nbd-server.c

   1 /*
   2  * Network Block Device - server
   3  *
   4  * Copyright 1996-1998 Pavel Machek, distribute under GPL
   5  *  <pavel@atrey.karlin.mff.cuni.cz>
   6  * Copyright 2001-2004 Wouter Verhelst <wouter@debian.org>
   7  * Copyright 2002 Anton Altaparmakov <aia21@cam.ac.uk>
   8  *
   9  * Version 1.0 - hopefully 64-bit-clean
  10  * Version 1.1 - merging enhancements from Josh Parsons, <josh@coombs.anu.edu.au>
  11  * Version 1.2 - autodetect size of block devices, thanx to Peter T. Breuer" <ptb@it.uc3m.es>
  12  * Version 1.5 - can compile on Unix systems that don't have 64 bit integer
  13  *      type, or don't have 64 bit file offsets by defining FS_32BIT
  14  *      in compile options for nbd-server *only*. This can be done
  15  *      with make FSCHOICE=-DFS_32BIT nbd-server. (I don't have the
  16  *      original autoconf input file, or I would make it a configure
  17  *      option.) Ken Yap <ken@nlc.net.au>.
  18  * Version 1.6 - fix autodetection of block device size and really make 64 bit
  19  *      clean on 32 bit machines. Anton Altaparmakov <aia21@cam.ac.uk>
  20  * Version 2.0 - Version synchronised with client
  21  * Version 2.1 - Reap zombie client processes when they exit. Removed
  22  *      (uncommented) the _IO magic, it's no longer necessary. Wouter
  23  *      Verhelst <wouter@debian.org>
  24  * Version 2.2 - Auto switch to read-only mode (usefull for floppies).
  25  * Version 2.3 - Fixed code so that Large File Support works. This
  26  *      removes the FS_32BIT compile-time directive; define
  27  *      _FILE_OFFSET_BITS=64 and _LARGEFILE_SOURCE if you used to be
  28  *      using FS_32BIT. This will allow you to use files >2GB instead of
  29  *      having to use the -m option. Wouter Verhelst <wouter@debian.org>
  30  * Version 2.4 - Added code to keep track of children, so that we can
  31  *      properly kill them from initscripts. Add a call to daemon(),
  32  *      so that processes don't think they have to wait for us, which is
  33  *      interesting for initscripts as well. Wouter Verhelst
  34  *      <wouter@debian.org>
  35  * Version 2.5 - Bugfix release: forgot to reset child_arraysize to
  36  *      zero after fork()ing, resulting in nbd-server going berserk
  37  *      when it receives a signal with at least one child open. Wouter
  38  *      Verhelst <wouter@debian.org>
  39  * 10/10/2003 - Added socket option SO_KEEPALIVE (sf.net bug 819235);
  40  *      rectified type of mainloop::size_host (sf.net bugs 814435 and
  41  *      817385); close the PID file after writing to it, so that the
  42  *      daemon can actually be found. Wouter Verhelst
  43  *      <wouter@debian.org>
  44  * 10/10/2003 - Size of the data "size_host" was wrong and so was not
  45  *      correctly put in network endianness. Many types were corrected
  46  *      (size_t and off_t instead of int).  <vspaceg@sourceforge.net>
  47  * Version 2.6 - Some code cleanup.
  48  * Version 2.7 - Better build system.
  49  * 11/02/2004 - Doxygenified the source, modularized it a bit. Needs a
  50  *      lot more work, but this is a start. Wouter Verhelst
  51  *      <wouter@debian.org>
  52  * 16/03/2010 - Add IPv6 support.
  53  *      Kitt Tientanopajai <kitt@kitty.in.th>
  54  *      Neutron Soutmun <neo.neutron@gmail.com>
  55  *      Suriya Soutmun <darksolar@gmail.com>
  56  */
  57
  58 /* Includes LFS defines, which defines behaviours of some of the following
  59  * headers, so must come before those */
  60 #include "lfs.h"
  61
  62 #include <sys/types.h>
  63 #include <sys/socket.h>
  64 #include <sys/stat.h>
  65 #include <sys/select.h>         /* select */
  66 #include <sys/wait.h>           /* wait */
  67 #ifdef HAVE_SYS_IOCTL_H
  68 #include <sys/ioctl.h>
  69 #endif
  70 #include <sys/param.h>
  71 #ifdef HAVE_SYS_MOUNT_H
  72 #include <sys/mount.h>          /* For BLKGETSIZE */
  73 #endif
  74 #include <signal.h>             /* sigaction */
  75 #include <errno.h>
  76 #include <netinet/tcp.h>
  77 #include <netinet/in.h>
  78 #include <netdb.h>
  79 #include <syslog.h>
  80 #include <unistd.h>
  81 #include <stdio.h>
  82 #include <stdlib.h>
  83 #include <string.h>
  84 #include <fcntl.h>
  85 #include <arpa/inet.h>
  86 #include <strings.h>
  87 #include <dirent.h>
  88 #include <unistd.h>
  89 #include <getopt.h>
  90 #include <pwd.h>
  91 #include <grp.h>
  92
  93 #include <glib.h>
  94
  95 /* used in cliserv.h, so must come first */
  96 #define MY_NAME "nbd_server"
  97 #include "cliserv.h"
  98
  99 /** Default position of the config file */
 100 #ifndef SYSCONFDIR
 101 #define SYSCONFDIR "/etc"
 102 #endif
 103 #define CFILE SYSCONFDIR "/nbd-server/config"
 104
 105 /** Where our config file actually is */
 106 gchar* config_file_pos;
 107
 108 /** What user we're running as */
 109 gchar* runuser=NULL;
 110 /** What group we're running as */
 111 gchar* rungroup=NULL;
 112 /** whether to export using the old negotiation protocol (port-based) */
 113 gboolean do_oldstyle=FALSE;
 114
 115 /** Logging macros, now nothing goes to syslog unless you say ISSERVER */
 116 #ifdef ISSERVER
 117 #define msg2(a,b) syslog(a,b)
 118 #define msg3(a,b,c) syslog(a,b,c)
 119 #define msg4(a,b,c,d) syslog(a,b,c,d)
 120 #else
 121 #define msg2(a,b) g_message(b)
 122 #define msg3(a,b,c) g_message(b,c)
 123 #define msg4(a,b,c,d) g_message(b,c,d)
 124 #endif
 125
 126 /* Debugging macros */
 127 //#define DODBG
 128 #ifdef DODBG
 129 #define DEBUG( a ) printf( a )
 130 #define DEBUG2( a,b ) printf( a,b )
 131 #define DEBUG3( a,b,c ) printf( a,b,c )
 132 #define DEBUG4( a,b,c,d ) printf( a,b,c,d )
 133 #else
 134 #define DEBUG( a )
 135 #define DEBUG2( a,b )
 136 #define DEBUG3( a,b,c )
 137 #define DEBUG4( a,b,c,d )
 138 #endif
 139 #ifndef PACKAGE_VERSION
 140 #define PACKAGE_VERSION ""
 141 #endif
 142 /**
 143  * The highest value a variable of type off_t can reach. This is a signed
 144  * integer, so set all bits except for the leftmost one.
 145  **/
 146 #define OFFT_MAX ~((off_t)1<<(sizeof(off_t)*8-1))
 147 #define LINELEN 256       /**< Size of static buffer used to read the
 148                                authorization file (yuck) */
 149 #define BUFSIZE (1024*1024) /**< Size of buffer that can hold requests */
 150 #define DIFFPAGESIZE 4096 /**< diff file uses those chunks */
 151 #define F_READONLY 1      /**< flag to tell us a file is readonly */
 152 #define F_MULTIFILE 2     /**< flag to tell us a file is exported using -m */
 153 #define F_COPYONWRITE 4   /**< flag to tell us a file is exported using
 154                             copyonwrite */
 155 #define F_AUTOREADONLY 8  /**< flag to tell us a file is set to autoreadonly */
 156 #define F_SPARSE 16       /**< flag to tell us copyronwrite should use a sparse file */
 157 #define F_SDP 32          /**< flag to tell us the export should be done using the Socket Direct Protocol for RDMA */
 158 #define F_SYNC 64         /**< Whether to fsync() after a write */
 159 GHashTable *children;
 160 char pidfname[256]; /**< name of our PID file */
 161 char pidftemplate[256]; /**< template to be used for the filename of the PID file */
 162 char default_authname[] = SYSCONFDIR "/nbd-server/allow"; /**< default name of allow file */
 163
 164 int modernsock=0;         /**< Socket for the modern handler. Not used
 165                                if a client was only specified on the
 166                                command line; only port used if
 167                                oldstyle is set to false (and then the
 168                                command-line client isn't used, gna gna) */
 169 char* modern_listen;      /**< listenaddr value for modernsock */
 170
 171 /**
 172  * Types of virtuatlization
 173  **/
 174 typedef enum {
 175         VIRT_NONE=0,    /**< No virtualization */
 176         VIRT_IPLIT,     /**< Literal IP address as part of the filename */
 177         VIRT_IPHASH,    /**< Replacing all dots in an ip address by a / before
 178                              doing the same as in IPLIT */
 179         VIRT_CIDR,      /**< Every subnet in its own directory */
 180 } VIRT_STYLE;
 181
 182 /**
 183  * Variables associated with a server.
 184  **/
 185 typedef struct {
 186         gchar* exportname;    /**< (unprocessed) filename of the file we're exporting */
 187         off_t expected_size; /**< size of the exported file as it was told to
 188                                us through configuration */
 189         gchar* listenaddr;   /**< The IP address we're listening on */
 190         unsigned int port;   /**< port we're exporting this file at */
 191         char* authname;      /**< filename of the authorization file */
 192         int flags;           /**< flags associated with this exported file */
 193         int socket;          /**< The socket of this server. */
 194         int socket_family;   /**< family of the socket */
 195         VIRT_STYLE virtstyle;/**< The style of virtualization, if any */
 196         uint8_t cidrlen;     /**< The length of the mask when we use
 197                                   CIDR-style virtualization */
 198         gchar* prerun;       /**< command to be ran after connecting a client,
 199                                   but before starting to serve */
 200         gchar* postrun;      /**< command that will be ran after the client
 201                                   disconnects */
 202         gchar* servename;    /**< name of the export as selected by nbd-client */
 203 } SERVER;
 204
 205 /**
 206  * Variables associated with a client socket.
