(PGC_POSTMASTER);
1204 #endif
1205
1206 /*
1207 * Write the external PID file if requested
1208 */
1209 if (external_pid_file)
1210 {
1211 FILE *fpidfile = fopen(external_pid_file, "w");
1212
1213 if (fpidfile)
1214 {
1215 fprintf(fpidfile, "%d\n", MyProcPid);
1216 fclose(fpidfile);
1217
1218 /* Make PID file world readable */
1219 if (chmod(external_pid_file, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH) != 0)
1220 write_stderr("%s: could not change permissions of external PID file \"%s\": %s\n",
1221 progname, external_pid_file, strerror(errno));
1222 }
1223 else
1224 write_stderr("%s: could not write external PID file \"%s\": %s\n",
1225 progname, external_pid_file, strerror(errno));
1226
1227 on_proc_exit(unlink_external_pid_file, 0);
1228 }
1229
1230 /*
1231 * Remove old temporary files. At this point there can be no other
1232 * Postgres processes running in this directory, so this should be safe.
1233 */
1234 RemovePgTempFiles();
1235
1236 /*
1237 * Forcibly remove the files signaling a standby promotion request.
1238 * Otherwise, the existence of those files triggers a promotion too early,
1239 * whether a user wants that or not.
1240 *
1241 * This removal of files is usually unnecessary because they can exist
1242 * only during a few moments during a standby promotion. However there is
1243 * a race condition: if pg_ctl promote is executed and creates the files
1244 * during a promotion, the files can stay around even after the server is
1245 * brought up to new master. Then, if new standby starts by using the
1246 * backup taken from that master, the files can exist at the server
1247 * startup and should be removed in order to avoid an unexpected
1248 * promotion.
1249 *
1250 * Note that promotion signal files need to be removed before the startup
1251 * process is invoked. Because, after that, they can be used by
1252 * postmaster's SIGUSR1 signal handler.
1253 */
1254 RemovePromoteSignalFiles();
1255
1256 /*
1257 * If enabled, start up syslogger collection subprocess
1258 */
1259 SysLoggerPID = SysLogger_Start();
1260
1261 /*
1262 * Reset whereToSendOutput from DestDebug (its starting state) to
1263 * DestNone. This stops ereport from sending log messages to stderr unless
1264 * Log_destination permits. We don't do this until the postmaster is
1265 * fully launched, since startup failures may as well be reported to
1266 * stderr.
1267 *
1268 * If we are in fact disabling logging to stderr, first emit a log message
1269 * saying so, to provide a breadcrumb trail for users who may not remember
1270 * that their logging is configured to go somewhere else.
1271 */
1272 if (!(Log_destination & LOG_DESTINATION_STDERR))
1273 ereport(LOG,
1274 (errmsg("ending log output to stderr"),
1275 errhint("Future log output will go to log destination \"%s\".",
1276 Log_destination_string)));
1277
1278 whereToSendOutput = DestNone;
1279
1280 /*
1281 * Initialize stats collection subsystem (this does NOT start the
1282 * collector process!)
1283 */
1284 pgstat_init();
1285
1286 /*
1287 * Initialize the autovacuum subsystem (again, no process start yet)
1288 */
1289 autovac_init();
1290
1291 /*
1292 * Load configuration files for client authentication.
1293 */
1294 if (!load_hba())
1295 {
1296 /*
1297 * It makes no sense to continue if we fail to load the HBA file,
1298 * since there is no way to connect to the database in this case.
1299 */
1300 ereport(FATAL,
1301 (errmsg("could not load pg_hba.conf")));
1302 }
1303 if (!load_ident())
1304 {
1305 /*
1306 * We can start up without the IDENT file, although it means that you
1307 * cannot log in using any of the authentication methods that need a
1308 * user name mapping. load_ident() already logged the details of error
1309 * to the log.
1310 */
1311 }
1312
1313 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1314
1315 /*
1316 * On Darwin, libintl replaces setlocale() with a version that calls
1317 * CFLocaleCopyCurrent() when its second argument is "" and every relevant
1318 * environment variable is unset or empty. CFLocaleCopyCurrent() makes
1319 * the process multithreaded. The postmaster calls sigprocmask() and
1320 * calls fork() without an immediate exec(), both of which have undefined
1321 * behavior in a multithreaded program. A multithreaded postmaster is the
1322 * normal case on Windows, which offers neither fork() nor sigprocmask().
1323 */
1324 if (pthread_is_threaded_np() != 0)
1325 ereport(FATAL,
1326 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1327 errmsg("postmaster became multithreaded during startup"),
1328 errhint("Set the LC_ALL environment variable to a valid locale.")));
1329 #endif
1330
1331 /*
1332 * Remember postmaster startup time
1333 */
1334 PgStartTime = GetCurrentTimestamp();
1335 /* PostmasterRandom wants its own copy */
1336 gettimeofday(&random_start_time, NULL);
1337
1338 /*
1339 * We're ready to rock and roll...
1340 */
1341 StartupPID = StartupDataBase();
1342 Assert(StartupPID != 0);
1343 StartupStatus = STARTUP_RUNNING;
1344 pmState = PM_STARTUP;
1345
1346 /* Some workers may be scheduled to start now */
1347 maybe_start_bgworkers();
1348
1349 status = ServerLoop();
1350
1351 /*
1352 * ServerLoop probably shouldn't ever return, but if it does, close down.
1353 */
1354 ExitPostmaster(status != STATUS_OK);
1355
1356 abort(); /* not reached */
1357 }
1358
1359
1360 /*
1361 * on_proc_exit callback to close server's listen sockets
1362 */
1363 static void
CloseServerPorts(int status,Datum arg)1364 CloseServerPorts(int status, Datum arg)
1365 {
1366 int i;
1367
1368 /*
1369 * First, explicitly close all the socket FDs. We used to just let this
1370 * happen implicitly at postmaster exit, but it's better to close them
1371 * before we remove the postmaster.pid lockfile; otherwise there's a race
1372 * condition if a new postmaster wants to re-use the TCP port number.
1373 */
1374 for (i = 0; i < MAXLISTEN; i++)
1375 {
1376 if (ListenSocket[i] != PGINVALID_SOCKET)
1377 {
1378 StreamClose(ListenSocket[i]);
1379 ListenSocket[i] = PGINVALID_SOCKET;
1380 }
1381 }
1382
1383 /*
1384 * Next, remove any filesystem entries for Unix sockets. To avoid race
1385 * conditions against incoming postmasters, this must happen after closing
1386 * the sockets and before removing lock files.
1387 */
1388 RemoveSocketFiles();
1389
1390 /*
1391 * We don't do anything about socket lock files here; those will be
1392 * removed in a later on_proc_exit callback.
1393 */
1394 }
1395
1396 /*
1397 * on_proc_exit callback to delete external_pid_file
1398 */
1399 static void
unlink_external_pid_file(int status,Datum arg)1400 unlink_external_pid_file(int status, Datum arg)
1401 {
1402 if (external_pid_file)
1403 unlink(external_pid_file);
1404 }
1405
1406
1407 /*
1408 * Compute and check the directory paths to files that are part of the
1409 * installation (as deduced from the postgres executable's own location)
1410 */
1411 static void
getInstallationPaths(const char * argv0)1412 getInstallationPaths(const char *argv0)
1413 {
1414 DIR *pdir;
1415
1416 /* Locate the postgres executable itself */
1417 if (find_my_exec(argv0, my_exec_path) < 0)
1418 elog(FATAL, "%s: could not locate my own executable path", argv0);
1419
1420 #ifdef EXEC_BACKEND
1421 /* Locate executable backend before we change working directory */
1422 if (find_other_exec(argv0, "postgres", PG_BACKEND_VERSIONSTR,
1423 postgres_exec_path) < 0)
1424 ereport(FATAL,
1425 (errmsg("%s: could not locate matching postgres executable",
1426 argv0)));
1427 #endif
1428
1429 /*
1430 * Locate the pkglib directory --- this has to be set early in case we try
1431 * to load any modules from it in response to postgresql.conf entries.
1432 */
1433 get_pkglib_path(my_exec_path, pkglib_path);
1434
1435 /*
1436 * Verify that there's a readable directory there; otherwise the Postgres
1437 * installation is incomplete or corrupt. (A typical cause of this
1438 * failure is that the postgres executable has been moved or hardlinked to
1439 * some directory that's not a sibling of the installation lib/
1440 * directory.)
1441 */
1442 pdir = AllocateDir(pkglib_path);
1443 if (pdir == NULL)
1444 ereport(ERROR,
1445 (errcode_for_file_access(),
1446 errmsg("could not open directory \"%s\": %m",
1447 pkglib_path),
1448 errhint("This may indicate an incomplete PostgreSQL installation, or that the file \"%s\" has been moved away from its proper location.",
1449 my_exec_path)));
1450 FreeDir(pdir);
1451
1452 /*
1453 * XXX is it worth similarly checking the share/ directory? If the lib/
1454 * directory is there, then share/ probably is too.
1455 */
1456 }
1457
1458
1459 /*
1460 * Validate the proposed data directory
1461 */
1462 static void
checkDataDir(void)1463 checkDataDir(void)
1464 {
1465 char path[MAXPGPATH];
1466 FILE *fp;
1467 struct stat stat_buf;
1468
1469 Assert(DataDir);
1470
1471 if (stat(DataDir, &stat_buf) != 0)
1472 {
1473 if (errno == ENOENT)
1474 ereport(FATAL,
1475 (errcode_for_file_access(),
1476 errmsg("data directory \"%s\" does not exist",
1477 DataDir)));
1478 else
1479 ereport(FATAL,
1480 (errcode_for_file_access(),
1481 errmsg("could not read permissions of directory \"%s\": %m",
1482 DataDir)));
1483 }
1484
1485 /* eventual chdir would fail anyway, but let's test ... */
1486 if (!S_ISDIR(stat_buf.st_mode))
1487 ereport(FATAL,
1488 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1489 errmsg("specified data directory \"%s\" is not a directory",
1490 DataDir)));
1491
1492 /*
1493 * Check that the directory belongs to my userid; if not, reject.
1494 *
1495 * This check is an essential part of the interlock that prevents two
1496 * postmasters from starting in the same directory (see CreateLockFile()).
1497 * Do not remove or weaken it.
1498 *
1499 * XXX can we safely enable this check on Windows?
1500 */
1501 #if !defined(WIN32) && !defined(__CYGWIN__)
1502 if (stat_buf.st_uid != geteuid())
1503 ereport(FATAL,
1504 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1505 errmsg("data directory \"%s\" has wrong ownership",
1506 DataDir),
1507 errhint("The server must be started by the user that owns the data directory.")));
1508 #endif
1509
1510 /*
1511 * Check if the directory has group or world access. If so, reject.
1512 *
1513 * It would be possible to allow weaker constraints (for example, allow
1514 * group access) but we cannot make a general assumption that that is
1515 * okay; for example there are platforms where nearly all users
1516 * customarily belong to the same group. Perhaps this test should be
1517 * configurable.
1518 *
1519 * XXX temporarily suppress check when on Windows, because there may not
1520 * be proper support for Unix-y file permissions. Need to think of a
1521 * reasonable check to apply on Windows.
1522 */
1523 #if !defined(WIN32) && !defined(__CYGWIN__)
1524 if (stat_buf.st_mode & (S_IRWXG | S_IRWXO))
1525 ereport(FATAL,
1526 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1527 errmsg("data directory \"%s\" has group or world access",
1528 DataDir),
1529 errdetail("Permissions should be u=rwx (0700).")));
1530 #endif
1531
1532 /* Look for PG_VERSION before looking for pg_control */
1533 ValidatePgVersion(DataDir);
1534
1535 snprintf(path, sizeof(path), "%s/global/pg_control", DataDir);
1536
1537 fp = AllocateFile(path, PG_BINARY_R);
1538 if (fp == NULL)
1539 {
1540 write_stderr("%s: could not find the database system\n"
1541 "Expected to find it in the directory \"%s\",\n"
1542 "but could not open file \"%s\": %s\n",
1543 progname, DataDir, path, strerror(errno));
1544 ExitPostmaster(2);
1545 }
1546 FreeFile(fp);
1547 }
1548
1549 /*
1550 * Determine how long should we let ServerLoop sleep.
1551 *
1552 * In normal conditions we wait at most one minute, to ensure that the other
1553 * background tasks handled by ServerLoop get done even when no requests are
1554 * arriving. However, if there are background workers waiting to be started,
1555 * we don't actually sleep so that they are quickly serviced. Other exception
1556 * cases are as shown in the code.
1557 */
1558 static void
DetermineSleepTime(struct timeval * timeout)1559 DetermineSleepTime(struct timeval * timeout)
1560 {
1561 TimestampTz next_wakeup = 0;
1562
1563 /*
1564 * Normal case: either there are no background workers at all, or we're in
1565 * a shutdown sequence (during which we ignore bgworkers altogether).
1566 */
1567 if (Shutdown > NoShutdown ||
1568 (!StartWorkerNeeded && !HaveCrashedWorker))
1569 {
1570 if (AbortStartTime != 0)
1571 {
1572 /* time left to abort; clamp to 0 in case it already expired */
1573 timeout->tv_sec = SIGKILL_CHILDREN_AFTER_SECS -
1574 (time(NULL) - AbortStartTime);
1575 timeout->tv_sec = Max(timeout->tv_sec, 0);
1576 timeout->tv_usec = 0;
1577 }
1578 else
1579 {
1580 timeout->tv_sec = 60;
1581 timeout->tv_usec = 0;
1582 }
1583 return;
1584 }
1585
1586 if (StartWorkerNeeded)
1587 {
1588 timeout->tv_sec = 0;
1589 timeout->tv_usec = 0;
1590 return;
1591 }
1592
1593 if (HaveCrashedWorker)
1594 {
1595 slist_mutable_iter siter;
1596
1597 /*
1598 * When there are crashed bgworkers, we sleep just long enough that
1599 * they are restarted when they request to be. Scan the list to
1600 * determine the minimum of all wakeup times according to most recent
1601 * crash time and requested restart interval.
1602 */
1603 slist_foreach_modify(siter, &BackgroundWorkerList)
1604 {
1605 RegisteredBgWorker *rw;
1606 TimestampTz this_wakeup;
1607
1608 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
1609
1610 if (rw->rw_crashed_at == 0)
1611 continue;
1612
1613 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART
1614 || rw->rw_terminate)
1615 {
1616 ForgetBackgroundWorker(&siter);
1617 continue;
1618 }
1619
1620 this_wakeup = TimestampTzPlusMilliseconds(rw->rw_crashed_at,
1621 1000L * rw->rw_worker.bgw_restart_time);
1622 if (next_wakeup == 0 || this_wakeup < next_wakeup)
1623 next_wakeup = this_wakeup;
1624 }
1625 }
1626
1627 if (next_wakeup != 0)
1628 {
1629 long secs;
1630 int microsecs;
1631
1632 TimestampDifference(GetCurrentTimestamp(), next_wakeup,
1633 &secs, µsecs);
1634 timeout->tv_sec = secs;
1635 timeout->tv_usec = microsecs;
1636
1637 /* Ensure we don't exceed one minute */
1638 if (timeout->tv_sec > 60)
1639 {
1640 timeout->tv_sec = 60;
1641 timeout->tv_usec = 0;
1642 }
1643 }
1644 else
1645 {
1646 timeout->tv_sec = 60;
1647 timeout->tv_usec = 0;
1648 }
1649 }
1650
1651 /*
1652 * Main idle loop of postmaster
1653 *
1654 * NB: Needs to be called with signals blocked
1655 */
1656 static int
ServerLoop(void)1657 ServerLoop(void)
1658 {
1659 fd_set readmask;
1660 int nSockets;
1661 time_t last_lockfile_recheck_time,
1662 last_touch_time;
1663
1664 last_lockfile_recheck_time = last_touch_time = time(NULL);
1665
1666 nSockets = initMasks(&readmask);
1667
1668 for (;;)
1669 {
1670 fd_set rmask;
1671 int selres;
1672 time_t now;
1673
1674 /*
1675 * Wait for a connection request to arrive.
1676 *
1677 * We block all signals except while sleeping. That makes it safe for
1678 * signal handlers, which again block all signals while executing, to
1679 * do nontrivial work.
1680 *
1681 * If we are in PM_WAIT_DEAD_END state, then we don't want to accept
1682 * any new connections, so we don't call select(), and just sleep.
1683 */
1684 memcpy((char *) &rmask, (char *) &readmask, sizeof(fd_set));
1685
1686 if (pmState == PM_WAIT_DEAD_END)
1687 {
1688 PG_SETMASK(&UnBlockSig);
1689
1690 pg_usleep(100000L); /* 100 msec seems reasonable */
1691 selres = 0;
1692
1693 PG_SETMASK(&BlockSig);
1694 }
1695 else
1696 {
1697 /* must set timeout each time; some OSes change it! */
1698 struct timeval timeout;
1699
1700 /* Needs to run with blocked signals! */
1701 DetermineSleepTime(&timeout);
1702
1703 PG_SETMASK(&UnBlockSig);
1704
1705 selres = select(nSockets, &rmask, NULL, NULL, &timeout);
1706
1707 PG_SETMASK(&BlockSig);
1708 }
1709
1710 /* Now check the select() result */
1711 if (selres < 0)
1712 {
1713 if (errno != EINTR && errno != EWOULDBLOCK)
1714 {
1715 ereport(LOG,
1716 (errcode_for_socket_access(),
1717 errmsg("select() failed in postmaster: %m")));
1718 return STATUS_ERROR;
1719 }
1720 }
1721
1722 /*
1723 * New connection pending on any of our sockets? If so, fork a child
1724 * process to deal with it.
1725 */
1726 if (selres > 0)
1727 {
1728 int i;
1729
1730 for (i = 0; i < MAXLISTEN; i++)
1731 {
1732 if (ListenSocket[i] == PGINVALID_SOCKET)
1733 break;
1734 if (FD_ISSET(ListenSocket[i], &rmask))
1735 {
1736 Port *port;
1737
1738 port = ConnCreate(ListenSocket[i]);
1739 if (port)
1740 {
1741 BackendStartup(port);
1742
1743 /*
1744 * We no longer need the open socket or port structure
1745 * in this process
1746 */
1747 StreamClose(port->sock);
1748 ConnFree(port);
1749 }
1750 }
1751 }
1752 }
1753
1754 /* If we have lost the log collector, try to start a new one */
1755 if (SysLoggerPID == 0 && Logging_collector)
1756 SysLoggerPID = SysLogger_Start();
1757
1758 /*
1759 * If no background writer process is running, and we are not in a
1760 * state that prevents it, start one. It doesn't matter if this
1761 * fails, we'll just try again later. Likewise for the checkpointer.
1762 */
1763 if (pmState == PM_RUN || pmState == PM_RECOVERY ||
1764 pmState == PM_HOT_STANDBY)
1765 {
1766 if (CheckpointerPID == 0)
1767 CheckpointerPID = StartCheckpointer();
1768 if (BgWriterPID == 0)
1769 BgWriterPID = StartBackgroundWriter();
1770 }
1771
1772 /*
1773 * Likewise, if we have lost the walwriter process, try to start a new
1774 * one. But this is needed only in normal operation (else we cannot
1775 * be writing any new WAL).
1776 */
1777 if (WalWriterPID == 0 && pmState == PM_RUN)
1778 WalWriterPID = StartWalWriter();
1779
1780 /*
1781 * If we have lost the autovacuum launcher, try to start a new one. We
1782 * don't want autovacuum to run in binary upgrade mode because
1783 * autovacuum might update relfrozenxid for empty tables before the
1784 * physical files are put in place.
1785 */
1786 if (!IsBinaryUpgrade && AutoVacPID == 0 &&
1787 (AutoVacuumingActive() || start_autovac_launcher) &&
1788 pmState == PM_RUN)
1789 {
1790 AutoVacPID = StartAutoVacLauncher();
1791 if (AutoVacPID != 0)
1792 start_autovac_launcher = false; /* signal processed */
1793 }
1794
1795 /* If we have lost the stats collector, try to start a new one */
1796 if (PgStatPID == 0 &&
1797 (pmState == PM_RUN || pmState == PM_HOT_STANDBY))
1798 PgStatPID = pgstat_start();
1799
1800 /* If we have lost the archiver, try to start a new one. */
1801 if (PgArchPID == 0 && PgArchStartupAllowed())
1802 PgArchPID = pgarch_start();
1803
1804 /* If we need to signal the autovacuum launcher, do so now */
1805 if (avlauncher_needs_signal)
1806 {
1807 avlauncher_needs_signal = false;
1808 if (AutoVacPID != 0)
1809 kill(AutoVacPID, SIGUSR2);
1810 }
1811
1812 /* If we need to start a WAL receiver, try to do that now */
1813 if (WalReceiverRequested)
1814 MaybeStartWalReceiver();
1815
1816 /* Get other worker processes running, if needed */
1817 if (StartWorkerNeeded || HaveCrashedWorker)
1818 maybe_start_bgworkers();
1819
1820 #ifdef HAVE_PTHREAD_IS_THREADED_NP
1821
1822 /*
1823 * With assertions enabled, check regularly for appearance of
1824 * additional threads. All builds check at start and exit.
1825 */
1826 Assert(pthread_is_threaded_np() == 0);
1827 #endif
1828
1829 /*
1830 * Lastly, check to see if it's time to do some things that we don't
1831 * want to do every single time through the loop, because they're a
1832 * bit expensive. Note that there's up to a minute of slop in when
1833 * these tasks will be performed, since DetermineSleepTime() will let
1834 * us sleep at most that long; except for SIGKILL timeout which has
1835 * special-case logic there.
1836 */
1837 now = time(NULL);
1838
1839 /*
1840 * If we already sent SIGQUIT to children and they are slow to shut
1841 * down, it's time to send them SIGKILL. This doesn't happen
1842 * normally, but under certain conditions backends can get stuck while
1843 * shutting down. This is a last measure to get them unwedged.
1844 *
1845 * Note we also do this during recovery from a process crash.
1846 */
1847 if ((Shutdown >= ImmediateShutdown || (FatalError && !SendStop)) &&
1848 AbortStartTime != 0 &&
1849 (now - AbortStartTime) >= SIGKILL_CHILDREN_AFTER_SECS)
1850 {
1851 /* We were gentle with them before. Not anymore */
1852 TerminateChildren(SIGKILL);
1853 /* reset flag so we don't SIGKILL again */
1854 AbortStartTime = 0;
1855 }
1856
1857 /*
1858 * Once a minute, verify that postmaster.pid hasn't been removed or
1859 * overwritten. If it has, we force a shutdown. This avoids having
1860 * postmasters and child processes hanging around after their database
1861 * is gone, and maybe causing problems if a new database cluster is
1862 * created in the same place. It also provides some protection
1863 * against a DBA foolishly removing postmaster.pid and manually
1864 * starting a new postmaster. Data corruption is likely to ensue from
1865 * that anyway, but we can minimize the damage by aborting ASAP.
1866 */
1867 if (now - last_lockfile_recheck_time >= 1 * SECS_PER_MINUTE)
1868 {
1869 if (!RecheckDataDirLockFile())
1870 {
1871 ereport(LOG,
1872 (errmsg("performing immediate shutdown because data directory lock file is invalid")));
1873 kill(MyProcPid, SIGQUIT);
1874 }
1875 last_lockfile_recheck_time = now;
1876 }
1877
1878 /*
1879 * Touch Unix socket and lock files every 58 minutes, to ensure that
1880 * they are not removed by overzealous /tmp-cleaning tasks. We assume
1881 * no one runs cleaners with cutoff times of less than an hour ...
