src/backend/libpq/pqcomm.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * pqcomm.c
   4  *        Communication functions between the Frontend and the Backend
   5  *
   6  * These routines handle the low-level details of communication between
   7  * frontend and backend.  They just shove data across the communication
   8  * channel, and are ignorant of the semantics of the data --- or would be,
   9  * except for major brain damage in the design of the old COPY OUT protocol.
  10  * Unfortunately, COPY OUT was designed to commandeer the communication
  11  * channel (it just transfers data without wrapping it into messages).
  12  * No other messages can be sent while COPY OUT is in progress; and if the
  13  * copy is aborted by an ereport(ERROR), we need to close out the copy so that
  14  * the frontend gets back into sync.  Therefore, these routines have to be
  15  * aware of COPY OUT state.  (New COPY-OUT is message-based and does *not*
  16  * set the DoingCopyOut flag.)
  17  *
  18  * NOTE: generally, it's a bad idea to emit outgoing messages directly with
  19  * pq_putbytes(), especially if the message would require multiple calls
  20  * to send.  Instead, use the routines in pqformat.c to construct the message
  21  * in a buffer and then emit it in one call to pq_putmessage.  This ensures
  22  * that the channel will not be clogged by an incomplete message if execution
  23  * is aborted by ereport(ERROR) partway through the message.  The only
  24  * non-libpq code that should call pq_putbytes directly is old-style COPY OUT.
  25  *
  26  * At one time, libpq was shared between frontend and backend, but now
  27  * the backend's "backend/libpq" is quite separate from "interfaces/libpq".
  28  * All that remains is similarities of names to trap the unwary...
  29  *
  30  * Portions Copyright (c) 1996-2010, PostgreSQL Global Development Group
  31  * Portions Copyright (c) 1994, Regents of the University of California
  32  *
  33  *      $PostgreSQL: pgsql/src/backend/libpq/pqcomm.c,v 1.202 2010/01/15 09:19:02 heikki Exp $
  34  *
  35  *-------------------------------------------------------------------------
  36  */
  37
  38 /*------------------------
  39  * INTERFACE ROUTINES
  40  *
  41  * setup/teardown:
  42  *              StreamServerPort        - Open postmaster's server port
  43  *              StreamConnection        - Create new connection with client
  44  *              StreamClose                     - Close a client/backend connection
  45  *              TouchSocketFile         - Protect socket file against /tmp cleaners
  46  *              pq_init                 - initialize libpq at backend startup
  47  *              pq_comm_reset   - reset libpq during error recovery
  48  *              pq_close                - shutdown libpq at backend exit
  49  *
  50  * low-level I/O:
  51  *              pq_getbytes             - get a known number of bytes from connection
  52  *              pq_getstring    - get a null terminated string from connection
  53  *              pq_getmessage   - get a message with length word from connection
  54  *              pq_getbyte              - get next byte from connection
  55  *              pq_peekbyte             - peek at next byte from connection
  56  *              pq_putbytes             - send bytes to connection (not flushed until pq_flush)
  57  *              pq_flush                - flush pending output
  58  *              pq_getbyte_if_available - get a byte if available without blocking
  59  *
  60  * message-level I/O (and old-style-COPY-OUT cruft):
  61  *              pq_putmessage   - send a normal message (suppressed in COPY OUT mode)
  62  *              pq_startcopyout - inform libpq that a COPY OUT transfer is beginning
  63  *              pq_endcopyout   - end a COPY OUT transfer
  64  *
  65  *------------------------
  66  */
  67 #include "postgres.h"
  68
  69 #include <signal.h>
  70 #include <fcntl.h>
  71 #include <grp.h>
  72 #include <unistd.h>
  73 #include <sys/file.h>
  74 #include <sys/socket.h>
  75 #include <sys/stat.h>
  76 #include <sys/time.h>
  77 #include <netdb.h>
  78 #include <netinet/in.h>
  79 #ifdef HAVE_NETINET_TCP_H
  80 #include <netinet/tcp.h>
  81 #endif
  82 #include <arpa/inet.h>
  83 #ifdef HAVE_UTIME_H
  84 #include <utime.h>
  85 #endif
  86
  87 #include "libpq/ip.h"
  88 #include "libpq/libpq.h"
  89 #include "miscadmin.h"
  90 #include "storage/ipc.h"
  91 #include "utils/guc.h"
  92
  93 /*
  94  * Configuration options
  95  */
  96 int                     Unix_socket_permissions;
  97 char       *Unix_socket_group;
  98
  99
 100 /* Where the Unix socket file is */
 101 static char sock_path[MAXPGPATH];
 102
 103
 104 /*
 105  * Buffers for low-level I/O
 106  */
 107
 108 #define PQ_BUFFER_SIZE 8192
 109
 110 static char PqSendBuffer[PQ_BUFFER_SIZE];
 111 static int      PqSendPointer;          /* Next index to store a byte in PqSendBuffer */
 112
 113 static char PqRecvBuffer[PQ_BUFFER_SIZE];
 114 static int      PqRecvPointer;          /* Next index to read a byte from PqRecvBuffer */
 115 static int      PqRecvLength;           /* End of data available in PqRecvBuffer */
 116
 117 /*
 118  * Message status
 119  */
 120 static bool PqCommBusy;
 121 static bool DoingCopyOut;
 122
 123
 124 /* Internal functions */
 125 static void pq_close(int code, Datum arg);
 126 static int      internal_putbytes(const char *s, size_t len);
 127 static int      internal_flush(void);
 128
 129 #ifdef HAVE_UNIX_SOCKETS
 130 static int      Lock_AF_UNIX(unsigned short portNumber, char *unixSocketName);
 131 static int      Setup_AF_UNIX(void);
 132 #endif   /* HAVE_UNIX_SOCKETS */
 133
 134
 135 /* --------------------------------
 136  *              pq_init - initialize libpq at backend startup
 137  * --------------------------------
 138  */
 139 void
 140 pq_init(void)
 141 {
 142         PqSendPointer = PqRecvPointer = PqRecvLength = 0;
 143         PqCommBusy = false;
 144         DoingCopyOut = false;
 145         on_proc_exit(pq_close, 0);
 146 }
 147
 148 /* --------------------------------
 149  *              pq_comm_reset - reset libpq during error recovery
 150  *
 151  * This is called from error recovery at the outer idle loop.  It's
 152  * just to get us out of trouble if we somehow manage to elog() from
 153  * inside a pqcomm.c routine (which ideally will never happen, but...)
