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         } else {
 519                 do_oldstyle = TRUE;
 520         }
 521         if(do_output) {
 522                 if(!serve) {
 523                         g_critical("Need a complete configuration on the command line to output a config file section!");
 524                         exit(EXIT_FAILURE);
 525                 }
 526                 dump_section(serve, section_header);
 527         }
 528         return serve;
 529 }
 530
 531 /**
 532  * Error codes for config file parsing
 533  **/
 534 typedef enum {
 535         CFILE_NOTFOUND,         /**< The configuration file is not found */
 536         CFILE_MISSING_GENERIC,  /**< The (required) group "generic" is missing */
 537         CFILE_KEY_MISSING,      /**< A (required) key is missing */
 538         CFILE_VALUE_INVALID,    /**< A value is syntactically invalid */
 539         CFILE_VALUE_UNSUPPORTED,/**< A value is not supported in this build */
 540         CFILE_PROGERR,          /**< Programmer error */
 541         CFILE_NO_EXPORTS,       /**< A config file was specified that does not
 542                                      define any exports */
 543         CFILE_INCORRECT_PORT,   /**< The reserved port was specified for an
 544                                      old-style export. */
 545 } CFILE_ERRORS;
 546
 547 /**
 548  * Remove a SERVER from memory. Used from the hash table
 549  **/
 550 void remove_server(gpointer s) {
 551         SERVER *server;
 552
 553         server=(SERVER*)s;
 554         g_free(server->exportname);
 555         if(server->authname)
 556                 g_free(server->authname);
 557         if(server->listenaddr)
 558                 g_free(server->listenaddr);
 559         if(server->prerun)
 560                 g_free(server->prerun);
 561         if(server->postrun)
 562                 g_free(server->postrun);
 563         g_free(server);
 564 }
 565
 566 /**
 567  * duplicate server
 568  * @param s the old server we want to duplicate
 569  * @return new duplicated server
 570  **/
 571 SERVER* dup_serve(SERVER *s) {
 572         SERVER *serve = NULL;
 573
 574         serve=g_new0(SERVER, 1);
 575         if(serve == NULL)
 576                 return NULL;
 577
 578         if(s->exportname)
 579                 serve->exportname = g_strdup(s->exportname);
 580
 581         serve->expected_size = s->expected_size;
 582
 583         if(s->listenaddr)
 584                 serve->listenaddr = g_strdup(s->listenaddr);
 585
 586         serve->port = s->port;
 587
 588         if(s->authname)
 589                 serve->authname = strdup(s->authname);
 590
 591         serve->flags = s->flags;
 592         serve->socket = serve->socket;
 593         serve->socket_family = serve->socket_family;
 594         serve->cidrlen = s->cidrlen;
 595
 596         if(s->prerun)
 597                 serve->prerun = g_strdup(s->prerun);
 598
 599         if(s->postrun)
 600                 serve->postrun = g_strdup(s->postrun);
 601
 602         if(s->servename)
 603                 serve->servename = g_strdup(s->servename);
 604
 605         return serve;
 606 }
 607
 608 /**
 609  * append new server to array
 610  * @param s server
 611  * @param a server array
 612  * @return 0 success, -1 error
 613  */
 614 int append_serve(SERVER *s, GArray *a) {
 615         SERVER *ns = NULL;
 616         struct addrinfo hints;
 617         struct addrinfo *ai = NULL;
 618         struct addrinfo *rp = NULL;
 619         char   host[NI_MAXHOST];
 620         gchar  *port = NULL;
 621         int e;
 622         int ret;
 623
 624         if(!s) {
 625                 err("Invalid parsing server");
 626                 return -1;
 627         }
 628
 629         port = g_strdup_printf("%d", s->port);
 630
 631         memset(&hints,'\0',sizeof(hints));
 632         hints.ai_family = AF_UNSPEC;
 633         hints.ai_socktype = SOCK_STREAM;
 634         hints.ai_flags = AI_ADDRCONFIG | AI_PASSIVE;
 635         hints.ai_protocol = IPPROTO_TCP;
 636
 637         e = getaddrinfo(s->listenaddr, port, &hints, &ai);
 638
 639         if (port)
 640                 g_free(port);
 641
 642         if(e == 0) {
 643                 for (rp = ai; rp != NULL; rp = rp->ai_next) {
 644                         e = getnameinfo(rp->ai_addr, rp->ai_addrlen, host, sizeof(host), NULL, 0, NI_NUMERICHOST);
 645
 646                         if (e != 0) { // error
 647                                 fprintf(stderr, "getnameinfo: %s\n", gai_strerror(e));
 648                                 continue;
 649                         }
 650
 651                         // duplicate server and set listenaddr to resolved IP address
 652                         ns = dup_serve (s);
 653                         if (ns) {
 654                                 ns->listenaddr = g_strdup(host);
 655                                 ns->socket_family = rp->ai_family;
 656                                 g_array_append_val(a, *ns);
 657                                 free(ns);
 658                                 ns = NULL;
 659                         }
 660                 }
 661
 662                 ret = 0;
 663         } else {
 664                 fprintf(stderr, "getaddrinfo failed on listen host/address: %s (%s)\n", s->listenaddr ? s->listenaddr : "any", gai_strerror(e));
 665                 ret = -1;
 666         }
 667
 668         if (ai)
 669                 freeaddrinfo(ai);
 670
 671         return ret;
 672 }
 673
 674 /**
 675  * Parse the config file.
 676  *
 677  * @param f the name of the config file
 678  * @param e a GError. @see CFILE_ERRORS for what error values this function can
 679  *      return.
