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         return serve;
 601 }
 602
 603 /**
 604  * append new server to array
 605  * @param s server
 606  * @param a server array
 607  * @return 0 success, -1 error
 608  */
 609 int append_serve(SERVER *s, GArray *a) {
 610         SERVER *ns = NULL;
 611         struct addrinfo hints;
 612         struct addrinfo *ai = NULL;
 613         struct addrinfo *rp = NULL;
 614         char   host[NI_MAXHOST];
 615         gchar  *port = NULL;
 616         int e;
 617         int ret;
 618
 619         if(!s) {
 620                 err("Invalid parsing server");
 621                 return -1;
 622         }
 623
 624         port = g_strdup_printf("%d", s->port);
 625
 626         memset(&hints,'\0',sizeof(hints));
 627         hints.ai_family = AF_UNSPEC;
 628         hints.ai_socktype = SOCK_STREAM;
 629         hints.ai_flags = AI_ADDRCONFIG | AI_PASSIVE;
 630         hints.ai_protocol = IPPROTO_TCP;
 631
 632         e = getaddrinfo(s->listenaddr, port, &hints, &ai);
 633
 634         if (port)
 635                 g_free(port);
 636
 637         if(e == 0) {
 638                 for (rp = ai; rp != NULL; rp = rp->ai_next) {
 639                         e = getnameinfo(rp->ai_addr, rp->ai_addrlen, host, sizeof(host), NULL, 0, NI_NUMERICHOST);
 640
 641                         if (e != 0) { // error
 642                                 fprintf(stderr, "getnameinfo: %s\n", gai_strerror(e));
 643                                 continue;
 644                         }
 645
 646                         // duplicate server and set listenaddr to resolved IP address
 647                         ns = dup_serve (s);
 648                         if (ns) {
 649                                 ns->listenaddr = g_strdup(host);
 650                                 ns->socket_family = rp->ai_family;
 651                                 g_array_append_val(a, *ns);
 652                                 free(ns);
 653                                 ns = NULL;
 654                         }
 655                 }
 656
 657                 ret = 0;
 658         } else {
 659                 fprintf(stderr, "getaddrinfo failed on listen host/address: %s (%s)\n", s->listenaddr ? s->listenaddr : "any", gai_strerror(e));
 660                 ret = -1;
 661         }
 662
 663         if (ai)
 664                 freeaddrinfo(ai);
 665
 666         return ret;
 667 }
 668
 669 /**
 670  * Parse the config file.
 671  *
 672  * @param f the name of the config file
 673  * @param e a GError. @see CFILE_ERRORS for what error values this function can
 674  *      return.
 675  * @return a Array of SERVER* pointers, If the config file is empty or does not
 676  *      exist, returns an empty GHashTable; if the config file contains an
 677  *      error, returns NULL, and e is set appropriately
 678  **/
 679 GArray* parse_cfile(gchar* f, GError** e) {
 680         const char* DEFAULT_ERROR = "Could not parse %s in group %s: %s";
 681         const char* MISSING_REQUIRED_ERROR = "Could not find required value %s in group %s: %s";
 682         SERVER s;
 683         gchar *virtstyle=NULL;
 684         PARAM lp[] = {
 685                 { "exportname", TRUE,   PARAM_STRING,   NULL, 0 },
 686                 { "port",       TRUE,   PARAM_INT,      NULL, 0 },
 687                 { "authfile",   FALSE,  PARAM_STRING,   NULL, 0 },
 688                 { "filesize",   FALSE,  PARAM_INT,      NULL, 0 },
 689                 { "virtstyle",  FALSE,  PARAM_STRING,   NULL, 0 },
 690                 { "prerun",     FALSE,  PARAM_STRING,   NULL, 0 },
 691                 { "postrun",    FALSE,  PARAM_STRING,   NULL, 0 },
 692                 { "readonly",   FALSE,  PARAM_BOOL,     NULL, F_READONLY },
 693                 { "multifile",  FALSE,  PARAM_BOOL,     NULL, F_MULTIFILE },
 694                 { "copyonwrite", FALSE, PARAM_BOOL,     NULL, F_COPYONWRITE },
 695                 { "sparse_cow", FALSE,  PARAM_BOOL,     NULL, F_SPARSE },
 696                 { "sdp",        FALSE,  PARAM_BOOL,     NULL, F_SDP },
 697                 { "sync",       FALSE,  PARAM_BOOL,     NULL, F_SYNC },
 698                 { "listenaddr", FALSE,  PARAM_STRING,   NULL, 0 },
 699         };
 700         const int lp_size=sizeof(lp)/sizeof(PARAM);
 701         PARAM gp[] = {
 702                 { "user",       FALSE, PARAM_STRING,    &runuser,       0 },
 703                 { "group",      FALSE, PARAM_STRING,    &rungroup,      0 },
 704                 { "oldstyle",   FALSE, PARAM_BOOL,      &do_oldstyle,   1 },
 705                 { "listenaddr", FALSE, PARAM_STRING,    &modern_listen, 0 },
 706         };
 707         PARAM* p=gp;
 708         int p_size=sizeof(gp)/sizeof(PARAM);
 709         GKeyFile *cfile;
 710         GError *err = NULL;
 711         const char *err_msg=NULL;
 712         GQuark errdomain;
 713         GArray *retval=NULL;
 714         gchar **groups;
 715         gboolean value;
 716         gchar* startgroup;
 717         gint i;
 718         gint j;
 719
 720         errdomain = g_quark_from_string("parse_cfile");
 721         cfile = g_key_file_new();
 722         retval = g_array_new(FALSE, TRUE, sizeof(SERVER));
 723         if(!g_key_file_load_from_file(cfile, f, G_KEY_FILE_KEEP_COMMENTS |
 724                         G_KEY_FILE_KEEP_TRANSLATIONS, &err)) {
 725                 g_set_error(e, errdomain, CFILE_NOTFOUND, "Could not open config file %s.", f);
 726                 g_key_file_free(cfile);
 727                 return retval;
 728         }
 729         startgroup = g_key_file_get_start_group(cfile);
 730         if(!startgroup || strcmp(startgroup, "generic")) {
 731                 g_set_error(e, errdomain, CFILE_MISSING_GENERIC, "Config file does not contain the [generic] group!");
 732                 g_key_file_free(cfile);
 733                 return NULL;
 734         }
 735         groups = g_key_file_get_groups(cfile, NULL);
 736         for(i=0;groups[i];i++) {
 737                 memset(&s, '\0', sizeof(SERVER));
 738                 lp[0].target=&(s.exportname);
 739                 lp[1].target=&(s.port);
 740                 lp[2].target=&(s.authname);
 741                 lp[3].target=&(s.expected_size);
 742                 lp[4].target=&(virtstyle);
 743                 lp[5].target=&(s.prerun);
 744                 lp[6].target=&(s.postrun);
 745                 lp[7].target=lp[8].target=lp[9].target=
 746                                 lp[10].target=lp[11].target=
 747                                 lp[12].target=&(s.flags);
 748                 lp[13].target=&(s.listenaddr);
 749
 750                 /* After the [generic] group, start parsing exports */
 751                 if(i==1) {
 752                         p=lp;
 753                         p_size=lp_size;
 754                 }
 755                 for(j=0;j<p_size;j++) {
 756                         g_assert(p[j].target != NULL);
 757                         g_assert(p[j].ptype==PARAM_INT||p[j].ptype==PARAM_STRING||p[j].ptype==PARAM_BOOL);
 758                         switch(p[j].ptype) {
 759                                 case PARAM_INT:
 760                                         *((gint*)p[j].target) =
 761                                                 g_key_file_get_integer(cfile,
 762                                                                 groups[i],
 763                                                                 p[j].paramname,
 764                                                                 &err);
 765                                         break;
 766                                 case PARAM_STRING:
 767                                         *((gchar**)p[j].target) =