 154  * --------------------------------
 155  */
 156 void
 157 pq_comm_reset(void)
 158 {
 159         /* Do not throw away pending data, but do reset the busy flag */
 160         PqCommBusy = false;
 161         /* We can abort any old-style COPY OUT, too */
 162         pq_endcopyout(true);
 163 }
 164
 165 /* --------------------------------
 166  *              pq_close - shutdown libpq at backend exit
 167  *
 168  * Note: in a standalone backend MyProcPort will be null,
 169  * don't crash during exit...
 170  * --------------------------------
 171  */
 172 static void
 173 pq_close(int code, Datum arg)
 174 {
 175         if (MyProcPort != NULL)
 176         {
 177 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
 178 #ifdef ENABLE_GSS
 179                 OM_uint32       min_s;
 180
 181                 /* Shutdown GSSAPI layer */
 182                 if (MyProcPort->gss->ctx != GSS_C_NO_CONTEXT)
 183                         gss_delete_sec_context(&min_s, &MyProcPort->gss->ctx, NULL);
 184
 185                 if (MyProcPort->gss->cred != GSS_C_NO_CREDENTIAL)
 186                         gss_release_cred(&min_s, &MyProcPort->gss->cred);
 187 #endif   /* ENABLE_GSS */
 188                 /* GSS and SSPI share the port->gss struct */
 189
 190                 free(MyProcPort->gss);
 191 #endif   /* ENABLE_GSS || ENABLE_SSPI */
 192
 193                 /* Cleanly shut down SSL layer */
 194                 secure_close(MyProcPort);
 195
 196                 /*
 197                  * Formerly we did an explicit close() here, but it seems better to
 198                  * leave the socket open until the process dies.  This allows clients
 199                  * to perform a "synchronous close" if they care --- wait till the
 200                  * transport layer reports connection closure, and you can be sure the
 201                  * backend has exited.
 202                  *
 203                  * We do set sock to PGINVALID_SOCKET to prevent any further I/O, though.
 204                  */
 205                 MyProcPort->sock = PGINVALID_SOCKET;
 206         }
 207 }
 208
 209
 210
 211 /*
 212  * Streams -- wrapper around Unix socket system calls
 213  *
 214  *
 215  *              Stream functions are used for vanilla TCP connection protocol.
 216  */
 217
 218
 219 /* StreamDoUnlink()
 220  * Shutdown routine for backend connection
 221  * If a Unix socket is used for communication, explicitly close it.
 222  */
 223 #ifdef HAVE_UNIX_SOCKETS
 224 static void
 225 StreamDoUnlink(int code, Datum arg)
 226 {
 227         Assert(sock_path[0]);
 228         unlink(sock_path);
 229 }
 230 #endif   /* HAVE_UNIX_SOCKETS */
 231
 232 /*
 233  * StreamServerPort -- open a "listening" port to accept connections.
 234  *
 235  * Successfully opened sockets are added to the ListenSocket[] array,
 236  * at the first position that isn't PGINVALID_SOCKET.
 237  *
 238  * RETURNS: STATUS_OK or STATUS_ERROR
 239  */
 240
 241 int
 242 StreamServerPort(int family, char *hostName, unsigned short portNumber,
 243                                  char *unixSocketName,
 244                                  pgsocket ListenSocket[], int MaxListen)
 245 {
 246         pgsocket        fd;
 247         int                     err;
 248         int                     maxconn;
 249         int                     ret;
 250         char            portNumberStr[32];
 251         const char *familyDesc;
 252         char            familyDescBuf[64];
 253         char       *service;
 254         struct addrinfo *addrs = NULL,
 255                            *addr;
 256         struct addrinfo hint;
 257         int                     listen_index = 0;
 258         int                     added = 0;
 259
 260 #if !defined(WIN32) || defined(IPV6_V6ONLY)
 261         int                     one = 1;
 262 #endif
 263
 264         /* Initialize hint structure */
 265         MemSet(&hint, 0, sizeof(hint));
 266         hint.ai_family = family;
 267         hint.ai_flags = AI_PASSIVE;
 268         hint.ai_socktype = SOCK_STREAM;
 269
 270 #ifdef HAVE_UNIX_SOCKETS
 271         if (family == AF_UNIX)
 272         {
 273                 /* Lock_AF_UNIX will also fill in sock_path. */
 274                 if (Lock_AF_UNIX(portNumber, unixSocketName) != STATUS_OK)
 275                         return STATUS_ERROR;
 276                 service = sock_path;
 277         }
 278         else
 279 #endif   /* HAVE_UNIX_SOCKETS */
 280         {
 281                 snprintf(portNumberStr, sizeof(portNumberStr), "%d", portNumber);
 282                 service = portNumberStr;
 283         }
 284
 285         ret = pg_getaddrinfo_all(hostName, service, &hint, &addrs);
 286         if (ret || !addrs)
 287         {
 288                 if (hostName)
 289                         ereport(LOG,
 290                                         (errmsg("could not translate host name \"%s\", service \"%s\" to address: %s",
 291                                                         hostName, service, gai_strerror(ret))));
 292                 else
 293                         ereport(LOG,
 294                                  (errmsg("could not translate service \"%s\" to address: %s",
 295                                                  service, gai_strerror(ret))));
 296                 if (addrs)
 297                         pg_freeaddrinfo_all(hint.ai_family, addrs);
 298                 return STATUS_ERROR;
 299         }
 300
 301         for (addr = addrs; addr; addr = addr->ai_next)
 302         {
 303                 if (!IS_AF_UNIX(family) && IS_AF_UNIX(addr->ai_family))
 304                 {
 305                         /*
 306                          * Only set up a unix domain socket when they really asked for it.
 307                          * The service/port is different in that case.
