Coverage Report

Created: 2025-09-27 06:52

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/src/postgres/src/backend/postmaster/postmaster.c
Line
Count
Source
1
/*-------------------------------------------------------------------------
2
 *
3
 * postmaster.c
4
 *    This program acts as a clearing house for requests to the
5
 *    POSTGRES system.  Frontend programs connect to the Postmaster,
6
 *    and postmaster forks a new backend process to handle the
7
 *    connection.
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) 1996-2025, PostgreSQL Global Development Group
36
 * Portions Copyright (c) 1994, Regents of the University of California
37
 *
38
 *
39
 * IDENTIFICATION
40
 *    src/backend/postmaster/postmaster.c
41
 *
42
 * NOTES
43
 *
44
 * Initialization:
45
 *    The Postmaster sets up shared memory data structures
46
 *    for the backends.
47
 *
48
 * Synchronization:
49
 *    The Postmaster shares memory with the backends but should avoid
50
 *    touching shared memory, so as not to become stuck if a crashing
51
 *    backend screws up locks or shared memory.  Likewise, the Postmaster
52
 *    should never block on messages from frontend clients.
53
 *
54
 * Garbage Collection:
55
 *    The Postmaster cleans up after backends if they have an emergency
56
 *    exit and/or core dump.
57
 *
58
 * Error Reporting:
59
 *    Use write_stderr() only for reporting "interactive" errors
60
 *    (essentially, bogus arguments on the command line).  Once the
61
 *    postmaster is launched, use ereport().
62
 *
63
 *-------------------------------------------------------------------------
64
 */
65
66
#include "postgres.h"
67
68
#include <unistd.h>
69
#include <signal.h>
70
#include <time.h>
71
#include <sys/wait.h>
72
#include <ctype.h>
73
#include <sys/stat.h>
74
#include <sys/socket.h>
75
#include <fcntl.h>
76
#include <sys/param.h>
77
#include <netdb.h>
78
#include <limits.h>
79
80
#ifdef USE_BONJOUR
81
#include <dns_sd.h>
82
#endif
83
84
#ifdef USE_SYSTEMD
85
#include <systemd/sd-daemon.h>
86
#endif
87
88
#ifdef HAVE_PTHREAD_IS_THREADED_NP
89
#include <pthread.h>
90
#endif
91
92
#include "access/xlog.h"
93
#include "access/xlog_internal.h"
94
#include "access/xlogrecovery.h"
95
#include "common/file_perm.h"
96
#include "common/pg_prng.h"
97
#include "lib/ilist.h"
98
#include "libpq/libpq.h"
99
#include "libpq/pqsignal.h"
100
#include "pg_getopt.h"
101
#include "pgstat.h"
102
#include "port/pg_bswap.h"
103
#include "postmaster/autovacuum.h"
104
#include "postmaster/bgworker_internals.h"
105
#include "postmaster/pgarch.h"
106
#include "postmaster/postmaster.h"
107
#include "postmaster/syslogger.h"
108
#include "postmaster/walsummarizer.h"
109
#include "replication/logicallauncher.h"
110
#include "replication/slotsync.h"
111
#include "replication/walsender.h"
112
#include "storage/aio_subsys.h"
113
#include "storage/fd.h"
114
#include "storage/io_worker.h"
115
#include "storage/ipc.h"
116
#include "storage/pmsignal.h"
117
#include "storage/proc.h"
118
#include "tcop/backend_startup.h"
119
#include "tcop/tcopprot.h"
120
#include "utils/datetime.h"
121
#include "utils/memutils.h"
122
#include "utils/pidfile.h"
123
#include "utils/timestamp.h"
124
#include "utils/varlena.h"
125
126
#ifdef EXEC_BACKEND
127
#include "common/file_utils.h"
128
#include "storage/pg_shmem.h"
129
#endif
130
131
132
/*
133
 * CountChildren and SignalChildren take a bitmask argument to represent
134
 * BackendTypes to count or signal.  Define a separate type and functions to
135
 * work with the bitmasks, to avoid accidentally passing a plain BackendType
136
 * in place of a bitmask or vice versa.
137
 */
138
typedef struct
139
{
140
  uint32    mask;
141
} BackendTypeMask;
142
143
StaticAssertDecl(BACKEND_NUM_TYPES < 32, "too many backend types for uint32");
144
145
static const BackendTypeMask BTYPE_MASK_ALL = {(1 << BACKEND_NUM_TYPES) - 1};
146
static const BackendTypeMask BTYPE_MASK_NONE = {0};
147
148
static inline BackendTypeMask
149
btmask(BackendType t)
150
0
{
151
0
  BackendTypeMask mask = {.mask = 1 << t};
152
153
0
  return mask;
154
0
}
155
156
static inline BackendTypeMask
157
btmask_add_n(BackendTypeMask mask, int nargs, BackendType *t)
158
0
{
159
0
  for (int i = 0; i < nargs; i++)
160
0
    mask.mask |= 1 << t[i];
161
0
  return mask;
162
0
}
163
164
#define btmask_add(mask, ...) \
165
0
  btmask_add_n(mask, \
166
0
    lengthof(((BackendType[]){__VA_ARGS__})), \
167
0
    (BackendType[]){__VA_ARGS__} \
168
0
  )
169
170
static inline BackendTypeMask
171
btmask_del(BackendTypeMask mask, BackendType t)
172
0
{
173
0
  mask.mask &= ~(1 << t);
174
0
  return mask;
175
0
}
176
177
static inline BackendTypeMask
178
btmask_all_except_n(int nargs, BackendType *t)
179
0
{
180
0
  BackendTypeMask mask = BTYPE_MASK_ALL;
181
182
0
  for (int i = 0; i < nargs; i++)
183
0
    mask = btmask_del(mask, t[i]);
184
0
  return mask;
185
0
}
186
187
#define btmask_all_except(...) \
188
0
  btmask_all_except_n( \
189
0
    lengthof(((BackendType[]){__VA_ARGS__})), \
190
0
    (BackendType[]){__VA_ARGS__} \
191
0
  )
192
193
static inline bool
194
btmask_contains(BackendTypeMask mask, BackendType t)
195
0
{
196
0
  return (mask.mask & (1 << t)) != 0;
197
0
}
198
199
200
BackgroundWorker *MyBgworkerEntry = NULL;
201
202
/* The socket number we are listening for connections on */
203
int     PostPortNumber = DEF_PGPORT;
204
205
/* The directory names for Unix socket(s) */
206
char     *Unix_socket_directories;
207
208
/* The TCP listen address(es) */
209
char     *ListenAddresses;
210
211
/*
212
 * SuperuserReservedConnections is the number of backends reserved for
213
 * superuser use, and ReservedConnections is the number of backends reserved
214
 * for use by roles with privileges of the pg_use_reserved_connections
215
 * predefined role.  These are taken out of the pool of MaxConnections backend
216
 * slots, so the number of backend slots available for roles that are neither
217
 * superuser nor have privileges of pg_use_reserved_connections is
218
 * (MaxConnections - SuperuserReservedConnections - ReservedConnections).
219
 *
220
 * If the number of remaining slots is less than or equal to
221
 * SuperuserReservedConnections, only superusers can make new connections.  If
222
 * the number of remaining slots is greater than SuperuserReservedConnections
223
 * but less than or equal to
224
 * (SuperuserReservedConnections + ReservedConnections), only superusers and
225
 * roles with privileges of pg_use_reserved_connections can make new
226
 * connections.  Note that pre-existing superuser and
227
 * pg_use_reserved_connections connections don't count against the limits.
228
 */
229
int     SuperuserReservedConnections;
230
int     ReservedConnections;
231
232
/* The socket(s) we're listening to. */
233
0
#define MAXLISTEN 64
234
static int  NumListenSockets = 0;
235
static pgsocket *ListenSockets = NULL;
236
237
/* still more option variables */
238
bool    EnableSSL = false;
239
240
int     PreAuthDelay = 0;
241
int     AuthenticationTimeout = 60;
242
243
bool    log_hostname;   /* for ps display and logging */
244
245
bool    enable_bonjour = false;
246
char     *bonjour_name;
247
bool    restart_after_crash = true;
248
bool    remove_temp_files_after_crash = true;
249
250
/*
251
 * When terminating child processes after fatal errors, like a crash of a
252
 * child process, we normally send SIGQUIT -- and most other comments in this
253
 * file are written on the assumption that we do -- but developers might
254
 * prefer to use SIGABRT to collect per-child core dumps.
255
 */
256
bool    send_abort_for_crash = false;
257
bool    send_abort_for_kill = false;
258
259
/* special child processes; NULL when not running */
260
static PMChild *StartupPMChild = NULL,
261
       *BgWriterPMChild = NULL,
262
       *CheckpointerPMChild = NULL,
263
       *WalWriterPMChild = NULL,
264
       *WalReceiverPMChild = NULL,
265
       *WalSummarizerPMChild = NULL,
266
       *AutoVacLauncherPMChild = NULL,
267
       *PgArchPMChild = NULL,
268
       *SysLoggerPMChild = NULL,
269
       *SlotSyncWorkerPMChild = NULL;
270
271
/* Startup process's status */
272
typedef enum
273
{
274
  STARTUP_NOT_RUNNING,
275
  STARTUP_RUNNING,
276
  STARTUP_SIGNALED,     /* we sent it a SIGQUIT or SIGKILL */
277
  STARTUP_CRASHED,
278
} StartupStatusEnum;
279
280
static StartupStatusEnum StartupStatus = STARTUP_NOT_RUNNING;
281
282
/* Startup/shutdown state */
283
0
#define     NoShutdown    0
284
0
#define     SmartShutdown 1
285
#define     FastShutdown  2
286
0
#define     ImmediateShutdown 3
287
288
static int  Shutdown = NoShutdown;
289
290
static bool FatalError = false; /* T if recovering from backend crash */
291
292
/*
293
 * We use a simple state machine to control startup, shutdown, and
294
 * crash recovery (which is rather like shutdown followed by startup).
295
 *
296
 * After doing all the postmaster initialization work, we enter PM_STARTUP
297
 * state and the startup process is launched. The startup process begins by
298
 * reading the control file and other preliminary initialization steps.
299
 * In a normal startup, or after crash recovery, the startup process exits
300
 * with exit code 0 and we switch to PM_RUN state.  However, archive recovery
301
 * is handled specially since it takes much longer and we would like to support
302
 * hot standby during archive recovery.
303
 *
304
 * When the startup process is ready to start archive recovery, it signals the
305
 * postmaster, and we switch to PM_RECOVERY state. The background writer and
306
 * checkpointer are launched, while the startup process continues applying WAL.
307
 * If Hot Standby is enabled, then, after reaching a consistent point in WAL
308
 * redo, startup process signals us again, and we switch to PM_HOT_STANDBY
309
 * state and begin accepting connections to perform read-only queries.  When
310
 * archive recovery is finished, the startup process exits with exit code 0
311
 * and we switch to PM_RUN state.
312
 *
313
 * Normal child backends can only be launched when we are in PM_RUN or
314
 * PM_HOT_STANDBY state.  (connsAllowed can also restrict launching.)
315
 * In other states we handle connection requests by launching "dead-end"
316
 * child processes, which will simply send the client an error message and
317
 * quit.  (We track these in the ActiveChildList so that we can know when they
318
 * are all gone; this is important because they're still connected to shared
319
 * memory, and would interfere with an attempt to destroy the shmem segment,
320
 * possibly leading to SHMALL failure when we try to make a new one.)
321
 * In PM_WAIT_DEAD_END state we are waiting for all the dead-end children
322
 * to drain out of the system, and therefore stop accepting connection
323
 * requests at all until the last existing child has quit (which hopefully
324
 * will not be very long).
325
 *
326
 * Notice that this state variable does not distinguish *why* we entered
327
 * states later than PM_RUN --- Shutdown and FatalError must be consulted
328
 * to find that out.  FatalError is never true in PM_RECOVERY, PM_HOT_STANDBY,
329
 * or PM_RUN states, nor in PM_WAIT_XLOG_SHUTDOWN states (because we don't
330
 * enter those states when trying to recover from a crash).  It can be true in
331
 * PM_STARTUP state, because we don't clear it until we've successfully
332
 * started WAL redo.
333
 */
334
typedef enum
335
{
336
  PM_INIT,          /* postmaster starting */
337
  PM_STARTUP,         /* waiting for startup subprocess */
338
  PM_RECOVERY,        /* in archive recovery mode */
339
  PM_HOT_STANDBY,       /* in hot standby mode */
340
  PM_RUN,           /* normal "database is alive" state */
341
  PM_STOP_BACKENDS,     /* need to stop remaining backends */
342
  PM_WAIT_BACKENDS,     /* waiting for live backends to exit */
343
  PM_WAIT_XLOG_SHUTDOWN,    /* waiting for checkpointer to do shutdown
344
                 * ckpt */
345
  PM_WAIT_XLOG_ARCHIVAL,    /* waiting for archiver and walsenders to
346
                 * finish */
347
  PM_WAIT_IO_WORKERS,     /* waiting for io workers to exit */
348
  PM_WAIT_CHECKPOINTER,   /* waiting for checkpointer to shut down */
349
  PM_WAIT_DEAD_END,     /* waiting for dead-end children to exit */
350
  PM_NO_CHILDREN,       /* all important children have exited */
351
} PMState;
352
353
static PMState pmState = PM_INIT;
354
355
/*
356
 * While performing a "smart shutdown", we restrict new connections but stay
357
 * in PM_RUN or PM_HOT_STANDBY state until all the client backends are gone.
358
 * connsAllowed is a sub-state indicator showing the active restriction.
359
 * It is of no interest unless pmState is PM_RUN or PM_HOT_STANDBY.
360
 */
361
static bool connsAllowed = true;
362
363
/* Start time of SIGKILL timeout during immediate shutdown or child crash */
364
/* Zero means timeout is not running */
365
static time_t AbortStartTime = 0;
366
367
/* Length of said timeout */
368
0
#define SIGKILL_CHILDREN_AFTER_SECS   5
369
370
static bool ReachedNormalRunning = false; /* T if we've reached PM_RUN */
371
372
bool    ClientAuthInProgress = false; /* T during new-client
373
                       * authentication */
374
375
bool    redirection_done = false; /* stderr redirected for syslogger? */
376
377
/* received START_AUTOVAC_LAUNCHER signal */
378
static bool start_autovac_launcher = false;
379
380
/* the launcher needs to be signaled to communicate some condition */
381
static bool avlauncher_needs_signal = false;
382
383
/* received START_WALRECEIVER signal */
384
static bool WalReceiverRequested = false;
385
386
/* set when there's a worker that needs to be started up */
387
static bool StartWorkerNeeded = true;
388
static bool HaveCrashedWorker = false;
389
390
/* set when signals arrive */
391
static volatile sig_atomic_t pending_pm_pmsignal;
392
static volatile sig_atomic_t pending_pm_child_exit;
393
static volatile sig_atomic_t pending_pm_reload_request;
394
static volatile sig_atomic_t pending_pm_shutdown_request;
395
static volatile sig_atomic_t pending_pm_fast_shutdown_request;
396
static volatile sig_atomic_t pending_pm_immediate_shutdown_request;
397
398
/* event multiplexing object */
399
static WaitEventSet *pm_wait_set;
400
401
#ifdef USE_SSL
402
/* Set when and if SSL has been initialized properly */
403
bool    LoadedSSL = false;
404
#endif
405
406
#ifdef USE_BONJOUR
407
static DNSServiceRef bonjour_sdref = NULL;
408
#endif
409
410
/* State for IO worker management. */
411
static int  io_worker_count = 0;
412
static PMChild *io_worker_children[MAX_IO_WORKERS];
413
414
/*
415
 * postmaster.c - function prototypes
416
 */
417
static void CloseServerPorts(int status, Datum arg);
418
static void unlink_external_pid_file(int status, Datum arg);
419
static void getInstallationPaths(const char *argv0);
420
static void checkControlFile(void);
421
static void handle_pm_pmsignal_signal(SIGNAL_ARGS);
422
static void handle_pm_child_exit_signal(SIGNAL_ARGS);
423
static void handle_pm_reload_request_signal(SIGNAL_ARGS);
424
static void handle_pm_shutdown_request_signal(SIGNAL_ARGS);
425
static void process_pm_pmsignal(void);
426
static void process_pm_child_exit(void);
427
static void process_pm_reload_request(void);
428
static void process_pm_shutdown_request(void);
429
static void dummy_handler(SIGNAL_ARGS);
430
static void CleanupBackend(PMChild *bp, int exitstatus);
431
static void HandleChildCrash(int pid, int exitstatus, const char *procname);
432
static void LogChildExit(int lev, const char *procname,
433
             int pid, int exitstatus);
434
static void PostmasterStateMachine(void);
435
static void UpdatePMState(PMState newState);
436
437
pg_noreturn static void ExitPostmaster(int status);
438
static int  ServerLoop(void);
439
static int  BackendStartup(ClientSocket *client_sock);
440
static void report_fork_failure_to_client(ClientSocket *client_sock, int errnum);
441
static CAC_state canAcceptConnections(BackendType backend_type);
442
static void signal_child(PMChild *pmchild, int signal);
443
static bool SignalChildren(int signal, BackendTypeMask targetMask);
444
static void TerminateChildren(int signal);
445
static int  CountChildren(BackendTypeMask targetMask);
446
static void LaunchMissingBackgroundProcesses(void);
447
static void maybe_start_bgworkers(void);
448
static bool maybe_reap_io_worker(int pid);
449
static void maybe_adjust_io_workers(void);
450
static bool CreateOptsFile(int argc, char *argv[], char *fullprogname);
451
static PMChild *StartChildProcess(BackendType type);
452
static void StartSysLogger(void);
453
static void StartAutovacuumWorker(void);
454
static bool StartBackgroundWorker(RegisteredBgWorker *rw);
455
static void InitPostmasterDeathWatchHandle(void);
456
457
#ifdef WIN32
458
#define WNOHANG 0       /* ignored, so any integer value will do */
459
460
static pid_t waitpid(pid_t pid, int *exitstatus, int options);
461
static void WINAPI pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired);
462
463
static HANDLE win32ChildQueue;
464
465
typedef struct
466
{
467
  HANDLE    waitHandle;
468
  HANDLE    procHandle;
469
  DWORD   procId;
470
} win32_deadchild_waitinfo;
471
#endif              /* WIN32 */
472
473
/* Macros to check exit status of a child process */
474
0
#define EXIT_STATUS_0(st)  ((st) == 0)
475
0
#define EXIT_STATUS_1(st)  (WIFEXITED(st) && WEXITSTATUS(st) == 1)
476
0
#define EXIT_STATUS_3(st)  (WIFEXITED(st) && WEXITSTATUS(st) == 3)
477
478
#ifndef WIN32
479
/*
480
 * File descriptors for pipe used to monitor if postmaster is alive.
481
 * First is POSTMASTER_FD_WATCH, second is POSTMASTER_FD_OWN.
482
 */
483
int     postmaster_alive_fds[2] = {-1, -1};
484
#else
485
/* Process handle of postmaster used for the same purpose on Windows */
486
HANDLE    PostmasterHandle;
487
#endif
488
489
/*
490
 * Postmaster main entry point
491
 */
492
void
493
PostmasterMain(int argc, char *argv[])
494
0
{
495
0
  int     opt;
496
0
  int     status;
497
0
  char     *userDoption = NULL;
498
0
  bool    listen_addr_saved = false;
499
0
  char     *output_config_variable = NULL;
500
501
0
  InitProcessGlobals();
502
503
0
  PostmasterPid = MyProcPid;
504
505
0
  IsPostmasterEnvironment = true;
506
507
  /*
508
   * Start our win32 signal implementation
509
   */
510
#ifdef WIN32
511
  pgwin32_signal_initialize();
512
#endif
513
514
  /*
515
   * We should not be creating any files or directories before we check the
516
   * data directory (see checkDataDir()), but just in case set the umask to
517
   * the most restrictive (owner-only) permissions.
518
   *
519
   * checkDataDir() will reset the umask based on the data directory
520
   * permissions.
521
   */
522
0
  umask(PG_MODE_MASK_OWNER);
523
524
  /*
525
   * By default, palloc() requests in the postmaster will be allocated in
526
   * the PostmasterContext, which is space that can be recycled by backends.
527
   * Allocated data that needs to be available to backends should be
528
   * allocated in TopMemoryContext.
529
   */
530
0
  PostmasterContext = AllocSetContextCreate(TopMemoryContext,
531
0
                        "Postmaster",
532
0
                        ALLOCSET_DEFAULT_SIZES);
533
0
  MemoryContextSwitchTo(PostmasterContext);
534
535
  /* Initialize paths to installation files */
536
0
  getInstallationPaths(argv[0]);
537
538
  /*
539
   * Set up signal handlers for the postmaster process.
540
   *
541
   * CAUTION: when changing this list, check for side-effects on the signal
542
   * handling setup of child processes.  See tcop/postgres.c,
543
   * bootstrap/bootstrap.c, postmaster/bgwriter.c, postmaster/walwriter.c,
544
   * postmaster/autovacuum.c, postmaster/pgarch.c, postmaster/syslogger.c,
545
   * postmaster/bgworker.c and postmaster/checkpointer.c.
546
   */
547
0
  pqinitmask();
548
0
  sigprocmask(SIG_SETMASK, &BlockSig, NULL);
549
550
0
  pqsignal(SIGHUP, handle_pm_reload_request_signal);
551
0
  pqsignal(SIGINT, handle_pm_shutdown_request_signal);
552
0
  pqsignal(SIGQUIT, handle_pm_shutdown_request_signal);
553
0
  pqsignal(SIGTERM, handle_pm_shutdown_request_signal);
554
0
  pqsignal(SIGALRM, SIG_IGN); /* ignored */
555
0
  pqsignal(SIGPIPE, SIG_IGN); /* ignored */
556
0
  pqsignal(SIGUSR1, handle_pm_pmsignal_signal);
557
0
  pqsignal(SIGUSR2, dummy_handler);  /* unused, reserve for children */
558
0
  pqsignal(SIGCHLD, handle_pm_child_exit_signal);
559
560
  /* This may configure SIGURG, depending on platform. */
561
0
  InitializeWaitEventSupport();
562
0
  InitProcessLocalLatch();
563
564
  /*
565
   * No other place in Postgres should touch SIGTTIN/SIGTTOU handling.  We
566
   * ignore those signals in a postmaster environment, so that there is no
567
   * risk of a child process freezing up due to writing to stderr.  But for
568
   * a standalone backend, their default handling is reasonable.  Hence, all
569
   * child processes should just allow the inherited settings to stand.