 768                                                 g_key_file_get_string(cfile,
 769                                                                 groups[i],
 770                                                                 p[j].paramname,
 771                                                                 &err);
 772                                         break;
 773                                 case PARAM_BOOL:
 774                                         value = g_key_file_get_boolean(cfile,
 775                                                         groups[i],
 776                                                         p[j].paramname, &err);
 777                                         if(!err) {
 778                                                 if(value) {
 779                                                         *((gint*)p[j].target) |= p[j].flagval;
 780                                                 } else {
 781                                                         *((gint*)p[j].target) &= ~(p[j].flagval);
 782                                                 }
 783                                         }
 784                                         break;
 785                         }
 786                         if(!strcmp(p[j].paramname, "port") && !strcmp(p[j].target, NBD_DEFAULT_PORT)) {
 787                                 g_set_error(e, errdomain, CFILE_INCORRECT_PORT, "Config file specifies default port for oldstyle export");
 788                                 g_key_file_free(cfile);
 789                                 return NULL;
 790                         }
 791                         if(err) {
 792                                 if(err->code == G_KEY_FILE_ERROR_KEY_NOT_FOUND) {
 793                                         if(!p[j].required) {
 794                                                 /* Ignore not-found error for optional values */
 795                                                 g_clear_error(&err);
 796                                                 continue;
 797                                         } else {
 798                                                 err_msg = MISSING_REQUIRED_ERROR;
 799                                         }
 800                                 } else {
 801                                         err_msg = DEFAULT_ERROR;
 802                                 }
 803                                 g_set_error(e, errdomain, CFILE_VALUE_INVALID, err_msg, p[j].paramname, groups[i], err->message);
 804                                 g_array_free(retval, TRUE);
 805                                 g_error_free(err);
 806                                 g_key_file_free(cfile);
 807                                 return NULL;
 808                         }
 809                 }
 810                 if(virtstyle) {
 811                         if(!strncmp(virtstyle, "none", 4)) {
 812                                 s.virtstyle=VIRT_NONE;
 813                         } else if(!strncmp(virtstyle, "ipliteral", 9)) {
 814                                 s.virtstyle=VIRT_IPLIT;
 815                         } else if(!strncmp(virtstyle, "iphash", 6)) {
 816                                 s.virtstyle=VIRT_IPHASH;
 817                         } else if(!strncmp(virtstyle, "cidrhash", 8)) {
 818                                 s.virtstyle=VIRT_CIDR;
 819                                 if(strlen(virtstyle)<10) {
 820                                         g_set_error(e, errdomain, CFILE_VALUE_INVALID, "Invalid value %s for parameter virtstyle in group %s: missing length", virtstyle, groups[i]);
 821                                         g_array_free(retval, TRUE);
 822                                         g_key_file_free(cfile);
 823                                         return NULL;
 824                                 }
 825                                 s.cidrlen=strtol(virtstyle+8, NULL, 0);
 826                         } else {
 827                                 g_set_error(e, errdomain, CFILE_VALUE_INVALID, "Invalid value %s for parameter virtstyle in group %s", virtstyle, groups[i]);
 828                                 g_array_free(retval, TRUE);
 829                                 g_key_file_free(cfile);
 830                                 return NULL;
 831                         }
 832                 } else {
 833                         s.virtstyle=VIRT_IPLIT;
 834                 }
 835                 /* Don't need to free this, it's not our string */
 836                 virtstyle=NULL;
 837                 /* Don't append values for the [generic] group */
 838                 if(i>0) {
 839                         s.socket_family = AF_UNSPEC;
 840                         s.servename = groups[i];
 841
 842                         append_serve(&s, retval);
 843                 } else {
 844                         if(!do_oldstyle) {
 845                                 lp[1].required = 0;
 846                         }
 847                 }
 848 #ifndef WITH_SDP
 849                 if(s.flags & F_SDP) {
 850                         g_set_error(e, errdomain, CFILE_VALUE_UNSUPPORTED, "This nbd-server was built without support for SDP, yet group %s uses it", groups[i]);
 851                         g_array_free(retval, TRUE);
 852                         g_key_file_free(cfile);
 853                         return NULL;
 854                 }
 855 #endif
 856         }
 857         if(i==1) {
 858                 g_set_error(e, errdomain, CFILE_NO_EXPORTS, "The config file does not specify any exports");
 859         }
 860         g_key_file_free(cfile);
 861         return retval;
 862 }
 863
 864 /**
 865  * Signal handler for SIGCHLD
 866  * @param s the signal we're handling (must be SIGCHLD, or something
 867  * is severely wrong)
 868  **/
 869 void sigchld_handler(int s) {
 870         int status;
 871         int* i;
 872         pid_t pid;
 873
 874         while((pid=waitpid(-1, &status, WNOHANG)) > 0) {
 875                 if(WIFEXITED(status)) {
 876                         msg3(LOG_INFO, "Child exited with %d", WEXITSTATUS(status));
 877                 }
 878                 i=g_hash_table_lookup(children, &pid);
 879                 if(!i) {
 880                         msg3(LOG_INFO, "SIGCHLD received for an unknown child with PID %ld", (long)pid);
 881                 } else {
 882                         DEBUG2("Removing %d from the list of children", pid);
 883                         g_hash_table_remove(children, &pid);
 884                 }
 885         }
 886 }
 887
 888 /**
 889  * Kill a child. Called from sigterm_handler::g_hash_table_foreach.
 890  *
 891  * @param key the key
 892  * @param value the value corresponding to the above key
 893  * @param user_data a pointer which we always set to 1, so that we know what
 894  * will happen next.
 895  **/
 896 void killchild(gpointer key, gpointer value, gpointer user_data) {
 897         pid_t *pid=value;
 898         int *parent=user_data;
 899
 900         kill(*pid, SIGTERM);
 901         *parent=1;
 902 }
 903
 904 /**
 905  * Handle SIGTERM and dispatch it to our children
 906  * @param s the signal we're handling (must be SIGTERM, or something
 907  * is severely wrong).
 908  **/
 909 void sigterm_handler(int s) {
 910         int parent=0;
 911
 912         g_hash_table_foreach(children, killchild, &parent);
 913
 914         if(parent) {
 915                 unlink(pidfname);
 916         }
 917
 918         exit(EXIT_SUCCESS);
 919 }
 920
 921 /**
 922  * Detect the size of a file.
 923  *
 924  * @param fhandle An open filedescriptor
 925  * @return the size of the file, or OFFT_MAX if detection was
 926  * impossible.