 680  * @return a Array of SERVER* pointers, If the config file is empty or does not
 681  *      exist, returns an empty GHashTable; if the config file contains an
 682  *      error, returns NULL, and e is set appropriately
 683  **/
 684 GArray* parse_cfile(gchar* f, GError** e) {
 685         const char* DEFAULT_ERROR = "Could not parse %s in group %s: %s";
 686         const char* MISSING_REQUIRED_ERROR = "Could not find required value %s in group %s: %s";
 687         SERVER s;
 688         gchar *virtstyle=NULL;
 689         PARAM lp[] = {
 690                 { "exportname", TRUE,   PARAM_STRING,   NULL, 0 },
 691                 { "port",       TRUE,   PARAM_INT,      NULL, 0 },
 692                 { "authfile",   FALSE,  PARAM_STRING,   NULL, 0 },
 693                 { "filesize",   FALSE,  PARAM_INT,      NULL, 0 },
 694                 { "virtstyle",  FALSE,  PARAM_STRING,   NULL, 0 },
 695                 { "prerun",     FALSE,  PARAM_STRING,   NULL, 0 },
 696                 { "postrun",    FALSE,  PARAM_STRING,   NULL, 0 },
 697                 { "readonly",   FALSE,  PARAM_BOOL,     NULL, F_READONLY },
 698                 { "multifile",  FALSE,  PARAM_BOOL,     NULL, F_MULTIFILE },
 699                 { "copyonwrite", FALSE, PARAM_BOOL,     NULL, F_COPYONWRITE },
 700                 { "sparse_cow", FALSE,  PARAM_BOOL,     NULL, F_SPARSE },
 701                 { "sdp",        FALSE,  PARAM_BOOL,     NULL, F_SDP },
 702                 { "sync",       FALSE,  PARAM_BOOL,     NULL, F_SYNC },
 703                 { "listenaddr", FALSE,  PARAM_STRING,   NULL, 0 },
 704         };
 705         const int lp_size=sizeof(lp)/sizeof(PARAM);
 706         PARAM gp[] = {
 707                 { "user",       FALSE, PARAM_STRING,    &runuser,       0 },
 708                 { "group",      FALSE, PARAM_STRING,    &rungroup,      0 },
 709                 { "oldstyle",   FALSE, PARAM_BOOL,      &do_oldstyle,   1 },
 710                 { "listenaddr", FALSE, PARAM_STRING,    &modern_listen, 0 },
 711         };
 712         PARAM* p=gp;
 713         int p_size=sizeof(gp)/sizeof(PARAM);
 714         GKeyFile *cfile;
 715         GError *err = NULL;
 716         const char *err_msg=NULL;
 717         GQuark errdomain;
 718         GArray *retval=NULL;
 719         gchar **groups;
 720         gboolean value;
 721         gchar* startgroup;
 722         gint i;
 723         gint j;
 724
 725         errdomain = g_quark_from_string("parse_cfile");
 726         cfile = g_key_file_new();
 727         retval = g_array_new(FALSE, TRUE, sizeof(SERVER));
 728         if(!g_key_file_load_from_file(cfile, f, G_KEY_FILE_KEEP_COMMENTS |
 729                         G_KEY_FILE_KEEP_TRANSLATIONS, &err)) {
 730                 g_set_error(e, errdomain, CFILE_NOTFOUND, "Could not open config file %s.", f);
 731                 g_key_file_free(cfile);
 732                 return retval;
 733         }
 734         startgroup = g_key_file_get_start_group(cfile);
 735         if(!startgroup || strcmp(startgroup, "generic")) {
 736                 g_set_error(e, errdomain, CFILE_MISSING_GENERIC, "Config file does not contain the [generic] group!");
 737                 g_key_file_free(cfile);
 738                 return NULL;
 739         }
 740         groups = g_key_file_get_groups(cfile, NULL);
 741         for(i=0;groups[i];i++) {
 742                 memset(&s, '\0', sizeof(SERVER));
 743                 lp[0].target=&(s.exportname);
 744                 lp[1].target=&(s.port);
 745                 lp[2].target=&(s.authname);
 746                 lp[3].target=&(s.expected_size);
 747                 lp[4].target=&(virtstyle);
 748                 lp[5].target=&(s.prerun);
 749                 lp[6].target=&(s.postrun);
 750                 lp[7].target=lp[8].target=lp[9].target=
 751                                 lp[10].target=lp[11].target=
 752                                 lp[12].target=&(s.flags);
 753                 lp[13].target=&(s.listenaddr);
 754
 755                 /* After the [generic] group, start parsing exports */
 756                 if(i==1) {
 757                         p=lp;
 758                         p_size=lp_size;
 759                 }
 760                 for(j=0;j<p_size;j++) {
 761                         g_assert(p[j].target != NULL);
 762                         g_assert(p[j].ptype==PARAM_INT||p[j].ptype==PARAM_STRING||p[j].ptype==PARAM_BOOL);
 763                         switch(p[j].ptype) {
 764                                 case PARAM_INT:
 765                                         *((gint*)p[j].target) =
 766                                                 g_key_file_get_integer(cfile,
 767                                                                 groups[i],
 768                                                                 p[j].paramname,
 769                                                                 &err);
 770                                         break;
 771                                 case PARAM_STRING:
 772                                         *((gchar**)p[j].target) =
 773                                                 g_key_file_get_string(cfile,
 774                                                                 groups[i],
 775                                                                 p[j].paramname,
 776                                                                 &err);
 777                                         break;
 778                                 case PARAM_BOOL:
 779                                         value = g_key_file_get_boolean(cfile,
 780                                                         groups[i],
 781                                                         p[j].paramname, &err);
 782                                         if(!err) {
 783                                                 if(value) {
 784                                                         *((gint*)p[j].target) |= p[j].flagval;
 785                                                 } else {
 786                                                         *((gint*)p[j].target) &= ~(p[j].flagval);
 787                                                 }
 788                                         }
 789                                         break;
 790                         }
 791                         if(!strcmp(p[j].paramname, "port") && !strcmp(p[j].target, NBD_DEFAULT_PORT)) {
 792                                 g_set_error(e, errdomain, CFILE_INCORRECT_PORT, "Config file specifies default port for oldstyle export");
 793                                 g_key_file_free(cfile);
 794                                 return NULL;
 795                         }
 796                         if(err) {
 797                                 if(err->code == G_KEY_FILE_ERROR_KEY_NOT_FOUND) {
 798                                         if(!p[j].required) {
 799                                                 /* Ignore not-found error for optional values */
 800                                                 g_clear_error(&err);
 801                                                 continue;
