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) 1996-2015, 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  *              TouchSocketFiles        - Protect socket files 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 /* Where the Unix socket files are (list of palloc'd strings) */
 106 static List *sock_paths = NIL;
 107
 108 /*
 109  * Buffers for low-level I/O.
 110  *
 111  * The receive buffer is fixed size. Send buffer is usually 8k, but can be
 112  * enlarged by pq_putmessage_noblock() if the message doesn't fit otherwise.
 113  */
 114
 115 #define PQ_SEND_BUFFER_SIZE 8192
 116 #define PQ_RECV_BUFFER_SIZE 8192
 117
 118 static char *PqSendBuffer;
 119 static int      PqSendBufferSize;       /* Size send buffer */
 120 static int      PqSendPointer;          /* Next index to store a byte in PqSendBuffer */
 121 static int      PqSendStart;            /* Next index to send a byte in PqSendBuffer */
 122
 123 static char PqRecvBuffer[PQ_RECV_BUFFER_SIZE];
 124 static int      PqRecvPointer;          /* Next index to read a byte from PqRecvBuffer */
 125 static int      PqRecvLength;           /* End of data available in PqRecvBuffer */
 126
 127 /*
 128  * Message status
 129  */
 130 static bool PqCommBusy;                 /* busy sending data to the client */
 131 static bool PqCommReadingMsg;   /* in the middle of reading a message */
 132 static bool DoingCopyOut;               /* in old-protocol COPY OUT processing */
 133
 134
 135 /* Internal functions */
 136 static void socket_comm_reset(void);
 137 static void socket_close(int code, Datum arg);
 138 static void socket_set_nonblocking(bool nonblocking);
 139 static int      socket_flush(void);
 140 static int      socket_flush_if_writable(void);
 141 static bool socket_is_send_pending(void);
 142 static int      socket_putmessage(char msgtype, const char *s, size_t len);
 143 static void socket_putmessage_noblock(char msgtype, const char *s, size_t len);
 144 static void socket_startcopyout(void);
 145 static void socket_endcopyout(bool errorAbort);
 146 static int      internal_putbytes(const char *s, size_t len);
 147 static int      internal_flush(void);
 148 static void socket_set_nonblocking(bool nonblocking);
 149
 150 #ifdef HAVE_UNIX_SOCKETS
 151 static int      Lock_AF_UNIX(char *unixSocketDir, char *unixSocketPath);
 152 static int      Setup_AF_UNIX(char *sock_path);
 153 #endif   /* HAVE_UNIX_SOCKETS */
 154
 155 static PQcommMethods PqCommSocketMethods = {
 156         socket_comm_reset,
 157         socket_flush,
 158         socket_flush_if_writable,
 159         socket_is_send_pending,
 160         socket_putmessage,
 161         socket_putmessage_noblock,
 162         socket_startcopyout,
 163         socket_endcopyout
 164 };
 165
 166 PQcommMethods *PqCommMethods = &PqCommSocketMethods;
 167
 168
 169
 170 /* --------------------------------
 171  *              pq_init - initialize libpq at backend startup
 172  * --------------------------------
 173  */
 174 void
 175 pq_init(void)
 176 {
 177         PqSendBufferSize = PQ_SEND_BUFFER_SIZE;
 178         PqSendBuffer = MemoryContextAlloc(TopMemoryContext, PqSendBufferSize);
 179         PqSendPointer = PqSendStart = PqRecvPointer = PqRecvLength = 0;
 180         PqCommBusy = false;
 181         PqCommReadingMsg = false;
 182         DoingCopyOut = false;
 183         on_proc_exit(socket_close, 0);
 184
 185         /*
 186          * In backends (as soon as forked) we operate the underlying socket in
 187          * nonblocking mode and use latches to implement blocking semantics if
 188          * needed. That allows us to provide safely interruptible reads and
 189          * writes.
 190          *
 191          * Use COMMERROR on failure, because ERROR would try to send the error to
 192          * the client, which might require changing the mode again, leading to
 193          * infinite recursion.
 194          */
 195 #ifndef WIN32
 196         if (!pg_set_noblock(MyProcPort->sock))
 197                 ereport(COMMERROR,
 198                                 (errmsg("could not set socket to nonblocking mode: %m")));
 199 #endif
 200
 201 }
 202
 203 /* --------------------------------
 204  *              socket_comm_reset - reset libpq during error recovery
 205  *
 206  * This is called from error recovery at the outer idle loop.  It's
 207  * just to get us out of trouble if we somehow manage to elog() from
 208  * inside a pqcomm.c routine (which ideally will never happen, but...)
 209  * --------------------------------
 210  */
 211 static void
 212 socket_comm_reset(void)
 213 {
 214         /* Do not throw away pending data, but do reset the busy flag */
 215         PqCommBusy = false;
 216         /* We can abort any old-style COPY OUT, too */
 217         pq_endcopyout(true);
 218 }
 219
 220 /* --------------------------------
 221  *              socket_close - shutdown libpq at backend exit
 222  *
 223  * This is the one pg_on_exit_callback in place during BackendInitialize().
 224  * That function's unusual signal handling constrains that this callback be
 225  * safe to run at any instant.
 226  * --------------------------------
 227  */
 228 static void
 229 socket_close(int code, Datum arg)
 230 {
 231         /* Nothing to do in a standalone backend, where MyProcPort is NULL. */
 232         if (MyProcPort != NULL)
 233         {
 234 #if defined(ENABLE_GSS) || defined(ENABLE_SSPI)
 235 #ifdef ENABLE_GSS
 236                 OM_uint32       min_s;
 237
 238                 /*
 239                  * Shutdown GSSAPI layer.  This section does nothing when interrupting
 240                  * BackendInitialize(), because pg_GSS_recvauth() makes first use of
 241                  * "ctx" and "cred".
 242                  */
 243                 if (MyProcPort->gss->ctx != GSS_C_NO_CONTEXT)
 244                         gss_delete_sec_context(&min_s, &MyProcPort->gss->ctx, NULL);
 245
 246                 if (MyProcPort->gss->cred != GSS_C_NO_CREDENTIAL)
 247                         gss_release_cred(&min_s, &MyProcPort->gss->cred);
 248 #endif   /* ENABLE_GSS */
 249
 250                 /*
 251                  * GSS and SSPI share the port->gss struct.  Since nowhere else does a
 252                  * postmaster child free this, doing so is safe when interrupting
 253                  * BackendInitialize().
 254                  */
 255                 free(MyProcPort->gss);
 256 #endif   /* ENABLE_GSS || ENABLE_SSPI */
 257
 258                 /*
 259                  * Cleanly shut down SSL layer.  Nowhere else does a postmaster child
 260                  * call this, so this is safe when interrupting BackendInitialize().
 261                  */
 262                 secure_close(MyProcPort);
 263
 264                 /*
 265                  * Formerly we did an explicit close() here, but it seems better to
 266                  * leave the socket open until the process dies.  This allows clients
 267                  * to perform a "synchronous close" if they care --- wait till the
 268                  * transport layer reports connection closure, and you can be sure the
 269                  * backend has exited.
 270                  *
 271                  * We do set sock to PGINVALID_SOCKET to prevent any further I/O,
 272                  * though.
 273                  */
 274                 MyProcPort->sock = PGINVALID_SOCKET;
 275         }
 276 }
 277
 278
 279
 280 /*
 281  * Streams -- wrapper around Unix socket system calls
 282  *
 283  *
 284  *              Stream functions are used for vanilla TCP connection protocol.
 285  */
 286
 287
 288 /* StreamDoUnlink()
 289  * Shutdown routine for backend connection
 290  * If any Unix sockets are used for communication, explicitly close them.
 291  */
 292 #ifdef HAVE_UNIX_SOCKETS
 293 static void
 294 StreamDoUnlink(int code, Datum arg)
 295 {
 296         ListCell   *l;
 297
 298         /* Loop through all created sockets... */
 299         foreach(l, sock_paths)
 300         {
 301                 char       *sock_path = (char *) lfirst(l);
 302
 303                 unlink(sock_path);
 304         }
 305         /* Since we're about to exit, no need to reclaim storage */
 306         sock_paths = NIL;
 307 }
 308 #endif   /* HAVE_UNIX_SOCKETS */
 309
 310 /*
 311  * StreamServerPort -- open a "listening" port to accept connections.