 207  **/
 208 typedef struct {
 209         int fhandle;      /**< file descriptor */
 210         off_t startoff;   /**< starting offset of this file */
 211 } FILE_INFO;
 212
 213 typedef struct {
 214         off_t exportsize;    /**< size of the file we're exporting */
 215         char *clientname;    /**< peer */
 216         char *exportname;    /**< (processed) filename of the file we're exporting */
 217         GArray *export;    /**< array of FILE_INFO of exported files;
 218                                array size is always 1 unless we're
 219                                doing the multiple file option */
 220         int net;             /**< The actual client socket */
 221         SERVER *server;      /**< The server this client is getting data from */
 222         char* difffilename;  /**< filename of the copy-on-write file, if any */
 223         int difffile;        /**< filedescriptor of copyonwrite file. @todo
 224                                shouldn't this be an array too? (cfr export) Or
 225                                make -m and -c mutually exclusive */
 226         u32 difffilelen;     /**< number of pages in difffile */
 227         u32 *difmap;         /**< see comment on the global difmap for this one */
 228         gboolean modern;     /**< client was negotiated using modern negotiation protocol */
 229 } CLIENT;
 230
 231 /**
 232  * Type of configuration file values
 233  **/
 234 typedef enum {
 235         PARAM_INT,              /**< This parameter is an integer */
 236         PARAM_STRING,           /**< This parameter is a string */
 237         PARAM_BOOL,             /**< This parameter is a boolean */
 238 } PARAM_TYPE;
 239
 240 /**
 241  * Configuration file values
 242  **/
 243 typedef struct {
 244         gchar *paramname;       /**< Name of the parameter, as it appears in
 245                                   the config file */
 246         gboolean required;      /**< Whether this is a required (as opposed to
 247                                   optional) parameter */
 248         PARAM_TYPE ptype;       /**< Type of the parameter. */
 249         gpointer target;        /**< Pointer to where the data of this
 250                                   parameter should be written. If ptype is
 251                                   PARAM_BOOL, the data is or'ed rather than
 252                                   overwritten. */
 253         gint flagval;           /**< Flag mask for this parameter in case ptype
 254                                   is PARAM_BOOL. */
 255 } PARAM;
 256
 257 /**
 258  * Check whether a client is allowed to connect. Works with an authorization
 259  * file which contains one line per machine, no wildcards.
 260  *
 261  * @param opts The client who's trying to connect.
 262  * @return 0 - authorization refused, 1 - OK
 263  **/
 264 int authorized_client(CLIENT *opts) {
 265         const char *ERRMSG="Invalid entry '%s' in authfile '%s', so, refusing all connections.";
 266         FILE *f ;
 267         char line[LINELEN];
 268         char *tmp;
 269         struct in_addr addr;
 270         struct in_addr client;
 271         struct in_addr cltemp;
 272         int len;
 273
 274         if ((f=fopen(opts->server->authname,"r"))==NULL) {
 275                 msg4(LOG_INFO,"Can't open authorization file %s (%s).",
 276                      opts->server->authname,strerror(errno)) ;
 277                 return 1 ;
 278         }
 279
 280         inet_aton(opts->clientname, &client);
 281         while (fgets(line,LINELEN,f)!=NULL) {
 282                 if((tmp=index(line, '/'))) {
 283                         if(strlen(line)<=tmp-line) {
 284                                 msg4(LOG_CRIT, ERRMSG, line, opts->server->authname);
 285                                 return 0;
 286                         }
 287                         *(tmp++)=0;
 288                         if(!inet_aton(line,&addr)) {
 289                                 msg4(LOG_CRIT, ERRMSG, line, opts->server->authname);
 290                                 return 0;
 291                         }
 292                         len=strtol(tmp, NULL, 0);
 293                         addr.s_addr>>=32-len;
 294                         addr.s_addr<<=32-len;
 295                         memcpy(&cltemp,&client,sizeof(client));
 296                         cltemp.s_addr>>=32-len;
 297                         cltemp.s_addr<<=32-len;
 298                         if(addr.s_addr == cltemp.s_addr) {
 299                                 return 1;
 300                         }
 301                 }
 302                 if (strncmp(line,opts->clientname,strlen(opts->clientname))==0) {
 303                         fclose(f);
 304                         return 1;
 305                 }
 306         }
 307         fclose(f);
 308         return 0;
 309 }
 310
 311 /**
 312  * Read data from a file descriptor into a buffer
 313  *
 314  * @param f a file descriptor
 315  * @param buf a buffer
 316  * @param len the number of bytes to be read
 317  **/
 318 inline void readit(int f, void *buf, size_t len) {
 319         ssize_t res;
 320         while (len > 0) {
 321                 DEBUG("*");
 322                 if ((res = read(f, buf, len)) <= 0) {
 323                         if(errno != EAGAIN) {
 324                                 err("Read failed: %m");
 325                         }
 326                 } else {
 327                         len -= res;
 328                         buf += res;
 329                 }
 330         }
 331 }
 332
 333 /**
 334  * Write data from a buffer into a filedescriptor
 335  *
 336  * @param f a file descriptor
 337  * @param buf a buffer containing data
 338  * @param len the number of bytes to be written
 339  **/
 340 inline void writeit(int f, void *buf, size_t len) {
 341         ssize_t res;
 342         while (len > 0) {
 343                 DEBUG("+");
 344                 if ((res = write(f, buf, len)) <= 0)
 345                         err("Send failed: %m");
 346                 len -= res;
 347                 buf += res;
 348         }
 349 }
 350
 351 /**
 352  * Print out a message about how to use nbd-server. Split out to a separate
 353  * function so that we can call it from multiple places
 354  */
 355 void usage() {
 356         printf("This is nbd-server version " VERSION "\n");
 357         printf("Usage: [ip:|ip6@]port file_to_export [size][kKmM] [-l authorize_file] [-r] [-m] [-c] [-C configuration file] [-p PID file name] [-o section name]\n"
 358                "\t-r|--read-only\t\tread only\n"
 359                "\t-m|--multi-file\t\tmultiple file\n"
 360                "\t-c|--copy-on-write\tcopy on write\n"
 361                "\t-C|--config-file\tspecify an alternate configuration file\n"
 362                "\t-l|--authorize-file\tfile with list of hosts that are allowed to\n\t\t\t\tconnect.\n"
 363                "\t-p|--pid-file\t\tspecify a filename to write our PID to\n"
 364                "\t-o|--output-config\toutput a config file section for what you\n\t\t\t\tspecified on the command line, with the\n\t\t\t\tspecified section name\n\n"
 365                "\tif port is set to 0, stdin is used (for running from inetd)\n"
 366                "\tif file_to_export contains '%%s', it is substituted with the IP\n"
 367                "\t\taddress of the machine trying to connect\n"
 368                "\tif ip is set, it contains the local IP address on which we're listening.\n\tif not, the server will listen on all local IP addresses\n");
 369         printf("Using configuration file %s\n", CFILE);
 370 }
 371
 372 /* Dumps a config file section of the given SERVER*, and exits. */
 373 void dump_section(SERVER* serve, gchar* section_header) {
 374         printf("[%s]\n", section_header);
 375         printf("\texportname = %s\n", serve->exportname);
 376         printf("\tlistenaddr = %s\n", serve->listenaddr);
 377         printf("\tport = %d\n", serve->port);
 378         if(serve->flags & F_READONLY) {
 379                 printf("\treadonly = true\n");
 380         }
 381         if(serve->flags & F_MULTIFILE) {
 382                 printf("\tmultifile = true\n");
 383         }
 384         if(serve->flags & F_COPYONWRITE) {
 385                 printf("\tcopyonwrite = true\n");
 386         }
 387         if(serve->expected_size) {
 388                 printf("\tfilesize = %lld\n", (long long int)serve->expected_size);
 389         }
 390         if(serve->authname) {
 391                 printf("\tauthfile = %s\n", serve->authname);
 392         }
 393         exit(EXIT_SUCCESS);
 394 }
 395
 396 /**
 397  * Parse the command line.