 308                          */
 309                         continue;
 310                 }
 311
 312                 /* See if there is still room to add 1 more socket. */
 313                 for (; listen_index < MaxListen; listen_index++)
 314                 {
 315                         if (ListenSocket[listen_index] == PGINVALID_SOCKET)
 316                                 break;
 317                 }
 318                 if (listen_index >= MaxListen)
 319                 {
 320                         ereport(LOG,
 321                                         (errmsg("could not bind to all requested addresses: MAXLISTEN (%d) exceeded",
 322                                                         MaxListen)));
 323                         break;
 324                 }
 325
 326                 /* set up family name for possible error messages */
 327                 switch (addr->ai_family)
 328                 {
 329                         case AF_INET:
 330                                 familyDesc = _("IPv4");
 331                                 break;
 332 #ifdef HAVE_IPV6
 333                         case AF_INET6:
 334                                 familyDesc = _("IPv6");
 335                                 break;
 336 #endif
 337 #ifdef HAVE_UNIX_SOCKETS
 338                         case AF_UNIX:
 339                                 familyDesc = _("Unix");
 340                                 break;
 341 #endif
 342                         default:
 343                                 snprintf(familyDescBuf, sizeof(familyDescBuf),
 344                                                  _("unrecognized address family %d"),
 345                                                  addr->ai_family);
 346                                 familyDesc = familyDescBuf;
 347                                 break;
 348                 }
 349
 350                 if ((fd = socket(addr->ai_family, SOCK_STREAM, 0)) < 0)
 351                 {
 352                         ereport(LOG,
 353                                         (errcode_for_socket_access(),
 354                         /* translator: %s is IPv4, IPv6, or Unix */
 355                                          errmsg("could not create %s socket: %m",
 356                                                         familyDesc)));
 357                         continue;
 358                 }
 359
 360 #ifndef WIN32
 361
 362                 /*
 363                  * Without the SO_REUSEADDR flag, a new postmaster can't be started
 364                  * right away after a stop or crash, giving "address already in use"
 365                  * error on TCP ports.
 366                  *
 367                  * On win32, however, this behavior only happens if the
 368                  * SO_EXLUSIVEADDRUSE is set. With SO_REUSEADDR, win32 allows multiple
 369                  * servers to listen on the same address, resulting in unpredictable
 370                  * behavior. With no flags at all, win32 behaves as Unix with
 371                  * SO_REUSEADDR.
 372                  */
 373                 if (!IS_AF_UNIX(addr->ai_family))
 374                 {
 375                         if ((setsockopt(fd, SOL_SOCKET, SO_REUSEADDR,
 376                                                         (char *) &one, sizeof(one))) == -1)
 377                         {
 378                                 ereport(LOG,
 379                                                 (errcode_for_socket_access(),
 380                                                  errmsg("setsockopt(SO_REUSEADDR) failed: %m")));
 381                                 closesocket(fd);
 382                                 continue;
 383                         }
 384                 }
 385 #endif
 386
 387 #ifdef IPV6_V6ONLY
 388                 if (addr->ai_family == AF_INET6)
 389                 {
 390                         if (setsockopt(fd, IPPROTO_IPV6, IPV6_V6ONLY,
 391                                                    (char *) &one, sizeof(one)) == -1)
 392                         {
 393                                 ereport(LOG,
 394                                                 (errcode_for_socket_access(),
 395                                                  errmsg("setsockopt(IPV6_V6ONLY) failed: %m")));
 396                                 closesocket(fd);
 397                                 continue;
 398                         }
 399                 }
 400 #endif
 401
 402                 /*
 403                  * Note: This might fail on some OS's, like Linux older than
 404                  * 2.4.21-pre3, that don't have the IPV6_V6ONLY socket option, and map
 405                  * ipv4 addresses to ipv6.      It will show ::ffff:ipv4 for all ipv4
 406                  * connections.
 407                  */
 408                 err = bind(fd, addr->ai_addr, addr->ai_addrlen);
 409                 if (err < 0)
 410                 {
 411                         ereport(LOG,
 412                                         (errcode_for_socket_access(),
 413                         /* translator: %s is IPv4, IPv6, or Unix */
 414                                          errmsg("could not bind %s socket: %m",
 415                                                         familyDesc),
 416                                          (IS_AF_UNIX(addr->ai_family)) ?
 417                                   errhint("Is another postmaster already running on port %d?"
 418                                                   " If not, remove socket file \"%s\" and retry.",
 419                                                   (int) portNumber, sock_path) :
 420                                   errhint("Is another postmaster already running on port %d?"
 421                                                   " If not, wait a few seconds and retry.",
 422                                                   (int) portNumber)));
 423                         closesocket(fd);
 424                         continue;
 425                 }
 426
 427 #ifdef HAVE_UNIX_SOCKETS
 428                 if (addr->ai_family == AF_UNIX)
 429                 {
 430                         if (Setup_AF_UNIX() != STATUS_OK)
 431                         {
 432                                 closesocket(fd);
 433                                 break;
 434                         }
 435                 }
 436 #endif
 437
 438                 /*
 439                  * Select appropriate accept-queue length limit.  PG_SOMAXCONN is only
 440                  * intended to provide a clamp on the request on platforms where an
 441                  * overly large request provokes a kernel error (are there any?).
 442                  */
 443                 maxconn = MaxBackends * 2;
 444                 if (maxconn > PG_SOMAXCONN)
 445                         maxconn = PG_SOMAXCONN;
 446
 447                 err = listen(fd, maxconn);
 448                 if (err < 0)
 449                 {
 450                         ereport(LOG,
 451                                         (errcode_for_socket_access(),
 452                         /* translator: %s is IPv4, IPv6, or Unix */
 453                                          errmsg("could not listen on %s socket: %m",
 454                                                         familyDesc)));
 455                         closesocket(fd);
 456                         continue;
 457                 }
 458                 ListenSocket[listen_index] = fd;
 459                 added++;
 460         }
 461
 462         pg_freeaddrinfo_all(hint.ai_family, addrs);
 463
 464         if (!added)
 465                 return STATUS_ERROR;
 466
 467         return STATUS_OK;
 468 }
 469
 470
 471 #ifdef HAVE_UNIX_SOCKETS
 472
 473 /*
 474  * Lock_AF_UNIX -- configure unix socket file path
 475  */
 476 static int
 477 Lock_AF_UNIX(unsigned short portNumber, char *unixSocketName)
 478 {
 479         UNIXSOCK_PATH(sock_path, portNumber, unixSocketName);
 480
 481         /*
 482          * Grab an interlock file associated with the socket file.
 483          */
 484         CreateSocketLockFile(sock_path, true);
 485
 486         /*
 487          * Once we have the interlock, we can safely delete any pre-existing
 488          * socket file to avoid failure at bind() time.
 489          */
 490         unlink(sock_path);
 491
 492         return STATUS_OK;
 493 }
 494
 495
 496 /*
 497  * Setup_AF_UNIX -- configure unix socket permissions
 498  */
 499 static int
 500 Setup_AF_UNIX(void)
 501 {
 502         /* Arrange to unlink the socket file at exit */
 503         on_proc_exit(StreamDoUnlink, 0);
 504
 505         /*
 506          * Fix socket ownership/permission if requested.  Note we must do this
 507          * before we listen() to avoid a window where unwanted connections could
 508          * get accepted.
