granicus.if.org Git - postgresql/blob - 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) 1996-2012, PostgreSQL Global Development Group
  31  * Portions Copyright (c) 1994, Regents of the University of California
  32  *
  33  *      src/backend/libpq/pqcomm.c
  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_flush_if_writable - flush pending output if writable without blocking
  59  *              pq_getbyte_if_available - get a byte if available without blocking
  60  *
  61  * message-level I/O (and old-style-COPY-OUT cruft):
  62  *              pq_putmessage   - send a normal message (suppressed in COPY OUT mode)
  63  *              pq_putmessage_noblock - buffer a normal message (suppressed in COPY OUT)
  64  *              pq_startcopyout - inform libpq that a COPY OUT transfer is beginning
  65  *              pq_endcopyout   - end a COPY OUT transfer
  66  *
  67  *------------------------
  68  */
  69 #include "postgres.h"
  70
  71 #include <signal.h>
  72 #include <fcntl.h>
  73 #include <grp.h>
  74 #include <unistd.h>
  75 #include <sys/file.h>
  76 #include <sys/socket.h>
  77 #include <sys/stat.h>
  78 #include <sys/time.h>
  79 #include <netdb.h>
  80 #include <netinet/in.h>
  81 #ifdef HAVE_NETINET_TCP_H
  82 #include <netinet/tcp.h>
  83 #endif
  84 #include <arpa/inet.h>
  85 #ifdef HAVE_UTIME_H
  86 #include <utime.h>
  87 #endif
  88 #ifdef WIN32_ONLY_COMPILER              /* mstcpip.h is missing on mingw */
  89 #include <mstcpip.h>
  90 #endif
  91
  92 #include "libpq/ip.h"
  93 #include "libpq/libpq.h"
  94 #include "miscadmin.h"
  95 #include "storage/ipc.h"
  96 #include "utils/guc.h"
  97 #include "utils/memutils.h"
  98
  99 /*
 100  * Configuration options
 101  */
 102 int                     Unix_socket_permissions;
 103 char       *Unix_socket_group;
 104
 105
 106 /* Where the Unix socket file is */
 107 static char sock_path[MAXPGPATH];
 108
 109
 110 /*
 111  * Buffers for low-level I/O.
 112  *
 113  * The receive buffer is fixed size. Send buffer is usually 8k, but can be
 114  * enlarged by pq_putmessage_noblock() if the message doesn't fit otherwise.
 115  */
 116
 117 #define PQ_SEND_BUFFER_SIZE 8192
 118 #define PQ_RECV_BUFFER_SIZE 8192
 119
 120 static char *PqSendBuffer;
 121 static int      PqSendBufferSize;       /* Size send buffer */
 122 static int      PqSendPointer;          /* Next index to store a byte in PqSendBuffer */
 123 static int      PqSendStart;            /* Next index to send a byte in PqSendBuffer */
 124
 125 static char PqRecvBuffer[PQ_RECV_BUFFER_SIZE];
 126 static int      PqRecvPointer;          /* Next index to read a byte from PqRecvBuffer */
 127 static int      PqRecvLength;           /* End of data available in PqRecvBuffer */
 128
 129 /*
 130  * Message status
 131  */
 132 static bool PqCommBusy;
 133 static bool DoingCopyOut;
 134
 135
 136 /* Internal functions */
 137 static void pq_close(int code, Datum arg);
 138 static int      internal_putbytes(const char *s, size_t len);
 139 static int      internal_flush(void);
 140 static void pq_set_nonblocking(bool nonblocking);
 141
 142 #ifdef HAVE_UNIX_SOCKETS
 143 static int      Lock_AF_UNIX(unsigned short portNumber, char *unixSocketName);
 144 static int      Setup_AF_UNIX(void);
 145 #endif   /* HAVE_UNIX_SOCKETS */
 146
 147
 148 /* --------------------------------
 149  *              pq_init - initialize libpq at backend startup
 150  * --------------------------------
 151  */
 152 void
 153 pq_init(void)
 154 {
 155         PqSendBufferSize = PQ_SEND_BUFFER_SIZE;
 156         PqSendBuffer = MemoryContextAlloc(TopMemoryContext, PqSendBufferSize);
 157         PqSendPointer = PqSendStart = PqRecvPointer = PqRecvLength = 0;
 158         PqCommBusy = false;
 159         DoingCopyOut = false;
 160         on_proc_exit(pq_close, 0);
 161 }
 162
 163 /* --------------------------------
 164  *              pq_comm_reset - reset libpq during error recovery
 165  *
 166  * This is called from error recovery at the outer idle loop.  It's
 167  * just to get us out of trouble if we somehow manage to elog() from
 168  * inside a pqcomm.c routine (which ideally will never happen, but...)
 169  * --------------------------------
 170  */
 171 void
 172 pq_comm_reset(void)
 173 {
 174         /* Do not throw away pending data, but do reset the busy flag */
 175         PqCommBusy = false;
 176         /* We can abort any old-style COPY OUT, too */
 177         pq_endcopyout(true);
 178 }
 179
 180 /* --------------------------------
 181  *              pq_close - shutdown libpq at backend exit
 182  *
 183  * Note: in a standalone backend MyProcPort will be null,
 184  * don't crash during exit...
 185  * --------------------------------
 186  */
 187 static void
 188 pq_close(int code, Datum arg)
 189 {
 190         if (MyProcPort != NULL)
 191         {
 192 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
 193 #ifdef ENABLE_GSS
 194                 OM_uint32       min_s;
 195
 196                 /* Shutdown GSSAPI layer */
 197                 if (MyProcPort->gss->ctx != GSS_C_NO_CONTEXT)
 198                         gss_delete_sec_context(&min_s, &MyProcPort->gss->ctx, NULL);
 199
 200                 if (MyProcPort->gss->cred != GSS_C_NO_CREDENTIAL)
 201                         gss_release_cred(&min_s, &MyProcPort->gss->cred);
 202 #endif   /* ENABLE_GSS */
 203                 /* GSS and SSPI share the port->gss struct */
 204
 205                 free(MyProcPort->gss);
 206 #endif   /* ENABLE_GSS || ENABLE_SSPI */
 207
 208                 /* Cleanly shut down SSL layer */
 209                 secure_close(MyProcPort);
 210
 211                 /*
 212                  * Formerly we did an explicit close() here, but it seems better to
 213                  * leave the socket open until the process dies.  This allows clients
 214                  * to perform a "synchronous close" if they care --- wait till the
 215                  * transport layer reports connection closure, and you can be sure the
 216                  * backend has exited.
 217                  *
 218                  * We do set sock to PGINVALID_SOCKET to prevent any further I/O,
 219                  * though.
 220                  */
 221                 MyProcPort->sock = PGINVALID_SOCKET;
 222         }
 223 }
 224
 225
 226
 227 /*
 228  * Streams -- wrapper around Unix socket system calls
 229  *
 230  *
 231  *              Stream functions are used for vanilla TCP connection protocol.
 232  */
 233
 234
 235 /* StreamDoUnlink()
 236  * Shutdown routine for backend connection
 237  * If a Unix socket is used for communication, explicitly close it.