570
   */
571
0
#ifdef SIGTTIN
572
0
  pqsignal(SIGTTIN, SIG_IGN); /* ignored */
573
0
#endif
574
0
#ifdef SIGTTOU
575
0
  pqsignal(SIGTTOU, SIG_IGN); /* ignored */
576
0
#endif
577
578
  /* ignore SIGXFSZ, so that ulimit violations work like disk full */
579
0
#ifdef SIGXFSZ
580
0
  pqsignal(SIGXFSZ, SIG_IGN); /* ignored */
581
0
#endif
582
583
  /* Begin accepting signals. */
584
0
  sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
585
586
  /*
587
   * Options setup
588
   */
589
0
  InitializeGUCOptions();
590
591
0
  opterr = 1;
592
593
  /*
594
   * Parse command-line options.  CAUTION: keep this in sync with
595
   * tcop/postgres.c (the option sets should not conflict) and with the
596
   * common help() function in main/main.c.
597
   */
598
0
  while ((opt = getopt(argc, argv, "B:bC:c:D:d:EeFf:h:ijk:lN:OPp:r:S:sTt:W:-:")) != -1)
599
0
  {
600
0
    switch (opt)
601
0
    {
602
0
      case 'B':
603
0
        SetConfigOption("shared_buffers", optarg, PGC_POSTMASTER, PGC_S_ARGV);
604
0
        break;
605
606
0
      case 'b':
607
        /* Undocumented flag used for binary upgrades */
608
0
        IsBinaryUpgrade = true;
609
0
        break;
610
611
0
      case 'C':
612
0
        output_config_variable = strdup(optarg);
613
0
        break;
614
615
0
      case '-':
616
617
        /*
618
         * Error if the user misplaced a special must-be-first option
619
         * for dispatching to a subprogram.  parse_dispatch_option()
620
         * returns DISPATCH_POSTMASTER if it doesn't find a match, so
621
         * error for anything else.
622
         */
623
0
        if (parse_dispatch_option(optarg) != DISPATCH_POSTMASTER)
624
0
          ereport(ERROR,
625
0
              (errcode(ERRCODE_SYNTAX_ERROR),
626
0
               errmsg("--%s must be first argument", optarg)));
627
628
        /* FALLTHROUGH */
629
0
      case 'c':
630
0
        {
631
0
          char     *name,
632
0
                 *value;
633
634
0
          ParseLongOption(optarg, &name, &value);
635
0
          if (!value)
636
0
          {
637
0
            if (opt == '-')
638
0
              ereport(ERROR,
639
0
                  (errcode(ERRCODE_SYNTAX_ERROR),
640
0
                   errmsg("--%s requires a value",
641
0
                      optarg)));
642
0
            else
643
0
              ereport(ERROR,
644
0
                  (errcode(ERRCODE_SYNTAX_ERROR),
645
0
                   errmsg("-c %s requires a value",
646
0
                      optarg)));
647
0
          }
648
649
0
          SetConfigOption(name, value, PGC_POSTMASTER, PGC_S_ARGV);
650
0
          pfree(name);
651
0
          pfree(value);
652
0
          break;
653
0
        }
654
655
0
      case 'D':
656
0
        userDoption = strdup(optarg);
657
0
        break;
658
659
0
      case 'd':
660
0
        set_debug_options(atoi(optarg), PGC_POSTMASTER, PGC_S_ARGV);
661
0
        break;
662
663
0
      case 'E':
664
0
        SetConfigOption("log_statement", "all", PGC_POSTMASTER, PGC_S_ARGV);
665
0
        break;
666
667
0
      case 'e':
668
0
        SetConfigOption("datestyle", "euro", PGC_POSTMASTER, PGC_S_ARGV);
669
0
        break;
670
671
0
      case 'F':
672
0
        SetConfigOption("fsync", "false", PGC_POSTMASTER, PGC_S_ARGV);
673
0
        break;
674
675
0
      case 'f':
676
0
        if (!set_plan_disabling_options(optarg, PGC_POSTMASTER, PGC_S_ARGV))
677
0
        {
678
0
          write_stderr("%s: invalid argument for option -f: \"%s\"\n",
679
0
                 progname, optarg);
680
0
          ExitPostmaster(1);
681
0
        }
682
0
        break;
683
684
0
      case 'h':
685
0
        SetConfigOption("listen_addresses", optarg, PGC_POSTMASTER, PGC_S_ARGV);
686
0
        break;
687
688
0
      case 'i':
689
0
        SetConfigOption("listen_addresses", "*", PGC_POSTMASTER, PGC_S_ARGV);
690
0
        break;
691
692
0
      case 'j':
693
        /* only used by interactive backend */
694
0
        break;
695
696
0
      case 'k':
697
0
        SetConfigOption("unix_socket_directories", optarg, PGC_POSTMASTER, PGC_S_ARGV);
698
0
        break;
699
700
0
      case 'l':
701
0
        SetConfigOption("ssl", "true", PGC_POSTMASTER, PGC_S_ARGV);
702
0
        break;
703
704
0
      case 'N':
705
0
        SetConfigOption("max_connections", optarg, PGC_POSTMASTER, PGC_S_ARGV);
706
0
        break;
707
708
0
      case 'O':
709
0
        SetConfigOption("allow_system_table_mods", "true", PGC_POSTMASTER, PGC_S_ARGV);
710
0
        break;
711
712
0
      case 'P':
713
0
        SetConfigOption("ignore_system_indexes", "true", PGC_POSTMASTER, PGC_S_ARGV);
714
0
        break;
715
716
0
      case 'p':
717
0
        SetConfigOption("port", optarg, PGC_POSTMASTER, PGC_S_ARGV);
718
0
        break;
719
720
0
      case 'r':
721
        /* only used by single-user backend */
722
0
        break;
723
724
0
      case 'S':
725
0
        SetConfigOption("work_mem", optarg, PGC_POSTMASTER, PGC_S_ARGV);
726
0
        break;
727
728
0
      case 's':
729
0
        SetConfigOption("log_statement_stats", "true", PGC_POSTMASTER, PGC_S_ARGV);
730
0
        break;
731
732
0
      case 'T':
733
734
        /*
735
         * This option used to be defined as sending SIGSTOP after a
736
         * backend crash, but sending SIGABRT seems more useful.
737
         */
738
0
        SetConfigOption("send_abort_for_crash", "true", PGC_POSTMASTER, PGC_S_ARGV);
739
0
        break;
740
741
0
      case 't':
742
0
        {
743
0
          const char *tmp = get_stats_option_name(optarg);
744
745
0
          if (tmp)
746
0
          {
747
0
            SetConfigOption(tmp, "true", PGC_POSTMASTER, PGC_S_ARGV);
748
0
          }
749
0
          else
750
0
          {
751
0
            write_stderr("%s: invalid argument for option -t: \"%s\"\n",
752
0
                   progname, optarg);
753
0
            ExitPostmaster(1);
754
0
          }
755
0
          break;
756
0
        }
757
758
0
      case 'W':
759
0
        SetConfigOption("post_auth_delay", optarg, PGC_POSTMASTER, PGC_S_ARGV);
760
0
        break;
761
762
0
      default:
763
0
        write_stderr("Try \"%s --help\" for more information.\n",
764
0
               progname);
765
0
        ExitPostmaster(1);
766
0
    }
767
0
  }
768
769
  /*
770
   * Postmaster accepts no non-option switch arguments.
771
   */
772
0
  if (optind < argc)
773
0
  {
774
0
    write_stderr("%s: invalid argument: \"%s\"\n",
775
0
           progname, argv[optind]);
776
0
    write_stderr("Try \"%s --help\" for more information.\n",
777
0
           progname);
778
0
    ExitPostmaster(1);
779
0
  }
780
781
  /*
782
   * Locate the proper configuration files and data directory, and read
783
   * postgresql.conf for the first time.
784
   */
785
0
  if (!SelectConfigFiles(userDoption, progname))
786
0
    ExitPostmaster(2);
787
788
0
  if (output_config_variable != NULL)
789
0
  {
790
    /*
791
     * If this is a runtime-computed GUC, it hasn't yet been initialized,
792
     * and the present value is not useful.  However, this is a convenient
793
     * place to print the value for most GUCs because it is safe to run
794
     * postmaster startup to this point even if the server is already
795
     * running.  For the handful of runtime-computed GUCs that we cannot
796
     * provide meaningful values for yet, we wait until later in
797
     * postmaster startup to print the value.  We won't be able to use -C
798
     * on running servers for those GUCs, but using this option now would
799
     * lead to incorrect results for them.
800
     */
801
0
    int     flags = GetConfigOptionFlags(output_config_variable, true);
802
803
0
    if ((flags & GUC_RUNTIME_COMPUTED) == 0)
804
0
    {
805
      /*
806
       * "-C guc" was specified, so print GUC's value and exit.  No
807
       * extra permission check is needed because the user is reading
808
       * inside the data dir.
809
       */
810
0
      const char *config_val = GetConfigOption(output_config_variable,
811
0
                           false, false);
812
813
0
      puts(config_val ? config_val : "");
814
0
      ExitPostmaster(0);
815
0
    }
816
817
    /*
818
     * A runtime-computed GUC will be printed later on.  As we initialize
819
     * a server startup sequence, silence any log messages that may show
820
     * up in the output generated.  FATAL and more severe messages are
821
     * useful to show, even if one would only expect at least PANIC.  LOG
822
     * entries are hidden.
823
     */
824
0
    SetConfigOption("log_min_messages", "FATAL", PGC_SUSET,
825
0
            PGC_S_OVERRIDE);
826
0
  }
827
828
  /* Verify that DataDir looks reasonable */
829
0
  checkDataDir();
830
831
  /* Check that pg_control exists */
832
0
  checkControlFile();
833
834
  /* And switch working directory into it */
835
0
  ChangeToDataDir();
836
837
  /*
838
   * Check for invalid combinations of GUC settings.
839
   */
840
0
  if (SuperuserReservedConnections + ReservedConnections >= MaxConnections)
841
0
  {
842
0
    write_stderr("%s: \"superuser_reserved_connections\" (%d) plus \"reserved_connections\" (%d) must be less than \"max_connections\" (%d)\n",
843
0
           progname,
844
0
           SuperuserReservedConnections, ReservedConnections,
845
0
           MaxConnections);
846
0
    ExitPostmaster(1);
847
0
  }
848
0
  if (XLogArchiveMode > ARCHIVE_MODE_OFF && wal_level == WAL_LEVEL_MINIMAL)
849
0
    ereport(ERROR,
850
0
        (errmsg("WAL archival cannot be enabled when \"wal_level\" is \"minimal\"")));
851
0
  if (max_wal_senders > 0 && wal_level == WAL_LEVEL_MINIMAL)
852
0
    ereport(ERROR,
853
0
        (errmsg("WAL streaming (\"max_wal_senders\" > 0) requires \"wal_level\" to be \"replica\" or \"logical\"")));
854
0
  if (summarize_wal && wal_level == WAL_LEVEL_MINIMAL)
855
0
    ereport(ERROR,
856
0
        (errmsg("WAL cannot be summarized when \"wal_level\" is \"minimal\"")));
857
0
  if (sync_replication_slots && wal_level < WAL_LEVEL_LOGICAL)
858
0
    ereport(ERROR,
859
0
        (errmsg("replication slot synchronization (\"sync_replication_slots\" = on) requires \"wal_level\" >= \"logical\"")));
860
861
  /*
862
   * Other one-time internal sanity checks can go here, if they are fast.
863
   * (Put any slow processing further down, after postmaster.pid creation.)
864
   */
865
0
  if (!CheckDateTokenTables())
866
0
  {
867
0
    write_stderr("%s: invalid datetoken tables, please fix\n", progname);
868
0
    ExitPostmaster(1);
869
0
  }
870
871
  /*
872
   * Now that we are done processing the postmaster arguments, reset
873
   * getopt(3) library so that it will work correctly in subprocesses.
874
   */
875
0
  optind = 1;
876
#ifdef HAVE_INT_OPTRESET
877
  optreset = 1;       /* some systems need this too */
878
#endif
879
880
  /* For debugging: display postmaster environment */
881
0
  if (message_level_is_interesting(DEBUG3))
882
0
  {
883
0
#if !defined(WIN32) || defined(_MSC_VER)
884
0
    extern char **environ;
885
0
#endif
886
0
    char    **p;
887
0
    StringInfoData si;
888
889
0
    initStringInfo(&si);
890
891
0
    appendStringInfoString(&si, "initial environment dump:");
892
0
    for (p = environ; *p; ++p)
893
0
      appendStringInfo(&si, "\n%s", *p);
894
895
0
    ereport(DEBUG3, errmsg_internal("%s", si.data));
896
0
    pfree(si.data);
897
0
  }
898
899
  /*
900
   * Create lockfile for data directory.
901
   *
902
   * We want to do this before we try to grab the input sockets, because the
903
   * data directory interlock is more reliable than the socket-file
904
   * interlock (thanks to whoever decided to put socket files in /tmp :-().
905
   * For the same reason, it's best to grab the TCP socket(s) before the
906
   * Unix socket(s).
907
   *
908
   * Also note that this internally sets up the on_proc_exit function that
909
   * is responsible for removing both data directory and socket lockfiles;
910
   * so it must happen before opening sockets so that at exit, the socket
911
   * lockfiles go away after CloseServerPorts runs.
912
   */
913
0
  CreateDataDirLockFile(true);
914
915
  /*
916
   * Read the control file (for error checking and config info).
917
   *
918
   * Since we verify the control file's CRC, this has a useful side effect
919
   * on machines where we need a run-time test for CRC support instructions.
920
   * The postmaster will do the test once at startup, and then its child
921
   * processes will inherit the correct function pointer and not need to
922
   * repeat the test.
923
   */
924
0
  LocalProcessControlFile(false);
925
926
  /*
927
   * Register the apply launcher.  It's probably a good idea to call this
928
   * before any modules had a chance to take the background worker slots.
929
   */
930
0
  ApplyLauncherRegister();
931
932
  /*
933
   * process any libraries that should be preloaded at postmaster start
934
   */
935
0
  process_shared_preload_libraries();
936
937
  /*
938
   * Initialize SSL library, if specified.
939
   */
940
#ifdef USE_SSL
941
  if (EnableSSL)
942
  {
943
    (void) secure_initialize(true);
944
    LoadedSSL = true;
945
  }
946
#endif
947
948
  /*
949
   * Now that loadable modules have had their chance to alter any GUCs,
950
   * calculate MaxBackends and initialize the machinery to track child
951
   * processes.
952
   */
953
0
  InitializeMaxBackends();
954
0
  InitPostmasterChildSlots();
955
956
  /*
957
   * Calculate the size of the PGPROC fast-path lock arrays.
958
   */
959
0
  InitializeFastPathLocks();
960
961
  /*
962
   * Give preloaded libraries a chance to request additional shared memory.
963
   */
964
0
  process_shmem_requests();
965
966
  /*
967
   * Now that loadable modules have had their chance to request additional
968
   * shared memory, determine the value of any runtime-computed GUCs that
969
   * depend on the amount of shared memory required.
970
   */
971
0
  InitializeShmemGUCs();
972
973
  /*
974
   * Now that modules have been loaded, we can process any custom resource
975
   * managers specified in the wal_consistency_checking GUC.
976
   */
977
0
  InitializeWalConsistencyChecking();
978
979
  /*
980
   * If -C was specified with a runtime-computed GUC, we held off printing
981
   * the value earlier, as the GUC was not yet initialized.  We handle -C
982
   * for most GUCs before we lock the data directory so that the option may
983
   * be used on a running server.  However, a handful of GUCs are runtime-
984
   * computed and do not have meaningful values until after locking the data
985
   * directory, and we cannot safely calculate their values earlier on a
986
   * running server.  At this point, such GUCs should be properly
987
   * initialized, and we haven't yet set up shared memory, so this is a good
988
   * time to handle the -C option for these special GUCs.
989
   */
990
0
  if (output_config_variable != NULL)
991
0
  {
992
0
    const char *config_val = GetConfigOption(output_config_variable,
993
0
                         false, false);
994
995
0
    puts(config_val ? config_val : "");
996
0
    ExitPostmaster(0);
997
0
  }
998
999
  /*
1000
   * Set up shared memory and semaphores.
1001
   *
1002
   * Note: if using SysV shmem and/or semas, each postmaster startup will
1003
   * normally choose the same IPC keys.  This helps ensure that we will
1004
   * clean up dead IPC objects if the postmaster crashes and is restarted.
1005
   */
1006
0
  CreateSharedMemoryAndSemaphores();
1007
1008
  /*
1009
   * Estimate number of openable files.  This must happen after setting up
1010
   * semaphores, because on some platforms semaphores count as open files.
1011
   */
1012
0
  set_max_safe_fds();
1013
1014
  /*
1015
   * Initialize pipe (or process handle on Windows) that allows children to
1016
   * wake up from sleep on postmaster death.
1017
   */
1018
0
  InitPostmasterDeathWatchHandle();
1019
1020
#ifdef WIN32
1021
1022
  /*
1023
   * Initialize I/O completion port used to deliver list of dead children.
1024
   */
1025
  win32ChildQueue = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 1);
1026
  if (win32ChildQueue == NULL)
1027
    ereport(FATAL,
1028
        (errmsg("could not create I/O completion port for child queue")));
1029
#endif
1030
1031
#ifdef EXEC_BACKEND
1032
  /* Write out nondefault GUC settings for child processes to use */
1033
  write_nondefault_variables(PGC_POSTMASTER);
1034
1035
  /*
1036
   * Clean out the temp directory used to transmit parameters to child
1037
   * processes (see internal_forkexec).  We must do this before launching
1038
   * any child processes, else we have a race condition: we could remove a
1039
   * parameter file before the child can read it.  It should be safe to do
1040
   * so now, because we verified earlier that there are no conflicting
1041
   * Postgres processes in this data directory.
1042
   */
1043
  RemovePgTempFilesInDir(PG_TEMP_FILES_DIR, true, false);
1044
#endif
1045
1046
  /*
1047
   * Forcibly remove the files signaling a standby promotion request.
1048
   * Otherwise, the existence of those files triggers a promotion too early,
1049
   * whether a user wants that or not.
1050
   *
1051
   * This removal of files is usually unnecessary because they can exist
1052
   * only during a few moments during a standby promotion. However there is
1053
   * a race condition: if pg_ctl promote is executed and creates the files
1054
   * during a promotion, the files can stay around even after the server is
1055
   * brought up to be the primary.  Then, if a new standby starts by using
1056
   * the backup taken from the new primary, the files can exist at server
1057
   * startup and must be removed in order to avoid an unexpected promotion.
1058
   *
1059
   * Note that promotion signal files need to be removed before the startup
1060
   * process is invoked. Because, after that, they can be used by
1061
   * postmaster's SIGUSR1 signal handler.
1062
   */
1063
0
  RemovePromoteSignalFiles();
1064
1065
  /* Do the same for logrotate signal file */
1066
0
  RemoveLogrotateSignalFiles();
1067
1068
  /* Remove any outdated file holding the current log filenames. */
1069
0
  if (unlink(LOG_METAINFO_DATAFILE) < 0 && errno != ENOENT)
1070
0
    ereport(LOG,
1071
0
        (errcode_for_file_access(),
1072
0
         errmsg("could not remove file \"%s\": %m",
1073
0
            LOG_METAINFO_DATAFILE)));
1074
1075
  /*
1076
   * If enabled, start up syslogger collection subprocess
1077
   */
1078
0
  if (Logging_collector)
1079
0
    StartSysLogger();
1080
1081
  /*
1082
   * Reset whereToSendOutput from DestDebug (its starting state) to
1083
   * DestNone. This stops ereport from sending log messages to stderr unless
1084
   * Log_destination permits.  We don't do this until the postmaster is
1085
   * fully launched, since startup failures may as well be reported to
1086
   * stderr.
1087
   *
1088
   * If we are in fact disabling logging to stderr, first emit a log message
1089
   * saying so, to provide a breadcrumb trail for users who may not remember
1090
   * that their logging is configured to go somewhere else.
1091
   */
1092
0
  if (!(Log_destination & LOG_DESTINATION_STDERR))
1093
0
    ereport(LOG,
1094
0
        (errmsg("ending log output to stderr"),
1095
0
         errhint("Future log output will go to log destination \"%s\".",
1096
0
             Log_destination_string)));
1097
1098
0
  whereToSendOutput = DestNone;
1099
1100
  /*
1101
   * Report server startup in log.  While we could emit this much earlier,
1102
   * it seems best to do so after starting the log collector, if we intend
1103
   * to use one.
1104
   */
1105
0
  ereport(LOG,
1106
0
      (errmsg("starting %s", PG_VERSION_STR)));
1107
1108
  /*
1109
   * Establish input sockets.
1110
   *
1111
   * First set up an on_proc_exit function that's charged with closing the
1112
   * sockets again at postmaster shutdown.
1113
   */
1114
0
  ListenSockets = palloc(MAXLISTEN * sizeof(pgsocket));
1115
0
  on_proc_exit(CloseServerPorts, 0);
1116
1117
0
  if (ListenAddresses)
1118
0
  {
1119
0
    char     *rawstring;
1120
0
    List     *elemlist;
1121
0
    ListCell   *l;
1122
0
    int     success = 0;
1123
1124
    /* Need a modifiable copy of ListenAddresses */
1125
0
    rawstring = pstrdup(ListenAddresses);
1126
1127
    /* Parse string into list of hostnames */
1128
0
    if (!SplitGUCList(rawstring, ',', &elemlist))
1129
0
    {
1130
      /* syntax error in list */
1131
0
      ereport(FATAL,
1132
0
          (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1133
0
           errmsg("invalid list syntax in parameter \"%s\"",
1134
0
              "listen_addresses")));
1135
0
    }
1136
1137
0
    foreach(l, elemlist)
1138
0
    {
1139
0
      char     *curhost = (char *) lfirst(l);
1140
1141
0
      if (strcmp(curhost, "*") == 0)
1142
0
        status = ListenServerPort(AF_UNSPEC, NULL,
1143
0
                      (unsigned short) PostPortNumber,
1144
0
                      NULL,
1145
0
                      ListenSockets,
1146
0
                      &NumListenSockets,
1147
0
                      MAXLISTEN);
1148
0
      else
1149
0
        status = ListenServerPort(AF_UNSPEC, curhost,
1150
0
                      (unsigned short) PostPortNumber,
1151
0
                      NULL,
1152
0
                      ListenSockets,
1153
0
                      &NumListenSockets,
1154
0
                      MAXLISTEN);
1155
1156
0
      if (status == STATUS_OK)
1157
0
      {
1158
0
        success++;
1159
        /* record the first successful host addr in lockfile */
1160
0
        if (!listen_addr_saved)
1161
0
        {
1162
0
          AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, curhost);
1163
0
          listen_addr_saved = true;
1164
0
        }
1165
0
      }
1166
0
      else
1167
0
        ereport(WARNING,
1168
0
            (errmsg("could not create listen socket for \"%s\"",
1169
0
                curhost)));
1170
0
    }
1171
1172
0
    if (!success && elemlist != NIL)
1173
0
      ereport(FATAL,
1174
0
          (errmsg("could not create any TCP/IP sockets")));
1175
1176
0
    list_free(elemlist);
1177
0
    pfree(rawstring);
1178
0
  }
1179
1180
#ifdef USE_BONJOUR
1181
  /* Register for Bonjour only if we opened TCP socket(s) */
1182
  if (enable_bonjour && NumListenSockets > 0)
1183
  {
1184
    DNSServiceErrorType err;
1185
1186
    /*
1187
     * We pass 0 for interface_index, which will result in registering on
1188
     * all "applicable" interfaces.  It's not entirely clear from the
1189
     * DNS-SD docs whether this would be appropriate if we have bound to
1190
     * just a subset of the available network interfaces.