 509          */
 510         Assert(Unix_socket_group);
 511         if (Unix_socket_group[0] != '\0')
 512         {
 513 #ifdef WIN32
 514                 elog(WARNING, "configuration item unix_socket_group is not supported on this platform");
 515 #else
 516                 char       *endptr;
 517                 unsigned long val;
 518                 gid_t           gid;
 519
 520                 val = strtoul(Unix_socket_group, &endptr, 10);
 521                 if (*endptr == '\0')
 522                 {                                               /* numeric group id */
 523                         gid = val;
 524                 }
 525                 else
 526                 {                                               /* convert group name to id */
 527                         struct group *gr;
 528
 529                         gr = getgrnam(Unix_socket_group);
 530                         if (!gr)
 531                         {
 532                                 ereport(LOG,
 533                                                 (errmsg("group \"%s\" does not exist",
 534                                                                 Unix_socket_group)));
 535                                 return STATUS_ERROR;
 536                         }
 537                         gid = gr->gr_gid;
 538                 }
 539                 if (chown(sock_path, -1, gid) == -1)
 540                 {
 541                         ereport(LOG,
 542                                         (errcode_for_file_access(),
 543                                          errmsg("could not set group of file \"%s\": %m",
 544                                                         sock_path)));
 545                         return STATUS_ERROR;
 546                 }
 547 #endif
 548         }
 549
 550         if (chmod(sock_path, Unix_socket_permissions) == -1)
 551         {
 552                 ereport(LOG,
 553                                 (errcode_for_file_access(),
 554                                  errmsg("could not set permissions of file \"%s\": %m",
 555                                                 sock_path)));
 556                 return STATUS_ERROR;
 557         }
 558         return STATUS_OK;
 559 }
 560 #endif   /* HAVE_UNIX_SOCKETS */
 561
 562
 563 /*
 564  * StreamConnection -- create a new connection with client using
 565  *              server port.  Set port->sock to the FD of the new connection.
 566  *
 567  * ASSUME: that this doesn't need to be non-blocking because
 568  *              the Postmaster uses select() to tell when the server master
 569  *              socket is ready for accept().
 570  *
 571  * RETURNS: STATUS_OK or STATUS_ERROR
 572  */
 573 int
 574 StreamConnection(pgsocket server_fd, Port *port)
 575 {
 576         /* accept connection and fill in the client (remote) address */
 577         port->raddr.salen = sizeof(port->raddr.addr);
 578         if ((port->sock = accept(server_fd,
 579                                                          (struct sockaddr *) & port->raddr.addr,
 580                                                          &port->raddr.salen)) < 0)
 581         {
 582                 ereport(LOG,
 583                                 (errcode_for_socket_access(),
 584                                  errmsg("could not accept new connection: %m")));
 585
 586                 /*
 587                  * If accept() fails then postmaster.c will still see the server
 588                  * socket as read-ready, and will immediately try again.  To avoid
 589                  * uselessly sucking lots of CPU, delay a bit before trying again.
 590                  * (The most likely reason for failure is being out of kernel file
 591                  * table slots; we can do little except hope some will get freed up.)
 592                  */
 593                 pg_usleep(100000L);             /* wait 0.1 sec */
 594                 return STATUS_ERROR;
 595         }
 596
 597 #ifdef SCO_ACCEPT_BUG
 598
 599         /*
 600          * UnixWare 7+ and OpenServer 5.0.4 are known to have this bug, but it
 601          * shouldn't hurt to catch it for all versions of those platforms.
 602          */
 603         if (port->raddr.addr.ss_family == 0)
 604                 port->raddr.addr.ss_family = AF_UNIX;
 605 #endif
 606
 607         /* fill in the server (local) address */
 608         port->laddr.salen = sizeof(port->laddr.addr);
 609         if (getsockname(port->sock,
 610                                         (struct sockaddr *) & port->laddr.addr,
 611                                         &port->laddr.salen) < 0)
 612         {
 613                 elog(LOG, "getsockname() failed: %m");
 614                 return STATUS_ERROR;
 615         }
 616
 617         /* select NODELAY and KEEPALIVE options if it's a TCP connection */
 618         if (!IS_AF_UNIX(port->laddr.addr.ss_family))
 619         {
 620                 int                     on;
 621
 622 #ifdef  TCP_NODELAY
 623                 on = 1;
 624                 if (setsockopt(port->sock, IPPROTO_TCP, TCP_NODELAY,
 625                                            (char *) &on, sizeof(on)) < 0)
 626                 {
 627                         elog(LOG, "setsockopt(TCP_NODELAY) failed: %m");
 628                         return STATUS_ERROR;
 629                 }
 630 #endif
 631                 on = 1;
 632                 if (setsockopt(port->sock, SOL_SOCKET, SO_KEEPALIVE,
 633                                            (char *) &on, sizeof(on)) < 0)
 634                 {
 635                         elog(LOG, "setsockopt(SO_KEEPALIVE) failed: %m");
 636                         return STATUS_ERROR;
 637                 }
 638
 639 #ifdef WIN32
 640
 641                 /*
 642                  * This is a Win32 socket optimization.  The ideal size is 32k.
 643                  * http://support.microsoft.com/kb/823764/EN-US/
 644                  */
 645                 on = PQ_BUFFER_SIZE * 4;
 646                 if (setsockopt(port->sock, SOL_SOCKET, SO_SNDBUF, (char *) &on,
 647                                            sizeof(on)) < 0)
 648                 {
 649                         elog(LOG, "setsockopt(SO_SNDBUF) failed: %m");
 650                         return STATUS_ERROR;
 651                 }
 652 #endif
 653
 654                 /*
 655                  * Also apply the current keepalive parameters.  If we fail to set a
 656                  * parameter, don't error out, because these aren't universally
 657                  * supported.  (Note: you might think we need to reset the GUC
 658                  * variables to 0 in such a case, but it's not necessary because the
 659                  * show hooks for these variables report the truth anyway.)
 660                  */
 661                 (void) pq_setkeepalivesidle(tcp_keepalives_idle, port);
 662                 (void) pq_setkeepalivesinterval(tcp_keepalives_interval, port);
 663                 (void) pq_setkeepalivescount(tcp_keepalives_count, port);
 664         }
 665
 666         return STATUS_OK;
 667 }
 668
 669 /*
 670  * StreamClose -- close a client/backend connection
 671  *
 672  * NOTE: this is NOT used to terminate a session; it is just used to release
 673  * the file descriptor in a process that should no longer have the socket
 674  * open.  (For example, the postmaster calls this after passing ownership
 675  * of the connection to a child process.)  It is expected that someone else
 676  * still has the socket open.  So, we only want to close the descriptor,
 677  * we do NOT want to send anything to the far end.