 312  *
 313  * family should be AF_UNIX or AF_UNSPEC; portNumber is the port number.
 314  * For AF_UNIX ports, hostName should be NULL and unixSocketDir must be
 315  * specified.  For TCP ports, hostName is either NULL for all interfaces or
 316  * the interface to listen on, and unixSocketDir is ignored (can be NULL).
 317  *
 318  * Successfully opened sockets are added to the ListenSocket[] array (of
 319  * length MaxListen), at the first position that isn't PGINVALID_SOCKET.
 320  *
 321  * RETURNS: STATUS_OK or STATUS_ERROR
 322  */
 323
 324 int
 325 StreamServerPort(int family, char *hostName, unsigned short portNumber,
 326                                  char *unixSocketDir,
 327                                  pgsocket ListenSocket[], int MaxListen)
 328 {
 329         pgsocket        fd;
 330         int                     err;
 331         int                     maxconn;
 332         int                     ret;
 333         char            portNumberStr[32];
 334         const char *familyDesc;
 335         char            familyDescBuf[64];
 336         char       *service;
 337         struct addrinfo *addrs = NULL,
 338                            *addr;
 339         struct addrinfo hint;
 340         int                     listen_index = 0;
 341         int                     added = 0;
 342
 343 #ifdef HAVE_UNIX_SOCKETS
 344         char            unixSocketPath[MAXPGPATH];
 345 #endif
 346 #if !defined(WIN32) || defined(IPV6_V6ONLY)
 347         int                     one = 1;
 348 #endif
 349
 350         /* Initialize hint structure */
 351         MemSet(&hint, 0, sizeof(hint));
 352         hint.ai_family = family;
 353         hint.ai_flags = AI_PASSIVE;
 354         hint.ai_socktype = SOCK_STREAM;
 355
 356 #ifdef HAVE_UNIX_SOCKETS
 357         if (family == AF_UNIX)
 358         {
 359                 /*
 360                  * Create unixSocketPath from portNumber and unixSocketDir and lock
 361                  * that file path
 362                  */
 363                 UNIXSOCK_PATH(unixSocketPath, portNumber, unixSocketDir);
 364                 if (strlen(unixSocketPath) >= UNIXSOCK_PATH_BUFLEN)
 365                 {
 366                         ereport(LOG,
 367                                         (errmsg("Unix-domain socket path \"%s\" is too long (maximum %d bytes)",
 368                                                         unixSocketPath,
 369                                                         (int) (UNIXSOCK_PATH_BUFLEN - 1))));
 370                         return STATUS_ERROR;
 371                 }
 372                 if (Lock_AF_UNIX(unixSocketDir, unixSocketPath) != STATUS_OK)
 373                         return STATUS_ERROR;
 374                 service = unixSocketPath;
 375         }
 376         else
 377 #endif   /* HAVE_UNIX_SOCKETS */
 378         {
 379                 snprintf(portNumberStr, sizeof(portNumberStr), "%d", portNumber);
 380                 service = portNumberStr;
 381         }
 382
 383         ret = pg_getaddrinfo_all(hostName, service, &hint, &addrs);
 384         if (ret || !addrs)
 385         {
 386                 if (hostName)
 387                         ereport(LOG,
 388                                         (errmsg("could not translate host name \"%s\", service \"%s\" to address: %s",
 389                                                         hostName, service, gai_strerror(ret))));
 390                 else
 391                         ereport(LOG,
 392                                  (errmsg("could not translate service \"%s\" to address: %s",
 393                                                  service, gai_strerror(ret))));
 394                 if (addrs)
 395                         pg_freeaddrinfo_all(hint.ai_family, addrs);
 396                 return STATUS_ERROR;
 397         }
 398
 399         for (addr = addrs; addr; addr = addr->ai_next)
 400         {
 401                 if (!IS_AF_UNIX(family) && IS_AF_UNIX(addr->ai_family))
 402                 {
 403                         /*
 404                          * Only set up a unix domain socket when they really asked for it.
 405                          * The service/port is different in that case.
 406                          */
 407                         continue;
 408                 }
 409
 410                 /* See if there is still room to add 1 more socket. */
 411                 for (; listen_index < MaxListen; listen_index++)
 412                 {
 413                         if (ListenSocket[listen_index] == PGINVALID_SOCKET)
 414                                 break;
 415                 }
 416                 if (listen_index >= MaxListen)
 417                 {
 418                         ereport(LOG,
 419                                         (errmsg("could not bind to all requested addresses: MAXLISTEN (%d) exceeded",
 420                                                         MaxListen)));
 421                         break;
 422                 }
 423
 424                 /* set up family name for possible error messages */
 425                 switch (addr->ai_family)
 426                 {
 427                         case AF_INET:
 428                                 familyDesc = _("IPv4");
 429                                 break;
 430 #ifdef HAVE_IPV6
 431                         case AF_INET6:
 432                                 familyDesc = _("IPv6");
 433                                 break;
 434 #endif
 435 #ifdef HAVE_UNIX_SOCKETS
 436                         case AF_UNIX:
 437                                 familyDesc = _("Unix");
 438                                 break;
 439 #endif
 440                         default:
 441                                 snprintf(familyDescBuf, sizeof(familyDescBuf),
 442                                                  _("unrecognized address family %d"),
 443                                                  addr->ai_family);
 444                                 familyDesc = familyDescBuf;
 445                                 break;
 446                 }
 447
 448                 if ((fd = socket(addr->ai_family, SOCK_STREAM, 0)) == PGINVALID_SOCKET)
 449                 {
 450                         ereport(LOG,
 451                                         (errcode_for_socket_access(),
 452                         /* translator: %s is IPv4, IPv6, or Unix */
 453                                          errmsg("could not create %s socket: %m",
 454                                                         familyDesc)));
 455                         continue;
 456                 }
 457
 458 #ifndef WIN32
 459
 460                 /*
 461                  * Without the SO_REUSEADDR flag, a new postmaster can't be started
 462                  * right away after a stop or crash, giving "address already in use"
 463                  * error on TCP ports.
 464                  *
 465                  * On win32, however, this behavior only happens if the
 466                  * SO_EXLUSIVEADDRUSE is set. With SO_REUSEADDR, win32 allows multiple
 467                  * servers to listen on the same address, resulting in unpredictable
 468                  * behavior. With no flags at all, win32 behaves as Unix with
 469                  * SO_REUSEADDR.
 470                  */
 471                 if (!IS_AF_UNIX(addr->ai_family))
 472                 {
 473                         if ((setsockopt(fd, SOL_SOCKET, SO_REUSEADDR,
 474                                                         (char *) &one, sizeof(one))) == -1)
 475                         {
 476                                 ereport(LOG,
 477                                                 (errcode_for_socket_access(),
 478                                                  errmsg("setsockopt(SO_REUSEADDR) failed: %m")));
 479                                 closesocket(fd);
 480                                 continue;
 481                         }
 482                 }
 483 #endif
 484
 485 #ifdef IPV6_V6ONLY
 486                 if (addr->ai_family == AF_INET6)
 487                 {
 488                         if (setsockopt(fd, IPPROTO_IPV6, IPV6_V6ONLY,
 489                                                    (char *) &one, sizeof(one)) == -1)
 490                         {
 491                                 ereport(LOG,
 492                                                 (errcode_for_socket_access(),
 493                                                  errmsg("setsockopt(IPV6_V6ONLY) failed: %m")));
 494                                 closesocket(fd);
 495                                 continue;
 496                         }
 497                 }
 498 #endif
 499
 500                 /*
 501                  * Note: This might fail on some OS's, like Linux older than
 502                  * 2.4.21-pre3, that don't have the IPV6_V6ONLY socket option, and map
 503                  * ipv4 addresses to ipv6.  It will show ::ffff:ipv4 for all ipv4
 504                  * connections.