 238  */
 239 #ifdef HAVE_UNIX_SOCKETS
 240 static void
 241 StreamDoUnlink(int code, Datum arg)
 242 {
 243         Assert(sock_path[0]);
 244         unlink(sock_path);
 245 }
 246 #endif   /* HAVE_UNIX_SOCKETS */
 247
 248 /*
 249  * StreamServerPort -- open a "listening" port to accept connections.
 250  *
 251  * Successfully opened sockets are added to the ListenSocket[] array,
 252  * at the first position that isn't PGINVALID_SOCKET.
 253  *
 254  * RETURNS: STATUS_OK or STATUS_ERROR
 255  */
 256
 257 int
 258 StreamServerPort(int family, char *hostName, unsigned short portNumber,
 259                                  char *unixSocketName,
 260                                  pgsocket ListenSocket[], int MaxListen)
 261 {
 262         pgsocket        fd;
 263         int                     err;
 264         int                     maxconn;
 265         int                     ret;
 266         char            portNumberStr[32];
 267         const char *familyDesc;
 268         char            familyDescBuf[64];
 269         char       *service;
 270         struct addrinfo *addrs = NULL,
 271                            *addr;
 272         struct addrinfo hint;
 273         int                     listen_index = 0;
 274         int                     added = 0;
 275
 276 #if !defined(WIN32) || defined(IPV6_V6ONLY)
 277         int                     one = 1;
 278 #endif
 279
 280         /* Initialize hint structure */
 281         MemSet(&hint, 0, sizeof(hint));
 282         hint.ai_family = family;
 283         hint.ai_flags = AI_PASSIVE;
 284         hint.ai_socktype = SOCK_STREAM;
 285
 286 #ifdef HAVE_UNIX_SOCKETS
 287         if (family == AF_UNIX)
 288         {
 289                 /* Lock_AF_UNIX will also fill in sock_path. */
 290                 if (Lock_AF_UNIX(portNumber, unixSocketName) != STATUS_OK)
 291                         return STATUS_ERROR;
 292                 service = sock_path;
 293         }
 294         else
 295 #endif   /* HAVE_UNIX_SOCKETS */
 296         {
 297                 snprintf(portNumberStr, sizeof(portNumberStr), "%d", portNumber);
 298                 service = portNumberStr;
 299         }
 300
 301         ret = pg_getaddrinfo_all(hostName, service, &hint, &addrs);
 302         if (ret || !addrs)
 303         {
 304                 if (hostName)
 305                         ereport(LOG,
 306                                         (errmsg("could not translate host name \"%s\", service \"%s\" to address: %s",
 307                                                         hostName, service, gai_strerror(ret))));
 308                 else
 309                         ereport(LOG,
 310                                  (errmsg("could not translate service \"%s\" to address: %s",
 311                                                  service, gai_strerror(ret))));
 312                 if (addrs)
 313                         pg_freeaddrinfo_all(hint.ai_family, addrs);
 314                 return STATUS_ERROR;
 315         }
 316
 317         for (addr = addrs; addr; addr = addr->ai_next)
 318         {
 319                 if (!IS_AF_UNIX(family) && IS_AF_UNIX(addr->ai_family))
 320                 {
 321                         /*
 322                          * Only set up a unix domain socket when they really asked for it.
 323                          * The service/port is different in that case.
 324                          */
 325                         continue;
 326                 }
 327
 328                 /* See if there is still room to add 1 more socket. */
 329                 for (; listen_index < MaxListen; listen_index++)
 330                 {
 331                         if (ListenSocket[listen_index] == PGINVALID_SOCKET)
 332                                 break;
 333                 }
 334                 if (listen_index >= MaxListen)
 335                 {
 336                         ereport(LOG,
 337                                         (errmsg("could not bind to all requested addresses: MAXLISTEN (%d) exceeded",
 338                                                         MaxListen)));
 339                         break;
 340                 }
 341
 342                 /* set up family name for possible error messages */
 343                 switch (addr->ai_family)
 344                 {
 345                         case AF_INET:
 346                                 familyDesc = _("IPv4");
 347                                 break;
 348 #ifdef HAVE_IPV6
 349                         case AF_INET6:
 350                                 familyDesc = _("IPv6");
 351                                 break;
 352 #endif
 353 #ifdef HAVE_UNIX_SOCKETS
 354                         case AF_UNIX:
 355                                 familyDesc = _("Unix");
 356                                 break;
 357 #endif
 358                         default:
 359                                 snprintf(familyDescBuf, sizeof(familyDescBuf),
 360                                                  _("unrecognized address family %d"),
 361                                                  addr->ai_family);
 362                                 familyDesc = familyDescBuf;
 363                                 break;
 364                 }
 365
 366                 if ((fd = socket(addr->ai_family, SOCK_STREAM, 0)) < 0)
 367                 {
 368                         ereport(LOG,
 369                                         (errcode_for_socket_access(),
 370                         /* translator: %s is IPv4, IPv6, or Unix */
 371                                          errmsg("could not create %s socket: %m",
 372                                                         familyDesc)));
 373                         continue;
 374                 }
 375
 376 #ifndef WIN32
 377
 378                 /*
 379                  * Without the SO_REUSEADDR flag, a new postmaster can't be started
 380                  * right away after a stop or crash, giving "address already in use"
 381                  * error on TCP ports.
 382                  *
 383                  * On win32, however, this behavior only happens if the
 384                  * SO_EXLUSIVEADDRUSE is set. With SO_REUSEADDR, win32 allows multiple
 385                  * servers to listen on the same address, resulting in unpredictable
 386                  * behavior. With no flags at all, win32 behaves as Unix with
 387                  * SO_REUSEADDR.
 388                  */
 389                 if (!IS_AF_UNIX(addr->ai_family))
 390                 {
 391                         if ((setsockopt(fd, SOL_SOCKET, SO_REUSEADDR,
 392                                                         (char *) &one, sizeof(one))) == -1)
 393                         {
 394                                 ereport(LOG,
 395                                                 (errcode_for_socket_access(),
 396                                                  errmsg("setsockopt(SO_REUSEADDR) failed: %m")));
 397                                 closesocket(fd);
 398                                 continue;
 399                         }
 400                 }
 401 #endif
 402
 403 #ifdef IPV6_V6ONLY
 404                 if (addr->ai_family == AF_INET6)
 405                 {
 406                         if (setsockopt(fd, IPPROTO_IPV6, IPV6_V6ONLY,
 407                                                    (char *) &one, sizeof(one)) == -1)
 408                         {
 409                                 ereport(LOG,
 410                                                 (errcode_for_socket_access(),
 411                                                  errmsg("setsockopt(IPV6_V6ONLY) failed: %m")));
 412                                 closesocket(fd);
 413                                 continue;
 414                         }
 415                 }
 416 #endif
 417
 418                 /*
 419                  * Note: This might fail on some OS's, like Linux older than
 420                  * 2.4.21-pre3, that don't have the IPV6_V6ONLY socket option, and map
 421                  * ipv4 addresses to ipv6.      It will show ::ffff:ipv4 for all ipv4
 422                  * connections.
 423                  */
 424                 err = bind(fd, addr->ai_addr, addr->ai_addrlen);
 425                 if (err < 0)
 426                 {
 427                         ereport(LOG,
 428                                         (errcode_for_socket_access(),
 429                         /* translator: %s is IPv4, IPv6, or Unix */
 430                                          errmsg("could not bind %s socket: %m",
 431                                                         familyDesc),
 432                                          (IS_AF_UNIX(addr->ai_family)) ?