 678  */
 679 void
 680 StreamClose(pgsocket sock)
 681 {
 682         closesocket(sock);
 683 }
 684
 685 /*
 686  * TouchSocketFile -- mark socket file as recently accessed
 687  *
 688  * This routine should be called every so often to ensure that the socket
 689  * file has a recent mod date (ordinary operations on sockets usually won't
 690  * change the mod date).  That saves it from being removed by
 691  * overenthusiastic /tmp-directory-cleaner daemons.  (Another reason we should
 692  * never have put the socket file in /tmp...)
 693  */
 694 void
 695 TouchSocketFile(void)
 696 {
 697         /* Do nothing if we did not create a socket... */
 698         if (sock_path[0] != '\0')
 699         {
 700                 /*
 701                  * utime() is POSIX standard, utimes() is a common alternative. If we
 702                  * have neither, there's no way to affect the mod or access time of
 703                  * the socket :-(
 704                  *
 705                  * In either path, we ignore errors; there's no point in complaining.
 706                  */
 707 #ifdef HAVE_UTIME
 708                 utime(sock_path, NULL);
 709 #else                                                   /* !HAVE_UTIME */
 710 #ifdef HAVE_UTIMES
 711                 utimes(sock_path, NULL);
 712 #endif   /* HAVE_UTIMES */
 713 #endif   /* HAVE_UTIME */
 714         }
 715 }
 716
 717
 718 /* --------------------------------
 719  * Low-level I/O routines begin here.
 720  *
 721  * These routines communicate with a frontend client across a connection
 722  * already established by the preceding routines.
 723  * --------------------------------
 724  */
 725
 726
 727 /* --------------------------------
 728  *              pq_recvbuf - load some bytes into the input buffer
 729  *
 730  *              returns 0 if OK, EOF if trouble
 731  * --------------------------------
 732  */
 733 static int
 734 pq_recvbuf(void)
 735 {
 736         if (PqRecvPointer > 0)
 737         {
 738                 if (PqRecvLength > PqRecvPointer)
 739                 {
 740                         /* still some unread data, left-justify it in the buffer */
 741                         memmove(PqRecvBuffer, PqRecvBuffer + PqRecvPointer,
 742                                         PqRecvLength - PqRecvPointer);
 743                         PqRecvLength -= PqRecvPointer;
 744                         PqRecvPointer = 0;
 745                 }
 746                 else
 747                         PqRecvLength = PqRecvPointer = 0;
 748         }
 749
 750         /* Can fill buffer from PqRecvLength and upwards */
 751         for (;;)
 752         {
 753                 int                     r;
 754
 755                 r = secure_read(MyProcPort, PqRecvBuffer + PqRecvLength,
 756                                                 PQ_BUFFER_SIZE - PqRecvLength);
 757
 758                 if (r < 0)
 759                 {
 760                         if (errno == EINTR)
 761                                 continue;               /* Ok if interrupted */
 762
 763                         /*
 764                          * Careful: an ereport() that tries to write to the client would
 765                          * cause recursion to here, leading to stack overflow and core
 766                          * dump!  This message must go *only* to the postmaster log.
 767                          */
 768                         ereport(COMMERROR,
 769                                         (errcode_for_socket_access(),
 770                                          errmsg("could not receive data from client: %m")));
 771                         return EOF;
 772                 }
 773                 if (r == 0)
 774                 {
 775                         /*
 776                          * EOF detected.  We used to write a log message here, but it's
 777                          * better to expect the ultimate caller to do that.
 778                          */
 779                         return EOF;
 780                 }
 781                 /* r contains number of bytes read, so just incr length */
 782                 PqRecvLength += r;
 783                 return 0;
 784         }
 785 }
 786
 787 /* --------------------------------
 788  *              pq_getbyte      - get a single byte from connection, or return EOF
 789  * --------------------------------
 790  */
 791 int
 792 pq_getbyte(void)
 793 {
 794         while (PqRecvPointer >= PqRecvLength)
 795         {
 796                 if (pq_recvbuf())               /* If nothing in buffer, then recv some */
 797                         return EOF;                     /* Failed to recv data */
 798         }
 799         return (unsigned char) PqRecvBuffer[PqRecvPointer++];
 800 }
 801
 802 /* --------------------------------
 803  *              pq_peekbyte             - peek at next byte from connection
 804  *
 805  *       Same as pq_getbyte() except we don't advance the pointer.
 806  * --------------------------------
 807  */
 808 int
 809 pq_peekbyte(void)
 810 {
 811         while (PqRecvPointer >= PqRecvLength)
 812         {
 813                 if (pq_recvbuf())               /* If nothing in buffer, then recv some */
 814                         return EOF;                     /* Failed to recv data */
 815         }
 816         return (unsigned char) PqRecvBuffer[PqRecvPointer];
 817 }
 818
 819
 820 /* --------------------------------
 821  *              pq_getbyte_if_available - get a single byte from connection,
 822  *                      if available
 823  *
 824  * The received byte is stored in *c. Returns 1 if a byte was read, 0 if
 825  * if no data was available, or EOF.
 826  * --------------------------------
 827  */
 828 int
 829 pq_getbyte_if_available(unsigned char *c)
 830 {
 831         int r;
 832
 833         if (PqRecvPointer < PqRecvLength)
 834         {
 835                 *c = PqRecvBuffer[PqRecvPointer++];
 836                 return 1;
 837         }
 838
 839         /* Temporarily put the socket into non-blocking mode */
 840         if (!pg_set_noblock(MyProcPort->sock))
 841                 ereport(ERROR,
 842                                 (errmsg("couldn't put socket to non-blocking mode: %m")));
 843         MyProcPort->noblock = true;
 844         PG_TRY();
 845         {
 846                 r = secure_read(MyProcPort, c, 1);
 847         }
 848         PG_CATCH();
 849         {
 850                 /*
 851                  * The rest of the backend code assumes the socket is in blocking
 852                  * mode, so treat failure as FATAL.