 505                  */
 506                 err = bind(fd, addr->ai_addr, addr->ai_addrlen);
 507                 if (err < 0)
 508                 {
 509                         ereport(LOG,
 510                                         (errcode_for_socket_access(),
 511                         /* translator: %s is IPv4, IPv6, or Unix */
 512                                          errmsg("could not bind %s socket: %m",
 513                                                         familyDesc),
 514                                          (IS_AF_UNIX(addr->ai_family)) ?
 515                                   errhint("Is another postmaster already running on port %d?"
 516                                                   " If not, remove socket file \"%s\" and retry.",
 517                                                   (int) portNumber, service) :
 518                                   errhint("Is another postmaster already running on port %d?"
 519                                                   " If not, wait a few seconds and retry.",
 520                                                   (int) portNumber)));
 521                         closesocket(fd);
 522                         continue;
 523                 }
 524
 525 #ifdef HAVE_UNIX_SOCKETS
 526                 if (addr->ai_family == AF_UNIX)
 527                 {
 528                         if (Setup_AF_UNIX(service) != STATUS_OK)
 529                         {
 530                                 closesocket(fd);
 531                                 break;
 532                         }
 533                 }
 534 #endif
 535
 536                 /*
 537                  * Select appropriate accept-queue length limit.  PG_SOMAXCONN is only
 538                  * intended to provide a clamp on the request on platforms where an
 539                  * overly large request provokes a kernel error (are there any?).
 540                  */
 541                 maxconn = MaxBackends * 2;
 542                 if (maxconn > PG_SOMAXCONN)
 543                         maxconn = PG_SOMAXCONN;
 544
 545                 err = listen(fd, maxconn);
 546                 if (err < 0)
 547                 {
 548                         ereport(LOG,
 549                                         (errcode_for_socket_access(),
 550                         /* translator: %s is IPv4, IPv6, or Unix */
 551                                          errmsg("could not listen on %s socket: %m",
 552                                                         familyDesc)));
 553                         closesocket(fd);
 554                         continue;
 555                 }
 556                 ListenSocket[listen_index] = fd;
 557                 added++;
 558         }
 559
 560         pg_freeaddrinfo_all(hint.ai_family, addrs);
 561
 562         if (!added)
 563                 return STATUS_ERROR;
 564
 565         return STATUS_OK;
 566 }
 567
 568
 569 #ifdef HAVE_UNIX_SOCKETS
 570
 571 /*
 572  * Lock_AF_UNIX -- configure unix socket file path
 573  */
 574 static int
 575 Lock_AF_UNIX(char *unixSocketDir, char *unixSocketPath)
 576 {
 577         /*
 578          * Grab an interlock file associated with the socket file.
 579          *
 580          * Note: there are two reasons for using a socket lock file, rather than
 581          * trying to interlock directly on the socket itself.  First, it's a lot
 582          * more portable, and second, it lets us remove any pre-existing socket
 583          * file without race conditions.
 584          */
 585         CreateSocketLockFile(unixSocketPath, true, unixSocketDir);
 586
 587         /*
 588          * Once we have the interlock, we can safely delete any pre-existing
 589          * socket file to avoid failure at bind() time.
 590          */
 591         unlink(unixSocketPath);
 592
 593         /*
 594          * Arrange to unlink the socket file(s) at proc_exit.  If this is the
 595          * first one, set up the on_proc_exit function to do it; then add this
 596          * socket file to the list of files to unlink.
 597          */
 598         if (sock_paths == NIL)
 599                 on_proc_exit(StreamDoUnlink, 0);
 600
 601         sock_paths = lappend(sock_paths, pstrdup(unixSocketPath));
 602
 603         return STATUS_OK;
 604 }
 605
 606
 607 /*
 608  * Setup_AF_UNIX -- configure unix socket permissions
 609  */
 610 static int
 611 Setup_AF_UNIX(char *sock_path)
 612 {
 613         /*
 614          * Fix socket ownership/permission if requested.  Note we must do this
 615          * before we listen() to avoid a window where unwanted connections could
 616          * get accepted.
 617          */
 618         Assert(Unix_socket_group);
 619         if (Unix_socket_group[0] != '\0')
 620         {
 621 #ifdef WIN32
 622                 elog(WARNING, "configuration item unix_socket_group is not supported on this platform");
 623 #else
 624                 char       *endptr;
 625                 unsigned long val;
 626                 gid_t           gid;
 627
 628                 val = strtoul(Unix_socket_group, &endptr, 10);
 629                 if (*endptr == '\0')
 630                 {                                               /* numeric group id */
 631                         gid = val;
 632                 }
 633                 else
 634                 {                                               /* convert group name to id */
 635                         struct group *gr;
 636
 637                         gr = getgrnam(Unix_socket_group);
 638                         if (!gr)
 639                         {
 640                                 ereport(LOG,
 641                                                 (errmsg("group \"%s\" does not exist",
 642                                                                 Unix_socket_group)));
 643                                 return STATUS_ERROR;
 644                         }
 645                         gid = gr->gr_gid;
 646                 }
 647                 if (chown(sock_path, -1, gid) == -1)
 648                 {
 649                         ereport(LOG,
 650                                         (errcode_for_file_access(),
 651                                          errmsg("could not set group of file \"%s\": %m",
 652                                                         sock_path)));
 653                         return STATUS_ERROR;
 654                 }
 655 #endif
 656         }
 657
 658         if (chmod(sock_path, Unix_socket_permissions) == -1)
 659         {
 660                 ereport(LOG,
 661                                 (errcode_for_file_access(),
 662                                  errmsg("could not set permissions of file \"%s\": %m",
 663                                                 sock_path)));
 664                 return STATUS_ERROR;
 665         }
 666         return STATUS_OK;
 667 }
 668 #endif   /* HAVE_UNIX_SOCKETS */
 669
 670
 671 /*
 672  * StreamConnection -- create a new connection with client using
 673  *              server port.  Set port->sock to the FD of the new connection.
 674  *
 675  * ASSUME: that this doesn't need to be non-blocking because
 676  *              the Postmaster uses select() to tell when the server master
 677  *              socket is ready for accept().
 678  *
 679  * RETURNS: STATUS_OK or STATUS_ERROR
 680  */
 681 int
 682 StreamConnection(pgsocket server_fd, Port *port)
 683 {
 684         /* accept connection and fill in the client (remote) address */
 685         port->raddr.salen = sizeof(port->raddr.addr);
 686         if ((port->sock = accept(server_fd,
 687                                                          (struct sockaddr *) & port->raddr.addr,
 688                                                          &port->raddr.salen)) == PGINVALID_SOCKET)
 689         {
 690                 ereport(LOG,
 691                                 (errcode_for_socket_access(),
 692                                  errmsg("could not accept new connection: %m")));
 693
 694                 /*
 695                  * If accept() fails then postmaster.c will still see the server
 696                  * socket as read-ready, and will immediately try again.  To avoid
 697                  * uselessly sucking lots of CPU, delay a bit before trying again.
 698                  * (The most likely reason for failure is being out of kernel file
 699                  * table slots; we can do little except hope some will get freed up.)
 700                  */
 701                 pg_usleep(100000L);             /* wait 0.1 sec */
 702                 return STATUS_ERROR;
 703         }
 704
 705 #ifdef SCO_ACCEPT_BUG
 706
 707         /*
 708          * UnixWare 7+ and OpenServer 5.0.4 are known to have this bug, but it
 709          * shouldn't hurt to catch it for all versions of those platforms.