 433                                   errhint("Is another postmaster already running on port %d?"
 434                                                   " If not, remove socket file \"%s\" and retry.",
 435                                                   (int) portNumber, sock_path) :
 436                                   errhint("Is another postmaster already running on port %d?"
 437                                                   " If not, wait a few seconds and retry.",
 438                                                   (int) portNumber)));
 439                         closesocket(fd);
 440                         continue;
 441                 }
 442
 443 #ifdef HAVE_UNIX_SOCKETS
 444                 if (addr->ai_family == AF_UNIX)
 445                 {
 446                         if (Setup_AF_UNIX() != STATUS_OK)
 447                         {
 448                                 closesocket(fd);
 449                                 break;
 450                         }
 451                 }
 452 #endif
 453
 454                 /*
 455                  * Select appropriate accept-queue length limit.  PG_SOMAXCONN is only
 456                  * intended to provide a clamp on the request on platforms where an
 457                  * overly large request provokes a kernel error (are there any?).
 458                  */
 459                 maxconn = MaxBackends * 2;
 460                 if (maxconn > PG_SOMAXCONN)
 461                         maxconn = PG_SOMAXCONN;
 462
 463                 err = listen(fd, maxconn);
 464                 if (err < 0)
 465                 {
 466                         ereport(LOG,
 467                                         (errcode_for_socket_access(),
 468                         /* translator: %s is IPv4, IPv6, or Unix */
 469                                          errmsg("could not listen on %s socket: %m",
 470                                                         familyDesc)));
 471                         closesocket(fd);
 472                         continue;
 473                 }
 474                 ListenSocket[listen_index] = fd;
 475                 added++;
 476         }
 477
 478         pg_freeaddrinfo_all(hint.ai_family, addrs);
 479
 480         if (!added)
 481                 return STATUS_ERROR;
 482
 483         return STATUS_OK;
 484 }
 485
 486
 487 #ifdef HAVE_UNIX_SOCKETS
 488
 489 /*
 490  * Lock_AF_UNIX -- configure unix socket file path
 491  */
 492 static int
 493 Lock_AF_UNIX(unsigned short portNumber, char *unixSocketName)
 494 {
 495         UNIXSOCK_PATH(sock_path, portNumber, unixSocketName);
 496
 497         /*
 498          * Grab an interlock file associated with the socket file.
 499          *
 500          * Note: there are two reasons for using a socket lock file, rather than
 501          * trying to interlock directly on the socket itself.  First, it's a lot
 502          * more portable, and second, it lets us remove any pre-existing socket
 503          * file without race conditions.
 504          */
 505         CreateSocketLockFile(sock_path, true);
 506
 507         /*
 508          * Once we have the interlock, we can safely delete any pre-existing
 509          * socket file to avoid failure at bind() time.
 510          */
 511         unlink(sock_path);
 512
 513         return STATUS_OK;
 514 }
 515
 516
 517 /*
 518  * Setup_AF_UNIX -- configure unix socket permissions
 519  */
 520 static int
 521 Setup_AF_UNIX(void)
 522 {
 523         /* Arrange to unlink the socket file at exit */
 524         on_proc_exit(StreamDoUnlink, 0);
 525
 526         /*
 527          * Fix socket ownership/permission if requested.  Note we must do this
 528          * before we listen() to avoid a window where unwanted connections could
 529          * get accepted.
 530          */
 531         Assert(Unix_socket_group);
 532         if (Unix_socket_group[0] != '\0')
 533         {
 534 #ifdef WIN32
 535                 elog(WARNING, "configuration item unix_socket_group is not supported on this platform");
 536 #else
 537                 char       *endptr;
 538                 unsigned long val;
 539                 gid_t           gid;
 540
 541                 val = strtoul(Unix_socket_group, &endptr, 10);
 542                 if (*endptr == '\0')
 543                 {                                               /* numeric group id */
 544                         gid = val;
 545                 }
 546                 else
 547                 {                                               /* convert group name to id */
 548                         struct group *gr;
 549
 550                         gr = getgrnam(Unix_socket_group);
 551                         if (!gr)
 552                         {
 553                                 ereport(LOG,
 554                                                 (errmsg("group \"%s\" does not exist",
 555                                                                 Unix_socket_group)));
 556                                 return STATUS_ERROR;
 557                         }
 558                         gid = gr->gr_gid;
 559                 }
 560                 if (chown(sock_path, -1, gid) == -1)
 561                 {
 562                         ereport(LOG,
 563                                         (errcode_for_file_access(),
 564                                          errmsg("could not set group of file \"%s\": %m",
 565                                                         sock_path)));
 566                         return STATUS_ERROR;
 567                 }
 568 #endif
 569         }
 570
 571         if (chmod(sock_path, Unix_socket_permissions) == -1)
 572         {
 573                 ereport(LOG,
 574                                 (errcode_for_file_access(),
 575                                  errmsg("could not set permissions of file \"%s\": %m",
 576                                                 sock_path)));
 577                 return STATUS_ERROR;
 578         }
 579         return STATUS_OK;
 580 }
 581 #endif   /* HAVE_UNIX_SOCKETS */
 582
 583
 584 /*
 585  * StreamConnection -- create a new connection with client using
 586  *              server port.  Set port->sock to the FD of the new connection.
 587  *
 588  * ASSUME: that this doesn't need to be non-blocking because
 589  *              the Postmaster uses select() to tell when the server master
 590  *              socket is ready for accept().
 591  *
 592  * RETURNS: STATUS_OK or STATUS_ERROR
 593  */
 594 int
 595 StreamConnection(pgsocket server_fd, Port *port)
 596 {
 597         /* accept connection and fill in the client (remote) address */
 598         port->raddr.salen = sizeof(port->raddr.addr);
 599         if ((port->sock = accept(server_fd,
 600                                                          (struct sockaddr *) & port->raddr.addr,
 601                                                          &port->raddr.salen)) < 0)
 602         {
 603                 ereport(LOG,
 604                                 (errcode_for_socket_access(),
 605                                  errmsg("could not accept new connection: %m")));
 606
 607                 /*
 608                  * If accept() fails then postmaster.c will still see the server
 609                  * socket as read-ready, and will immediately try again.  To avoid
 610                  * uselessly sucking lots of CPU, delay a bit before trying again.
 611                  * (The most likely reason for failure is being out of kernel file
 612                  * table slots; we can do little except hope some will get freed up.)
 613                  */
 614                 pg_usleep(100000L);             /* wait 0.1 sec */
 615                 return STATUS_ERROR;
 616         }
 617
 618 #ifdef SCO_ACCEPT_BUG
 619
 620         /*
 621          * UnixWare 7+ and OpenServer 5.0.4 are known to have this bug, but it
 622          * shouldn't hurt to catch it for all versions of those platforms.