 853                  */
 854                 if (!pg_set_block(MyProcPort->sock))
 855                         ereport(FATAL,
 856                                         (errmsg("couldn't put socket to blocking mode: %m")));
 857                 MyProcPort->noblock = false;
 858                 PG_RE_THROW();
 859         }
 860         PG_END_TRY();
 861         if (!pg_set_block(MyProcPort->sock))
 862                 ereport(FATAL,
 863                                 (errmsg("couldn't put socket to blocking mode: %m")));
 864         MyProcPort->noblock = false;
 865
 866         return r;
 867 }
 868
 869 /* --------------------------------
 870  *              pq_getbytes             - get a known number of bytes from connection
 871  *
 872  *              returns 0 if OK, EOF if trouble
 873  * --------------------------------
 874  */
 875 int
 876 pq_getbytes(char *s, size_t len)
 877 {
 878         size_t          amount;
 879
 880         while (len > 0)
 881         {
 882                 while (PqRecvPointer >= PqRecvLength)
 883                 {
 884                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
 885                                 return EOF;             /* Failed to recv data */
 886                 }
 887                 amount = PqRecvLength - PqRecvPointer;
 888                 if (amount > len)
 889                         amount = len;
 890                 memcpy(s, PqRecvBuffer + PqRecvPointer, amount);
 891                 PqRecvPointer += amount;
 892                 s += amount;
 893                 len -= amount;
 894         }
 895         return 0;
 896 }
 897
 898 /* --------------------------------
 899  *              pq_discardbytes         - throw away a known number of bytes
 900  *
 901  *              same as pq_getbytes except we do not copy the data to anyplace.
 902  *              this is used for resynchronizing after read errors.
 903  *
 904  *              returns 0 if OK, EOF if trouble
 905  * --------------------------------
 906  */
 907 static int
 908 pq_discardbytes(size_t len)
 909 {
 910         size_t          amount;
 911
 912         while (len > 0)
 913         {
 914                 while (PqRecvPointer >= PqRecvLength)
 915                 {
 916                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
 917                                 return EOF;             /* Failed to recv data */
 918                 }
 919                 amount = PqRecvLength - PqRecvPointer;
 920                 if (amount > len)
 921                         amount = len;
 922                 PqRecvPointer += amount;
 923                 len -= amount;
 924         }
 925         return 0;
 926 }
 927
 928 /* --------------------------------
 929  *              pq_getstring    - get a null terminated string from connection
 930  *
 931  *              The return value is placed in an expansible StringInfo, which has
 932  *              already been initialized by the caller.
 933  *
 934  *              This is used only for dealing with old-protocol clients.  The idea
 935  *              is to produce a StringInfo that looks the same as we would get from
 936  *              pq_getmessage() with a newer client; we will then process it with
 937  *              pq_getmsgstring.  Therefore, no character set conversion is done here,
 938  *              even though this is presumably useful only for text.
 939  *
 940  *              returns 0 if OK, EOF if trouble
 941  * --------------------------------
 942  */
 943 int
 944 pq_getstring(StringInfo s)
 945 {
 946         int                     i;
 947
 948         resetStringInfo(s);
 949
 950         /* Read until we get the terminating '\0' */
 951         for (;;)
 952         {
 953                 while (PqRecvPointer >= PqRecvLength)
 954                 {
 955                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
 956                                 return EOF;             /* Failed to recv data */
 957                 }
 958
 959                 for (i = PqRecvPointer; i < PqRecvLength; i++)
 960                 {
 961                         if (PqRecvBuffer[i] == '\0')
 962                         {
 963                                 /* include the '\0' in the copy */
 964                                 appendBinaryStringInfo(s, PqRecvBuffer + PqRecvPointer,
 965                                                                            i - PqRecvPointer + 1);
 966                                 PqRecvPointer = i + 1;  /* advance past \0 */
 967                                 return 0;
 968                         }
 969                 }
 970
 971                 /* If we're here we haven't got the \0 in the buffer yet. */
 972                 appendBinaryStringInfo(s, PqRecvBuffer + PqRecvPointer,
 973                                                            PqRecvLength - PqRecvPointer);
 974                 PqRecvPointer = PqRecvLength;
 975         }
 976 }
 977
 978
 979 /* --------------------------------
 980  *              pq_getmessage   - get a message with length word from connection
 981  *
 982  *              The return value is placed in an expansible StringInfo, which has
 983  *              already been initialized by the caller.
 984  *              Only the message body is placed in the StringInfo; the length word
 985  *              is removed.  Also, s->cursor is initialized to zero for convenience
 986  *              in scanning the message contents.
 987  *
 988  *              If maxlen is not zero, it is an upper limit on the length of the
 989  *              message we are willing to accept.  We abort the connection (by
 990  *              returning EOF) if client tries to send more than that.
 991  *
 992  *              returns 0 if OK, EOF if trouble
 993  * --------------------------------
 994  */
 995 int
 996 pq_getmessage(StringInfo s, int maxlen)
 997 {
 998         int32           len;
 999
1000         resetStringInfo(s);
1001
1002         /* Read message length word */
1003         if (pq_getbytes((char *) &len, 4) == EOF)
1004         {
1005                 ereport(COMMERROR,
1006                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1007                                  errmsg("unexpected EOF within message length word")));
1008                 return EOF;
1009         }
1010
1011         len = ntohl(len);
1012
1013         if (len < 4 ||
1014                 (maxlen > 0 && len > maxlen))
1015         {
1016                 ereport(COMMERROR,
1017                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1018                                  errmsg("invalid message length")));
1019                 return EOF;
1020         }
1021
1022         len -= 4;                                       /* discount length itself */
1023
1024         if (len > 0)
1025         {
1026                 /*
1027                  * Allocate space for message.  If we run out of room (ridiculously
1028                  * large message), we will elog(ERROR), but we want to discard the
1029                  * message body so as not to lose communication sync.