 802                                         } else {
 803                                                 err_msg = MISSING_REQUIRED_ERROR;
 804                                         }
 805                                 } else {
 806                                         err_msg = DEFAULT_ERROR;
 807                                 }
 808                                 g_set_error(e, errdomain, CFILE_VALUE_INVALID, err_msg, p[j].paramname, groups[i], err->message);
 809                                 g_array_free(retval, TRUE);
 810                                 g_error_free(err);
 811                                 g_key_file_free(cfile);
 812                                 return NULL;
 813                         }
 814                 }
 815                 if(virtstyle) {
 816                         if(!strncmp(virtstyle, "none", 4)) {
 817                                 s.virtstyle=VIRT_NONE;
 818                         } else if(!strncmp(virtstyle, "ipliteral", 9)) {
 819                                 s.virtstyle=VIRT_IPLIT;
 820                         } else if(!strncmp(virtstyle, "iphash", 6)) {
 821                                 s.virtstyle=VIRT_IPHASH;
 822                         } else if(!strncmp(virtstyle, "cidrhash", 8)) {
 823                                 s.virtstyle=VIRT_CIDR;
 824                                 if(strlen(virtstyle)<10) {
 825                                         g_set_error(e, errdomain, CFILE_VALUE_INVALID, "Invalid value %s for parameter virtstyle in group %s: missing length", virtstyle, groups[i]);
 826                                         g_array_free(retval, TRUE);
 827                                         g_key_file_free(cfile);
 828                                         return NULL;
 829                                 }
 830                                 s.cidrlen=strtol(virtstyle+8, NULL, 0);
 831                         } else {
 832                                 g_set_error(e, errdomain, CFILE_VALUE_INVALID, "Invalid value %s for parameter virtstyle in group %s", virtstyle, groups[i]);
 833                                 g_array_free(retval, TRUE);
 834                                 g_key_file_free(cfile);
 835                                 return NULL;
 836                         }
 837                 } else {
 838                         s.virtstyle=VIRT_IPLIT;
 839                 }
 840                 /* Don't need to free this, it's not our string */
 841                 virtstyle=NULL;
 842                 /* Don't append values for the [generic] group */
 843                 if(i>0) {
 844                         s.socket_family = AF_UNSPEC;
 845                         s.servename = groups[i];
 846
 847                         append_serve(&s, retval);
 848                 } else {
 849                         if(!do_oldstyle) {
 850                                 lp[1].required = 0;
 851                         }
 852                 }
 853 #ifndef WITH_SDP
 854                 if(s.flags & F_SDP) {
 855                         g_set_error(e, errdomain, CFILE_VALUE_UNSUPPORTED, "This nbd-server was built without support for SDP, yet group %s uses it", groups[i]);
 856                         g_array_free(retval, TRUE);
 857                         g_key_file_free(cfile);
 858                         return NULL;
 859                 }
 860 #endif
 861         }
 862         if(i==1) {
 863                 g_set_error(e, errdomain, CFILE_NO_EXPORTS, "The config file does not specify any exports");
 864         }
 865         g_key_file_free(cfile);
 866         return retval;
 867 }
 868
 869 /**
 870  * Signal handler for SIGCHLD
 871  * @param s the signal we're handling (must be SIGCHLD, or something
 872  * is severely wrong)
 873  **/
 874 void sigchld_handler(int s) {
 875         int status;
 876         int* i;
 877         pid_t pid;
 878
 879         while((pid=waitpid(-1, &status, WNOHANG)) > 0) {
 880                 if(WIFEXITED(status)) {
 881                         msg3(LOG_INFO, "Child exited with %d", WEXITSTATUS(status));
 882                 }
 883                 i=g_hash_table_lookup(children, &pid);
 884                 if(!i) {
 885                         msg3(LOG_INFO, "SIGCHLD received for an unknown child with PID %ld", (long)pid);
 886                 } else {
 887                         DEBUG2("Removing %d from the list of children", pid);
 888                         g_hash_table_remove(children, &pid);
 889                 }
 890         }
 891 }
 892
 893 /**
 894  * Kill a child. Called from sigterm_handler::g_hash_table_foreach.
 895  *
 896  * @param key the key
 897  * @param value the value corresponding to the above key
 898  * @param user_data a pointer which we always set to 1, so that we know what
 899  * will happen next.
 900  **/
 901 void killchild(gpointer key, gpointer value, gpointer user_data) {
 902         pid_t *pid=value;
 903         int *parent=user_data;
 904
 905         kill(*pid, SIGTERM);
 906         *parent=1;
 907 }
 908
 909 /**
 910  * Handle SIGTERM and dispatch it to our children
 911  * @param s the signal we're handling (must be SIGTERM, or something
 912  * is severely wrong).
 913  **/
 914 void sigterm_handler(int s) {
 915         int parent=0;
 916
 917         g_hash_table_foreach(children, killchild, &parent);
 918
 919         if(parent) {
 920                 unlink(pidfname);
 921         }
 922
 923         exit(EXIT_SUCCESS);
 924 }
 925
 926 /**
 927  * Detect the size of a file.
 928  *
 929  * @param fhandle An open filedescriptor
 930  * @return the size of the file, or OFFT_MAX if detection was
 931  * impossible.
 932  **/
 933 off_t size_autodetect(int fhandle) {
 934         off_t es;
 935         u64 bytes;
 936         struct stat stat_buf;
 937         int error;
 938
 939 #ifdef HAVE_SYS_MOUNT_H
 940 #ifdef HAVE_SYS_IOCTL_H
 941 #ifdef BLKGETSIZE64
 942         DEBUG("looking for export size with ioctl BLKGETSIZE64\n");
 943         if (!ioctl(fhandle, BLKGETSIZE64, &bytes) && bytes) {
 944                 return (off_t)bytes;
 945         }
 946 #endif /* BLKGETSIZE64 */
 947 #endif /* HAVE_SYS_IOCTL_H */
 948 #endif /* HAVE_SYS_MOUNT_H */
 949
 950         DEBUG("looking for fhandle size with fstat\n");
 951         stat_buf.st_size = 0;
 952         error = fstat(fhandle, &stat_buf);
 953         if (!error) {
 954                 if(stat_buf.st_size > 0)
 955                         return (off_t)stat_buf.st_size;
 956         } else {
 957                 err("fstat failed: %m");
 958         }
 959
 960         DEBUG("looking for fhandle size with lseek SEEK_END\n");
 961         es = lseek(fhandle, (off_t)0, SEEK_END);
 962         if (es > ((off_t)0)) {
 963                 return es;
 964         } else {
 965                 DEBUG2("lseek failed: %d", errno==EBADF?1:(errno==ESPIPE?2:(errno==EINVAL?3:4)));
 966         }
 967
 968         err("Could not find size of exported block device: %m");
 969         return OFFT_MAX;
 970 }
 971
 972 /**
 973  * Get the file handle and offset, given an export offset.