1191
     */
1192
    err = DNSServiceRegister(&bonjour_sdref,
1193
                 0,
1194
                 0,
1195
                 bonjour_name,
1196
                 "_postgresql._tcp.",
1197
                 NULL,
1198
                 NULL,
1199
                 pg_hton16(PostPortNumber),
1200
                 0,
1201
                 NULL,
1202
                 NULL,
1203
                 NULL);
1204
    if (err != kDNSServiceErr_NoError)
1205
      ereport(LOG,
1206
          (errmsg("DNSServiceRegister() failed: error code %ld",
1207
              (long) err)));
1208
1209
    /*
1210
     * We don't bother to read the mDNS daemon's reply, and we expect that
1211
     * it will automatically terminate our registration when the socket is
1212
     * closed at postmaster termination.  So there's nothing more to be
1213
     * done here.  However, the bonjour_sdref is kept around so that
1214
     * forked children can close their copies of the socket.
1215
     */
1216
  }
1217
#endif
1218
1219
0
  if (Unix_socket_directories)
1220
0
  {
1221
0
    char     *rawstring;
1222
0
    List     *elemlist;
1223
0
    ListCell   *l;
1224
0
    int     success = 0;
1225
1226
    /* Need a modifiable copy of Unix_socket_directories */
1227
0
    rawstring = pstrdup(Unix_socket_directories);
1228
1229
    /* Parse string into list of directories */
1230
0
    if (!SplitDirectoriesString(rawstring, ',', &elemlist))
1231
0
    {
1232
      /* syntax error in list */
1233
0
      ereport(FATAL,
1234
0
          (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1235
0
           errmsg("invalid list syntax in parameter \"%s\"",
1236
0
              "unix_socket_directories")));
1237
0
    }
1238
1239
0
    foreach(l, elemlist)
1240
0
    {
1241
0
      char     *socketdir = (char *) lfirst(l);
1242
1243
0
      status = ListenServerPort(AF_UNIX, NULL,
1244
0
                    (unsigned short) PostPortNumber,
1245
0
                    socketdir,
1246
0
                    ListenSockets,
1247
0
                    &NumListenSockets,
1248
0
                    MAXLISTEN);
1249
1250
0
      if (status == STATUS_OK)
1251
0
      {
1252
0
        success++;
1253
        /* record the first successful Unix socket in lockfile */
1254
0
        if (success == 1)
1255
0
          AddToDataDirLockFile(LOCK_FILE_LINE_SOCKET_DIR, socketdir);
1256
0
      }
1257
0
      else
1258
0
        ereport(WARNING,
1259
0
            (errmsg("could not create Unix-domain socket in directory \"%s\"",
1260
0
                socketdir)));
1261
0
    }
1262
1263
0
    if (!success && elemlist != NIL)
1264
0
      ereport(FATAL,
1265
0
          (errmsg("could not create any Unix-domain sockets")));
1266
1267
0
    list_free_deep(elemlist);
1268
0
    pfree(rawstring);
1269
0
  }
1270
1271
  /*
1272
   * check that we have some socket to listen on
1273
   */
1274
0
  if (NumListenSockets == 0)
1275
0
    ereport(FATAL,
1276
0
        (errmsg("no socket created for listening")));
1277
1278
  /*
1279
   * If no valid TCP ports, write an empty line for listen address,
1280
   * indicating the Unix socket must be used.  Note that this line is not
1281
   * added to the lock file until there is a socket backing it.
1282
   */
1283
0
  if (!listen_addr_saved)
1284
0
    AddToDataDirLockFile(LOCK_FILE_LINE_LISTEN_ADDR, "");
1285
1286
  /*
1287
   * Record postmaster options.  We delay this till now to avoid recording
1288
   * bogus options (eg, unusable port number).
1289
   */
1290
0
  if (!CreateOptsFile(argc, argv, my_exec_path))
1291
0
    ExitPostmaster(1);
1292
1293
  /*
1294
   * Write the external PID file if requested
1295
   */
1296
0
  if (external_pid_file)
1297
0
  {
1298
0
    FILE     *fpidfile = fopen(external_pid_file, "w");
1299
1300
0
    if (fpidfile)
1301
0
    {
1302
0
      fprintf(fpidfile, "%d\n", MyProcPid);
1303
0
      fclose(fpidfile);
1304
1305
      /* Make PID file world readable */
1306
0
      if (chmod(external_pid_file, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH) != 0)
1307
0
        write_stderr("%s: could not change permissions of external PID file \"%s\": %m\n",
1308
0
               progname, external_pid_file);
1309
0
    }
1310
0
    else
1311
0
      write_stderr("%s: could not write external PID file \"%s\": %m\n",
1312
0
             progname, external_pid_file);
1313
1314
0
    on_proc_exit(unlink_external_pid_file, 0);
1315
0
  }
1316
1317
  /*
1318
   * Remove old temporary files.  At this point there can be no other
1319
   * Postgres processes running in this directory, so this should be safe.
1320
   */
1321
0
  RemovePgTempFiles();
1322
1323
  /*
1324
   * Initialize the autovacuum subsystem (again, no process start yet)
1325
   */
1326
0
  autovac_init();
1327
1328
  /*
1329
   * Load configuration files for client authentication.
1330
   */
1331
0
  if (!load_hba())
1332
0
  {
1333
    /*
1334
     * It makes no sense to continue if we fail to load the HBA file,
1335
     * since there is no way to connect to the database in this case.
1336
     */
1337
0
    ereport(FATAL,
1338
    /* translator: %s is a configuration file */
1339
0
        (errmsg("could not load %s", HbaFileName)));
1340
0
  }
1341
0
  if (!load_ident())
1342
0
  {
1343
    /*
1344
     * We can start up without the IDENT file, although it means that you
1345
     * cannot log in using any of the authentication methods that need a
1346
     * user name mapping. load_ident() already logged the details of error
1347
     * to the log.
1348
     */
1349
0
  }
1350
1351
#ifdef HAVE_PTHREAD_IS_THREADED_NP
1352
1353
  /*
1354
   * On macOS, libintl replaces setlocale() with a version that calls
1355
   * CFLocaleCopyCurrent() when its second argument is "" and every relevant
1356
   * environment variable is unset or empty.  CFLocaleCopyCurrent() makes
1357
   * the process multithreaded.  The postmaster calls sigprocmask() and
1358
   * calls fork() without an immediate exec(), both of which have undefined
1359
   * behavior in a multithreaded program.  A multithreaded postmaster is the
1360
   * normal case on Windows, which offers neither fork() nor sigprocmask().
1361
   * Currently, macOS is the only platform having pthread_is_threaded_np(),
1362
   * so we need not worry whether this HINT is appropriate elsewhere.
1363
   */
1364
  if (pthread_is_threaded_np() != 0)
1365
    ereport(FATAL,
1366
        (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1367
         errmsg("postmaster became multithreaded during startup"),
1368
         errhint("Set the LC_ALL environment variable to a valid locale.")));
1369
#endif
1370
1371
  /*
1372
   * Remember postmaster startup time
1373
   */
1374
0
  PgStartTime = GetCurrentTimestamp();
1375
1376
  /*
1377
   * Report postmaster status in the postmaster.pid file, to allow pg_ctl to
1378
   * see what's happening.
1379
   */
1380
0
  AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STARTING);
1381
1382
0
  UpdatePMState(PM_STARTUP);
1383
1384
  /* Make sure we can perform I/O while starting up. */
1385
0
  maybe_adjust_io_workers();
1386
1387
  /* Start bgwriter and checkpointer so they can help with recovery */
1388
0
  if (CheckpointerPMChild == NULL)
1389
0
    CheckpointerPMChild = StartChildProcess(B_CHECKPOINTER);
1390
0
  if (BgWriterPMChild == NULL)
1391
0
    BgWriterPMChild = StartChildProcess(B_BG_WRITER);
1392
1393
  /*
1394
   * We're ready to rock and roll...
1395
   */
1396
0
  StartupPMChild = StartChildProcess(B_STARTUP);
1397
0
  Assert(StartupPMChild != NULL);
1398
0
  StartupStatus = STARTUP_RUNNING;
1399
1400
  /* Some workers may be scheduled to start now */
1401
0
  maybe_start_bgworkers();
1402
1403
0
  status = ServerLoop();
1404
1405
  /*
1406
   * ServerLoop probably shouldn't ever return, but if it does, close down.
1407
   */
1408
0
  ExitPostmaster(status != STATUS_OK);
1409
1410
0
  abort();          /* not reached */
1411
0
}
1412
1413
1414
/*
1415
 * on_proc_exit callback to close server's listen sockets
1416
 */
1417
static void
1418
CloseServerPorts(int status, Datum arg)
1419
{
1420
  int     i;
1421
1422
  /*
1423
   * First, explicitly close all the socket FDs.  We used to just let this
1424
   * happen implicitly at postmaster exit, but it's better to close them
1425
   * before we remove the postmaster.pid lockfile; otherwise there's a race
1426
   * condition if a new postmaster wants to re-use the TCP port number.
1427
   */
1428
  for (i = 0; i < NumListenSockets; i++)
1429
  {
1430
    if (closesocket(ListenSockets[i]) != 0)
1431
      elog(LOG, "could not close listen socket: %m");
1432
  }
1433
  NumListenSockets = 0;
1434
1435
  /*
1436
   * Next, remove any filesystem entries for Unix sockets.  To avoid race
1437
   * conditions against incoming postmasters, this must happen after closing
1438
   * the sockets and before removing lock files.
1439
   */
1440
  RemoveSocketFiles();
1441
1442
  /*
1443
   * We don't do anything about socket lock files here; those will be
1444
   * removed in a later on_proc_exit callback.
1445
   */
1446
}
1447
1448
/*
1449
 * on_proc_exit callback to delete external_pid_file
1450
 */
1451
static void
1452
unlink_external_pid_file(int status, Datum arg)
1453
0
{
1454
0
  if (external_pid_file)
1455
0
    unlink(external_pid_file);
1456
0
}
1457
1458
1459
/*
1460
 * Compute and check the directory paths to files that are part of the
1461
 * installation (as deduced from the postgres executable's own location)
1462
 */
1463
static void
1464
getInstallationPaths(const char *argv0)
1465
0
{
1466
0
  DIR      *pdir;
1467
1468
  /* Locate the postgres executable itself */
1469
0
  if (find_my_exec(argv0, my_exec_path) < 0)
1470
0
    ereport(FATAL,
1471
0
        (errmsg("%s: could not locate my own executable path", argv0)));
1472
1473
#ifdef EXEC_BACKEND
1474
  /* Locate executable backend before we change working directory */
1475
  if (find_other_exec(argv0, "postgres", PG_BACKEND_VERSIONSTR,
1476
            postgres_exec_path) < 0)
1477
    ereport(FATAL,
1478
        (errmsg("%s: could not locate matching postgres executable",
1479
            argv0)));
1480
#endif
1481
1482
  /*
1483
   * Locate the pkglib directory --- this has to be set early in case we try
1484
   * to load any modules from it in response to postgresql.conf entries.
1485
   */
1486
0
  get_pkglib_path(my_exec_path, pkglib_path);
1487
1488
  /*
1489
   * Verify that there's a readable directory there; otherwise the Postgres
1490
   * installation is incomplete or corrupt.  (A typical cause of this
1491
   * failure is that the postgres executable has been moved or hardlinked to
1492
   * some directory that's not a sibling of the installation lib/
1493
   * directory.)
1494
   */
1495
0
  pdir = AllocateDir(pkglib_path);
1496
0
  if (pdir == NULL)
1497
0
    ereport(ERROR,
1498
0
        (errcode_for_file_access(),
1499
0
         errmsg("could not open directory \"%s\": %m",
1500
0
            pkglib_path),
1501
0
         errhint("This may indicate an incomplete PostgreSQL installation, or that the file \"%s\" has been moved away from its proper location.",
1502
0
             my_exec_path)));
1503
0
  FreeDir(pdir);
1504
1505
  /*
1506
   * It's not worth checking the share/ directory.  If the lib/ directory is
1507
   * there, then share/ probably is too.
1508
   */
1509
0
}
1510
1511
/*
1512
 * Check that pg_control exists in the correct location in the data directory.
1513
 *
1514
 * No attempt is made to validate the contents of pg_control here.  This is
1515
 * just a sanity check to see if we are looking at a real data directory.
1516
 */
1517
static void
1518
checkControlFile(void)
1519
0
{
1520
0
  char    path[MAXPGPATH];
1521
0
  FILE     *fp;
1522
1523
0
  snprintf(path, sizeof(path), "%s/%s", DataDir, XLOG_CONTROL_FILE);
1524
1525
0
  fp = AllocateFile(path, PG_BINARY_R);
1526
0
  if (fp == NULL)
1527
0
  {
1528
0
    write_stderr("%s: could not find the database system\n"
1529
0
           "Expected to find it in the directory \"%s\",\n"
1530
0
           "but could not open file \"%s\": %m\n",
1531
0
           progname, DataDir, path);
1532
0
    ExitPostmaster(2);
1533
0
  }
1534
0
  FreeFile(fp);
1535
0
}
1536
1537
/*
1538
 * Determine how long should we let ServerLoop sleep, in milliseconds.
1539
 *
1540
 * In normal conditions we wait at most one minute, to ensure that the other
1541
 * background tasks handled by ServerLoop get done even when no requests are
1542
 * arriving.  However, if there are background workers waiting to be started,
1543
 * we don't actually sleep so that they are quickly serviced.  Other exception
1544
 * cases are as shown in the code.
1545
 */
1546
static int
1547
DetermineSleepTime(void)
1548
0
{
1549
0
  TimestampTz next_wakeup = 0;
1550
1551
  /*
1552
   * Normal case: either there are no background workers at all, or we're in
1553
   * a shutdown sequence (during which we ignore bgworkers altogether).
1554
   */
1555
0
  if (Shutdown > NoShutdown ||
1556
0
    (!StartWorkerNeeded && !HaveCrashedWorker))
1557
0
  {
1558
0
    if (AbortStartTime != 0)
1559
0
    {
1560
0
      int     seconds;
1561
1562
      /* time left to abort; clamp to 0 in case it already expired */
1563
0
      seconds = SIGKILL_CHILDREN_AFTER_SECS -
1564
0
        (time(NULL) - AbortStartTime);
1565
1566
0
      return Max(seconds * 1000, 0);
1567
0
    }
1568
0
    else
1569
0
      return 60 * 1000;
1570
0
  }
1571
1572
0
  if (StartWorkerNeeded)
1573
0
    return 0;
1574
1575
0
  if (HaveCrashedWorker)
1576
0
  {
1577
0
    dlist_mutable_iter iter;
1578
1579
    /*
1580
     * When there are crashed bgworkers, we sleep just long enough that
1581
     * they are restarted when they request to be.  Scan the list to
1582
     * determine the minimum of all wakeup times according to most recent
1583
     * crash time and requested restart interval.
1584
     */
1585
0
    dlist_foreach_modify(iter, &BackgroundWorkerList)
1586
0
    {
1587
0
      RegisteredBgWorker *rw;
1588
0
      TimestampTz this_wakeup;
1589
1590
0
      rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur);
1591
1592
0
      if (rw->rw_crashed_at == 0)
1593
0
        continue;
1594
1595
0
      if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART
1596
0
        || rw->rw_terminate)
1597
0
      {
1598
0
        ForgetBackgroundWorker(rw);
1599
0
        continue;
1600
0
      }
1601
1602
0
      this_wakeup = TimestampTzPlusMilliseconds(rw->rw_crashed_at,
1603
0
                            1000L * rw->rw_worker.bgw_restart_time);
1604
0
      if (next_wakeup == 0 || this_wakeup < next_wakeup)
1605
0
        next_wakeup = this_wakeup;
1606
0
    }
1607
0
  }
1608
1609
0
  if (next_wakeup != 0)
1610
0
  {
1611
0
    int     ms;
1612
1613
    /* result of TimestampDifferenceMilliseconds is in [0, INT_MAX] */
1614
0
    ms = (int) TimestampDifferenceMilliseconds(GetCurrentTimestamp(),
1615
0
                           next_wakeup);
1616
0
    return Min(60 * 1000, ms);
1617
0
  }
1618
1619
0
  return 60 * 1000;
1620
0
}
1621
1622
/*
1623
 * Activate or deactivate notifications of server socket events.  Since we
1624
 * don't currently have a way to remove events from an existing WaitEventSet,
1625
 * we'll just destroy and recreate the whole thing.  This is called during
1626
 * shutdown so we can wait for backends to exit without accepting new
1627
 * connections, and during crash reinitialization when we need to start
1628
 * listening for new connections again.  The WaitEventSet will be freed in fork
1629
 * children by ClosePostmasterPorts().
1630
 */
1631
static void
1632
ConfigurePostmasterWaitSet(bool accept_connections)
1633
0
{
1634
0
  if (pm_wait_set)
1635
0
    FreeWaitEventSet(pm_wait_set);
1636
0
  pm_wait_set = NULL;
1637
1638
0
  pm_wait_set = CreateWaitEventSet(NULL,
1639
0
                   accept_connections ? (1 + NumListenSockets) : 1);
1640
0
  AddWaitEventToSet(pm_wait_set, WL_LATCH_SET, PGINVALID_SOCKET, MyLatch,
1641
0
            NULL);
1642
1643
0
  if (accept_connections)
1644
0
  {
1645
0
    for (int i = 0; i < NumListenSockets; i++)
1646
0
      AddWaitEventToSet(pm_wait_set, WL_SOCKET_ACCEPT, ListenSockets[i],
1647
0
                NULL, NULL);
1648
0
  }
1649
0
}
1650
1651
/*
1652
 * Main idle loop of postmaster
1653
 */
1654
static int
1655
ServerLoop(void)
1656
0
{
1657
0
  time_t    last_lockfile_recheck_time,
1658
0
        last_touch_time;
1659
0
  WaitEvent events[MAXLISTEN];
1660
0
  int     nevents;
1661
1662
0
  ConfigurePostmasterWaitSet(true);
1663
0
  last_lockfile_recheck_time = last_touch_time = time(NULL);
1664
1665
0
  for (;;)
1666
0
  {
1667
0
    time_t    now;
1668
1669
0
    nevents = WaitEventSetWait(pm_wait_set,
1670
0
                   DetermineSleepTime(),
1671
0
                   events,
1672
0
                   lengthof(events),
1673
0
                   0 /* postmaster posts no wait_events */ );
1674
1675
    /*
1676
     * Latch set by signal handler, or new connection pending on any of
1677
     * our sockets? If the latter, fork a child process to deal with it.
1678
     */
1679
0
    for (int i = 0; i < nevents; i++)
1680
0
    {
1681
0
      if (events[i].events & WL_LATCH_SET)
1682
0
        ResetLatch(MyLatch);
1683
1684
      /*
1685
       * The following requests are handled unconditionally, even if we
1686
       * didn't see WL_LATCH_SET.  This gives high priority to shutdown
1687
       * and reload requests where the latch happens to appear later in
1688
       * events[] or will be reported by a later call to
1689
       * WaitEventSetWait().
1690
       */
1691
0
      if (pending_pm_shutdown_request)
1692
0
        process_pm_shutdown_request();
1693
0
      if (pending_pm_reload_request)
1694
0
        process_pm_reload_request();
1695
0
      if (pending_pm_child_exit)
1696
0
        process_pm_child_exit();
1697
0
      if (pending_pm_pmsignal)
1698
0
        process_pm_pmsignal();
1699
1700
0
      if (events[i].events & WL_SOCKET_ACCEPT)
1701
0
      {
1702
0
        ClientSocket s;
1703
1704
0
        if (AcceptConnection(events[i].fd, &s) == STATUS_OK)
1705
0
          BackendStartup(&s);
1706
1707
        /* We no longer need the open socket in this process */
1708
0
        if (s.sock != PGINVALID_SOCKET)
1709
0
        {
1710
0
          if (closesocket(s.sock) != 0)
1711
0
            elog(LOG, "could not close client socket: %m");
1712
0
        }
1713
0
      }
1714
0
    }
1715
1716
    /*
1717
     * If we need to launch any background processes after changing state
1718
     * or because some exited, do so now.
1719
     */
1720
0
    LaunchMissingBackgroundProcesses();
1721
1722
    /* If we need to signal the autovacuum launcher, do so now */
1723
0
    if (avlauncher_needs_signal)
1724
0
    {
1725
0
      avlauncher_needs_signal = false;
1726
0
      if (AutoVacLauncherPMChild != NULL)
1727
0
        signal_child(AutoVacLauncherPMChild, SIGUSR2);
1728
0
    }
1729
1730
#ifdef HAVE_PTHREAD_IS_THREADED_NP
1731
1732
    /*
1733
     * With assertions enabled, check regularly for appearance of
1734
     * additional threads.  All builds check at start and exit.
1735
     */
1736
    Assert(pthread_is_threaded_np() == 0);
1737
#endif
1738
1739
    /*
1740
     * Lastly, check to see if it's time to do some things that we don't
1741
     * want to do every single time through the loop, because they're a
1742
     * bit expensive.  Note that there's up to a minute of slop in when
1743
     * these tasks will be performed, since DetermineSleepTime() will let
1744
     * us sleep at most that long; except for SIGKILL timeout which has
1745
     * special-case logic there.
1746
     */
1747
0
    now = time(NULL);
1748
1749
    /*
1750
     * If we already sent SIGQUIT to children and they are slow to shut
1751
     * down, it's time to send them SIGKILL (or SIGABRT if requested).
1752
     * This doesn't happen normally, but under certain conditions backends
1753
     * can get stuck while shutting down.  This is a last measure to get
1754
     * them unwedged.