 398  *
 399  * @param argc the argc argument to main()
 400  * @param argv the argv argument to main()
 401  **/
 402 SERVER* cmdline(int argc, char *argv[]) {
 403         int i=0;
 404         int nonspecial=0;
 405         int c;
 406         struct option long_options[] = {
 407                 {"read-only", no_argument, NULL, 'r'},
 408                 {"multi-file", no_argument, NULL, 'm'},
 409                 {"copy-on-write", no_argument, NULL, 'c'},
 410                 {"authorize-file", required_argument, NULL, 'l'},
 411                 {"config-file", required_argument, NULL, 'C'},
 412                 {"pid-file", required_argument, NULL, 'p'},
 413                 {"output-config", required_argument, NULL, 'o'},
 414                 {0,0,0,0}
 415         };
 416         SERVER *serve;
 417         off_t es;
 418         size_t last;
 419         char suffix;
 420         gboolean do_output=FALSE;
 421         gchar* section_header="";
 422         gchar** addr_port;
 423
 424         if(argc==1) {
 425                 return NULL;
 426         }
 427         serve=g_new0(SERVER, 1);
 428         serve->authname = g_strdup(default_authname);
 429         serve->virtstyle=VIRT_IPLIT;
 430         while((c=getopt_long(argc, argv, "-C:cl:mo:rp:", long_options, &i))>=0) {
 431                 switch (c) {
 432                 case 1:
 433                         /* non-option argument */
 434                         switch(nonspecial++) {
 435                         case 0:
 436                                 if(strchr(optarg, ':') == strrchr(optarg, ':')) {
 437                                         addr_port=g_strsplit(optarg, ":", 2);
 438
 439                                         /* Check for "@" - maybe user using this separator
 440                                                  for IPv4 address */
 441                                         if(!addr_port[1]) {
 442                                                 g_strfreev(addr_port);
 443                                                 addr_port=g_strsplit(optarg, "@", 2);
 444                                         }
 445                                 } else {
 446                                         addr_port=g_strsplit(optarg, "@", 2);
 447                                 }
 448
 449                                 if(addr_port[1]) {
 450                                         serve->port=strtol(addr_port[1], NULL, 0);
 451                                         serve->listenaddr=g_strdup(addr_port[0]);
 452                                 } else {
 453                                         serve->listenaddr=NULL;
 454                                         serve->port=strtol(addr_port[0], NULL, 0);
 455                                 }
 456                                 g_strfreev(addr_port);
 457                                 break;
 458                         case 1:
 459                                 serve->exportname = g_strdup(optarg);
 460                                 if(serve->exportname[0] != '/') {
 461                                         fprintf(stderr, "E: The to be exported file needs to be an absolute filename!\n");
 462                                         exit(EXIT_FAILURE);
 463                                 }
 464                                 break;
 465                         case 2:
 466                                 last=strlen(optarg)-1;
 467                                 suffix=optarg[last];
 468                                 if (suffix == 'k' || suffix == 'K' ||
 469                                     suffix == 'm' || suffix == 'M')
 470                                         optarg[last] = '\0';
 471                                 es = (off_t)atoll(optarg);
 472                                 switch (suffix) {
 473                                         case 'm':
 474                                         case 'M':  es <<= 10;
 475                                         case 'k':
 476                                         case 'K':  es <<= 10;
 477                                         default :  break;
 478                                 }
 479                                 serve->expected_size = es;
 480                                 break;
 481                         }
 482                         break;
 483                 case 'r':
 484                         serve->flags |= F_READONLY;
 485                         break;
 486                 case 'm':
 487                         serve->flags |= F_MULTIFILE;
 488                         break;
 489                 case 'o':
 490                         do_output = TRUE;
 491                         section_header = g_strdup(optarg);
 492                         break;
 493                 case 'p':
 494                         strncpy(pidftemplate, optarg, 256);
 495                         break;
 496                 case 'c':
 497                         serve->flags |=F_COPYONWRITE;
 498                         break;
 499                 case 'C':
 500                         g_free(config_file_pos);
 501                         config_file_pos=g_strdup(optarg);
 502                         break;
 503                 case 'l':
 504                         g_free(serve->authname);
 505                         serve->authname=g_strdup(optarg);
 506                         break;
 507                 default:
 508                         usage();
 509                         exit(EXIT_FAILURE);
 510                         break;
 511                 }
 512         }
 513         /* What's left: the port to export, the name of the to be exported
 514          * file, and, optionally, the size of the file, in that order. */
 515         if(nonspecial<2) {
 516                 g_free(serve);
 517                 serve=NULL;
 518         }
 519         if(do_output) {
 520                 if(!serve) {
 521                         g_critical("Need a complete configuration on the command line to output a config file section!");
 522                         exit(EXIT_FAILURE);
 523                 }
 524                 dump_section(serve, section_header);
 525         }
 526         return serve;
 527 }
 528
 529 /**
 530  * Error codes for config file parsing
 531  **/
 532 typedef enum {
 533         CFILE_NOTFOUND,         /**< The configuration file is not found */
 534         CFILE_MISSING_GENERIC,  /**< The (required) group "generic" is missing */
 535         CFILE_KEY_MISSING,      /**< A (required) key is missing */
 536         CFILE_VALUE_INVALID,    /**< A value is syntactically invalid */
 537         CFILE_VALUE_UNSUPPORTED,/**< A value is not supported in this build */
 538         CFILE_PROGERR,          /**< Programmer error */
 539         CFILE_NO_EXPORTS,       /**< A config file was specified that does not
 540                                      define any exports */
 541         CFILE_INCORRECT_PORT,   /**< The reserved port was specified for an
 542                                      old-style export. */
 543 } CFILE_ERRORS;
 544
 545 /**
 546  * Remove a SERVER from memory. Used from the hash table
 547  **/
 548 void remove_server(gpointer s) {
 549         SERVER *server;
 550
 551         server=(SERVER*)s;
 552         g_free(server->exportname);
 553         if(server->authname)
 554                 g_free(server->authname);
 555         if(server->listenaddr)
 556                 g_free(server->listenaddr);
 557         if(server->prerun)
 558                 g_free(server->prerun);
 559         if(server->postrun)
 560                 g_free(server->postrun);
 561         g_free(server);
 562 }
 563
 564 /**
 565  * duplicate server
 566  * @param s the old server we want to duplicate
 567  * @return new duplicated server
 568  **/
 569 SERVER* dup_serve(SERVER *s) {
 570         SERVER *serve = NULL;
 571
 572         serve=g_new0(SERVER, 1);
 573         if(serve == NULL)
 574                 return NULL;
 575
 576         if(s->exportname)
 577                 serve->exportname = g_strdup(s->exportname);
 578
 579         serve->expected_size = s->expected_size;
 580
 581         if(s->listenaddr)
 582                 serve->listenaddr = g_strdup(s->listenaddr);
 583
 584         serve->port = s->port;
 585
 586         if(s->authname)
 587                 serve->authname = strdup(s->authname);
 588
 589         serve->flags = s->flags;
 590         serve->socket = serve->socket;
 591         serve->socket_family = serve->socket_family;
 592         serve->cidrlen = s->cidrlen;
 593
 594         if(s->prerun)
 595                 serve->prerun = g_strdup(s->prerun);
 596
 597         if(s->postrun)
 598                 serve->postrun = g_strdup(s->postrun);
 599
 600         if(s->servename)
 601                 serve->servename = g_strdup(s->servename);
 602
 603         return serve;
 604 }
 605
 606 /**
 607  * append new server to array
 608  * @param s server
 609  * @param a server array
 610  * @return 0 success, -1 error
 611  */
 612 int append_serve(SERVER *s, GArray *a) {
 613         SERVER *ns = NULL;
 614         struct addrinfo hints;
 615         struct addrinfo *ai = NULL;
 616         struct addrinfo *rp = NULL;
 617         char   host[NI_MAXHOST];
 618         gchar  *port = NULL;
 619         int e;
 620         int ret;
 621
 622         if(!s) {
 623                 err("Invalid parsing server");
 624                 return -1;
 625         }
 626
 627         port = g_strdup_printf("%d", s->port);
 628
 629         memset(&hints,'\0',sizeof(hints));
 630         hints.ai_family = AF_UNSPEC;
 631         hints.ai_socktype = SOCK_STREAM;
 632         hints.ai_flags = AI_ADDRCONFIG | AI_PASSIVE;
 633         hints.ai_protocol = IPPROTO_TCP;
 634
 635         e = getaddrinfo(s->listenaddr, port, &hints, &ai);
 636
 637         if (port)
 638                 g_free(port);
 639
 640         if(e == 0) {
 641                 for (rp = ai; rp != NULL; rp = rp->ai_next) {
 642                         e = getnameinfo(rp->ai_addr, rp->ai_addrlen, host, sizeof(host), NULL, 0, NI_NUMERICHOST);
 643
 644                         if (e != 0) { // error
 645                                 fprintf(stderr, "getnameinfo: %s\n", gai_strerror(e));
 646                                 continue;
 647                         }
 648
 649                         // duplicate server and set listenaddr to resolved IP address
 650                         ns = dup_serve (s);
 651                         if (ns) {
 652                                 ns->listenaddr = g_strdup(host);
 653                                 ns->socket_family = rp->ai_family;
 654                                 g_array_append_val(a, *ns);
 655                                 free(ns);
 656                                 ns = NULL;
 657                         }
 658                 }
 659
 660                 ret = 0;
 661         } else {
 662                 fprintf(stderr, "getaddrinfo failed on listen host/address: %s (%s)\n", s->listenaddr ? s->listenaddr : "any", gai_strerror(e));
 663                 ret = -1;
 664         }
 665
 666         if (ai)
 667                 freeaddrinfo(ai);
 668
 669         return ret;
 670 }
 671
 672 /**
 673  * Parse the config file.
 674  *
 675  * @param f the name of the config file
 676  * @param e a GError. @see CFILE_ERRORS for what error values this function can
 677  *      return.
 678  * @return a Array of SERVER* pointers, If the config file is empty or does not
 679  *      exist, returns an empty GHashTable; if the config file contains an
 680  *      error, returns NULL, and e is set appropriately
 681  **/
 682 GArray* parse_cfile(gchar* f, GError** e) {
 683         const char* DEFAULT_ERROR = "Could not parse %s in group %s: %s";
 684         const char* MISSING_REQUIRED_ERROR = "Could not find required value %s in group %s: %s";
 685         SERVER s;
 686         gchar *virtstyle=NULL;
 687         PARAM lp[] = {
 688                 { "exportname", TRUE,   PARAM_STRING,   NULL, 0 },
 689                 { "port",       TRUE,   PARAM_INT,      NULL, 0 },
 690                 { "authfile",   FALSE,  PARAM_STRING,   NULL, 0 },
 691                 { "filesize",   FALSE,  PARAM_INT,      NULL, 0 },
 692                 { "virtstyle",  FALSE,  PARAM_STRING,   NULL, 0 },
 693                 { "prerun",     FALSE,  PARAM_STRING,   NULL, 0 },
 694                 { "postrun",    FALSE,  PARAM_STRING,   NULL, 0 },
 695                 { "readonly",   FALSE,  PARAM_BOOL,     NULL, F_READONLY },
 696                 { "multifile",  FALSE,  PARAM_BOOL,     NULL, F_MULTIFILE },
 697                 { "copyonwrite", FALSE, PARAM_BOOL,     NULL, F_COPYONWRITE },
 698                 { "sparse_cow", FALSE,  PARAM_BOOL,     NULL, F_SPARSE },
 699                 { "sdp",        FALSE,  PARAM_BOOL,     NULL, F_SDP },
 700                 { "sync",       FALSE,  PARAM_BOOL,     NULL, F_SYNC },
 701                 { "listenaddr", FALSE,  PARAM_STRING,   NULL, 0 },
 702         };
 703         const int lp_size=sizeof(lp)/sizeof(PARAM);
 704         PARAM gp[] = {
 705                 { "user",       FALSE, PARAM_STRING,    &runuser,       0 },
 706                 { "group",      FALSE, PARAM_STRING,    &rungroup,      0 },
 707                 { "oldstyle",   FALSE, PARAM_BOOL,      &do_oldstyle,   1 },
 708                 { "listenaddr", FALSE, PARAM_STRING,    &modern_listen, 0 },
 709         };
 710         PARAM* p=gp;
 711         int p_size=sizeof(gp)/sizeof(PARAM);
 712         GKeyFile *cfile;
 713         GError *err = NULL;
 714         const char *err_msg=NULL;
 715         GQuark errdomain;
 716         GArray *retval=NULL;
 717         gchar **groups;
 718         gboolean value;