 974  *
 975  * @param export An array of export files
 976  * @param a The offset to get corresponding file/offset for
 977  * @param fhandle [out] File descriptor
 978  * @param foffset [out] Offset into fhandle
 979  * @param maxbytes [out] Tells how many bytes can be read/written
 980  * from fhandle starting at foffset (0 if there is no limit)
 981  * @return 0 on success, -1 on failure
 982  **/
 983 int get_filepos(GArray* export, off_t a, int* fhandle, off_t* foffset, size_t* maxbytes ) {
 984         /* Negative offset not allowed */
 985         if(a < 0)
 986                 return -1;
 987
 988         /* Binary search for last file with starting offset <= a */
 989         FILE_INFO fi;
 990         int start = 0;
 991         int end = export->len - 1;
 992         while( start <= end ) {
 993                 int mid = (start + end) / 2;
 994                 fi = g_array_index(export, FILE_INFO, mid);
 995                 if( fi.startoff < a ) {
 996                         start = mid + 1;
 997                 } else if( fi.startoff > a ) {
 998                         end = mid - 1;
 999                 } else {
1000                         start = end = mid;
1001                         break;
1002                 }
1003         }
1004
1005         /* end should never go negative, since first startoff is 0 and a >= 0 */
1006         g_assert(end >= 0);
1007
1008         fi = g_array_index(export, FILE_INFO, end);
1009         *fhandle = fi.fhandle;
1010         *foffset = a - fi.startoff;
1011         *maxbytes = 0;
1012         if( end+1 < export->len ) {
1013                 FILE_INFO fi_next = g_array_index(export, FILE_INFO, end+1);
1014                 *maxbytes = fi_next.startoff - a;
1015         }
1016
1017         return 0;
1018 }
1019
1020 /**
1021  * seek to a position in a file, with error handling.
1022  * @param handle a filedescriptor
1023  * @param a position to seek to
1024  * @todo get rid of this; lastpoint is a global variable right now, but it
1025  * shouldn't be. If we pass it on as a parameter, that makes things a *lot*
1026  * easier.
1027  **/
1028 void myseek(int handle,off_t a) {
1029         if (lseek(handle, a, SEEK_SET) < 0) {
1030                 err("Can not seek locally!\n");
1031         }
1032 }
1033
1034 /**
1035  * Write an amount of bytes at a given offset to the right file. This
1036  * abstracts the write-side of the multiple file option.
1037  *
1038  * @param a The offset where the write should start
1039  * @param buf The buffer to write from
1040  * @param len The length of buf
1041  * @param client The client we're serving for
1042  * @return The number of bytes actually written, or -1 in case of an error
1043  **/
1044 ssize_t rawexpwrite(off_t a, char *buf, size_t len, CLIENT *client) {
1045         int fhandle;
1046         off_t foffset;
1047         size_t maxbytes;
1048         ssize_t retval;
1049
1050         if(get_filepos(client->export, a, &fhandle, &foffset, &maxbytes))
1051                 return -1;
1052         if(maxbytes && len > maxbytes)
1053                 len = maxbytes;
1054
1055         DEBUG4("(WRITE to fd %d offset %llu len %u), ", fhandle, foffset, len);
1056
1057         myseek(fhandle, foffset);
1058         retval = write(fhandle, buf, len);
1059         if(client->server->flags & F_SYNC) {
1060                 fsync(fhandle);
1061         }
1062         return retval;
1063 }
1064
1065 /**
1066  * Call rawexpwrite repeatedly until all data has been written.
1067  * @return 0 on success, nonzero on failure
1068  **/
1069 int rawexpwrite_fully(off_t a, char *buf, size_t len, CLIENT *client) {
1070         ssize_t ret=0;
1071
1072         while(len > 0 && (ret=rawexpwrite(a, buf, len, client)) > 0 ) {
1073                 a += ret;
1074                 buf += ret;
1075                 len -= ret;
1076         }
1077         return (ret < 0 || len != 0);
1078 }
1079
1080 /**
1081  * Read an amount of bytes at a given offset from the right file. This
1082  * abstracts the read-side of the multiple files option.
1083  *
1084  * @param a The offset where the read should start
1085  * @param buf A buffer to read into
1086  * @param len The size of buf
1087  * @param client The client we're serving for
1088  * @return The number of bytes actually read, or -1 in case of an
1089  * error.
1090  **/
1091 ssize_t rawexpread(off_t a, char *buf, size_t len, CLIENT *client) {
1092         int fhandle;
1093         off_t foffset;
1094         size_t maxbytes;
1095
1096         if(get_filepos(client->export, a, &fhandle, &foffset, &maxbytes))
1097                 return -1;
1098         if(maxbytes && len > maxbytes)
1099                 len = maxbytes;
1100
1101         DEBUG4("(READ from fd %d offset %llu len %u), ", fhandle, foffset, len);
1102
1103         myseek(fhandle, foffset);
1104         return read(fhandle, buf, len);
1105 }
1106
1107 /**
1108  * Call rawexpread repeatedly until all data has been read.
1109  * @return 0 on success, nonzero on failure
1110  **/
1111 int rawexpread_fully(off_t a, char *buf, size_t len, CLIENT *client) {
1112         ssize_t ret=0;
1113
1114         while(len > 0 && (ret=rawexpread(a, buf, len, client)) > 0 ) {
1115                 a += ret;
1116                 buf += ret;
1117                 len -= ret;
1118         }
1119         return (ret < 0 || len != 0);
1120 }
1121
1122 /**
1123  * Read an amount of bytes at a given offset from the right file. This
1124  * abstracts the read-side of the copyonwrite stuff, and calls
1125  * rawexpread() with the right parameters to do the actual work.
1126  * @param a The offset where the read should start
1127  * @param buf A buffer to read into
1128  * @param len The size of buf
1129  * @param client The client we're going to read for
1130  * @return 0 on success, nonzero on failure
1131  **/
1132 int expread(off_t a, char *buf, size_t len, CLIENT *client) {
1133         off_t rdlen, offset;
1134         off_t mapcnt, mapl, maph, pagestart;
1135
1136         if (!(client->server->flags & F_COPYONWRITE))
1137                 return(rawexpread_fully(a, buf, len, client));
1138         DEBUG3("Asked to read %d bytes at %llu.\n", len, (unsigned long long)a);
1139
1140         mapl=a/DIFFPAGESIZE; maph=(a+len-1)/DIFFPAGESIZE;
1141
1142         for (mapcnt=mapl;mapcnt<=maph;mapcnt++) {
1143                 pagestart=mapcnt*DIFFPAGESIZE;
1144                 offset=a-pagestart;
1145                 rdlen=(0<DIFFPAGESIZE-offset && len<(size_t)(DIFFPAGESIZE-offset)) ?