 710          */
 711         if (port->raddr.addr.ss_family == 0)
 712                 port->raddr.addr.ss_family = AF_UNIX;
 713 #endif
 714
 715         /* fill in the server (local) address */
 716         port->laddr.salen = sizeof(port->laddr.addr);
 717         if (getsockname(port->sock,
 718                                         (struct sockaddr *) & port->laddr.addr,
 719                                         &port->laddr.salen) < 0)
 720         {
 721                 elog(LOG, "getsockname() failed: %m");
 722                 return STATUS_ERROR;
 723         }
 724
 725         /* select NODELAY and KEEPALIVE options if it's a TCP connection */
 726         if (!IS_AF_UNIX(port->laddr.addr.ss_family))
 727         {
 728                 int                     on;
 729 #ifdef WIN32
 730                 int                     oldopt;
 731                 int                     optlen;
 732                 int                     newopt;
 733 #endif
 734
 735 #ifdef  TCP_NODELAY
 736                 on = 1;
 737                 if (setsockopt(port->sock, IPPROTO_TCP, TCP_NODELAY,
 738                                            (char *) &on, sizeof(on)) < 0)
 739                 {
 740                         elog(LOG, "setsockopt(TCP_NODELAY) failed: %m");
 741                         return STATUS_ERROR;
 742                 }
 743 #endif
 744                 on = 1;
 745                 if (setsockopt(port->sock, SOL_SOCKET, SO_KEEPALIVE,
 746                                            (char *) &on, sizeof(on)) < 0)
 747                 {
 748                         elog(LOG, "setsockopt(SO_KEEPALIVE) failed: %m");
 749                         return STATUS_ERROR;
 750                 }
 751
 752 #ifdef WIN32
 753
 754                 /*
 755                  * This is a Win32 socket optimization.  The OS send buffer should be
 756                  * large enough to send the whole Postgres send buffer in one go, or
 757                  * performance suffers.  The Postgres send buffer can be enlarged if a
 758                  * very large message needs to be sent, but we won't attempt to
 759                  * enlarge the OS buffer if that happens, so somewhat arbitrarily
 760                  * ensure that the OS buffer is at least PQ_SEND_BUFFER_SIZE * 4.
 761                  * (That's 32kB with the current default).
 762                  *
 763                  * The default OS buffer size used to be 8kB in earlier Windows
 764                  * versions, but was raised to 64kB in Windows 2012.  So it shouldn't
 765                  * be necessary to change it in later versions anymore.  Changing it
 766                  * unnecessarily can even reduce performance, because setting
 767                  * SO_SNDBUF in the application disables the "dynamic send buffering"
 768                  * feature that was introduced in Windows 7.  So before fiddling with
 769                  * SO_SNDBUF, check if the current buffer size is already large enough
 770                  * and only increase it if necessary.
 771                  *
 772                  * See https://support.microsoft.com/kb/823764/EN-US/ and
 773                  * https://msdn.microsoft.com/en-us/library/bb736549%28v=vs.85%29.aspx
 774                  */
 775                 optlen = sizeof(oldopt);
 776                 if (getsockopt(server_fd, SOL_SOCKET, SO_SNDBUF, (char *) &oldopt,
 777                                            &optlen) < 0)
 778                 {
 779                         elog(LOG, "getsockopt(SO_SNDBUF) failed: %m");
 780                         return STATUS_ERROR;
 781                 }
 782                 newopt = PQ_SEND_BUFFER_SIZE * 4;
 783                 if (oldopt < newopt)
 784                 {
 785                         if (setsockopt(port->sock, SOL_SOCKET, SO_SNDBUF, (char *) &newopt,
 786                                                    sizeof(newopt)) < 0)
 787                         {
 788                                 elog(LOG, "setsockopt(SO_SNDBUF) failed: %m");
 789                                 return STATUS_ERROR;
 790                         }
 791                 }
 792 #endif
 793
 794                 /*
 795                  * Also apply the current keepalive parameters.  If we fail to set a
 796                  * parameter, don't error out, because these aren't universally
 797                  * supported.  (Note: you might think we need to reset the GUC
 798                  * variables to 0 in such a case, but it's not necessary because the
 799                  * show hooks for these variables report the truth anyway.)
 800                  */
 801                 (void) pq_setkeepalivesidle(tcp_keepalives_idle, port);
 802                 (void) pq_setkeepalivesinterval(tcp_keepalives_interval, port);
 803                 (void) pq_setkeepalivescount(tcp_keepalives_count, port);
 804         }
 805
 806         return STATUS_OK;
 807 }
 808
 809 /*
 810  * StreamClose -- close a client/backend connection
 811  *
 812  * NOTE: this is NOT used to terminate a session; it is just used to release
 813  * the file descriptor in a process that should no longer have the socket
 814  * open.  (For example, the postmaster calls this after passing ownership
 815  * of the connection to a child process.)  It is expected that someone else
 816  * still has the socket open.  So, we only want to close the descriptor,
 817  * we do NOT want to send anything to the far end.
 818  */
 819 void
 820 StreamClose(pgsocket sock)
 821 {
 822         closesocket(sock);
 823 }
 824
 825 /*
 826  * TouchSocketFiles -- mark socket files as recently accessed
 827  *
 828  * This routine should be called every so often to ensure that the socket
 829  * files have a recent mod date (ordinary operations on sockets usually won't
 830  * change the mod date).  That saves them from being removed by
 831  * overenthusiastic /tmp-directory-cleaner daemons.  (Another reason we should
 832  * never have put the socket file in /tmp...)
 833  */
 834 void
 835 TouchSocketFiles(void)
 836 {
 837         ListCell   *l;
 838
 839         /* Loop through all created sockets... */
 840         foreach(l, sock_paths)
 841         {
 842                 char       *sock_path = (char *) lfirst(l);
 843
 844                 /*
 845                  * utime() is POSIX standard, utimes() is a common alternative. If we
 846                  * have neither, there's no way to affect the mod or access time of
 847                  * the socket :-(
 848                  *
 849                  * In either path, we ignore errors; there's no point in complaining.
 850                  */
 851 #ifdef HAVE_UTIME
 852                 utime(sock_path, NULL);
 853 #else                                                   /* !HAVE_UTIME */
 854 #ifdef HAVE_UTIMES
 855                 utimes(sock_path, NULL);
 856 #endif   /* HAVE_UTIMES */
 857 #endif   /* HAVE_UTIME */
 858         }
 859 }
 860
 861
 862 /* --------------------------------
 863  * Low-level I/O routines begin here.
 864  *
 865  * These routines communicate with a frontend client across a connection
 866  * already established by the preceding routines.
 867  * --------------------------------
 868  */
 869
 870 /* --------------------------------
 871  *                        socket_set_nonblocking - set socket blocking/non-blocking
 872  *
 873  * Sets the socket non-blocking if nonblocking is TRUE, or sets it
 874  * blocking otherwise.
 875  * --------------------------------
 876  */
 877 static void
 878 socket_set_nonblocking(bool nonblocking)
 879 {
 880         if (MyProcPort == NULL)
 881                 ereport(ERROR,
 882                                 (errcode(ERRCODE_CONNECTION_DOES_NOT_EXIST),
 883                                  errmsg("there is no client connection")));
 884
 885         MyProcPort->noblock = nonblocking;
 886 }
 887
 888 /* --------------------------------
 889  *              pq_recvbuf - load some bytes into the input buffer
 890  *
 891  *              returns 0 if OK, EOF if trouble
 892  * --------------------------------
 893  */
 894 static int
 895 pq_recvbuf(void)
 896 {
 897         if (PqRecvPointer > 0)
 898         {
 899                 if (PqRecvLength > PqRecvPointer)
 900                 {
 901                         /* still some unread data, left-justify it in the buffer */
 902                         memmove(PqRecvBuffer, PqRecvBuffer + PqRecvPointer,
 903                                         PqRecvLength - PqRecvPointer);
 904                         PqRecvLength -= PqRecvPointer;
 905                         PqRecvPointer = 0;
 906                 }
 907                 else
 908                         PqRecvLength = PqRecvPointer = 0;
 909         }
 910
 911         /* Ensure that we're in blocking mode */
 912         socket_set_nonblocking(false);
 913
 914         /* Can fill buffer from PqRecvLength and upwards */
 915         for (;;)
 916         {
 917                 int                     r;
 918
 919                 r = secure_read(MyProcPort, PqRecvBuffer + PqRecvLength,
 920                                                 PQ_RECV_BUFFER_SIZE - PqRecvLength);
 921
 922                 if (r < 0)
 923                 {
 924                         if (errno == EINTR)
 925                                 continue;               /* Ok if interrupted */
 926
 927                         /*
 928                          * Careful: an ereport() that tries to write to the client would
 929                          * cause recursion to here, leading to stack overflow and core
 930                          * dump!  This message must go *only* to the postmaster log.