 719         gchar* startgroup;
 720         gint i;
 721         gint j;
 722
 723         errdomain = g_quark_from_string("parse_cfile");
 724         cfile = g_key_file_new();
 725         retval = g_array_new(FALSE, TRUE, sizeof(SERVER));
 726         if(!g_key_file_load_from_file(cfile, f, G_KEY_FILE_KEEP_COMMENTS |
 727                         G_KEY_FILE_KEEP_TRANSLATIONS, &err)) {
 728                 g_set_error(e, errdomain, CFILE_NOTFOUND, "Could not open config file %s.", f);
 729                 g_key_file_free(cfile);
 730                 return retval;
 731         }
 732         startgroup = g_key_file_get_start_group(cfile);
 733         if(!startgroup || strcmp(startgroup, "generic")) {
 734                 g_set_error(e, errdomain, CFILE_MISSING_GENERIC, "Config file does not contain the [generic] group!");
 735                 g_key_file_free(cfile);
 736                 return NULL;
 737         }
 738         groups = g_key_file_get_groups(cfile, NULL);
 739         for(i=0;groups[i];i++) {
 740                 memset(&s, '\0', sizeof(SERVER));
 741                 lp[0].target=&(s.exportname);
 742                 lp[1].target=&(s.port);
 743                 lp[2].target=&(s.authname);
 744                 lp[3].target=&(s.expected_size);
 745                 lp[4].target=&(virtstyle);
 746                 lp[5].target=&(s.prerun);
 747                 lp[6].target=&(s.postrun);
 748                 lp[7].target=lp[8].target=lp[9].target=
 749                                 lp[10].target=lp[11].target=
 750                                 lp[12].target=&(s.flags);
 751                 lp[13].target=&(s.listenaddr);
 752
 753                 /* After the [generic] group, start parsing exports */
 754                 if(i==1) {
 755                         p=lp;
 756                         p_size=lp_size;
 757                 }
 758                 for(j=0;j<p_size;j++) {
 759                         g_assert(p[j].target != NULL);
 760                         g_assert(p[j].ptype==PARAM_INT||p[j].ptype==PARAM_STRING||p[j].ptype==PARAM_BOOL);
 761                         switch(p[j].ptype) {
 762                                 case PARAM_INT:
 763                                         *((gint*)p[j].target) =
 764                                                 g_key_file_get_integer(cfile,
 765                                                                 groups[i],
 766                                                                 p[j].paramname,
 767                                                                 &err);
 768                                         break;
 769                                 case PARAM_STRING:
 770                                         *((gchar**)p[j].target) =
 771                                                 g_key_file_get_string(cfile,
 772                                                                 groups[i],
 773                                                                 p[j].paramname,
 774                                                                 &err);
 775                                         break;
 776                                 case PARAM_BOOL:
 777                                         value = g_key_file_get_boolean(cfile,
 778                                                         groups[i],
 779                                                         p[j].paramname, &err);
 780                                         if(!err) {
 781                                                 if(value) {
 782                                                         *((gint*)p[j].target) |= p[j].flagval;
 783                                                 } else {
 784                                                         *((gint*)p[j].target) &= ~(p[j].flagval);
 785                                                 }
 786                                         }
 787                                         break;
 788                         }
 789                         if(!strcmp(p[j].paramname, "port") && !strcmp(p[j].target, NBD_DEFAULT_PORT)) {
 790                                 g_set_error(e, errdomain, CFILE_INCORRECT_PORT, "Config file specifies default port for oldstyle export");
 791                                 g_key_file_free(cfile);
 792                                 return NULL;
 793                         }
 794                         if(err) {
 795                                 if(err->code == G_KEY_FILE_ERROR_KEY_NOT_FOUND) {
 796                                         if(!p[j].required) {
 797                                                 /* Ignore not-found error for optional values */
 798                                                 g_clear_error(&err);
 799                                                 continue;
 800                                         } else {
 801                                                 err_msg = MISSING_REQUIRED_ERROR;
 802                                         }
 803                                 } else {
 804                                         err_msg = DEFAULT_ERROR;
 805                                 }
 806                                 g_set_error(e, errdomain, CFILE_VALUE_INVALID, err_msg, p[j].paramname, groups[i], err->message);
 807                                 g_array_free(retval, TRUE);
 808                                 g_error_free(err);
 809                                 g_key_file_free(cfile);
 810                                 return NULL;
 811                         }
 812                 }
 813                 if(virtstyle) {
 814                         if(!strncmp(virtstyle, "none", 4)) {
 815                                 s.virtstyle=VIRT_NONE;
 816                         } else if(!strncmp(virtstyle, "ipliteral", 9)) {
 817                                 s.virtstyle=VIRT_IPLIT;
 818                         } else if(!strncmp(virtstyle, "iphash", 6)) {
 819                                 s.virtstyle=VIRT_IPHASH;
 820                         } else if(!strncmp(virtstyle, "cidrhash", 8)) {
 821                                 s.virtstyle=VIRT_CIDR;
 822                                 if(strlen(virtstyle)<10) {
 823                                         g_set_error(e, errdomain, CFILE_VALUE_INVALID, "Invalid value %s for parameter virtstyle in group %s: missing length", virtstyle, groups[i]);
 824                                         g_array_free(retval, TRUE);
 825                                         g_key_file_free(cfile);
 826                                         return NULL;
 827                                 }
 828                                 s.cidrlen=strtol(virtstyle+8, NULL, 0);
 829                         } else {
 830                                 g_set_error(e, errdomain, CFILE_VALUE_INVALID, "Invalid value %s for parameter virtstyle in group %s", virtstyle, groups[i]);
 831                                 g_array_free(retval, TRUE);
 832                                 g_key_file_free(cfile);
 833                                 return NULL;
 834                         }
 835                 } else {
 836                         s.virtstyle=VIRT_IPLIT;
 837                 }
 838                 /* Don't need to free this, it's not our string */
 839                 virtstyle=NULL;
 840                 /* Don't append values for the [generic] group */
 841                 if(i>0) {
 842                         s.socket_family = AF_UNSPEC;
 843                         s.servename = groups[i];
 844
 845                         append_serve(&s, retval);
 846                 } else {
 847                         if(!do_oldstyle) {
 848                                 lp[1].required = 0;
 849                         }
 850                 }
 851 #ifndef WITH_SDP
 852                 if(s.flags & F_SDP) {
 853                         g_set_error(e, errdomain, CFILE_VALUE_UNSUPPORTED, "This nbd-server was built without support for SDP, yet group %s uses it", groups[i]);
 854                         g_array_free(retval, TRUE);
 855                         g_key_file_free(cfile);
 856                         return NULL;
 857                 }
 858 #endif
 859         }
 860         if(i==1) {
 861                 g_set_error(e, errdomain, CFILE_NO_EXPORTS, "The config file does not specify any exports");
 862         }
 863         g_key_file_free(cfile);
 864         return retval;
 865 }
 866
 867 /**
 868  * Signal handler for SIGCHLD
 869  * @param s the signal we're handling (must be SIGCHLD, or something
 870  * is severely wrong)
 871  **/
 872 void sigchld_handler(int s) {
 873         int status;
 874         int* i;
 875         pid_t pid;
 876
 877         while((pid=waitpid(-1, &status, WNOHANG)) > 0) {
 878                 if(WIFEXITED(status)) {
 879                         msg3(LOG_INFO, "Child exited with %d", WEXITSTATUS(status));
 880                 }
 881                 i=g_hash_table_lookup(children, &pid);
 882                 if(!i) {
 883                         msg3(LOG_INFO, "SIGCHLD received for an unknown child with PID %ld", (long)pid);
 884                 } else {
 885                         DEBUG2("Removing %d from the list of children", pid);
 886                         g_hash_table_remove(children, &pid);
 887                 }
 888         }
 889 }
 890
 891 /**
 892  * Kill a child. Called from sigterm_handler::g_hash_table_foreach.
 893  *
 894  * @param key the key
 895  * @param value the value corresponding to the above key
 896  * @param user_data a pointer which we always set to 1, so that we know what
 897  * will happen next.
 898  **/
 899 void killchild(gpointer key, gpointer value, gpointer user_data) {
 900         pid_t *pid=value;
 901         int *parent=user_data;
 902
 903         kill(*pid, SIGTERM);
 904         *parent=1;
 905 }
 906
 907 /**
 908  * Handle SIGTERM and dispatch it to our children
 909  * @param s the signal we're handling (must be SIGTERM, or something
 910  * is severely wrong).
 911  **/
 912 void sigterm_handler(int s) {
 913         int parent=0;
 914
 915         g_hash_table_foreach(children, killchild, &parent);
 916
 917         if(parent) {
 918                 unlink(pidfname);
 919         }
 920
 921         exit(EXIT_SUCCESS);
 922 }
 923
 924 /**
 925  * Detect the size of a file.
 926  *
 927  * @param fhandle An open filedescriptor
 928  * @return the size of the file, or OFFT_MAX if detection was
 929  * impossible.
 930  **/
 931 off_t size_autodetect(int fhandle) {
 932         off_t es;
 933         u64 bytes;
 934         struct stat stat_buf;
 935         int error;
 936
 937 #ifdef HAVE_SYS_MOUNT_H
 938 #ifdef HAVE_SYS_IOCTL_H
 939 #ifdef BLKGETSIZE64
 940         DEBUG("looking for export size with ioctl BLKGETSIZE64\n");
 941         if (!ioctl(fhandle, BLKGETSIZE64, &bytes) && bytes) {
 942                 return (off_t)bytes;
 943         }
 944 #endif /* BLKGETSIZE64 */
 945 #endif /* HAVE_SYS_IOCTL_H */
 946 #endif /* HAVE_SYS_MOUNT_H */
 947
 948         DEBUG("looking for fhandle size with fstat\n");
 949         stat_buf.st_size = 0;
 950         error = fstat(fhandle, &stat_buf);
 951         if (!error) {
 952                 if(stat_buf.st_size > 0)
 953                         return (off_t)stat_buf.st_size;
 954         } else {
 955                 err("fstat failed: %m");
 956         }
 957
 958         DEBUG("looking for fhandle size with lseek SEEK_END\n");
 959         es = lseek(fhandle, (off_t)0, SEEK_END);
 960         if (es > ((off_t)0)) {
 961                 return es;
 962         } else {
 963                 DEBUG2("lseek failed: %d", errno==EBADF?1:(errno==ESPIPE?2:(errno==EINVAL?3:4)));
 964         }
 965
 966         err("Could not find size of exported block device: %m");
 967         return OFFT_MAX;
 968 }
 969
 970 /**
 971  * Get the file handle and offset, given an export offset.
 972  *
 973  * @param export An array of export files
 974  * @param a The offset to get corresponding file/offset for
 975  * @param fhandle [out] File descriptor
 976  * @param foffset [out] Offset into fhandle
 977  * @param maxbytes [out] Tells how many bytes can be read/written
 978  * from fhandle starting at foffset (0 if there is no limit)
 979  * @return 0 on success, -1 on failure
 980  **/
 981 int get_filepos(GArray* export, off_t a, int* fhandle, off_t* foffset, size_t* maxbytes ) {
 982         /* Negative offset not allowed */
 983         if(a < 0)
 984                 return -1;
 985
 986         /* Binary search for last file with starting offset <= a */
 987         FILE_INFO fi;
 988         int start = 0;
 989         int end = export->len - 1;
 990         while( start <= end ) {
 991                 int mid = (start + end) / 2;
 992                 fi = g_array_index(export, FILE_INFO, mid);
 993                 if( fi.startoff < a ) {
 994                         start = mid + 1;
 995                 } else if( fi.startoff > a ) {
 996                         end = mid - 1;
 997                 } else {
 998                         start = end = mid;
 999                         break;
1000                 }
1001         }
1002
1003         /* end should never go negative, since first startoff is 0 and a >= 0 */
1004         g_assert(end >= 0);
1005
1006         fi = g_array_index(export, FILE_INFO, end);
1007         *fhandle = fi.fhandle;
1008         *foffset = a - fi.startoff;
1009         *maxbytes = 0;
1010         if( end+1 < export->len ) {
1011                 FILE_INFO fi_next = g_array_index(export, FILE_INFO, end+1);
1012                 *maxbytes = fi_next.startoff - a;
1013         }
1014
1015         return 0;
1016 }
1017
1018 /**
1019  * seek to a position in a file, with error handling.