1146                         len : (size_t)DIFFPAGESIZE-offset;
1147                 if (client->difmap[mapcnt]!=(u32)(-1)) { /* the block is already there */
1148                         DEBUG3("Page %llu is at %lu\n", (unsigned long long)mapcnt,
1149                                (unsigned long)(client->difmap[mapcnt]));
1150                         myseek(client->difffile, client->difmap[mapcnt]*DIFFPAGESIZE+offset);
1151                         if (read(client->difffile, buf, rdlen) != rdlen) return -1;
1152                 } else { /* the block is not there */
1153                         DEBUG2("Page %llu is not here, we read the original one\n",
1154                                (unsigned long long)mapcnt);
1155                         if(rawexpread_fully(a, buf, rdlen, client)) return -1;
1156                 }
1157                 len-=rdlen; a+=rdlen; buf+=rdlen;
1158         }
1159         return 0;
1160 }
1161
1162 /**
1163  * Write an amount of bytes at a given offset to the right file. This
1164  * abstracts the write-side of the copyonwrite option, and calls
1165  * rawexpwrite() with the right parameters to do the actual work.
1166  *
1167  * @param a The offset where the write should start
1168  * @param buf The buffer to write from
1169  * @param len The length of buf
1170  * @param client The client we're going to write for.
1171  * @return 0 on success, nonzero on failure
1172  **/
1173 int expwrite(off_t a, char *buf, size_t len, CLIENT *client) {
1174         char pagebuf[DIFFPAGESIZE];
1175         off_t mapcnt,mapl,maph;
1176         off_t wrlen,rdlen;
1177         off_t pagestart;
1178         off_t offset;
1179
1180         if (!(client->server->flags & F_COPYONWRITE))
1181                 return(rawexpwrite_fully(a, buf, len, client));
1182         DEBUG3("Asked to write %d bytes at %llu.\n", len, (unsigned long long)a);
1183
1184         mapl=a/DIFFPAGESIZE ; maph=(a+len-1)/DIFFPAGESIZE ;
1185
1186         for (mapcnt=mapl;mapcnt<=maph;mapcnt++) {
1187                 pagestart=mapcnt*DIFFPAGESIZE ;
1188                 offset=a-pagestart ;
1189                 wrlen=(0<DIFFPAGESIZE-offset && len<(size_t)(DIFFPAGESIZE-offset)) ?
1190                         len : (size_t)DIFFPAGESIZE-offset;
1191
1192                 if (client->difmap[mapcnt]!=(u32)(-1)) { /* the block is already there */
1193                         DEBUG3("Page %llu is at %lu\n", (unsigned long long)mapcnt,
1194                                (unsigned long)(client->difmap[mapcnt])) ;
1195                         myseek(client->difffile,
1196                                         client->difmap[mapcnt]*DIFFPAGESIZE+offset);
1197                         if (write(client->difffile, buf, wrlen) != wrlen) return -1 ;
1198                 } else { /* the block is not there */
1199                         myseek(client->difffile,client->difffilelen*DIFFPAGESIZE) ;
1200                         client->difmap[mapcnt]=(client->server->flags&F_SPARSE)?mapcnt:client->difffilelen++;
1201                         DEBUG3("Page %llu is not here, we put it at %lu\n",
1202                                (unsigned long long)mapcnt,
1203                                (unsigned long)(client->difmap[mapcnt]));
1204                         rdlen=DIFFPAGESIZE ;
1205                         if (rawexpread_fully(pagestart, pagebuf, rdlen, client))
1206                                 return -1;
1207                         memcpy(pagebuf+offset,buf,wrlen) ;
1208                         if (write(client->difffile, pagebuf, DIFFPAGESIZE) !=
1209                                         DIFFPAGESIZE)
1210                                 return -1;
1211                 }
1212                 len-=wrlen ; a+=wrlen ; buf+=wrlen ;
1213         }
1214         return 0;
1215 }
1216
1217 /**
1218  * Do the initial negotiation.
1219  *
1220  * @param client The client we're negotiating with.
1221  **/
1222 CLIENT* negotiate(int net, CLIENT *client, GArray* servers) {
1223         char zeros[128];
1224         uint64_t size_host;
1225         uint32_t flags = NBD_FLAG_HAS_FLAGS;
1226         uint16_t smallflags = 0;
1227         uint64_t magic;
1228
1229         memset(zeros, '\0', sizeof(zeros));
1230         if(!client || !client->modern) {
1231                 if (write(net, INIT_PASSWD, 8) < 0) {
1232                         err_nonfatal("Negotiation failed: %m");
1233                         if(client)
1234                                 exit(EXIT_FAILURE);
1235                 }
1236                 if(client && client->modern) {
1237                         magic = htonll(opts_magic);
1238                 } else {
1239                         magic = htonll(cliserv_magic);
1240                 }
1241                 if (write(net, &magic, sizeof(magic)) < 0) {
1242                         err_nonfatal("Negotiation failed: %m");
1243                         if(client)
1244                                 exit(EXIT_FAILURE);
1245                 }
1246         }
1247         if(!client) {
1248                 uint32_t reserved;
1249                 uint32_t opt;
1250                 uint32_t namelen;
1251                 char* name;
1252                 int i;
1253
1254                 if(!servers)
1255                         err("programmer error");
1256                 write(net, &smallflags, sizeof(uint16_t));
1257                 read(net, &reserved, sizeof(reserved));
1258                 read(net, &magic, sizeof(magic));
1259                 magic = ntohll(magic);
1260                 if(magic != opts_magic) {
1261                         close(net);
1262                         return NULL;
1263                 }
1264                 read(net, &opt, sizeof(opt));
1265                 opt = ntohl(opt);
1266                 if(opt != NBD_OPT_EXPORT_NAME) {
1267                         close(net);
1268                         return NULL;
1269                 }
1270                 read(net, &namelen, sizeof(namelen));
1271                 namelen = ntohl(namelen);
1272                 name = malloc(namelen+1);
1273                 name[namelen+1]=0;
1274                 read(net, name, namelen);
1275                 for(i=0; i<servers->len; i++) {
1276                         SERVER* serve = &(g_array_index(servers, SERVER, i));
1277                         if(!strcmp(serve->servename, name)) {
1278                                 CLIENT* client = g_new0(CLIENT, 1);
1279                                 client->server = serve;
1280                                 client->exportsize = OFFT_MAX;
1281                                 client->net = net;
1282                                 client->modern = TRUE;
1283                                 return client;
1284                         }
1285                 }
1286         }
1287         size_host = htonll((u64)(client->exportsize));
1288         if (write(net, &size_host, 8) < 0)
1289                 err("Negotiation failed: %m");
1290         if (client->server->flags & F_READONLY)
1291                 flags |= NBD_FLAG_READ_ONLY;
1292         if (!client->modern) {
1293                 flags = htonl(flags);
1294                 if (write(client->net, &flags, 4) < 0)
1295                         err("Negotiation failed: %m");
1296         } else {
1297                 smallflags = (uint16_t)(flags & ~((uint16_t)0));
1298                 smallflags = htons(smallflags);
1299                 if (write(client->net, &smallflags, sizeof(smallflags)) < 0) {
1300                         err("Negotiation failed: %m");
1301                 }
1302         }
1303         if (write(client->net, zeros, 124) < 0)
1304                 err("Negotiation failed: %m");
1305         return NULL;
1306 }
1307
1308 /** sending macro. */
1309 #define SEND(net,reply) writeit( net, &reply, sizeof( reply ));
1310 /** error macro. */
1311 #define ERROR(client,reply,errcode) { reply.error = htonl(errcode); SEND(client->net,reply); reply.error = 0; }
1312 /**
1313  * Serve a file to a single client.