 931                          */
 932                         ereport(COMMERROR,
 933                                         (errcode_for_socket_access(),
 934                                          errmsg("could not receive data from client: %m")));
 935                         return EOF;
 936                 }
 937                 if (r == 0)
 938                 {
 939                         /*
 940                          * EOF detected.  We used to write a log message here, but it's
 941                          * better to expect the ultimate caller to do that.
 942                          */
 943                         return EOF;
 944                 }
 945                 /* r contains number of bytes read, so just incr length */
 946                 PqRecvLength += r;
 947                 return 0;
 948         }
 949 }
 950
 951 /* --------------------------------
 952  *              pq_getbyte      - get a single byte from connection, or return EOF
 953  * --------------------------------
 954  */
 955 int
 956 pq_getbyte(void)
 957 {
 958         Assert(PqCommReadingMsg);
 959
 960         while (PqRecvPointer >= PqRecvLength)
 961         {
 962                 if (pq_recvbuf())               /* If nothing in buffer, then recv some */
 963                         return EOF;                     /* Failed to recv data */
 964         }
 965         return (unsigned char) PqRecvBuffer[PqRecvPointer++];
 966 }
 967
 968 /* --------------------------------
 969  *              pq_peekbyte             - peek at next byte from connection
 970  *
 971  *       Same as pq_getbyte() except we don't advance the pointer.
 972  * --------------------------------
 973  */
 974 int
 975 pq_peekbyte(void)
 976 {
 977         Assert(PqCommReadingMsg);
 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         return (unsigned char) PqRecvBuffer[PqRecvPointer];
 985 }
 986
 987 /* --------------------------------
 988  *              pq_getbyte_if_available - get a single byte from connection,
 989  *                      if available
 990  *
 991  * The received byte is stored in *c. Returns 1 if a byte was read,
 992  * 0 if no data was available, or EOF if trouble.
 993  * --------------------------------
 994  */
 995 int
 996 pq_getbyte_if_available(unsigned char *c)
 997 {
 998         int                     r;
 999
1000         Assert(PqCommReadingMsg);
1001
1002         if (PqRecvPointer < PqRecvLength)
1003         {
1004                 *c = PqRecvBuffer[PqRecvPointer++];
1005                 return 1;
1006         }
1007
1008         /* Put the socket into non-blocking mode */
1009         socket_set_nonblocking(true);
1010
1011         r = secure_read(MyProcPort, c, 1);
1012         if (r < 0)
1013         {
1014                 /*
1015                  * Ok if no data available without blocking or interrupted (though
1016                  * EINTR really shouldn't happen with a non-blocking socket). Report
1017                  * other errors.
1018                  */
1019                 if (errno == EAGAIN || errno == EWOULDBLOCK || errno == EINTR)
1020                         r = 0;
1021                 else
1022                 {
1023                         /*
1024                          * Careful: an ereport() that tries to write to the client would
1025                          * cause recursion to here, leading to stack overflow and core
1026                          * dump!  This message must go *only* to the postmaster log.
1027                          */
1028                         ereport(COMMERROR,
1029                                         (errcode_for_socket_access(),
1030                                          errmsg("could not receive data from client: %m")));
1031                         r = EOF;
1032                 }
1033         }
1034         else if (r == 0)
1035         {
1036                 /* EOF detected */
1037                 r = EOF;
1038         }
1039
1040         return r;
1041 }
1042
1043 /* --------------------------------
1044  *              pq_getbytes             - get a known number of bytes from connection
1045  *
1046  *              returns 0 if OK, EOF if trouble
1047  * --------------------------------
1048  */
1049 int
1050 pq_getbytes(char *s, size_t len)
1051 {
1052         size_t          amount;
1053
1054         Assert(PqCommReadingMsg);
1055
1056         while (len > 0)
1057         {
1058                 while (PqRecvPointer >= PqRecvLength)
1059                 {
1060                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
1061                                 return EOF;             /* Failed to recv data */
1062                 }
1063                 amount = PqRecvLength - PqRecvPointer;
1064                 if (amount > len)
1065                         amount = len;
1066                 memcpy(s, PqRecvBuffer + PqRecvPointer, amount);
1067                 PqRecvPointer += amount;
1068                 s += amount;
1069                 len -= amount;
1070         }
1071         return 0;
1072 }
1073
1074 /* --------------------------------
1075  *              pq_discardbytes         - throw away a known number of bytes
1076  *
1077  *              same as pq_getbytes except we do not copy the data to anyplace.
1078  *              this is used for resynchronizing after read errors.
1079  *
1080  *              returns 0 if OK, EOF if trouble
1081  * --------------------------------
1082  */
1083 static int
1084 pq_discardbytes(size_t len)
1085 {
1086         size_t          amount;
1087
1088         Assert(PqCommReadingMsg);
1089
1090         while (len > 0)
1091         {
1092                 while (PqRecvPointer >= PqRecvLength)
1093                 {
1094                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
1095                                 return EOF;             /* Failed to recv data */
1096                 }
1097                 amount = PqRecvLength - PqRecvPointer;
1098                 if (amount > len)
1099                         amount = len;
1100                 PqRecvPointer += amount;
1101                 len -= amount;
1102         }
1103         return 0;
1104 }
1105
1106 /* --------------------------------
1107  *              pq_getstring    - get a null terminated string from connection
1108  *
1109  *              The return value is placed in an expansible StringInfo, which has
1110  *              already been initialized by the caller.
1111  *
1112  *              This is used only for dealing with old-protocol clients.  The idea
1113  *              is to produce a StringInfo that looks the same as we would get from
1114  *              pq_getmessage() with a newer client; we will then process it with
1115  *              pq_getmsgstring.  Therefore, no character set conversion is done here,
1116  *              even though this is presumably useful only for text.
1117  *
1118  *              returns 0 if OK, EOF if trouble
1119  * --------------------------------
1120  */
1121 int
1122 pq_getstring(StringInfo s)
1123 {
1124         int                     i;
1125
1126         Assert(PqCommReadingMsg);
1127
1128         resetStringInfo(s);
1129
1130         /* Read until we get the terminating '\0' */
1131         for (;;)
1132         {
1133                 while (PqRecvPointer >= PqRecvLength)
1134                 {
1135                         if (pq_recvbuf())       /* If nothing in buffer, then recv some */
1136                                 return EOF;             /* Failed to recv data */
1137                 }
1138
1139                 for (i = PqRecvPointer; i < PqRecvLength; i++)
1140                 {
1141                         if (PqRecvBuffer[i] == '\0')
1142                         {
1143                                 /* include the '\0' in the copy */
1144                                 appendBinaryStringInfo(s, PqRecvBuffer + PqRecvPointer,
1145                                                                            i - PqRecvPointer + 1);
1146                                 PqRecvPointer = i + 1;  /* advance past \0 */
1147                                 return 0;
1148                         }
1149                 }
1150
1151                 /* If we're here we haven't got the \0 in the buffer yet. */
1152                 appendBinaryStringInfo(s, PqRecvBuffer + PqRecvPointer,
1153                                                            PqRecvLength - PqRecvPointer);
1154                 PqRecvPointer = PqRecvLength;
1155         }
1156 }
1157
1158
1159 /* --------------------------------
1160  *              pq_startmsgread - begin reading a message from the client.
1161  *
1162  *              This must be called before any of the pq_get* functions.
1163  * --------------------------------
1164  */
1165 void
1166 pq_startmsgread(void)
1167 {
1168         /*
1169          * There shouldn't be a read active already, but let's check just to be
1170          * sure.
1171          */
1172         if (PqCommReadingMsg)
1173                 ereport(FATAL,
1174                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1175                    errmsg("terminating connection because protocol sync was lost")));
1176
1177         PqCommReadingMsg = true;
1178 }
1179
1180
1181 /* --------------------------------
1182  *              pq_endmsgread   - finish reading message.