 623          */
 624         if (port->raddr.addr.ss_family == 0)
 625                 port->raddr.addr.ss_family = AF_UNIX;
 626 #endif
 627
 628         /* fill in the server (local) address */
 629         port->laddr.salen = sizeof(port->laddr.addr);
 630         if (getsockname(port->sock,
 631                                         (struct sockaddr *) & port->laddr.addr,
 632                                         &port->laddr.salen) < 0)
 633         {
 634                 elog(LOG, "getsockname() failed: %m");
 635                 return STATUS_ERROR;
 636         }
 637
 638         /* select NODELAY and KEEPALIVE options if it's a TCP connection */
 639         if (!IS_AF_UNIX(port->laddr.addr.ss_family))
 640         {
 641                 int                     on;
 642
 643 #ifdef  TCP_NODELAY
 644                 on = 1;
 645                 if (setsockopt(port->sock, IPPROTO_TCP, TCP_NODELAY,
 646                                            (char *) &on, sizeof(on)) < 0)
 647                 {
 648                         elog(LOG, "setsockopt(TCP_NODELAY) failed: %m");
 649                         return STATUS_ERROR;
 650                 }
 651 #endif
 652                 on = 1;
 653                 if (setsockopt(port->sock, SOL_SOCKET, SO_KEEPALIVE,
 654                                            (char *) &on, sizeof(on)) < 0)
 655                 {
 656                         elog(LOG, "setsockopt(SO_KEEPALIVE) failed: %m");
 657                         return STATUS_ERROR;
 658                 }
 659
 660 #ifdef WIN32
 661
 662                 /*
 663                  * This is a Win32 socket optimization.  The ideal size is 32k.
 664                  * http://support.microsoft.com/kb/823764/EN-US/
 665                  */
 666                 on = PQ_SEND_BUFFER_SIZE * 4;
 667                 if (setsockopt(port->sock, SOL_SOCKET, SO_SNDBUF, (char *) &on,
 668                                            sizeof(on)) < 0)
 669                 {
 670                         elog(LOG, "setsockopt(SO_SNDBUF) failed: %m");
 671                         return STATUS_ERROR;
 672                 }
 673 #endif
 674
 675                 /*
 676                  * Also apply the current keepalive parameters.  If we fail to set a
 677                  * parameter, don't error out, because these aren't universally
 678                  * supported.  (Note: you might think we need to reset the GUC
 679                  * variables to 0 in such a case, but it's not necessary because the
 680                  * show hooks for these variables report the truth anyway.)
 681                  */
 682                 (void) pq_setkeepalivesidle(tcp_keepalives_idle, port);
 683                 (void) pq_setkeepalivesinterval(tcp_keepalives_interval, port);
 684                 (void) pq_setkeepalivescount(tcp_keepalives_count, port);
 685         }
 686
 687         return STATUS_OK;
 688 }
 689
 690 /*
 691  * StreamClose -- close a client/backend connection
 692  *
 693  * NOTE: this is NOT used to terminate a session; it is just used to release
 694  * the file descriptor in a process that should no longer have the socket
 695  * open.  (For example, the postmaster calls this after passing ownership
 696  * of the connection to a child process.)  It is expected that someone else
 697  * still has the socket open.  So, we only want to close the descriptor,
 698  * we do NOT want to send anything to the far end.
 699  */
 700 void
 701 StreamClose(pgsocket sock)
 702 {
 703         closesocket(sock);
 704 }
 705
 706 /*
 707  * TouchSocketFile -- mark socket file as recently accessed
 708  *
 709  * This routine should be called every so often to ensure that the socket
 710  * file has a recent mod date (ordinary operations on sockets usually won't
 711  * change the mod date).  That saves it from being removed by
 712  * overenthusiastic /tmp-directory-cleaner daemons.  (Another reason we should
 713  * never have put the socket file in /tmp...)
 714  */
 715 void
 716 TouchSocketFile(void)
 717 {
 718         /* Do nothing if we did not create a socket... */
 719         if (sock_path[0] != '\0')
 720         {
 721                 /*
 722                  * utime() is POSIX standard, utimes() is a common alternative. If we
 723                  * have neither, there's no way to affect the mod or access time of
 724                  * the socket :-(
 725                  *
 726                  * In either path, we ignore errors; there's no point in complaining.
 727                  */
 728 #ifdef HAVE_UTIME
 729                 utime(sock_path, NULL);
 730 #else                                                   /* !HAVE_UTIME */
 731 #ifdef HAVE_UTIMES
 732                 utimes(sock_path, NULL);
 733 #endif   /* HAVE_UTIMES */
 734 #endif   /* HAVE_UTIME */
 735         }
 736 }
 737
 738
 739 /* --------------------------------
 740  * Low-level I/O routines begin here.
 741  *
 742  * These routines communicate with a frontend client across a connection
 743  * already established by the preceding routines.
 744  * --------------------------------
 745  */
 746
 747 /* --------------------------------
 748  *                        pq_set_nonblocking - set socket blocking/non-blocking
 749  *
 750  * Sets the socket non-blocking if nonblocking is TRUE, or sets it
 751  * blocking otherwise.
 752  * --------------------------------
 753  */
 754 static void
 755 pq_set_nonblocking(bool nonblocking)
 756 {
 757         if (MyProcPort->noblock == nonblocking)
 758                 return;
 759
 760 #ifdef WIN32
 761         pgwin32_noblock = nonblocking ? 1 : 0;
 762 #else
 763
 764         /*
 765          * Use COMMERROR on failure, because ERROR would try to send the error to
 766          * the client, which might require changing the mode again, leading to
 767          * infinite recursion.
 768          */
 769         if (nonblocking)
 770         {
 771                 if (!pg_set_noblock(MyProcPort->sock))
 772                         ereport(COMMERROR,
 773                                   (errmsg("could not set socket to non-blocking mode: %m")));
 774         }
 775         else
 776         {
 777                 if (!pg_set_block(MyProcPort->sock))
 778                         ereport(COMMERROR,
 779                                         (errmsg("could not set socket to blocking mode: %m")));
 780         }
 781 #endif
 782         MyProcPort->noblock = nonblocking;
 783 }
 784
 785 /* --------------------------------
 786  *              pq_recvbuf - load some bytes into the input buffer
 787  *
 788  *              returns 0 if OK, EOF if trouble
 789  * --------------------------------
 790  */
 791 static int
 792 pq_recvbuf(void)
 793 {
 794         if (PqRecvPointer > 0)
 795         {
 796                 if (PqRecvLength > PqRecvPointer)
 797                 {
 798                         /* still some unread data, left-justify it in the buffer */
 799                         memmove(PqRecvBuffer, PqRecvBuffer + PqRecvPointer,
 800                                         PqRecvLength - PqRecvPointer);
 801                         PqRecvLength -= PqRecvPointer;
 802                         PqRecvPointer = 0;
 803                 }
 804                 else
 805                         PqRecvLength = PqRecvPointer = 0;
 806         }
 807
 808         /* Ensure that we're in blocking mode */
 809         pq_set_nonblocking(false);
 810
 811         /* Can fill buffer from PqRecvLength and upwards */
 812         for (;;)
 813         {
 814                 int                     r;
 815
 816                 r = secure_read(MyProcPort, PqRecvBuffer + PqRecvLength,
 817                                                 PQ_RECV_BUFFER_SIZE - PqRecvLength);
 818
 819                 if (r < 0)
 820                 {
 821                         if (errno == EINTR)
 822                                 continue;               /* Ok if interrupted */
 823
 824                         /*
 825                          * Careful: an ereport() that tries to write to the client would
 826                          * cause recursion to here, leading to stack overflow and core
 827                          * dump!  This message must go *only* to the postmaster log.