1020  * @param handle a filedescriptor
1021  * @param a position to seek to
1022  * @todo get rid of this; lastpoint is a global variable right now, but it
1023  * shouldn't be. If we pass it on as a parameter, that makes things a *lot*
1024  * easier.
1025  **/
1026 void myseek(int handle,off_t a) {
1027         if (lseek(handle, a, SEEK_SET) < 0) {
1028                 err("Can not seek locally!\n");
1029         }
1030 }
1031
1032 /**
1033  * Write an amount of bytes at a given offset to the right file. This
1034  * abstracts the write-side of the multiple file option.
1035  *
1036  * @param a The offset where the write should start
1037  * @param buf The buffer to write from
1038  * @param len The length of buf
1039  * @param client The client we're serving for
1040  * @return The number of bytes actually written, or -1 in case of an error
1041  **/
1042 ssize_t rawexpwrite(off_t a, char *buf, size_t len, CLIENT *client) {
1043         int fhandle;
1044         off_t foffset;
1045         size_t maxbytes;
1046         ssize_t retval;
1047
1048         if(get_filepos(client->export, a, &fhandle, &foffset, &maxbytes))
1049                 return -1;
1050         if(maxbytes && len > maxbytes)
1051                 len = maxbytes;
1052
1053         DEBUG4("(WRITE to fd %d offset %llu len %u), ", fhandle, foffset, len);
1054
1055         myseek(fhandle, foffset);
1056         retval = write(fhandle, buf, len);
1057         if(client->server->flags & F_SYNC) {
1058                 fsync(fhandle);
1059         }
1060         return retval;
1061 }
1062
1063 /**
1064  * Call rawexpwrite repeatedly until all data has been written.
1065  * @return 0 on success, nonzero on failure
1066  **/
1067 int rawexpwrite_fully(off_t a, char *buf, size_t len, CLIENT *client) {
1068         ssize_t ret=0;
1069
1070         while(len > 0 && (ret=rawexpwrite(a, buf, len, client)) > 0 ) {
1071                 a += ret;
1072                 buf += ret;
1073                 len -= ret;
1074         }
1075         return (ret < 0 || len != 0);
1076 }
1077
1078 /**
1079  * Read an amount of bytes at a given offset from the right file. This
1080  * abstracts the read-side of the multiple files option.
1081  *
1082  * @param a The offset where the read should start
1083  * @param buf A buffer to read into
1084  * @param len The size of buf
1085  * @param client The client we're serving for
1086  * @return The number of bytes actually read, or -1 in case of an
1087  * error.
1088  **/
1089 ssize_t rawexpread(off_t a, char *buf, size_t len, CLIENT *client) {
1090         int fhandle;
1091         off_t foffset;
1092         size_t maxbytes;
1093
1094         if(get_filepos(client->export, a, &fhandle, &foffset, &maxbytes))
1095                 return -1;
1096         if(maxbytes && len > maxbytes)
1097                 len = maxbytes;
1098
1099         DEBUG4("(READ from fd %d offset %llu len %u), ", fhandle, foffset, len);
1100
1101         myseek(fhandle, foffset);
1102         return read(fhandle, buf, len);
1103 }
1104
1105 /**
1106  * Call rawexpread repeatedly until all data has been read.
1107  * @return 0 on success, nonzero on failure
1108  **/
1109 int rawexpread_fully(off_t a, char *buf, size_t len, CLIENT *client) {
1110         ssize_t ret=0;
1111
1112         while(len > 0 && (ret=rawexpread(a, buf, len, client)) > 0 ) {
1113                 a += ret;
1114                 buf += ret;
1115                 len -= ret;
1116         }
1117         return (ret < 0 || len != 0);
1118 }
1119
1120 /**
1121  * Read an amount of bytes at a given offset from the right file. This
1122  * abstracts the read-side of the copyonwrite stuff, and calls
1123  * rawexpread() with the right parameters to do the actual work.
1124  * @param a The offset where the read should start
1125  * @param buf A buffer to read into
1126  * @param len The size of buf
1127  * @param client The client we're going to read for
1128  * @return 0 on success, nonzero on failure
1129  **/
1130 int expread(off_t a, char *buf, size_t len, CLIENT *client) {
1131         off_t rdlen, offset;
1132         off_t mapcnt, mapl, maph, pagestart;
1133
1134         if (!(client->server->flags & F_COPYONWRITE))
1135                 return(rawexpread_fully(a, buf, len, client));
1136         DEBUG3("Asked to read %d bytes at %llu.\n", len, (unsigned long long)a);
1137
1138         mapl=a/DIFFPAGESIZE; maph=(a+len-1)/DIFFPAGESIZE;
1139
1140         for (mapcnt=mapl;mapcnt<=maph;mapcnt++) {
1141                 pagestart=mapcnt*DIFFPAGESIZE;
1142                 offset=a-pagestart;
1143                 rdlen=(0<DIFFPAGESIZE-offset && len<(size_t)(DIFFPAGESIZE-offset)) ?
1144                         len : (size_t)DIFFPAGESIZE-offset;
1145                 if (client->difmap[mapcnt]!=(u32)(-1)) { /* the block is already there */
1146                         DEBUG3("Page %llu is at %lu\n", (unsigned long long)mapcnt,
1147                                (unsigned long)(client->difmap[mapcnt]));
1148                         myseek(client->difffile, client->difmap[mapcnt]*DIFFPAGESIZE+offset);
1149                         if (read(client->difffile, buf, rdlen) != rdlen) return -1;
1150                 } else { /* the block is not there */
1151                         DEBUG2("Page %llu is not here, we read the original one\n",
1152                                (unsigned long long)mapcnt);
1153                         if(rawexpread_fully(a, buf, rdlen, client)) return -1;
1154                 }
1155                 len-=rdlen; a+=rdlen; buf+=rdlen;
1156         }
1157         return 0;
1158 }
1159
1160 /**
1161  * Write an amount of bytes at a given offset to the right file. This
1162  * abstracts the write-side of the copyonwrite option, and calls
1163  * rawexpwrite() with the right parameters to do the actual work.
1164  *
1165  * @param a The offset where the write should start
1166  * @param buf The buffer to write from
1167  * @param len The length of buf
1168  * @param client The client we're going to write for.
1169  * @return 0 on success, nonzero on failure
1170  **/
1171 int expwrite(off_t a, char *buf, size_t len, CLIENT *client) {
1172         char pagebuf[DIFFPAGESIZE];
1173         off_t mapcnt,mapl,maph;
1174         off_t wrlen,rdlen;
1175         off_t pagestart;
1176         off_t offset;
1177
1178         if (!(client->server->flags & F_COPYONWRITE))
1179                 return(rawexpwrite_fully(a, buf, len, client));
1180         DEBUG3("Asked to write %d bytes at %llu.\n", len, (unsigned long long)a);
1181
1182         mapl=a/DIFFPAGESIZE ; maph=(a+len-1)/DIFFPAGESIZE ;
1183
1184         for (mapcnt=mapl;mapcnt<=maph;mapcnt++) {
1185                 pagestart=mapcnt*DIFFPAGESIZE ;
1186                 offset=a-pagestart ;
1187                 wrlen=(0<DIFFPAGESIZE-offset && len<(size_t)(DIFFPAGESIZE-offset)) ?
1188                         len : (size_t)DIFFPAGESIZE-offset;
1189
1190                 if (client->difmap[mapcnt]!=(u32)(-1)) { /* the block is already there */
1191                         DEBUG3("Page %llu is at %lu\n", (unsigned long long)mapcnt,
1192                                (unsigned long)(client->difmap[mapcnt])) ;
1193                         myseek(client->difffile,
1194                                         client->difmap[mapcnt]*DIFFPAGESIZE+offset);
1195                         if (write(client->difffile, buf, wrlen) != wrlen) return -1 ;
1196                 } else { /* the block is not there */
1197                         myseek(client->difffile,client->difffilelen*DIFFPAGESIZE) ;
1198                         client->difmap[mapcnt]=(client->server->flags&F_SPARSE)?mapcnt:client->difffilelen++;
1199                         DEBUG3("Page %llu is not here, we put it at %lu\n",
1200                                (unsigned long long)mapcnt,
1201                                (unsigned long)(client->difmap[mapcnt]));
1202                         rdlen=DIFFPAGESIZE ;
1203                         if (rawexpread_fully(pagestart, pagebuf, rdlen, client))
1204                                 return -1;
1205                         memcpy(pagebuf+offset,buf,wrlen) ;
1206                         if (write(client->difffile, pagebuf, DIFFPAGESIZE) !=
1207                                         DIFFPAGESIZE)
1208                                 return -1;
1209                 }
1210                 len-=wrlen ; a+=wrlen ; buf+=wrlen ;
1211         }
1212         return 0;
1213 }
1214
1215 /**
1216  * Do the initial negotiation.
1217  *
1218  * @param client The client we're negotiating with.