1183  *
1184  *              This must be called after reading a V2 protocol message with
1185  *              pq_getstring() and friends, to indicate that we have read the whole
1186  *              message. In V3 protocol, pq_getmessage() does this implicitly.
1187  * --------------------------------
1188  */
1189 void
1190 pq_endmsgread(void)
1191 {
1192         Assert(PqCommReadingMsg);
1193
1194         PqCommReadingMsg = false;
1195 }
1196
1197 /* --------------------------------
1198  *              pq_is_reading_msg - are we currently reading a message?
1199  *
1200  * This is used in error recovery at the outer idle loop to detect if we have
1201  * lost protocol sync, and need to terminate the connection. pq_startmsgread()
1202  * will check for that too, but it's nicer to detect it earlier.
1203  * --------------------------------
1204  */
1205 bool
1206 pq_is_reading_msg(void)
1207 {
1208         return PqCommReadingMsg;
1209 }
1210
1211 /* --------------------------------
1212  *              pq_getmessage   - get a message with length word from connection
1213  *
1214  *              The return value is placed in an expansible StringInfo, which has
1215  *              already been initialized by the caller.
1216  *              Only the message body is placed in the StringInfo; the length word
1217  *              is removed.  Also, s->cursor is initialized to zero for convenience
1218  *              in scanning the message contents.
1219  *
1220  *              If maxlen is not zero, it is an upper limit on the length of the
1221  *              message we are willing to accept.  We abort the connection (by
1222  *              returning EOF) if client tries to send more than that.
1223  *
1224  *              returns 0 if OK, EOF if trouble
1225  * --------------------------------
1226  */
1227 int
1228 pq_getmessage(StringInfo s, int maxlen)
1229 {
1230         int32           len;
1231
1232         Assert(PqCommReadingMsg);
1233
1234         resetStringInfo(s);
1235
1236         /* Read message length word */
1237         if (pq_getbytes((char *) &len, 4) == EOF)
1238         {
1239                 ereport(COMMERROR,
1240                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1241                                  errmsg("unexpected EOF within message length word")));
1242                 return EOF;
1243         }
1244
1245         len = ntohl(len);
1246
1247         if (len < 4 ||
1248                 (maxlen > 0 && len > maxlen))
1249         {
1250                 ereport(COMMERROR,
1251                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1252                                  errmsg("invalid message length")));
1253                 return EOF;
1254         }
1255
1256         len -= 4;                                       /* discount length itself */
1257
1258         if (len > 0)
1259         {
1260                 /*
1261                  * Allocate space for message.  If we run out of room (ridiculously
1262                  * large message), we will elog(ERROR), but we want to discard the
1263                  * message body so as not to lose communication sync.
1264                  */
1265                 PG_TRY();
1266                 {
1267                         enlargeStringInfo(s, len);
1268                 }
1269                 PG_CATCH();
1270                 {
1271                         if (pq_discardbytes(len) == EOF)
1272                                 ereport(COMMERROR,
1273                                                 (errcode(ERRCODE_PROTOCOL_VIOLATION),
1274                                                  errmsg("incomplete message from client")));
1275
1276                         /* we discarded the rest of the message so we're back in sync. */
1277                         PqCommReadingMsg = false;
1278                         PG_RE_THROW();
1279                 }
1280                 PG_END_TRY();
1281
1282                 /* And grab the message */
1283                 if (pq_getbytes(s->data, len) == EOF)
1284                 {
1285                         ereport(COMMERROR,
1286                                         (errcode(ERRCODE_PROTOCOL_VIOLATION),
1287                                          errmsg("incomplete message from client")));
1288                         return EOF;
1289                 }
1290                 s->len = len;
1291                 /* Place a trailing null per StringInfo convention */
1292                 s->data[len] = '\0';
1293         }
1294
1295         /* finished reading the message. */
1296         PqCommReadingMsg = false;
1297
1298         return 0;
1299 }
1300
1301
1302 /* --------------------------------
1303  *              pq_putbytes             - send bytes to connection (not flushed until pq_flush)
1304  *
1305  *              returns 0 if OK, EOF if trouble
1306  * --------------------------------
1307  */
1308 int
1309 pq_putbytes(const char *s, size_t len)
1310 {
1311         int                     res;
1312
1313         /* Should only be called by old-style COPY OUT */
1314         Assert(DoingCopyOut);
1315         /* No-op if reentrant call */
1316         if (PqCommBusy)
1317                 return 0;
1318         PqCommBusy = true;
1319         res = internal_putbytes(s, len);
1320         PqCommBusy = false;
1321         return res;
1322 }
1323
1324 static int
1325 internal_putbytes(const char *s, size_t len)
1326 {
1327         size_t          amount;
1328
1329         while (len > 0)
1330         {
1331                 /* If buffer is full, then flush it out */
1332                 if (PqSendPointer >= PqSendBufferSize)
1333                 {
1334                         socket_set_nonblocking(false);
1335                         if (internal_flush())
1336                                 return EOF;
1337                 }
1338                 amount = PqSendBufferSize - PqSendPointer;
1339                 if (amount > len)
1340                         amount = len;
1341                 memcpy(PqSendBuffer + PqSendPointer, s, amount);
1342                 PqSendPointer += amount;
1343                 s += amount;
1344                 len -= amount;
1345         }
1346         return 0;
1347 }
1348
1349 /* --------------------------------
1350  *              socket_flush            - flush pending output
1351  *
1352  *              returns 0 if OK, EOF if trouble
1353  * --------------------------------
1354  */
1355 static int
1356 socket_flush(void)
1357 {
1358         int                     res;
1359
1360         /* No-op if reentrant call */
1361         if (PqCommBusy)
1362                 return 0;
1363         PqCommBusy = true;
1364         socket_set_nonblocking(false);
1365         res = internal_flush();
1366         PqCommBusy = false;
1367         return res;
1368 }
1369
1370 /* --------------------------------
1371  *              internal_flush - flush pending output
1372  *
1373  * Returns 0 if OK (meaning everything was sent, or operation would block
1374  * and the socket is in non-blocking mode), or EOF if trouble.
1375  * --------------------------------
1376  */
1377 static int
1378 internal_flush(void)
1379 {
1380         static int      last_reported_send_errno = 0;
1381
1382         char       *bufptr = PqSendBuffer + PqSendStart;
1383         char       *bufend = PqSendBuffer + PqSendPointer;
1384
1385         while (bufptr < bufend)
1386         {
1387                 int                     r;
1388
1389                 r = secure_write(MyProcPort, bufptr, bufend - bufptr);
1390
1391                 if (r <= 0)
1392                 {
1393                         if (errno == EINTR)
1394                                 continue;               /* Ok if we were interrupted */
1395
1396                         /*
1397                          * Ok if no data writable without blocking, and the socket is in
1398                          * non-blocking mode.
1399                          */
1400                         if (errno == EAGAIN ||
1401                                 errno == EWOULDBLOCK)
1402                         {
1403                                 return 0;
1404                         }
1405
1406                         /*
1407                          * Careful: an ereport() that tries to write to the client would
1408                          * cause recursion to here, leading to stack overflow and core
1409                          * dump!  This message must go *only* to the postmaster log.
1410                          *
1411                          * If a client disconnects while we're in the midst of output, we
1412                          * might write quite a bit of data before we get to a safe query
1413                          * abort point.  So, suppress duplicate log messages.
1414                          */
1415                         if (errno != last_reported_send_errno)
1416                         {
1417                                 last_reported_send_errno = errno;
1418                                 ereport(COMMERROR,
1419                                                 (errcode_for_socket_access(),
1420                                                  errmsg("could not send data to client: %m")));
1421                         }
1422
1423                         /*
1424                          * We drop the buffered data anyway so that processing can
1425                          * continue, even though we'll probably quit soon. We also set a
1426                          * flag that'll cause the next CHECK_FOR_INTERRUPTS to terminate
1427                          * the connection.