 927  **/
 928 off_t size_autodetect(int fhandle) {
 929         off_t es;
 930         u64 bytes;
 931         struct stat stat_buf;
 932         int error;
 933
 934 #ifdef HAVE_SYS_MOUNT_H
 935 #ifdef HAVE_SYS_IOCTL_H
 936 #ifdef BLKGETSIZE64
 937         DEBUG("looking for export size with ioctl BLKGETSIZE64\n");
 938         if (!ioctl(fhandle, BLKGETSIZE64, &bytes) && bytes) {
 939                 return (off_t)bytes;
 940         }
 941 #endif /* BLKGETSIZE64 */
 942 #endif /* HAVE_SYS_IOCTL_H */
 943 #endif /* HAVE_SYS_MOUNT_H */
 944
 945         DEBUG("looking for fhandle size with fstat\n");
 946         stat_buf.st_size = 0;
 947         error = fstat(fhandle, &stat_buf);
 948         if (!error) {
 949                 if(stat_buf.st_size > 0)
 950                         return (off_t)stat_buf.st_size;
 951         } else {
 952                 err("fstat failed: %m");
 953         }
 954
 955         DEBUG("looking for fhandle size with lseek SEEK_END\n");
 956         es = lseek(fhandle, (off_t)0, SEEK_END);
 957         if (es > ((off_t)0)) {
 958                 return es;
 959         } else {
 960                 DEBUG2("lseek failed: %d", errno==EBADF?1:(errno==ESPIPE?2:(errno==EINVAL?3:4)));
 961         }
 962
 963         err("Could not find size of exported block device: %m");
 964         return OFFT_MAX;
 965 }
 966
 967 /**
 968  * Get the file handle and offset, given an export offset.
 969  *
 970  * @param export An array of export files
 971  * @param a The offset to get corresponding file/offset for
 972  * @param fhandle [out] File descriptor
 973  * @param foffset [out] Offset into fhandle
 974  * @param maxbytes [out] Tells how many bytes can be read/written
 975  * from fhandle starting at foffset (0 if there is no limit)
 976  * @return 0 on success, -1 on failure
 977  **/
 978 int get_filepos(GArray* export, off_t a, int* fhandle, off_t* foffset, size_t* maxbytes ) {
 979         /* Negative offset not allowed */
 980         if(a < 0)
 981                 return -1;
 982
 983         /* Binary search for last file with starting offset <= a */
 984         FILE_INFO fi;
 985         int start = 0;
 986         int end = export->len - 1;
 987         while( start <= end ) {
 988                 int mid = (start + end) / 2;
 989                 fi = g_array_index(export, FILE_INFO, mid);
 990                 if( fi.startoff < a ) {
 991                         start = mid + 1;
 992                 } else if( fi.startoff > a ) {
 993                         end = mid - 1;
 994                 } else {
 995                         start = end = mid;
 996                         break;
 997                 }
 998         }
 999
1000         /* end should never go negative, since first startoff is 0 and a >= 0 */
1001         g_assert(end >= 0);
1002
1003         fi = g_array_index(export, FILE_INFO, end);
1004         *fhandle = fi.fhandle;
1005         *foffset = a - fi.startoff;
1006         *maxbytes = 0;
1007         if( end+1 < export->len ) {
1008                 FILE_INFO fi_next = g_array_index(export, FILE_INFO, end+1);
1009                 *maxbytes = fi_next.startoff - a;
1010         }
1011
1012         return 0;
1013 }
1014
1015 /**
1016  * seek to a position in a file, with error handling.
1017  * @param handle a filedescriptor
1018  * @param a position to seek to
1019  * @todo get rid of this; lastpoint is a global variable right now, but it
1020  * shouldn't be. If we pass it on as a parameter, that makes things a *lot*
1021  * easier.
1022  **/
1023 void myseek(int handle,off_t a) {
1024         if (lseek(handle, a, SEEK_SET) < 0) {
1025                 err("Can not seek locally!\n");
1026         }
1027 }
1028
1029 /**
1030  * Write an amount of bytes at a given offset to the right file. This
1031  * abstracts the write-side of the multiple file option.
1032  *
1033  * @param a The offset where the write should start
1034  * @param buf The buffer to write from
1035  * @param len The length of buf
1036  * @param client The client we're serving for
1037  * @return The number of bytes actually written, or -1 in case of an error
1038  **/
1039 ssize_t rawexpwrite(off_t a, char *buf, size_t len, CLIENT *client) {
1040         int fhandle;
1041         off_t foffset;
1042         size_t maxbytes;
1043         ssize_t retval;
1044
1045         if(get_filepos(client->export, a, &fhandle, &foffset, &maxbytes))
1046                 return -1;
1047         if(maxbytes && len > maxbytes)
1048                 len = maxbytes;
1049
1050         DEBUG4("(WRITE to fd %d offset %llu len %u), ", fhandle, foffset, len);
1051
1052         myseek(fhandle, foffset);
1053         retval = write(fhandle, buf, len);
1054         if(client->server->flags & F_SYNC) {
1055                 fsync(fhandle);
1056         }
1057         return retval;
1058 }
1059
1060 /**
1061  * Call rawexpwrite repeatedly until all data has been written.
1062  * @return 0 on success, nonzero on failure
1063  **/
1064 int rawexpwrite_fully(off_t a, char *buf, size_t len, CLIENT *client) {
1065         ssize_t ret=0;
1066
1067         while(len > 0 && (ret=rawexpwrite(a, buf, len, client)) > 0 ) {
1068                 a += ret;
1069                 buf += ret;
1070                 len -= ret;
1071         }
1072         return (ret < 0 || len != 0);
1073 }
1074
1075 /**
1076  * Read an amount of bytes at a given offset from the right file. This
1077  * abstracts the read-side of the multiple files option.
1078  *
1079  * @param a The offset where the read should start
1080  * @param buf A buffer to read into
1081  * @param len The size of buf
1082  * @param client The client we're serving for
1083  * @return The number of bytes actually read, or -1 in case of an
1084  * error.
1085  **/
1086 ssize_t rawexpread(off_t a, char *buf, size_t len, CLIENT *client) {
1087         int fhandle;
1088         off_t foffset;
1089         size_t maxbytes;
1090
1091         if(get_filepos(client->export, a, &fhandle, &foffset, &maxbytes))
1092                 return -1;
1093         if(maxbytes && len > maxbytes)
1094                 len = maxbytes;
1095
1096         DEBUG4("(READ from fd %d offset %llu len %u), ", fhandle, foffset, len);
1097
1098         myseek(fhandle, foffset);
1099         return read(fhandle, buf, len);
1100 }
1101
1102 /**
1103  * Call rawexpread repeatedly until all data has been read.
1104  * @return 0 on success, nonzero on failure
1105  **/
1106 int rawexpread_fully(off_t a, char *buf, size_t len, CLIENT *client) {
1107         ssize_t ret=0;
1108
1109         while(len > 0 && (ret=rawexpread(a, buf, len, client)) > 0 ) {
1110                 a += ret;
1111                 buf += ret;
1112                 len -= ret;
1113         }
1114         return (ret < 0 || len != 0);
1115 }
1116
1117 /**
1118  * Read an amount of bytes at a given offset from the right file. This
1119  * abstracts the read-side of the copyonwrite stuff, and calls
1120  * rawexpread() with the right parameters to do the actual work.
1121  * @param a The offset where the read should start
1122  * @param buf A buffer to read into
1123  * @param len The size of buf
1124  * @param client The client we're going to read for
1125  * @return 0 on success, nonzero on failure
1126  **/
1127 int expread(off_t a, char *buf, size_t len, CLIENT *client) {
1128         off_t rdlen, offset;
1129         off_t mapcnt, mapl, maph, pagestart;
1130
1131         if (!(client->server->flags & F_COPYONWRITE))
1132                 return(rawexpread_fully(a, buf, len, client));
1133         DEBUG3("Asked to read %d bytes at %llu.\n", len, (unsigned long long)a);
1134
1135         mapl=a/DIFFPAGESIZE; maph=(a+len-1)/DIFFPAGESIZE;
1136
1137         for (mapcnt=mapl;mapcnt<=maph;mapcnt++) {
1138                 pagestart=mapcnt*DIFFPAGESIZE;
1139                 offset=a-pagestart;
1140                 rdlen=(0<DIFFPAGESIZE-offset && len<(size_t)(DIFFPAGESIZE-offset)) ?