1314  *
1315  * @todo This beast needs to be split up in many tiny little manageable
1316  * pieces. Preferably with a chainsaw.
1317  *
1318  * @param client The client we're going to serve to.
1319  * @return when the client disconnects
1320  **/
1321 int mainloop(CLIENT *client) {
1322         struct nbd_request request;
1323         struct nbd_reply reply;
1324         gboolean go_on=TRUE;
1325 #ifdef DODBG
1326         int i = 0;
1327 #endif
1328         negotiate(client->net, client, NULL);
1329         DEBUG("Entering request loop!\n");
1330         reply.magic = htonl(NBD_REPLY_MAGIC);
1331         reply.error = 0;
1332         while (go_on) {
1333                 char buf[BUFSIZE];
1334                 size_t len;
1335 #ifdef DODBG
1336                 i++;
1337                 printf("%d: ", i);
1338 #endif
1339                 readit(client->net, &request, sizeof(request));
1340                 request.from = ntohll(request.from);
1341                 request.type = ntohl(request.type);
1342
1343                 if (request.type==NBD_CMD_DISC) {
1344                         msg2(LOG_INFO, "Disconnect request received.");
1345                         if (client->server->flags & F_COPYONWRITE) {
1346                                 if (client->difmap) g_free(client->difmap) ;
1347                                 close(client->difffile);
1348                                 unlink(client->difffilename);
1349                                 free(client->difffilename);
1350                         }
1351                         go_on=FALSE;
1352                         continue;
1353                 }
1354
1355                 len = ntohl(request.len);
1356
1357                 if (request.magic != htonl(NBD_REQUEST_MAGIC))
1358                         err("Not enough magic.");
1359                 if (len > BUFSIZE + sizeof(struct nbd_reply))
1360                         err("Request too big!");
1361 #ifdef DODBG
1362                 printf("%s from %llu (%llu) len %d, ", request.type ? "WRITE" :
1363                                 "READ", (unsigned long long)request.from,
1364                                 (unsigned long long)request.from / 512, len);
1365 #endif
1366                 memcpy(reply.handle, request.handle, sizeof(reply.handle));
1367                 if ((request.from + len) > (OFFT_MAX)) {
1368                         DEBUG("[Number too large!]");
1369                         ERROR(client, reply, EINVAL);
1370                         continue;
1371                 }
1372
1373                 if (((ssize_t)((off_t)request.from + len) > client->exportsize)) {
1374                         DEBUG("[RANGE!]");
1375                         ERROR(client, reply, EINVAL);
1376                         continue;
1377                 }
1378
1379                 if (request.type==NBD_CMD_WRITE) {
1380                         DEBUG("wr: net->buf, ");
1381                         readit(client->net, buf, len);
1382                         DEBUG("buf->exp, ");
1383                         if ((client->server->flags & F_READONLY) ||
1384                             (client->server->flags & F_AUTOREADONLY)) {
1385                                 DEBUG("[WRITE to READONLY!]");
1386                                 ERROR(client, reply, EPERM);
1387                                 continue;
1388                         }
1389                         if (expwrite(request.from, buf, len, client)) {
1390                                 DEBUG("Write failed: %m" );
1391                                 ERROR(client, reply, errno);
1392                                 continue;
1393                         }
1394                         SEND(client->net, reply);
1395                         DEBUG("OK!\n");
1396                         continue;
1397                 }
1398                 /* READ */
1399
1400                 DEBUG("exp->buf, ");
1401                 if (expread(request.from, buf + sizeof(struct nbd_reply), len, client)) {
1402                         DEBUG("Read failed: %m");
1403                         ERROR(client, reply, errno);
1404                         continue;
1405                 }
1406
1407                 DEBUG("buf->net, ");
1408                 memcpy(buf, &reply, sizeof(struct nbd_reply));
1409                 writeit(client->net, buf, len + sizeof(struct nbd_reply));
1410                 DEBUG("OK!\n");
1411         }
1412         return 0;
1413 }
1414
1415 /**
1416  * Set up client export array, which is an array of FILE_INFO.
1417  * Also, split a single exportfile into multiple ones, if that was asked.
1418  * @param client information on the client which we want to setup export for
1419  **/
1420 void setupexport(CLIENT* client) {
1421         int i;
1422         off_t laststartoff = 0, lastsize = 0;
1423         int multifile = (client->server->flags & F_MULTIFILE);
1424
1425         client->export = g_array_new(TRUE, TRUE, sizeof(FILE_INFO));
1426
1427         /* If multi-file, open as many files as we can.
1428          * If not, open exactly one file.