1030                  */
1031                 PG_TRY();
1032                 {
1033                         enlargeStringInfo(s, len);
1034                 }
1035                 PG_CATCH();
1036                 {
1037                         if (pq_discardbytes(len) == EOF)
1038                                 ereport(COMMERROR,
1039                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1040                                                  errmsg("incomplete message from client")));
1041                         PG_RE_THROW();
1042                 }
1043                 PG_END_TRY();
1044
1045                 /* And grab the message */
1046                 if (pq_getbytes(s->data, len) == EOF)
1047                 {
1048                         ereport(COMMERROR,
1049                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
1050                                          errmsg("incomplete message from client")));
1051                         return EOF;
1052                 }
1053                 s->len = len;
1054                 /* Place a trailing null per StringInfo convention */
1055                 s->data[len] = '\0';
1056         }
1057
1058         return 0;
1059 }
1060
1061
1062 /* --------------------------------
1063  *              pq_putbytes             - send bytes to connection (not flushed until pq_flush)
1064  *
1065  *              returns 0 if OK, EOF if trouble
1066  * --------------------------------
1067  */
1068 int
1069 pq_putbytes(const char *s, size_t len)
1070 {
1071         int                     res;
1072
1073         /* Should only be called by old-style COPY OUT */
1074         Assert(DoingCopyOut);
1075         /* No-op if reentrant call */
1076         if (PqCommBusy)
1077                 return 0;
1078         PqCommBusy = true;
1079         res = internal_putbytes(s, len);
1080         PqCommBusy = false;
1081         return res;
1082 }
1083
1084 static int
1085 internal_putbytes(const char *s, size_t len)
1086 {
1087         size_t          amount;
1088
1089         while (len > 0)
1090         {
1091                 /* If buffer is full, then flush it out */
1092                 if (PqSendPointer >= PQ_BUFFER_SIZE)
1093                         if (internal_flush())
1094                                 return EOF;
1095                 amount = PQ_BUFFER_SIZE - PqSendPointer;
1096                 if (amount > len)
1097                         amount = len;
1098                 memcpy(PqSendBuffer + PqSendPointer, s, amount);
1099                 PqSendPointer += amount;
1100                 s += amount;
1101                 len -= amount;
1102         }
1103         return 0;
1104 }
1105
1106 /* --------------------------------
1107  *              pq_flush                - flush pending output
1108  *
1109  *              returns 0 if OK, EOF if trouble
1110  * --------------------------------
1111  */
1112 int
1113 pq_flush(void)
1114 {
1115         int                     res;
1116
1117         /* No-op if reentrant call */
1118         if (PqCommBusy)
1119                 return 0;
1120         PqCommBusy = true;
1121         res = internal_flush();
1122         PqCommBusy = false;
1123         return res;
1124 }
1125
1126 static int
1127 internal_flush(void)
1128 {
1129         static int      last_reported_send_errno = 0;
1130
1131         char       *bufptr = PqSendBuffer;
1132         char       *bufend = PqSendBuffer + PqSendPointer;
1133
1134         while (bufptr < bufend)
1135         {
1136                 int                     r;
1137
1138                 r = secure_write(MyProcPort, bufptr, bufend - bufptr);
1139
1140                 if (r <= 0)
1141                 {
1142                         if (errno == EINTR)
1143                                 continue;               /* Ok if we were interrupted */
1144
1145                         /*
1146                          * Careful: an ereport() that tries to write to the client would
1147                          * cause recursion to here, leading to stack overflow and core
1148                          * dump!  This message must go *only* to the postmaster log.
1149                          *
1150                          * If a client disconnects while we're in the midst of output, we
1151                          * might write quite a bit of data before we get to a safe query
1152                          * abort point.  So, suppress duplicate log messages.
1153                          */
1154                         if (errno != last_reported_send_errno)
1155                         {
1156                                 last_reported_send_errno = errno;
1157                                 ereport(COMMERROR,
1158                                                 (errcode_for_socket_access(),
1159                                                  errmsg("could not send data to client: %m")));
1160                         }
1161
1162                         /*
1163                          * We drop the buffered data anyway so that processing can
1164                          * continue, even though we'll probably quit soon.
1165                          */
1166                         PqSendPointer = 0;
1167                         return EOF;
1168                 }
1169
1170                 last_reported_send_errno = 0;   /* reset after any successful send */
1171                 bufptr += r;
1172         }
1173
1174         PqSendPointer = 0;
1175         return 0;
1176 }
1177
1178
1179 /* --------------------------------
1180  * Message-level I/O routines begin here.
1181  *
1182  * These routines understand about the old-style COPY OUT protocol.
1183  * --------------------------------
1184  */
1185
1186
1187 /* --------------------------------
1188  *              pq_putmessage   - send a normal message (suppressed in COPY OUT mode)
1189  *
1190  *              If msgtype is not '\0', it is a message type code to place before
1191  *              the message body.  If msgtype is '\0', then the message has no type
1192  *              code (this is only valid in pre-3.0 protocols).
1193  *
1194  *              len is the length of the message body data at *s.  In protocol 3.0
1195  *              and later, a message length word (equal to len+4 because it counts
1196  *              itself too) is inserted by this routine.
1197  *
1198  *              All normal messages are suppressed while old-style COPY OUT is in
1199  *              progress.  (In practice only a few notice messages might get emitted
1200  *              then; dropping them is annoying, but at least they will still appear
1201  *              in the postmaster log.)
1202  *
1203  *              We also suppress messages generated while pqcomm.c is busy.  This
1204  *              avoids any possibility of messages being inserted within other
1205  *              messages.  The only known trouble case arises if SIGQUIT occurs
1206  *              during a pqcomm.c routine --- quickdie() will try to send a warning
1207  *              message, and the most reasonable approach seems to be to drop it.
1208  *
1209  *              returns 0 if OK, EOF if trouble
1210  * --------------------------------
1211  */
1212 int
1213 pq_putmessage(char msgtype, const char *s, size_t len)
1214 {
1215         if (DoingCopyOut || PqCommBusy)
1216                 return 0;
1217         PqCommBusy = true;
1218         if (msgtype)
1219                 if (internal_putbytes(&msgtype, 1))
1220                         goto fail;
1221         if (PG_PROTOCOL_MAJOR(FrontendProtocol) >= 3)
1222         {
1223                 uint32          n32;
1224
1225                 n32 = htonl((uint32) (len + 4));
1226                 if (internal_putbytes((char *) &n32, 4))
1227                         goto fail;
1228         }
1229         if (internal_putbytes(s, len))
1230                 goto fail;
1231         PqCommBusy = false;
1232         return 0;
1233
1234 fail:
1235         PqCommBusy = false;
1236         return EOF;
1237 }
1238
1239 /* --------------------------------
1240  *              pq_startcopyout - inform libpq that an old-style COPY OUT transfer
1241  *                      is beginning
1242  * --------------------------------
1243  */
1244 void
1245 pq_startcopyout(void)
1246 {
1247         DoingCopyOut = true;
1248 }
1249
1250 /* --------------------------------
1251  *              pq_endcopyout   - end an old-style COPY OUT transfer
1252  *
1253  *              If errorAbort is indicated, we are aborting a COPY OUT due to an error,
1254  *              and must send a terminator line.  Since a partial data line might have
1255  *              been emitted, send a couple of newlines first (the first one could
1256  *              get absorbed by a backslash...)  Note that old-style COPY OUT does
1257  *              not allow binary transfers, so a textual terminator is always correct.
