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