1428                          */
1429                         PqSendStart = PqSendPointer = 0;
1430                         ClientConnectionLost = 1;
1431                         InterruptPending = 1;
1432                         return EOF;
1433                 }
1434
1435                 last_reported_send_errno = 0;   /* reset after any successful send */
1436                 bufptr += r;
1437                 PqSendStart += r;
1438         }
1439
1440         PqSendStart = PqSendPointer = 0;
1441         return 0;
1442 }
1443
1444 /* --------------------------------
1445  *              pq_flush_if_writable - flush pending output if writable without blocking
1446  *
1447  * Returns 0 if OK, or EOF if trouble.
1448  * --------------------------------
1449  */
1450 static int
1451 socket_flush_if_writable(void)
1452 {
1453         int                     res;
1454
1455         /* Quick exit if nothing to do */
1456         if (PqSendPointer == PqSendStart)
1457                 return 0;
1458
1459         /* No-op if reentrant call */
1460         if (PqCommBusy)
1461                 return 0;
1462
1463         /* Temporarily put the socket into non-blocking mode */
1464         socket_set_nonblocking(true);
1465
1466         PqCommBusy = true;
1467         res = internal_flush();
1468         PqCommBusy = false;
1469         return res;
1470 }
1471
1472 /* --------------------------------
1473  *      socket_is_send_pending  - is there any pending data in the output buffer?
1474  * --------------------------------
1475  */
1476 static bool
1477 socket_is_send_pending(void)
1478 {
1479         return (PqSendStart < PqSendPointer);
1480 }
1481
1482 /* --------------------------------
1483  * Message-level I/O routines begin here.
1484  *
1485  * These routines understand about the old-style COPY OUT protocol.
1486  * --------------------------------
1487  */
1488
1489
1490 /* --------------------------------
1491  *              socket_putmessage - send a normal message (suppressed in COPY OUT mode)
1492  *
1493  *              If msgtype is not '\0', it is a message type code to place before
1494  *              the message body.  If msgtype is '\0', then the message has no type
1495  *              code (this is only valid in pre-3.0 protocols).
1496  *
1497  *              len is the length of the message body data at *s.  In protocol 3.0
1498  *              and later, a message length word (equal to len+4 because it counts
1499  *              itself too) is inserted by this routine.
1500  *
1501  *              All normal messages are suppressed while old-style COPY OUT is in
1502  *              progress.  (In practice only a few notice messages might get emitted
1503  *              then; dropping them is annoying, but at least they will still appear
1504  *              in the postmaster log.)
1505  *
1506  *              We also suppress messages generated while pqcomm.c is busy.  This
1507  *              avoids any possibility of messages being inserted within other
1508  *              messages.  The only known trouble case arises if SIGQUIT occurs
1509  *              during a pqcomm.c routine --- quickdie() will try to send a warning
1510  *              message, and the most reasonable approach seems to be to drop it.
1511  *
1512  *              returns 0 if OK, EOF if trouble
1513  * --------------------------------
1514  */
1515 static int
1516 socket_putmessage(char msgtype, const char *s, size_t len)
1517 {
1518         if (DoingCopyOut || PqCommBusy)
1519                 return 0;
1520         PqCommBusy = true;
1521         if (msgtype)
1522                 if (internal_putbytes(&msgtype, 1))
1523                         goto fail;
1524         if (PG_PROTOCOL_MAJOR(FrontendProtocol) >= 3)
1525         {
1526                 uint32          n32;
1527
1528                 n32 = htonl((uint32) (len + 4));
1529                 if (internal_putbytes((char *) &n32, 4))
1530                         goto fail;
1531         }
1532         if (internal_putbytes(s, len))
1533                 goto fail;
1534         PqCommBusy = false;
1535         return 0;
1536
1537 fail:
1538         PqCommBusy = false;
1539         return EOF;
1540 }
1541
1542 /* --------------------------------
1543  *              pq_putmessage_noblock   - like pq_putmessage, but never blocks
1544  *
1545  *              If the output buffer is too small to hold the message, the buffer
1546  *              is enlarged.
1547  */
1548 static void
1549 socket_putmessage_noblock(char msgtype, const char *s, size_t len)
1550 {
1551         int res         PG_USED_FOR_ASSERTS_ONLY;
1552         int                     required;
1553
1554         /*
1555          * Ensure we have enough space in the output buffer for the message header
1556          * as well as the message itself.
1557          */
1558         required = PqSendPointer + 1 + 4 + len;
1559         if (required > PqSendBufferSize)
1560         {
1561                 PqSendBuffer = repalloc(PqSendBuffer, required);
1562                 PqSendBufferSize = required;
1563         }
1564         res = pq_putmessage(msgtype, s, len);
1565         Assert(res == 0);                       /* should not fail when the message fits in
1566                                                                  * buffer */
1567 }
1568
1569
1570 /* --------------------------------
1571  *              socket_startcopyout - inform libpq that an old-style COPY OUT transfer
1572  *                      is beginning
1573  * --------------------------------
1574  */
1575 static void
1576 socket_startcopyout(void)
1577 {
1578         DoingCopyOut = true;
1579 }
1580
1581 /* --------------------------------
1582  *              socket_endcopyout       - end an old-style COPY OUT transfer
1583  *
1584  *              If errorAbort is indicated, we are aborting a COPY OUT due to an error,
1585  *              and must send a terminator line.  Since a partial data line might have
1586  *              been emitted, send a couple of newlines first (the first one could
1587  *              get absorbed by a backslash...)  Note that old-style COPY OUT does
1588  *              not allow binary transfers, so a textual terminator is always correct.
1589  * --------------------------------
1590  */
1591 static void
1592 socket_endcopyout(bool errorAbort)
1593 {
1594         if (!DoingCopyOut)
1595                 return;
1596         if (errorAbort)
1597                 pq_putbytes("\n\n\\.\n", 5);
1598         /* in non-error case, copy.c will have emitted the terminator line */
1599         DoingCopyOut = false;
1600 }
1601
1602 /*
1603  * Support for TCP Keepalive parameters
1604  */
1605
1606 /*
1607  * On Windows, we need to set both idle and interval at the same time.
1608  * We also cannot reset them to the default (setting to zero will
1609  * actually set them to zero, not default), therefore we fallback to
1610  * the out-of-the-box default instead.