1141                         len : (size_t)DIFFPAGESIZE-offset;
1142                 if (client->difmap[mapcnt]!=(u32)(-1)) { /* the block is already there */
1143                         DEBUG3("Page %llu is at %lu\n", (unsigned long long)mapcnt,
1144                                (unsigned long)(client->difmap[mapcnt]));
1145                         myseek(client->difffile, client->difmap[mapcnt]*DIFFPAGESIZE+offset);
1146                         if (read(client->difffile, buf, rdlen) != rdlen) return -1;
1147                 } else { /* the block is not there */
1148                         DEBUG2("Page %llu is not here, we read the original one\n",
1149                                (unsigned long long)mapcnt);
1150                         if(rawexpread_fully(a, buf, rdlen, client)) return -1;
1151                 }
1152                 len-=rdlen; a+=rdlen; buf+=rdlen;
1153         }
1154         return 0;
1155 }
1156
1157 /**
1158  * Write an amount of bytes at a given offset to the right file. This
1159  * abstracts the write-side of the copyonwrite option, and calls
1160  * rawexpwrite() with the right parameters to do the actual work.
1161  *
1162  * @param a The offset where the write should start
1163  * @param buf The buffer to write from
1164  * @param len The length of buf
1165  * @param client The client we're going to write for.
1166  * @return 0 on success, nonzero on failure
1167  **/
1168 int expwrite(off_t a, char *buf, size_t len, CLIENT *client) {
1169         char pagebuf[DIFFPAGESIZE];
1170         off_t mapcnt,mapl,maph;
1171         off_t wrlen,rdlen;
1172         off_t pagestart;
1173         off_t offset;
1174
1175         if (!(client->server->flags & F_COPYONWRITE))
1176                 return(rawexpwrite_fully(a, buf, len, client));
1177         DEBUG3("Asked to write %d bytes at %llu.\n", len, (unsigned long long)a);
1178
1179         mapl=a/DIFFPAGESIZE ; maph=(a+len-1)/DIFFPAGESIZE ;
1180
1181         for (mapcnt=mapl;mapcnt<=maph;mapcnt++) {
1182                 pagestart=mapcnt*DIFFPAGESIZE ;
1183                 offset=a-pagestart ;
1184                 wrlen=(0<DIFFPAGESIZE-offset && len<(size_t)(DIFFPAGESIZE-offset)) ?
1185                         len : (size_t)DIFFPAGESIZE-offset;
1186
1187                 if (client->difmap[mapcnt]!=(u32)(-1)) { /* the block is already there */
1188                         DEBUG3("Page %llu is at %lu\n", (unsigned long long)mapcnt,
1189                                (unsigned long)(client->difmap[mapcnt])) ;
1190                         myseek(client->difffile,
1191                                         client->difmap[mapcnt]*DIFFPAGESIZE+offset);
1192                         if (write(client->difffile, buf, wrlen) != wrlen) return -1 ;
1193                 } else { /* the block is not there */
1194                         myseek(client->difffile,client->difffilelen*DIFFPAGESIZE) ;
1195                         client->difmap[mapcnt]=(client->server->flags&F_SPARSE)?mapcnt:client->difffilelen++;
1196                         DEBUG3("Page %llu is not here, we put it at %lu\n",
1197                                (unsigned long long)mapcnt,
1198                                (unsigned long)(client->difmap[mapcnt]));
1199                         rdlen=DIFFPAGESIZE ;
1200                         if (rawexpread_fully(pagestart, pagebuf, rdlen, client))
1201                                 return -1;
1202                         memcpy(pagebuf+offset,buf,wrlen) ;
1203                         if (write(client->difffile, pagebuf, DIFFPAGESIZE) !=
1204                                         DIFFPAGESIZE)
1205                                 return -1;
1206                 }
1207                 len-=wrlen ; a+=wrlen ; buf+=wrlen ;
1208         }
1209         return 0;
1210 }
1211
1212 /**
1213  * Do the initial negotiation.
1214  *
1215  * @param client The client we're negotiating with.
1216  **/
1217 CLIENT* negotiate(int net, CLIENT *client, GArray* servers) {
1218         char zeros[128];
1219         uint64_t size_host;
1220         uint32_t flags = NBD_FLAG_HAS_FLAGS;
1221         uint16_t smallflags = 0;
1222         uint64_t magic;
1223
1224         memset(zeros, '\0', sizeof(zeros));
1225         if(!client || !client->modern) {
1226                 if (write(net, INIT_PASSWD, 8) < 0) {
1227                         err_nonfatal("Negotiation failed: %m");
1228                         if(client)
1229                                 exit(EXIT_FAILURE);
1230                 }
1231                 magic = htonll(opts_magic);
1232                 if (write(net, &magic, sizeof(magic)) < 0) {
1233                         err_nonfatal("Negotiation failed: %m");
1234                         if(client)
1235                                 exit(EXIT_FAILURE);
1236                 }
1237         }
1238         if(!client) {
1239                 uint64_t reserved;
1240                 uint32_t opt;
1241                 uint64_t namelen;
1242                 char* name;
1243                 int i;
1244
1245                 if(!servers)
1246                         err("programmer error");
1247                 write(net, &smallflags, sizeof(uint16_t));
1248                 read(net, &reserved, sizeof(reserved));
1249                 read(net, &magic, sizeof(magic));
1250                 magic = ntohll(magic);
1251                 if(magic != cliserv_magic) {
1252                         close(net);
1253                         return NULL;
1254                 }
1255                 read(net, &opt, sizeof(opt));
1256                 opt = ntohl(opt);
1257                 if(opt != NBD_OPT_EXPORT_NAME) {
1258                         close(net);
1259                         return NULL;
1260                 }
1261                 read(net, &namelen, sizeof(namelen));
1262                 namelen = ntohll(namelen);
1263                 name = malloc(namelen+1);
1264                 name[namelen+1]=0;
1265                 read(net, &name, namelen);
1266                 for(i=0; i<servers->len; i++) {
1267                         SERVER* serve = &(g_array_index(servers, SERVER, i));
1268                         if(!strcmp(serve->servename, name)) {
1269                                 CLIENT* client = g_new0(CLIENT, 1);
1270                                 client->server = serve;
1271                                 client->exportsize = OFFT_MAX;
1272                                 client->net = net;
1273                                 client->modern = TRUE;
1274                                 return client;
1275                         }
1276                 }
1277         }
1278         size_host = htonll((u64)(client->exportsize));
1279         if (write(net, &size_host, 8) < 0)
1280                 err("Negotiation failed: %m");
1281         if (client->server->flags & F_READONLY)
1282                 flags |= NBD_FLAG_READ_ONLY;
1283         flags = htonl(flags);
1284         if (write(client->net, &flags, 4) < 0)
1285                 err("Negotiation failed: %m");
1286         if (write(client->net, zeros, 124) < 0)
1287                 err("Negotiation failed: %m");
1288         return NULL;
1289 }
1290
1291 /** sending macro. */
1292 #define SEND(net,reply) writeit( net, &reply, sizeof( reply ));
1293 /** error macro. */
1294 #define ERROR(client,reply,errcode) { reply.error = htonl(errcode); SEND(client->net,reply); reply.error = 0; }
1295 /**
1296  * Serve a file to a single client.