1258  * --------------------------------
1259  */
1260 void
1261 pq_endcopyout(bool errorAbort)
1262 {
1263         if (!DoingCopyOut)
1264                 return;
1265         if (errorAbort)
1266                 pq_putbytes("\n\n\\.\n", 5);
1267         /* in non-error case, copy.c will have emitted the terminator line */
1268         DoingCopyOut = false;
1269 }
1270
1271
1272 /*
1273  * Support for TCP Keepalive parameters
1274  */
1275
1276 int
1277 pq_getkeepalivesidle(Port *port)
1278 {
1279 #ifdef TCP_KEEPIDLE
1280         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1281                 return 0;
1282
1283         if (port->keepalives_idle != 0)
1284                 return port->keepalives_idle;
1285
1286         if (port->default_keepalives_idle == 0)
1287         {
1288                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_idle);
1289
1290                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPIDLE,
1291                                            (char *) &port->default_keepalives_idle,
1292                                            &size) < 0)
1293                 {
1294                         elog(LOG, "getsockopt(TCP_KEEPIDLE) failed: %m");
1295                         port->default_keepalives_idle = -1; /* don't know */
1296                 }
1297         }
1298
1299         return port->default_keepalives_idle;
1300 #else
1301         return 0;
1302 #endif
1303 }
1304
1305 int
1306 pq_setkeepalivesidle(int idle, Port *port)
1307 {
1308         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1309                 return STATUS_OK;
1310
1311 #ifdef TCP_KEEPIDLE
1312         if (idle == port->keepalives_idle)
1313                 return STATUS_OK;
1314
1315         if (port->default_keepalives_idle <= 0)
1316         {
1317                 if (pq_getkeepalivesidle(port) < 0)
1318                 {
1319                         if (idle == 0)
1320                                 return STATUS_OK;               /* default is set but unknown */
1321                         else
1322                                 return STATUS_ERROR;
1323                 }
1324         }
1325
1326         if (idle == 0)
1327                 idle = port->default_keepalives_idle;
1328
1329         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPIDLE,
1330                                    (char *) &idle, sizeof(idle)) < 0)
1331         {
1332                 elog(LOG, "setsockopt(TCP_KEEPIDLE) failed: %m");
1333                 return STATUS_ERROR;
1334         }
1335
1336         port->keepalives_idle = idle;
1337 #else
1338         if (idle != 0)
1339         {
1340                 elog(LOG, "setsockopt(TCP_KEEPIDLE) not supported");
1341                 return STATUS_ERROR;
1342         }
1343 #endif
1344
1345         return STATUS_OK;
1346 }
1347
1348 int
1349 pq_getkeepalivesinterval(Port *port)
1350 {
1351 #ifdef TCP_KEEPINTVL
1352         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1353                 return 0;
1354
1355         if (port->keepalives_interval != 0)
1356                 return port->keepalives_interval;
1357
1358         if (port->default_keepalives_interval == 0)
1359         {
1360                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_interval);
1361
1362                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPINTVL,
1363                                            (char *) &port->default_keepalives_interval,
1364                                            &size) < 0)
1365                 {
1366                         elog(LOG, "getsockopt(TCP_KEEPINTVL) failed: %m");
1367                         port->default_keepalives_interval = -1;         /* don't know */
1368                 }
1369         }
1370
1371         return port->default_keepalives_interval;
1372 #else
1373         return 0;
1374 #endif
1375 }
1376
1377 int
1378 pq_setkeepalivesinterval(int interval, Port *port)
1379 {
1380         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1381                 return STATUS_OK;
1382
1383 #ifdef TCP_KEEPINTVL
1384         if (interval == port->keepalives_interval)
1385                 return STATUS_OK;
1386
1387         if (port->default_keepalives_interval <= 0)
1388         {
1389                 if (pq_getkeepalivesinterval(port) < 0)
1390                 {
1391                         if (interval == 0)
1392                                 return STATUS_OK;               /* default is set but unknown */
1393                         else
1394                                 return STATUS_ERROR;
1395                 }
1396         }
1397
1398         if (interval == 0)
1399                 interval = port->default_keepalives_interval;
1400
1401         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPINTVL,
1402                                    (char *) &interval, sizeof(interval)) < 0)
1403         {
1404                 elog(LOG, "setsockopt(TCP_KEEPINTVL) failed: %m");
1405                 return STATUS_ERROR;
1406         }
1407
1408         port->keepalives_interval = interval;
1409 #else
1410         if (interval != 0)
1411         {
1412                 elog(LOG, "setsockopt(TCP_KEEPINTVL) not supported");
1413                 return STATUS_ERROR;
1414         }
1415 #endif
1416
1417         return STATUS_OK;
1418 }
1419
1420 int
1421 pq_getkeepalivescount(Port *port)
1422 {
1423 #ifdef TCP_KEEPCNT
1424         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1425                 return 0;
1426
1427         if (port->keepalives_count != 0)
1428                 return port->keepalives_count;
1429
1430         if (port->default_keepalives_count == 0)
1431         {
1432                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_count);
1433
1434                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPCNT,
1435                                            (char *) &port->default_keepalives_count,
1436                                            &size) < 0)
1437                 {
1438                         elog(LOG, "getsockopt(TCP_KEEPCNT) failed: %m");
1439                         port->default_keepalives_count = -1;            /* don't know */
1440                 }
1441         }
1442
1443         return port->default_keepalives_count;
1444 #else
1445         return 0;
1446 #endif
1447 }
1448
1449 int
1450 pq_setkeepalivescount(int count, Port *port)
1451 {
1452         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1453                 return STATUS_OK;
1454
1455 #ifdef TCP_KEEPCNT
1456         if (count == port->keepalives_count)
1457                 return STATUS_OK;
1458
1459         if (port->default_keepalives_count <= 0)
1460         {
1461                 if (pq_getkeepalivescount(port) < 0)
1462                 {
1463                         if (count == 0)
1464                                 return STATUS_OK;               /* default is set but unknown */
1465                         else
1466                                 return STATUS_ERROR;
1467                 }
1468         }
1469
1470         if (count == 0)
1471                 count = port->default_keepalives_count;
1472
1473         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPCNT,
1474                                    (char *) &count, sizeof(count)) < 0)
1475         {
1476                 elog(LOG, "setsockopt(TCP_KEEPCNT) failed: %m");
1477                 return STATUS_ERROR;
1478         }
1479
1480         port->keepalives_count = count;
1481 #else
1482         if (count != 0)
1483         {
1484                 elog(LOG, "setsockopt(TCP_KEEPCNT) not supported");
1485                 return STATUS_ERROR;
1486         }
1487 #endif
1488
1489         return STATUS_OK;
1490 }