1755
     *
1756
     * Note we also do this during recovery from a process crash.
1757
     */
1758
0
    if ((Shutdown >= ImmediateShutdown || FatalError) &&
1759
0
      AbortStartTime != 0 &&
1760
0
      (now - AbortStartTime) >= SIGKILL_CHILDREN_AFTER_SECS)
1761
0
    {
1762
      /* We were gentle with them before. Not anymore */
1763
0
      ereport(LOG,
1764
      /* translator: %s is SIGKILL or SIGABRT */
1765
0
          (errmsg("issuing %s to recalcitrant children",
1766
0
              send_abort_for_kill ? "SIGABRT" : "SIGKILL")));
1767
0
      TerminateChildren(send_abort_for_kill ? SIGABRT : SIGKILL);
1768
      /* reset flag so we don't SIGKILL again */
1769
0
      AbortStartTime = 0;
1770
0
    }
1771
1772
    /*
1773
     * Once a minute, verify that postmaster.pid hasn't been removed or
1774
     * overwritten.  If it has, we force a shutdown.  This avoids having
1775
     * postmasters and child processes hanging around after their database
1776
     * is gone, and maybe causing problems if a new database cluster is
1777
     * created in the same place.  It also provides some protection
1778
     * against a DBA foolishly removing postmaster.pid and manually
1779
     * starting a new postmaster.  Data corruption is likely to ensue from
1780
     * that anyway, but we can minimize the damage by aborting ASAP.
1781
     */
1782
0
    if (now - last_lockfile_recheck_time >= 1 * SECS_PER_MINUTE)
1783
0
    {
1784
0
      if (!RecheckDataDirLockFile())
1785
0
      {
1786
0
        ereport(LOG,
1787
0
            (errmsg("performing immediate shutdown because data directory lock file is invalid")));
1788
0
        kill(MyProcPid, SIGQUIT);
1789
0
      }
1790
0
      last_lockfile_recheck_time = now;
1791
0
    }
1792
1793
    /*
1794
     * Touch Unix socket and lock files every 58 minutes, to ensure that
1795
     * they are not removed by overzealous /tmp-cleaning tasks.  We assume
1796
     * no one runs cleaners with cutoff times of less than an hour ...
1797
     */
1798
0
    if (now - last_touch_time >= 58 * SECS_PER_MINUTE)
1799
0
    {
1800
0
      TouchSocketFiles();
1801
0
      TouchSocketLockFiles();
1802
0
      last_touch_time = now;
1803
0
    }
1804
0
  }
1805
0
}
1806
1807
/*
1808
 * canAcceptConnections --- check to see if database state allows connections
1809
 * of the specified type.  backend_type can be B_BACKEND or B_AUTOVAC_WORKER.
1810
 * (Note that we don't yet know whether a normal B_BACKEND connection might
1811
 * turn into a walsender.)
1812
 */
1813
static CAC_state
1814
canAcceptConnections(BackendType backend_type)
1815
0
{
1816
0
  CAC_state result = CAC_OK;
1817
1818
0
  Assert(backend_type == B_BACKEND || backend_type == B_AUTOVAC_WORKER);
1819
1820
  /*
1821
   * Can't start backends when in startup/shutdown/inconsistent recovery
1822
   * state.  We treat autovac workers the same as user backends for this
1823
   * purpose.
1824
   */
1825
0
  if (pmState != PM_RUN && pmState != PM_HOT_STANDBY)
1826
0
  {
1827
0
    if (Shutdown > NoShutdown)
1828
0
      return CAC_SHUTDOWN; /* shutdown is pending */
1829
0
    else if (!FatalError && pmState == PM_STARTUP)
1830
0
      return CAC_STARTUP; /* normal startup */
1831
0
    else if (!FatalError && pmState == PM_RECOVERY)
1832
0
      return CAC_NOTHOTSTANDBY; /* not yet ready for hot standby */
1833
0
    else
1834
0
      return CAC_RECOVERY; /* else must be crash recovery */
1835
0
  }
1836
1837
  /*
1838
   * "Smart shutdown" restrictions are applied only to normal connections,
1839
   * not to autovac workers.
1840
   */
1841
0
  if (!connsAllowed && backend_type == B_BACKEND)
1842
0
    return CAC_SHUTDOWN; /* shutdown is pending */
1843
1844
0
  return result;
1845
0
}
1846
1847
/*
1848
 * ClosePostmasterPorts -- close all the postmaster's open sockets
1849
 *
1850
 * This is called during child process startup to release file descriptors
1851
 * that are not needed by that child process.  The postmaster still has
1852
 * them open, of course.
1853
 *
1854
 * Note: we pass am_syslogger as a boolean because we don't want to set
1855
 * the global variable yet when this is called.
1856
 */
1857
void
1858
ClosePostmasterPorts(bool am_syslogger)
1859
0
{
1860
  /* Release resources held by the postmaster's WaitEventSet. */
1861
0
  if (pm_wait_set)
1862
0
  {
1863
0
    FreeWaitEventSetAfterFork(pm_wait_set);
1864
0
    pm_wait_set = NULL;
1865
0
  }
1866
1867
0
#ifndef WIN32
1868
1869
  /*
1870
   * Close the write end of postmaster death watch pipe. It's important to
1871
   * do this as early as possible, so that if postmaster dies, others won't
1872
   * think that it's still running because we're holding the pipe open.
1873
   */
1874
0
  if (close(postmaster_alive_fds[POSTMASTER_FD_OWN]) != 0)
1875
0
    ereport(FATAL,
1876
0
        (errcode_for_file_access(),
1877
0
         errmsg_internal("could not close postmaster death monitoring pipe in child process: %m")));
1878
0
  postmaster_alive_fds[POSTMASTER_FD_OWN] = -1;
1879
  /* Notify fd.c that we released one pipe FD. */
1880
0
  ReleaseExternalFD();
1881
0
#endif
1882
1883
  /*
1884
   * Close the postmaster's listen sockets.  These aren't tracked by fd.c,
1885
   * so we don't call ReleaseExternalFD() here.
1886
   *
1887
   * The listen sockets are marked as FD_CLOEXEC, so this isn't needed in
1888
   * EXEC_BACKEND mode.
1889
   */
1890
0
#ifndef EXEC_BACKEND
1891
0
  if (ListenSockets)
1892
0
  {
1893
0
    for (int i = 0; i < NumListenSockets; i++)
1894
0
    {
1895
0
      if (closesocket(ListenSockets[i]) != 0)
1896
0
        elog(LOG, "could not close listen socket: %m");
1897
0
    }
1898
0
    pfree(ListenSockets);
1899
0
  }
1900
0
  NumListenSockets = 0;
1901
0
  ListenSockets = NULL;
1902
0
#endif
1903
1904
  /*
1905
   * If using syslogger, close the read side of the pipe.  We don't bother
1906
   * tracking this in fd.c, either.
1907
   */
1908
0
  if (!am_syslogger)
1909
0
  {
1910
0
#ifndef WIN32
1911
0
    if (syslogPipe[0] >= 0)
1912
0
      close(syslogPipe[0]);
1913
0
    syslogPipe[0] = -1;
1914
#else
1915
    if (syslogPipe[0])
1916
      CloseHandle(syslogPipe[0]);
1917
    syslogPipe[0] = 0;
1918
#endif
1919
0
  }
1920
1921
#ifdef USE_BONJOUR
1922
  /* If using Bonjour, close the connection to the mDNS daemon */
1923
  if (bonjour_sdref)
1924
    close(DNSServiceRefSockFD(bonjour_sdref));
1925
#endif
1926
0
}
1927
1928
1929
/*
1930
 * InitProcessGlobals -- set MyStartTime[stamp], random seeds
1931
 *
1932
 * Called early in the postmaster and every backend.
1933
 */
1934
void
1935
InitProcessGlobals(void)
1936
0
{
1937
0
  MyStartTimestamp = GetCurrentTimestamp();
1938
0
  MyStartTime = timestamptz_to_time_t(MyStartTimestamp);
1939
1940
  /*
1941
   * Set a different global seed in every process.  We want something
1942
   * unpredictable, so if possible, use high-quality random bits for the
1943
   * seed.  Otherwise, fall back to a seed based on timestamp and PID.
1944
   */
1945
0
  if (unlikely(!pg_prng_strong_seed(&pg_global_prng_state)))
1946
0
  {
1947
0
    uint64    rseed;
1948
1949
    /*
1950
     * Since PIDs and timestamps tend to change more frequently in their
1951
     * least significant bits, shift the timestamp left to allow a larger
1952
     * total number of seeds in a given time period.  Since that would
1953
     * leave only 20 bits of the timestamp that cycle every ~1 second,
1954
     * also mix in some higher bits.
1955
     */
1956
0
    rseed = ((uint64) MyProcPid) ^
1957
0
      ((uint64) MyStartTimestamp << 12) ^
1958
0
      ((uint64) MyStartTimestamp >> 20);
1959
1960
0
    pg_prng_seed(&pg_global_prng_state, rseed);
1961
0
  }
1962
1963
  /*
1964
   * Also make sure that we've set a good seed for random(3).  Use of that
1965
   * is deprecated in core Postgres, but extensions might use it.
1966
   */
1967
0
#ifndef WIN32
1968
0
  srandom(pg_prng_uint32(&pg_global_prng_state));
1969
0
#endif
1970
0
}
1971
1972
/*
1973
 * Child processes use SIGUSR1 to notify us of 'pmsignals'.  pg_ctl uses
1974
 * SIGUSR1 to ask postmaster to check for logrotate and promote files.
1975
 */
1976
static void
1977
handle_pm_pmsignal_signal(SIGNAL_ARGS)
1978
0
{
1979
0
  pending_pm_pmsignal = true;
1980
0
  SetLatch(MyLatch);
1981
0
}
1982
1983
/*
1984
 * pg_ctl uses SIGHUP to request a reload of the configuration files.
1985
 */
1986
static void
1987
handle_pm_reload_request_signal(SIGNAL_ARGS)
1988
0
{
1989
0
  pending_pm_reload_request = true;
1990
0
  SetLatch(MyLatch);
1991
0
}
1992
1993
/*
1994
 * Re-read config files, and tell children to do same.
1995
 */
1996
static void
1997
process_pm_reload_request(void)
1998
{
1999
  pending_pm_reload_request = false;
2000
2001
  ereport(DEBUG2,
2002
      (errmsg_internal("postmaster received reload request signal")));
2003
2004
  if (Shutdown <= SmartShutdown)
2005
  {
2006
    ereport(LOG,
2007
        (errmsg("received SIGHUP, reloading configuration files")));
2008
    ProcessConfigFile(PGC_SIGHUP);
2009
    SignalChildren(SIGHUP, btmask_all_except(B_DEAD_END_BACKEND));
2010
2011
    /* Reload authentication config files too */
2012
    if (!load_hba())
2013
      ereport(LOG,
2014
      /* translator: %s is a configuration file */
2015
          (errmsg("%s was not reloaded", HbaFileName)));
2016
2017
    if (!load_ident())
2018
      ereport(LOG,
2019
          (errmsg("%s was not reloaded", IdentFileName)));
2020
2021
#ifdef USE_SSL
2022
    /* Reload SSL configuration as well */
2023
    if (EnableSSL)
2024
    {
2025
      if (secure_initialize(false) == 0)
2026
        LoadedSSL = true;
2027
      else
2028
        ereport(LOG,
2029
            (errmsg("SSL configuration was not reloaded")));
2030
    }
2031
    else
2032
    {
2033
      secure_destroy();
2034
      LoadedSSL = false;
2035
    }
2036
#endif
2037
2038
#ifdef EXEC_BACKEND
2039
    /* Update the starting-point file for future children */
2040
    write_nondefault_variables(PGC_SIGHUP);
2041
#endif
2042
  }
2043
}
2044
2045
/*
2046
 * pg_ctl uses SIGTERM, SIGINT and SIGQUIT to request different types of
2047
 * shutdown.
2048
 */
2049
static void
2050
handle_pm_shutdown_request_signal(SIGNAL_ARGS)
2051
0
{
2052
0
  switch (postgres_signal_arg)
2053
0
  {
2054
0
    case SIGTERM:
2055
      /* smart is implied if the other two flags aren't set */
2056
0
      pending_pm_shutdown_request = true;
2057
0
      break;
2058
0
    case SIGINT:
2059
0
      pending_pm_fast_shutdown_request = true;
2060
0
      pending_pm_shutdown_request = true;
2061
0
      break;
2062
0
    case SIGQUIT:
2063
0
      pending_pm_immediate_shutdown_request = true;
2064
0
      pending_pm_shutdown_request = true;
2065
0
      break;
2066
0
  }
2067
0
  SetLatch(MyLatch);
2068
0
}
2069
2070
/*
2071
 * Process shutdown request.
2072
 */
2073
static void
2074
process_pm_shutdown_request(void)
2075
{
2076
  int     mode;
2077
2078
  ereport(DEBUG2,
2079
      (errmsg_internal("postmaster received shutdown request signal")));
2080
2081
  pending_pm_shutdown_request = false;
2082
2083
  /*
2084
   * If more than one shutdown request signal arrived since the last server
2085
   * loop, take the one that is the most immediate.  That matches the
2086
   * priority that would apply if we processed them one by one in any order.
2087
   */
2088
  if (pending_pm_immediate_shutdown_request)
2089
  {
2090
    pending_pm_immediate_shutdown_request = false;
2091
    pending_pm_fast_shutdown_request = false;
2092
    mode = ImmediateShutdown;
2093
  }
2094
  else if (pending_pm_fast_shutdown_request)
2095
  {
2096
    pending_pm_fast_shutdown_request = false;
2097
    mode = FastShutdown;
2098
  }
2099
  else
2100
    mode = SmartShutdown;
2101
2102
  switch (mode)
2103
  {
2104
    case SmartShutdown:
2105
2106
      /*
2107
       * Smart Shutdown:
2108
       *
2109
       * Wait for children to end their work, then shut down.
2110
       */
2111
      if (Shutdown >= SmartShutdown)
2112
        break;
2113
      Shutdown = SmartShutdown;
2114
      ereport(LOG,
2115
          (errmsg("received smart shutdown request")));
2116
2117
      /* Report status */
2118
      AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2119
#ifdef USE_SYSTEMD
2120
      sd_notify(0, "STOPPING=1");
2121
#endif
2122
2123
      /*
2124
       * If we reached normal running, we go straight to waiting for
2125
       * client backends to exit.  If already in PM_STOP_BACKENDS or a
2126
       * later state, do not change it.
2127
       */
2128
      if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
2129
        connsAllowed = false;
2130
      else if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2131
      {
2132
        /* There should be no clients, so proceed to stop children */
2133
        UpdatePMState(PM_STOP_BACKENDS);
2134
      }
2135
2136
      /*
2137
       * Now wait for online backup mode to end and backends to exit. If
2138
       * that is already the case, PostmasterStateMachine will take the
2139
       * next step.
2140
       */
2141
      PostmasterStateMachine();
2142
      break;
2143
2144
    case FastShutdown:
2145
2146
      /*
2147
       * Fast Shutdown:
2148
       *
2149
       * Abort all children with SIGTERM (rollback active transactions
2150
       * and exit) and shut down when they are gone.
2151
       */
2152
      if (Shutdown >= FastShutdown)
2153
        break;
2154
      Shutdown = FastShutdown;
2155
      ereport(LOG,
2156
          (errmsg("received fast shutdown request")));
2157
2158
      /* Report status */
2159
      AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2160
#ifdef USE_SYSTEMD
2161
      sd_notify(0, "STOPPING=1");
2162
#endif
2163
2164
      if (pmState == PM_STARTUP || pmState == PM_RECOVERY)
2165
      {
2166
        /* Just shut down background processes silently */
2167
        UpdatePMState(PM_STOP_BACKENDS);
2168
      }
2169
      else if (pmState == PM_RUN ||
2170
           pmState == PM_HOT_STANDBY)
2171
      {
2172
        /* Report that we're about to zap live client sessions */
2173
        ereport(LOG,
2174
            (errmsg("aborting any active transactions")));
2175
        UpdatePMState(PM_STOP_BACKENDS);
2176
      }
2177
2178
      /*
2179
       * PostmasterStateMachine will issue any necessary signals, or
2180
       * take the next step if no child processes need to be killed.
2181
       */
2182
      PostmasterStateMachine();
2183
      break;
2184
2185
    case ImmediateShutdown:
2186
2187
      /*
2188
       * Immediate Shutdown:
2189
       *
2190
       * abort all children with SIGQUIT, wait for them to exit,
2191
       * terminate remaining ones with SIGKILL, then exit without
2192
       * attempt to properly shut down the data base system.
2193
       */
2194
      if (Shutdown >= ImmediateShutdown)
2195
        break;
2196
      Shutdown = ImmediateShutdown;
2197
      ereport(LOG,
2198
          (errmsg("received immediate shutdown request")));
2199
2200
      /* Report status */
2201
      AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STOPPING);
2202
#ifdef USE_SYSTEMD
2203
      sd_notify(0, "STOPPING=1");
2204
#endif
2205
2206
      /* tell children to shut down ASAP */
2207
      /* (note we don't apply send_abort_for_crash here) */
2208
      SetQuitSignalReason(PMQUIT_FOR_STOP);
2209
      TerminateChildren(SIGQUIT);
2210
      UpdatePMState(PM_WAIT_BACKENDS);
2211
2212
      /* set stopwatch for them to die */
2213
      AbortStartTime = time(NULL);
2214
2215
      /*
2216
       * Now wait for backends to exit.  If there are none,
2217
       * PostmasterStateMachine will take the next step.
2218
       */
2219
      PostmasterStateMachine();
2220
      break;
2221
  }
2222
}
2223
2224
static void
2225
handle_pm_child_exit_signal(SIGNAL_ARGS)
2226
0
{
2227
0
  pending_pm_child_exit = true;
2228
0
  SetLatch(MyLatch);
2229
0
}
2230
2231
/*
2232
 * Cleanup after a child process dies.
2233
 */
2234
static void
2235
process_pm_child_exit(void)
2236
0
{
2237
0
  int     pid;      /* process id of dead child process */
2238
0
  int     exitstatus;   /* its exit status */
2239
2240
0
  pending_pm_child_exit = false;
2241
2242
0
  ereport(DEBUG4,
2243
0
      (errmsg_internal("reaping dead processes")));
2244
2245
0
  while ((pid = waitpid(-1, &exitstatus, WNOHANG)) > 0)
2246
0
  {
2247
0
    PMChild    *pmchild;
2248
2249
    /*
2250
     * Check if this child was a startup process.
2251
     */
2252
0
    if (StartupPMChild && pid == StartupPMChild->pid)
2253
0
    {
2254
0
      ReleasePostmasterChildSlot(StartupPMChild);
2255
0
      StartupPMChild = NULL;
2256
2257
      /*
2258
       * Startup process exited in response to a shutdown request (or it
2259
       * completed normally regardless of the shutdown request).
2260
       */
2261
0
      if (Shutdown > NoShutdown &&
2262
0
        (EXIT_STATUS_0(exitstatus) || EXIT_STATUS_1(exitstatus)))
2263
0
      {
2264
0
        StartupStatus = STARTUP_NOT_RUNNING;
2265
0
        UpdatePMState(PM_WAIT_BACKENDS);
2266
        /* PostmasterStateMachine logic does the rest */
2267
0
        continue;
2268
0
      }
2269
2270
0
      if (EXIT_STATUS_3(exitstatus))
2271
0
      {
2272
0
        ereport(LOG,
2273
0
            (errmsg("shutdown at recovery target")));
2274
0
        StartupStatus = STARTUP_NOT_RUNNING;
2275
0
        Shutdown = Max(Shutdown, SmartShutdown);
2276
0
        TerminateChildren(SIGTERM);
2277
0
        UpdatePMState(PM_WAIT_BACKENDS);
2278
        /* PostmasterStateMachine logic does the rest */
2279
0
        continue;
2280
0
      }
2281
2282
      /*
2283
       * Unexpected exit of startup process (including FATAL exit)
2284
       * during PM_STARTUP is treated as catastrophic. There are no
2285
       * other processes running yet, so we can just exit.
2286
       */
2287
0
      if (pmState == PM_STARTUP &&
2288
0
        StartupStatus != STARTUP_SIGNALED &&
2289
0
        !EXIT_STATUS_0(exitstatus))
2290
0
      {
2291
0
        LogChildExit(LOG, _("startup process"),
2292
0
               pid, exitstatus);
2293
0
        ereport(LOG,
2294
0
            (errmsg("aborting startup due to startup process failure")));
2295
0
        ExitPostmaster(1);
2296
0
      }
2297
2298
      /*
2299
       * After PM_STARTUP, any unexpected exit (including FATAL exit) of
2300
       * the startup process is catastrophic, so kill other children,
2301
       * and set StartupStatus so we don't try to reinitialize after
2302
       * they're gone.  Exception: if StartupStatus is STARTUP_SIGNALED,
2303
       * then we previously sent the startup process a SIGQUIT; so
2304
       * that's probably the reason it died, and we do want to try to
2305
       * restart in that case.
2306
       *
2307
       * This stanza also handles the case where we sent a SIGQUIT
2308
       * during PM_STARTUP due to some dead-end child crashing: in that
2309
       * situation, if the startup process dies on the SIGQUIT, we need
2310
       * to transition to PM_WAIT_BACKENDS state which will allow
2311
       * PostmasterStateMachine to restart the startup process.  (On the
2312
       * other hand, the startup process might complete normally, if we
2313
       * were too late with the SIGQUIT.  In that case we'll fall
2314
       * through and commence normal operations.)
2315
       */
2316
0
      if (!EXIT_STATUS_0(exitstatus))
2317
0
      {
2318
0
        if (StartupStatus == STARTUP_SIGNALED)
2319
0
        {
2320
0
          StartupStatus = STARTUP_NOT_RUNNING;
2321
0
          if (pmState == PM_STARTUP)
2322
0
            UpdatePMState(PM_WAIT_BACKENDS);
2323
0
        }
2324
0
        else
2325
0
          StartupStatus = STARTUP_CRASHED;
2326
0
        HandleChildCrash(pid, exitstatus,
2327
0
                 _("startup process"));
2328
0
        continue;
2329
0
      }
2330
2331
      /*
2332
       * Startup succeeded, commence normal operations
2333
       */
2334
0
      StartupStatus = STARTUP_NOT_RUNNING;
2335
0
      FatalError = false;
2336
0
      AbortStartTime = 0;
2337
0
      ReachedNormalRunning = true;
2338
0
      UpdatePMState(PM_RUN);
2339
0
      connsAllowed = true;
2340
2341
      /*
2342
       * At the next iteration of the postmaster's main loop, we will
2343
       * crank up the background tasks like the autovacuum launcher and
2344
       * background workers that were not started earlier already.