1882 */
1883 if (now - last_touch_time >= 58 * SECS_PER_MINUTE)
1884 {
1885 TouchSocketFiles();
1886 TouchSocketLockFiles();
1887 last_touch_time = now;
1888 }
1889 }
1890 }
1891
1892 /*
1893 * Initialise the masks for select() for the ports we are listening on.
1894 * Return the number of sockets to listen on.
1895 */
1896 static int
initMasks(fd_set * rmask)1897 initMasks(fd_set *rmask)
1898 {
1899 int maxsock = -1;
1900 int i;
1901
1902 FD_ZERO(rmask);
1903
1904 for (i = 0; i < MAXLISTEN; i++)
1905 {
1906 int fd = ListenSocket[i];
1907
1908 if (fd == PGINVALID_SOCKET)
1909 break;
1910 FD_SET(fd, rmask);
1911
1912 if (fd > maxsock)
1913 maxsock = fd;
1914 }
1915
1916 return maxsock + 1;
1917 }
1918
1919
1920 /*
1921 * Read a client's startup packet and do something according to it.
1922 *
1923 * Returns STATUS_OK or STATUS_ERROR, or might call ereport(FATAL) and
1924 * not return at all.
1925 *
1926 * (Note that ereport(FATAL) stuff is sent to the client, so only use it
1927 * if that's what you want. Return STATUS_ERROR if you don't want to
1928 * send anything to the client, which would typically be appropriate
1929 * if we detect a communications failure.)
1930 */
1931 static int
ProcessStartupPacket(Port * port,bool SSLdone)1932 ProcessStartupPacket(Port *port, bool SSLdone)
1933 {
1934 int32 len;
1935 void *buf;
1936 ProtocolVersion proto;
1937 MemoryContext oldcontext;
1938
1939 pq_startmsgread();
1940 if (pq_getbytes((char *) &len, 4) == EOF)
1941 {
1942 /*
1943 * EOF after SSLdone probably means the client didn't like our
1944 * response to NEGOTIATE_SSL_CODE. That's not an error condition, so
1945 * don't clutter the log with a complaint.
1946 */
1947 if (!SSLdone)
1948 ereport(COMMERROR,
1949 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1950 errmsg("incomplete startup packet")));
1951 return STATUS_ERROR;
1952 }
1953
1954 len = ntohl(len);
1955 len -= 4;
1956
1957 if (len < (int32) sizeof(ProtocolVersion) ||
1958 len > MAX_STARTUP_PACKET_LENGTH)
1959 {
1960 ereport(COMMERROR,
1961 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1962 errmsg("invalid length of startup packet")));
1963 return STATUS_ERROR;
1964 }
1965
1966 /*
1967 * Allocate at least the size of an old-style startup packet, plus one
1968 * extra byte, and make sure all are zeroes. This ensures we will have
1969 * null termination of all strings, in both fixed- and variable-length
1970 * packet layouts.
1971 */
1972 if (len <= (int32) sizeof(StartupPacket))
1973 buf = palloc0(sizeof(StartupPacket) + 1);
1974 else
1975 buf = palloc0(len + 1);
1976
1977 if (pq_getbytes(buf, len) == EOF)
1978 {
1979 ereport(COMMERROR,
1980 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1981 errmsg("incomplete startup packet")));
1982 return STATUS_ERROR;
1983 }
1984 pq_endmsgread();
1985
1986 /*
1987 * The first field is either a protocol version number or a special
1988 * request code.
1989 */
1990 port->proto = proto = ntohl(*((ProtocolVersion *) buf));
1991
1992 if (proto == CANCEL_REQUEST_CODE)
1993 {
1994 processCancelRequest(port, buf);
1995 /* Not really an error, but we don't want to proceed further */
1996 return STATUS_ERROR;
1997 }
1998
1999 if (proto == NEGOTIATE_SSL_CODE && !SSLdone)
2000 {
2001 char SSLok;
2002
2003 #ifdef USE_SSL
2004 /* No SSL when disabled or on Unix sockets */
2005 if (!EnableSSL || IS_AF_UNIX(port->laddr.addr.ss_family))
2006 SSLok = 'N';
2007 else
2008 SSLok = 'S'; /* Support for SSL */
2009 #else
2010 SSLok = 'N'; /* No support for SSL */
2011 #endif
2012
2013 retry1:
2014 if (send(port->sock, &SSLok, 1, 0) != 1)
2015 {
2016 if (errno == EINTR)
2017 goto retry1; /* if interrupted, just retry */
2018 ereport(COMMERROR,
2019 (errcode_for_socket_access(),
2020 errmsg("failed to send SSL negotiation response: %m")));
2021 return STATUS_ERROR; /* close the connection */
2022 }
2023
2024 #ifdef USE_SSL
2025 if (SSLok == 'S' && secure_open_server(port) == -1)
2026 return STATUS_ERROR;
2027 #endif
2028
2029 /*
2030 * At this point we should have no data already buffered. If we do,
2031 * it was received before we performed the SSL handshake, so it wasn't
2032 * encrypted and indeed may have been injected by a man-in-the-middle.
2033 * We report this case to the client.
2034 */
2035 if (pq_buffer_has_data())
2036 ereport(FATAL,
2037 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2038 errmsg("received unencrypted data after SSL request"),
2039 errdetail("This could be either a client-software bug or evidence of an attempted man-in-the-middle attack.")));
2040
2041 /* regular startup packet, cancel, etc packet should follow... */
2042 /* but not another SSL negotiation request */
2043 return ProcessStartupPacket(port, true);
2044 }
2045
2046 /* Could add additional special packet types here */
2047
2048 /*
2049 * Set FrontendProtocol now so that ereport() knows what format to send if
2050 * we fail during startup.
2051 */
2052 FrontendProtocol = proto;
2053
2054 /* Check that the major protocol version is in range. */
2055 if (PG_PROTOCOL_MAJOR(proto) < PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST) ||
2056 PG_PROTOCOL_MAJOR(proto) > PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST))
2057 ereport(FATAL,
2058 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2059 errmsg("unsupported frontend protocol %u.%u: server supports %u.0 to %u.%u",
2060 PG_PROTOCOL_MAJOR(proto), PG_PROTOCOL_MINOR(proto),
2061 PG_PROTOCOL_MAJOR(PG_PROTOCOL_EARLIEST),
2062 PG_PROTOCOL_MAJOR(PG_PROTOCOL_LATEST),
2063 PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST))));
2064
2065 /*
2066 * Now fetch parameters out of startup packet and save them into the Port
2067 * structure. All data structures attached to the Port struct must be
2068 * allocated in TopMemoryContext so that they will remain available in a
2069 * running backend (even after PostmasterContext is destroyed). We need
2070 * not worry about leaking this storage on failure, since we aren't in the
2071 * postmaster process anymore.
2072 */
2073 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
2074
2075 if (PG_PROTOCOL_MAJOR(proto) >= 3)
2076 {
2077 int32 offset = sizeof(ProtocolVersion);
2078 List *unrecognized_protocol_options = NIL;
2079
2080 /*
2081 * Scan packet body for name/option pairs. We can assume any string
2082 * beginning within the packet body is null-terminated, thanks to
2083 * zeroing extra byte above.
2084 */
2085 port->guc_options = NIL;
2086
2087 while (offset < len)
2088 {
2089 char *nameptr = ((char *) buf) + offset;
2090 int32 valoffset;
2091 char *valptr;
2092
2093 if (*nameptr == '\0')
2094 break; /* found packet terminator */
2095 valoffset = offset + strlen(nameptr) + 1;
2096 if (valoffset >= len)
2097 break; /* missing value, will complain below */
2098 valptr = ((char *) buf) + valoffset;
2099
2100 if (strcmp(nameptr, "database") == 0)
2101 port->database_name = pstrdup(valptr);
2102 else if (strcmp(nameptr, "user") == 0)
2103 port->user_name = pstrdup(valptr);
2104 else if (strcmp(nameptr, "options") == 0)
2105 port->cmdline_options = pstrdup(valptr);
2106 else if (strcmp(nameptr, "replication") == 0)
2107 {
2108 /*
2109 * Due to backward compatibility concerns the replication
2110 * parameter is a hybrid beast which allows the value to be
2111 * either boolean or the string 'database'. The latter
2112 * connects to a specific database which is e.g. required for
2113 * logical decoding while.
2114 */
2115 if (strcmp(valptr, "database") == 0)
2116 {
2117 am_walsender = true;
2118 am_db_walsender = true;
2119 }
2120 else if (!parse_bool(valptr, &am_walsender))
2121 ereport(FATAL,
2122 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
2123 errmsg("invalid value for parameter \"%s\": \"%s\"",
2124 "replication",
2125 valptr),
2126 errhint("Valid values are: \"false\", 0, \"true\", 1, \"database\".")));
2127 }
2128 else if (strncmp(nameptr, "_pq_.", 5) == 0)
2129 {
2130 /*
2131 * Any option beginning with _pq_. is reserved for use as a
2132 * protocol-level option, but at present no such options are
2133 * defined.
2134 */
2135 unrecognized_protocol_options =
2136 lappend(unrecognized_protocol_options, pstrdup(nameptr));
2137 }
2138 else
2139 {
2140 /* Assume it's a generic GUC option */
2141 port->guc_options = lappend(port->guc_options,
2142 pstrdup(nameptr));
2143 port->guc_options = lappend(port->guc_options,
2144 pstrdup(valptr));
2145 }
2146 offset = valoffset + strlen(valptr) + 1;
2147 }
2148
2149 /*
2150 * If we didn't find a packet terminator exactly at the end of the
2151 * given packet length, complain.
2152 */
2153 if (offset != len - 1)
2154 ereport(FATAL,
2155 (errcode(ERRCODE_PROTOCOL_VIOLATION),
2156 errmsg("invalid startup packet layout: expected terminator as last byte")));
2157
2158 /*
2159 * If the client requested a newer protocol version or if the client
2160 * requested any protocol options we didn't recognize, let them know
2161 * the newest minor protocol version we do support and the names of any
2162 * unrecognized options.
2163 */
2164 if (PG_PROTOCOL_MINOR(proto) > PG_PROTOCOL_MINOR(PG_PROTOCOL_LATEST) ||
2165 unrecognized_protocol_options != NIL)
2166 SendNegotiateProtocolVersion(unrecognized_protocol_options);
2167 }
2168 else
2169 {
2170 /*
2171 * Get the parameters from the old-style, fixed-width-fields startup
2172 * packet as C strings. The packet destination was cleared first so a
2173 * short packet has zeros silently added. We have to be prepared to
2174 * truncate the pstrdup result for oversize fields, though.
2175 */
2176 StartupPacket *packet = (StartupPacket *) buf;
2177
2178 port->database_name = pstrdup(packet->database);
2179 if (strlen(port->database_name) > sizeof(packet->database))
2180 port->database_name[sizeof(packet->database)] = '\0';
2181 port->user_name = pstrdup(packet->user);
2182 if (strlen(port->user_name) > sizeof(packet->user))
2183 port->user_name[sizeof(packet->user)] = '\0';
2184 port->cmdline_options = pstrdup(packet->options);
2185 if (strlen(port->cmdline_options) > sizeof(packet->options))
2186 port->cmdline_options[sizeof(packet->options)] = '\0';
2187 port->guc_options = NIL;
2188 }
2189
2190 /* Check a user name was given. */
2191 if (port->user_name == NULL || port->user_name[0] == '\0')
2192 ereport(FATAL,
2193 (errcode(ERRCODE_INVALID_AUTHORIZATION_SPECIFICATION),
2194 errmsg("no PostgreSQL user name specified in startup packet")));
2195
2196 /* The database defaults to the user name. */
2197 if (port->database_name == NULL || port->database_name[0] == '\0')
2198 port->database_name = pstrdup(port->user_name);
2199
2200 if (Db_user_namespace)
2201 {
2202 /*
2203 * If user@, it is a global user, remove '@'. We only want to do this
2204 * if there is an '@' at the end and no earlier in the user string or
2205 * they may fake as a local user of another database attaching to this
2206 * database.
2207 */
2208 if (strchr(port->user_name, '@') ==
2209 port->user_name + strlen(port->user_name) - 1)
2210 *strchr(port->user_name, '@') = '\0';
2211 else
2212 {
2213 /* Append '@' and dbname */
2214 port->user_name = psprintf("%s@%s", port->user_name, port->database_name);
2215 }
2216 }
2217
2218 /*
2219 * Truncate given database and user names to length of a Postgres name.
2220 * This avoids lookup failures when overlength names are given.
2221 */
2222 if (strlen(port->database_name) >= NAMEDATALEN)
2223 port->database_name[NAMEDATALEN - 1] = '\0';
2224 if (strlen(port->user_name) >= NAMEDATALEN)
2225 port->user_name[NAMEDATALEN - 1] = '\0';
2226
2227 /*
2228 * Normal walsender backends, e.g. for streaming replication, are not
2229 * connected to a particular database. But walsenders used for logical
2230 * replication need to connect to a specific database. We allow streaming
2231 * replication commands to be issued even if connected to a database as it
2232 * can make sense to first make a basebackup and then stream changes
2233 * starting from that.
2234 */
2235 if (am_walsender && !am_db_walsender)
2236 port->database_name[0] = '\0';
2237
2238 /*
2239 * Done putting stuff in TopMemoryContext.
2240 */
2241 MemoryContextSwitchTo(oldcontext);
2242
2243 /*
2244 * If we're going to reject the connection due to database state, say so
2245 * now instead of wasting cycles on an authentication exchange. (This also
2246 * allows a pg_ping utility to be written.)
2247 */
2248 switch (port->canAcceptConnections)
2249 {
2250 case CAC_STARTUP:
2251 ereport(FATAL,
2252 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2253 errmsg("the database system is starting up")));
2254 break;
2255 case CAC_SHUTDOWN:
2256 ereport(FATAL,
2257 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2258 errmsg("the database system is shutting down")));
2259 break;
2260 case CAC_RECOVERY:
2261 ereport(FATAL,
2262 (errcode(ERRCODE_CANNOT_CONNECT_NOW),
2263 errmsg("the database system is in recovery mode")));
2264 break;
2265 case CAC_TOOMANY:
2266 ereport(FATAL,
2267 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
2268 errmsg("sorry, too many clients already")));
2269 break;
2270 case CAC_SUPERUSER:
2271 /* OK for now, will check in InitPostgres */
2272 break;
2273 case CAC_OK:
2274 break;
2275 }
2276
2277 return STATUS_OK;
2278 }
2279
2280 /*
2281 * Send a NegotiateProtocolVersion to the client. This lets the client know
2282 * that they have requested a newer minor protocol version than we are able
2283 * to speak. We'll speak the highest version we know about; the client can,
2284 * of course, abandon the connection if that's a problem.
2285 *
2286 * We also include in the response a list of protocol options we didn't
2287 * understand. This allows clients to include optional parameters that might
2288 * be present either in newer protocol versions or third-party protocol
2289 * extensions without fear of having to reconnect if those options are not
2290 * understood, while at the same time making certain that the client is aware
2291 * of which options were actually accepted.
2292 */
2293 static void
SendNegotiateProtocolVersion(List * unrecognized_protocol_options)2294 SendNegotiateProtocolVersion(List *unrecognized_protocol_options)
2295 {
2296 StringInfoData buf;
2297 ListCell *lc;
2298
2299 pq_beginmessage(&buf, 'v'); /* NegotiateProtocolVersion */
2300 pq_sendint(&buf, PG_PROTOCOL_LATEST, 4);
2301 pq_sendint(&buf, list_length(unrecognized_protocol_options), 4);
2302 foreach(lc, unrecognized_protocol_options)
2303 pq_sendstring(&buf, lfirst(lc));
2304 pq_endmessage(&buf);
2305
2306 /* no need to flush, some other message will follow */
2307 }
2308
2309 /*
2310 * The client has sent a cancel request packet, not a normal
2311 * start-a-new-connection packet. Perform the necessary processing.
2312 * Nothing is sent back to the client.
2313 */
2314 static void
processCancelRequest(Port * port,void * pkt)2315 processCancelRequest(Port *port, void *pkt)
2316 {
2317 CancelRequestPacket *canc = (CancelRequestPacket *) pkt;
2318 int backendPID;
2319 long cancelAuthCode;
2320 Backend *bp;
2321
2322 #ifndef EXEC_BACKEND
2323 dlist_iter iter;
2324 #else
2325 int i;
2326 #endif
2327
2328 backendPID = (int) ntohl(canc->backendPID);
2329 cancelAuthCode = (long) ntohl(canc->cancelAuthCode);
2330
2331 /*
2332 * See if we have a matching backend. In the EXEC_BACKEND case, we can no
2333 * longer access the postmaster's own backend list, and must rely on the
2334 * duplicate array in shared memory.
2335 */
2336 #ifndef EXEC_BACKEND
2337 dlist_foreach(iter, &BackendList)
2338 {
2339 bp = dlist_container(Backend, elem, iter.cur);
2340 #else
2341 for (i = MaxLivePostmasterChildren() - 1; i >= 0; i--)
2342 {
2343 bp = (Backend *) &ShmemBackendArray[i];
2344 #endif
2345 if (bp->pid == backendPID)
2346 {
2347 if (bp->cancel_key == cancelAuthCode)
2348 {
2349 /* Found a match; signal that backend to cancel current op */
2350 ereport(DEBUG2,
2351 (errmsg_internal("processing cancel request: sending SIGINT to process %d",
2352 backendPID)));
2353 signal_child(bp->pid, SIGINT);
2354 }
2355 else
2356 /* Right PID, wrong key: no way, Jose */
2357 ereport(LOG,
2358 (errmsg("wrong key in cancel request for process %d",
2359 backendPID)));
2360 return;
2361 }
2362 }
2363
2364 /* No matching backend */
2365 ereport(LOG,
2366 (errmsg("PID %d in cancel request did not match any process",
2367 backendPID)));
2368 }
2369
2370 /*
2371 * canAcceptConnections --- check to see if database state allows connections
2372 * of the specified type. backend_type can be BACKEND_TYPE_NORMAL,
2373 * BACKEND_TYPE_AUTOVAC, or BACKEND_TYPE_BGWORKER. (Note that we don't yet
2374 * know whether a NORMAL connection might turn into a walsender.)
2375 */
2376 static CAC_state
2377 canAcceptConnections(int backend_type)
2378 {
2379 CAC_state result = CAC_OK;
2380
2381 /*
2382 * Can't start backends when in startup/shutdown/inconsistent recovery
2383 * state. We treat autovac workers the same as user backends for this
2384 * purpose. However, bgworkers are excluded from this test; we expect
2385 * bgworker_should_start_now() decided whether the DB state allows them.
2386 */
2387 if (pmState != PM_RUN && pmState != PM_HOT_STANDBY &&
2388 backend_type != BACKEND_TYPE_BGWORKER)
2389 {
2390 if (Shutdown > NoShutdown)
2391 return CAC_SHUTDOWN; /* shutdown is pending */
2392 else if (!FatalError &&
2393 (pmState == PM_STARTUP ||
2394 pmState == PM_RECOVERY))
2395 return CAC_STARTUP; /* normal startup */
2396 else
2397 return CAC_RECOVERY; /* else must be crash recovery */
2398 }
2399
2400 /*
2401 * "Smart shutdown" restrictions are applied only to normal connections,
2402 * not to autovac workers or bgworkers. When only superusers can connect,
2403 * we return CAC_SUPERUSER to indicate that superuserness must be checked
2404 * later. Note that neither CAC_OK nor CAC_SUPERUSER can safely be
2405 * returned until we have checked for too many children.
2406 */
2407 if (connsAllowed != ALLOW_ALL_CONNS &&
2408 backend_type == BACKEND_TYPE_NORMAL)
2409 {
2410 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
2411 result = CAC_SUPERUSER; /* allow superusers only */
2412 else
2413 return CAC_SHUTDOWN; /* shutdown is pending */
2414 }
2415
2416 /*
2417 * Don't start too many children.
2418 *
2419 * We allow more connections here than we can have backends because some
2420 * might still be authenticating; they might fail auth, or some existing
2421 * backend might exit before the auth cycle is completed. The exact
2422 * MaxBackends limit is enforced when a new backend tries to join the
2423 * shared-inval backend array.
2424 *
2425 * The limit here must match the sizes of the per-child-process arrays;
2426 * see comments for MaxLivePostmasterChildren().
2427 */
2428 if (CountChildren(BACKEND_TYPE_ALL) >= MaxLivePostmasterChildren())
2429 result = CAC_TOOMANY;
2430
2431 return result;
2432 }
2433
2434
2435 /*
2436 * ConnCreate -- create a local connection data structure
2437 *
2438 * Returns NULL on failure, other than out-of-memory which is fatal.
2439 */
2440 static Port *
2441 ConnCreate(int serverFd)
2442 {
2443 Port *port;
2444
2445 if (!(port = (Port *) calloc(1, sizeof(Port))))
2446 {
2447 ereport(LOG,
2448 (errcode(ERRCODE_OUT_OF_MEMORY),
2449 errmsg("out of memory")));
2450 ExitPostmaster(1);
2451 }
2452
2453 if (StreamConnection(serverFd, port) != STATUS_OK)
2454 {
2455 if (port->sock != PGINVALID_SOCKET)
2456 StreamClose(port->sock);
2457 ConnFree(port);
2458 return NULL;
2459 }
2460
2461 /*
2462 * Precompute password salt values to use for this connection. It's
2463 * slightly annoying to do this long in advance of knowing whether we'll
2464 * need 'em or not, but we must do the random() calls before we fork, not
2465 * after. Else the postmaster's random sequence won't get advanced, and
2466 * all backends would end up using the same salt...
2467 */
2468 RandomSalt(port->md5Salt);
2469
2470 /*
2471 * Allocate GSSAPI specific state struct
2472 */
2473 #ifndef EXEC_BACKEND
2474 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
2475 port->gss = (pg_gssinfo *) calloc(1, sizeof(pg_gssinfo));
2476 if (!port->gss)
2477 {
2478 ereport(LOG,
2479 (errcode(ERRCODE_OUT_OF_MEMORY),
2480 errmsg("out of memory")));
2481 ExitPostmaster(1);
2482 }
2483 #endif
2484 #endif
2485
2486 return port;
2487 }
2488
2489
2490 /*
2491 * ConnFree -- free a local connection data structure
2492 */
2493 static void
2494 ConnFree(Port *conn)
2495 {
2496 #ifdef USE_SSL
2497 secure_close(conn);
2498 #endif
2499 if (conn->gss)
2500 free(conn->gss);
2501 free(conn);
2502 }
2503
2504
2505 /*
2506 * ClosePostmasterPorts -- close all the postmaster's open sockets
2507 *
2508 * This is called during child process startup to release file descriptors
2509 * that are not needed by that child process. The postmaster still has
2510 * them open, of course.
2511 *
2512 * Note: we pass am_syslogger as a boolean because we don't want to set
2513 * the global variable yet when this is called.
2514 */
2515 void
2516 ClosePostmasterPorts(bool am_syslogger)
2517 {
2518 int i;
2519
2520 #ifndef WIN32
2521
2522 /*
2523 * Close the write end of postmaster death watch pipe. It's important to
2524 * do this as early as possible, so that if postmaster dies, others won't
2525 * think that it's still running because we're holding the pipe open.