1219  **/
1220 CLIENT* negotiate(int net, CLIENT *client, GArray* servers) {
1221         char zeros[128];
1222         uint64_t size_host;
1223         uint32_t flags = NBD_FLAG_HAS_FLAGS;
1224         uint16_t smallflags = 0;
1225         uint64_t magic;
1226
1227         memset(zeros, '\0', sizeof(zeros));
1228         if(!client || !client->modern) {
1229                 if (write(net, INIT_PASSWD, 8) < 0) {
1230                         err_nonfatal("Negotiation failed: %m");
1231                         if(client)
1232                                 exit(EXIT_FAILURE);
1233                 }
1234                 magic = htonll(opts_magic);
1235                 if (write(net, &magic, sizeof(magic)) < 0) {
1236                         err_nonfatal("Negotiation failed: %m");
1237                         if(client)
1238                                 exit(EXIT_FAILURE);
1239                 }
1240         }
1241         if(!client) {
1242                 uint64_t reserved;
1243                 uint32_t opt;
1244                 uint64_t namelen;
1245                 char* name;
1246                 int i;
1247
1248                 if(!servers)
1249                         err("programmer error");
1250                 write(net, &smallflags, sizeof(uint16_t));
1251                 read(net, &reserved, sizeof(reserved));
1252                 read(net, &magic, sizeof(magic));
1253                 magic = ntohll(magic);
1254                 if(magic != cliserv_magic) {
1255                         close(net);
1256                         return NULL;
1257                 }
1258                 read(net, &opt, sizeof(opt));
1259                 opt = ntohl(opt);
1260                 if(opt != NBD_OPT_EXPORT_NAME) {
1261                         close(net);
1262                         return NULL;
1263                 }
1264                 read(net, &namelen, sizeof(namelen));
1265                 namelen = ntohll(namelen);
1266                 name = malloc(namelen+1);
1267                 name[namelen+1]=0;
1268                 read(net, name, namelen);
1269                 for(i=0; i<servers->len; i++) {
1270                         SERVER* serve = &(g_array_index(servers, SERVER, i));
1271                         if(!strcmp(serve->servename, name)) {
1272                                 CLIENT* client = g_new0(CLIENT, 1);
1273                                 client->server = serve;
1274                                 client->exportsize = OFFT_MAX;
1275                                 client->net = net;
1276                                 client->modern = TRUE;
1277                                 return client;
1278                         }
1279                 }
1280         }
1281         size_host = htonll((u64)(client->exportsize));
1282         if (write(net, &size_host, 8) < 0)
1283                 err("Negotiation failed: %m");
1284         if (client->server->flags & F_READONLY)
1285                 flags |= NBD_FLAG_READ_ONLY;
1286         if (!client->modern) {
1287         flags = htonl(flags);
1288                 if (write(client->net, &flags, 4) < 0)
1289                         err("Negotiation failed: %m");
1290         } else {
1291                 smallflags = (uint16_t)(flags & ~((uint16_t)0));
1292                 if (write(client->net, &smallflags, sizeof(smallflags)) < 0) {
1293                         err("Negotiation failed: %m");
1294                 }
1295         }
1296         if (write(client->net, zeros, 124) < 0)
1297                 err("Negotiation failed: %m");
1298         return NULL;
1299 }
1300
1301 /** sending macro. */
1302 #define SEND(net,reply) writeit( net, &reply, sizeof( reply ));
1303 /** error macro. */
1304 #define ERROR(client,reply,errcode) { reply.error = htonl(errcode); SEND(client->net,reply); reply.error = 0; }
1305 /**
1306  * Serve a file to a single client.
1307  *
1308  * @todo This beast needs to be split up in many tiny little manageable
1309  * pieces. Preferably with a chainsaw.
1310  *
1311  * @param client The client we're going to serve to.
1312  * @return when the client disconnects
1313  **/
1314 int mainloop(CLIENT *client) {
1315         struct nbd_request request;
1316         struct nbd_reply reply;
1317         gboolean go_on=TRUE;
1318 #ifdef DODBG
1319         int i = 0;
1320 #endif
1321         negotiate(client->net, client, NULL);
1322         DEBUG("Entering request loop!\n");
1323         reply.magic = htonl(NBD_REPLY_MAGIC);
1324         reply.error = 0;
1325         while (go_on) {
1326                 char buf[BUFSIZE];
1327                 size_t len;
1328 #ifdef DODBG
1329                 i++;
1330                 printf("%d: ", i);
1331 #endif
1332                 readit(client->net, &request, sizeof(request));
1333                 request.from = ntohll(request.from);
1334                 request.type = ntohl(request.type);
1335
1336                 if (request.type==NBD_CMD_DISC) {
1337                         msg2(LOG_INFO, "Disconnect request received.");
1338                         if (client->server->flags & F_COPYONWRITE) {
1339                                 if (client->difmap) g_free(client->difmap) ;
1340                                 close(client->difffile);
1341                                 unlink(client->difffilename);
1342                                 free(client->difffilename);
1343                         }
1344                         go_on=FALSE;
1345                         continue;
1346                 }
1347
1348                 len = ntohl(request.len);
1349
1350                 if (request.magic != htonl(NBD_REQUEST_MAGIC))
1351                         err("Not enough magic.");
1352                 if (len > BUFSIZE + sizeof(struct nbd_reply))
1353                         err("Request too big!");
1354 #ifdef DODBG
1355                 printf("%s from %llu (%llu) len %d, ", request.type ? "WRITE" :
1356                                 "READ", (unsigned long long)request.from,
1357                                 (unsigned long long)request.from / 512, len);
1358 #endif
1359                 memcpy(reply.handle, request.handle, sizeof(reply.handle));
1360                 if ((request.from + len) > (OFFT_MAX)) {
1361                         DEBUG("[Number too large!]");
1362                         ERROR(client, reply, EINVAL);
1363                         continue;
1364                 }
1365
1366                 if (((ssize_t)((off_t)request.from + len) > client->exportsize)) {
1367                         DEBUG("[RANGE!]");
1368                         ERROR(client, reply, EINVAL);
1369                         continue;
1370                 }
1371
1372                 if (request.type==NBD_CMD_WRITE) {
1373                         DEBUG("wr: net->buf, ");
1374                         readit(client->net, buf, len);
1375                         DEBUG("buf->exp, ");
1376                         if ((client->server->flags & F_READONLY) ||
1377                             (client->server->flags & F_AUTOREADONLY)) {
1378                                 DEBUG("[WRITE to READONLY!]");
1379                                 ERROR(client, reply, EPERM);
1380                                 continue;
1381                         }
1382                         if (expwrite(request.from, buf, len, client)) {
1383                                 DEBUG("Write failed: %m" );
1384                                 ERROR(client, reply, errno);
1385                                 continue;
1386                         }
1387                         SEND(client->net, reply);
1388                         DEBUG("OK!\n");
1389                         continue;
1390                 }
1391                 /* READ */
1392
1393                 DEBUG("exp->buf, ");
1394                 if (expread(request.from, buf + sizeof(struct nbd_reply), len, client)) {
1395                         DEBUG("Read failed: %m");
1396                         ERROR(client, reply, errno);
1397                         continue;
1398                 }
1399
1400                 DEBUG("buf->net, ");
1401                 memcpy(buf, &reply, sizeof(struct nbd_reply));
1402                 writeit(client->net, buf, len + sizeof(struct nbd_reply));
1403                 DEBUG("OK!\n");
1404         }
1405         return 0;
1406 }
1407
1408 /**
1409  * Set up client export array, which is an array of FILE_INFO.
1410  * Also, split a single exportfile into multiple ones, if that was asked.
1411  * @param client information on the client which we want to setup export for
1412  **/
1413 void setupexport(CLIENT* client) {
1414         int i;
1415         off_t laststartoff = 0, lastsize = 0;
1416         int multifile = (client->server->flags & F_MULTIFILE);
1417
1418         client->export = g_array_new(TRUE, TRUE, sizeof(FILE_INFO));
1419
1420         /* If multi-file, open as many files as we can.
1421          * If not, open exactly one file.
1422          * Calculate file sizes as we go to get total size. */
1423         for(i=0; ; i++) {
1424                 FILE_INFO fi;
1425                 gchar *tmpname;
1426                 gchar* error_string;
1427                 mode_t mode = (client->server->flags & F_READONLY) ? O_RDONLY : O_RDWR;
1428
1429                 if(multifile) {
1430                         tmpname=g_strdup_printf("%s.%d", client->exportname, i);
1431                 } else {
1432                         tmpname=g_strdup(client->exportname);
1433                 }
1434                 DEBUG2( "Opening %s\n", tmpname );
1435                 fi.fhandle = open(tmpname, mode);
1436                 if(fi.fhandle == -1 && mode == O_RDWR) {
1437                         /* Try again because maybe media was read-only */
1438                         fi.fhandle = open(tmpname, O_RDONLY);
1439                         if(fi.fhandle != -1) {
1440                                 /* Opening the base file in copyonwrite mode is
1441                                  * okay */
1442                                 if(!(client->server->flags & F_COPYONWRITE)) {
1443                                         client->server->flags |= F_AUTOREADONLY;
1444                                         client->server->flags |= F_READONLY;
1445                                 }
1446                         }
1447                 }
1448                 if(fi.fhandle == -1) {
1449                         if(multifile && i>0)
1450                                 break;
1451                         error_string=g_strdup_printf(
1452                                 "Could not open exported file %s: %%m",
1453                                 tmpname);
1454                         err(error_string);
1455                 }
1456                 fi.startoff = laststartoff + lastsize;
1457                 g_array_append_val(client->export, fi);
1458                 g_free(tmpname);
1459
1460                 /* Starting offset and size of this file will be used to
1461                  * calculate starting offset of next file */
1462                 laststartoff = fi.startoff;
1463                 lastsize = size_autodetect(fi.fhandle);
1464
1465                 if(!multifile)
1466                         break;
1467         }
1468
1469         /* Set export size to total calculated size */
1470         client->exportsize = laststartoff + lastsize;
1471
1472         /* Export size may be overridden */
1473         if(client->server->expected_size) {
1474                 /* desired size must be <= total calculated size */
1475                 if(client->server->expected_size > client->exportsize) {
1476                         err("Size of exported file is too big\n");
1477                 }
1478
1479                 client->exportsize = client->server->expected_size;
1480         }
1481
1482         msg3(LOG_INFO, "Size of exported file/device is %llu", (unsigned long long)client->exportsize);
1483         if(multifile) {
1484                 msg3(LOG_INFO, "Total number of files: %d", i);
1485         }
1486 }
1487
1488 int copyonwrite_prepare(CLIENT* client) {
1489         off_t i;
1490         if ((client->difffilename = malloc(1024))==NULL)
1491                 err("Failed to allocate string for diff file name");
1492         snprintf(client->difffilename, 1024, "%s-%s-%d.diff",client->exportname,client->clientname,
1493                 (int)getpid()) ;
1494         client->difffilename[1023]='\0';
1495         msg3(LOG_INFO,"About to create map and diff file %s",client->difffilename) ;
1496         client->difffile=open(client->difffilename,O_RDWR | O_CREAT | O_TRUNC,0600) ;
1497         if (client->difffile<0) err("Could not create diff file (%m)") ;
1498         if ((client->difmap=calloc(client->exportsize/DIFFPAGESIZE,sizeof(u32)))==NULL)
1499                 err("Could not allocate memory") ;
1500         for (i=0;i<client->exportsize/DIFFPAGESIZE;i++) client->difmap[i]=(u32)-1 ;
1501
1502         return 0;
1503 }
1504
1505 /**
1506  * Run a command. This is used for the ``prerun'' and ``postrun'' config file
1507  * options
1508  *
1509  * @param command the command to be ran. Read from the config file
1510  * @param file the file name we're about to export
1511  **/
1512 int do_run(gchar* command, gchar* file) {
1513         gchar* cmd;
1514         int retval=0;
1515
1516         if(command && *command) {
1517                 cmd = g_strdup_printf(command, file);
1518                 retval=system(cmd);
1519                 g_free(cmd);
1520         }
1521         return retval;
1522 }
1523
1524 /**
1525  * Serve a connection.