2345
       */
2346
0
      StartWorkerNeeded = true;
2347
2348
      /* at this point we are really open for business */
2349
0
      ereport(LOG,
2350
0
          (errmsg("database system is ready to accept connections")));
2351
2352
      /* Report status */
2353
0
      AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
2354
#ifdef USE_SYSTEMD
2355
      sd_notify(0, "READY=1");
2356
#endif
2357
2358
0
      continue;
2359
0
    }
2360
2361
    /*
2362
     * Was it the bgwriter?  Normal exit can be ignored; we'll start a new
2363
     * one at the next iteration of the postmaster's main loop, if
2364
     * necessary.  Any other exit condition is treated as a crash.
2365
     */
2366
0
    if (BgWriterPMChild && pid == BgWriterPMChild->pid)
2367
0
    {
2368
0
      ReleasePostmasterChildSlot(BgWriterPMChild);
2369
0
      BgWriterPMChild = NULL;
2370
0
      if (!EXIT_STATUS_0(exitstatus))
2371
0
        HandleChildCrash(pid, exitstatus,
2372
0
                 _("background writer process"));
2373
0
      continue;
2374
0
    }
2375
2376
    /*
2377
     * Was it the checkpointer?
2378
     */
2379
0
    if (CheckpointerPMChild && pid == CheckpointerPMChild->pid)
2380
0
    {
2381
0
      ReleasePostmasterChildSlot(CheckpointerPMChild);
2382
0
      CheckpointerPMChild = NULL;
2383
0
      if (EXIT_STATUS_0(exitstatus) && pmState == PM_WAIT_CHECKPOINTER)
2384
0
      {
2385
        /*
2386
         * OK, we saw normal exit of the checkpointer after it's been
2387
         * told to shut down.  We know checkpointer wrote a shutdown
2388
         * checkpoint, otherwise we'd still be in
2389
         * PM_WAIT_XLOG_SHUTDOWN state.
2390
         *
2391
         * At this point only dead-end children and logger should be
2392
         * left.
2393
         */
2394
0
        UpdatePMState(PM_WAIT_DEAD_END);
2395
0
        ConfigurePostmasterWaitSet(false);
2396
0
        SignalChildren(SIGTERM, btmask_all_except(B_LOGGER));
2397
0
      }
2398
0
      else
2399
0
      {
2400
        /*
2401
         * Any unexpected exit of the checkpointer (including FATAL
2402
         * exit) is treated as a crash.
2403
         */
2404
0
        HandleChildCrash(pid, exitstatus,
2405
0
                 _("checkpointer process"));
2406
0
      }
2407
2408
0
      continue;
2409
0
    }
2410
2411
    /*
2412
     * Was it the wal writer?  Normal exit can be ignored; we'll start a
2413
     * new one at the next iteration of the postmaster's main loop, if
2414
     * necessary.  Any other exit condition is treated as a crash.
2415
     */
2416
0
    if (WalWriterPMChild && pid == WalWriterPMChild->pid)
2417
0
    {
2418
0
      ReleasePostmasterChildSlot(WalWriterPMChild);
2419
0
      WalWriterPMChild = NULL;
2420
0
      if (!EXIT_STATUS_0(exitstatus))
2421
0
        HandleChildCrash(pid, exitstatus,
2422
0
                 _("WAL writer process"));
2423
0
      continue;
2424
0
    }
2425
2426
    /*
2427
     * Was it the wal receiver?  If exit status is zero (normal) or one
2428
     * (FATAL exit), we assume everything is all right just like normal
2429
     * backends.  (If we need a new wal receiver, we'll start one at the
2430
     * next iteration of the postmaster's main loop.)
2431
     */
2432
0
    if (WalReceiverPMChild && pid == WalReceiverPMChild->pid)
2433
0
    {
2434
0
      ReleasePostmasterChildSlot(WalReceiverPMChild);
2435
0
      WalReceiverPMChild = NULL;
2436
0
      if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
2437
0
        HandleChildCrash(pid, exitstatus,
2438
0
                 _("WAL receiver process"));
2439
0
      continue;
2440
0
    }
2441
2442
    /*
2443
     * Was it the wal summarizer? Normal exit can be ignored; we'll start
2444
     * a new one at the next iteration of the postmaster's main loop, if
2445
     * necessary.  Any other exit condition is treated as a crash.
2446
     */
2447
0
    if (WalSummarizerPMChild && pid == WalSummarizerPMChild->pid)
2448
0
    {
2449
0
      ReleasePostmasterChildSlot(WalSummarizerPMChild);
2450
0
      WalSummarizerPMChild = NULL;
2451
0
      if (!EXIT_STATUS_0(exitstatus))
2452
0
        HandleChildCrash(pid, exitstatus,
2453
0
                 _("WAL summarizer process"));
2454
0
      continue;
2455
0
    }
2456
2457
    /*
2458
     * Was it the autovacuum launcher?  Normal exit can be ignored; we'll
2459
     * start a new one at the next iteration of the postmaster's main
2460
     * loop, if necessary.  Any other exit condition is treated as a
2461
     * crash.
2462
     */
2463
0
    if (AutoVacLauncherPMChild && pid == AutoVacLauncherPMChild->pid)
2464
0
    {
2465
0
      ReleasePostmasterChildSlot(AutoVacLauncherPMChild);
2466
0
      AutoVacLauncherPMChild = NULL;
2467
0
      if (!EXIT_STATUS_0(exitstatus))
2468
0
        HandleChildCrash(pid, exitstatus,
2469
0
                 _("autovacuum launcher process"));
2470
0
      continue;
2471
0
    }
2472
2473
    /*
2474
     * Was it the archiver?  If exit status is zero (normal) or one (FATAL
2475
     * exit), we assume everything is all right just like normal backends
2476
     * and just try to start a new one on the next cycle of the
2477
     * postmaster's main loop, to retry archiving remaining files.
2478
     */
2479
0
    if (PgArchPMChild && pid == PgArchPMChild->pid)
2480
0
    {
2481
0
      ReleasePostmasterChildSlot(PgArchPMChild);
2482
0
      PgArchPMChild = NULL;
2483
0
      if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
2484
0
        HandleChildCrash(pid, exitstatus,
2485
0
                 _("archiver process"));
2486
0
      continue;
2487
0
    }
2488
2489
    /* Was it the system logger?  If so, try to start a new one */
2490
0
    if (SysLoggerPMChild && pid == SysLoggerPMChild->pid)
2491
0
    {
2492
0
      ReleasePostmasterChildSlot(SysLoggerPMChild);
2493
0
      SysLoggerPMChild = NULL;
2494
2495
      /* for safety's sake, launch new logger *first* */
2496
0
      if (Logging_collector)
2497
0
        StartSysLogger();
2498
2499
0
      if (!EXIT_STATUS_0(exitstatus))
2500
0
        LogChildExit(LOG, _("system logger process"),
2501
0
               pid, exitstatus);
2502
0
      continue;
2503
0
    }
2504
2505
    /*
2506
     * Was it the slot sync worker? Normal exit or FATAL exit can be
2507
     * ignored (FATAL can be caused by libpqwalreceiver on receiving
2508
     * shutdown request by the startup process during promotion); we'll
2509
     * start a new one at the next iteration of the postmaster's main
2510
     * loop, if necessary. Any other exit condition is treated as a crash.
2511
     */
2512
0
    if (SlotSyncWorkerPMChild && pid == SlotSyncWorkerPMChild->pid)
2513
0
    {
2514
0
      ReleasePostmasterChildSlot(SlotSyncWorkerPMChild);
2515
0
      SlotSyncWorkerPMChild = NULL;
2516
0
      if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
2517
0
        HandleChildCrash(pid, exitstatus,
2518
0
                 _("slot sync worker process"));
2519
0
      continue;
2520
0
    }
2521
2522
    /* Was it an IO worker? */
2523
0
    if (maybe_reap_io_worker(pid))
2524
0
    {
2525
0
      if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
2526
0
        HandleChildCrash(pid, exitstatus, _("io worker"));
2527
2528
0
      maybe_adjust_io_workers();
2529
0
      continue;
2530
0
    }
2531
2532
    /*
2533
     * Was it a backend or a background worker?
2534
     */
2535
0
    pmchild = FindPostmasterChildByPid(pid);
2536
0
    if (pmchild)
2537
0
    {
2538
0
      CleanupBackend(pmchild, exitstatus);
2539
0
    }
2540
2541
    /*
2542
     * We don't know anything about this child process.  That's highly
2543
     * unexpected, as we do track all the child processes that we fork.
2544
     */
2545
0
    else
2546
0
    {
2547
0
      if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
2548
0
        HandleChildCrash(pid, exitstatus, _("untracked child process"));
2549
0
      else
2550
0
        LogChildExit(LOG, _("untracked child process"), pid, exitstatus);
2551
0
    }
2552
0
  }             /* loop over pending child-death reports */
2553
2554
  /*
2555
   * After cleaning out the SIGCHLD queue, see if we have any state changes
2556
   * or actions to make.
2557
   */
2558
0
  PostmasterStateMachine();
2559
0
}
2560
2561
/*
2562
 * CleanupBackend -- cleanup after terminated backend or background worker.
2563
 *
2564
 * Remove all local state associated with the child process and release its
2565
 * PMChild slot.
2566
 */
2567
static void
2568
CleanupBackend(PMChild *bp,
2569
         int exitstatus)  /* child's exit status. */
2570
0
{
2571
0
  char    namebuf[MAXPGPATH];
2572
0
  const char *procname;
2573
0
  bool    crashed = false;
2574
0
  bool    logged = false;
2575
0
  pid_t   bp_pid;
2576
0
  bool    bp_bgworker_notify;
2577
0
  BackendType bp_bkend_type;
2578
0
  RegisteredBgWorker *rw;
2579
2580
  /* Construct a process name for the log message */
2581
0
  if (bp->bkend_type == B_BG_WORKER)
2582
0
  {
2583
0
    snprintf(namebuf, MAXPGPATH, _("background worker \"%s\""),
2584
0
         bp->rw->rw_worker.bgw_type);
2585
0
    procname = namebuf;
2586
0
  }
2587
0
  else
2588
0
    procname = _(GetBackendTypeDesc(bp->bkend_type));
2589
2590
  /*
2591
   * If a backend dies in an ugly way then we must signal all other backends
2592
   * to quickdie.  If exit status is zero (normal) or one (FATAL exit), we
2593
   * assume everything is all right and proceed to remove the backend from
2594
   * the active child list.
2595
   */
2596
0
  if (!EXIT_STATUS_0(exitstatus) && !EXIT_STATUS_1(exitstatus))
2597
0
    crashed = true;
2598
2599
#ifdef WIN32
2600
2601
  /*
2602
   * On win32, also treat ERROR_WAIT_NO_CHILDREN (128) as nonfatal case,
2603
   * since that sometimes happens under load when the process fails to start
2604
   * properly (long before it starts using shared memory). Microsoft reports
2605
   * it is related to mutex failure:
2606
   * http://archives.postgresql.org/pgsql-hackers/2010-09/msg00790.php
2607
   */
2608
  if (exitstatus == ERROR_WAIT_NO_CHILDREN)
2609
  {
2610
    LogChildExit(LOG, procname, bp->pid, exitstatus);
2611
    logged = true;
2612
    crashed = false;
2613
  }
2614
#endif
2615
2616
  /*
2617
   * Release the PMChild entry.
2618
   *
2619
   * If the process attached to shared memory, this also checks that it
2620
   * detached cleanly.
2621
   */
2622
0
  bp_pid = bp->pid;
2623
0
  bp_bgworker_notify = bp->bgworker_notify;
2624
0
  bp_bkend_type = bp->bkend_type;
2625
0
  rw = bp->rw;
2626
0
  if (!ReleasePostmasterChildSlot(bp))
2627
0
  {
2628
    /*
2629
     * Uh-oh, the child failed to clean itself up.  Treat as a crash after
2630
     * all.
2631
     */
2632
0
    crashed = true;
2633
0
  }
2634
0
  bp = NULL;
2635
2636
  /*
2637
   * In a crash case, exit immediately without resetting background worker
2638
   * state. However, if restart_after_crash is enabled, the background
2639
   * worker state (e.g., rw_pid) still needs be reset so the worker can
2640
   * restart after crash recovery. This reset is handled in
2641
   * ResetBackgroundWorkerCrashTimes(), not here.
2642
   */
2643
0
  if (crashed)
2644
0
  {
2645
0
    HandleChildCrash(bp_pid, exitstatus, procname);
2646
0
    return;
2647
0
  }
2648
2649
  /*
2650
   * This backend may have been slated to receive SIGUSR1 when some
2651
   * background worker started or stopped.  Cancel those notifications, as
2652
   * we don't want to signal PIDs that are not PostgreSQL backends.  This
2653
   * gets skipped in the (probably very common) case where the backend has
2654
   * never requested any such notifications.
2655
   */
2656
0
  if (bp_bgworker_notify)
2657
0
    BackgroundWorkerStopNotifications(bp_pid);
2658
2659
  /*
2660
   * If it was a background worker, also update its RegisteredBgWorker
2661
   * entry.
2662
   */
2663
0
  if (bp_bkend_type == B_BG_WORKER)
2664
0
  {
2665
0
    if (!EXIT_STATUS_0(exitstatus))
2666
0
    {
2667
      /* Record timestamp, so we know when to restart the worker. */
2668
0
      rw->rw_crashed_at = GetCurrentTimestamp();
2669
0
    }
2670
0
    else
2671
0
    {
2672
      /* Zero exit status means terminate */
2673
0
      rw->rw_crashed_at = 0;
2674
0
      rw->rw_terminate = true;
2675
0
    }
2676
2677
0
    rw->rw_pid = 0;
2678
0
    ReportBackgroundWorkerExit(rw); /* report child death */
2679
2680
0
    if (!logged)
2681
0
    {
2682
0
      LogChildExit(EXIT_STATUS_0(exitstatus) ? DEBUG1 : LOG,
2683
0
             procname, bp_pid, exitstatus);
2684
0
      logged = true;
2685
0
    }
2686
2687
    /* have it be restarted */
2688
0
    HaveCrashedWorker = true;
2689
0
  }
2690
2691
0
  if (!logged)
2692
0
    LogChildExit(DEBUG2, procname, bp_pid, exitstatus);
2693
0
}
2694
2695
/*
2696
 * Transition into FatalError state, in response to something bad having
2697
 * happened. Commonly the caller will have logged the reason for entering
2698
 * FatalError state.
2699
 *
2700
 * This should only be called when not already in FatalError or
2701
 * ImmediateShutdown state.
2702
 */
2703
static void
2704
HandleFatalError(QuitSignalReason reason, bool consider_sigabrt)
2705
0
{
2706
0
  int     sigtosend;
2707
2708
0
  Assert(!FatalError);
2709
0
  Assert(Shutdown != ImmediateShutdown);
2710
2711
0
  SetQuitSignalReason(reason);
2712
2713
0
  if (consider_sigabrt && send_abort_for_crash)
2714
0
    sigtosend = SIGABRT;
2715
0
  else
2716
0
    sigtosend = SIGQUIT;
2717
2718
  /*
2719
   * Signal all other child processes to exit.
2720
   *
2721
   * We could exclude dead-end children here, but at least when sending
2722
   * SIGABRT it seems better to include them.
2723
   */
2724
0
  TerminateChildren(sigtosend);
2725
2726
0
  FatalError = true;
2727
2728
  /*
2729
   * Choose the appropriate new state to react to the fatal error. Unless we
2730
   * were already in the process of shutting down, we go through
2731
   * PM_WAIT_BACKENDS. For errors during the shutdown sequence, we directly
2732
   * switch to PM_WAIT_DEAD_END.
2733
   */
2734
0
  switch (pmState)
2735
0
  {
2736
0
    case PM_INIT:
2737
      /* shouldn't have any children */
2738
0
      Assert(false);
2739
0
      break;
2740
0
    case PM_STARTUP:
2741
      /* should have been handled in process_pm_child_exit */
2742
0
      Assert(false);
2743
0
      break;
2744
2745
      /* wait for children to die */
2746
0
    case PM_RECOVERY:
2747
0
    case PM_HOT_STANDBY:
2748
0
    case PM_RUN:
2749
0
    case PM_STOP_BACKENDS:
2750
0
      UpdatePMState(PM_WAIT_BACKENDS);
2751
0
      break;
2752
2753
0
    case PM_WAIT_BACKENDS:
2754
      /* there might be more backends to wait for */
2755
0
      break;
2756
2757
0
    case PM_WAIT_XLOG_SHUTDOWN:
2758
0
    case PM_WAIT_XLOG_ARCHIVAL:
2759
0
    case PM_WAIT_CHECKPOINTER:
2760
0
    case PM_WAIT_IO_WORKERS:
2761
2762
      /*
2763
       * NB: Similar code exists in PostmasterStateMachine()'s handling
2764
       * of FatalError in PM_STOP_BACKENDS/PM_WAIT_BACKENDS states.
2765
       */
2766
0
      ConfigurePostmasterWaitSet(false);
2767
0
      UpdatePMState(PM_WAIT_DEAD_END);
2768
0
      break;
2769
2770
0
    case PM_WAIT_DEAD_END:
2771
0
    case PM_NO_CHILDREN:
2772
0
      break;
2773
0
  }
2774
2775
  /*
2776
   * .. and if this doesn't happen quickly enough, now the clock is ticking
2777
   * for us to kill them without mercy.
2778
   */
2779
0
  if (AbortStartTime == 0)
2780
0
    AbortStartTime = time(NULL);
2781
0
}
2782
2783
/*
2784
 * HandleChildCrash -- cleanup after failed backend, bgwriter, checkpointer,
2785
 * walwriter, autovacuum, archiver, slot sync worker, or background worker.
2786
 *
2787
 * The objectives here are to clean up our local state about the child
2788
 * process, and to signal all other remaining children to quickdie.
2789
 *
2790
 * The caller has already released its PMChild slot.
2791
 */
2792
static void
2793
HandleChildCrash(int pid, int exitstatus, const char *procname)
2794
{
2795
  /*
2796
   * We only log messages and send signals if this is the first process
2797
   * crash and we're not doing an immediate shutdown; otherwise, we're only
2798
   * here to update postmaster's idea of live processes.  If we have already
2799
   * signaled children, nonzero exit status is to be expected, so don't
2800
   * clutter log.
2801
   */
2802
  if (FatalError || Shutdown == ImmediateShutdown)
2803
    return;
2804
2805
  LogChildExit(LOG, procname, pid, exitstatus);
2806
  ereport(LOG,
2807
      (errmsg("terminating any other active server processes")));
2808
2809
  /*
2810
   * Switch into error state. The crashed process has already been removed
2811
   * from ActiveChildList.
2812
   */
2813
  HandleFatalError(PMQUIT_FOR_CRASH, true);
2814
}
2815
2816
/*
2817
 * Log the death of a child process.
2818
 */
2819
static void
2820
LogChildExit(int lev, const char *procname, int pid, int exitstatus)
2821
0
{
2822
  /*
2823
   * size of activity_buffer is arbitrary, but set equal to default
2824
   * track_activity_query_size
2825
   */
2826
0
  char    activity_buffer[1024];
2827
0
  const char *activity = NULL;
2828
2829
0
  if (!EXIT_STATUS_0(exitstatus))
2830
0
    activity = pgstat_get_crashed_backend_activity(pid,
2831
0
                             activity_buffer,
2832
0
                             sizeof(activity_buffer));
2833
2834
0
  if (WIFEXITED(exitstatus))
2835
0
    ereport(lev,
2836
2837
    /*------
2838
      translator: %s is a noun phrase describing a child process, such as
2839
      "server process" */
2840
0
        (errmsg("%s (PID %d) exited with exit code %d",
2841
0
            procname, pid, WEXITSTATUS(exitstatus)),
2842
0
         activity ? errdetail("Failed process was running: %s", activity) : 0));
2843
0
  else if (WIFSIGNALED(exitstatus))
2844
0
  {
2845
#if defined(WIN32)
2846
    ereport(lev,
2847
2848
    /*------
2849
      translator: %s is a noun phrase describing a child process, such as
2850
      "server process" */
2851
        (errmsg("%s (PID %d) was terminated by exception 0x%X",
2852
            procname, pid, WTERMSIG(exitstatus)),
2853
         errhint("See C include file \"ntstatus.h\" for a description of the hexadecimal value."),
2854
         activity ? errdetail("Failed process was running: %s", activity) : 0));
2855
#else
2856
0
    ereport(lev,
2857
2858
    /*------
2859
      translator: %s is a noun phrase describing a child process, such as
2860
      "server process" */
2861
0
        (errmsg("%s (PID %d) was terminated by signal %d: %s",
2862
0
            procname, pid, WTERMSIG(exitstatus),
2863
0
            pg_strsignal(WTERMSIG(exitstatus))),
2864
0
         activity ? errdetail("Failed process was running: %s", activity) : 0));
2865
0
#endif
2866
0
  }
2867
0
  else
2868
0
    ereport(lev,
2869
2870
    /*------
2871
      translator: %s is a noun phrase describing a child process, such as
2872
      "server process" */
2873
0
        (errmsg("%s (PID %d) exited with unrecognized status %d",
2874
0
            procname, pid, exitstatus),
2875
0
         activity ? errdetail("Failed process was running: %s", activity) : 0));
2876
0
}
2877
2878
/*
2879
 * Advance the postmaster's state machine and take actions as appropriate
2880
 *
2881
 * This is common code for process_pm_shutdown_request(),
2882
 * process_pm_child_exit() and process_pm_pmsignal(), which process the signals
2883
 * that might mean we need to change state.
2884
 */
2885
static void
2886
PostmasterStateMachine(void)
2887
0
{
2888
  /* If we're doing a smart shutdown, try to advance that state. */
2889
0
  if (pmState == PM_RUN || pmState == PM_HOT_STANDBY)
2890
0
  {
2891
0
    if (!connsAllowed)
2892
0
    {
2893
      /*
2894
       * This state ends when we have no normal client backends running.
2895
       * Then we're ready to stop other children.
2896
       */
2897
0
      if (CountChildren(btmask(B_BACKEND)) == 0)
2898
0
        UpdatePMState(PM_STOP_BACKENDS);
2899
0
    }
2900
0
  }
2901
2902
  /*
2903
   * In the PM_WAIT_BACKENDS state, wait for all the regular backends and
2904
   * processes like autovacuum and background workers that are comparable to
2905
   * backends to exit.