1297  *
1298  * @todo This beast needs to be split up in many tiny little manageable
1299  * pieces. Preferably with a chainsaw.
1300  *
1301  * @param client The client we're going to serve to.
1302  * @return when the client disconnects
1303  **/
1304 int mainloop(CLIENT *client) {
1305         struct nbd_request request;
1306         struct nbd_reply reply;
1307         gboolean go_on=TRUE;
1308 #ifdef DODBG
1309         int i = 0;
1310 #endif
1311         negotiate(client->net, client, NULL);
1312         DEBUG("Entering request loop!\n");
1313         reply.magic = htonl(NBD_REPLY_MAGIC);
1314         reply.error = 0;
1315         while (go_on) {
1316                 char buf[BUFSIZE];
1317                 size_t len;
1318 #ifdef DODBG
1319                 i++;
1320                 printf("%d: ", i);
1321 #endif
1322                 readit(client->net, &request, sizeof(request));
1323                 request.from = ntohll(request.from);
1324                 request.type = ntohl(request.type);
1325
1326                 if (request.type==NBD_CMD_DISC) {
1327                         msg2(LOG_INFO, "Disconnect request received.");
1328                         if (client->server->flags & F_COPYONWRITE) {
1329                                 if (client->difmap) g_free(client->difmap) ;
1330                                 close(client->difffile);
1331                                 unlink(client->difffilename);
1332                                 free(client->difffilename);
1333                         }
1334                         go_on=FALSE;
1335                         continue;
1336                 }
1337
1338                 len = ntohl(request.len);
1339
1340                 if (request.magic != htonl(NBD_REQUEST_MAGIC))
1341                         err("Not enough magic.");
1342                 if (len > BUFSIZE + sizeof(struct nbd_reply))
1343                         err("Request too big!");
1344 #ifdef DODBG
1345                 printf("%s from %llu (%llu) len %d, ", request.type ? "WRITE" :
1346                                 "READ", (unsigned long long)request.from,
1347                                 (unsigned long long)request.from / 512, len);
1348 #endif
1349                 memcpy(reply.handle, request.handle, sizeof(reply.handle));
1350                 if ((request.from + len) > (OFFT_MAX)) {
1351                         DEBUG("[Number too large!]");
1352                         ERROR(client, reply, EINVAL);
1353                         continue;
1354                 }
1355
1356                 if (((ssize_t)((off_t)request.from + len) > client->exportsize)) {
1357                         DEBUG("[RANGE!]");
1358                         ERROR(client, reply, EINVAL);
1359                         continue;
1360                 }
1361
1362                 if (request.type==NBD_CMD_WRITE) {
1363                         DEBUG("wr: net->buf, ");
1364                         readit(client->net, buf, len);
1365                         DEBUG("buf->exp, ");
1366                         if ((client->server->flags & F_READONLY) ||
1367                             (client->server->flags & F_AUTOREADONLY)) {
1368                                 DEBUG("[WRITE to READONLY!]");
1369                                 ERROR(client, reply, EPERM);
1370                                 continue;
1371                         }
1372                         if (expwrite(request.from, buf, len, client)) {
1373                                 DEBUG("Write failed: %m" );
1374                                 ERROR(client, reply, errno);
1375                                 continue;
1376                         }
1377                         SEND(client->net, reply);
1378                         DEBUG("OK!\n");
1379                         continue;
1380                 }
1381                 /* READ */
1382
1383                 DEBUG("exp->buf, ");
1384                 if (expread(request.from, buf + sizeof(struct nbd_reply), len, client)) {
1385                         DEBUG("Read failed: %m");
1386                         ERROR(client, reply, errno);
1387                         continue;
1388                 }
1389
1390                 DEBUG("buf->net, ");
1391                 memcpy(buf, &reply, sizeof(struct nbd_reply));
1392                 writeit(client->net, buf, len + sizeof(struct nbd_reply));
1393                 DEBUG("OK!\n");
1394         }
1395         return 0;
1396 }
1397
1398 /**
1399  * Set up client export array, which is an array of FILE_INFO.
1400  * Also, split a single exportfile into multiple ones, if that was asked.
1401  * @param client information on the client which we want to setup export for
1402  **/
1403 void setupexport(CLIENT* client) {
1404         int i;
1405         off_t laststartoff = 0, lastsize = 0;
1406         int multifile = (client->server->flags & F_MULTIFILE);
1407
1408         client->export = g_array_new(TRUE, TRUE, sizeof(FILE_INFO));
1409
1410         /* If multi-file, open as many files as we can.
1411          * If not, open exactly one file.
1412          * Calculate file sizes as we go to get total size. */
1413         for(i=0; ; i++) {
1414                 FILE_INFO fi;
1415                 gchar *tmpname;
1416                 gchar* error_string;
1417                 mode_t mode = (client->server->flags & F_READONLY) ? O_RDONLY : O_RDWR;
1418
1419                 if(multifile) {
1420                         tmpname=g_strdup_printf("%s.%d", client->exportname, i);
1421                 } else {
1422                         tmpname=g_strdup(client->exportname);
1423                 }
1424                 DEBUG2( "Opening %s\n", tmpname );
1425                 fi.fhandle = open(tmpname, mode);
1426                 if(fi.fhandle == -1 && mode == O_RDWR) {
1427                         /* Try again because maybe media was read-only */
1428                         fi.fhandle = open(tmpname, O_RDONLY);
1429                         if(fi.fhandle != -1) {
1430                                 /* Opening the base file in copyonwrite mode is
1431                                  * okay */
1432                                 if(!(client->server->flags & F_COPYONWRITE)) {
1433                                         client->server->flags |= F_AUTOREADONLY;
1434                                         client->server->flags |= F_READONLY;
1435                                 }
1436                         }
1437                 }
1438                 if(fi.fhandle == -1) {
1439                         if(multifile && i>0)
1440                                 break;
1441                         error_string=g_strdup_printf(
1442                                 "Could not open exported file %s: %%m",
1443                                 tmpname);
1444                         err(error_string);
1445                 }
1446                 fi.startoff = laststartoff + lastsize;
1447                 g_array_append_val(client->export, fi);
1448                 g_free(tmpname);
1449
1450                 /* Starting offset and size of this file will be used to
1451                  * calculate starting offset of next file */
1452                 laststartoff = fi.startoff;
1453                 lastsize = size_autodetect(fi.fhandle);
1454
1455                 if(!multifile)
1456                         break;
1457         }
1458
1459         /* Set export size to total calculated size */
1460         client->exportsize = laststartoff + lastsize;
1461
1462         /* Export size may be overridden */
1463         if(client->server->expected_size) {
1464                 /* desired size must be <= total calculated size */
1465                 if(client->server->expected_size > client->exportsize) {
1466                         err("Size of exported file is too big\n");
1467                 }
1468
1469                 client->exportsize = client->server->expected_size;
1470         }
1471
1472         msg3(LOG_INFO, "Size of exported file/device is %llu", (unsigned long long)client->exportsize);
1473         if(multifile) {
1474                 msg3(LOG_INFO, "Total number of files: %d", i);
1475         }
1476 }
1477
1478 int copyonwrite_prepare(CLIENT* client) {
1479         off_t i;
1480         if ((client->difffilename = malloc(1024))==NULL)
1481                 err("Failed to allocate string for diff file name");
1482         snprintf(client->difffilename, 1024, "%s-%s-%d.diff",client->exportname,client->clientname,
1483                 (int)getpid()) ;
1484         client->difffilename[1023]='\0';
1485         msg3(LOG_INFO,"About to create map and diff file %s",client->difffilename) ;
1486         client->difffile=open(client->difffilename,O_RDWR | O_CREAT | O_TRUNC,0600) ;
1487         if (client->difffile<0) err("Could not create diff file (%m)") ;
1488         if ((client->difmap=calloc(client->exportsize/DIFFPAGESIZE,sizeof(u32)))==NULL)
1489                 err("Could not allocate memory") ;
1490         for (i=0;i<client->exportsize/DIFFPAGESIZE;i++) client->difmap[i]=(u32)-1 ;
1491
1492         return 0;
1493 }
1494
1495 /**
1496  * Run a command. This is used for the ``prerun'' and ``postrun'' config file
1497  * options
1498  *
1499  * @param command the command to be ran. Read from the config file
1500  * @param file the file name we're about to export
1501  **/
1502 int do_run(gchar* command, gchar* file) {
1503         gchar* cmd;
1504         int retval=0;
1505
1506         if(command && *command) {
1507                 cmd = g_strdup_printf(command, file);
1508                 retval=system(cmd);
1509                 g_free(cmd);
1510         }
1511         return retval;
1512 }
1513
1514 /**
1515  * Serve a connection.