2526 */
2527 if (close(postmaster_alive_fds[POSTMASTER_FD_OWN]))
2528 ereport(FATAL,
2529 (errcode_for_file_access(),
2530 errmsg_internal("could not close postmaster death monitoring pipe in child process: %m")));
2531 postmaster_alive_fds[POSTMASTER_FD_OWN] = -1;
2532 #endif
2533
2534 /* Close the listen sockets */
2535 for (i = 0; i < MAXLISTEN; i++)
2536 {
2537 if (ListenSocket[i] != PGINVALID_SOCKET)
2538 {
2539 StreamClose(ListenSocket[i]);
2540 ListenSocket[i] = PGINVALID_SOCKET;
2541 }
2542 }
2543
2544 /* If using syslogger, close the read side of the pipe */
2545 if (!am_syslogger)
2546 {
2547 #ifndef WIN32
2548 if (syslogPipe[0] >= 0)
2549 close(syslogPipe[0]);
2550 syslogPipe[0] = -1;
2551 #else
2552 if (syslogPipe[0])
2553 CloseHandle(syslogPipe[0]);
2554 syslogPipe[0] = 0;
2555 #endif
2556 }
2557
2558 #ifdef USE_BONJOUR
2559 /* If using Bonjour, close the connection to the mDNS daemon */
2560 if (bonjour_sdref)
2561 close(DNSServiceRefSockFD(bonjour_sdref));
2562 #endif
2563 }
2564
2565
2566 /*
2567 * reset_shared -- reset shared memory and semaphores
2568 */
2569 static void
2570 reset_shared(int port)
2571 {
2572 /*
2573 * Create or re-create shared memory and semaphores.
2574 *
2575 * Note: in each "cycle of life" we will normally assign the same IPC keys
2576 * (if using SysV shmem and/or semas), since the port number is used to
2577 * determine IPC keys. This helps ensure that we will clean up dead IPC
2578 * objects if the postmaster crashes and is restarted.
2579 */
2580 CreateSharedMemoryAndSemaphores(port);
2581 }
2582
2583
2584 /*
2585 * SIGHUP -- reread config files, and tell children to do same
2586 */
2587 static void
2588 SIGHUP_handler(SIGNAL_ARGS)
2589 {
2590 int save_errno = errno;
2591
2592 /*
2593 * We rely on the signal mechanism to have blocked all signals ... except
2594 * on Windows, which lacks sigaction(), so we have to do it manually.
2595 */
2596 #ifdef WIN32
2597 PG_SETMASK(&BlockSig);
2598 #endif
2599
2600 if (Shutdown <= SmartShutdown)
2601 {
2602 ereport(LOG,
2603 (errmsg("received SIGHUP, reloading configuration files")));
2604 ProcessConfigFile(PGC_SIGHUP);
2605 SignalChildren(SIGHUP);
2606 if (StartupPID != 0)
2607 signal_child(StartupPID, SIGHUP);
2608 if (BgWriterPID != 0)
2609 signal_child(BgWriterPID, SIGHUP);
2610 if (CheckpointerPID != 0)
2611 signal_child(CheckpointerPID, SIGHUP);
2612 if (WalWriterPID != 0)
2613 signal_child(WalWriterPID, SIGHUP);
2614 if (WalReceiverPID != 0)
2615 signal_child(WalReceiverPID, SIGHUP);
2616 if (AutoVacPID != 0)
2617 signal_child(AutoVacPID, SIGHUP);
2618 if (PgArchPID != 0)
2619 signal_child(PgArchPID, SIGHUP);
2620 if (SysLoggerPID != 0)
2621 signal_child(SysLoggerPID, SIGHUP);
2622 if (PgStatPID != 0)
2623 signal_child(PgStatPID, SIGHUP);
2624
2625 /* Reload authentication config files too */
2626 if (!load_hba())
2627 ereport(WARNING,
2628 (errmsg("pg_hba.conf not reloaded")));
2629
2630 if (!load_ident())
2631 ereport(WARNING,
2632 (errmsg("pg_ident.conf not reloaded")));
2633
2634 #ifdef EXEC_BACKEND
2635 /* Update the starting-point file for future children */
2636 write_nondefault_variables(PGC_SIGHUP);
2637 #endif
2638 }
2639
2640 #ifdef WIN32
2641 PG_SETMASK(&UnBlockSig);
2642 #endif
2643
2644 errno = save_errno;
2645 }
2646
2647
2648 /*
2649 * pmdie -- signal handler for processing various postmaster signals.
2650 */
2651 static void
2652 pmdie(SIGNAL_ARGS)
2653 {
2654 int save_errno = errno;
2655
2656 /*
2657 * We rely on the signal mechanism to have blocked all signals ... except
2658 * on Windows, which lacks sigaction(), so we have to do it manually.
2659 */
2660 #ifdef WIN32
2661 PG_SETMASK(&BlockSig);
2662 #endif
2663
2664 ereport(DEBUG2,
2665 (errmsg_internal("postmaster received signal %d",
2666 postgres_signal_arg)));
2667
2668 switch (postgres_signal_arg)
2669 {
2670 case SIGTERM:
2671
2672 /*
2673 * Smart Shutdown:
2674 *
2675 * Wait for children to end their work, then shut down.
2676 */
2677 if (Shutdown >= SmartShutdown)
2678 break;
2679 Shutdown = SmartShutdown;
2680 ereport(LOG,
2681 (errmsg("received smart shutdown request")));
2682 #ifdef USE_SYSTEMD
2683 sd_notify(0, "STOPPING=1");
2684 #endif
2685
2686 /*
2687 * If we reached normal running, we have to wait for any online
2688 * backup mode to end; otherwise go straight to waiting for client
2689 * backends to exit. (The difference is that in the former state,
2690 * we'll still let in new superuser clients, so that somebody can
2691 * end the online backup mode.) If already in PM_STOP_BACKENDS or
2692 * a later state, do not change it.
2693 */
2694 if (pmState == PM_RUN)
2695 connsAllowed = ALLOW_SUPERUSER_CONNS;
2696 else if (pmState == PM_HOT_STANDBY)
2697 connsAllowed = ALLOW_NO_CONNS;
2698 else if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2699 {
2700 /* There should be no clients, so proceed to stop children */
2701 pmState = PM_STOP_BACKENDS;
2702 }
2703
2704 /*
2705 * Now wait for online backup mode to end and backends to exit. If
2706 * that is already the case, PostmasterStateMachine will take the
2707 * next step.
2708 */
2709 PostmasterStateMachine();
2710 break;
2711
2712 case SIGINT:
2713
2714 /*
2715 * Fast Shutdown:
2716 *
2717 * Abort all children with SIGTERM (rollback active transactions
2718 * and exit) and shut down when they are gone.
2719 */
2720 if (Shutdown >= FastShutdown)
2721 break;
2722 Shutdown = FastShutdown;
2723 ereport(LOG,
2724 (errmsg("received fast shutdown request")));
2725 #ifdef USE_SYSTEMD
2726 sd_notify(0, "STOPPING=1");
2727 #endif
2728
2729 if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2730 {
2731 /* Just shut down background processes silently */
2732 pmState = PM_STOP_BACKENDS;
2733 }
2734 else if (pmState == PM_RUN ||
2735 pmState == PM_HOT_STANDBY)
2736 {
2737 /* Report that we're about to zap live client sessions */
2738 ereport(LOG,
2739 (errmsg("aborting any active transactions")));
2740 pmState = PM_STOP_BACKENDS;
2741 }
2742
2743 /*
2744 * PostmasterStateMachine will issue any necessary signals, or
2745 * take the next step if no child processes need to be killed.
2746 */
2747 PostmasterStateMachine();
2748 break;
2749
2750 case SIGQUIT:
2751
2752 /*
2753 * Immediate Shutdown:
2754 *
2755 * abort all children with SIGQUIT, wait for them to exit,
2756 * terminate remaining ones with SIGKILL, then exit without
2757 * attempt to properly shut down the data base system.
2758 */
2759 if (Shutdown >= ImmediateShutdown)
2760 break;
2761 Shutdown = ImmediateShutdown;
2762 ereport(LOG,
2763 (errmsg("received immediate shutdown request")));
2764 #ifdef USE_SYSTEMD
2765 sd_notify(0, "STOPPING=1");
2766 #endif
2767
2768 TerminateChildren(SIGQUIT);
2769 pmState = PM_WAIT_BACKENDS;
2770
2771 /* set stopwatch for them to die */
2772 AbortStartTime = time(NULL);
2773
2774 /*
2775 * Now wait for backends to exit. If there are none,
2776 * PostmasterStateMachine will take the next step.
2777 */
2778 PostmasterStateMachine();
2779 break;
2780 }
2781
2782 #ifdef WIN32
2783 PG_SETMASK(&UnBlockSig);
2784 #endif
2785
2786 errno = save_errno;
2787 }
2788
2789 /*
2790 * Reaper -- signal handler to cleanup after a child process dies.
2791 */
2792 static void
2793 reaper(SIGNAL_ARGS)
2794 {
2795 int save_errno = errno;
2796 int pid; /* process id of dead child process */
2797 int exitstatus; /* its exit status */
2798
2799 /*
2800 * We rely on the signal mechanism to have blocked all signals ... except
2801 * on Windows, which lacks sigaction(), so we have to do it manually.
2802 */
2803 #ifdef WIN32
2804 PG_SETMASK(&BlockSig);
2805 #endif
2806
2807 ereport(DEBUG4,
2808 (errmsg_internal("reaping dead processes")));
2809
2810 while ((pid = waitpid(-1, &exitstatus, WNOHANG)) > 0)
2811 {
2812 /*
2813 * Check if this child was a startup process.
2814 */
2815 if (pid == StartupPID)
2816 {
2817 StartupPID = 0;
2818
2819 /*
2820 * Startup process exited in response to a shutdown request (or it
2821 * completed normally regardless of the shutdown request).
2822 */
2823 if (Shutdown > NoShutdown &&
2824 (EXIT_STATUS_0(exitstatus) || EXIT_STATUS_1(exitstatus)))
2825 {
2826 StartupStatus = STARTUP_NOT_RUNNING;
2827 pmState = PM_WAIT_BACKENDS;
2828 /* PostmasterStateMachine logic does the rest */
2829 continue;
2830 }
2831
2832 if (EXIT_STATUS_3(exitstatus))
2833 {
2834 ereport(LOG,
2835 (errmsg("shutdown at recovery target")));
2836 StartupStatus = STARTUP_NOT_RUNNING;
2837 Shutdown = Max(Shutdown, SmartShutdown);
2838 TerminateChildren(SIGTERM);
2839 pmState = PM_WAIT_BACKENDS;
2840 /* PostmasterStateMachine logic does the rest */
2841 continue;
2842 }
2843
2844 /*
2845 * Unexpected exit of startup process (including FATAL exit)
2846 * during PM_STARTUP is treated as catastrophic. There are no
2847 * other processes running yet, so we can just exit.
2848 */
2849 if (pmState == PM_STARTUP && !EXIT_STATUS_0(exitstatus))
2850 {
2851 LogChildExit(LOG, _("startup process"),
2852 pid, exitstatus);
2853 ereport(LOG,
2854 (errmsg("aborting startup due to startup process failure")));
2855 ExitPostmaster(1);
2856 }
2857
2858 /*
2859 * After PM_STARTUP, any unexpected exit (including FATAL exit) of
2860 * the startup process is catastrophic, so kill other children,
2861 * and set StartupStatus so we don't try to reinitialize after
2862 * they're gone. Exception: if StartupStatus is STARTUP_SIGNALED,
2863 * then we previously sent the startup process a SIGQUIT; so
2864 * that's probably the reason it died, and we do want to try to
2865 * restart in that case.
2866 */
2867 if (!EXIT_STATUS_0(exitstatus))
2868 {
2869 if (StartupStatus == STARTUP_SIGNALED)
2870 StartupStatus = STARTUP_NOT_RUNNING;
2871 else
2872 StartupStatus = STARTUP_CRASHED;
2873 HandleChildCrash(pid, exitstatus,
2874 _("startup process"));
2875 continue;
2876 }
2877
2878 /*
2879 * Startup succeeded, commence normal operations
2880 */
2881 StartupStatus = STARTUP_NOT_RUNNING;
2882 FatalError = false;
2883 Assert(AbortStartTime == 0);
2884 ReachedNormalRunning = true;
2885 pmState = PM_RUN;
2886 connsAllowed = ALLOW_ALL_CONNS;
2887
2888 /*
2889 * Crank up the background tasks, if we didn't do that already
2890 * when we entered consistent recovery state. It doesn't matter
2891 * if this fails, we'll just try again later.
2892 */
2893 if (CheckpointerPID == 0)
2894 CheckpointerPID = StartCheckpointer();
2895 if (BgWriterPID == 0)
2896 BgWriterPID = StartBackgroundWriter();
2897 if (WalWriterPID == 0)
2898 WalWriterPID = StartWalWriter();
2899
2900 /*
2901 * Likewise, start other special children as needed. In a restart
2902 * situation, some of them may be alive already.
2903 */
2904 if (!IsBinaryUpgrade && AutoVacuumingActive() && AutoVacPID == 0)
2905 AutoVacPID = StartAutoVacLauncher();
2906 if (PgArchStartupAllowed() && PgArchPID == 0)
2907 PgArchPID = pgarch_start();
2908 if (PgStatPID == 0)
2909 PgStatPID = pgstat_start();
2910
2911 /* workers may be scheduled to start now */
2912 maybe_start_bgworkers();
2913
2914 /* at this point we are really open for business */
2915 ereport(LOG,
2916 (errmsg("database system is ready to accept connections")));
2917
2918 #ifdef USE_SYSTEMD
2919 sd_notify(0, "READY=1");
2920 #endif
2921
2922 continue;
2923 }
2924
2925 /*
2926 * Was it the bgwriter? Normal exit can be ignored; we'll start a new
2927 * one at the next iteration of the postmaster's main loop, if
2928 * necessary. Any other exit condition is treated as a crash.
2929 */
2930 if (pid == BgWriterPID)
2931 {
2932 BgWriterPID = 0;
2933 if (!EXIT_STATUS_0(exitstatus))
2934 HandleChildCrash(pid, exitstatus,
2935 _("background writer process"));
2936 continue;
2937 }
2938
2939 /*
2940 * Was it the checkpointer?
2941 */
2942 if (pid == CheckpointerPID)
2943 {
2944 CheckpointerPID = 0;
2945 if (EXIT_STATUS_0(exitstatus) && pmState == PM_SHUTDOWN)
2946 {
2947 /*
2948 * OK, we saw normal exit of the checkpointer after it's been
2949 * told to shut down. We expect that it wrote a shutdown
2950 * checkpoint. (If for some reason it didn't, recovery will
2951 * occur on next postmaster start.)
2952 *
2953 * At this point we should have no normal backend children
2954 * left (else we'd not be in PM_SHUTDOWN state) but we might
2955 * have dead_end children to wait for.
2956 *
2957 * If we have an archiver subprocess, tell it to do a last
2958 * archive cycle and quit. Likewise, if we have walsender
2959 * processes, tell them to send any remaining WAL and quit.
2960 */
2961 Assert(Shutdown > NoShutdown);
2962
2963 /* Waken archiver for the last time */
2964 if (PgArchPID != 0)
2965 signal_child(PgArchPID, SIGUSR2);
2966
2967 /*
2968 * Waken walsenders for the last time. No regular backends
2969 * should be around anymore.
2970 */
2971 SignalChildren(SIGUSR2);
2972
2973 pmState = PM_SHUTDOWN_2;
2974
2975 /*
2976 * We can also shut down the stats collector now; there's
2977 * nothing left for it to do.
2978 */
2979 if (PgStatPID != 0)
2980 signal_child(PgStatPID, SIGQUIT);
2981 }
2982 else
2983 {
2984 /*
2985 * Any unexpected exit of the checkpointer (including FATAL
2986 * exit) is treated as a crash.
2987 */
2988 HandleChildCrash(pid, exitstatus,
2989 _("checkpointer process"));
2990 }
2991
2992 continue;
2993 }
2994
2995 /*
2996 * Was it the wal writer? Normal exit can be ignored; we'll start a
2997 * new one at the next iteration of the postmaster's main loop, if
2998 * necessary. Any other exit condition is treated as a crash.
2999 */
3000 if (pid == WalWriterPID)
3001 {
3002 WalWriterPID = 0;
3003 if (!EXIT_STATUS_0(exitstatus))
3004 HandleChildCrash(pid, exitstatus,
3005 _("WAL writer process"));
3006 continue;
3007 }
3008
3009 /*
3010 * Was it the wal receiver? If exit status is zero (normal) or one
3011 * (FATAL exit), we assume everything is all right just like normal
3012 * backends. (If we need a new wal receiver, we'll start one at the
3013 * next iteration of the postmaster's main loop.)
3014 */
3015 if (pid == WalReceiverPID)
3016 {
3017 WalReceiverPID = 0;
3018 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3019 HandleChildCrash(pid, exitstatus,
3020 _("WAL receiver process"));
3021 continue;
3022 }
3023
3024 /*
3025 * Was it the autovacuum launcher? Normal exit can be ignored; we'll
3026 * start a new one at the next iteration of the postmaster's main
3027 * loop, if necessary. Any other exit condition is treated as a
3028 * crash.
3029 */
3030 if (pid == AutoVacPID)
3031 {
3032 AutoVacPID = 0;
3033 if (!EXIT_STATUS_0(exitstatus))
3034 HandleChildCrash(pid, exitstatus,
3035 _("autovacuum launcher process"));
3036 continue;
3037 }
3038
3039 /*
3040 * Was it the archiver? If so, just try to start a new one; no need
3041 * to force reset of the rest of the system. (If fail, we'll try
3042 * again in future cycles of the main loop.). Unless we were waiting
3043 * for it to shut down; don't restart it in that case, and
3044 * PostmasterStateMachine() will advance to the next shutdown step.
3045 */
3046 if (pid == PgArchPID)
3047 {
3048 PgArchPID = 0;
3049 if (!EXIT_STATUS_0(exitstatus))
3050 LogChildExit(LOG, _("archiver process"),
3051 pid, exitstatus);
3052 if (PgArchStartupAllowed())
3053 PgArchPID = pgarch_start();
3054 continue;
3055 }
3056
3057 /*
3058 * Was it the statistics collector? If so, just try to start a new
3059 * one; no need to force reset of the rest of the system. (If fail,
3060 * we'll try again in future cycles of the main loop.)
3061 */
3062 if (pid == PgStatPID)
3063 {
3064 PgStatPID = 0;
3065 if (!EXIT_STATUS_0(exitstatus))
3066 LogChildExit(LOG, _("statistics collector process"),
3067 pid, exitstatus);
3068 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3069 PgStatPID = pgstat_start();
3070 continue;
3071 }
3072
3073 /* Was it the system logger? If so, try to start a new one */
3074 if (pid == SysLoggerPID)
3075 {
3076 SysLoggerPID = 0;
3077 /* for safety's sake, launch new logger *first* */
3078 SysLoggerPID = SysLogger_Start();
3079 if (!EXIT_STATUS_0(exitstatus))
3080 LogChildExit(LOG, _("system logger process"),
3081 pid, exitstatus);
3082 continue;
3083 }
3084
3085 /* Was it one of our background workers? */
3086 if (CleanupBackgroundWorker(pid, exitstatus))
3087 {
3088 /* have it be restarted */
3089 HaveCrashedWorker = true;
3090 continue;
3091 }
3092
3093 /*
3094 * Else do standard backend child cleanup.
3095 */
3096 CleanupBackend(pid, exitstatus);
3097 } /* loop over pending child-death reports */
3098
3099 /*
3100 * After cleaning out the SIGCHLD queue, see if we have any state changes
3101 * or actions to make.
3102 */
3103 PostmasterStateMachine();
3104
3105 /* Done with signal handler */
3106 #ifdef WIN32
3107 PG_SETMASK(&UnBlockSig);
3108 #endif
3109
3110 errno = save_errno;
3111 }
3112
3113 /*
3114 * Scan the bgworkers list and see if the given PID (which has just stopped
3115 * or crashed) is in it. Handle its shutdown if so, and return true. If not a
3116 * bgworker, return false.
3117 *
3118 * This is heavily based on CleanupBackend. One important difference is that
3119 * we don't know yet that the dying process is a bgworker, so we must be silent
3120 * until we're sure it is.
3121 */
3122 static bool
3123 CleanupBackgroundWorker(int pid,
3124 int exitstatus) /* child's exit status */
3125 {
3126 char namebuf[MAXPGPATH];
3127 slist_iter iter;
3128
3129 slist_foreach(iter, &BackgroundWorkerList)
3130 {
3131 RegisteredBgWorker *rw;
3132
3133 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
3134
3135 if (rw->rw_pid != pid)
3136 continue;
3137
3138 #ifdef WIN32
3139 /* see CleanupBackend */
3140 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3141 exitstatus = 0;
3142 #endif
3143
3144 snprintf(namebuf, MAXPGPATH, "%s: %s", _("worker process"),
3145 rw->rw_worker.bgw_name);
3146
3147 if (!EXIT_STATUS_0(exitstatus))
3148 {
3149 /* Record timestamp, so we know when to restart the worker. */
3150 rw->rw_crashed_at = GetCurrentTimestamp();
3151 }
3152 else
3153 {
3154 /* Zero exit status means terminate */
3155 rw->rw_crashed_at = 0;
3156 rw->rw_terminate = true;
3157 }
3158
3159 /*
3160 * Additionally, for shared-memory-connected workers, just like a
3161 * backend, any exit status other than 0 or 1 is considered a crash
3162 * and causes a system-wide restart.
3163 */
3164 if ((rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3165 {
3166 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3167 {
3168 HandleChildCrash(pid, exitstatus, namebuf);
3169 return true;
3170 }
3171 }
3172
3173 /*
3174 * We must release the postmaster child slot whether this worker is
3175 * connected to shared memory or not, but we only treat it as a crash
3176 * if it is in fact connected.
3177 */
3178 if (!ReleasePostmasterChildSlot(rw->rw_child_slot) &&
3179 (rw->rw_worker.bgw_flags & BGWORKER_SHMEM_ACCESS) != 0)
3180 {
3181 HandleChildCrash(pid, exitstatus, namebuf);
3182 return true;
3183 }
3184
3185 /* Get it out of the BackendList and clear out remaining data */
3186 dlist_delete(&rw->rw_backend->elem);
3187 #ifdef EXEC_BACKEND
3188 ShmemBackendArrayRemove(rw->rw_backend);
3189 #endif
3190
3191 /*
3192 * It's possible that this background worker started some OTHER
3193 * background worker and asked to be notified when that worker started
3194 * or stopped. If so, cancel any notifications destined for the
3195 * now-dead backend.
3196 */
3197 if (rw->rw_backend->bgworker_notify)
3198 BackgroundWorkerStopNotifications(rw->rw_pid);
3199 free(rw->rw_backend);
3200 rw->rw_backend = NULL;
3201 rw->rw_pid = 0;
3202 rw->rw_child_slot = 0;
3203 ReportBackgroundWorkerPID(rw); /* report child death */
3204
3205 LogChildExit(EXIT_STATUS_0(exitstatus) ? DEBUG1 : LOG,
3206 namebuf, pid, exitstatus);
3207
3208 return true;
3209 }
3210
3211 return false;
3212 }
3213
3214 /*
3215 * CleanupBackend -- cleanup after terminated backend.