2906
   *
2907
   * PM_STOP_BACKENDS is a transient state that means the same as
2908
   * PM_WAIT_BACKENDS, but we signal the processes first, before waiting for
2909
   * them.  Treating it as a distinct pmState allows us to share this code
2910
   * across multiple shutdown code paths.
2911
   */
2912
0
  if (pmState == PM_STOP_BACKENDS || pmState == PM_WAIT_BACKENDS)
2913
0
  {
2914
0
    BackendTypeMask targetMask = BTYPE_MASK_NONE;
2915
2916
    /*
2917
     * PM_WAIT_BACKENDS state ends when we have no regular backends, no
2918
     * autovac launcher or workers, and no bgworkers (including
2919
     * unconnected ones).
2920
     */
2921
0
    targetMask = btmask_add(targetMask,
2922
0
                B_BACKEND,
2923
0
                B_AUTOVAC_LAUNCHER,
2924
0
                B_AUTOVAC_WORKER,
2925
0
                B_BG_WORKER);
2926
2927
    /*
2928
     * No walwriter, bgwriter, slot sync worker, or WAL summarizer either.
2929
     */
2930
0
    targetMask = btmask_add(targetMask,
2931
0
                B_WAL_WRITER,
2932
0
                B_BG_WRITER,
2933
0
                B_SLOTSYNC_WORKER,
2934
0
                B_WAL_SUMMARIZER);
2935
2936
    /* If we're in recovery, also stop startup and walreceiver procs */
2937
0
    targetMask = btmask_add(targetMask,
2938
0
                B_STARTUP,
2939
0
                B_WAL_RECEIVER);
2940
2941
    /*
2942
     * If we are doing crash recovery or an immediate shutdown then we
2943
     * expect archiver, checkpointer, io workers and walsender to exit as
2944
     * well, otherwise not.
2945
     */
2946
0
    if (FatalError || Shutdown >= ImmediateShutdown)
2947
0
      targetMask = btmask_add(targetMask,
2948
0
                  B_CHECKPOINTER,
2949
0
                  B_ARCHIVER,
2950
0
                  B_IO_WORKER,
2951
0
                  B_WAL_SENDER);
2952
2953
    /*
2954
     * Normally archiver, checkpointer, IO workers and walsenders will
2955
     * continue running; they will be terminated later after writing the
2956
     * checkpoint record.  We also let dead-end children to keep running
2957
     * for now.  The syslogger process exits last.
2958
     *
2959
     * This assertion checks that we have covered all backend types,
2960
     * either by including them in targetMask, or by noting here that they
2961
     * are allowed to continue running.
2962
     */
2963
#ifdef USE_ASSERT_CHECKING
2964
    {
2965
      BackendTypeMask remainMask = BTYPE_MASK_NONE;
2966
2967
      remainMask = btmask_add(remainMask,
2968
                  B_DEAD_END_BACKEND,
2969
                  B_LOGGER);
2970
2971
      /*
2972
       * Archiver, checkpointer, IO workers, and walsender may or may
2973
       * not be in targetMask already.
2974
       */
2975
      remainMask = btmask_add(remainMask,
2976
                  B_ARCHIVER,
2977
                  B_CHECKPOINTER,
2978
                  B_IO_WORKER,
2979
                  B_WAL_SENDER);
2980
2981
      /* these are not real postmaster children */
2982
      remainMask = btmask_add(remainMask,
2983
                  B_INVALID,
2984
                  B_STANDALONE_BACKEND);
2985
2986
      /* All types should be included in targetMask or remainMask */
2987
      Assert((remainMask.mask | targetMask.mask) == BTYPE_MASK_ALL.mask);
2988
    }
2989
#endif
2990
2991
    /* If we had not yet signaled the processes to exit, do so now */
2992
0
    if (pmState == PM_STOP_BACKENDS)
2993
0
    {
2994
      /*
2995
       * Forget any pending requests for background workers, since we're
2996
       * no longer willing to launch any new workers.  (If additional
2997
       * requests arrive, BackgroundWorkerStateChange will reject them.)
2998
       */
2999
0
      ForgetUnstartedBackgroundWorkers();
3000
3001
0
      SignalChildren(SIGTERM, targetMask);
3002
3003
0
      UpdatePMState(PM_WAIT_BACKENDS);
3004
0
    }
3005
3006
    /* Are any of the target processes still running? */
3007
0
    if (CountChildren(targetMask) == 0)
3008
0
    {
3009
0
      if (Shutdown >= ImmediateShutdown || FatalError)
3010
0
      {
3011
        /*
3012
         * Stop any dead-end children and stop creating new ones.
3013
         *
3014
         * NB: Similar code exists in HandleFatalError(), when the
3015
         * error happens in pmState > PM_WAIT_BACKENDS.
3016
         */
3017
0
        UpdatePMState(PM_WAIT_DEAD_END);
3018
0
        ConfigurePostmasterWaitSet(false);
3019
0
        SignalChildren(SIGQUIT, btmask(B_DEAD_END_BACKEND));
3020
3021
        /*
3022
         * We already SIGQUIT'd auxiliary processes (other than
3023
         * logger), if any, when we started immediate shutdown or
3024
         * entered FatalError state.
3025
         */
3026
0
      }
3027
0
      else
3028
0
      {
3029
        /*
3030
         * If we get here, we are proceeding with normal shutdown. All
3031
         * the regular children are gone, and it's time to tell the
3032
         * checkpointer to do a shutdown checkpoint.
3033
         */
3034
0
        Assert(Shutdown > NoShutdown);
3035
        /* Start the checkpointer if not running */
3036
0
        if (CheckpointerPMChild == NULL)
3037
0
          CheckpointerPMChild = StartChildProcess(B_CHECKPOINTER);
3038
        /* And tell it to write the shutdown checkpoint */
3039
0
        if (CheckpointerPMChild != NULL)
3040
0
        {
3041
0
          signal_child(CheckpointerPMChild, SIGINT);
3042
0
          UpdatePMState(PM_WAIT_XLOG_SHUTDOWN);
3043
0
        }
3044
0
        else
3045
0
        {
3046
          /*
3047
           * If we failed to fork a checkpointer, just shut down.
3048
           * Any required cleanup will happen at next restart. We
3049
           * set FatalError so that an "abnormal shutdown" message
3050
           * gets logged when we exit.
3051
           *
3052
           * We don't consult send_abort_for_crash here, as it's
3053
           * unlikely that dumping cores would illuminate the reason
3054
           * for checkpointer fork failure.
3055
           *
3056
           * XXX: It may be worth to introduce a different PMQUIT
3057
           * value that signals that the cluster is in a bad state,
3058
           * without a process having crashed. But right now this
3059
           * path is very unlikely to be reached, so it isn't
3060
           * obviously worthwhile adding a distinct error message in
3061
           * quickdie().
3062
           */
3063
0
          HandleFatalError(PMQUIT_FOR_CRASH, false);
3064
0
        }
3065
0
      }
3066
0
    }
3067
0
  }
3068
3069
  /*
3070
   * The state transition from PM_WAIT_XLOG_SHUTDOWN to
3071
   * PM_WAIT_XLOG_ARCHIVAL is in process_pm_pmsignal(), in response to
3072
   * PMSIGNAL_XLOG_IS_SHUTDOWN.
3073
   */
3074
3075
0
  if (pmState == PM_WAIT_XLOG_ARCHIVAL)
3076
0
  {
3077
    /*
3078
     * PM_WAIT_XLOG_ARCHIVAL state ends when there are no children other
3079
     * than checkpointer, io workers and dead-end children left. There
3080
     * shouldn't be any regular backends left by now anyway; what we're
3081
     * really waiting for is for walsenders and archiver to exit.
3082
     */
3083
0
    if (CountChildren(btmask_all_except(B_CHECKPOINTER, B_IO_WORKER,
3084
0
                      B_LOGGER, B_DEAD_END_BACKEND)) == 0)
3085
0
    {
3086
0
      UpdatePMState(PM_WAIT_IO_WORKERS);
3087
0
      SignalChildren(SIGUSR2, btmask(B_IO_WORKER));
3088
0
    }
3089
0
  }
3090
3091
0
  if (pmState == PM_WAIT_IO_WORKERS)
3092
0
  {
3093
    /*
3094
     * PM_WAIT_IO_WORKERS state ends when there's only checkpointer and
3095
     * dead-end children left.
3096
     */
3097
0
    if (io_worker_count == 0)
3098
0
    {
3099
0
      UpdatePMState(PM_WAIT_CHECKPOINTER);
3100
3101
      /*
3102
       * Now that the processes mentioned above are gone, tell
3103
       * checkpointer to shut down too. That allows checkpointer to
3104
       * perform some last bits of cleanup without other processes
3105
       * interfering.
3106
       */
3107
0
      if (CheckpointerPMChild != NULL)
3108
0
        signal_child(CheckpointerPMChild, SIGUSR2);
3109
0
    }
3110
0
  }
3111
3112
  /*
3113
   * The state transition from PM_WAIT_CHECKPOINTER to PM_WAIT_DEAD_END is
3114
   * in process_pm_child_exit().
3115
   */
3116
3117
0
  if (pmState == PM_WAIT_DEAD_END)
3118
0
  {
3119
    /*
3120
     * PM_WAIT_DEAD_END state ends when all other children are gone except
3121
     * for the logger.  During normal shutdown, all that remains are
3122
     * dead-end backends, but in FatalError processing we jump straight
3123
     * here with more processes remaining.  Note that they have already
3124
     * been sent appropriate shutdown signals, either during a normal
3125
     * state transition leading up to PM_WAIT_DEAD_END, or during
3126
     * FatalError processing.
3127
     *
3128
     * The reason we wait is to protect against a new postmaster starting
3129
     * conflicting subprocesses; this isn't an ironclad protection, but it
3130
     * at least helps in the shutdown-and-immediately-restart scenario.
3131
     */
3132
0
    if (CountChildren(btmask_all_except(B_LOGGER)) == 0)
3133
0
    {
3134
      /* These other guys should be dead already */
3135
0
      Assert(StartupPMChild == NULL);
3136
0
      Assert(WalReceiverPMChild == NULL);
3137
0
      Assert(WalSummarizerPMChild == NULL);
3138
0
      Assert(BgWriterPMChild == NULL);
3139
0
      Assert(CheckpointerPMChild == NULL);
3140
0
      Assert(WalWriterPMChild == NULL);
3141
0
      Assert(AutoVacLauncherPMChild == NULL);
3142
0
      Assert(SlotSyncWorkerPMChild == NULL);
3143
      /* syslogger is not considered here */
3144
0
      UpdatePMState(PM_NO_CHILDREN);
3145
0
    }
3146
0
  }
3147
3148
  /*
3149
   * If we've been told to shut down, we exit as soon as there are no
3150
   * remaining children.  If there was a crash, cleanup will occur at the
3151
   * next startup.  (Before PostgreSQL 8.3, we tried to recover from the
3152
   * crash before exiting, but that seems unwise if we are quitting because
3153
   * we got SIGTERM from init --- there may well not be time for recovery
3154
   * before init decides to SIGKILL us.)
3155
   *
3156
   * Note that the syslogger continues to run.  It will exit when it sees
3157
   * EOF on its input pipe, which happens when there are no more upstream
3158
   * processes.
3159
   */
3160
0
  if (Shutdown > NoShutdown && pmState == PM_NO_CHILDREN)
3161
0
  {
3162
0
    if (FatalError)
3163
0
    {
3164
0
      ereport(LOG, (errmsg("abnormal database system shutdown")));
3165
0
      ExitPostmaster(1);
3166
0
    }
3167
0
    else
3168
0
    {
3169
      /*
3170
       * Normal exit from the postmaster is here.  We don't need to log
3171
       * anything here, since the UnlinkLockFiles proc_exit callback
3172
       * will do so, and that should be the last user-visible action.
3173
       */
3174
0
      ExitPostmaster(0);
3175
0
    }
3176
0
  }
3177
3178
  /*
3179
   * If the startup process failed, or the user does not want an automatic
3180
   * restart after backend crashes, wait for all non-syslogger children to
3181
   * exit, and then exit postmaster.  We don't try to reinitialize when the
3182
   * startup process fails, because more than likely it will just fail again
3183
   * and we will keep trying forever.
3184
   */
3185
0
  if (pmState == PM_NO_CHILDREN)
3186
0
  {
3187
0
    if (StartupStatus == STARTUP_CRASHED)
3188
0
    {
3189
0
      ereport(LOG,
3190
0
          (errmsg("shutting down due to startup process failure")));
3191
0
      ExitPostmaster(1);
3192
0
    }
3193
0
    if (!restart_after_crash)
3194
0
    {
3195
0
      ereport(LOG,
3196
0
          (errmsg("shutting down because \"restart_after_crash\" is off")));
3197
0
      ExitPostmaster(1);
3198
0
    }
3199
0
  }
3200
3201
  /*
3202
   * If we need to recover from a crash, wait for all non-syslogger children
3203
   * to exit, then reset shmem and start the startup process.
3204
   */
3205
0
  if (FatalError && pmState == PM_NO_CHILDREN)
3206
0
  {
3207
0
    ereport(LOG,
3208
0
        (errmsg("all server processes terminated; reinitializing")));
3209
3210
    /* remove leftover temporary files after a crash */
3211
0
    if (remove_temp_files_after_crash)
3212
0
      RemovePgTempFiles();
3213
3214
    /* allow background workers to immediately restart */
3215
0
    ResetBackgroundWorkerCrashTimes();
3216
3217
0
    shmem_exit(1);
3218
3219
    /* re-read control file into local memory */
3220
0
    LocalProcessControlFile(true);
3221
3222
    /* re-create shared memory and semaphores */
3223
0
    CreateSharedMemoryAndSemaphores();
3224
3225
0
    UpdatePMState(PM_STARTUP);
3226
3227
    /* Make sure we can perform I/O while starting up. */
3228
0
    maybe_adjust_io_workers();
3229
3230
0
    StartupPMChild = StartChildProcess(B_STARTUP);
3231
0
    Assert(StartupPMChild != NULL);
3232
0
    StartupStatus = STARTUP_RUNNING;
3233
    /* crash recovery started, reset SIGKILL flag */
3234
0
    AbortStartTime = 0;
3235
3236
    /* start accepting server socket connection events again */
3237
0
    ConfigurePostmasterWaitSet(true);
3238
0
  }
3239
0
}
3240
3241
static const char *
3242
pmstate_name(PMState state)
3243
0
{
3244
0
#define PM_TOSTR_CASE(sym) case sym: return #sym
3245
0
  switch (state)
3246
0
  {
3247
0
      PM_TOSTR_CASE(PM_INIT);
3248
0
      PM_TOSTR_CASE(PM_STARTUP);
3249
0
      PM_TOSTR_CASE(PM_RECOVERY);
3250
0
      PM_TOSTR_CASE(PM_HOT_STANDBY);
3251
0
      PM_TOSTR_CASE(PM_RUN);
3252
0
      PM_TOSTR_CASE(PM_STOP_BACKENDS);
3253
0
      PM_TOSTR_CASE(PM_WAIT_BACKENDS);
3254
0
      PM_TOSTR_CASE(PM_WAIT_XLOG_SHUTDOWN);
3255
0
      PM_TOSTR_CASE(PM_WAIT_XLOG_ARCHIVAL);
3256
0
      PM_TOSTR_CASE(PM_WAIT_IO_WORKERS);
3257
0
      PM_TOSTR_CASE(PM_WAIT_DEAD_END);
3258
0
      PM_TOSTR_CASE(PM_WAIT_CHECKPOINTER);
3259
0
      PM_TOSTR_CASE(PM_NO_CHILDREN);
3260
0
  }
3261
0
#undef PM_TOSTR_CASE
3262
3263
0
  pg_unreachable();
3264
0
  return "";         /* silence compiler */
3265
0
}
3266
3267
/*
3268
 * Simple wrapper for updating pmState. The main reason to have this wrapper
3269
 * is that it makes it easy to log all state transitions.
3270
 */
3271
static void
3272
UpdatePMState(PMState newState)
3273
{
3274
  elog(DEBUG1, "updating PMState from %s to %s",
3275
     pmstate_name(pmState), pmstate_name(newState));
3276
  pmState = newState;
3277
}
3278
3279
/*
3280
 * Launch background processes after state change, or relaunch after an
3281
 * existing process has exited.
3282
 *
3283
 * Check the current pmState and the status of any background processes.  If
3284
 * there are any background processes missing that should be running in the
3285
 * current state, but are not, launch them.
3286
 */
3287
static void
3288
LaunchMissingBackgroundProcesses(void)
3289
0
{
3290
  /* Syslogger is active in all states */
3291
0
  if (SysLoggerPMChild == NULL && Logging_collector)
3292
0
    StartSysLogger();
3293
3294
  /*
3295
   * The number of configured workers might have changed, or a prior start
3296
   * of a worker might have failed. Check if we need to start/stop any
3297
   * workers.
3298
   *
3299
   * A config file change will always lead to this function being called, so
3300
   * we always will process the config change in a timely manner.
3301
   */
3302
0
  maybe_adjust_io_workers();
3303
3304
  /*
3305
   * The checkpointer and the background writer are active from the start,
3306
   * until shutdown is initiated.
3307
   *
3308
   * (If the checkpointer is not running when we enter the
3309
   * PM_WAIT_XLOG_SHUTDOWN state, it is launched one more time to perform
3310
   * the shutdown checkpoint.  That's done in PostmasterStateMachine(), not
3311
   * here.)
3312
   */
3313
0
  if (pmState == PM_RUN || pmState == PM_RECOVERY ||
3314
0
    pmState == PM_HOT_STANDBY || pmState == PM_STARTUP)
3315
0
  {
3316
0
    if (CheckpointerPMChild == NULL)
3317
0
      CheckpointerPMChild = StartChildProcess(B_CHECKPOINTER);
3318
0
    if (BgWriterPMChild == NULL)
3319
0
      BgWriterPMChild = StartChildProcess(B_BG_WRITER);
3320
0
  }
3321
3322
  /*
3323
   * WAL writer is needed only in normal operation (else we cannot be
3324
   * writing any new WAL).
3325
   */
3326
0
  if (WalWriterPMChild == NULL && pmState == PM_RUN)
3327
0
    WalWriterPMChild = StartChildProcess(B_WAL_WRITER);
3328
3329
  /*
3330
   * We don't want autovacuum to run in binary upgrade mode because
3331
   * autovacuum might update relfrozenxid for empty tables before the
3332
   * physical files are put in place.
3333
   */
3334
0
  if (!IsBinaryUpgrade && AutoVacLauncherPMChild == NULL &&
3335
0
    (AutoVacuumingActive() || start_autovac_launcher) &&
3336
0
    pmState == PM_RUN)
3337
0
  {
3338
0
    AutoVacLauncherPMChild = StartChildProcess(B_AUTOVAC_LAUNCHER);
3339
0
    if (AutoVacLauncherPMChild != NULL)
3340
0
      start_autovac_launcher = false; /* signal processed */
3341
0
  }
3342
3343
  /*
3344
   * If WAL archiving is enabled always, we are allowed to start archiver
3345
   * even during recovery.
3346
   */
3347
0
  if (PgArchPMChild == NULL &&
3348
0
    ((XLogArchivingActive() && pmState == PM_RUN) ||
3349
0
     (XLogArchivingAlways() && (pmState == PM_RECOVERY || pmState == PM_HOT_STANDBY))) &&
3350
0
    PgArchCanRestart())
3351
0
    PgArchPMChild = StartChildProcess(B_ARCHIVER);
3352
3353
  /*
3354
   * If we need to start a slot sync worker, try to do that now
3355
   *
3356
   * We allow to start the slot sync worker when we are on a hot standby,
3357
   * fast or immediate shutdown is not in progress, slot sync parameters are
3358
   * configured correctly, and it is the first time of worker's launch, or
3359
   * enough time has passed since the worker was launched last.
3360
   */
3361
0
  if (SlotSyncWorkerPMChild == NULL && pmState == PM_HOT_STANDBY &&
3362
0
    Shutdown <= SmartShutdown && sync_replication_slots &&
3363
0
    ValidateSlotSyncParams(LOG) && SlotSyncWorkerCanRestart())
3364
0
    SlotSyncWorkerPMChild = StartChildProcess(B_SLOTSYNC_WORKER);
3365
3366
  /*
3367
   * If we need to start a WAL receiver, try to do that now
3368
   *
3369
   * Note: if a walreceiver process is already running, it might seem that
3370
   * we should clear WalReceiverRequested.  However, there's a race
3371
   * condition if the walreceiver terminates and the startup process
3372
   * immediately requests a new one: it's quite possible to get the signal
3373
   * for the request before reaping the dead walreceiver process.  Better to
3374
   * risk launching an extra walreceiver than to miss launching one we need.
3375
   * (The walreceiver code has logic to recognize that it should go away if
3376
   * not needed.)
3377
   */
3378
0
  if (WalReceiverRequested)
3379
0
  {
3380
0
    if (WalReceiverPMChild == NULL &&
3381
0
      (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
3382
0
       pmState == PM_HOT_STANDBY) &&
3383
0
      Shutdown <= SmartShutdown)
3384
0
    {
3385
0
      WalReceiverPMChild = StartChildProcess(B_WAL_RECEIVER);
3386
0
      if (WalReceiverPMChild != 0)
3387
0
        WalReceiverRequested = false;
3388
      /* else leave the flag set, so we'll try again later */
3389
0
    }
3390
0
  }
3391
3392
  /* If we need to start a WAL summarizer, try to do that now */
3393
0
  if (summarize_wal && WalSummarizerPMChild == NULL &&
3394
0
    (pmState == PM_RUN || pmState == PM_HOT_STANDBY) &&
3395
0
    Shutdown <= SmartShutdown)
3396
0
    WalSummarizerPMChild = StartChildProcess(B_WAL_SUMMARIZER);
3397
3398
  /* Get other worker processes running, if needed */
3399
0
  if (StartWorkerNeeded || HaveCrashedWorker)
3400
0
    maybe_start_bgworkers();
3401
0
}
3402
3403
/*
3404
 * Return string representation of signal.
3405
 *
3406
 * Because this is only implemented for signals we already rely on in this
3407
 * file we don't need to deal with unimplemented or same-numeric-value signals
3408
 * (as we'd e.g. have to for EWOULDBLOCK / EAGAIN).