1516  *
1517  * @todo allow for multithreading, perhaps use libevent. Not just yet, though;
1518  * follow the road map.
1519  *
1520  * @param client a connected client
1521  **/
1522 void serveconnection(CLIENT *client) {
1523         if(do_run(client->server->prerun, client->exportname)) {
1524                 exit(EXIT_FAILURE);
1525         }
1526         setupexport(client);
1527
1528         if (client->server->flags & F_COPYONWRITE) {
1529                 copyonwrite_prepare(client);
1530         }
1531
1532         setmysockopt(client->net);
1533
1534         mainloop(client);
1535         do_run(client->server->postrun, client->exportname);
1536 }
1537
1538 /**
1539  * Find the name of the file we have to serve. This will use g_strdup_printf
1540  * to put the IP address of the client inside a filename containing
1541  * "%s" (in the form as specified by the "virtstyle" option). That name
1542  * is then written to client->exportname.
1543  *
1544  * @param net A socket connected to an nbd client
1545  * @param client information about the client. The IP address in human-readable
1546  * format will be written to a new char* buffer, the address of which will be
1547  * stored in client->clientname.
1548  **/
1549 void set_peername(int net, CLIENT *client) {
1550         struct sockaddr_storage addrin;
1551         struct sockaddr_storage netaddr;
1552         struct sockaddr_in  *netaddr4 = NULL;
1553         struct sockaddr_in6 *netaddr6 = NULL;
1554         size_t addrinlen = sizeof( addrin );
1555         struct addrinfo hints;
1556         struct addrinfo *ai = NULL;
1557         char peername[NI_MAXHOST];
1558         char netname[NI_MAXHOST];
1559         char *tmp = NULL;
1560         int i;
1561         int e;
1562         int shift;
1563
1564         if (getpeername(net, (struct sockaddr *) &addrin, (socklen_t *)&addrinlen) < 0)
1565                 err("getsockname failed: %m");
1566
1567         getnameinfo((struct sockaddr *)&addrin, (socklen_t)addrinlen,
1568                 peername, sizeof (peername), NULL, 0, NI_NUMERICHOST);
1569
1570         memset(&hints, '\0', sizeof (hints));
1571         hints.ai_flags = AI_ADDRCONFIG;
1572         e = getaddrinfo(peername, NULL, &hints, &ai);
1573
1574         if(e != 0) {
1575                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1576                 freeaddrinfo(ai);
1577                 return;
1578         }
1579
1580         switch(client->server->virtstyle) {
1581                 case VIRT_NONE:
1582                         client->exportname=g_strdup(client->server->exportname);
1583                         break;
1584                 case VIRT_IPHASH:
1585                         for(i=0;i<strlen(peername);i++) {
1586                                 if(peername[i]=='.') {
1587                                         peername[i]='/';
1588                                 }
1589                         }
1590                 case VIRT_IPLIT:
1591                         client->exportname=g_strdup_printf(client->server->exportname, peername);
1592                         break;
1593                 case VIRT_CIDR:
1594                         memcpy(&netaddr, &addrin, addrinlen);
1595                         if(ai->ai_family == AF_INET) {
1596                                 netaddr4 = (struct sockaddr_in *)&netaddr;
1597                                 (netaddr4->sin_addr).s_addr>>=32-(client->server->cidrlen);
1598                                 (netaddr4->sin_addr).s_addr<<=32-(client->server->cidrlen);
1599
1600                                 getnameinfo((struct sockaddr *) netaddr4, (socklen_t) addrinlen,
1601                                                         netname, sizeof (netname), NULL, 0, NI_NUMERICHOST);
1602                                 tmp=g_strdup_printf("%s/%s", netname, peername);
1603                         }else if(ai->ai_family == AF_INET6) {
1604                                 netaddr6 = (struct sockaddr_in6 *)&netaddr;
1605
1606                                 shift = 128-(client->server->cidrlen);
1607                                 i = 3;
1608                                 while(shift >= 32) {
1609                                         ((netaddr6->sin6_addr).s6_addr32[i])=0;
1610                                         shift-=32;
1611                                         i--;
1612                                 }
1613                                 (netaddr6->sin6_addr).s6_addr32[i]>>=shift;
1614                                 (netaddr6->sin6_addr).s6_addr32[i]<<=shift;
1615
1616                                 getnameinfo((struct sockaddr *)netaddr6, (socklen_t)addrinlen,
1617                                             netname, sizeof(netname), NULL, 0, NI_NUMERICHOST);
1618                                 tmp=g_strdup_printf("%s/%s", netname, peername);
1619                         }
1620
1621                         if(tmp != NULL)
1622                           client->exportname=g_strdup_printf(client->server->exportname, tmp);
1623
1624                         break;
1625         }
1626
1627         freeaddrinfo(ai);
1628         msg4(LOG_INFO, "connect from %s, assigned file is %s",
1629              peername, client->exportname);
1630         client->clientname=g_strdup(peername);
1631 }
1632
1633 /**
1634  * Destroy a pid_t*
1635  * @param data a pointer to pid_t which should be freed
1636  **/
1637 void destroy_pid_t(gpointer data) {
1638         g_free(data);
1639 }
1640
1641 /**
1642  * Loop through the available servers, and serve them. Never returns.
1643  **/
1644 int serveloop(GArray* servers) {
1645         struct sockaddr_storage addrin;
1646         socklen_t addrinlen=sizeof(addrin);
1647         int i;
1648         int max;
1649         int sock;
1650         fd_set mset;
1651         fd_set rset;
1652
1653         /*
1654          * Set up the master fd_set. The set of descriptors we need
1655          * to select() for never changes anyway and it buys us a *lot*
1656          * of time to only build this once. However, if we ever choose
1657          * to not fork() for clients anymore, we may have to revisit
1658          * this.