3216 *
3217 * Remove all local state associated with backend.
3218 *
3219 * If you change this, see also CleanupBackgroundWorker.
3220 */
3221 static void
3222 CleanupBackend(int pid,
3223 int exitstatus) /* child's exit status. */
3224 {
3225 dlist_mutable_iter iter;
3226
3227 LogChildExit(DEBUG2, _("server process"), pid, exitstatus);
3228
3229 /*
3230 * If a backend dies in an ugly way then we must signal all other backends
3231 * to quickdie. If exit status is zero (normal) or one (FATAL exit), we
3232 * assume everything is all right and proceed to remove the backend from
3233 * the active backend list.
3234 */
3235
3236 #ifdef WIN32
3237
3238 /*
3239 * On win32, also treat ERROR_WAIT_NO_CHILDREN (128) as nonfatal case,
3240 * since that sometimes happens under load when the process fails to start
3241 * properly (long before it starts using shared memory). Microsoft reports
3242 * it is related to mutex failure:
3243 * http://archives.postgresql.org/pgsql-hackers/2010-09/msg00790.php
3244 */
3245 if (exitstatus == ERROR_WAIT_NO_CHILDREN)
3246 {
3247 LogChildExit(LOG, _("server process"), pid, exitstatus);
3248 exitstatus = 0;
3249 }
3250 #endif
3251
3252 if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
3253 {
3254 HandleChildCrash(pid, exitstatus, _("server process"));
3255 return;
3256 }
3257
3258 dlist_foreach_modify(iter, &BackendList)
3259 {
3260 Backend *bp = dlist_container(Backend, elem, iter.cur);
3261
3262 if (bp->pid == pid)
3263 {
3264 if (!bp->dead_end)
3265 {
3266 if (!ReleasePostmasterChildSlot(bp->child_slot))
3267 {
3268 /*
3269 * Uh-oh, the child failed to clean itself up. Treat as a
3270 * crash after all.
3271 */
3272 HandleChildCrash(pid, exitstatus, _("server process"));
3273 return;
3274 }
3275 #ifdef EXEC_BACKEND
3276 ShmemBackendArrayRemove(bp);
3277 #endif
3278 }
3279 if (bp->bgworker_notify)
3280 {
3281 /*
3282 * This backend may have been slated to receive SIGUSR1 when
3283 * some background worker started or stopped. Cancel those
3284 * notifications, as we don't want to signal PIDs that are not
3285 * PostgreSQL backends. This gets skipped in the (probably
3286 * very common) case where the backend has never requested any
3287 * such notifications.
3288 */
3289 BackgroundWorkerStopNotifications(bp->pid);
3290 }
3291 dlist_delete(iter.cur);
3292 free(bp);
3293 break;
3294 }
3295 }
3296 }
3297
3298 /*
3299 * HandleChildCrash -- cleanup after failed backend, bgwriter, checkpointer,
3300 * walwriter, autovacuum, or background worker.
3301 *
3302 * The objectives here are to clean up our local state about the child
3303 * process, and to signal all other remaining children to quickdie.
3304 */
3305 static void
3306 HandleChildCrash(int pid, int exitstatus, const char *procname)
3307 {
3308 dlist_mutable_iter iter;
3309 slist_iter siter;
3310 Backend *bp;
3311 bool take_action;
3312
3313 /*
3314 * We only log messages and send signals if this is the first process
3315 * crash and we're not doing an immediate shutdown; otherwise, we're only
3316 * here to update postmaster's idea of live processes. If we have already
3317 * signalled children, nonzero exit status is to be expected, so don't
3318 * clutter log.
3319 */
3320 take_action = !FatalError && Shutdown != ImmediateShutdown;
3321
3322 if (take_action)
3323 {
3324 LogChildExit(LOG, procname, pid, exitstatus);
3325 ereport(LOG,
3326 (errmsg("terminating any other active server processes")));
3327 }
3328
3329 /* Process background workers. */
3330 slist_foreach(siter, &BackgroundWorkerList)
3331 {
3332 RegisteredBgWorker *rw;
3333
3334 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
3335 if (rw->rw_pid == 0)
3336 continue; /* not running */
3337 if (rw->rw_pid == pid)
3338 {
3339 /*
3340 * Found entry for freshly-dead worker, so remove it.
3341 */
3342 (void) ReleasePostmasterChildSlot(rw->rw_child_slot);
3343 dlist_delete(&rw->rw_backend->elem);
3344 #ifdef EXEC_BACKEND
3345 ShmemBackendArrayRemove(rw->rw_backend);
3346 #endif
3347 free(rw->rw_backend);
3348 rw->rw_backend = NULL;
3349 rw->rw_pid = 0;
3350 rw->rw_child_slot = 0;
3351 /* don't reset crashed_at */
3352 /* don't report child stop, either */
3353 /* Keep looping so we can signal remaining workers */
3354 }
3355 else
3356 {
3357 /*
3358 * This worker is still alive. Unless we did so already, tell it
3359 * to commit hara-kiri.
3360 *
3361 * SIGQUIT is the special signal that says exit without proc_exit
3362 * and let the user know what's going on. But if SendStop is set
3363 * (-s on command line), then we send SIGSTOP instead, so that we
3364 * can get core dumps from all backends by hand.
3365 */
3366 if (take_action)
3367 {
3368 ereport(DEBUG2,
3369 (errmsg_internal("sending %s to process %d",
3370 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3371 (int) rw->rw_pid)));
3372 signal_child(rw->rw_pid, (SendStop ? SIGSTOP : SIGQUIT));
3373 }
3374 }
3375 }
3376
3377 /* Process regular backends */
3378 dlist_foreach_modify(iter, &BackendList)
3379 {
3380 bp = dlist_container(Backend, elem, iter.cur);
3381
3382 if (bp->pid == pid)
3383 {
3384 /*
3385 * Found entry for freshly-dead backend, so remove it.
3386 */
3387 if (!bp->dead_end)
3388 {
3389 (void) ReleasePostmasterChildSlot(bp->child_slot);
3390 #ifdef EXEC_BACKEND
3391 ShmemBackendArrayRemove(bp);
3392 #endif
3393 }
3394 dlist_delete(iter.cur);
3395 free(bp);
3396 /* Keep looping so we can signal remaining backends */
3397 }
3398 else
3399 {
3400 /*
3401 * This backend is still alive. Unless we did so already, tell it
3402 * to commit hara-kiri.
3403 *
3404 * SIGQUIT is the special signal that says exit without proc_exit
3405 * and let the user know what's going on. But if SendStop is set
3406 * (-s on command line), then we send SIGSTOP instead, so that we
3407 * can get core dumps from all backends by hand.
3408 *
3409 * We could exclude dead_end children here, but at least in the
3410 * SIGSTOP case it seems better to include them.
3411 *
3412 * Background workers were already processed above; ignore them
3413 * here.
3414 */
3415 if (bp->bkend_type == BACKEND_TYPE_BGWORKER)
3416 continue;
3417
3418 if (take_action)
3419 {
3420 ereport(DEBUG2,
3421 (errmsg_internal("sending %s to process %d",
3422 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3423 (int) bp->pid)));
3424 signal_child(bp->pid, (SendStop ? SIGSTOP : SIGQUIT));
3425 }
3426 }
3427 }
3428
3429 /* Take care of the startup process too */
3430 if (pid == StartupPID)
3431 {
3432 StartupPID = 0;
3433 StartupStatus = STARTUP_CRASHED;
3434 }
3435 else if (StartupPID != 0 && take_action)
3436 {
3437 ereport(DEBUG2,
3438 (errmsg_internal("sending %s to process %d",
3439 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3440 (int) StartupPID)));
3441 signal_child(StartupPID, (SendStop ? SIGSTOP : SIGQUIT));
3442 StartupStatus = STARTUP_SIGNALED;
3443 }
3444
3445 /* Take care of the bgwriter too */
3446 if (pid == BgWriterPID)
3447 BgWriterPID = 0;
3448 else if (BgWriterPID != 0 && take_action)
3449 {
3450 ereport(DEBUG2,
3451 (errmsg_internal("sending %s to process %d",
3452 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3453 (int) BgWriterPID)));
3454 signal_child(BgWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3455 }
3456
3457 /* Take care of the checkpointer too */
3458 if (pid == CheckpointerPID)
3459 CheckpointerPID = 0;
3460 else if (CheckpointerPID != 0 && take_action)
3461 {
3462 ereport(DEBUG2,
3463 (errmsg_internal("sending %s to process %d",
3464 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3465 (int) CheckpointerPID)));
3466 signal_child(CheckpointerPID, (SendStop ? SIGSTOP : SIGQUIT));
3467 }
3468
3469 /* Take care of the walwriter too */
3470 if (pid == WalWriterPID)
3471 WalWriterPID = 0;
3472 else if (WalWriterPID != 0 && take_action)
3473 {
3474 ereport(DEBUG2,
3475 (errmsg_internal("sending %s to process %d",
3476 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3477 (int) WalWriterPID)));
3478 signal_child(WalWriterPID, (SendStop ? SIGSTOP : SIGQUIT));
3479 }
3480
3481 /* Take care of the walreceiver too */
3482 if (pid == WalReceiverPID)
3483 WalReceiverPID = 0;
3484 else if (WalReceiverPID != 0 && take_action)
3485 {
3486 ereport(DEBUG2,
3487 (errmsg_internal("sending %s to process %d",
3488 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3489 (int) WalReceiverPID)));
3490 signal_child(WalReceiverPID, (SendStop ? SIGSTOP : SIGQUIT));
3491 }
3492
3493 /* Take care of the autovacuum launcher too */
3494 if (pid == AutoVacPID)
3495 AutoVacPID = 0;
3496 else if (AutoVacPID != 0 && take_action)
3497 {
3498 ereport(DEBUG2,
3499 (errmsg_internal("sending %s to process %d",
3500 (SendStop ? "SIGSTOP" : "SIGQUIT"),
3501 (int) AutoVacPID)));
3502 signal_child(AutoVacPID, (SendStop ? SIGSTOP : SIGQUIT));
3503 }
3504
3505 /*
3506 * Force a power-cycle of the pgarch process too. (This isn't absolutely
3507 * necessary, but it seems like a good idea for robustness, and it
3508 * simplifies the state-machine logic in the case where a shutdown request
3509 * arrives during crash processing.)
3510 */
3511 if (PgArchPID != 0 && take_action)
3512 {
3513 ereport(DEBUG2,
3514 (errmsg_internal("sending %s to process %d",
3515 "SIGQUIT",
3516 (int) PgArchPID)));
3517 signal_child(PgArchPID, SIGQUIT);
3518 }
3519
3520 /*
3521 * Force a power-cycle of the pgstat process too. (This isn't absolutely
3522 * necessary, but it seems like a good idea for robustness, and it
3523 * simplifies the state-machine logic in the case where a shutdown request
3524 * arrives during crash processing.)
3525 */
3526 if (PgStatPID != 0 && take_action)
3527 {
3528 ereport(DEBUG2,
3529 (errmsg_internal("sending %s to process %d",
3530 "SIGQUIT",
3531 (int) PgStatPID)));
3532 signal_child(PgStatPID, SIGQUIT);
3533 allow_immediate_pgstat_restart();
3534 }
3535
3536 /* We do NOT restart the syslogger */
3537
3538 if (Shutdown != ImmediateShutdown)
3539 FatalError = true;
3540
3541 /* We now transit into a state of waiting for children to die */
3542 if (pmState == PM_RECOVERY ||
3543 pmState == PM_HOT_STANDBY ||
3544 pmState == PM_RUN ||
3545 pmState == PM_STOP_BACKENDS ||
3546 pmState == PM_SHUTDOWN)
3547 pmState = PM_WAIT_BACKENDS;
3548
3549 /*
3550 * .. and if this doesn't happen quickly enough, now the clock is ticking
3551 * for us to kill them without mercy.
3552 */
3553 if (AbortStartTime == 0)
3554 AbortStartTime = time(NULL);
3555 }
3556
3557 /*
3558 * Log the death of a child process.
3559 */
3560 static void
3561 LogChildExit(int lev, const char *procname, int pid, int exitstatus)
3562 {
3563 /*
3564 * size of activity_buffer is arbitrary, but set equal to default
3565 * track_activity_query_size
3566 */
3567 char activity_buffer[1024];
3568 const char *activity = NULL;
3569
3570 if (!EXIT_STATUS_0(exitstatus))
3571 activity = pgstat_get_crashed_backend_activity(pid,
3572 activity_buffer,
3573 sizeof(activity_buffer));
3574
3575 if (WIFEXITED(exitstatus))
3576 ereport(lev,
3577
3578 /*------
3579 translator: %s is a noun phrase describing a child process, such as
3580 "server process" */
3581 (errmsg("%s (PID %d) exited with exit code %d",
3582 procname, pid, WEXITSTATUS(exitstatus)),
3583 activity ? errdetail("Failed process was running: %s", activity) : 0));
3584 else if (WIFSIGNALED(exitstatus))
3585 {
3586 #if defined(WIN32)
3587 ereport(lev,
3588
3589 /*------
3590 translator: %s is a noun phrase describing a child process, such as
3591 "server process" */
3592 (errmsg("%s (PID %d) was terminated by exception 0x%X",
3593 procname, pid, WTERMSIG(exitstatus)),
3594 errhint("See C include file \"ntstatus.h\" for a description of the hexadecimal value."),
3595 activity ? errdetail("Failed process was running: %s", activity) : 0));
3596 #else
3597 ereport(lev,
3598
3599 /*------
3600 translator: %s is a noun phrase describing a child process, such as
3601 "server process" */
3602 (errmsg("%s (PID %d) was terminated by signal %d: %s",
3603 procname, pid, WTERMSIG(exitstatus),
3604 pg_strsignal(WTERMSIG(exitstatus))),
3605 activity ? errdetail("Failed process was running: %s", activity) : 0));
3606 #endif
3607 }
3608 else
3609 ereport(lev,
3610
3611 /*------
3612 translator: %s is a noun phrase describing a child process, such as
3613 "server process" */
3614 (errmsg("%s (PID %d) exited with unrecognized status %d",
3615 procname, pid, exitstatus),
3616 activity ? errdetail("Failed process was running: %s", activity) : 0));
3617 }
3618
3619 /*
3620 * Advance the postmaster's state machine and take actions as appropriate
3621 *
3622 * This is common code for pmdie(), reaper() and sigusr1_handler(), which
3623 * receive the signals that might mean we need to change state.
3624 */
3625 static void
3626 PostmasterStateMachine(void)
3627 {
3628 /* If we're doing a smart shutdown, try to advance that state. */
3629 if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
3630 {
3631 if (connsAllowed == ALLOW_SUPERUSER_CONNS)
3632 {
3633 /*
3634 * ALLOW_SUPERUSER_CONNS state ends as soon as online backup mode
3635 * is not active.
3636 */
3637 if (!BackupInProgress())
3638 connsAllowed = ALLOW_NO_CONNS;
3639 }
3640
3641 if (connsAllowed == ALLOW_NO_CONNS)
3642 {
3643 /*
3644 * ALLOW_NO_CONNS state ends when we have no normal client
3645 * backends running. Then we're ready to stop other children.
3646 */
3647 if (CountChildren(BACKEND_TYPE_NORMAL) == 0)
3648 pmState = PM_STOP_BACKENDS;
3649 }
3650 }
3651
3652 /*
3653 * If we're ready to do so, signal child processes to shut down. (This
3654 * isn't a persistent state, but treating it as a distinct pmState allows
3655 * us to share this code across multiple shutdown code paths.)
3656 */
3657 if (pmState == PM_STOP_BACKENDS)
3658 {
3659 /* Signal all backend children except walsenders */
3660 SignalSomeChildren(SIGTERM,
3661 BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND);
3662 /* and the autovac launcher too */
3663 if (AutoVacPID != 0)
3664 signal_child(AutoVacPID, SIGTERM);
3665 /* and the bgwriter too */
3666 if (BgWriterPID != 0)
3667 signal_child(BgWriterPID, SIGTERM);
3668 /* and the walwriter too */
3669 if (WalWriterPID != 0)
3670 signal_child(WalWriterPID, SIGTERM);
3671 /* If we're in recovery, also stop startup and walreceiver procs */
3672 if (StartupPID != 0)
3673 signal_child(StartupPID, SIGTERM);
3674 if (WalReceiverPID != 0)
3675 signal_child(WalReceiverPID, SIGTERM);
3676 /* checkpointer, archiver, stats, and syslogger may continue for now */
3677
3678 /* Now transition to PM_WAIT_BACKENDS state to wait for them to die */
3679 pmState = PM_WAIT_BACKENDS;
3680 }
3681
3682 /*
3683 * If we are in a state-machine state that implies waiting for backends to
3684 * exit, see if they're all gone, and change state if so.
3685 */
3686 if (pmState == PM_WAIT_BACKENDS)
3687 {
3688 /*
3689 * PM_WAIT_BACKENDS state ends when we have no regular backends
3690 * (including autovac workers), no bgworkers (including unconnected
3691 * ones), and no walwriter, autovac launcher or bgwriter. If we are
3692 * doing crash recovery or an immediate shutdown then we expect the
3693 * checkpointer to exit as well, otherwise not. The archiver, stats,
3694 * and syslogger processes are disregarded since they are not
3695 * connected to shared memory; we also disregard dead_end children
3696 * here. Walsenders are also disregarded, they will be terminated
3697 * later after writing the checkpoint record, like the archiver
3698 * process.
3699 */
3700 if (CountChildren(BACKEND_TYPE_ALL - BACKEND_TYPE_WALSND) == 0 &&
3701 StartupPID == 0 &&
3702 WalReceiverPID == 0 &&
3703 BgWriterPID == 0 &&
3704 (CheckpointerPID == 0 ||
3705 (!FatalError && Shutdown < ImmediateShutdown)) &&
3706 WalWriterPID == 0 &&
3707 AutoVacPID == 0)
3708 {
3709 if (Shutdown >= ImmediateShutdown || FatalError)
3710 {
3711 /*
3712 * Start waiting for dead_end children to die. This state
3713 * change causes ServerLoop to stop creating new ones.
3714 */
3715 pmState = PM_WAIT_DEAD_END;
3716
3717 /*
3718 * We already SIGQUIT'd the archiver and stats processes, if
3719 * any, when we started immediate shutdown or entered
3720 * FatalError state.
3721 */
3722 }
3723 else
3724 {
3725 /*
3726 * If we get here, we are proceeding with normal shutdown. All
3727 * the regular children are gone, and it's time to tell the
3728 * checkpointer to do a shutdown checkpoint.
3729 */
3730 Assert(Shutdown > NoShutdown);
3731 /* Start the checkpointer if not running */
3732 if (CheckpointerPID == 0)
3733 CheckpointerPID = StartCheckpointer();
3734 /* And tell it to shut down */
3735 if (CheckpointerPID != 0)
3736 {
3737 signal_child(CheckpointerPID, SIGUSR2);
3738 pmState = PM_SHUTDOWN;
3739 }
3740 else
3741 {
3742 /*
3743 * If we failed to fork a checkpointer, just shut down.
3744 * Any required cleanup will happen at next restart. We
3745 * set FatalError so that an "abnormal shutdown" message
3746 * gets logged when we exit.
3747 */
3748 FatalError = true;
3749 pmState = PM_WAIT_DEAD_END;
3750
3751 /* Kill the walsenders, archiver and stats collector too */
3752 SignalChildren(SIGQUIT);
3753 if (PgArchPID != 0)
3754 signal_child(PgArchPID, SIGQUIT);
3755 if (PgStatPID != 0)
3756 signal_child(PgStatPID, SIGQUIT);
3757 }
3758 }
3759 }
3760 }
3761
3762 if (pmState == PM_SHUTDOWN_2)
3763 {
3764 /*
3765 * PM_SHUTDOWN_2 state ends when there's no other children than
3766 * dead_end children left. There shouldn't be any regular backends
3767 * left by now anyway; what we're really waiting for is walsenders and
3768 * archiver.
3769 */
3770 if (PgArchPID == 0 && CountChildren(BACKEND_TYPE_ALL) == 0)
3771 {
3772 pmState = PM_WAIT_DEAD_END;
3773 }
3774 }
3775
3776 if (pmState == PM_WAIT_DEAD_END)
3777 {
3778 /*
3779 * PM_WAIT_DEAD_END state ends when the BackendList is entirely empty
3780 * (ie, no dead_end children remain), and the archiver and stats
3781 * collector are gone too.
3782 *
3783 * The reason we wait for those two is to protect them against a new
3784 * postmaster starting conflicting subprocesses; this isn't an
3785 * ironclad protection, but it at least helps in the
3786 * shutdown-and-immediately-restart scenario. Note that they have
3787 * already been sent appropriate shutdown signals, either during a
3788 * normal state transition leading up to PM_WAIT_DEAD_END, or during
3789 * FatalError processing.
3790 */
3791 if (dlist_is_empty(&BackendList) &&
3792 PgArchPID == 0 && PgStatPID == 0)
3793 {
3794 /* These other guys should be dead already */
3795 Assert(StartupPID == 0);
3796 Assert(WalReceiverPID == 0);
3797 Assert(BgWriterPID == 0);
3798 Assert(CheckpointerPID == 0);
3799 Assert(WalWriterPID == 0);
3800 Assert(AutoVacPID == 0);
3801 /* syslogger is not considered here */
3802 pmState = PM_NO_CHILDREN;
3803 }
3804 }
3805
3806 /*
3807 * If we've been told to shut down, we exit as soon as there are no
3808 * remaining children. If there was a crash, cleanup will occur at the
3809 * next startup. (Before PostgreSQL 8.3, we tried to recover from the
3810 * crash before exiting, but that seems unwise if we are quitting because
3811 * we got SIGTERM from init --- there may well not be time for recovery
3812 * before init decides to SIGKILL us.)
3813 *
3814 * Note that the syslogger continues to run. It will exit when it sees
3815 * EOF on its input pipe, which happens when there are no more upstream
3816 * processes.
3817 */
3818 if (Shutdown > NoShutdown && pmState == PM_NO_CHILDREN)
3819 {
3820 if (FatalError)
3821 {
3822 ereport(LOG, (errmsg("abnormal database system shutdown")));
3823 ExitPostmaster(1);
3824 }
3825 else
3826 {
3827 /*
3828 * Terminate exclusive backup mode to avoid recovery after a clean
3829 * fast shutdown. Since an exclusive backup can only be taken
3830 * during normal running (and not, for example, while running
3831 * under Hot Standby) it only makes sense to do this if we reached
3832 * normal running. If we're still in recovery, the backup file is
3833 * one we're recovering *from*, and we must keep it around so that
3834 * recovery restarts from the right place.
3835 */
3836 if (ReachedNormalRunning)
3837 CancelBackup();
3838
3839 /* Normal exit from the postmaster is here */
3840 ExitPostmaster(0);
3841 }
3842 }
3843
3844 /*
3845 * If the startup process failed, or the user does not want an automatic
3846 * restart after backend crashes, wait for all non-syslogger children to
3847 * exit, and then exit postmaster. We don't try to reinitialize when the
3848 * startup process fails, because more than likely it will just fail again
3849 * and we will keep trying forever.
3850 */
3851 if (pmState == PM_NO_CHILDREN &&
3852 (StartupStatus == STARTUP_CRASHED || !restart_after_crash))
3853 ExitPostmaster(1);
3854
3855 /*
3856 * If we need to recover from a crash, wait for all non-syslogger children
3857 * to exit, then reset shmem and StartupDataBase.