 828                          */
 829                         ereport(COMMERROR,
 830                                         (errcode_for_socket_access(),
 831                                          errmsg("could not receive data from client: %m")));
 832                         return EOF;
 833                 }
 834                 if (r == 0)
 835                 {
 836                         /*
 837                          * EOF detected.  We used to write a log message here, but it's
 838                          * better to expect the ultimate caller to do that.
 839                          */
 840                         return EOF;
 841                 }
 842                 /* r contains number of bytes read, so just incr length */
 843                 PqRecvLength += r;
 844                 return 0;
 845         }
 846 }
 847
 848 /* --------------------------------
 849  *              pq_getbyte      - get a single byte from connection, or return EOF
 850  * --------------------------------
 851  */
 852 int
 853 pq_getbyte(void)
 854 {
 855         while (PqRecvPointer >= PqRecvLength)
 856         {
 857                 if (pq_recvbuf())               /* If nothing in buffer, then recv some */
 858                         return EOF;                     /* Failed to recv data */
 859         }
 860         return (unsigned char) PqRecvBuffer[PqRecvPointer++];
 861 }
 862
 863 /* --------------------------------
 864  *              pq_peekbyte             - peek at next byte from connection
 865  *
 866  *       Same as pq_getbyte() except we don't advance the pointer.
 867  * --------------------------------
 868  */
 869 int
 870 pq_peekbyte(void)
 871 {
 872         while (PqRecvPointer >= PqRecvLength)
 873         {
 874                 if (pq_recvbuf())               /* If nothing in buffer, then recv some */
 875                         return EOF;                     /* Failed to recv data */
 876         }
 877         return (unsigned char) PqRecvBuffer[PqRecvPointer];
 878 }
 879
 880 /* --------------------------------
 881  *              pq_getbyte_if_available - get a single byte from connection,
 882  *                      if available
 883  *
 884  * The received byte is stored in *c. Returns 1 if a byte was read,
 885  * 0 if no data was available, or EOF if trouble.
 886  * --------------------------------
 887  */
 888 int
 889 pq_getbyte_if_available(unsigned char *c)
 890 {
 891         int                     r;
 892
 893         if (PqRecvPointer < PqRecvLength)
 894         {
 895                 *c = PqRecvBuffer[PqRecvPointer++];
 896                 return 1;
 897         }
 898
 899         /* Put the socket into non-blocking mode */
 900         pq_set_nonblocking(true);
 901
 902         r = secure_read(MyProcPort, c, 1);
 903         if (r < 0)
 904         {
 905                 /*
 906                  * Ok if no data available without blocking or interrupted (though
 907                  * EINTR really shouldn't happen with a non-blocking socket). Report
 908                  * other errors.
 909                  */
 910                 if (errno == EAGAIN || errno == EWOULDBLOCK || errno == EINTR)
 911                         r = 0;
 912                 else
 913                 {
 914                         /*
 915                          * Careful: an ereport() that tries to write to the client would
 916                          * cause recursion to here, leading to stack overflow and core
 917                          * dump!  This message must go *only* to the postmaster log.
 918                          */
 919                         ereport(COMMERROR,
 920                                         (errcode_for_socket_access(),
 921                                          errmsg("could not receive data from client: %m")));
 922                         r = EOF;
 923                 }
 924         }
 925         else if (r == 0)
 926         {
 927                 /* EOF detected */
 928                 r = EOF;
 929         }
 930
 931         return r;
 932 }
 933
 934 /* --------------------------------
 935  *              pq_getbytes             - get a known number of bytes from connection
 936  *
 937  *              returns 0 if OK, EOF if trouble
 938  * --------------------------------
 939  */
 940 int
 941 pq_getbytes(char *s, size_t len)
 942 {
 943         size_t          amount;
 944
 945         while (len > 0)
 946         {
 947                 while (PqRecvPointer >= PqRecvLength)
 948                 {
 949                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
 950                                 return EOF;             /* Failed to recv data */
 951                 }
 952                 amount = PqRecvLength - PqRecvPointer;
 953                 if (amount > len)
 954                         amount = len;
 955                 memcpy(s, PqRecvBuffer + PqRecvPointer, amount);
 956                 PqRecvPointer += amount;
 957                 s += amount;
 958                 len -= amount;
 959         }
 960         return 0;
 961 }
 962
 963 /* --------------------------------
 964  *              pq_discardbytes         - throw away a known number of bytes
 965  *
 966  *              same as pq_getbytes except we do not copy the data to anyplace.
 967  *              this is used for resynchronizing after read errors.
 968  *
 969  *              returns 0 if OK, EOF if trouble
 970  * --------------------------------
 971  */
 972 static int
 973 pq_discardbytes(size_t len)
 974 {
 975         size_t          amount;
 976
 977         while (len > 0)
 978         {
 979                 while (PqRecvPointer >= PqRecvLength)
 980                 {
 981                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
 982                                 return EOF;             /* Failed to recv data */
 983                 }
 984                 amount = PqRecvLength - PqRecvPointer;
 985                 if (amount > len)
 986                         amount = len;
 987                 PqRecvPointer += amount;
 988                 len -= amount;
 989         }
 990         return 0;
 991 }
 992
 993 /* --------------------------------
 994  *              pq_getstring    - get a null terminated string from connection
 995  *
 996  *              The return value is placed in an expansible StringInfo, which has
 997  *              already been initialized by the caller.
 998  *
 999  *              This is used only for dealing with old-protocol clients.  The idea
1000  *              is to produce a StringInfo that looks the same as we would get from
1001  *              pq_getmessage() with a newer client; we will then process it with
1002  *              pq_getmsgstring.  Therefore, no character set conversion is done here,
1003  *              even though this is presumably useful only for text.
1004  *
1005  *              returns 0 if OK, EOF if trouble
1006  * --------------------------------
1007  */
1008 int
1009 pq_getstring(StringInfo s)
1010 {
1011         int                     i;
1012
1013         resetStringInfo(s);
1014
1015         /* Read until we get the terminating '\0' */
1016         for (;;)
1017         {
1018                 while (PqRecvPointer >= PqRecvLength)
1019                 {
1020                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
1021                                 return EOF;             /* Failed to recv data */
1022                 }
1023
1024                 for (i = PqRecvPointer; i < PqRecvLength; i++)
1025                 {
1026                         if (PqRecvBuffer[i] == '\0')
1027                         {
1028                                 /* include the '\0' in the copy */
1029                                 appendBinaryStringInfo(s, PqRecvBuffer + PqRecvPointer,
1030                                                                            i - PqRecvPointer + 1);
1031                                 PqRecvPointer = i + 1;  /* advance past \0 */
1032                                 return 0;
1033                         }
1034                 }
1035
1036                 /* If we're here we haven't got the \0 in the buffer yet. */
1037                 appendBinaryStringInfo(s, PqRecvBuffer + PqRecvPointer,
1038                                                            PqRecvLength - PqRecvPointer);
1039                 PqRecvPointer = PqRecvLength;
1040         }
1041 }
1042
1043
1044 /* --------------------------------
1045  *              pq_getmessage   - get a message with length word from connection
1046  *
1047  *              The return value is placed in an expansible StringInfo, which has
1048  *              already been initialized by the caller.
1049  *              Only the message body is placed in the StringInfo; the length word
1050  *              is removed.  Also, s->cursor is initialized to zero for convenience
1051  *              in scanning the message contents.