1611  */
1612 #if defined(WIN32) && defined(SIO_KEEPALIVE_VALS)
1613 static int
1614 pq_setkeepaliveswin32(Port *port, int idle, int interval)
1615 {
1616         struct tcp_keepalive ka;
1617         DWORD           retsize;
1618
1619         if (idle <= 0)
1620                 idle = 2 * 60 * 60;             /* default = 2 hours */
1621         if (interval <= 0)
1622                 interval = 1;                   /* default = 1 second */
1623
1624         ka.onoff = 1;
1625         ka.keepalivetime = idle * 1000;
1626         ka.keepaliveinterval = interval * 1000;
1627
1628         if (WSAIoctl(port->sock,
1629                                  SIO_KEEPALIVE_VALS,
1630                                  (LPVOID) &ka,
1631                                  sizeof(ka),
1632                                  NULL,
1633                                  0,
1634                                  &retsize,
1635                                  NULL,
1636                                  NULL)
1637                 != 0)
1638         {
1639                 elog(LOG, "WSAIoctl(SIO_KEEPALIVE_VALS) failed: %ui",
1640                          WSAGetLastError());
1641                 return STATUS_ERROR;
1642         }
1643         if (port->keepalives_idle != idle)
1644                 port->keepalives_idle = idle;
1645         if (port->keepalives_interval != interval)
1646                 port->keepalives_interval = interval;
1647         return STATUS_OK;
1648 }
1649 #endif
1650
1651 int
1652 pq_getkeepalivesidle(Port *port)
1653 {
1654 #if defined(TCP_KEEPIDLE) || defined(TCP_KEEPALIVE) || defined(WIN32)
1655         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1656                 return 0;
1657
1658         if (port->keepalives_idle != 0)
1659                 return port->keepalives_idle;
1660
1661         if (port->default_keepalives_idle == 0)
1662         {
1663 #ifndef WIN32
1664                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_idle);
1665
1666 #ifdef TCP_KEEPIDLE
1667                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPIDLE,
1668                                            (char *) &port->default_keepalives_idle,
1669                                            &size) < 0)
1670                 {
1671                         elog(LOG, "getsockopt(TCP_KEEPIDLE) failed: %m");
1672                         port->default_keepalives_idle = -1; /* don't know */
1673                 }
1674 #else
1675                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPALIVE,
1676                                            (char *) &port->default_keepalives_idle,
1677                                            &size) < 0)
1678                 {
1679                         elog(LOG, "getsockopt(TCP_KEEPALIVE) failed: %m");
1680                         port->default_keepalives_idle = -1; /* don't know */
1681                 }
1682 #endif   /* TCP_KEEPIDLE */
1683 #else                                                   /* WIN32 */
1684                 /* We can't get the defaults on Windows, so return "don't know" */
1685                 port->default_keepalives_idle = -1;
1686 #endif   /* WIN32 */
1687         }
1688
1689         return port->default_keepalives_idle;
1690 #else
1691         return 0;
1692 #endif
1693 }
1694
1695 int
1696 pq_setkeepalivesidle(int idle, Port *port)
1697 {
1698         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1699                 return STATUS_OK;
1700
1701 #if defined(TCP_KEEPIDLE) || defined(TCP_KEEPALIVE) || defined(SIO_KEEPALIVE_VALS)
1702         if (idle == port->keepalives_idle)
1703                 return STATUS_OK;
1704
1705 #ifndef WIN32
1706         if (port->default_keepalives_idle <= 0)
1707         {
1708                 if (pq_getkeepalivesidle(port) < 0)
1709                 {
1710                         if (idle == 0)
1711                                 return STATUS_OK;               /* default is set but unknown */
1712                         else
1713                                 return STATUS_ERROR;
1714                 }
1715         }
1716
1717         if (idle == 0)
1718                 idle = port->default_keepalives_idle;
1719
1720 #ifdef TCP_KEEPIDLE
1721         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPIDLE,
1722                                    (char *) &idle, sizeof(idle)) < 0)
1723         {
1724                 elog(LOG, "setsockopt(TCP_KEEPIDLE) failed: %m");
1725                 return STATUS_ERROR;
1726         }
1727 #else
1728         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPALIVE,
1729                                    (char *) &idle, sizeof(idle)) < 0)
1730         {
1731                 elog(LOG, "setsockopt(TCP_KEEPALIVE) failed: %m");
1732                 return STATUS_ERROR;
1733         }
1734 #endif
1735
1736         port->keepalives_idle = idle;
1737 #else                                                   /* WIN32 */
1738         return pq_setkeepaliveswin32(port, idle, port->keepalives_interval);
1739 #endif
1740 #else                                                   /* TCP_KEEPIDLE || SIO_KEEPALIVE_VALS */
1741         if (idle != 0)
1742         {
1743                 elog(LOG, "setting the keepalive idle time is not supported");
1744                 return STATUS_ERROR;
1745         }
1746 #endif
1747         return STATUS_OK;
1748 }
1749
1750 int
1751 pq_getkeepalivesinterval(Port *port)
1752 {
1753 #if defined(TCP_KEEPINTVL) || defined(SIO_KEEPALIVE_VALS)
1754         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1755                 return 0;
1756
1757         if (port->keepalives_interval != 0)
1758                 return port->keepalives_interval;
1759
1760         if (port->default_keepalives_interval == 0)
1761         {
1762 #ifndef WIN32
1763                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_interval);
1764
1765                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPINTVL,
1766                                            (char *) &port->default_keepalives_interval,
1767                                            &size) < 0)
1768                 {
1769                         elog(LOG, "getsockopt(TCP_KEEPINTVL) failed: %m");
1770                         port->default_keepalives_interval = -1;         /* don't know */
1771                 }
1772 #else
1773                 /* We can't get the defaults on Windows, so return "don't know" */
1774                 port->default_keepalives_interval = -1;
1775 #endif   /* WIN32 */
1776         }
1777
1778         return port->default_keepalives_interval;
1779 #else
1780         return 0;
1781 #endif
1782 }
1783
1784 int
1785 pq_setkeepalivesinterval(int interval, Port *port)
1786 {
1787         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1788                 return STATUS_OK;
1789
1790 #if defined(TCP_KEEPINTVL) || defined (SIO_KEEPALIVE_VALS)
1791         if (interval == port->keepalives_interval)
1792                 return STATUS_OK;
1793
1794 #ifndef WIN32
1795         if (port->default_keepalives_interval <= 0)
1796         {
1797                 if (pq_getkeepalivesinterval(port) < 0)
1798                 {
1799                         if (interval == 0)
1800                                 return STATUS_OK;               /* default is set but unknown */
1801                         else
1802                                 return STATUS_ERROR;
1803                 }
1804         }
1805
1806         if (interval == 0)
1807                 interval = port->default_keepalives_interval;
1808
1809         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPINTVL,
1810                                    (char *) &interval, sizeof(interval)) < 0)
1811         {
1812                 elog(LOG, "setsockopt(TCP_KEEPINTVL) failed: %m");
1813                 return STATUS_ERROR;
1814         }
1815
1816         port->keepalives_interval = interval;
1817 #else                                                   /* WIN32 */
1818         return pq_setkeepaliveswin32(port, port->keepalives_idle, interval);
1819 #endif
1820 #else
1821         if (interval != 0)
1822         {
1823                 elog(LOG, "setsockopt(TCP_KEEPINTVL) not supported");
1824                 return STATUS_ERROR;
1825         }
1826 #endif
1827
1828         return STATUS_OK;
1829 }
1830
1831 int
1832 pq_getkeepalivescount(Port *port)
1833 {
1834 #ifdef TCP_KEEPCNT
1835         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1836                 return 0;
1837
1838         if (port->keepalives_count != 0)
1839                 return port->keepalives_count;
1840
1841         if (port->default_keepalives_count == 0)
1842         {
1843                 ACCEPT_TYPE_ARG3 size = sizeof(port->default_keepalives_count);
1844
1845                 if (getsockopt(port->sock, IPPROTO_TCP, TCP_KEEPCNT,
1846                                            (char *) &port->default_keepalives_count,
1847                                            &size) < 0)
1848                 {
1849                         elog(LOG, "getsockopt(TCP_KEEPCNT) failed: %m");
1850                         port->default_keepalives_count = -1;            /* don't know */
1851                 }
1852         }
1853
1854         return port->default_keepalives_count;
1855 #else
1856         return 0;
1857 #endif
1858 }
1859
1860 int
1861 pq_setkeepalivescount(int count, Port *port)
1862 {
1863         if (port == NULL || IS_AF_UNIX(port->laddr.addr.ss_family))
1864                 return STATUS_OK;
1865
1866 #ifdef TCP_KEEPCNT
1867         if (count == port->keepalives_count)
1868                 return STATUS_OK;
1869
1870         if (port->default_keepalives_count <= 0)
1871         {
1872                 if (pq_getkeepalivescount(port) < 0)
1873                 {
1874                         if (count == 0)
1875                                 return STATUS_OK;               /* default is set but unknown */
1876                         else
1877                                 return STATUS_ERROR;
1878                 }
1879         }
1880
1881         if (count == 0)
1882                 count = port->default_keepalives_count;
1883
1884         if (setsockopt(port->sock, IPPROTO_TCP, TCP_KEEPCNT,
1885                                    (char *) &count, sizeof(count)) < 0)
1886         {
1887                 elog(LOG, "setsockopt(TCP_KEEPCNT) failed: %m");
1888                 return STATUS_ERROR;
1889         }
1890
1891         port->keepalives_count = count;
1892 #else
1893         if (count != 0)
1894         {
1895                 elog(LOG, "setsockopt(TCP_KEEPCNT) not supported");
1896                 return STATUS_ERROR;
1897         }
1898 #endif
1899
1900         return STATUS_OK;
1901 }