3858 */
3859 if (FatalError && pmState == PM_NO_CHILDREN)
3860 {
3861 ereport(LOG,
3862 (errmsg("all server processes terminated; reinitializing")));
3863
3864 /* allow background workers to immediately restart */
3865 ResetBackgroundWorkerCrashTimes();
3866
3867 shmem_exit(1);
3868 reset_shared(PostPortNumber);
3869
3870 StartupPID = StartupDataBase();
3871 Assert(StartupPID != 0);
3872 StartupStatus = STARTUP_RUNNING;
3873 pmState = PM_STARTUP;
3874 /* crash recovery started, reset SIGKILL flag */
3875 AbortStartTime = 0;
3876 }
3877 }
3878
3879
3880 /*
3881 * Send a signal to a postmaster child process
3882 *
3883 * On systems that have setsid(), each child process sets itself up as a
3884 * process group leader. For signals that are generally interpreted in the
3885 * appropriate fashion, we signal the entire process group not just the
3886 * direct child process. This allows us to, for example, SIGQUIT a blocked
3887 * archive_recovery script, or SIGINT a script being run by a backend via
3888 * system().
3889 *
3890 * There is a race condition for recently-forked children: they might not
3891 * have executed setsid() yet. So we signal the child directly as well as
3892 * the group. We assume such a child will handle the signal before trying
3893 * to spawn any grandchild processes. We also assume that signaling the
3894 * child twice will not cause any problems.
3895 */
3896 static void
3897 signal_child(pid_t pid, int signal)
3898 {
3899 if (kill(pid, signal) < 0)
3900 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) pid, signal);
3901 #ifdef HAVE_SETSID
3902 switch (signal)
3903 {
3904 case SIGINT:
3905 case SIGTERM:
3906 case SIGQUIT:
3907 case SIGSTOP:
3908 case SIGKILL:
3909 if (kill(-pid, signal) < 0)
3910 elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) (-pid), signal);
3911 break;
3912 default:
3913 break;
3914 }
3915 #endif
3916 }
3917
3918 /*
3919 * Send a signal to the targeted children (but NOT special children;
3920 * dead_end children are never signaled, either).
3921 */
3922 static bool
3923 SignalSomeChildren(int signal, int target)
3924 {
3925 dlist_iter iter;
3926 bool signaled = false;
3927
3928 dlist_foreach(iter, &BackendList)
3929 {
3930 Backend *bp = dlist_container(Backend, elem, iter.cur);
3931
3932 if (bp->dead_end)
3933 continue;
3934
3935 /*
3936 * Since target == BACKEND_TYPE_ALL is the most common case, we test
3937 * it first and avoid touching shared memory for every child.
3938 */
3939 if (target != BACKEND_TYPE_ALL)
3940 {
3941 /*
3942 * Assign bkend_type for any recently announced WAL Sender
3943 * processes.
3944 */
3945 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
3946 IsPostmasterChildWalSender(bp->child_slot))
3947 bp->bkend_type = BACKEND_TYPE_WALSND;
3948
3949 if (!(target & bp->bkend_type))
3950 continue;
3951 }
3952
3953 ereport(DEBUG4,
3954 (errmsg_internal("sending signal %d to process %d",
3955 signal, (int) bp->pid)));
3956 signal_child(bp->pid, signal);
3957 signaled = true;
3958 }
3959 return signaled;
3960 }
3961
3962 /*
3963 * Send a termination signal to children. This considers all of our children
3964 * processes, except syslogger and dead_end backends.
3965 */
3966 static void
3967 TerminateChildren(int signal)
3968 {
3969 SignalChildren(signal);
3970 if (StartupPID != 0)
3971 {
3972 signal_child(StartupPID, signal);
3973 if (signal == SIGQUIT || signal == SIGKILL)
3974 StartupStatus = STARTUP_SIGNALED;
3975 }
3976 if (BgWriterPID != 0)
3977 signal_child(BgWriterPID, signal);
3978 if (CheckpointerPID != 0)
3979 signal_child(CheckpointerPID, signal);
3980 if (WalWriterPID != 0)
3981 signal_child(WalWriterPID, signal);
3982 if (WalReceiverPID != 0)
3983 signal_child(WalReceiverPID, signal);
3984 if (AutoVacPID != 0)
3985 signal_child(AutoVacPID, signal);
3986 if (PgArchPID != 0)
3987 signal_child(PgArchPID, signal);
3988 if (PgStatPID != 0)
3989 signal_child(PgStatPID, signal);
3990 }
3991
3992 /*
3993 * BackendStartup -- start backend process
3994 *
3995 * returns: STATUS_ERROR if the fork failed, STATUS_OK otherwise.
3996 *
3997 * Note: if you change this code, also consider StartAutovacuumWorker.
3998 */
3999 static int
4000 BackendStartup(Port *port)
4001 {
4002 Backend *bn; /* for backend cleanup */
4003 pid_t pid;
4004
4005 /*
4006 * Create backend data structure. Better before the fork() so we can
4007 * handle failure cleanly.
4008 */
4009 bn = (Backend *) malloc(sizeof(Backend));
4010 if (!bn)
4011 {
4012 ereport(LOG,
4013 (errcode(ERRCODE_OUT_OF_MEMORY),
4014 errmsg("out of memory")));
4015 return STATUS_ERROR;
4016 }
4017
4018 /*
4019 * Compute the cancel key that will be assigned to this backend. The
4020 * backend will have its own copy in the forked-off process' value of
4021 * MyCancelKey, so that it can transmit the key to the frontend.
4022 */
4023 MyCancelKey = PostmasterRandom();
4024 bn->cancel_key = MyCancelKey;
4025
4026 /* Pass down canAcceptConnections state */
4027 port->canAcceptConnections = canAcceptConnections(BACKEND_TYPE_NORMAL);
4028 bn->dead_end = (port->canAcceptConnections != CAC_OK &&
4029 port->canAcceptConnections != CAC_SUPERUSER);
4030
4031 /*
4032 * Unless it's a dead_end child, assign it a child slot number
4033 */
4034 if (!bn->dead_end)
4035 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
4036 else
4037 bn->child_slot = 0;
4038
4039 /* Hasn't asked to be notified about any bgworkers yet */
4040 bn->bgworker_notify = false;
4041
4042 #ifdef EXEC_BACKEND
4043 pid = backend_forkexec(port);
4044 #else /* !EXEC_BACKEND */
4045 pid = fork_process();
4046 if (pid == 0) /* child */
4047 {
4048 free(bn);
4049
4050 /* Detangle from postmaster */
4051 InitPostmasterChild();
4052
4053 /* Close the postmaster's sockets */
4054 ClosePostmasterPorts(false);
4055
4056 /* Perform additional initialization and collect startup packet */
4057 BackendInitialize(port);
4058
4059 /* And run the backend */
4060 BackendRun(port);
4061 }
4062 #endif /* EXEC_BACKEND */
4063
4064 if (pid < 0)
4065 {
4066 /* in parent, fork failed */
4067 int save_errno = errno;
4068
4069 if (!bn->dead_end)
4070 (void) ReleasePostmasterChildSlot(bn->child_slot);
4071 free(bn);
4072 errno = save_errno;
4073 ereport(LOG,
4074 (errmsg("could not fork new process for connection: %m")));
4075 report_fork_failure_to_client(port, save_errno);
4076 return STATUS_ERROR;
4077 }
4078
4079 /* in parent, successful fork */
4080 ereport(DEBUG2,
4081 (errmsg_internal("forked new backend, pid=%d socket=%d",
4082 (int) pid, (int) port->sock)));
4083
4084 /*
4085 * Everything's been successful, it's safe to add this backend to our list
4086 * of backends.
4087 */
4088 bn->pid = pid;
4089 bn->bkend_type = BACKEND_TYPE_NORMAL; /* Can change later to WALSND */
4090 dlist_push_head(&BackendList, &bn->elem);
4091
4092 #ifdef EXEC_BACKEND
4093 if (!bn->dead_end)
4094 ShmemBackendArrayAdd(bn);
4095 #endif
4096
4097 return STATUS_OK;
4098 }
4099
4100 /*
4101 * Try to report backend fork() failure to client before we close the
4102 * connection. Since we do not care to risk blocking the postmaster on
4103 * this connection, we set the connection to non-blocking and try only once.
4104 *
4105 * This is grungy special-purpose code; we cannot use backend libpq since
4106 * it's not up and running.
4107 */
4108 static void
4109 report_fork_failure_to_client(Port *port, int errnum)
4110 {
4111 char buffer[1000];
4112 int rc;
4113
4114 /* Format the error message packet (always V2 protocol) */
4115 snprintf(buffer, sizeof(buffer), "E%s%s\n",
4116 _("could not fork new process for connection: "),
4117 strerror(errnum));
4118
4119 /* Set port to non-blocking. Don't do send() if this fails */
4120 if (!pg_set_noblock(port->sock))
4121 return;
4122
4123 /* We'll retry after EINTR, but ignore all other failures */
4124 do
4125 {
4126 rc = send(port->sock, buffer, strlen(buffer) + 1, 0);
4127 } while (rc < 0 && errno == EINTR);
4128 }
4129
4130
4131 /*
4132 * BackendInitialize -- initialize an interactive (postmaster-child)
4133 * backend process, and collect the client's startup packet.
4134 *
4135 * returns: nothing. Will not return at all if there's any failure.
4136 *
4137 * Note: this code does not depend on having any access to shared memory.
4138 * In the EXEC_BACKEND case, we are physically attached to shared memory
4139 * but have not yet set up most of our local pointers to shmem structures.
4140 */
4141 static void
4142 BackendInitialize(Port *port)
4143 {
4144 int status;
4145 int ret;
4146 char remote_host[NI_MAXHOST];
4147 char remote_port[NI_MAXSERV];
4148 char remote_ps_data[NI_MAXHOST];
4149
4150 /* Save port etc. for ps status */
4151 MyProcPort = port;
4152
4153 /*
4154 * PreAuthDelay is a debugging aid for investigating problems in the
4155 * authentication cycle: it can be set in postgresql.conf to allow time to
4156 * attach to the newly-forked backend with a debugger. (See also
4157 * PostAuthDelay, which we allow clients to pass through PGOPTIONS, but it
4158 * is not honored until after authentication.)
4159 */
4160 if (PreAuthDelay > 0)
4161 pg_usleep(PreAuthDelay * 1000000L);
4162
4163 /* This flag will remain set until InitPostgres finishes authentication */
4164 ClientAuthInProgress = true; /* limit visibility of log messages */
4165
4166 /* save process start time */
4167 port->SessionStartTime = GetCurrentTimestamp();
4168 MyStartTime = timestamptz_to_time_t(port->SessionStartTime);
4169
4170 /* set these to empty in case they are needed before we set them up */
4171 port->remote_host = "";
4172 port->remote_port = "";
4173
4174 /*
4175 * Initialize libpq and enable reporting of ereport errors to the client.
4176 * Must do this now because authentication uses libpq to send messages.
4177 */
4178 pq_init(); /* initialize libpq to talk to client */
4179 whereToSendOutput = DestRemote; /* now safe to ereport to client */
4180
4181 /*
4182 * We arrange to do proc_exit(1) if we receive SIGTERM or timeout while
4183 * trying to collect the startup packet; while SIGQUIT results in
4184 * _exit(2). Otherwise the postmaster cannot shutdown the database FAST
4185 * or IMMED cleanly if a buggy client fails to send the packet promptly.
4186 *
4187 * XXX this is pretty dangerous; signal handlers should not call anything
4188 * as complex as proc_exit() directly. We minimize the hazard by not
4189 * keeping these handlers active for longer than we must. However, it
4190 * seems necessary to be able to escape out of DNS lookups as well as the
4191 * startup packet reception proper, so we can't narrow the scope further
4192 * than is done here.
4193 *
4194 * XXX it follows that the remainder of this function must tolerate losing
4195 * control at any instant. Likewise, any pg_on_exit_callback registered
4196 * before or during this function must be prepared to execute at any
4197 * instant between here and the end of this function. Furthermore,
4198 * affected callbacks execute partially or not at all when a second
4199 * exit-inducing signal arrives after proc_exit_prepare() decrements
4200 * on_proc_exit_index. (Thanks to that mechanic, callbacks need not
4201 * anticipate more than one call.) This is fragile; it ought to instead
4202 * follow the norm of handling interrupts at selected, safe opportunities.
4203 */
4204 pqsignal(SIGTERM, process_startup_packet_die);
4205 pqsignal(SIGQUIT, process_startup_packet_quickdie);
4206 InitializeTimeouts(); /* establishes SIGALRM handler */
4207 PG_SETMASK(&StartupBlockSig);
4208
4209 /*
4210 * Get the remote host name and port for logging and status display.
4211 */
4212 remote_host[0] = '\0';
4213 remote_port[0] = '\0';
4214 if ((ret = pg_getnameinfo_all(&port->raddr.addr, port->raddr.salen,
4215 remote_host, sizeof(remote_host),
4216 remote_port, sizeof(remote_port),
4217 (log_hostname ? 0 : NI_NUMERICHOST) | NI_NUMERICSERV)) != 0)
4218 ereport(WARNING,
4219 (errmsg_internal("pg_getnameinfo_all() failed: %s",
4220 gai_strerror(ret))));
4221 if (remote_port[0] == '\0')
4222 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s", remote_host);
4223 else
4224 snprintf(remote_ps_data, sizeof(remote_ps_data), "%s(%s)", remote_host, remote_port);
4225
4226 /*
4227 * Save remote_host and remote_port in port structure (after this, they
4228 * will appear in log_line_prefix data for log messages).
4229 */
4230 port->remote_host = strdup(remote_host);
4231 port->remote_port = strdup(remote_port);
4232
4233 /* And now we can issue the Log_connections message, if wanted */
4234 if (Log_connections)
4235 {
4236 if (remote_port[0])
4237 ereport(LOG,
4238 (errmsg("connection received: host=%s port=%s",
4239 remote_host,
4240 remote_port)));
4241 else
4242 ereport(LOG,
4243 (errmsg("connection received: host=%s",
4244 remote_host)));
4245 }
4246
4247 /*
4248 * If we did a reverse lookup to name, we might as well save the results
4249 * rather than possibly repeating the lookup during authentication.
4250 *
4251 * Note that we don't want to specify NI_NAMEREQD above, because then we'd
4252 * get nothing useful for a client without an rDNS entry. Therefore, we
4253 * must check whether we got a numeric IPv4 or IPv6 address, and not save
4254 * it into remote_hostname if so. (This test is conservative and might
4255 * sometimes classify a hostname as numeric, but an error in that
4256 * direction is safe; it only results in a possible extra lookup.)
4257 */
4258 if (log_hostname &&
4259 ret == 0 &&
4260 strspn(remote_host, "0123456789.") < strlen(remote_host) &&
4261 strspn(remote_host, "0123456789ABCDEFabcdef:") < strlen(remote_host))
4262 port->remote_hostname = strdup(remote_host);
4263
4264 /*
4265 * Ready to begin client interaction. We will give up and proc_exit(1)
4266 * after a time delay, so that a broken client can't hog a connection
4267 * indefinitely. PreAuthDelay and any DNS interactions above don't count
4268 * against the time limit.
4269 *
4270 * Note: AuthenticationTimeout is applied here while waiting for the
4271 * startup packet, and then again in InitPostgres for the duration of any
4272 * authentication operations. So a hostile client could tie up the
4273 * process for nearly twice AuthenticationTimeout before we kick him off.
4274 *
4275 * Note: because PostgresMain will call InitializeTimeouts again, the
4276 * registration of STARTUP_PACKET_TIMEOUT will be lost. This is okay
4277 * since we never use it again after this function.
4278 */
4279 RegisterTimeout(STARTUP_PACKET_TIMEOUT, StartupPacketTimeoutHandler);
4280 enable_timeout_after(STARTUP_PACKET_TIMEOUT, AuthenticationTimeout * 1000);
4281
4282 /*
4283 * Receive the startup packet (which might turn out to be a cancel request
4284 * packet).
4285 */
4286 status = ProcessStartupPacket(port, false);
4287
4288 /*
4289 * Disable the timeout, and prevent SIGTERM/SIGQUIT again.
4290 */
4291 disable_timeout(STARTUP_PACKET_TIMEOUT, false);
4292 PG_SETMASK(&BlockSig);
4293
4294 /*
4295 * Stop here if it was bad or a cancel packet. ProcessStartupPacket
4296 * already did any appropriate error reporting.
4297 */
4298 if (status != STATUS_OK)
4299 proc_exit(0);
4300
4301 /*
4302 * Now that we have the user and database name, we can set the process
4303 * title for ps. It's good to do this as early as possible in startup.
4304 *
4305 * For a walsender, the ps display is set in the following form:
4306 *
4307 * postgres: wal sender process <user> <host> <activity>
4308 *
4309 * To achieve that, we pass "wal sender process" as username and username
4310 * as dbname to init_ps_display(). XXX: should add a new variant of
4311 * init_ps_display() to avoid abusing the parameters like this.
4312 */
4313 if (am_walsender)
4314 init_ps_display("wal sender process", port->user_name, remote_ps_data,
4315 update_process_title ? "authentication" : "");
4316 else
4317 init_ps_display(port->user_name, port->database_name, remote_ps_data,
4318 update_process_title ? "authentication" : "");
4319 }
4320
4321
4322 /*
4323 * BackendRun -- set up the backend's argument list and invoke PostgresMain()
4324 *
4325 * returns:
4326 * Shouldn't return at all.
4327 * If PostgresMain() fails, return status.
4328 */
4329 static void
4330 BackendRun(Port *port)
4331 {
4332 char **av;
4333 int maxac;
4334 int ac;
4335 long secs;
4336 int usecs;
4337 int i;
4338
4339 /*
4340 * Don't want backend to be able to see the postmaster random number
4341 * generator state. We have to clobber the static random_seed *and* start
4342 * a new random sequence in the random() library function.
4343 */
4344 random_seed = 0;
4345 random_start_time.tv_usec = 0;
4346 /* slightly hacky way to convert timestamptz into integers */
4347 TimestampDifference(0, port->SessionStartTime, &secs, &usecs);
4348 srandom((unsigned int) (MyProcPid ^ (usecs << 12) ^ secs));
4349
4350 /*
4351 * Now, build the argv vector that will be given to PostgresMain.
4352 *
4353 * The maximum possible number of commandline arguments that could come
4354 * from ExtraOptions is (strlen(ExtraOptions) + 1) / 2; see
4355 * pg_split_opts().
4356 */
4357 maxac = 2; /* for fixed args supplied below */
4358 maxac += (strlen(ExtraOptions) + 1) / 2;
4359
4360 av = (char **) MemoryContextAlloc(TopMemoryContext,
4361 maxac * sizeof(char *));
4362 ac = 0;
4363
4364 av[ac++] = "postgres";
4365
4366 /*
4367 * Pass any backend switches specified with -o on the postmaster's own
4368 * command line. We assume these are secure.
4369 */
4370 pg_split_opts(av, &ac, ExtraOptions);
4371
4372 av[ac] = NULL;
4373
4374 Assert(ac < maxac);
4375
4376 /*
4377 * Debug: print arguments being passed to backend
4378 */
4379 ereport(DEBUG3,
4380 (errmsg_internal("%s child[%d]: starting with (",
4381 progname, (int) getpid())));
4382 for (i = 0; i < ac; ++i)
4383 ereport(DEBUG3,
4384 (errmsg_internal("\t%s", av[i])));
4385 ereport(DEBUG3,
4386 (errmsg_internal(")")));
4387
4388 /*
4389 * Make sure we aren't in PostmasterContext anymore. (We can't delete it
4390 * just yet, though, because InitPostgres will need the HBA data.)
4391 */
4392 MemoryContextSwitchTo(TopMemoryContext);
4393
4394 PostgresMain(ac, av, port->database_name, port->user_name);
4395 }
4396
4397
4398 #ifdef EXEC_BACKEND
4399
4400 /*
4401 * postmaster_forkexec -- fork and exec a postmaster subprocess
4402 *
4403 * The caller must have set up the argv array already, except for argv[2]
4404 * which will be filled with the name of the temp variable file.
4405 *
4406 * Returns the child process PID, or -1 on fork failure (a suitable error
4407 * message has been logged on failure).
4408 *
4409 * All uses of this routine will dispatch to SubPostmasterMain in the
4410 * child process.
4411 */
4412 pid_t
4413 postmaster_forkexec(int argc, char *argv[])
4414 {
4415 Port port;
4416
4417 /* This entry point passes dummy values for the Port variables */
4418 memset(&port, 0, sizeof(port));
4419 return internal_forkexec(argc, argv, &port);
4420 }
4421
4422 /*
4423 * backend_forkexec -- fork/exec off a backend process
4424 *
4425 * Some operating systems (WIN32) don't have fork() so we have to simulate
4426 * it by storing parameters that need to be passed to the child and
4427 * then create a new child process.
4428 *
4429 * returns the pid of the fork/exec'd process, or -1 on failure
4430 */
4431 static pid_t
4432 backend_forkexec(Port *port)
4433 {
4434 char *av[4];
4435 int ac = 0;
4436
4437 av[ac++] = "postgres";
4438 av[ac++] = "--forkbackend";
4439 av[ac++] = NULL; /* filled in by internal_forkexec */
4440
4441 av[ac] = NULL;
4442 Assert(ac < lengthof(av));
4443
4444 return internal_forkexec(ac, av, port);
4445 }
4446
4447 #ifndef WIN32
4448
4449 /*
4450 * internal_forkexec non-win32 implementation
4451 *
4452 * - writes out backend variables to the parameter file
4453 * - fork():s, and then exec():s the child process
4454 */
4455 static pid_t
4456 internal_forkexec(int argc, char *argv[], Port *port)
4457 {
4458 static unsigned long tmpBackendFileNum = 0;
4459 pid_t pid;
4460 char tmpfilename[MAXPGPATH];
4461 BackendParameters param;
4462 FILE *fp;
4463
4464 if (!save_backend_variables(¶m, port))
4465 return -1; /* log made by save_backend_variables */
4466
4467 /* Calculate name for temp file */
4468 snprintf(tmpfilename, MAXPGPATH, "%s/%s.backend_var.%d.%lu",
4469 PG_TEMP_FILES_DIR, PG_TEMP_FILE_PREFIX,
4470 MyProcPid, ++tmpBackendFileNum);
4471
4472 /* Open file */
4473 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4474 if (!fp)
4475 {
4476 /*
4477 * As in OpenTemporaryFileInTablespace, try to make the temp-file
4478 * directory
4479 */
4480 mkdir(PG_TEMP_FILES_DIR, S_IRWXU);
4481
4482 fp = AllocateFile(tmpfilename, PG_BINARY_W);
4483 if (!fp)
4484 {
4485 ereport(LOG,
4486 (errcode_for_file_access(),
4487 errmsg("could not create file \"%s\": %m",
4488 tmpfilename)));
4489 return -1;
4490 }
4491 }
4492
4493 if (fwrite(¶m, sizeof(param), 1, fp) != 1)
4494 {
4495 ereport(LOG,
4496 (errcode_for_file_access(),
4497 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4498 FreeFile(fp);
4499 return -1;
4500 }
4501
4502 /* Release file */
4503 if (FreeFile(fp))
4504 {
4505 ereport(LOG,
4506 (errcode_for_file_access(),
4507 errmsg("could not write to file \"%s\": %m", tmpfilename)));
4508 return -1;
4509 }
4510
4511 /* Make sure caller set up argv properly */
4512 Assert(argc >= 3);
4513 Assert(argv[argc] == NULL);
4514 Assert(strncmp(argv[1], "--fork", 6) == 0);
4515 Assert(argv[2] == NULL);
4516
4517 /* Insert temp file name after --fork argument */
4518 argv[2] = tmpfilename;
4519
4520 /* Fire off execv in child */
4521 if ((pid = fork_process()) == 0)
4522 {
4523 if (execv(postgres_exec_path, argv) < 0)
4524 {
4525 ereport(LOG,
4526 (errmsg("could not execute server process \"%s\": %m",
4527 postgres_exec_path)));
4528 /* We're already in the child process here, can't return */
4529 exit(1);
4530 }
4531 }
4532
4533 return pid; /* Parent returns pid, or -1 on fork failure */
4534 }
4535 #else /* WIN32 */
4536
4537 /*
4538 * internal_forkexec win32 implementation
4539 *
4540 * - starts backend using CreateProcess(), in suspended state
4541 * - writes out backend variables to the parameter file
4542 * - during this, duplicates handles and sockets required for
4543 * inheritance into the new process
4544 * - resumes execution of the new process once the backend parameter
4545 * file is complete.