1052  *
1053  *              If maxlen is not zero, it is an upper limit on the length of the
1054  *              message we are willing to accept.  We abort the connection (by
1055  *              returning EOF) if client tries to send more than that.
1056  *
1057  *              returns 0 if OK, EOF if trouble
1058  * --------------------------------
1059  */
1060 int
1061 pq_getmessage(StringInfo s, int maxlen)
1062 {
1063         int32           len;
1064
1065         resetStringInfo(s);
1066
1067         /* Read message length word */
1068         if (pq_getbytes((char *) &len, 4) == EOF)
1069         {
1070                 ereport(COMMERROR,
1071                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1072                                  errmsg("unexpected EOF within message length word")));
1073                 return EOF;
1074         }
1075
1076         len = ntohl(len);
1077
1078         if (len < 4 ||
1079                 (maxlen > 0 && len > maxlen))
1080         {
1081                 ereport(COMMERROR,
1082                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1083                                  errmsg("invalid message length")));
1084                 return EOF;
1085         }
1086
1087         len -= 4;                                       /* discount length itself */
1088
1089         if (len > 0)
1090         {
1091                 /*
1092                  * Allocate space for message.  If we run out of room (ridiculously
1093                  * large message), we will elog(ERROR), but we want to discard the
1094                  * message body so as not to lose communication sync.
1095                  */
1096                 PG_TRY();
1097                 {
1098                         enlargeStringInfo(s, len);
1099                 }
1100                 PG_CATCH();
1101                 {
1102                         if (pq_discardbytes(len) == EOF)
1103                                 ereport(COMMERROR,
1104                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1105                                                  errmsg("incomplete message from client")));
1106                         PG_RE_THROW();
1107                 }
1108                 PG_END_TRY();
1109
1110                 /* And grab the message */
1111                 if (pq_getbytes(s->data, len) == EOF)
1112                 {
1113                         ereport(COMMERROR,
1114                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
1115                                          errmsg("incomplete message from client")));
1116                         return EOF;
1117                 }
1118                 s->len = len;
1119                 /* Place a trailing null per StringInfo convention */
1120                 s->data[len] = '\0';
1121         }
1122
1123         return 0;
1124 }
1125
1126
1127 /* --------------------------------
1128  *              pq_putbytes             - send bytes to connection (not flushed until pq_flush)
1129  *
1130  *              returns 0 if OK, EOF if trouble
1131  * --------------------------------
1132  */
1133 int
1134 pq_putbytes(const char *s, size_t len)
1135 {
1136         int                     res;
1137
1138         /* Should only be called by old-style COPY OUT */
1139         Assert(DoingCopyOut);
1140         /* No-op if reentrant call */
1141         if (PqCommBusy)
1142                 return 0;
1143         PqCommBusy = true;
1144         res = internal_putbytes(s, len);
1145         PqCommBusy = false;
1146         return res;
1147 }
1148
1149 static int
1150 internal_putbytes(const char *s, size_t len)
1151 {
1152         size_t          amount;
1153
1154         while (len > 0)
1155         {
1156                 /* If buffer is full, then flush it out */
1157                 if (PqSendPointer >= PqSendBufferSize)
1158                 {
1159                         pq_set_nonblocking(false);
1160                         if (internal_flush())
1161                                 return EOF;
1162                 }
1163                 amount = PqSendBufferSize - PqSendPointer;
1164                 if (amount > len)
1165                         amount = len;
1166                 memcpy(PqSendBuffer + PqSendPointer, s, amount);
1167                 PqSendPointer += amount;
1168                 s += amount;
1169                 len -= amount;
1170         }
1171         return 0;
1172 }
1173
1174 /* --------------------------------
1175  *              pq_flush                - flush pending output
1176  *
1177  *              returns 0 if OK, EOF if trouble
1178  * --------------------------------
1179  */
1180 int
1181 pq_flush(void)
1182 {
1183         int                     res;
1184
1185         /* No-op if reentrant call */
1186         if (PqCommBusy)
1187                 return 0;
1188         PqCommBusy = true;
1189         pq_set_nonblocking(false);
1190         res = internal_flush();
1191         PqCommBusy = false;
1192         return res;
1193 }
1194
1195 /* --------------------------------
1196  *              internal_flush - flush pending output
1197  *
1198  * Returns 0 if OK (meaning everything was sent, or operation would block
1199  * and the socket is in non-blocking mode), or EOF if trouble.
1200  * --------------------------------
1201  */
1202 static int
1203 internal_flush(void)
1204 {
1205         static int      last_reported_send_errno = 0;
1206
1207         char       *bufptr = PqSendBuffer + PqSendStart;
1208         char       *bufend = PqSendBuffer + PqSendPointer;
1209
1210         while (bufptr < bufend)
1211         {
1212                 int                     r;
1213
1214                 r = secure_write(MyProcPort, bufptr, bufend - bufptr);
1215
1216                 if (r <= 0)
1217                 {
1218                         if (errno == EINTR)
1219                                 continue;               /* Ok if we were interrupted */
1220
1221                         /*
1222                          * Ok if no data writable without blocking, and the socket is in
1223                          * non-blocking mode.
1224                          */
1225                         if (errno == EAGAIN ||
1226                                 errno == EWOULDBLOCK)
1227                         {
1228                                 return 0;
1229                         }
1230
1231                         /*
1232                          * Careful: an ereport() that tries to write to the client would
1233                          * cause recursion to here, leading to stack overflow and core
1234                          * dump!  This message must go *only* to the postmaster log.
1235                          *
1236                          * If a client disconnects while we're in the midst of output, we
1237                          * might write quite a bit of data before we get to a safe query
1238                          * abort point.  So, suppress duplicate log messages.
1239                          */
1240                         if (errno != last_reported_send_errno)
1241                         {
1242                                 last_reported_send_errno = errno;
1243                                 ereport(COMMERROR,
1244                                                 (errcode_for_socket_access(),
1245                                                  errmsg("could not send data to client: %m")));
1246                         }
1247
1248                         /*
1249                          * We drop the buffered data anyway so that processing can
1250                          * continue, even though we'll probably quit soon. We also set a
1251                          * flag that'll cause the next CHECK_FOR_INTERRUPTS to terminate
1252                          * the connection.
1253                          */
1254                         PqSendStart = PqSendPointer = 0;
1255                         ClientConnectionLost = 1;
1256                         InterruptPending = 1;
1257                         return EOF;
1258                 }
1259
1260                 last_reported_send_errno = 0;   /* reset after any successful send */
1261                 bufptr += r;
1262                 PqSendStart += r;
1263         }
1264
1265         PqSendStart = PqSendPointer = 0;
1266         return 0;
1267 }
1268
1269 /* --------------------------------
1270  *              pq_flush_if_writable - flush pending output if writable without blocking
1271  *
1272  * Returns 0 if OK, or EOF if trouble.
1273  * --------------------------------
1274  */
1275 int
1276 pq_flush_if_writable(void)
1277 {
1278         int                     res;
1279
1280         /* Quick exit if nothing to do */
1281         if (PqSendPointer == PqSendStart)
1282                 return 0;
1283
1284         /* No-op if reentrant call */
1285         if (PqCommBusy)
1286                 return 0;
1287
1288         /* Temporarily put the socket into non-blocking mode */
1289         pq_set_nonblocking(true);
1290
1291         PqCommBusy = true;
1292         res = internal_flush();
1293         PqCommBusy = false;
1294         return res;
1295 }
1296
1297 /* --------------------------------
1298  *              pq_is_send_pending      - is there any pending data in the output buffer?