1659          */
1660         max=0;
1661         FD_ZERO(&mset);
1662         for(i=0;i<servers->len;i++) {
1663                 sock=(g_array_index(servers, SERVER, i)).socket;
1664                 FD_SET(sock, &mset);
1665                 max=sock>max?sock:max;
1666         }
1667         if(modernsock) {
1668                 FD_SET(modernsock, &mset);
1669                 max=modernsock>max?sock:max;
1670         }
1671         for(;;) {
1672                 CLIENT *client = NULL;
1673                 pid_t *pid;
1674
1675                 memcpy(&rset, &mset, sizeof(fd_set));
1676                 if(select(max+1, &rset, NULL, NULL, NULL)>0) {
1677                         int net = 0;
1678                         SERVER* serve;
1679
1680                         DEBUG("accept, ");
1681                         if(FD_ISSET(modernsock, &rset)) {
1682                                 if((net=accept(modernsock, (struct sockaddr *) &addrin, &addrinlen)) < 0)
1683                                         err("accept: %m");
1684                                 client = negotiate(net, NULL, servers);
1685                                 if(!client) {
1686                                         err_nonfatal("negotiation failed");
1687                                         close(net);
1688                                 }
1689                         }
1690                         for(i=0;i<servers->len && !net;i++) {
1691                                 serve=&(g_array_index(servers, SERVER, i));
1692                                 if(FD_ISSET(serve->socket, &rset)) {
1693                                         if ((net=accept(serve->socket, (struct sockaddr *) &addrin, &addrinlen)) < 0)
1694                                                 err("accept: %m");
1695                                 }
1696                         }
1697                         if(net) {
1698                                 int sock_flags;
1699
1700                                 if((sock_flags = fcntl(net, F_GETFL, 0))==-1) {
1701                                         err("fcntl F_GETFL");
1702                                 }
1703                                 if(fcntl(net, F_SETFL, sock_flags &~O_NONBLOCK)==-1) {
1704                                         err("fcntl F_SETFL ~O_NONBLOCK");
1705                                 }
1706                                 if(!client) {
1707                                         client = g_new0(CLIENT, 1);
1708                                         client->server=serve;
1709                                         client->exportsize=OFFT_MAX;
1710                                         client->net=net;
1711                                 }
1712                                 set_peername(net, client);
1713                                 if (!authorized_client(client)) {
1714                                         msg2(LOG_INFO,"Unauthorized client") ;
1715                                         close(net);
1716                                         continue;
1717                                 }
1718                                 msg2(LOG_INFO,"Authorized client") ;
1719                                 pid=g_malloc(sizeof(pid_t));
1720 #ifndef NOFORK
1721                                 if ((*pid=fork())<0) {
1722                                         msg3(LOG_INFO,"Could not fork (%s)",strerror(errno)) ;
1723                                         close(net);
1724                                         continue;
1725                                 }
1726                                 if (*pid>0) { /* parent */
1727                                         close(net);
1728                                         g_hash_table_insert(children, pid, pid);
1729                                         continue;
1730                                 }
1731                                 /* child */
1732                                 g_hash_table_destroy(children);
1733                                 for(i=0;i<servers->len;i++) {
1734                                         serve=&g_array_index(servers, SERVER, i);
1735                                         close(serve->socket);
1736                                 }
1737                                 /* FALSE does not free the
1738                                 actual data. This is required,
1739                                 because the client has a
1740                                 direct reference into that
1741                                 data, and otherwise we get a
1742                                 segfault... */
1743                                 g_array_free(servers, FALSE);
1744 #endif // NOFORK
1745                                 msg2(LOG_INFO,"Starting to serve");
1746                                 serveconnection(client);
1747                                 exit(EXIT_SUCCESS);
1748                         }
1749                 }
1750         }
1751 }
1752
1753 void dosockopts(int socket) {
1754 #ifndef sun
1755         int yes=1;
1756 #else
1757         char yes='1';
1758 #endif /* sun */
1759         int sock_flags;
1760
1761         /* lose the pesky "Address already in use" error message */
1762         if (setsockopt(socket,SOL_SOCKET,SO_REUSEADDR,&yes,sizeof(int)) == -1) {
1763                 err("setsockopt SO_REUSEADDR");
1764         }
1765         if (setsockopt(socket,SOL_SOCKET,SO_KEEPALIVE,&yes,sizeof(int)) == -1) {
1766                 err("setsockopt SO_KEEPALIVE");
1767         }
1768
1769         /* make the listening socket non-blocking */
1770         if ((sock_flags = fcntl(socket, F_GETFL, 0)) == -1) {
1771                 err("fcntl F_GETFL");
1772         }
1773         if (fcntl(socket, F_SETFL, sock_flags | O_NONBLOCK) == -1) {
1774                 err("fcntl F_SETFL O_NONBLOCK");
1775         }
1776 }
1777
1778 /**
1779  * Connect a server's socket.
1780  *
1781  * @param serve the server we want to connect.
1782  **/
1783 int setup_serve(SERVER *serve) {
1784         struct addrinfo hints;
1785         struct addrinfo *ai = NULL;
1786         gchar *port = NULL;
1787         int e;
1788
1789         if(!do_oldstyle) {
1790                 return serve->servename ? 1 : 0;
1791         }
1792         memset(&hints,'\0',sizeof(hints));
1793         hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG | AI_NUMERICSERV;
1794         hints.ai_socktype = SOCK_STREAM;
1795         hints.ai_family = serve->socket_family;
1796
1797         port = g_strdup_printf ("%d", serve->port);
1798         if (port == NULL)
1799                 return 0;
1800
1801         e = getaddrinfo(serve->listenaddr,port,&hints,&ai);
1802
1803         g_free(port);
1804
1805         if(e != 0) {
1806                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1807                 serve->socket = -1;
1808                 freeaddrinfo(ai);
1809                 exit(EXIT_FAILURE);
1810         }
1811
1812         if(serve->socket_family == AF_UNSPEC)
1813                 serve->socket_family = ai->ai_family;
1814
1815 #ifdef WITH_SDP
1816         if ((serve->flags) && F_SDP) {
1817                 if (ai->ai_family == AF_INET)
1818                         ai->ai_family = AF_INET_SDP;
1819                 else (ai->ai_family == AF_INET6)
1820                         ai->ai_family = AF_INET6_SDP;
1821         }
1822 #endif
1823         if ((serve->socket = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol)) < 0)
1824                 err("socket: %m");
1825
1826         dosockopts(serve->socket);
1827
1828         DEBUG("Waiting for connections... bind, ");
1829         e = bind(serve->socket, ai->ai_addr, ai->ai_addrlen);
1830         if (e != 0 && errno != EADDRINUSE)
1831                 err("bind: %m");
1832         DEBUG("listen, ");
1833         if (listen(serve->socket, 1) < 0)
1834                 err("listen: %m");
1835
1836         freeaddrinfo (ai);
1837         if(serve->servename) {
1838                 return 1;
1839         } else {
1840                 return 0;
1841         }
1842 }
1843
1844 void open_modern(void) {
1845         struct addrinfo hints;
1846         struct addrinfo* ai = NULL;
1847         struct sock_flags;
1848         int e;
1849
1850         memset(&hints, '\0', sizeof(hints));
1851         hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG;
1852         hints.ai_socktype = SOCK_STREAM;
1853         hints.ai_family = AF_UNSPEC;
1854         hints.ai_protocol = IPPROTO_TCP;
1855         e = getaddrinfo(modern_listen, NBD_DEFAULT_PORT, &hints, &ai);
1856         if(e != 0) {
1857                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1858                 exit(EXIT_FAILURE);
1859         }
1860         if((modernsock = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol))<0) {
1861                 err("socket: %m");
1862         }
1863
1864         dosockopts(modernsock);
1865
1866         if(bind(modernsock, ai->ai_addr, ai->ai_addrlen)) {
1867                 err("bind: %m");
1868         }
1869         if(listen(modernsock, 10) <0) {
1870                 err("listen: %m");
1871         }
1872
1873         freeaddrinfo(ai);
1874 }
1875
1876 /**
1877  * Connect our servers.