4546 */
4547 static pid_t
4548 internal_forkexec(int argc, char *argv[], Port *port)
4549 {
4550 int retry_count = 0;
4551 STARTUPINFO si;
4552 PROCESS_INFORMATION pi;
4553 int i;
4554 int j;
4555 char cmdLine[MAXPGPATH * 2];
4556 HANDLE paramHandle;
4557 BackendParameters *param;
4558 SECURITY_ATTRIBUTES sa;
4559 char paramHandleStr[32];
4560 win32_deadchild_waitinfo *childinfo;
4561
4562 /* Make sure caller set up argv properly */
4563 Assert(argc >= 3);
4564 Assert(argv[argc] == NULL);
4565 Assert(strncmp(argv[1], "--fork", 6) == 0);
4566 Assert(argv[2] == NULL);
4567
4568 /* Resume here if we need to retry */
4569 retry:
4570
4571 /* Set up shared memory for parameter passing */
4572 ZeroMemory(&sa, sizeof(sa));
4573 sa.nLength = sizeof(sa);
4574 sa.bInheritHandle = TRUE;
4575 paramHandle = CreateFileMapping(INVALID_HANDLE_VALUE,
4576 &sa,
4577 PAGE_READWRITE,
4578 0,
4579 sizeof(BackendParameters),
4580 NULL);
4581 if (paramHandle == INVALID_HANDLE_VALUE)
4582 {
4583 elog(LOG, "could not create backend parameter file mapping: error code %lu",
4584 GetLastError());
4585 return -1;
4586 }
4587
4588 param = MapViewOfFile(paramHandle, FILE_MAP_WRITE, 0, 0, sizeof(BackendParameters));
4589 if (!param)
4590 {
4591 elog(LOG, "could not map backend parameter memory: error code %lu",
4592 GetLastError());
4593 CloseHandle(paramHandle);
4594 return -1;
4595 }
4596
4597 /* Insert temp file name after --fork argument */
4598 #ifdef _WIN64
4599 sprintf(paramHandleStr, "%llu", (LONG_PTR) paramHandle);
4600 #else
4601 sprintf(paramHandleStr, "%lu", (DWORD) paramHandle);
4602 #endif
4603 argv[2] = paramHandleStr;
4604
4605 /* Format the cmd line */
4606 cmdLine[sizeof(cmdLine) - 1] = '\0';
4607 cmdLine[sizeof(cmdLine) - 2] = '\0';
4608 snprintf(cmdLine, sizeof(cmdLine) - 1, "\"%s\"", postgres_exec_path);
4609 i = 0;
4610 while (argv[++i] != NULL)
4611 {
4612 j = strlen(cmdLine);
4613 snprintf(cmdLine + j, sizeof(cmdLine) - 1 - j, " \"%s\"", argv[i]);
4614 }
4615 if (cmdLine[sizeof(cmdLine) - 2] != '\0')
4616 {
4617 elog(LOG, "subprocess command line too long");
4618 return -1;
4619 }
4620
4621 memset(&pi, 0, sizeof(pi));
4622 memset(&si, 0, sizeof(si));
4623 si.cb = sizeof(si);
4624
4625 /*
4626 * Create the subprocess in a suspended state. This will be resumed later,
4627 * once we have written out the parameter file.
4628 */
4629 if (!CreateProcess(NULL, cmdLine, NULL, NULL, TRUE, CREATE_SUSPENDED,
4630 NULL, NULL, &si, &pi))
4631 {
4632 elog(LOG, "CreateProcess call failed: %m (error code %lu)",
4633 GetLastError());
4634 return -1;
4635 }
4636
4637 if (!save_backend_variables(param, port, pi.hProcess, pi.dwProcessId))
4638 {
4639 /*
4640 * log made by save_backend_variables, but we have to clean up the
4641 * mess with the half-started process
4642 */
4643 if (!TerminateProcess(pi.hProcess, 255))
4644 ereport(LOG,
4645 (errmsg_internal("could not terminate unstarted process: error code %lu",
4646 GetLastError())));
4647 CloseHandle(pi.hProcess);
4648 CloseHandle(pi.hThread);
4649 return -1; /* log made by save_backend_variables */
4650 }
4651
4652 /* Drop the parameter shared memory that is now inherited to the backend */
4653 if (!UnmapViewOfFile(param))
4654 elog(LOG, "could not unmap view of backend parameter file: error code %lu",
4655 GetLastError());
4656 if (!CloseHandle(paramHandle))
4657 elog(LOG, "could not close handle to backend parameter file: error code %lu",
4658 GetLastError());
4659
4660 /*
4661 * Reserve the memory region used by our main shared memory segment before
4662 * we resume the child process. Normally this should succeed, but if ASLR
4663 * is active then it might sometimes fail due to the stack or heap having
4664 * gotten mapped into that range. In that case, just terminate the
4665 * process and retry.
4666 */
4667 if (!pgwin32_ReserveSharedMemoryRegion(pi.hProcess))
4668 {
4669 /* pgwin32_ReserveSharedMemoryRegion already made a log entry */
4670 if (!TerminateProcess(pi.hProcess, 255))
4671 ereport(LOG,
4672 (errmsg_internal("could not terminate process that failed to reserve memory: error code %lu",
4673 GetLastError())));
4674 CloseHandle(pi.hProcess);
4675 CloseHandle(pi.hThread);
4676 if (++retry_count < 100)
4677 goto retry;
4678 ereport(LOG,
4679 (errmsg("giving up after too many tries to reserve shared memory"),
4680 errhint("This might be caused by ASLR or antivirus software.")));
4681 return -1;
4682 }
4683
4684 /*
4685 * Now that the backend variables are written out, we start the child
4686 * thread so it can start initializing while we set up the rest of the
4687 * parent state.
4688 */
4689 if (ResumeThread(pi.hThread) == -1)
4690 {
4691 if (!TerminateProcess(pi.hProcess, 255))
4692 {
4693 ereport(LOG,
4694 (errmsg_internal("could not terminate unstartable process: error code %lu",
4695 GetLastError())));
4696 CloseHandle(pi.hProcess);
4697 CloseHandle(pi.hThread);
4698 return -1;
4699 }
4700 CloseHandle(pi.hProcess);
4701 CloseHandle(pi.hThread);
4702 ereport(LOG,
4703 (errmsg_internal("could not resume thread of unstarted process: error code %lu",
4704 GetLastError())));
4705 return -1;
4706 }
4707
4708 /*
4709 * Queue a waiter for to signal when this child dies. The wait will be
4710 * handled automatically by an operating system thread pool.
4711 *
4712 * Note: use malloc instead of palloc, since it needs to be thread-safe.
4713 * Struct will be free():d from the callback function that runs on a
4714 * different thread.
4715 */
4716 childinfo = malloc(sizeof(win32_deadchild_waitinfo));
4717 if (!childinfo)
4718 ereport(FATAL,
4719 (errcode(ERRCODE_OUT_OF_MEMORY),
4720 errmsg("out of memory")));
4721
4722 childinfo->procHandle = pi.hProcess;
4723 childinfo->procId = pi.dwProcessId;
4724
4725 if (!RegisterWaitForSingleObject(&childinfo->waitHandle,
4726 pi.hProcess,
4727 pgwin32_deadchild_callback,
4728 childinfo,
4729 INFINITE,
4730 WT_EXECUTEONLYONCE | WT_EXECUTEINWAITTHREAD))
4731 ereport(FATAL,
4732 (errmsg_internal("could not register process for wait: error code %lu",
4733 GetLastError())));
4734
4735 /* Don't close pi.hProcess here - the wait thread needs access to it */
4736
4737 CloseHandle(pi.hThread);
4738
4739 return pi.dwProcessId;
4740 }
4741 #endif /* WIN32 */
4742
4743
4744 /*
4745 * SubPostmasterMain -- Get the fork/exec'd process into a state equivalent
4746 * to what it would be if we'd simply forked on Unix, and then
4747 * dispatch to the appropriate place.
4748 *
4749 * The first two command line arguments are expected to be "--forkFOO"
4750 * (where FOO indicates which postmaster child we are to become), and
4751 * the name of a variables file that we can read to load data that would
4752 * have been inherited by fork() on Unix. Remaining arguments go to the
4753 * subprocess FooMain() routine.
4754 */
4755 void
4756 SubPostmasterMain(int argc, char *argv[])
4757 {
4758 Port port;
4759
4760 /* In EXEC_BACKEND case we will not have inherited these settings */
4761 IsPostmasterEnvironment = true;
4762 whereToSendOutput = DestNone;
4763
4764 /* Setup as postmaster child */
4765 InitPostmasterChild();
4766
4767 /* Setup essential subsystems (to ensure elog() behaves sanely) */
4768 InitializeGUCOptions();
4769
4770 /* Check we got appropriate args */
4771 if (argc < 3)
4772 elog(FATAL, "invalid subpostmaster invocation");
4773
4774 /* Read in the variables file */
4775 memset(&port, 0, sizeof(Port));
4776 read_backend_variables(argv[2], &port);
4777
4778 /* Close the postmaster's sockets (as soon as we know them) */
4779 ClosePostmasterPorts(strcmp(argv[1], "--forklog") == 0);
4780
4781 /*
4782 * Set reference point for stack-depth checking
4783 */
4784 set_stack_base();
4785
4786 /*
4787 * Set up memory area for GSS information. Mirrors the code in ConnCreate
4788 * for the non-exec case.
4789 */
4790 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
4791 port.gss = (pg_gssinfo *) calloc(1, sizeof(pg_gssinfo));
4792 if (!port.gss)
4793 ereport(FATAL,
4794 (errcode(ERRCODE_OUT_OF_MEMORY),
4795 errmsg("out of memory")));
4796 #endif
4797
4798 /*
4799 * If appropriate, physically re-attach to shared memory segment. We want
4800 * to do this before going any further to ensure that we can attach at the
4801 * same address the postmaster used. On the other hand, if we choose not
4802 * to re-attach, we may have other cleanup to do.
4803 *
4804 * If testing EXEC_BACKEND on Linux, you should run this as root before
4805 * starting the postmaster:
4806 *
4807 * echo 0 >/proc/sys/kernel/randomize_va_space
4808 *
4809 * This prevents using randomized stack and code addresses that cause the
4810 * child process's memory map to be different from the parent's, making it
4811 * sometimes impossible to attach to shared memory at the desired address.
4812 * Return the setting to its old value (usually '1' or '2') when finished.
4813 */
4814 if (strcmp(argv[1], "--forkbackend") == 0 ||
4815 strcmp(argv[1], "--forkavlauncher") == 0 ||
4816 strcmp(argv[1], "--forkavworker") == 0 ||
4817 strcmp(argv[1], "--forkboot") == 0 ||
4818 strncmp(argv[1], "--forkbgworker=", 15) == 0)
4819 PGSharedMemoryReAttach();
4820 else
4821 PGSharedMemoryNoReAttach();
4822
4823 /* autovacuum needs this set before calling InitProcess */
4824 if (strcmp(argv[1], "--forkavlauncher") == 0)
4825 AutovacuumLauncherIAm();
4826 if (strcmp(argv[1], "--forkavworker") == 0)
4827 AutovacuumWorkerIAm();
4828
4829 /*
4830 * Start our win32 signal implementation. This has to be done after we
4831 * read the backend variables, because we need to pick up the signal pipe
4832 * from the parent process.
4833 */
4834 #ifdef WIN32
4835 pgwin32_signal_initialize();
4836 #endif
4837
4838 /* In EXEC_BACKEND case we will not have inherited these settings */
4839 pqinitmask();
4840 PG_SETMASK(&BlockSig);
4841
4842 /* Read in remaining GUC variables */
4843 read_nondefault_variables();
4844
4845 /*
4846 * Reload any libraries that were preloaded by the postmaster. Since we
4847 * exec'd this process, those libraries didn't come along with us; but we
4848 * should load them into all child processes to be consistent with the
4849 * non-EXEC_BACKEND behavior.
4850 */
4851 process_shared_preload_libraries();
4852
4853 /* Run backend or appropriate child */
4854 if (strcmp(argv[1], "--forkbackend") == 0)
4855 {
4856 Assert(argc == 3); /* shouldn't be any more args */
4857
4858 /*
4859 * Need to reinitialize the SSL library in the backend, since the
4860 * context structures contain function pointers and cannot be passed
4861 * through the parameter file.
4862 *
4863 * XXX should we do this in all child processes? For the moment it's
4864 * enough to do it in backend children.
4865 */
4866 #ifdef USE_SSL
4867 if (EnableSSL)
4868 secure_initialize();
4869 #endif
4870
4871 /*
4872 * Perform additional initialization and collect startup packet.
4873 *
4874 * We want to do this before InitProcess() for a couple of reasons: 1.
4875 * so that we aren't eating up a PGPROC slot while waiting on the
4876 * client. 2. so that if InitProcess() fails due to being out of
4877 * PGPROC slots, we have already initialized libpq and are able to
4878 * report the error to the client.
4879 */
4880 BackendInitialize(&port);
4881
4882 /* Restore basic shared memory pointers */
4883 InitShmemAccess(UsedShmemSegAddr);
4884
4885 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4886 InitProcess();
4887
4888 /* Attach process to shared data structures */
4889 CreateSharedMemoryAndSemaphores(0);
4890
4891 /* And run the backend */
4892 BackendRun(&port); /* does not return */
4893 }
4894 if (strcmp(argv[1], "--forkboot") == 0)
4895 {
4896 /* Restore basic shared memory pointers */
4897 InitShmemAccess(UsedShmemSegAddr);
4898
4899 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4900 InitAuxiliaryProcess();
4901
4902 /* Attach process to shared data structures */
4903 CreateSharedMemoryAndSemaphores(0);
4904
4905 AuxiliaryProcessMain(argc - 2, argv + 2); /* does not return */
4906 }
4907 if (strcmp(argv[1], "--forkavlauncher") == 0)
4908 {
4909 /* Restore basic shared memory pointers */
4910 InitShmemAccess(UsedShmemSegAddr);
4911
4912 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4913 InitProcess();
4914
4915 /* Attach process to shared data structures */
4916 CreateSharedMemoryAndSemaphores(0);
4917
4918 AutoVacLauncherMain(argc - 2, argv + 2); /* does not return */
4919 }
4920 if (strcmp(argv[1], "--forkavworker") == 0)
4921 {
4922 /* Restore basic shared memory pointers */
4923 InitShmemAccess(UsedShmemSegAddr);
4924
4925 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4926 InitProcess();
4927
4928 /* Attach process to shared data structures */
4929 CreateSharedMemoryAndSemaphores(0);
4930
4931 AutoVacWorkerMain(argc - 2, argv + 2); /* does not return */
4932 }
4933 if (strncmp(argv[1], "--forkbgworker=", 15) == 0)
4934 {
4935 int shmem_slot;
4936
4937 /* do this as early as possible; in particular, before InitProcess() */
4938 IsBackgroundWorker = true;
4939
4940 /* Restore basic shared memory pointers */
4941 InitShmemAccess(UsedShmemSegAddr);
4942
4943 /* Need a PGPROC to run CreateSharedMemoryAndSemaphores */
4944 InitProcess();
4945
4946 /* Attach process to shared data structures */
4947 CreateSharedMemoryAndSemaphores(0);
4948
4949 /* Fetch MyBgworkerEntry from shared memory */
4950 shmem_slot = atoi(argv[1] + 15);
4951 MyBgworkerEntry = BackgroundWorkerEntry(shmem_slot);
4952
4953 StartBackgroundWorker();
4954 }
4955 if (strcmp(argv[1], "--forkarch") == 0)
4956 {
4957 /* Do not want to attach to shared memory */
4958
4959 PgArchiverMain(argc, argv); /* does not return */
4960 }
4961 if (strcmp(argv[1], "--forkcol") == 0)
4962 {
4963 /* Do not want to attach to shared memory */
4964
4965 PgstatCollectorMain(argc, argv); /* does not return */
4966 }
4967 if (strcmp(argv[1], "--forklog") == 0)
4968 {
4969 /* Do not want to attach to shared memory */
4970
4971 SysLoggerMain(argc, argv); /* does not return */
4972 }
4973
4974 abort(); /* shouldn't get here */
4975 }
4976 #endif /* EXEC_BACKEND */
4977
4978
4979 /*
4980 * ExitPostmaster -- cleanup
4981 *
4982 * Do NOT call exit() directly --- always go through here!
4983 */
4984 static void
4985 ExitPostmaster(int status)
4986 {
4987 #ifdef HAVE_PTHREAD_IS_THREADED_NP
4988
4989 /*
4990 * There is no known cause for a postmaster to become multithreaded after
4991 * startup. Recheck to account for the possibility of unknown causes.
4992 * This message uses LOG level, because an unclean shutdown at this point
4993 * would usually not look much different from a clean shutdown.
4994 */
4995 if (pthread_is_threaded_np() != 0)
4996 ereport(LOG,
4997 (errcode(ERRCODE_INTERNAL_ERROR),
4998 errmsg_internal("postmaster became multithreaded"),
4999 errdetail("Please report this to <pgsql-bugs@postgresql.org>.")));
5000 #endif
5001
5002 /* should cleanup shared memory and kill all backends */
5003
5004 /*
5005 * Not sure of the semantics here. When the Postmaster dies, should the
5006 * backends all be killed? probably not.
5007 *
5008 * MUST -- vadim 05-10-1999
5009 */
5010
5011 proc_exit(status);
5012 }
5013
5014 /*
5015 * sigusr1_handler - handle signal conditions from child processes
5016 */
5017 static void
5018 sigusr1_handler(SIGNAL_ARGS)
5019 {
5020 int save_errno = errno;
5021
5022 /*
5023 * We rely on the signal mechanism to have blocked all signals ... except
5024 * on Windows, which lacks sigaction(), so we have to do it manually.
5025 */
5026 #ifdef WIN32
5027 PG_SETMASK(&BlockSig);
5028 #endif
5029
5030 /* Process background worker state change. */
5031 if (CheckPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE))
5032 {
5033 BackgroundWorkerStateChange();
5034 StartWorkerNeeded = true;
5035 }
5036
5037 /*
5038 * RECOVERY_STARTED and BEGIN_HOT_STANDBY signals are ignored in
5039 * unexpected states. If the startup process quickly starts up, completes
5040 * recovery, exits, we might process the death of the startup process
5041 * first. We don't want to go back to recovery in that case.
5042 */
5043 if (CheckPostmasterSignal(PMSIGNAL_RECOVERY_STARTED) &&
5044 pmState == PM_STARTUP && Shutdown == NoShutdown)
5045 {
5046 /* WAL redo has started. We're out of reinitialization. */
5047 FatalError = false;
5048 Assert(AbortStartTime == 0);
5049
5050 /*
5051 * Crank up the background tasks. It doesn't matter if this fails,
5052 * we'll just try again later.
5053 */
5054 Assert(CheckpointerPID == 0);
5055 CheckpointerPID = StartCheckpointer();
5056 Assert(BgWriterPID == 0);
5057 BgWriterPID = StartBackgroundWriter();
5058
5059 /*
5060 * Start the archiver if we're responsible for (re-)archiving received
5061 * files.
5062 */
5063 Assert(PgArchPID == 0);
5064 if (XLogArchivingAlways())
5065 PgArchPID = pgarch_start();
5066
5067 #ifdef USE_SYSTEMD
5068 if (!EnableHotStandby)
5069 sd_notify(0, "READY=1");
5070 #endif
5071
5072 pmState = PM_RECOVERY;
5073 }
5074 if (CheckPostmasterSignal(PMSIGNAL_BEGIN_HOT_STANDBY) &&
5075 pmState == PM_RECOVERY && Shutdown == NoShutdown)
5076 {
5077 /*
5078 * Likewise, start other special children as needed.
5079 */
5080 Assert(PgStatPID == 0);
5081 PgStatPID = pgstat_start();
5082
5083 ereport(LOG,
5084 (errmsg("database system is ready to accept read only connections")));
5085
5086 #ifdef USE_SYSTEMD
5087 sd_notify(0, "READY=1");
5088 #endif
5089
5090 pmState = PM_HOT_STANDBY;
5091 connsAllowed = ALLOW_ALL_CONNS;
5092
5093 /* Some workers may be scheduled to start now */
5094 StartWorkerNeeded = true;
5095 }
5096
5097 if (StartWorkerNeeded || HaveCrashedWorker)
5098 maybe_start_bgworkers();
5099
5100 if (CheckPostmasterSignal(PMSIGNAL_WAKEN_ARCHIVER) &&
5101 PgArchPID != 0)
5102 {
5103 /*
5104 * Send SIGUSR1 to archiver process, to wake it up and begin archiving
5105 * next transaction log file.
5106 */
5107 signal_child(PgArchPID, SIGUSR1);
5108 }
5109
5110 if (CheckPostmasterSignal(PMSIGNAL_ROTATE_LOGFILE) &&
5111 SysLoggerPID != 0)
5112 {
5113 /* Tell syslogger to rotate logfile */
5114 signal_child(SysLoggerPID, SIGUSR1);
5115 }
5116
5117 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER) &&
5118 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5119 {
5120 /*
5121 * Start one iteration of the autovacuum daemon, even if autovacuuming
5122 * is nominally not enabled. This is so we can have an active defense
5123 * against transaction ID wraparound. We set a flag for the main loop
5124 * to do it rather than trying to do it here --- this is because the
5125 * autovac process itself may send the signal, and we want to handle
5126 * that by launching another iteration as soon as the current one
5127 * completes.
5128 */
5129 start_autovac_launcher = true;
5130 }
5131
5132 if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER) &&
5133 Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
5134 {
5135 /* The autovacuum launcher wants us to start a worker process. */
5136 StartAutovacuumWorker();
5137 }
5138
5139 if (CheckPostmasterSignal(PMSIGNAL_START_WALRECEIVER))
5140 {
5141 /* Startup Process wants us to start the walreceiver process. */
5142 /* Start immediately if possible, else remember request for later. */
5143 WalReceiverRequested = true;
5144 MaybeStartWalReceiver();
5145 }
5146
5147 /*
5148 * Try to advance postmaster's state machine, if a child requests it.