1526  *
1527  * @todo allow for multithreading, perhaps use libevent. Not just yet, though;
1528  * follow the road map.
1529  *
1530  * @param client a connected client
1531  **/
1532 void serveconnection(CLIENT *client) {
1533         if(do_run(client->server->prerun, client->exportname)) {
1534                 exit(EXIT_FAILURE);
1535         }
1536         setupexport(client);
1537
1538         if (client->server->flags & F_COPYONWRITE) {
1539                 copyonwrite_prepare(client);
1540         }
1541
1542         setmysockopt(client->net);
1543
1544         mainloop(client);
1545         do_run(client->server->postrun, client->exportname);
1546 }
1547
1548 /**
1549  * Find the name of the file we have to serve. This will use g_strdup_printf
1550  * to put the IP address of the client inside a filename containing
1551  * "%s" (in the form as specified by the "virtstyle" option). That name
1552  * is then written to client->exportname.
1553  *
1554  * @param net A socket connected to an nbd client
1555  * @param client information about the client. The IP address in human-readable
1556  * format will be written to a new char* buffer, the address of which will be
1557  * stored in client->clientname.
1558  **/
1559 void set_peername(int net, CLIENT *client) {
1560         struct sockaddr_storage addrin;
1561         struct sockaddr_storage netaddr;
1562         struct sockaddr_in  *netaddr4 = NULL;
1563         struct sockaddr_in6 *netaddr6 = NULL;
1564         size_t addrinlen = sizeof( addrin );
1565         struct addrinfo hints;
1566         struct addrinfo *ai = NULL;
1567         char peername[NI_MAXHOST];
1568         char netname[NI_MAXHOST];
1569         char *tmp = NULL;
1570         int i;
1571         int e;
1572         int shift;
1573
1574         if (getpeername(net, (struct sockaddr *) &addrin, (socklen_t *)&addrinlen) < 0)
1575                 err("getsockname failed: %m");
1576
1577         getnameinfo((struct sockaddr *)&addrin, (socklen_t)addrinlen,
1578                 peername, sizeof (peername), NULL, 0, NI_NUMERICHOST);
1579
1580         memset(&hints, '\0', sizeof (hints));
1581         hints.ai_flags = AI_ADDRCONFIG;
1582         e = getaddrinfo(peername, NULL, &hints, &ai);
1583
1584         if(e != 0) {
1585                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1586                 freeaddrinfo(ai);
1587                 return;
1588         }
1589
1590         switch(client->server->virtstyle) {
1591                 case VIRT_NONE:
1592                         client->exportname=g_strdup(client->server->exportname);
1593                         break;
1594                 case VIRT_IPHASH:
1595                         for(i=0;i<strlen(peername);i++) {
1596                                 if(peername[i]=='.') {
1597                                         peername[i]='/';
1598                                 }
1599                         }
1600                 case VIRT_IPLIT:
1601                         client->exportname=g_strdup_printf(client->server->exportname, peername);
1602                         break;
1603                 case VIRT_CIDR:
1604                         memcpy(&netaddr, &addrin, addrinlen);
1605                         if(ai->ai_family == AF_INET) {
1606                                 netaddr4 = (struct sockaddr_in *)&netaddr;
1607                                 (netaddr4->sin_addr).s_addr>>=32-(client->server->cidrlen);
1608                                 (netaddr4->sin_addr).s_addr<<=32-(client->server->cidrlen);
1609
1610                                 getnameinfo((struct sockaddr *) netaddr4, (socklen_t) addrinlen,
1611                                                         netname, sizeof (netname), NULL, 0, NI_NUMERICHOST);
1612                                 tmp=g_strdup_printf("%s/%s", netname, peername);
1613                         }else if(ai->ai_family == AF_INET6) {
1614                                 netaddr6 = (struct sockaddr_in6 *)&netaddr;
1615
1616                                 shift = 128-(client->server->cidrlen);
1617                                 i = 3;
1618                                 while(shift >= 32) {
1619                                         ((netaddr6->sin6_addr).s6_addr32[i])=0;
1620                                         shift-=32;
1621                                         i--;
1622                                 }
1623                                 (netaddr6->sin6_addr).s6_addr32[i]>>=shift;
1624                                 (netaddr6->sin6_addr).s6_addr32[i]<<=shift;
1625
1626                                 getnameinfo((struct sockaddr *)netaddr6, (socklen_t)addrinlen,
1627                                             netname, sizeof(netname), NULL, 0, NI_NUMERICHOST);
1628                                 tmp=g_strdup_printf("%s/%s", netname, peername);
1629                         }
1630
1631                         if(tmp != NULL)
1632                           client->exportname=g_strdup_printf(client->server->exportname, tmp);
1633
1634                         break;
1635         }
1636
1637         freeaddrinfo(ai);
1638         msg4(LOG_INFO, "connect from %s, assigned file is %s",
1639              peername, client->exportname);
1640         client->clientname=g_strdup(peername);
1641 }
1642
1643 /**
1644  * Destroy a pid_t*
1645  * @param data a pointer to pid_t which should be freed
1646  **/
1647 void destroy_pid_t(gpointer data) {
1648         g_free(data);
1649 }
1650
1651 /**
1652  * Loop through the available servers, and serve them. Never returns.
1653  **/
1654 int serveloop(GArray* servers) {
1655         struct sockaddr_storage addrin;
1656         socklen_t addrinlen=sizeof(addrin);
1657         int i;
1658         int max;
1659         int sock;
1660         fd_set mset;
1661         fd_set rset;
1662
1663         /*
1664          * Set up the master fd_set. The set of descriptors we need
1665          * to select() for never changes anyway and it buys us a *lot*
1666          * of time to only build this once. However, if we ever choose
1667          * to not fork() for clients anymore, we may have to revisit
1668          * this.
1669          */
1670         max=0;
1671         FD_ZERO(&mset);
1672         for(i=0;i<servers->len;i++) {
1673                 if((sock=(g_array_index(servers, SERVER, i)).socket)) {
1674                         FD_SET(sock, &mset);
1675                         max=sock>max?sock:max;
1676                 }
1677         }
1678         if(modernsock) {
1679                 FD_SET(modernsock, &mset);
1680                 max=modernsock>max?modernsock:max;
1681         }
1682         for(;;) {
1683                 CLIENT *client = NULL;
1684                 pid_t *pid;
1685
1686                 memcpy(&rset, &mset, sizeof(fd_set));
1687                 if(select(max+1, &rset, NULL, NULL, NULL)>0) {
1688                         int net = 0;
1689                         SERVER* serve;
1690
1691                         DEBUG("accept, ");
1692                         if(FD_ISSET(modernsock, &rset)) {
1693                                 if((net=accept(modernsock, (struct sockaddr *) &addrin, &addrinlen)) < 0)
1694                                         err("accept: %m");
1695                                 client = negotiate(net, NULL, servers);
1696                                 if(!client) {
1697                                         err_nonfatal("negotiation failed");
1698                                         close(net);
1699                                 }
1700                         }
1701                         for(i=0;i<servers->len && !net;i++) {
1702                                 serve=&(g_array_index(servers, SERVER, i));
1703                                 if(FD_ISSET(serve->socket, &rset)) {
1704                                         if ((net=accept(serve->socket, (struct sockaddr *) &addrin, &addrinlen)) < 0)
1705                                                 err("accept: %m");
1706                                 }
1707                         }
1708                         if(net) {
1709                                 int sock_flags;
1710
1711                                 if((sock_flags = fcntl(net, F_GETFL, 0))==-1) {
1712                                         err("fcntl F_GETFL");
1713                                 }
1714                                 if(fcntl(net, F_SETFL, sock_flags &~O_NONBLOCK)==-1) {
1715                                         err("fcntl F_SETFL ~O_NONBLOCK");
1716                                 }
1717                                 if(!client) {
1718                                         client = g_new0(CLIENT, 1);
1719                                         client->server=serve;
1720                                         client->exportsize=OFFT_MAX;
1721                                         client->net=net;
1722                                 }
1723                                 set_peername(net, client);
1724                                 if (!authorized_client(client)) {
1725                                         msg2(LOG_INFO,"Unauthorized client") ;
1726                                         close(net);
1727                                         continue;
1728                                 }
1729                                 msg2(LOG_INFO,"Authorized client") ;
1730                                 pid=g_malloc(sizeof(pid_t));
1731 #ifndef NOFORK
1732                                 if ((*pid=fork())<0) {
1733                                         msg3(LOG_INFO,"Could not fork (%s)",strerror(errno)) ;
1734                                         close(net);
1735                                         continue;
1736                                 }
1737                                 if (*pid>0) { /* parent */
1738                                         close(net);
1739                                         g_hash_table_insert(children, pid, pid);
1740                                         continue;
1741                                 }
1742                                 /* child */
1743                                 g_hash_table_destroy(children);
1744                                 for(i=0;i<servers->len;i++) {
1745                                         serve=&g_array_index(servers, SERVER, i);
1746                                         close(serve->socket);
1747                                 }
1748                                 /* FALSE does not free the
1749                                 actual data. This is required,
1750                                 because the client has a
1751                                 direct reference into that
1752                                 data, and otherwise we get a
1753                                 segfault... */
1754                                 g_array_free(servers, FALSE);
1755 #endif // NOFORK
1756                                 msg2(LOG_INFO,"Starting to serve");
1757                                 serveconnection(client);
1758                                 exit(EXIT_SUCCESS);
1759                         }
1760                 }
1761         }
1762 }
1763
1764 void dosockopts(int socket) {
1765 #ifndef sun
1766         int yes=1;
1767 #else
1768         char yes='1';
1769 #endif /* sun */
1770         int sock_flags;
1771
1772         /* lose the pesky "Address already in use" error message */
1773         if (setsockopt(socket,SOL_SOCKET,SO_REUSEADDR,&yes,sizeof(int)) == -1) {
1774                 err("setsockopt SO_REUSEADDR");
1775         }
1776         if (setsockopt(socket,SOL_SOCKET,SO_KEEPALIVE,&yes,sizeof(int)) == -1) {
1777                 err("setsockopt SO_KEEPALIVE");
1778         }
1779
1780         /* make the listening socket non-blocking */
1781         if ((sock_flags = fcntl(socket, F_GETFL, 0)) == -1) {
1782                 err("fcntl F_GETFL");
1783         }
1784         if (fcntl(socket, F_SETFL, sock_flags | O_NONBLOCK) == -1) {
1785                 err("fcntl F_SETFL O_NONBLOCK");
1786         }
1787 }
1788
1789 /**
1790  * Connect a server's socket.