1429          * Calculate file sizes as we go to get total size. */
1430         for(i=0; ; i++) {
1431                 FILE_INFO fi;
1432                 gchar *tmpname;
1433                 gchar* error_string;
1434                 mode_t mode = (client->server->flags & F_READONLY) ? O_RDONLY : O_RDWR;
1435
1436                 if(multifile) {
1437                         tmpname=g_strdup_printf("%s.%d", client->exportname, i);
1438                 } else {
1439                         tmpname=g_strdup(client->exportname);
1440                 }
1441                 DEBUG2( "Opening %s\n", tmpname );
1442                 fi.fhandle = open(tmpname, mode);
1443                 if(fi.fhandle == -1 && mode == O_RDWR) {
1444                         /* Try again because maybe media was read-only */
1445                         fi.fhandle = open(tmpname, O_RDONLY);
1446                         if(fi.fhandle != -1) {
1447                                 /* Opening the base file in copyonwrite mode is
1448                                  * okay */
1449                                 if(!(client->server->flags & F_COPYONWRITE)) {
1450                                         client->server->flags |= F_AUTOREADONLY;
1451                                         client->server->flags |= F_READONLY;
1452                                 }
1453                         }
1454                 }
1455                 if(fi.fhandle == -1) {
1456                         if(multifile && i>0)
1457                                 break;
1458                         error_string=g_strdup_printf(
1459                                 "Could not open exported file %s: %%m",
1460                                 tmpname);
1461                         err(error_string);
1462                 }
1463                 fi.startoff = laststartoff + lastsize;
1464                 g_array_append_val(client->export, fi);
1465                 g_free(tmpname);
1466
1467                 /* Starting offset and size of this file will be used to
1468                  * calculate starting offset of next file */
1469                 laststartoff = fi.startoff;
1470                 lastsize = size_autodetect(fi.fhandle);
1471
1472                 if(!multifile)
1473                         break;
1474         }
1475
1476         /* Set export size to total calculated size */
1477         client->exportsize = laststartoff + lastsize;
1478
1479         /* Export size may be overridden */
1480         if(client->server->expected_size) {
1481                 /* desired size must be <= total calculated size */
1482                 if(client->server->expected_size > client->exportsize) {
1483                         err("Size of exported file is too big\n");
1484                 }
1485
1486                 client->exportsize = client->server->expected_size;
1487         }
1488
1489         msg3(LOG_INFO, "Size of exported file/device is %llu", (unsigned long long)client->exportsize);
1490         if(multifile) {
1491                 msg3(LOG_INFO, "Total number of files: %d", i);
1492         }
1493 }
1494
1495 int copyonwrite_prepare(CLIENT* client) {
1496         off_t i;
1497         if ((client->difffilename = malloc(1024))==NULL)
1498                 err("Failed to allocate string for diff file name");
1499         snprintf(client->difffilename, 1024, "%s-%s-%d.diff",client->exportname,client->clientname,
1500                 (int)getpid()) ;
1501         client->difffilename[1023]='\0';
1502         msg3(LOG_INFO,"About to create map and diff file %s",client->difffilename) ;
1503         client->difffile=open(client->difffilename,O_RDWR | O_CREAT | O_TRUNC,0600) ;
1504         if (client->difffile<0) err("Could not create diff file (%m)") ;
1505         if ((client->difmap=calloc(client->exportsize/DIFFPAGESIZE,sizeof(u32)))==NULL)
1506                 err("Could not allocate memory") ;
1507         for (i=0;i<client->exportsize/DIFFPAGESIZE;i++) client->difmap[i]=(u32)-1 ;
1508
1509         return 0;
1510 }
1511
1512 /**
1513  * Run a command. This is used for the ``prerun'' and ``postrun'' config file
1514  * options
1515  *
1516  * @param command the command to be ran. Read from the config file
1517  * @param file the file name we're about to export
1518  **/
1519 int do_run(gchar* command, gchar* file) {
1520         gchar* cmd;
1521         int retval=0;
1522
1523         if(command && *command) {
1524                 cmd = g_strdup_printf(command, file);
1525                 retval=system(cmd);
1526                 g_free(cmd);
1527         }
1528         return retval;
1529 }
1530
1531 /**
1532  * Serve a connection.
1533  *
1534  * @todo allow for multithreading, perhaps use libevent. Not just yet, though;
1535  * follow the road map.
1536  *
1537  * @param client a connected client
1538  **/
1539 void serveconnection(CLIENT *client) {
1540         if(do_run(client->server->prerun, client->exportname)) {
1541                 exit(EXIT_FAILURE);
1542         }
1543         setupexport(client);
1544
1545         if (client->server->flags & F_COPYONWRITE) {
1546                 copyonwrite_prepare(client);
1547         }
1548
1549         setmysockopt(client->net);
1550
1551         mainloop(client);
1552         do_run(client->server->postrun, client->exportname);
1553 }
1554
1555 /**
1556  * Find the name of the file we have to serve. This will use g_strdup_printf
1557  * to put the IP address of the client inside a filename containing
1558  * "%s" (in the form as specified by the "virtstyle" option). That name
1559  * is then written to client->exportname.
1560  *
1561  * @param net A socket connected to an nbd client
1562  * @param client information about the client. The IP address in human-readable
1563  * format will be written to a new char* buffer, the address of which will be
1564  * stored in client->clientname.