5149 *
5150 * Be careful about the order of this action relative to sigusr1_handler's
5151 * other actions. Generally, this should be after other actions, in case
5152 * they have effects PostmasterStateMachine would need to know about.
5153 * However, we should do it before the CheckPromoteSignal step, which
5154 * cannot have any (immediate) effect on the state machine, but does
5155 * depend on what state we're in now.
5156 */
5157 if (CheckPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE))
5158 {
5159 PostmasterStateMachine();
5160 }
5161
5162 if (StartupPID != 0 &&
5163 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5164 pmState == PM_HOT_STANDBY) &&
5165 CheckPromoteSignal())
5166 {
5167 /* Tell startup process to finish recovery */
5168 signal_child(StartupPID, SIGUSR2);
5169 }
5170
5171 #ifdef WIN32
5172 PG_SETMASK(&UnBlockSig);
5173 #endif
5174
5175 errno = save_errno;
5176 }
5177
5178 /*
5179 * SIGTERM while processing startup packet.
5180 * Clean up and exit(1).
5181 *
5182 * Running proc_exit() from a signal handler is pretty unsafe, since we
5183 * can't know what code we've interrupted. But the alternative of using
5184 * _exit(2) is also unpalatable, since it'd mean that a "fast shutdown"
5185 * would cause a database crash cycle (forcing WAL replay at restart)
5186 * if any sessions are in authentication. So we live with it for now.
5187 *
5188 * One might be tempted to try to send a message indicating why we are
5189 * disconnecting. However, that would make this even more unsafe. Also,
5190 * it seems undesirable to provide clues about the database's state to
5191 * a client that has not yet completed authentication.
5192 */
5193 static void
5194 process_startup_packet_die(SIGNAL_ARGS)
5195 {
5196 proc_exit(1);
5197 }
5198
5199 /*
5200 * SIGQUIT while processing startup packet.
5201 *
5202 * Some backend has bought the farm,
5203 * so we need to stop what we're doing and exit.
5204 */
5205 static void
5206 process_startup_packet_quickdie(SIGNAL_ARGS)
5207 {
5208 /*
5209 * We DO NOT want to run proc_exit() or atexit() callbacks; they wouldn't
5210 * be safe to run from a signal handler. Just nail the windows shut and
5211 * get out of town.
5212 *
5213 * Note we do _exit(2) not _exit(1). This is to force the postmaster into
5214 * a system reset cycle if someone sends a manual SIGQUIT to a random
5215 * backend. (While it might be safe to do _exit(1), since this session
5216 * shouldn't have touched shared memory yet, there seems little point in
5217 * taking any risks.)
5218 */
5219 _exit(2);
5220 }
5221
5222 /*
5223 * Dummy signal handler
5224 *
5225 * We use this for signals that we don't actually use in the postmaster,
5226 * but we do use in backends. If we were to SIG_IGN such signals in the
5227 * postmaster, then a newly started backend might drop a signal that arrives
5228 * before it's able to reconfigure its signal processing. (See notes in
5229 * tcop/postgres.c.)
5230 */
5231 static void
5232 dummy_handler(SIGNAL_ARGS)
5233 {
5234 }
5235
5236 /*
5237 * Timeout while processing startup packet.
5238 * As for process_startup_packet_die(), we clean up and exit(1).
5239 *
5240 * This is theoretically just as hazardous as in process_startup_packet_die(),
5241 * although in practice we're almost certainly waiting for client input,
5242 * which greatly reduces the risk.
5243 */
5244 static void
5245 StartupPacketTimeoutHandler(void)
5246 {
5247 proc_exit(1);
5248 }
5249
5250
5251 /*
5252 * RandomSalt
5253 */
5254 static void
5255 RandomSalt(char *md5Salt)
5256 {
5257 long rand;
5258
5259 /*
5260 * We use % 255, sacrificing one possible byte value, so as to ensure that
5261 * all bits of the random() value participate in the result. While at it,
5262 * add one to avoid generating any null bytes.
5263 */
5264 rand = PostmasterRandom();
5265 md5Salt[0] = (rand % 255) + 1;
5266 rand = PostmasterRandom();
5267 md5Salt[1] = (rand % 255) + 1;
5268 rand = PostmasterRandom();
5269 md5Salt[2] = (rand % 255) + 1;
5270 rand = PostmasterRandom();
5271 md5Salt[3] = (rand % 255) + 1;
5272 }
5273
5274 /*
5275 * PostmasterRandom
5276 *
5277 * Caution: use this only for values needed during connection-request
5278 * processing. Otherwise, the intended property of having an unpredictable
5279 * delay between random_start_time and random_stop_time will be broken.
5280 */
5281 static long
5282 PostmasterRandom(void)
5283 {
5284 /*
5285 * Select a random seed at the time of first receiving a request.
5286 */
5287 if (random_seed == 0)
5288 {
5289 do
5290 {
5291 struct timeval random_stop_time;
5292
5293 gettimeofday(&random_stop_time, NULL);
5294
5295 /*
5296 * We are not sure how much precision is in tv_usec, so we swap
5297 * the high and low 16 bits of 'random_stop_time' and XOR them
5298 * with 'random_start_time'. On the off chance that the result is
5299 * 0, we loop until it isn't.
5300 */
5301 random_seed = random_start_time.tv_usec ^
5302 ((random_stop_time.tv_usec << 16) |
5303 ((random_stop_time.tv_usec >> 16) & 0xffff));
5304 }
5305 while (random_seed == 0);
5306
5307 srandom(random_seed);
5308 }
5309
5310 return random();
5311 }
5312
5313 /*
5314 * Count up number of child processes of specified types (dead_end children
5315 * are always excluded).
5316 */
5317 static int
5318 CountChildren(int target)
5319 {
5320 dlist_iter iter;
5321 int cnt = 0;
5322
5323 dlist_foreach(iter, &BackendList)
5324 {
5325 Backend *bp = dlist_container(Backend, elem, iter.cur);
5326
5327 if (bp->dead_end)
5328 continue;
5329
5330 /*
5331 * Since target == BACKEND_TYPE_ALL is the most common case, we test
5332 * it first and avoid touching shared memory for every child.
5333 */
5334 if (target != BACKEND_TYPE_ALL)
5335 {
5336 /*
5337 * Assign bkend_type for any recently announced WAL Sender
5338 * processes.
5339 */
5340 if (bp->bkend_type == BACKEND_TYPE_NORMAL &&
5341 IsPostmasterChildWalSender(bp->child_slot))
5342 bp->bkend_type = BACKEND_TYPE_WALSND;
5343
5344 if (!(target & bp->bkend_type))
5345 continue;
5346 }
5347
5348 cnt++;
5349 }
5350 return cnt;
5351 }
5352
5353
5354 /*
5355 * StartChildProcess -- start an auxiliary process for the postmaster
5356 *
5357 * "type" determines what kind of child will be started. All child types
5358 * initially go to AuxiliaryProcessMain, which will handle common setup.
5359 *
5360 * Return value of StartChildProcess is subprocess' PID, or 0 if failed
5361 * to start subprocess.
5362 */
5363 static pid_t
5364 StartChildProcess(AuxProcType type)
5365 {
5366 pid_t pid;
5367 char *av[10];
5368 int ac = 0;
5369 char typebuf[32];
5370
5371 /*
5372 * Set up command-line arguments for subprocess
5373 */
5374 av[ac++] = "postgres";
5375
5376 #ifdef EXEC_BACKEND
5377 av[ac++] = "--forkboot";
5378 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5379 #endif
5380
5381 snprintf(typebuf, sizeof(typebuf), "-x%d", type);
5382 av[ac++] = typebuf;
5383
5384 av[ac] = NULL;
5385 Assert(ac < lengthof(av));
5386
5387 #ifdef EXEC_BACKEND
5388 pid = postmaster_forkexec(ac, av);
5389 #else /* !EXEC_BACKEND */
5390 pid = fork_process();
5391
5392 if (pid == 0) /* child */
5393 {
5394 InitPostmasterChild();
5395
5396 /* Close the postmaster's sockets */
5397 ClosePostmasterPorts(false);
5398
5399 /* Release postmaster's working memory context */
5400 MemoryContextSwitchTo(TopMemoryContext);
5401 MemoryContextDelete(PostmasterContext);
5402 PostmasterContext = NULL;
5403
5404 AuxiliaryProcessMain(ac, av);
5405 ExitPostmaster(0);
5406 }
5407 #endif /* EXEC_BACKEND */
5408
5409 if (pid < 0)
5410 {
5411 /* in parent, fork failed */
5412 int save_errno = errno;
5413
5414 errno = save_errno;
5415 switch (type)
5416 {
5417 case StartupProcess:
5418 ereport(LOG,
5419 (errmsg("could not fork startup process: %m")));
5420 break;
5421 case BgWriterProcess:
5422 ereport(LOG,
5423 (errmsg("could not fork background writer process: %m")));
5424 break;
5425 case CheckpointerProcess:
5426 ereport(LOG,
5427 (errmsg("could not fork checkpointer process: %m")));
5428 break;
5429 case WalWriterProcess:
5430 ereport(LOG,
5431 (errmsg("could not fork WAL writer process: %m")));
5432 break;
5433 case WalReceiverProcess:
5434 ereport(LOG,
5435 (errmsg("could not fork WAL receiver process: %m")));
5436 break;
5437 default:
5438 ereport(LOG,
5439 (errmsg("could not fork process: %m")));
5440 break;
5441 }
5442
5443 /*
5444 * fork failure is fatal during startup, but there's no need to choke
5445 * immediately if starting other child types fails.
5446 */
5447 if (type == StartupProcess)
5448 ExitPostmaster(1);
5449 return 0;
5450 }
5451
5452 /*
5453 * in parent, successful fork
5454 */
5455 return pid;
5456 }
5457
5458 /*
5459 * StartAutovacuumWorker
5460 * Start an autovac worker process.
5461 *
5462 * This function is here because it enters the resulting PID into the
5463 * postmaster's private backends list.
5464 *
5465 * NB -- this code very roughly matches BackendStartup.
5466 */
5467 static void
5468 StartAutovacuumWorker(void)
5469 {
5470 Backend *bn;
5471
5472 /*
5473 * If not in condition to run a process, don't try, but handle it like a
5474 * fork failure. This does not normally happen, since the signal is only
5475 * supposed to be sent by autovacuum launcher when it's OK to do it, but
5476 * we have to check to avoid race-condition problems during DB state
5477 * changes.
5478 */
5479 if (canAcceptConnections(BACKEND_TYPE_AUTOVAC) == CAC_OK)
5480 {
5481 bn = (Backend *) malloc(sizeof(Backend));
5482 if (bn)
5483 {
5484 /*
5485 * Compute the cancel key that will be assigned to this session.
5486 * We probably don't need cancel keys for autovac workers, but
5487 * we'd better have something random in the field to prevent
5488 * unfriendly people from sending cancels to them.
5489 */
5490 MyCancelKey = PostmasterRandom();
5491 bn->cancel_key = MyCancelKey;
5492
5493 /* Autovac workers are not dead_end and need a child slot */
5494 bn->dead_end = false;
5495 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5496 bn->bgworker_notify = false;
5497
5498 bn->pid = StartAutoVacWorker();
5499 if (bn->pid > 0)
5500 {
5501 bn->bkend_type = BACKEND_TYPE_AUTOVAC;
5502 dlist_push_head(&BackendList, &bn->elem);
5503 #ifdef EXEC_BACKEND
5504 ShmemBackendArrayAdd(bn);
5505 #endif
5506 /* all OK */
5507 return;
5508 }
5509
5510 /*
5511 * fork failed, fall through to report -- actual error message was
5512 * logged by StartAutoVacWorker
5513 */
5514 (void) ReleasePostmasterChildSlot(bn->child_slot);
5515 free(bn);
5516 }
5517 else
5518 ereport(LOG,
5519 (errcode(ERRCODE_OUT_OF_MEMORY),
5520 errmsg("out of memory")));
5521 }
5522
5523 /*
5524 * Report the failure to the launcher, if it's running. (If it's not, we
5525 * might not even be connected to shared memory, so don't try to call
5526 * AutoVacWorkerFailed.) Note that we also need to signal it so that it
5527 * responds to the condition, but we don't do that here, instead waiting
5528 * for ServerLoop to do it. This way we avoid a ping-pong signalling in
5529 * quick succession between the autovac launcher and postmaster in case
5530 * things get ugly.
5531 */
5532 if (AutoVacPID != 0)
5533 {
5534 AutoVacWorkerFailed();
5535 avlauncher_needs_signal = true;
5536 }
5537 }
5538
5539 /*
5540 * MaybeStartWalReceiver
5541 * Start the WAL receiver process, if not running and our state allows.
5542 *
5543 * Note: if WalReceiverPID is already nonzero, it might seem that we should
5544 * clear WalReceiverRequested. However, there's a race condition if the
5545 * walreceiver terminates and the startup process immediately requests a new
5546 * one: it's quite possible to get the signal for the request before reaping
5547 * the dead walreceiver process. Better to risk launching an extra
5548 * walreceiver than to miss launching one we need. (The walreceiver code
5549 * has logic to recognize that it should go away if not needed.)
5550 */
5551 static void
5552 MaybeStartWalReceiver(void)
5553 {
5554 if (WalReceiverPID == 0 &&
5555 (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
5556 pmState == PM_HOT_STANDBY) &&
5557 Shutdown <= SmartShutdown)
5558 {
5559 WalReceiverPID = StartWalReceiver();
5560 if (WalReceiverPID != 0)
5561 WalReceiverRequested = false;
5562 /* else leave the flag set, so we'll try again later */
5563 }
5564 }
5565
5566
5567 /*
5568 * Create the opts file
5569 */
5570 static bool
5571 CreateOptsFile(int argc, char *argv[], char *fullprogname)
5572 {
5573 FILE *fp;
5574 int i;
5575
5576 #define OPTS_FILE "postmaster.opts"
5577
5578 if ((fp = fopen(OPTS_FILE, "w")) == NULL)
5579 {
5580 elog(LOG, "could not create file \"%s\": %m", OPTS_FILE);
5581 return false;
5582 }
5583
5584 fprintf(fp, "%s", fullprogname);
5585 for (i = 1; i < argc; i++)
5586 fprintf(fp, " \"%s\"", argv[i]);
5587 fputs("\n", fp);
5588
5589 if (fclose(fp))
5590 {
5591 elog(LOG, "could not write file \"%s\": %m", OPTS_FILE);
5592 return false;
5593 }
5594
5595 return true;
5596 }
5597
5598
5599 /*
5600 * MaxLivePostmasterChildren
5601 *
5602 * This reports the number of entries needed in per-child-process arrays
5603 * (the PMChildFlags array, and if EXEC_BACKEND the ShmemBackendArray).
5604 * These arrays include regular backends, autovac workers, walsenders
5605 * and background workers, but not special children nor dead_end children.
5606 * This allows the arrays to have a fixed maximum size, to wit the same
5607 * too-many-children limit enforced by canAcceptConnections(). The exact value
5608 * isn't too critical as long as it's more than MaxBackends.
5609 */
5610 int
5611 MaxLivePostmasterChildren(void)
5612 {
5613 return 2 * (MaxConnections + autovacuum_max_workers + 1 +
5614 max_worker_processes);
5615 }
5616
5617 /*
5618 * Connect background worker to a database.
5619 */
5620 void
5621 BackgroundWorkerInitializeConnection(char *dbname, char *username)
5622 {
5623 BackgroundWorker *worker = MyBgworkerEntry;
5624
5625 /* XXX is this the right errcode? */
5626 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5627 ereport(FATAL,
5628 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5629 errmsg("database connection requirement not indicated during registration")));
5630
5631 InitPostgres(dbname, InvalidOid, username, InvalidOid, NULL);
5632
5633 /* it had better not gotten out of "init" mode yet */
5634 if (!IsInitProcessingMode())
5635 ereport(ERROR,
5636 (errmsg("invalid processing mode in background worker")));
5637 SetProcessingMode(NormalProcessing);
5638 }
5639
5640 /*
5641 * Connect background worker to a database using OIDs.
5642 */
5643 void
5644 BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid)
5645 {
5646 BackgroundWorker *worker = MyBgworkerEntry;
5647
5648 /* XXX is this the right errcode? */
5649 if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
5650 ereport(FATAL,
5651 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5652 errmsg("database connection requirement not indicated during registration")));
5653
5654 InitPostgres(NULL, dboid, NULL, useroid, NULL);
5655
5656 /* it had better not gotten out of "init" mode yet */
5657 if (!IsInitProcessingMode())
5658 ereport(ERROR,
5659 (errmsg("invalid processing mode in background worker")));
5660 SetProcessingMode(NormalProcessing);
5661 }
5662
5663 /*
5664 * Block/unblock signals in a background worker
5665 */
5666 void
5667 BackgroundWorkerBlockSignals(void)
5668 {
5669 PG_SETMASK(&BlockSig);
5670 }
5671
5672 void
5673 BackgroundWorkerUnblockSignals(void)
5674 {
5675 PG_SETMASK(&UnBlockSig);
5676 }
5677
5678 #ifdef EXEC_BACKEND
5679 static pid_t
5680 bgworker_forkexec(int shmem_slot)
5681 {
5682 char *av[10];
5683 int ac = 0;
5684 char forkav[MAXPGPATH];
5685
5686 snprintf(forkav, MAXPGPATH, "--forkbgworker=%d", shmem_slot);
5687
5688 av[ac++] = "postgres";
5689 av[ac++] = forkav;
5690 av[ac++] = NULL; /* filled in by postmaster_forkexec */
5691 av[ac] = NULL;
5692
5693 Assert(ac < lengthof(av));
5694
5695 return postmaster_forkexec(ac, av);
5696 }
5697 #endif
5698
5699 /*
5700 * Start a new bgworker.
5701 * Starting time conditions must have been checked already.
5702 *
5703 * Returns true on success, false on failure.
5704 * In either case, update the RegisteredBgWorker's state appropriately.
5705 *
5706 * This code is heavily based on autovacuum.c, q.v.
5707 */
5708 static bool
5709 do_start_bgworker(RegisteredBgWorker *rw)
5710 {
5711 pid_t worker_pid;
5712
5713 Assert(rw->rw_pid == 0);
5714
5715 /*
5716 * Allocate and assign the Backend element. Note we must do this before
5717 * forking, so that we can handle failures (out of memory or child-process
5718 * slots) cleanly.
5719 *
5720 * Treat failure as though the worker had crashed. That way, the
5721 * postmaster will wait a bit before attempting to start it again; if we
5722 * tried again right away, most likely we'd find ourselves hitting the
5723 * same resource-exhaustion condition.
5724 */
5725 if (!assign_backendlist_entry(rw))
5726 {
5727 rw->rw_crashed_at = GetCurrentTimestamp();
5728 return false;
5729 }
5730
5731 ereport(DEBUG1,
5732 (errmsg("starting background worker process \"%s\"",
5733 rw->rw_worker.bgw_name)));
5734
5735 #ifdef EXEC_BACKEND
5736 switch ((worker_pid = bgworker_forkexec(rw->rw_shmem_slot)))
5737 #else
5738 switch ((worker_pid = fork_process()))
5739 #endif
5740 {
5741 case -1:
5742 /* in postmaster, fork failed ... */
5743 ereport(LOG,
5744 (errmsg("could not fork worker process: %m")));
5745 /* undo what assign_backendlist_entry did */
5746 ReleasePostmasterChildSlot(rw->rw_child_slot);
5747 rw->rw_child_slot = 0;
5748 free(rw->rw_backend);
5749 rw->rw_backend = NULL;
5750 /* mark entry as crashed, so we'll try again later */
5751 rw->rw_crashed_at = GetCurrentTimestamp();
5752 break;
5753
5754 #ifndef EXEC_BACKEND
5755 case 0:
5756 /* in postmaster child ... */
5757 InitPostmasterChild();
5758
5759 /* Close the postmaster's sockets */
5760 ClosePostmasterPorts(false);
5761
5762 /*
5763 * Before blowing away PostmasterContext, save this bgworker's
5764 * data where it can find it.
5765 */
5766 MyBgworkerEntry = (BackgroundWorker *)
5767 MemoryContextAlloc(TopMemoryContext, sizeof(BackgroundWorker));
5768 memcpy(MyBgworkerEntry, &rw->rw_worker, sizeof(BackgroundWorker));
5769
5770 /* Release postmaster's working memory context */
5771 MemoryContextSwitchTo(TopMemoryContext);
5772 MemoryContextDelete(PostmasterContext);
5773 PostmasterContext = NULL;
5774
5775 StartBackgroundWorker();
5776
5777 exit(1); /* should not get here */
5778 break;
5779 #endif
5780 default:
5781 /* in postmaster, fork successful ... */
5782 rw->rw_pid = worker_pid;
5783 rw->rw_backend->pid = rw->rw_pid;
5784 ReportBackgroundWorkerPID(rw);
5785 /* add new worker to lists of backends */
5786 dlist_push_head(&BackendList, &rw->rw_backend->elem);
5787 #ifdef EXEC_BACKEND
5788 ShmemBackendArrayAdd(rw->rw_backend);
5789 #endif
5790 return true;
5791 }
5792
5793 return false;
5794 }
5795
5796 /*
5797 * Does the current postmaster state require starting a worker with the
5798 * specified start_time?
5799 */
5800 static bool
5801 bgworker_should_start_now(BgWorkerStartTime start_time)
5802 {
5803 switch (pmState)
5804 {
5805 case PM_NO_CHILDREN:
5806 case PM_WAIT_DEAD_END:
5807 case PM_SHUTDOWN_2:
5808 case PM_SHUTDOWN:
5809 case PM_WAIT_BACKENDS:
5810 case PM_STOP_BACKENDS:
5811 break;
5812
5813 case PM_RUN:
5814 if (start_time == BgWorkerStart_RecoveryFinished)
5815 return true;
5816 /* fall through */
5817
5818 case PM_HOT_STANDBY:
5819 if (start_time == BgWorkerStart_ConsistentState)
5820 return true;
5821 /* fall through */
5822
5823 case PM_RECOVERY:
5824 case PM_STARTUP:
5825 case PM_INIT:
5826 if (start_time == BgWorkerStart_PostmasterStart)
5827 return true;
5828 /* fall through */
5829
5830 }
5831
5832 return false;
5833 }
5834
5835 /*
5836 * Allocate the Backend struct for a connected background worker, but don't
5837 * add it to the list of backends just yet.
5838 *
5839 * On failure, return false without changing any worker state.
5840 *
5841 * Some info from the Backend is copied into the passed rw.
5842 */
5843 static bool
5844 assign_backendlist_entry(RegisteredBgWorker *rw)
5845 {
5846 Backend *bn;
5847
5848 /*
5849 * Check that database state allows another connection. Currently the
5850 * only possible failure is CAC_TOOMANY, so we just log an error message
5851 * based on that rather than checking the error code precisely.
5852 */
5853 if (canAcceptConnections(BACKEND_TYPE_BGWORKER) != CAC_OK)
5854 {
5855 ereport(LOG,
5856 (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
5857 errmsg("no slot available for new worker process")));
5858 return false;
5859 }
5860
5861 bn = malloc(sizeof(Backend));
5862 if (bn == NULL)
5863 {
5864 ereport(LOG,
5865 (errcode(ERRCODE_OUT_OF_MEMORY),
5866 errmsg("out of memory")));
5867 return false;
5868 }
5869
5870 /*
5871 * Compute the cancel key that will be assigned to this session. We
5872 * probably don't need cancel keys for background workers, but we'd better
5873 * have something random in the field to prevent unfriendly people from
5874 * sending cancels to them.