1791  *
1792  * @param serve the server we want to connect.
1793  **/
1794 int setup_serve(SERVER *serve) {
1795         struct addrinfo hints;
1796         struct addrinfo *ai = NULL;
1797         gchar *port = NULL;
1798         int e;
1799
1800         if(!do_oldstyle) {
1801                 return serve->servename ? 1 : 0;
1802         }
1803         memset(&hints,'\0',sizeof(hints));
1804         hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG | AI_NUMERICSERV;
1805         hints.ai_socktype = SOCK_STREAM;
1806         hints.ai_family = serve->socket_family;
1807
1808         port = g_strdup_printf ("%d", serve->port);
1809         if (port == NULL)
1810                 return 0;
1811
1812         e = getaddrinfo(serve->listenaddr,port,&hints,&ai);
1813
1814         g_free(port);
1815
1816         if(e != 0) {
1817                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1818                 serve->socket = -1;
1819                 freeaddrinfo(ai);
1820                 exit(EXIT_FAILURE);
1821         }
1822
1823         if(serve->socket_family == AF_UNSPEC)
1824                 serve->socket_family = ai->ai_family;
1825
1826 #ifdef WITH_SDP
1827         if ((serve->flags) && F_SDP) {
1828                 if (ai->ai_family == AF_INET)
1829                         ai->ai_family = AF_INET_SDP;
1830                 else (ai->ai_family == AF_INET6)
1831                         ai->ai_family = AF_INET6_SDP;
1832         }
1833 #endif
1834         if ((serve->socket = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol)) < 0)
1835                 err("socket: %m");
1836
1837         dosockopts(serve->socket);
1838
1839         DEBUG("Waiting for connections... bind, ");
1840         e = bind(serve->socket, ai->ai_addr, ai->ai_addrlen);
1841         if (e != 0 && errno != EADDRINUSE)
1842                 err("bind: %m");
1843         DEBUG("listen, ");
1844         if (listen(serve->socket, 1) < 0)
1845                 err("listen: %m");
1846
1847         freeaddrinfo (ai);
1848         if(serve->servename) {
1849                 return 1;
1850         } else {
1851                 return 0;
1852         }
1853 }
1854
1855 void open_modern(void) {
1856         struct addrinfo hints;
1857         struct addrinfo* ai = NULL;
1858         struct sock_flags;
1859         int e;
1860
1861         memset(&hints, '\0', sizeof(hints));
1862         hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG;
1863         hints.ai_socktype = SOCK_STREAM;
1864         hints.ai_family = AF_UNSPEC;
1865         hints.ai_protocol = IPPROTO_TCP;
1866         e = getaddrinfo(modern_listen, NBD_DEFAULT_PORT, &hints, &ai);
1867         if(e != 0) {
1868                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1869                 exit(EXIT_FAILURE);
1870         }
1871         if((modernsock = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol))<0) {
1872                 err("socket: %m");
1873         }
1874
1875         dosockopts(modernsock);
1876
1877         if(bind(modernsock, ai->ai_addr, ai->ai_addrlen)) {
1878                 err("bind: %m");
1879         }
1880         if(listen(modernsock, 10) <0) {
1881                 err("listen: %m");
1882         }
1883
1884         freeaddrinfo(ai);
1885 }
1886
1887 /**
1888  * Connect our servers.
1889  **/
1890 void setup_servers(GArray* servers) {
1891         int i;
1892         struct sigaction sa;
1893         int want_modern=0;
1894
1895         for(i=0;i<servers->len;i++) {
1896                 want_modern |= setup_serve(&(g_array_index(servers, SERVER, i)));
1897         }
1898         if(want_modern) {
1899                 open_modern();
1900         }
1901         children=g_hash_table_new_full(g_int_hash, g_int_equal, NULL, destroy_pid_t);
1902
1903         sa.sa_handler = sigchld_handler;
1904         sigemptyset(&sa.sa_mask);
1905         sa.sa_flags = SA_RESTART;
1906         if(sigaction(SIGCHLD, &sa, NULL) == -1)
1907                 err("sigaction: %m");
1908         sa.sa_handler = sigterm_handler;
1909         sigemptyset(&sa.sa_mask);
1910         sa.sa_flags = SA_RESTART;
1911         if(sigaction(SIGTERM, &sa, NULL) == -1)
1912                 err("sigaction: %m");
1913 }
1914
1915 /**
1916  * Go daemon (unless we specified at compile time that we didn't want this)
1917  * @param serve the first server of our configuration. If its port is zero,
1918  *      then do not daemonize, because we're doing inetd then. This parameter
1919  *      is only used to create a PID file of the form
1920  *      /var/run/nbd-server.&lt;port&gt;.pid; it's not modified in any way.
1921  **/
1922 #if !defined(NODAEMON) && !defined(NOFORK)
1923 void daemonize(SERVER* serve) {
1924         FILE*pidf;
1925
1926         if(serve && !(serve->port)) {
1927                 return;
1928         }
1929         if(daemon(0,0)<0) {
1930                 err("daemon");
1931         }
1932         if(!*pidftemplate) {
1933                 if(serve) {
1934                         strncpy(pidftemplate, "/var/run/nbd-server.%d.pid", 255);
1935                 } else {
1936                         strncpy(pidftemplate, "/var/run/nbd-server.pid", 255);
1937                 }
1938         }
1939         snprintf(pidfname, 255, pidftemplate, serve ? serve->port : 0);
1940         pidf=fopen(pidfname, "w");
1941         if(pidf) {
1942                 fprintf(pidf,"%d\n", (int)getpid());
1943                 fclose(pidf);
1944         } else {
1945                 perror("fopen");
1946                 fprintf(stderr, "Not fatal; continuing");
1947         }
1948 }
1949 #else
1950 #define daemonize(serve)
1951 #endif /* !defined(NODAEMON) && !defined(NOFORK) */
1952
1953 /*
1954  * Everything beyond this point (in the file) is run in non-daemon mode.
1955  * The stuff above daemonize() isn't.
1956  */
1957
1958 void serve_err(SERVER* serve, const char* msg) G_GNUC_NORETURN;
1959
1960 void serve_err(SERVER* serve, const char* msg) {
1961         g_message("Export of %s on port %d failed:", serve->exportname,
1962                         serve->port);
1963         err(msg);
1964 }
1965
1966 /**
1967  * Set up user-ID and/or group-ID
1968  **/
1969 void dousers(void) {
1970         struct passwd *pw;
1971         struct group *gr;
1972         gchar* str;
1973         if(rungroup) {
1974                 gr=getgrnam(rungroup);
1975                 if(!gr) {
1976                         str = g_strdup_printf("Invalid group name: %s", rungroup);
1977                         err(str);
1978                 }
1979                 if(setgid(gr->gr_gid)<0) {
1980                         err("Could not set GID: %m");
1981                 }
1982         }
1983         if(runuser) {
1984                 pw=getpwnam(runuser);
1985                 if(!pw) {
1986                         str = g_strdup_printf("Invalid user name: %s", runuser);
1987                         err(str);
1988                 }
1989                 if(setuid(pw->pw_uid)<0) {
1990                         err("Could not set UID: %m");
1991                 }
1992         }
1993 }
1994
1995 #ifndef ISSERVER
1996 void glib_message_syslog_redirect(const gchar *log_domain,
1997                                   GLogLevelFlags log_level,
1998                                   const gchar *message,
1999                                   gpointer user_data)
2000 {
2001     int level=LOG_DEBUG;
2002
2003     switch( log_level )
2004     {
2005       case G_LOG_FLAG_FATAL:
2006       case G_LOG_LEVEL_CRITICAL:
2007       case G_LOG_LEVEL_ERROR:
2008         level=LOG_ERR;
2009         break;
2010       case G_LOG_LEVEL_WARNING:
2011         level=LOG_WARNING;
2012         break;
2013       case G_LOG_LEVEL_MESSAGE:
2014       case G_LOG_LEVEL_INFO:
2015         level=LOG_INFO;
2016         break;
2017       case G_LOG_LEVEL_DEBUG:
2018         level=LOG_DEBUG;
2019       default:
2020         level=LOG_ERR;
2021     }
2022     syslog(level, message);
2023 }
2024 #endif
2025
2026 /**
2027  * Main entry point...
2028  **/
2029 int main(int argc, char *argv[]) {
2030         SERVER *serve;
2031         GArray *servers;
2032         GError *err=NULL;
2033
2034         if (sizeof( struct nbd_request )!=28) {
2035                 fprintf(stderr,"Bad size of structure. Alignment problems?\n");
2036                 exit(EXIT_FAILURE) ;
2037         }
2038
2039         memset(pidftemplate, '\0', 256);
2040
2041         logging();
2042         config_file_pos = g_strdup(CFILE);
2043         serve=cmdline(argc, argv);
2044         servers = parse_cfile(config_file_pos, &err);
2045
2046         if(serve) {
2047                 serve->socket_family = AF_UNSPEC;
2048
2049                 append_serve(serve, servers);
2050
2051                 if (!(serve->port)) {
2052                         CLIENT *client;
2053 #ifndef ISSERVER
2054                         /* You really should define ISSERVER if you're going to use
2055                          * inetd mode, but if you don't, closing stdout and stderr
2056                          * (which inetd had connected to the client socket) will let it
2057                          * work. */
2058                         close(1);
2059                         close(2);
2060                         open("/dev/null", O_WRONLY);
2061                         open("/dev/null", O_WRONLY);
2062                         g_log_set_default_handler( glib_message_syslog_redirect, NULL );
2063 #endif
2064                         client=g_malloc(sizeof(CLIENT));
2065                         client->server=serve;
2066                         client->net=0;
2067                         client->exportsize=OFFT_MAX;
2068                         set_peername(0,client);
2069                         serveconnection(client);
2070                         return 0;
2071                 }
2072         }
2073
2074         if(!servers || !servers->len) {
2075                 g_warning("Could not parse config file: %s",
2076                                 err ? err->message : "Unknown error");
2077         }
2078         if(serve) {
2079                 g_warning("Specifying an export on the command line is deprecated.");
2080                 g_warning("Please use a configuration file instead.");
2081         }
2082
2083         if((!serve) && (!servers||!servers->len)) {
2084                 g_message("Nothing to do! Bye!");
2085                 exit(EXIT_FAILURE);
2086         }
2087         daemonize(serve);
2088         setup_servers(servers);
2089         dousers();
2090         serveloop(servers);
2091         return 0 ;
2092 }