1299  * --------------------------------
1300  */
1301 bool
1302 pq_is_send_pending(void)
1303 {
1304         return (PqSendStart < PqSendPointer);
1305 }
1306
1307 /* --------------------------------
1308  * Message-level I/O routines begin here.
1309  *
1310  * These routines understand about the old-style COPY OUT protocol.
1311  * --------------------------------
1312  */
1313
1314
1315 /* --------------------------------
1316  *              pq_putmessage   - send a normal message (suppressed in COPY OUT mode)
1317  *
1318  *              If msgtype is not '\0', it is a message type code to place before
1319  *              the message body.  If msgtype is '\0', then the message has no type
1320  *              code (this is only valid in pre-3.0 protocols).
1321  *
1322  *              len is the length of the message body data at *s.  In protocol 3.0
1323  *              and later, a message length word (equal to len+4 because it counts
1324  *              itself too) is inserted by this routine.
1325  *
1326  *              All normal messages are suppressed while old-style COPY OUT is in
1327  *              progress.  (In practice only a few notice messages might get emitted
1328  *              then; dropping them is annoying, but at least they will still appear
1329  *              in the postmaster log.)
1330  *
1331  *              We also suppress messages generated while pqcomm.c is busy.  This
1332  *              avoids any possibility of messages being inserted within other
1333  *              messages.  The only known trouble case arises if SIGQUIT occurs
1334  *              during a pqcomm.c routine --- quickdie() will try to send a warning
1335  *              message, and the most reasonable approach seems to be to drop it.
1336  *
1337  *              returns 0 if OK, EOF if trouble
1338  * --------------------------------
1339  */
1340 int
1341 pq_putmessage(char msgtype, const char *s, size_t len)
1342 {
1343         if (DoingCopyOut || PqCommBusy)
1344                 return 0;
1345         PqCommBusy = true;
1346         if (msgtype)
1347                 if (internal_putbytes(&msgtype, 1))
1348                         goto fail;
1349         if (PG_PROTOCOL_MAJOR(FrontendProtocol) >= 3)
1350         {
1351                 uint32          n32;
1352
1353                 n32 = htonl((uint32) (len + 4));
1354                 if (internal_putbytes((char *) &n32, 4))
1355                         goto fail;
1356         }
1357         if (internal_putbytes(s, len))
1358                 goto fail;
1359         PqCommBusy = false;
1360         return 0;
1361
1362 fail:
1363         PqCommBusy = false;
1364         return EOF;
1365 }
1366
1367 /* --------------------------------
1368  *              pq_putmessage_noblock   - like pq_putmessage, but never blocks
1369  *
1370  *              If the output buffer is too small to hold the message, the buffer
1371  *              is enlarged.
1372  */
1373 void
1374 pq_putmessage_noblock(char msgtype, const char *s, size_t len)
1375 {
1376         int res         PG_USED_FOR_ASSERTS_ONLY;
1377         int                     required;
1378
1379         /*
1380          * Ensure we have enough space in the output buffer for the message header
1381          * as well as the message itself.
1382          */
1383         required = PqSendPointer + 1 + 4 + len;
1384         if (required > PqSendBufferSize)
1385         {
1386                 PqSendBuffer = repalloc(PqSendBuffer, required);
1387                 PqSendBufferSize = required;
1388         }
1389         res = pq_putmessage(msgtype, s, len);
1390         Assert(res == 0);                       /* should not fail when the message fits in
1391                                                                  * buffer */
1392 }
1393
1394
1395 /* --------------------------------
1396  *              pq_startcopyout - inform libpq that an old-style COPY OUT transfer
1397  *                      is beginning
1398  * --------------------------------
1399  */
1400 void
1401 pq_startcopyout(void)
1402 {
1403         DoingCopyOut = true;
1404 }
1405
1406 /* --------------------------------
1407  *              pq_endcopyout   - end an old-style COPY OUT transfer
1408  *
1409  *              If errorAbort is indicated, we are aborting a COPY OUT due to an error,
1410  *              and must send a terminator line.  Since a partial data line might have
1411  *              been emitted, send a couple of newlines first (the first one could
1412  *              get absorbed by a backslash...)  Note that old-style COPY OUT does
1413  *              not allow binary transfers, so a textual terminator is always correct.
1414  * --------------------------------
1415  */
1416 void
1417 pq_endcopyout(bool errorAbort)
1418 {
1419         if (!DoingCopyOut)
1420                 return;
1421         if (errorAbort)
1422                 pq_putbytes("\n\n\\.\n", 5);
1423         /* in non-error case, copy.c will have emitted the terminator line */
1424         DoingCopyOut = false;
1425 }
1426
1427
1428 /*
1429  * Support for TCP Keepalive parameters
1430  */
1431
1432 /*
1433  * On Windows, we need to set both idle and interval at the same time.
1434  * We also cannot reset them to the default (setting to zero will
1435  * actually set them to zero, not default), therefor we fallback to
1436  * the out-of-the-box default instead.
1437  */
1438 #if defined(WIN32) && defined(SIO_KEEPALIVE_VALS)
1439 static int
1440 pq_setkeepaliveswin32(Port *port, int idle, int interval)
1441 {
1442         struct tcp_keepalive ka;
1443         DWORD           retsize;
1444
1445         if (idle <= 0)
1446                 idle = 2 * 60 * 60;             /* default = 2 hours */
1447         if (interval <= 0)
1448                 interval = 1;                   /* default = 1 second */
1449
1450         ka.onoff = 1;
1451         ka.keepalivetime = idle * 1000;
1452         ka.keepaliveinterval = interval * 1000;
1453
1454         if (WSAIoctl(port->sock,
1455                                  SIO_KEEPALIVE_VALS,
1456                                  (LPVOID) &ka,
1457                                  sizeof(ka),
1458                                  NULL,
1459                                  0,
1460                                  &retsize,
1461                                  NULL,
1462                                  NULL)
1463                 != 0)
1464         {
1465                 elog(LOG, "WSAIoctl(SIO_KEEPALIVE_VALS) failed: %ui",
1466                          WSAGetLastError());
1467                 return STATUS_ERROR;
1468         }
1469         if (port->keepalives_idle != idle)
1470                 port->keepalives_idle = idle;
1471         if (port->keepalives_interval != interval)
1472                 port->keepalives_interval = interval;
1473         return STATUS_OK;
1474 }
1475 #endif
1476
1477 int
1478 pq_getkeepalivesidle(Port *port)
1479 {
1480 #if defined(TCP_KEEPIDLE) || defined(TCP_KEEPALIVE) || defined(WIN32)
1481         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1482                 return 0;
1483
1484         if (port->keepalives_idle != 0)
1485                 return port->keepalives_idle;
1486
1487         if (port->default_keepalives_idle == 0)
1488         {
1489 #ifndef WIN32
1490                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_idle);
1491
1492 #ifdef TCP_KEEPIDLE
1493                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPIDLE,
1494                                            (char *) &port->default_keepalives_idle,
1495                                            &size) < 0)
1496                 {
1497                         elog(LOG, "getsockopt(TCP_KEEPIDLE) failed: %m");
1498                         port->default_keepalives_idle = -1; /* don't know */
1499                 }
1500 #else
1501                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPALIVE,
1502                                            (char *) &port->default_keepalives_idle,
1503                                            &size) < 0)