5875 */
5876 MyCancelKey = PostmasterRandom();
5877 bn->cancel_key = MyCancelKey;
5878
5879 bn->child_slot = MyPMChildSlot = AssignPostmasterChildSlot();
5880 bn->bkend_type = BACKEND_TYPE_BGWORKER;
5881 bn->dead_end = false;
5882 bn->bgworker_notify = false;
5883
5884 rw->rw_backend = bn;
5885 rw->rw_child_slot = bn->child_slot;
5886
5887 return true;
5888 }
5889
5890 /*
5891 * If the time is right, start background worker(s).
5892 *
5893 * As a side effect, the bgworker control variables are set or reset
5894 * depending on whether more workers may need to be started.
5895 *
5896 * We limit the number of workers started per call, to avoid consuming the
5897 * postmaster's attention for too long when many such requests are pending.
5898 * As long as StartWorkerNeeded is true, ServerLoop will not block and will
5899 * call this function again after dealing with any other issues.
5900 */
5901 static void
5902 maybe_start_bgworkers(void)
5903 {
5904 #define MAX_BGWORKERS_TO_LAUNCH 100
5905 int num_launched = 0;
5906 TimestampTz now = 0;
5907 slist_mutable_iter iter;
5908
5909 /*
5910 * During crash recovery, we have no need to be called until the state
5911 * transition out of recovery.
5912 */
5913 if (FatalError)
5914 {
5915 StartWorkerNeeded = false;
5916 HaveCrashedWorker = false;
5917 return;
5918 }
5919
5920 /* Don't need to be called again unless we find a reason for it below */
5921 StartWorkerNeeded = false;
5922 HaveCrashedWorker = false;
5923
5924 slist_foreach_modify(iter, &BackgroundWorkerList)
5925 {
5926 RegisteredBgWorker *rw;
5927
5928 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
5929
5930 /* ignore if already running */
5931 if (rw->rw_pid != 0)
5932 continue;
5933
5934 /* if marked for death, clean up and remove from list */
5935 if (rw->rw_terminate)
5936 {
5937 ForgetBackgroundWorker(&iter);
5938 continue;
5939 }
5940
5941 /*
5942 * If this worker has crashed previously, maybe it needs to be
5943 * restarted (unless on registration it specified it doesn't want to
5944 * be restarted at all). Check how long ago did a crash last happen.
5945 * If the last crash is too recent, don't start it right away; let it
5946 * be restarted once enough time has passed.
5947 */
5948 if (rw->rw_crashed_at != 0)
5949 {
5950 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
5951 {
5952 int notify_pid;
5953
5954 notify_pid = rw->rw_worker.bgw_notify_pid;
5955
5956 ForgetBackgroundWorker(&iter);
5957
5958 /* Report worker is gone now. */
5959 if (notify_pid != 0)
5960 kill(notify_pid, SIGUSR1);
5961
5962 continue;
5963 }
5964
5965 /* read system time only when needed */
5966 if (now == 0)
5967 now = GetCurrentTimestamp();
5968
5969 if (!TimestampDifferenceExceeds(rw->rw_crashed_at, now,
5970 rw->rw_worker.bgw_restart_time * 1000))
5971 {
5972 /* Set flag to remember that we have workers to start later */
5973 HaveCrashedWorker = true;
5974 continue;
5975 }
5976 }
5977
5978 if (bgworker_should_start_now(rw->rw_worker.bgw_start_time))
5979 {
5980 /* reset crash time before trying to start worker */
5981 rw->rw_crashed_at = 0;
5982
5983 /*
5984 * Try to start the worker.
5985 *
5986 * On failure, give up processing workers for now, but set
5987 * StartWorkerNeeded so we'll come back here on the next iteration
5988 * of ServerLoop to try again. (We don't want to wait, because
5989 * there might be additional ready-to-run workers.) We could set
5990 * HaveCrashedWorker as well, since this worker is now marked
5991 * crashed, but there's no need because the next run of this
5992 * function will do that.
5993 */
5994 if (!do_start_bgworker(rw))
5995 {
5996 StartWorkerNeeded = true;
5997 return;
5998 }
5999
6000 /*
6001 * If we've launched as many workers as allowed, quit, but have
6002 * ServerLoop call us again to look for additional ready-to-run
6003 * workers. There might not be any, but we'll find out the next
6004 * time we run.
6005 */
6006 if (++num_launched >= MAX_BGWORKERS_TO_LAUNCH)
6007 {
6008 StartWorkerNeeded = true;
6009 return;
6010 }
6011 }
6012 }
6013 }
6014
6015 /*
6016 * When a backend asks to be notified about worker state changes, we
6017 * set a flag in its backend entry. The background worker machinery needs
6018 * to know when such backends exit.
6019 */
6020 bool
6021 PostmasterMarkPIDForWorkerNotify(int pid)
6022 {
6023 dlist_iter iter;
6024 Backend *bp;
6025
6026 dlist_foreach(iter, &BackendList)
6027 {
6028 bp = dlist_container(Backend, elem, iter.cur);
6029 if (bp->pid == pid)
6030 {
6031 bp->bgworker_notify = true;
6032 return true;
6033 }
6034 }
6035 return false;
6036 }
6037
6038 #ifdef EXEC_BACKEND
6039
6040 /*
6041 * The following need to be available to the save/restore_backend_variables
6042 * functions. They are marked NON_EXEC_STATIC in their home modules.
6043 */
6044 extern slock_t *ShmemLock;
6045 extern slock_t *ProcStructLock;
6046 extern PGPROC *AuxiliaryProcs;
6047 extern PMSignalData *PMSignalState;
6048 extern pgsocket pgStatSock;
6049 extern pg_time_t first_syslogger_file_time;
6050
6051 #ifndef WIN32
6052 #define write_inheritable_socket(dest, src, childpid) ((*(dest) = (src)), true)
6053 #define read_inheritable_socket(dest, src) (*(dest) = *(src))
6054 #else
6055 static bool write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE child);
6056 static bool write_inheritable_socket(InheritableSocket *dest, SOCKET src,
6057 pid_t childPid);
6058 static void read_inheritable_socket(SOCKET *dest, InheritableSocket *src);
6059 #endif
6060
6061
6062 /* Save critical backend variables into the BackendParameters struct */
6063 #ifndef WIN32
6064 static bool
6065 save_backend_variables(BackendParameters *param, Port *port)
6066 #else
6067 static bool
6068 save_backend_variables(BackendParameters *param, Port *port,
6069 HANDLE childProcess, pid_t childPid)
6070 #endif
6071 {
6072 memcpy(¶m->port, port, sizeof(Port));
6073 if (!write_inheritable_socket(¶m->portsocket, port->sock, childPid))
6074 return false;
6075
6076 strlcpy(param->DataDir, DataDir, MAXPGPATH);
6077
6078 memcpy(¶m->ListenSocket, &ListenSocket, sizeof(ListenSocket));
6079
6080 param->MyCancelKey = MyCancelKey;
6081 param->MyPMChildSlot = MyPMChildSlot;
6082
6083 #ifdef WIN32
6084 param->ShmemProtectiveRegion = ShmemProtectiveRegion;
6085 #endif
6086 param->UsedShmemSegID = UsedShmemSegID;
6087 param->UsedShmemSegAddr = UsedShmemSegAddr;
6088
6089 param->ShmemLock = ShmemLock;
6090 param->ShmemVariableCache = ShmemVariableCache;
6091 param->ShmemBackendArray = ShmemBackendArray;
6092
6093 #ifndef HAVE_SPINLOCKS
6094 param->SpinlockSemaArray = SpinlockSemaArray;
6095 #endif
6096 param->NamedLWLockTrancheRequests = NamedLWLockTrancheRequests;
6097 param->NamedLWLockTrancheArray = NamedLWLockTrancheArray;
6098 param->MainLWLockArray = MainLWLockArray;
6099 param->ProcStructLock = ProcStructLock;
6100 param->ProcGlobal = ProcGlobal;
6101 param->AuxiliaryProcs = AuxiliaryProcs;
6102 param->PreparedXactProcs = PreparedXactProcs;
6103 param->PMSignalState = PMSignalState;
6104 if (!write_inheritable_socket(¶m->pgStatSock, pgStatSock, childPid))
6105 return false;
6106
6107 param->PostmasterPid = PostmasterPid;
6108 param->PgStartTime = PgStartTime;
6109 param->PgReloadTime = PgReloadTime;
6110 param->first_syslogger_file_time = first_syslogger_file_time;
6111
6112 param->redirection_done = redirection_done;
6113 param->IsBinaryUpgrade = IsBinaryUpgrade;
6114 param->max_safe_fds = max_safe_fds;
6115
6116 param->MaxBackends = MaxBackends;
6117
6118 #ifdef WIN32
6119 param->PostmasterHandle = PostmasterHandle;
6120 if (!write_duplicated_handle(¶m->initial_signal_pipe,
6121 pgwin32_create_signal_listener(childPid),
6122 childProcess))
6123 return false;
6124 #else
6125 memcpy(¶m->postmaster_alive_fds, &postmaster_alive_fds,
6126 sizeof(postmaster_alive_fds));
6127 #endif
6128
6129 memcpy(¶m->syslogPipe, &syslogPipe, sizeof(syslogPipe));
6130
6131 strlcpy(param->my_exec_path, my_exec_path, MAXPGPATH);
6132
6133 strlcpy(param->pkglib_path, pkglib_path, MAXPGPATH);
6134
6135 strlcpy(param->ExtraOptions, ExtraOptions, MAXPGPATH);
6136
6137 return true;
6138 }
6139
6140
6141 #ifdef WIN32
6142 /*
6143 * Duplicate a handle for usage in a child process, and write the child
6144 * process instance of the handle to the parameter file.
6145 */
6146 static bool
6147 write_duplicated_handle(HANDLE *dest, HANDLE src, HANDLE childProcess)
6148 {
6149 HANDLE hChild = INVALID_HANDLE_VALUE;
6150
6151 if (!DuplicateHandle(GetCurrentProcess(),
6152 src,
6153 childProcess,
6154 &hChild,
6155 0,
6156 TRUE,
6157 DUPLICATE_CLOSE_SOURCE | DUPLICATE_SAME_ACCESS))
6158 {
6159 ereport(LOG,
6160 (errmsg_internal("could not duplicate handle to be written to backend parameter file: error code %lu",
6161 GetLastError())));
6162 return false;
6163 }
6164
6165 *dest = hChild;
6166 return true;
6167 }
6168
6169 /*
6170 * Duplicate a socket for usage in a child process, and write the resulting
6171 * structure to the parameter file.
6172 * This is required because a number of LSPs (Layered Service Providers) very
6173 * common on Windows (antivirus, firewalls, download managers etc) break
6174 * straight socket inheritance.
6175 */
6176 static bool
6177 write_inheritable_socket(InheritableSocket *dest, SOCKET src, pid_t childpid)
6178 {
6179 dest->origsocket = src;
6180 if (src != 0 && src != PGINVALID_SOCKET)
6181 {
6182 /* Actual socket */
6183 if (WSADuplicateSocket(src, childpid, &dest->wsainfo) != 0)
6184 {
6185 ereport(LOG,
6186 (errmsg("could not duplicate socket %d for use in backend: error code %d",
6187 (int) src, WSAGetLastError())));
6188 return false;
6189 }
6190 }
6191 return true;
6192 }
6193
6194 /*
6195 * Read a duplicate socket structure back, and get the socket descriptor.
6196 */
6197 static void
6198 read_inheritable_socket(SOCKET *dest, InheritableSocket *src)
6199 {
6200 SOCKET s;
6201
6202 if (src->origsocket == PGINVALID_SOCKET || src->origsocket == 0)
6203 {
6204 /* Not a real socket! */
6205 *dest = src->origsocket;
6206 }
6207 else
6208 {
6209 /* Actual socket, so create from structure */
6210 s = WSASocket(FROM_PROTOCOL_INFO,
6211 FROM_PROTOCOL_INFO,
6212 FROM_PROTOCOL_INFO,
6213 &src->wsainfo,
6214 0,
6215 0);
6216 if (s == INVALID_SOCKET)
6217 {
6218 write_stderr("could not create inherited socket: error code %d\n",
6219 WSAGetLastError());
6220 exit(1);
6221 }
6222 *dest = s;
6223
6224 /*
6225 * To make sure we don't get two references to the same socket, close
6226 * the original one. (This would happen when inheritance actually
6227 * works..
6228 */
6229 closesocket(src->origsocket);
6230 }
6231 }
6232 #endif
6233
6234 static void
6235 read_backend_variables(char *id, Port *port)
6236 {
6237 BackendParameters param;
6238
6239 #ifndef WIN32
6240 /* Non-win32 implementation reads from file */
6241 FILE *fp;
6242
6243 /* Open file */
6244 fp = AllocateFile(id, PG_BINARY_R);
6245 if (!fp)
6246 {
6247 write_stderr("could not open backend variables file \"%s\": %s\n",
6248 id, strerror(errno));
6249 exit(1);
6250 }
6251
6252 if (fread(¶m, sizeof(param), 1, fp) != 1)
6253 {
6254 write_stderr("could not read from backend variables file \"%s\": %s\n",
6255 id, strerror(errno));
6256 exit(1);
6257 }
6258
6259 /* Release file */
6260 FreeFile(fp);
6261 if (unlink(id) != 0)
6262 {
6263 write_stderr("could not remove file \"%s\": %s\n",
6264 id, strerror(errno));
6265 exit(1);
6266 }
6267 #else
6268 /* Win32 version uses mapped file */
6269 HANDLE paramHandle;
6270 BackendParameters *paramp;
6271
6272 #ifdef _WIN64
6273 paramHandle = (HANDLE) _atoi64(id);
6274 #else
6275 paramHandle = (HANDLE) atol(id);
6276 #endif
6277 paramp = MapViewOfFile(paramHandle, FILE_MAP_READ, 0, 0, 0);
6278 if (!paramp)
6279 {
6280 write_stderr("could not map view of backend variables: error code %lu\n",
6281 GetLastError());
6282 exit(1);
6283 }
6284
6285 memcpy(¶m, paramp, sizeof(BackendParameters));
6286
6287 if (!UnmapViewOfFile(paramp))
6288 {
6289 write_stderr("could not unmap view of backend variables: error code %lu\n",
6290 GetLastError());
6291 exit(1);
6292 }
6293
6294 if (!CloseHandle(paramHandle))
6295 {
6296 write_stderr("could not close handle to backend parameter variables: error code %lu\n",
6297 GetLastError());
6298 exit(1);
6299 }
6300 #endif
6301
6302 restore_backend_variables(¶m, port);
6303 }
6304
6305 /* Restore critical backend variables from the BackendParameters struct */
6306 static void
6307 restore_backend_variables(BackendParameters *param, Port *port)
6308 {
6309 memcpy(port, ¶m->port, sizeof(Port));
6310 read_inheritable_socket(&port->sock, ¶m->portsocket);
6311
6312 SetDataDir(param->DataDir);
6313
6314 memcpy(&ListenSocket, ¶m->ListenSocket, sizeof(ListenSocket));
6315
6316 MyCancelKey = param->MyCancelKey;
6317 MyPMChildSlot = param->MyPMChildSlot;
6318
6319 #ifdef WIN32
6320 ShmemProtectiveRegion = param->ShmemProtectiveRegion;
6321 #endif
6322 UsedShmemSegID = param->UsedShmemSegID;
6323 UsedShmemSegAddr = param->UsedShmemSegAddr;
6324
6325 ShmemLock = param->ShmemLock;
6326 ShmemVariableCache = param->ShmemVariableCache;
6327 ShmemBackendArray = param->ShmemBackendArray;
6328
6329 #ifndef HAVE_SPINLOCKS
6330 SpinlockSemaArray = param->SpinlockSemaArray;
6331 #endif
6332 NamedLWLockTrancheRequests = param->NamedLWLockTrancheRequests;
6333 NamedLWLockTrancheArray = param->NamedLWLockTrancheArray;
6334 MainLWLockArray = param->MainLWLockArray;
6335 ProcStructLock = param->ProcStructLock;
6336 ProcGlobal = param->ProcGlobal;
6337 AuxiliaryProcs = param->AuxiliaryProcs;
6338 PreparedXactProcs = param->PreparedXactProcs;
6339 PMSignalState = param->PMSignalState;
6340 read_inheritable_socket(&pgStatSock, ¶m->pgStatSock);
6341
6342 PostmasterPid = param->PostmasterPid;
6343 PgStartTime = param->PgStartTime;
6344 PgReloadTime = param->PgReloadTime;
6345 first_syslogger_file_time = param->first_syslogger_file_time;
6346
6347 redirection_done = param->redirection_done;
6348 IsBinaryUpgrade = param->IsBinaryUpgrade;
6349 max_safe_fds = param->max_safe_fds;
6350
6351 MaxBackends = param->MaxBackends;
6352
6353 #ifdef WIN32
6354 PostmasterHandle = param->PostmasterHandle;
6355 pgwin32_initial_signal_pipe = param->initial_signal_pipe;
6356 #else
6357 memcpy(&postmaster_alive_fds, ¶m->postmaster_alive_fds,
6358 sizeof(postmaster_alive_fds));
6359 #endif
6360
6361 memcpy(&syslogPipe, ¶m->syslogPipe, sizeof(syslogPipe));
6362
6363 strlcpy(my_exec_path, param->my_exec_path, MAXPGPATH);
6364
6365 strlcpy(pkglib_path, param->pkglib_path, MAXPGPATH);
6366
6367 strlcpy(ExtraOptions, param->ExtraOptions, MAXPGPATH);
6368 }
6369
6370
6371 Size
6372 ShmemBackendArraySize(void)
6373 {
6374 return mul_size(MaxLivePostmasterChildren(), sizeof(Backend));
6375 }
6376
6377 void
6378 ShmemBackendArrayAllocation(void)
6379 {
6380 Size size = ShmemBackendArraySize();
6381
6382 ShmemBackendArray = (Backend *) ShmemAlloc(size);
6383 /* Mark all slots as empty */
6384 memset(ShmemBackendArray, 0, size);
6385 }
6386
6387 static void
6388 ShmemBackendArrayAdd(Backend *bn)
6389 {
6390 /* The array slot corresponding to my PMChildSlot should be free */
6391 int i = bn->child_slot - 1;
6392
6393 Assert(ShmemBackendArray[i].pid == 0);
6394 ShmemBackendArray[i] = *bn;
6395 }
6396
6397 static void
6398 ShmemBackendArrayRemove(Backend *bn)
6399 {
6400 int i = bn->child_slot - 1;
6401
6402 Assert(ShmemBackendArray[i].pid == bn->pid);
6403 /* Mark the slot as empty */
6404 ShmemBackendArray[i].pid = 0;
6405 }
6406 #endif /* EXEC_BACKEND */
6407
6408
6409 #ifdef WIN32
6410
6411 /*
6412 * Subset implementation of waitpid() for Windows. We assume pid is -1
6413 * (that is, check all child processes) and options is WNOHANG (don't wait).
6414 */
6415 static pid_t
6416 waitpid(pid_t pid, int *exitstatus, int options)
6417 {
6418 DWORD dwd;
6419 ULONG_PTR key;
6420 OVERLAPPED *ovl;
6421
6422 /*
6423 * Check if there are any dead children. If there are, return the pid of
6424 * the first one that died.
6425 */
6426 if (GetQueuedCompletionStatus(win32ChildQueue, &dwd, &key, &ovl, 0))
6427 {
6428 *exitstatus = (int) key;
6429 return dwd;
6430 }
6431
6432 return -1;
6433 }
6434
6435 /*
6436 * Note! Code below executes on a thread pool! All operations must
6437 * be thread safe! Note that elog() and friends must *not* be used.
6438 */
6439 static void WINAPI
6440 pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired)
6441 {
6442 win32_deadchild_waitinfo *childinfo = (win32_deadchild_waitinfo *) lpParameter;
6443 DWORD exitcode;
6444
6445 if (TimerOrWaitFired)
6446 return; /* timeout. Should never happen, since we use
6447 * INFINITE as timeout value. */
6448
6449 /*
6450 * Remove handle from wait - required even though it's set to wait only
6451 * once
6452 */
6453 UnregisterWaitEx(childinfo->waitHandle, NULL);
6454
6455 if (!GetExitCodeProcess(childinfo->procHandle, &exitcode))
6456 {
6457 /*
6458 * Should never happen. Inform user and set a fixed exitcode.
6459 */
6460 write_stderr("could not read exit code for process\n");
6461 exitcode = 255;
6462 }
6463
6464 if (!PostQueuedCompletionStatus(win32ChildQueue, childinfo->procId, (ULONG_PTR) exitcode, NULL))
6465 write_stderr("could not post child completion status\n");
6466
6467 /*
6468 * Handle is per-process, so we close it here instead of in the
6469 * originating thread
6470 */
6471 CloseHandle(childinfo->procHandle);
6472
6473 /*
6474 * Free struct that was allocated before the call to
6475 * RegisterWaitForSingleObject()
6476 */
6477 free(childinfo);
6478
6479 /* Queue SIGCHLD signal */
6480 pg_queue_signal(SIGCHLD);
6481 }
6482 #endif /* WIN32 */
6483
6484 /*
6485 * Initialize one and only handle for monitoring postmaster death.
6486 *
6487 * Called once in the postmaster, so that child processes can subsequently
6488 * monitor if their parent is dead.
6489 */
6490 static void
6491 InitPostmasterDeathWatchHandle(void)
6492 {
6493 #ifndef WIN32
6494
6495 /*
6496 * Create a pipe. Postmaster holds the write end of the pipe open
6497 * (POSTMASTER_FD_OWN), and children hold the read end. Children can pass
6498 * the read file descriptor to select() to wake up in case postmaster
6499 * dies, or check for postmaster death with a (read() == 0). Children must
6500 * close the write end as soon as possible after forking, because EOF
6501 * won't be signaled in the read end until all processes have closed the
6502 * write fd. That is taken care of in ClosePostmasterPorts().
6503 */
6504 Assert(MyProcPid == PostmasterPid);
6505 if (pipe(postmaster_alive_fds) < 0)
6506 ereport(FATAL,
6507 (errcode_for_file_access(),
6508 errmsg_internal("could not create pipe to monitor postmaster death: %m")));
6509
6510 /*
6511 * Set O_NONBLOCK to allow testing for the fd's presence with a read()
6512 * call.
6513 */
6514 if (fcntl(postmaster_alive_fds[POSTMASTER_FD_WATCH], F_SETFL, O_NONBLOCK) == -1)
6515 ereport(FATAL,
6516 (errcode_for_socket_access(),
6517 errmsg_internal("could not set postmaster death monitoring pipe to nonblocking mode: %m")));
6518 #else
6519
6520 /*
6521 * On Windows, we use a process handle for the same purpose.
6522 */
6523 if (DuplicateHandle(GetCurrentProcess(),
6524 GetCurrentProcess(),
6525 GetCurrentProcess(),
6526 &PostmasterHandle,
6527 0,
6528 TRUE,
6529 DUPLICATE_SAME_ACCESS) == 0)
6530 ereport(FATAL,
6531 (errmsg_internal("could not duplicate postmaster handle: error code %lu",
6532 GetLastError())));
6533 #endif /* WIN32 */
6534 }
6535