3409
 */
3410
static const char *
3411
pm_signame(int signal)
3412
0
{
3413
0
#define PM_TOSTR_CASE(sym) case sym: return #sym
3414
0
  switch (signal)
3415
0
  {
3416
0
      PM_TOSTR_CASE(SIGABRT);
3417
0
      PM_TOSTR_CASE(SIGCHLD);
3418
0
      PM_TOSTR_CASE(SIGHUP);
3419
0
      PM_TOSTR_CASE(SIGINT);
3420
0
      PM_TOSTR_CASE(SIGKILL);
3421
0
      PM_TOSTR_CASE(SIGQUIT);
3422
0
      PM_TOSTR_CASE(SIGTERM);
3423
0
      PM_TOSTR_CASE(SIGUSR1);
3424
0
      PM_TOSTR_CASE(SIGUSR2);
3425
0
    default:
3426
      /* all signals sent by postmaster should be listed here */
3427
0
      Assert(false);
3428
0
      return "(unknown)";
3429
0
  }
3430
0
#undef PM_TOSTR_CASE
3431
3432
0
  return "";         /* silence compiler */
3433
0
}
3434
3435
/*
3436
 * Send a signal to a postmaster child process
3437
 *
3438
 * On systems that have setsid(), each child process sets itself up as a
3439
 * process group leader.  For signals that are generally interpreted in the
3440
 * appropriate fashion, we signal the entire process group not just the
3441
 * direct child process.  This allows us to, for example, SIGQUIT a blocked
3442
 * archive_recovery script, or SIGINT a script being run by a backend via
3443
 * system().
3444
 *
3445
 * There is a race condition for recently-forked children: they might not
3446
 * have executed setsid() yet.  So we signal the child directly as well as
3447
 * the group.  We assume such a child will handle the signal before trying
3448
 * to spawn any grandchild processes.  We also assume that signaling the
3449
 * child twice will not cause any problems.
3450
 */
3451
static void
3452
signal_child(PMChild *pmchild, int signal)
3453
0
{
3454
0
  pid_t   pid = pmchild->pid;
3455
3456
0
  ereport(DEBUG3,
3457
0
      (errmsg_internal("sending signal %d/%s to %s process with pid %d",
3458
0
               signal, pm_signame(signal),
3459
0
               GetBackendTypeDesc(pmchild->bkend_type),
3460
0
               (int) pmchild->pid)));
3461
3462
0
  if (kill(pid, signal) < 0)
3463
0
    elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) pid, signal);
3464
0
#ifdef HAVE_SETSID
3465
0
  switch (signal)
3466
0
  {
3467
0
    case SIGINT:
3468
0
    case SIGTERM:
3469
0
    case SIGQUIT:
3470
0
    case SIGKILL:
3471
0
    case SIGABRT:
3472
0
      if (kill(-pid, signal) < 0)
3473
0
        elog(DEBUG3, "kill(%ld,%d) failed: %m", (long) (-pid), signal);
3474
0
      break;
3475
0
    default:
3476
0
      break;
3477
0
  }
3478
0
#endif
3479
0
}
3480
3481
/*
3482
 * Send a signal to the targeted children.
3483
 */
3484
static bool
3485
SignalChildren(int signal, BackendTypeMask targetMask)
3486
0
{
3487
0
  dlist_iter  iter;
3488
0
  bool    signaled = false;
3489
3490
0
  dlist_foreach(iter, &ActiveChildList)
3491
0
  {
3492
0
    PMChild    *bp = dlist_container(PMChild, elem, iter.cur);
3493
3494
    /*
3495
     * If we need to distinguish between B_BACKEND and B_WAL_SENDER, check
3496
     * if any B_BACKEND backends have recently announced that they are
3497
     * actually WAL senders.
3498
     */
3499
0
    if (btmask_contains(targetMask, B_WAL_SENDER) != btmask_contains(targetMask, B_BACKEND) &&
3500
0
      bp->bkend_type == B_BACKEND)
3501
0
    {
3502
0
      if (IsPostmasterChildWalSender(bp->child_slot))
3503
0
        bp->bkend_type = B_WAL_SENDER;
3504
0
    }
3505
3506
0
    if (!btmask_contains(targetMask, bp->bkend_type))
3507
0
      continue;
3508
3509
0
    signal_child(bp, signal);
3510
0
    signaled = true;
3511
0
  }
3512
0
  return signaled;
3513
0
}
3514
3515
/*
3516
 * Send a termination signal to children.  This considers all of our children
3517
 * processes, except syslogger.
3518
 */
3519
static void
3520
TerminateChildren(int signal)
3521
0
{
3522
0
  SignalChildren(signal, btmask_all_except(B_LOGGER));
3523
0
  if (StartupPMChild != NULL)
3524
0
  {
3525
0
    if (signal == SIGQUIT || signal == SIGKILL || signal == SIGABRT)
3526
0
      StartupStatus = STARTUP_SIGNALED;
3527
0
  }
3528
0
}
3529
3530
/*
3531
 * BackendStartup -- start backend process
3532
 *
3533
 * returns: STATUS_ERROR if the fork failed, STATUS_OK otherwise.
3534
 *
3535
 * Note: if you change this code, also consider StartAutovacuumWorker and
3536
 * StartBackgroundWorker.
3537
 */
3538
static int
3539
BackendStartup(ClientSocket *client_sock)
3540
{
3541
  PMChild    *bn = NULL;
3542
  pid_t   pid;
3543
  BackendStartupData startup_data;
3544
  CAC_state cac;
3545
3546
  /*
3547
   * Capture time that Postmaster got a socket from accept (for logging
3548
   * connection establishment and setup total duration).
3549
   */
3550
  startup_data.socket_created = GetCurrentTimestamp();
3551
3552
  /*
3553
   * Allocate and assign the child slot.  Note we must do this before
3554
   * forking, so that we can handle failures (out of memory or child-process
3555
   * slots) cleanly.
3556
   */
3557
  cac = canAcceptConnections(B_BACKEND);
3558
  if (cac == CAC_OK)
3559
  {
3560
    /* Can change later to B_WAL_SENDER */
3561
    bn = AssignPostmasterChildSlot(B_BACKEND);
3562
    if (!bn)
3563
    {
3564
      /*
3565
       * Too many regular child processes; launch a dead-end child
3566
       * process instead.
3567
       */
3568
      cac = CAC_TOOMANY;
3569
    }
3570
  }
3571
  if (!bn)
3572
  {
3573
    bn = AllocDeadEndChild();
3574
    if (!bn)
3575
    {
3576
      ereport(LOG,
3577
          (errcode(ERRCODE_OUT_OF_MEMORY),
3578
           errmsg("out of memory")));
3579
      return STATUS_ERROR;
3580
    }
3581
  }
3582
3583
  /* Pass down canAcceptConnections state */
3584
  startup_data.canAcceptConnections = cac;
3585
  bn->rw = NULL;
3586
3587
  /* Hasn't asked to be notified about any bgworkers yet */
3588
  bn->bgworker_notify = false;
3589
3590
  pid = postmaster_child_launch(bn->bkend_type, bn->child_slot,
3591
                  &startup_data, sizeof(startup_data),
3592
                  client_sock);
3593
  if (pid < 0)
3594
  {
3595
    /* in parent, fork failed */
3596
    int     save_errno = errno;
3597
3598
    (void) ReleasePostmasterChildSlot(bn);
3599
    errno = save_errno;
3600
    ereport(LOG,
3601
        (errmsg("could not fork new process for connection: %m")));
3602
    report_fork_failure_to_client(client_sock, save_errno);
3603
    return STATUS_ERROR;
3604
  }
3605
3606
  /* in parent, successful fork */
3607
  ereport(DEBUG2,
3608
      (errmsg_internal("forked new %s, pid=%d socket=%d",
3609
               GetBackendTypeDesc(bn->bkend_type),
3610
               (int) pid, (int) client_sock->sock)));
3611
3612
  /*
3613
   * Everything's been successful, it's safe to add this backend to our list
3614
   * of backends.
3615
   */
3616
  bn->pid = pid;
3617
  return STATUS_OK;
3618
}
3619
3620
/*
3621
 * Try to report backend fork() failure to client before we close the
3622
 * connection.  Since we do not care to risk blocking the postmaster on
3623
 * this connection, we set the connection to non-blocking and try only once.
3624
 *
3625
 * This is grungy special-purpose code; we cannot use backend libpq since
3626
 * it's not up and running.
3627
 */
3628
static void
3629
report_fork_failure_to_client(ClientSocket *client_sock, int errnum)
3630
0
{
3631
0
  char    buffer[1000];
3632
0
  int     rc;
3633
3634
  /* Format the error message packet (always V2 protocol) */
3635
0
  snprintf(buffer, sizeof(buffer), "E%s%s\n",
3636
0
       _("could not fork new process for connection: "),
3637
0
       strerror(errnum));
3638
3639
  /* Set port to non-blocking.  Don't do send() if this fails */
3640
0
  if (!pg_set_noblock(client_sock->sock))
3641
0
    return;
3642
3643
  /* We'll retry after EINTR, but ignore all other failures */
3644
0
  do
3645
0
  {
3646
0
    rc = send(client_sock->sock, buffer, strlen(buffer) + 1, 0);
3647
0
  } while (rc < 0 && errno == EINTR);
3648
0
}
3649
3650
/*
3651
 * ExitPostmaster -- cleanup
3652
 *
3653
 * Do NOT call exit() directly --- always go through here!
3654
 */
3655
static void
3656
ExitPostmaster(int status)
3657
0
{
3658
#ifdef HAVE_PTHREAD_IS_THREADED_NP
3659
3660
  /*
3661
   * There is no known cause for a postmaster to become multithreaded after
3662
   * startup.  However, we might reach here via an error exit before
3663
   * reaching the test in PostmasterMain, so provide the same hint as there.
3664
   * This message uses LOG level, because an unclean shutdown at this point
3665
   * would usually not look much different from a clean shutdown.
3666
   */
3667
  if (pthread_is_threaded_np() != 0)
3668
    ereport(LOG,
3669
        (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
3670
         errmsg("postmaster became multithreaded"),
3671
         errhint("Set the LC_ALL environment variable to a valid locale.")));
3672
#endif
3673
3674
  /* should cleanup shared memory and kill all backends */
3675
3676
  /*
3677
   * Not sure of the semantics here.  When the Postmaster dies, should the
3678
   * backends all be killed? probably not.
3679
   *
3680
   * MUST   -- vadim 05-10-1999
3681
   */
3682
3683
0
  proc_exit(status);
3684
0
}
3685
3686
/*
3687
 * Handle pmsignal conditions representing requests from backends,
3688
 * and check for promote and logrotate requests from pg_ctl.
3689
 */
3690
static void
3691
process_pm_pmsignal(void)
3692
0
{
3693
0
  bool    request_state_update = false;
3694
3695
0
  pending_pm_pmsignal = false;
3696
3697
0
  ereport(DEBUG2,
3698
0
      (errmsg_internal("postmaster received pmsignal signal")));
3699
3700
  /*
3701
   * RECOVERY_STARTED and BEGIN_HOT_STANDBY signals are ignored in
3702
   * unexpected states. If the startup process quickly starts up, completes
3703
   * recovery, exits, we might process the death of the startup process
3704
   * first. We don't want to go back to recovery in that case.
3705
   */
3706
0
  if (CheckPostmasterSignal(PMSIGNAL_RECOVERY_STARTED) &&
3707
0
    pmState == PM_STARTUP && Shutdown == NoShutdown)
3708
0
  {
3709
    /* WAL redo has started. We're out of reinitialization. */
3710
0
    FatalError = false;
3711
0
    AbortStartTime = 0;
3712
0
    reachedConsistency = false;
3713
3714
    /*
3715
     * Start the archiver if we're responsible for (re-)archiving received
3716
     * files.
3717
     */
3718
0
    Assert(PgArchPMChild == NULL);
3719
0
    if (XLogArchivingAlways())
3720
0
      PgArchPMChild = StartChildProcess(B_ARCHIVER);
3721
3722
    /*
3723
     * If we aren't planning to enter hot standby mode later, treat
3724
     * RECOVERY_STARTED as meaning we're out of startup, and report status
3725
     * accordingly.
3726
     */
3727
0
    if (!EnableHotStandby)
3728
0
    {
3729
0
      AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STANDBY);
3730
#ifdef USE_SYSTEMD
3731
      sd_notify(0, "READY=1");
3732
#endif
3733
0
    }
3734
3735
0
    UpdatePMState(PM_RECOVERY);
3736
0
  }
3737
3738
0
  if (CheckPostmasterSignal(PMSIGNAL_RECOVERY_CONSISTENT) &&
3739
0
    pmState == PM_RECOVERY && Shutdown == NoShutdown)
3740
0
  {
3741
0
    reachedConsistency = true;
3742
0
  }
3743
3744
0
  if (CheckPostmasterSignal(PMSIGNAL_BEGIN_HOT_STANDBY) &&
3745
0
    (pmState == PM_RECOVERY && Shutdown == NoShutdown))
3746
0
  {
3747
0
    ereport(LOG,
3748
0
        (errmsg("database system is ready to accept read-only connections")));
3749
3750
    /* Report status */
3751
0
    AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_READY);
3752
#ifdef USE_SYSTEMD
3753
    sd_notify(0, "READY=1");
3754
#endif
3755
3756
0
    UpdatePMState(PM_HOT_STANDBY);
3757
0
    connsAllowed = true;
3758
3759
    /* Some workers may be scheduled to start now */
3760
0
    StartWorkerNeeded = true;
3761
0
  }
3762
3763
  /* Process background worker state changes. */
3764
0
  if (CheckPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE))
3765
0
  {
3766
    /* Accept new worker requests only if not stopping. */
3767
0
    BackgroundWorkerStateChange(pmState < PM_STOP_BACKENDS);
3768
0
    StartWorkerNeeded = true;
3769
0
  }
3770
3771
  /* Tell syslogger to rotate logfile if requested */
3772
0
  if (SysLoggerPMChild != NULL)
3773
0
  {
3774
0
    if (CheckLogrotateSignal())
3775
0
    {
3776
0
      signal_child(SysLoggerPMChild, SIGUSR1);
3777
0
      RemoveLogrotateSignalFiles();
3778
0
    }
3779
0
    else if (CheckPostmasterSignal(PMSIGNAL_ROTATE_LOGFILE))
3780
0
    {
3781
0
      signal_child(SysLoggerPMChild, SIGUSR1);
3782
0
    }
3783
0
  }
3784
3785
0
  if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER) &&
3786
0
    Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
3787
0
  {
3788
    /*
3789
     * Start one iteration of the autovacuum daemon, even if autovacuuming
3790
     * is nominally not enabled.  This is so we can have an active defense
3791
     * against transaction ID wraparound.  We set a flag for the main loop
3792
     * to do it rather than trying to do it here --- this is because the
3793
     * autovac process itself may send the signal, and we want to handle
3794
     * that by launching another iteration as soon as the current one
3795
     * completes.
3796
     */
3797
0
    start_autovac_launcher = true;
3798
0
  }
3799
3800
0
  if (CheckPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER) &&
3801
0
    Shutdown <= SmartShutdown && pmState < PM_STOP_BACKENDS)
3802
0
  {
3803
    /* The autovacuum launcher wants us to start a worker process. */
3804
0
    StartAutovacuumWorker();
3805
0
  }
3806
3807
0
  if (CheckPostmasterSignal(PMSIGNAL_START_WALRECEIVER))
3808
0
  {
3809
    /* Startup Process wants us to start the walreceiver process. */
3810
0
    WalReceiverRequested = true;
3811
0
  }
3812
3813
0
  if (CheckPostmasterSignal(PMSIGNAL_XLOG_IS_SHUTDOWN))
3814
0
  {
3815
    /* Checkpointer completed the shutdown checkpoint */
3816
0
    if (pmState == PM_WAIT_XLOG_SHUTDOWN)
3817
0
    {
3818
      /*
3819
       * If we have an archiver subprocess, tell it to do a last archive
3820
       * cycle and quit. Likewise, if we have walsender processes, tell
3821
       * them to send any remaining WAL and quit.
3822
       */
3823
0
      Assert(Shutdown > NoShutdown);
3824
3825
      /* Waken archiver for the last time */
3826
0
      if (PgArchPMChild != NULL)
3827
0
        signal_child(PgArchPMChild, SIGUSR2);
3828
3829
      /*
3830
       * Waken walsenders for the last time. No regular backends should
3831
       * be around anymore.
3832
       */
3833
0
      SignalChildren(SIGUSR2, btmask(B_WAL_SENDER));
3834
3835
0
      UpdatePMState(PM_WAIT_XLOG_ARCHIVAL);
3836
0
    }
3837
0
    else if (!FatalError && Shutdown != ImmediateShutdown)
3838
0
    {
3839
      /*
3840
       * Checkpointer only ought to perform the shutdown checkpoint
3841
       * during shutdown.  If somehow checkpointer did so in another
3842
       * situation, we have no choice but to crash-restart.
3843
       *
3844
       * It's possible however that we get PMSIGNAL_XLOG_IS_SHUTDOWN
3845
       * outside of PM_WAIT_XLOG_SHUTDOWN if an orderly shutdown was
3846
       * "interrupted" by a crash or an immediate shutdown.
3847
       */
3848
0
      ereport(LOG,
3849
0
          (errmsg("WAL was shut down unexpectedly")));
3850
3851
      /*
3852
       * Doesn't seem likely to help to take send_abort_for_crash into
3853
       * account here.
3854
       */
3855
0
      HandleFatalError(PMQUIT_FOR_CRASH, false);
3856
0
    }
3857
3858
    /*
3859
     * Need to run PostmasterStateMachine() to check if we already can go
3860
     * to the next state.
3861
     */
3862
0
    request_state_update = true;
3863
0
  }
3864
3865
  /*
3866
   * Try to advance postmaster's state machine, if a child requests it.
3867
   */
3868
0
  if (CheckPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE))
3869
0
  {
3870
0
    request_state_update = true;
3871
0
  }
3872
3873
  /*
3874
   * Be careful about the order of this action relative to this function's
3875
   * other actions.  Generally, this should be after other actions, in case
3876
   * they have effects PostmasterStateMachine would need to know about.
3877
   * However, we should do it before the CheckPromoteSignal step, which
3878
   * cannot have any (immediate) effect on the state machine, but does
3879
   * depend on what state we're in now.
3880
   */
3881
0
  if (request_state_update)
3882
0
  {
3883
0
    PostmasterStateMachine();
3884
0
  }
3885
3886
0
  if (StartupPMChild != NULL &&
3887
0
    (pmState == PM_STARTUP || pmState == PM_RECOVERY ||
3888
0
     pmState == PM_HOT_STANDBY) &&
3889
0
    CheckPromoteSignal())
3890
0
  {
3891
    /*
3892
     * Tell startup process to finish recovery.
3893
     *
3894
     * Leave the promote signal file in place and let the Startup process
3895
     * do the unlink.
3896
     */
3897
0
    signal_child(StartupPMChild, SIGUSR2);
3898
0
  }
3899
0
}
3900
3901
/*
3902
 * Dummy signal handler
3903
 *
3904
 * We use this for signals that we don't actually use in the postmaster,
3905
 * but we do use in backends.  If we were to SIG_IGN such signals in the
3906
 * postmaster, then a newly started backend might drop a signal that arrives
3907
 * before it's able to reconfigure its signal processing.  (See notes in
3908
 * tcop/postgres.c.)
3909
 */
3910
static void
3911
dummy_handler(SIGNAL_ARGS)
3912
0
{
3913
0
}
3914
3915
/*
3916
 * Count up number of child processes of specified types.
3917
 */
3918
static int
3919
CountChildren(BackendTypeMask targetMask)
3920
{
3921
  dlist_iter  iter;
3922
  int     cnt = 0;
3923
3924
  dlist_foreach(iter, &ActiveChildList)
3925
  {
3926
    PMChild    *bp = dlist_container(PMChild, elem, iter.cur);
3927
3928
    /*
3929
     * If we need to distinguish between B_BACKEND and B_WAL_SENDER, check
3930
     * if any B_BACKEND backends have recently announced that they are
3931
     * actually WAL senders.
3932
     */
3933
    if (btmask_contains(targetMask, B_WAL_SENDER) != btmask_contains(targetMask, B_BACKEND) &&
3934
      bp->bkend_type == B_BACKEND)
3935
    {
3936
      if (IsPostmasterChildWalSender(bp->child_slot))
3937
        bp->bkend_type = B_WAL_SENDER;
3938
    }
3939
3940
    if (!btmask_contains(targetMask, bp->bkend_type))
3941
      continue;
3942
3943
    ereport(DEBUG4,
3944
        (errmsg_internal("%s process %d is still running",
3945
                 GetBackendTypeDesc(bp->bkend_type), (int) bp->pid)));
3946
3947
    cnt++;
3948
  }
3949
  return cnt;
3950
}
3951
3952
3953
/*
3954
 * StartChildProcess -- start an auxiliary process for the postmaster
3955
 *
3956
 * "type" determines what kind of child will be started.  All child types
3957
 * initially go to AuxiliaryProcessMain, which will handle common setup.
3958
 *
3959
 * Return value of StartChildProcess is subprocess' PMChild entry, or NULL on
3960
 * failure.
3961
 */
3962
static PMChild *
3963
StartChildProcess(BackendType type)
3964
0
{
3965
0
  PMChild    *pmchild;
3966
0
  pid_t   pid;
3967
3968
0
  pmchild = AssignPostmasterChildSlot(type);
3969
0
  if (!pmchild)
3970
0
  {
3971
0
    if (type == B_AUTOVAC_WORKER)
3972
0
      ereport(LOG,
3973
0
          (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
3974
0
           errmsg("no slot available for new autovacuum worker process")));
3975
0
    else
3976
0
    {
3977
      /* shouldn't happen because we allocate enough slots */
3978
0
      elog(LOG, "no postmaster child slot available for aux process");
3979
0
    }
3980
0
    return NULL;
3981
0
  }
3982
3983
0
  pid = postmaster_child_launch(type, pmchild->child_slot, NULL, 0, NULL);
3984
0
  if (pid < 0)
3985
0
  {
3986
    /* in parent, fork failed */
3987
0
    ReleasePostmasterChildSlot(pmchild);
3988
0
    ereport(LOG,
3989
0
        (errmsg("could not fork \"%s\" process: %m", PostmasterChildName(type))));
3990
3991
    /*
3992
     * fork failure is fatal during startup, but there's no need to choke
3993
     * immediately if starting other child types fails.
3994
     */
3995
0
    if (type == B_STARTUP)
3996
0
      ExitPostmaster(1);
3997
0
    return NULL;
3998
0
  }
3999
4000
  /* in parent, successful fork */
4001
0
  pmchild->pid = pid;
4002
0
  return pmchild;
4003
0
}
4004
4005
/*
4006
 * StartSysLogger -- start the syslogger process
4007
 */
4008
void
4009
StartSysLogger(void)
4010
0
{
4011
0
  Assert(SysLoggerPMChild == NULL);
4012
4013
0
  SysLoggerPMChild = AssignPostmasterChildSlot(B_LOGGER);
4014
0
  if (!SysLoggerPMChild)
4015
0
    elog(PANIC, "no postmaster child slot available for syslogger");
4016
0
  SysLoggerPMChild->pid = SysLogger_Start(SysLoggerPMChild->child_slot);
4017
0
  if (SysLoggerPMChild->pid == 0)
4018
0
  {
4019
0
    ReleasePostmasterChildSlot(SysLoggerPMChild);
4020
0
    SysLoggerPMChild = NULL;
4021
0
  }
4022
0
}
4023
4024
/*
4025
 * StartAutovacuumWorker
4026
 *    Start an autovac worker process.