1504                 {
1505                         elog(LOG, "getsockopt(TCP_KEEPALIVE) failed: %m");
1506                         port->default_keepalives_idle = -1; /* don't know */
1507                 }
1508 #endif   /* TCP_KEEPIDLE */
1509 #else                                                   /* WIN32 */
1510                 /* We can't get the defaults on Windows, so return "don't know" */
1511                 port->default_keepalives_idle = -1;
1512 #endif   /* WIN32 */
1513         }
1514
1515         return port->default_keepalives_idle;
1516 #else
1517         return 0;
1518 #endif
1519 }
1520
1521 int
1522 pq_setkeepalivesidle(int idle, Port *port)
1523 {
1524         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1525                 return STATUS_OK;
1526
1527 #if defined(TCP_KEEPIDLE) || defined(TCP_KEEPALIVE) || defined(SIO_KEEPALIVE_VALS)
1528         if (idle == port->keepalives_idle)
1529                 return STATUS_OK;
1530
1531 #ifndef WIN32
1532         if (port->default_keepalives_idle <= 0)
1533         {
1534                 if (pq_getkeepalivesidle(port) < 0)
1535                 {
1536                         if (idle == 0)
1537                                 return STATUS_OK;               /* default is set but unknown */
1538                         else
1539                                 return STATUS_ERROR;
1540                 }
1541         }
1542
1543         if (idle == 0)
1544                 idle = port->default_keepalives_idle;
1545
1546 #ifdef TCP_KEEPIDLE
1547         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPIDLE,
1548                                    (char *) &idle, sizeof(idle)) < 0)
1549         {
1550                 elog(LOG, "setsockopt(TCP_KEEPIDLE) failed: %m");
1551                 return STATUS_ERROR;
1552         }
1553 #else
1554         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPALIVE,
1555                                    (char *) &idle, sizeof(idle)) < 0)
1556         {
1557                 elog(LOG, "setsockopt(TCP_KEEPALIVE) failed: %m");
1558                 return STATUS_ERROR;
1559         }
1560 #endif
1561
1562         port->keepalives_idle = idle;
1563 #else                                                   /* WIN32 */
1564         return pq_setkeepaliveswin32(port, idle, port->keepalives_interval);
1565 #endif
1566 #else                                                   /* TCP_KEEPIDLE || SIO_KEEPALIVE_VALS */
1567         if (idle != 0)
1568         {
1569                 elog(LOG, "setting the keepalive idle time is not supported");
1570                 return STATUS_ERROR;
1571         }
1572 #endif
1573         return STATUS_OK;
1574 }
1575
1576 int
1577 pq_getkeepalivesinterval(Port *port)
1578 {
1579 #if defined(TCP_KEEPINTVL) || defined(SIO_KEEPALIVE_VALS)
1580         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1581                 return 0;
1582
1583         if (port->keepalives_interval != 0)
1584                 return port->keepalives_interval;
1585
1586         if (port->default_keepalives_interval == 0)
1587         {
1588 #ifndef WIN32
1589                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_interval);
1590
1591                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPINTVL,
1592                                            (char *) &port->default_keepalives_interval,
1593                                            &size) < 0)
1594                 {
1595                         elog(LOG, "getsockopt(TCP_KEEPINTVL) failed: %m");
1596                         port->default_keepalives_interval = -1;         /* don't know */
1597                 }
1598 #else
1599                 /* We can't get the defaults on Windows, so return "don't know" */
1600                 port->default_keepalives_interval = -1;
1601 #endif   /* WIN32 */
1602         }
1603
1604         return port->default_keepalives_interval;
1605 #else
1606         return 0;
1607 #endif
1608 }
1609
1610 int
1611 pq_setkeepalivesinterval(int interval, Port *port)
1612 {
1613         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1614                 return STATUS_OK;
1615
1616 #if defined(TCP_KEEPINTVL) || defined (SIO_KEEPALIVE_VALS)
1617         if (interval == port->keepalives_interval)
1618                 return STATUS_OK;
1619
1620 #ifndef WIN32
1621         if (port->default_keepalives_interval <= 0)
1622         {
1623                 if (pq_getkeepalivesinterval(port) < 0)
1624                 {
1625                         if (interval == 0)
1626                                 return STATUS_OK;               /* default is set but unknown */
1627                         else
1628                                 return STATUS_ERROR;
1629                 }
1630         }
1631
1632         if (interval == 0)
1633                 interval = port->default_keepalives_interval;
1634
1635         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPINTVL,
1636                                    (char *) &interval, sizeof(interval)) < 0)
1637         {
1638                 elog(LOG, "setsockopt(TCP_KEEPINTVL) failed: %m");
1639                 return STATUS_ERROR;
1640         }
1641
1642         port->keepalives_interval = interval;
1643 #else                                                   /* WIN32 */
1644         return pq_setkeepaliveswin32(port, port->keepalives_idle, interval);
1645 #endif
1646 #else
1647         if (interval != 0)
1648         {
1649                 elog(LOG, "setsockopt(TCP_KEEPINTVL) not supported");
1650                 return STATUS_ERROR;
1651         }
1652 #endif
1653
1654         return STATUS_OK;
1655 }
1656
1657 int
1658 pq_getkeepalivescount(Port *port)
1659 {
1660 #ifdef TCP_KEEPCNT
1661         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1662                 return 0;
1663
1664         if (port->keepalives_count != 0)
1665                 return port->keepalives_count;
1666
1667         if (port->default_keepalives_count == 0)
1668         {
1669                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_count);
1670
1671                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPCNT,
1672                                            (char *) &port->default_keepalives_count,
1673                                            &size) < 0)
1674                 {
1675                         elog(LOG, "getsockopt(TCP_KEEPCNT) failed: %m");
1676                         port->default_keepalives_count = -1;            /* don't know */
1677                 }
1678         }
1679
1680         return port->default_keepalives_count;
1681 #else
1682         return 0;
1683 #endif
1684 }
1685
1686 int
1687 pq_setkeepalivescount(int count, Port *port)
1688 {
1689         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1690                 return STATUS_OK;
1691
1692 #ifdef TCP_KEEPCNT
1693         if (count == port->keepalives_count)
1694                 return STATUS_OK;
1695
1696         if (port->default_keepalives_count <= 0)
1697         {
1698                 if (pq_getkeepalivescount(port) < 0)
1699                 {
1700                         if (count == 0)
1701                                 return STATUS_OK;               /* default is set but unknown */
1702                         else
1703                                 return STATUS_ERROR;
1704                 }
1705         }
1706
1707         if (count == 0)
1708                 count = port->default_keepalives_count;
1709
1710         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPCNT,
1711                                    (char *) &count, sizeof(count)) < 0)
1712         {
1713                 elog(LOG, "setsockopt(TCP_KEEPCNT) failed: %m");
1714                 return STATUS_ERROR;
1715         }
1716
1717         port->keepalives_count = count;
1718 #else
1719         if (count != 0)
1720         {
1721                 elog(LOG, "setsockopt(TCP_KEEPCNT) not supported");
1722                 return STATUS_ERROR;
1723         }
1724 #endif
1725
1726         return STATUS_OK;
1727 }