1878  **/
1879 void setup_servers(GArray* servers) {
1880         int i;
1881         struct sigaction sa;
1882         int want_modern=0;
1883
1884         for(i=0;i<servers->len;i++) {
1885                 want_modern |= setup_serve(&(g_array_index(servers, SERVER, i)));
1886         }
1887         if(want_modern) {
1888                 open_modern();
1889         }
1890         children=g_hash_table_new_full(g_int_hash, g_int_equal, NULL, destroy_pid_t);
1891
1892         sa.sa_handler = sigchld_handler;
1893         sigemptyset(&sa.sa_mask);
1894         sa.sa_flags = SA_RESTART;
1895         if(sigaction(SIGCHLD, &sa, NULL) == -1)
1896                 err("sigaction: %m");
1897         sa.sa_handler = sigterm_handler;
1898         sigemptyset(&sa.sa_mask);
1899         sa.sa_flags = SA_RESTART;
1900         if(sigaction(SIGTERM, &sa, NULL) == -1)
1901                 err("sigaction: %m");
1902 }
1903
1904 /**
1905  * Go daemon (unless we specified at compile time that we didn't want this)
1906  * @param serve the first server of our configuration. If its port is zero,
1907  *      then do not daemonize, because we're doing inetd then. This parameter
1908  *      is only used to create a PID file of the form
1909  *      /var/run/nbd-server.&lt;port&gt;.pid; it's not modified in any way.
1910  **/
1911 #if !defined(NODAEMON) && !defined(NOFORK)
1912 void daemonize(SERVER* serve) {
1913         FILE*pidf;
1914
1915         if(serve && !(serve->port)) {
1916                 return;
1917         }
1918         if(daemon(0,0)<0) {
1919                 err("daemon");
1920         }
1921         if(!*pidftemplate) {
1922                 if(serve) {
1923                         strncpy(pidftemplate, "/var/run/nbd-server.%d.pid", 255);
1924                 } else {
1925                         strncpy(pidftemplate, "/var/run/nbd-server.pid", 255);
1926                 }
1927         }
1928         snprintf(pidfname, 255, pidftemplate, serve ? serve->port : 0);
1929         pidf=fopen(pidfname, "w");
1930         if(pidf) {
1931                 fprintf(pidf,"%d\n", (int)getpid());
1932                 fclose(pidf);
1933         } else {
1934                 perror("fopen");
1935                 fprintf(stderr, "Not fatal; continuing");
1936         }
1937 }
1938 #else
1939 #define daemonize(serve)
1940 #endif /* !defined(NODAEMON) && !defined(NOFORK) */
1941
1942 /*
1943  * Everything beyond this point (in the file) is run in non-daemon mode.
1944  * The stuff above daemonize() isn't.
1945  */
1946
1947 void serve_err(SERVER* serve, const char* msg) G_GNUC_NORETURN;
1948
1949 void serve_err(SERVER* serve, const char* msg) {
1950         g_message("Export of %s on port %d failed:", serve->exportname,
1951                         serve->port);
1952         err(msg);
1953 }
1954
1955 /**
1956  * Set up user-ID and/or group-ID
1957  **/
1958 void dousers(void) {
1959         struct passwd *pw;
1960         struct group *gr;
1961         gchar* str;
1962         if(rungroup) {
1963                 gr=getgrnam(rungroup);
1964                 if(!gr) {
1965                         str = g_strdup_printf("Invalid group name: %s", rungroup);
1966                         err(str);
1967                 }
1968                 if(setgid(gr->gr_gid)<0) {
1969                         err("Could not set GID: %m");
1970                 }
1971         }
1972         if(runuser) {
1973                 pw=getpwnam(runuser);
1974                 if(!pw) {
1975                         str = g_strdup_printf("Invalid user name: %s", runuser);
1976                         err(str);
1977                 }
1978                 if(setuid(pw->pw_uid)<0) {
1979                         err("Could not set UID: %m");
1980                 }
1981         }
1982 }
1983
1984 #ifndef ISSERVER
1985 void glib_message_syslog_redirect(const gchar *log_domain,
1986                                   GLogLevelFlags log_level,
1987                                   const gchar *message,
1988                                   gpointer user_data)
1989 {
1990     int level=LOG_DEBUG;
1991
1992     switch( log_level )
1993     {
1994       case G_LOG_FLAG_FATAL:
1995       case G_LOG_LEVEL_CRITICAL:
1996       case G_LOG_LEVEL_ERROR:
1997         level=LOG_ERR;
1998         break;
1999       case G_LOG_LEVEL_WARNING:
2000         level=LOG_WARNING;
2001         break;
2002       case G_LOG_LEVEL_MESSAGE:
2003       case G_LOG_LEVEL_INFO:
2004         level=LOG_INFO;
2005         break;
2006       case G_LOG_LEVEL_DEBUG:
2007         level=LOG_DEBUG;
2008       default:
2009         level=LOG_ERR;
2010     }
2011     syslog(level, message);
2012 }
2013 #endif
2014
2015 /**
2016  * Main entry point...
2017  **/
2018 int main(int argc, char *argv[]) {
2019         SERVER *serve;
2020         GArray *servers;
2021         GError *err=NULL;
2022
2023         if (sizeof( struct nbd_request )!=28) {
2024                 fprintf(stderr,"Bad size of structure. Alignment problems?\n");
2025                 exit(EXIT_FAILURE) ;
2026         }
2027
2028         memset(pidftemplate, '\0', 256);
2029
2030         logging();
2031         config_file_pos = g_strdup(CFILE);
2032         serve=cmdline(argc, argv);
2033         servers = parse_cfile(config_file_pos, &err);
2034
2035         if(serve) {
2036                 serve->socket_family = AF_UNSPEC;
2037
2038                 append_serve(serve, servers);
2039
2040                 if (!(serve->port)) {
2041                         CLIENT *client;
2042 #ifndef ISSERVER
2043                         /* You really should define ISSERVER if you're going to use
2044                          * inetd mode, but if you don't, closing stdout and stderr
2045                          * (which inetd had connected to the client socket) will let it
2046                          * work. */
2047                         close(1);
2048                         close(2);
2049                         open("/dev/null", O_WRONLY);
2050                         open("/dev/null", O_WRONLY);
2051                         g_log_set_default_handler( glib_message_syslog_redirect, NULL );
2052 #endif
2053                         client=g_malloc(sizeof(CLIENT));
2054                         client->server=serve;
2055                         client->net=0;
2056                         client->exportsize=OFFT_MAX;
2057                         set_peername(0,client);
2058                         serveconnection(client);
2059                         return 0;
2060                 }
2061         }
2062
2063         if(!servers || !servers->len) {
2064                 g_warning("Could not parse config file: %s",
2065                                 err ? err->message : "Unknown error");
2066         }
2067         if(serve) {
2068                 g_warning("Specifying an export on the command line is deprecated.");
2069                 g_warning("Please use a configuration file instead.");
2070         }
2071
2072         if((!serve) && (!servers||!servers->len)) {
2073                 g_message("Nothing to do! Bye!");
2074                 exit(EXIT_FAILURE);
2075         }
2076         daemonize(serve);
2077         setup_servers(servers);
2078         dousers();
2079         serveloop(servers);
2080         return 0 ;
2081 }