4027
 *
4028
 * This function is here because it enters the resulting PID into the
4029
 * postmaster's private backends list.
4030
 *
4031
 * NB -- this code very roughly matches BackendStartup.
4032
 */
4033
static void
4034
StartAutovacuumWorker(void)
4035
0
{
4036
0
  PMChild    *bn;
4037
4038
  /*
4039
   * If not in condition to run a process, don't try, but handle it like a
4040
   * fork failure.  This does not normally happen, since the signal is only
4041
   * supposed to be sent by autovacuum launcher when it's OK to do it, but
4042
   * we have to check to avoid race-condition problems during DB state
4043
   * changes.
4044
   */
4045
0
  if (canAcceptConnections(B_AUTOVAC_WORKER) == CAC_OK)
4046
0
  {
4047
0
    bn = StartChildProcess(B_AUTOVAC_WORKER);
4048
0
    if (bn)
4049
0
    {
4050
0
      bn->bgworker_notify = false;
4051
0
      bn->rw = NULL;
4052
0
      return;
4053
0
    }
4054
0
    else
4055
0
    {
4056
      /*
4057
       * fork failed, fall through to report -- actual error message was
4058
       * logged by StartChildProcess
4059
       */
4060
0
    }
4061
0
  }
4062
4063
  /*
4064
   * Report the failure to the launcher, if it's running.  (If it's not, we
4065
   * might not even be connected to shared memory, so don't try to call
4066
   * AutoVacWorkerFailed.)  Note that we also need to signal it so that it
4067
   * responds to the condition, but we don't do that here, instead waiting
4068
   * for ServerLoop to do it.  This way we avoid a ping-pong signaling in
4069
   * quick succession between the autovac launcher and postmaster in case
4070
   * things get ugly.
4071
   */
4072
0
  if (AutoVacLauncherPMChild != NULL)
4073
0
  {
4074
0
    AutoVacWorkerFailed();
4075
0
    avlauncher_needs_signal = true;
4076
0
  }
4077
0
}
4078
4079
4080
/*
4081
 * Create the opts file
4082
 */
4083
static bool
4084
CreateOptsFile(int argc, char *argv[], char *fullprogname)
4085
{
4086
  FILE     *fp;
4087
  int     i;
4088
4089
#define OPTS_FILE "postmaster.opts"
4090
4091
  if ((fp = fopen(OPTS_FILE, "w")) == NULL)
4092
  {
4093
    ereport(LOG,
4094
        (errcode_for_file_access(),
4095
         errmsg("could not create file \"%s\": %m", OPTS_FILE)));
4096
    return false;
4097
  }
4098
4099
  fprintf(fp, "%s", fullprogname);
4100
  for (i = 1; i < argc; i++)
4101
    fprintf(fp, " \"%s\"", argv[i]);
4102
  fputs("\n", fp);
4103
4104
  if (fclose(fp))
4105
  {
4106
    ereport(LOG,
4107
        (errcode_for_file_access(),
4108
         errmsg("could not write file \"%s\": %m", OPTS_FILE)));
4109
    return false;
4110
  }
4111
4112
  return true;
4113
}
4114
4115
4116
/*
4117
 * Start a new bgworker.
4118
 * Starting time conditions must have been checked already.
4119
 *
4120
 * Returns true on success, false on failure.
4121
 * In either case, update the RegisteredBgWorker's state appropriately.
4122
 *
4123
 * NB -- this code very roughly matches BackendStartup.
4124
 */
4125
static bool
4126
StartBackgroundWorker(RegisteredBgWorker *rw)
4127
{
4128
  PMChild    *bn;
4129
  pid_t   worker_pid;
4130
4131
  Assert(rw->rw_pid == 0);
4132
4133
  /*
4134
   * Allocate and assign the child slot.  Note we must do this before
4135
   * forking, so that we can handle failures (out of memory or child-process
4136
   * slots) cleanly.
4137
   *
4138
   * Treat failure as though the worker had crashed.  That way, the
4139
   * postmaster will wait a bit before attempting to start it again; if we
4140
   * tried again right away, most likely we'd find ourselves hitting the
4141
   * same resource-exhaustion condition.
4142
   */
4143
  bn = AssignPostmasterChildSlot(B_BG_WORKER);
4144
  if (bn == NULL)
4145
  {
4146
    ereport(LOG,
4147
        (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
4148
         errmsg("no slot available for new background worker process")));
4149
    rw->rw_crashed_at = GetCurrentTimestamp();
4150
    return false;
4151
  }
4152
  bn->rw = rw;
4153
  bn->bkend_type = B_BG_WORKER;
4154
  bn->bgworker_notify = false;
4155
4156
  ereport(DEBUG1,
4157
      (errmsg_internal("starting background worker process \"%s\"",
4158
               rw->rw_worker.bgw_name)));
4159
4160
  worker_pid = postmaster_child_launch(B_BG_WORKER, bn->child_slot,
4161
                     &rw->rw_worker, sizeof(BackgroundWorker), NULL);
4162
  if (worker_pid == -1)
4163
  {
4164
    /* in postmaster, fork failed ... */
4165
    ereport(LOG,
4166
        (errmsg("could not fork background worker process: %m")));
4167
    /* undo what AssignPostmasterChildSlot did */
4168
    ReleasePostmasterChildSlot(bn);
4169
4170
    /* mark entry as crashed, so we'll try again later */
4171
    rw->rw_crashed_at = GetCurrentTimestamp();
4172
    return false;
4173
  }
4174
4175
  /* in postmaster, fork successful ... */
4176
  rw->rw_pid = worker_pid;
4177
  bn->pid = rw->rw_pid;
4178
  ReportBackgroundWorkerPID(rw);
4179
  return true;
4180
}
4181
4182
/*
4183
 * Does the current postmaster state require starting a worker with the
4184
 * specified start_time?
4185
 */
4186
static bool
4187
bgworker_should_start_now(BgWorkerStartTime start_time)
4188
0
{
4189
0
  switch (pmState)
4190
0
  {
4191
0
    case PM_NO_CHILDREN:
4192
0
    case PM_WAIT_CHECKPOINTER:
4193
0
    case PM_WAIT_DEAD_END:
4194
0
    case PM_WAIT_XLOG_ARCHIVAL:
4195
0
    case PM_WAIT_XLOG_SHUTDOWN:
4196
0
    case PM_WAIT_IO_WORKERS:
4197
0
    case PM_WAIT_BACKENDS:
4198
0
    case PM_STOP_BACKENDS:
4199
0
      break;
4200
4201
0
    case PM_RUN:
4202
0
      if (start_time == BgWorkerStart_RecoveryFinished)
4203
0
        return true;
4204
      /* fall through */
4205
4206
0
    case PM_HOT_STANDBY:
4207
0
      if (start_time == BgWorkerStart_ConsistentState)
4208
0
        return true;
4209
      /* fall through */
4210
4211
0
    case PM_RECOVERY:
4212
0
    case PM_STARTUP:
4213
0
    case PM_INIT:
4214
0
      if (start_time == BgWorkerStart_PostmasterStart)
4215
0
        return true;
4216
      /* fall through */
4217
0
  }
4218
4219
0
  return false;
4220
0
}
4221
4222
/*
4223
 * If the time is right, start background worker(s).
4224
 *
4225
 * As a side effect, the bgworker control variables are set or reset
4226
 * depending on whether more workers may need to be started.
4227
 *
4228
 * We limit the number of workers started per call, to avoid consuming the
4229
 * postmaster's attention for too long when many such requests are pending.
4230
 * As long as StartWorkerNeeded is true, ServerLoop will not block and will
4231
 * call this function again after dealing with any other issues.
4232
 */
4233
static void
4234
maybe_start_bgworkers(void)
4235
0
{
4236
0
#define MAX_BGWORKERS_TO_LAUNCH 100
4237
0
  int     num_launched = 0;
4238
0
  TimestampTz now = 0;
4239
0
  dlist_mutable_iter iter;
4240
4241
  /*
4242
   * During crash recovery, we have no need to be called until the state
4243
   * transition out of recovery.
4244
   */
4245
0
  if (FatalError)
4246
0
  {
4247
0
    StartWorkerNeeded = false;
4248
0
    HaveCrashedWorker = false;
4249
0
    return;
4250
0
  }
4251
4252
  /* Don't need to be called again unless we find a reason for it below */
4253
0
  StartWorkerNeeded = false;
4254
0
  HaveCrashedWorker = false;
4255
4256
0
  dlist_foreach_modify(iter, &BackgroundWorkerList)
4257
0
  {
4258
0
    RegisteredBgWorker *rw;
4259
4260
0
    rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur);
4261
4262
    /* ignore if already running */
4263
0
    if (rw->rw_pid != 0)
4264
0
      continue;
4265
4266
    /* if marked for death, clean up and remove from list */
4267
0
    if (rw->rw_terminate)
4268
0
    {
4269
0
      ForgetBackgroundWorker(rw);
4270
0
      continue;
4271
0
    }
4272
4273
    /*
4274
     * If this worker has crashed previously, maybe it needs to be
4275
     * restarted (unless on registration it specified it doesn't want to
4276
     * be restarted at all).  Check how long ago did a crash last happen.
4277
     * If the last crash is too recent, don't start it right away; let it
4278
     * be restarted once enough time has passed.
4279
     */
4280
0
    if (rw->rw_crashed_at != 0)
4281
0
    {
4282
0
      if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
4283
0
      {
4284
0
        int     notify_pid;
4285
4286
0
        notify_pid = rw->rw_worker.bgw_notify_pid;
4287
4288
0
        ForgetBackgroundWorker(rw);
4289
4290
        /* Report worker is gone now. */
4291
0
        if (notify_pid != 0)
4292
0
          kill(notify_pid, SIGUSR1);
4293
4294
0
        continue;
4295
0
      }
4296
4297
      /* read system time only when needed */
4298
0
      if (now == 0)
4299
0
        now = GetCurrentTimestamp();
4300
4301
0
      if (!TimestampDifferenceExceeds(rw->rw_crashed_at, now,
4302
0
                      rw->rw_worker.bgw_restart_time * 1000))
4303
0
      {
4304
        /* Set flag to remember that we have workers to start later */
4305
0
        HaveCrashedWorker = true;
4306
0
        continue;
4307
0
      }
4308
0
    }
4309
4310
0
    if (bgworker_should_start_now(rw->rw_worker.bgw_start_time))
4311
0
    {
4312
      /* reset crash time before trying to start worker */
4313
0
      rw->rw_crashed_at = 0;
4314
4315
      /*
4316
       * Try to start the worker.
4317
       *
4318
       * On failure, give up processing workers for now, but set
4319
       * StartWorkerNeeded so we'll come back here on the next iteration
4320
       * of ServerLoop to try again.  (We don't want to wait, because
4321
       * there might be additional ready-to-run workers.)  We could set
4322
       * HaveCrashedWorker as well, since this worker is now marked
4323
       * crashed, but there's no need because the next run of this
4324
       * function will do that.
4325
       */
4326
0
      if (!StartBackgroundWorker(rw))
4327
0
      {
4328
0
        StartWorkerNeeded = true;
4329
0
        return;
4330
0
      }
4331
4332
      /*
4333
       * If we've launched as many workers as allowed, quit, but have
4334
       * ServerLoop call us again to look for additional ready-to-run
4335
       * workers.  There might not be any, but we'll find out the next
4336
       * time we run.
4337
       */
4338
0
      if (++num_launched >= MAX_BGWORKERS_TO_LAUNCH)
4339
0
      {
4340
0
        StartWorkerNeeded = true;
4341
0
        return;
4342
0
      }
4343
0
    }
4344
0
  }
4345
0
}
4346
4347
static bool
4348
maybe_reap_io_worker(int pid)
4349
0
{
4350
0
  for (int i = 0; i < MAX_IO_WORKERS; ++i)
4351
0
  {
4352
0
    if (io_worker_children[i] &&
4353
0
      io_worker_children[i]->pid == pid)
4354
0
    {
4355
0
      ReleasePostmasterChildSlot(io_worker_children[i]);
4356
4357
0
      --io_worker_count;
4358
0
      io_worker_children[i] = NULL;
4359
0
      return true;
4360
0
    }
4361
0
  }
4362
0
  return false;
4363
0
}
4364
4365
/*
4366
 * Start or stop IO workers, to close the gap between the number of running
4367
 * workers and the number of configured workers.  Used to respond to change of
4368
 * the io_workers GUC (by increasing and decreasing the number of workers), as
4369
 * well as workers terminating in response to errors (by starting
4370
 * "replacement" workers).
4371
 */
4372
static void
4373
maybe_adjust_io_workers(void)
4374
0
{
4375
0
  if (!pgaio_workers_enabled())
4376
0
    return;
4377
4378
  /*
4379
   * If we're in final shutting down state, then we're just waiting for all
4380
   * processes to exit.
4381
   */
4382
0
  if (pmState >= PM_WAIT_IO_WORKERS)
4383
0
    return;
4384
4385
  /* Don't start new workers during an immediate shutdown either. */
4386
0
  if (Shutdown >= ImmediateShutdown)
4387
0
    return;
4388
4389
  /*
4390
   * Don't start new workers if we're in the shutdown phase of a crash
4391
   * restart. But we *do* need to start if we're already starting up again.
4392
   */
4393
0
  if (FatalError && pmState >= PM_STOP_BACKENDS)
4394
0
    return;
4395
4396
0
  Assert(pmState < PM_WAIT_IO_WORKERS);
4397
4398
  /* Not enough running? */
4399
0
  while (io_worker_count < io_workers)
4400
0
  {
4401
0
    PMChild    *child;
4402
0
    int     i;
4403
4404
    /* find unused entry in io_worker_children array */
4405
0
    for (i = 0; i < MAX_IO_WORKERS; ++i)
4406
0
    {
4407
0
      if (io_worker_children[i] == NULL)
4408
0
        break;
4409
0
    }
4410
0
    if (i == MAX_IO_WORKERS)
4411
0
      elog(ERROR, "could not find a free IO worker slot");
4412
4413
    /* Try to launch one. */
4414
0
    child = StartChildProcess(B_IO_WORKER);
4415
0
    if (child != NULL)
4416
0
    {
4417
0
      io_worker_children[i] = child;
4418
0
      ++io_worker_count;
4419
0
    }
4420
0
    else
4421
0
      break;       /* try again next time */
4422
0
  }
4423
4424
  /* Too many running? */
4425
0
  if (io_worker_count > io_workers)
4426
0
  {
4427
    /* ask the IO worker in the highest slot to exit */
4428
0
    for (int i = MAX_IO_WORKERS - 1; i >= 0; --i)
4429
0
    {
4430
0
      if (io_worker_children[i] != NULL)
4431
0
      {
4432
0
        kill(io_worker_children[i]->pid, SIGUSR2);
4433
0
        break;
4434
0
      }
4435
0
    }
4436
0
  }
4437
0
}
4438
4439
4440
/*
4441
 * When a backend asks to be notified about worker state changes, we
4442
 * set a flag in its backend entry.  The background worker machinery needs
4443
 * to know when such backends exit.
4444
 */
4445
bool
4446
PostmasterMarkPIDForWorkerNotify(int pid)
4447
0
{
4448
0
  dlist_iter  iter;
4449
0
  PMChild    *bp;
4450
4451
0
  dlist_foreach(iter, &ActiveChildList)
4452
0
  {
4453
0
    bp = dlist_container(PMChild, elem, iter.cur);
4454
0
    if (bp->pid == pid)
4455
0
    {
4456
0
      bp->bgworker_notify = true;
4457
0
      return true;
4458
0
    }
4459
0
  }
4460
0
  return false;
4461
0
}
4462
4463
#ifdef WIN32
4464
4465
/*
4466
 * Subset implementation of waitpid() for Windows.  We assume pid is -1
4467
 * (that is, check all child processes) and options is WNOHANG (don't wait).
4468
 */
4469
static pid_t
4470
waitpid(pid_t pid, int *exitstatus, int options)
4471
{
4472
  win32_deadchild_waitinfo *childinfo;
4473
  DWORD   exitcode;
4474
  DWORD   dwd;
4475
  ULONG_PTR key;
4476
  OVERLAPPED *ovl;
4477
4478
  /* Try to consume one win32_deadchild_waitinfo from the queue. */
4479
  if (!GetQueuedCompletionStatus(win32ChildQueue, &dwd, &key, &ovl, 0))
4480
  {
4481
    errno = EAGAIN;
4482
    return -1;
4483
  }
4484
4485
  childinfo = (win32_deadchild_waitinfo *) key;
4486
  pid = childinfo->procId;
4487
4488
  /*
4489
   * Remove handle from wait - required even though it's set to wait only
4490
   * once
4491
   */
4492
  UnregisterWaitEx(childinfo->waitHandle, NULL);
4493
4494
  if (!GetExitCodeProcess(childinfo->procHandle, &exitcode))
4495
  {
4496
    /*
4497
     * Should never happen. Inform user and set a fixed exitcode.
4498
     */
4499
    write_stderr("could not read exit code for process\n");
4500
    exitcode = 255;
4501
  }
4502
  *exitstatus = exitcode;
4503
4504
  /*
4505
   * Close the process handle.  Only after this point can the PID can be
4506
   * recycled by the kernel.
4507
   */
4508
  CloseHandle(childinfo->procHandle);
4509
4510
  /*
4511
   * Free struct that was allocated before the call to
4512
   * RegisterWaitForSingleObject()
4513
   */
4514
  pfree(childinfo);
4515
4516
  return pid;
4517
}
4518
4519
/*
4520
 * Note! Code below executes on a thread pool! All operations must
4521
 * be thread safe! Note that elog() and friends must *not* be used.
4522
 */
4523
static void WINAPI
4524
pgwin32_deadchild_callback(PVOID lpParameter, BOOLEAN TimerOrWaitFired)
4525
{
4526
  /* Should never happen, since we use INFINITE as timeout value. */
4527
  if (TimerOrWaitFired)
4528
    return;
4529
4530
  /*
4531
   * Post the win32_deadchild_waitinfo object for waitpid() to deal with. If
4532
   * that fails, we leak the object, but we also leak a whole process and
4533
   * get into an unrecoverable state, so there's not much point in worrying
4534
   * about that.  We'd like to panic, but we can't use that infrastructure
4535
   * from this thread.
4536
   */
4537
  if (!PostQueuedCompletionStatus(win32ChildQueue,
4538
                  0,
4539
                  (ULONG_PTR) lpParameter,
4540
                  NULL))
4541
    write_stderr("could not post child completion status\n");
4542
4543
  /* Queue SIGCHLD signal. */
4544
  pg_queue_signal(SIGCHLD);
4545
}
4546
4547
/*
4548
 * Queue a waiter to signal when this child dies.  The wait will be handled
4549
 * automatically by an operating system thread pool.  The memory and the
4550
 * process handle will be freed by a later call to waitpid().
4551
 */
4552
void
4553
pgwin32_register_deadchild_callback(HANDLE procHandle, DWORD procId)
4554
{
4555
  win32_deadchild_waitinfo *childinfo;
4556
4557
  childinfo = palloc(sizeof(win32_deadchild_waitinfo));
4558
  childinfo->procHandle = procHandle;
4559
  childinfo->procId = procId;
4560
4561
  if (!RegisterWaitForSingleObject(&childinfo->waitHandle,
4562
                   procHandle,
4563
                   pgwin32_deadchild_callback,
4564
                   childinfo,
4565
                   INFINITE,
4566
                   WT_EXECUTEONLYONCE | WT_EXECUTEINWAITTHREAD))
4567
    ereport(FATAL,
4568
        (errmsg_internal("could not register process for wait: error code %lu",
4569
                 GetLastError())));
4570
}
4571
4572
#endif              /* WIN32 */
4573
4574
/*
4575
 * Initialize one and only handle for monitoring postmaster death.
4576
 *
4577
 * Called once in the postmaster, so that child processes can subsequently
4578
 * monitor if their parent is dead.
4579
 */
4580
static void
4581
InitPostmasterDeathWatchHandle(void)
4582
0
{
4583
0
#ifndef WIN32
4584
4585
  /*
4586
   * Create a pipe. Postmaster holds the write end of the pipe open
4587
   * (POSTMASTER_FD_OWN), and children hold the read end. Children can pass
4588
   * the read file descriptor to select() to wake up in case postmaster
4589
   * dies, or check for postmaster death with a (read() == 0). Children must
4590
   * close the write end as soon as possible after forking, because EOF
4591
   * won't be signaled in the read end until all processes have closed the
4592
   * write fd. That is taken care of in ClosePostmasterPorts().
4593
   */
4594
0
  Assert(MyProcPid == PostmasterPid);
4595
0
  if (pipe(postmaster_alive_fds) < 0)
4596
0
    ereport(FATAL,
4597
0
        (errcode_for_file_access(),
4598
0
         errmsg_internal("could not create pipe to monitor postmaster death: %m")));
4599
4600
  /* Notify fd.c that we've eaten two FDs for the pipe. */
4601
0
  ReserveExternalFD();
4602
0
  ReserveExternalFD();
4603
4604
  /*
4605
   * Set O_NONBLOCK to allow testing for the fd's presence with a read()
4606
   * call.
4607
   */
4608
0
  if (fcntl(postmaster_alive_fds[POSTMASTER_FD_WATCH], F_SETFL, O_NONBLOCK) == -1)
4609
0
    ereport(FATAL,
4610
0
        (errcode_for_socket_access(),
4611
0
         errmsg_internal("could not set postmaster death monitoring pipe to nonblocking mode: %m")));
4612
#else
4613
4614
  /*
4615
   * On Windows, we use a process handle for the same purpose.
4616
   */
4617
  if (DuplicateHandle(GetCurrentProcess(),
4618
            GetCurrentProcess(),
4619
            GetCurrentProcess(),
4620
            &PostmasterHandle,
4621
            0,
4622
            TRUE,
4623
            DUPLICATE_SAME_ACCESS) == 0)
4624
    ereport(FATAL,
4625
        (errmsg_internal("could not duplicate postmaster handle: error code %lu",
4626
                 GetLastError())));
4627
#endif              /* WIN32 */
4628
0
}