1565  **/
1566 void set_peername(int net, CLIENT *client) {
1567         struct sockaddr_storage addrin;
1568         struct sockaddr_storage netaddr;
1569         struct sockaddr_in  *netaddr4 = NULL;
1570         struct sockaddr_in6 *netaddr6 = NULL;
1571         size_t addrinlen = sizeof( addrin );
1572         struct addrinfo hints;
1573         struct addrinfo *ai = NULL;
1574         char peername[NI_MAXHOST];
1575         char netname[NI_MAXHOST];
1576         char *tmp = NULL;
1577         int i;
1578         int e;
1579         int shift;
1580
1581         if (getpeername(net, (struct sockaddr *) &addrin, (socklen_t *)&addrinlen) < 0)
1582                 err("getsockname failed: %m");
1583
1584         getnameinfo((struct sockaddr *)&addrin, (socklen_t)addrinlen,
1585                 peername, sizeof (peername), NULL, 0, NI_NUMERICHOST);
1586
1587         memset(&hints, '\0', sizeof (hints));
1588         hints.ai_flags = AI_ADDRCONFIG;
1589         e = getaddrinfo(peername, NULL, &hints, &ai);
1590
1591         if(e != 0) {
1592                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1593                 freeaddrinfo(ai);
1594                 return;
1595         }
1596
1597         switch(client->server->virtstyle) {
1598                 case VIRT_NONE:
1599                         client->exportname=g_strdup(client->server->exportname);
1600                         break;
1601                 case VIRT_IPHASH:
1602                         for(i=0;i<strlen(peername);i++) {
1603                                 if(peername[i]=='.') {
1604                                         peername[i]='/';
1605                                 }
1606                         }
1607                 case VIRT_IPLIT:
1608                         client->exportname=g_strdup_printf(client->server->exportname, peername);
1609                         break;
1610                 case VIRT_CIDR:
1611                         memcpy(&netaddr, &addrin, addrinlen);
1612                         if(ai->ai_family == AF_INET) {
1613                                 netaddr4 = (struct sockaddr_in *)&netaddr;
1614                                 (netaddr4->sin_addr).s_addr>>=32-(client->server->cidrlen);
1615                                 (netaddr4->sin_addr).s_addr<<=32-(client->server->cidrlen);
1616
1617                                 getnameinfo((struct sockaddr *) netaddr4, (socklen_t) addrinlen,
1618                                                         netname, sizeof (netname), NULL, 0, NI_NUMERICHOST);
1619                                 tmp=g_strdup_printf("%s/%s", netname, peername);
1620                         }else if(ai->ai_family == AF_INET6) {
1621                                 netaddr6 = (struct sockaddr_in6 *)&netaddr;
1622
1623                                 shift = 128-(client->server->cidrlen);
1624                                 i = 3;
1625                                 while(shift >= 32) {
1626                                         ((netaddr6->sin6_addr).s6_addr32[i])=0;
1627                                         shift-=32;
1628                                         i--;
1629                                 }
1630                                 (netaddr6->sin6_addr).s6_addr32[i]>>=shift;
1631                                 (netaddr6->sin6_addr).s6_addr32[i]<<=shift;
1632
1633                                 getnameinfo((struct sockaddr *)netaddr6, (socklen_t)addrinlen,
1634                                             netname, sizeof(netname), NULL, 0, NI_NUMERICHOST);
1635                                 tmp=g_strdup_printf("%s/%s", netname, peername);
1636                         }
1637
1638                         if(tmp != NULL)
1639                           client->exportname=g_strdup_printf(client->server->exportname, tmp);
1640
1641                         break;
1642         }
1643
1644         freeaddrinfo(ai);
1645         msg4(LOG_INFO, "connect from %s, assigned file is %s",
1646              peername, client->exportname);
1647         client->clientname=g_strdup(peername);
1648 }
1649
1650 /**
1651  * Destroy a pid_t*
1652  * @param data a pointer to pid_t which should be freed
1653  **/
1654 void destroy_pid_t(gpointer data) {
1655         g_free(data);
1656 }
1657
1658 /**
1659  * Loop through the available servers, and serve them. Never returns.
1660  **/
1661 int serveloop(GArray* servers) {
1662         struct sockaddr_storage addrin;
1663         socklen_t addrinlen=sizeof(addrin);
1664         int i;
1665         int max;
1666         int sock;
1667         fd_set mset;
1668         fd_set rset;
1669
1670         /*
1671          * Set up the master fd_set. The set of descriptors we need
1672          * to select() for never changes anyway and it buys us a *lot*
1673          * of time to only build this once. However, if we ever choose
1674          * to not fork() for clients anymore, we may have to revisit
1675          * this.
1676          */
1677         max=0;
1678         FD_ZERO(&mset);
1679         for(i=0;i<servers->len;i++) {
1680                 if((sock=(g_array_index(servers, SERVER, i)).socket)) {
1681                         FD_SET(sock, &mset);
1682                         max=sock>max?sock:max;
1683                 }
1684         }
1685         if(modernsock) {
1686                 FD_SET(modernsock, &mset);
1687                 max=modernsock>max?modernsock:max;
1688         }
1689         for(;;) {
1690                 CLIENT *client = NULL;
1691                 pid_t *pid;
1692
1693                 memcpy(&rset, &mset, sizeof(fd_set));
1694                 if(select(max+1, &rset, NULL, NULL, NULL)>0) {
1695                         int net = 0;
1696                         SERVER* serve;
1697
1698                         DEBUG("accept, ");
1699                         if(FD_ISSET(modernsock, &rset)) {
1700                                 if((net=accept(modernsock, (struct sockaddr *) &addrin, &addrinlen)) < 0)
1701                                         err("accept: %m");
1702                                 client = negotiate(net, NULL, servers);
1703                                 if(!client) {
1704                                         err_nonfatal("negotiation failed");
1705                                         close(net);
1706                                 }
1707                         }
1708                         for(i=0;i<servers->len && !net;i++) {
1709                                 serve=&(g_array_index(servers, SERVER, i));
1710                                 if(FD_ISSET(serve->socket, &rset)) {
1711                                         if ((net=accept(serve->socket, (struct sockaddr *) &addrin, &addrinlen)) < 0)
1712                                                 err("accept: %m");
1713                                 }
1714                         }
1715                         if(net) {
1716                                 int sock_flags;
1717
1718                                 if((sock_flags = fcntl(net, F_GETFL, 0))==-1) {
1719                                         err("fcntl F_GETFL");
1720                                 }
1721                                 if(fcntl(net, F_SETFL, sock_flags &~O_NONBLOCK)==-1) {
1722                                         err("fcntl F_SETFL ~O_NONBLOCK");
1723                                 }
1724                                 if(!client) {
1725                                         client = g_new0(CLIENT, 1);
1726                                         client->server=serve;
1727                                         client->exportsize=OFFT_MAX;
1728                                         client->net=net;
1729                                 }
1730                                 set_peername(net, client);
1731                                 if (!authorized_client(client)) {
1732                                         msg2(LOG_INFO,"Unauthorized client") ;
1733                                         close(net);
1734                                         continue;
1735                                 }
1736                                 msg2(LOG_INFO,"Authorized client") ;
1737                                 pid=g_malloc(sizeof(pid_t));
1738 #ifndef NOFORK
1739                                 if ((*pid=fork())<0) {
1740                                         msg3(LOG_INFO,"Could not fork (%s)",strerror(errno)) ;
1741                                         close(net);
1742                                         continue;
1743                                 }
1744                                 if (*pid>0) { /* parent */
1745                                         close(net);
1746                                         g_hash_table_insert(children, pid, pid);
1747                                         continue;
1748                                 }
1749                                 /* child */
1750                                 g_hash_table_destroy(children);
1751                                 for(i=0;i<servers->len;i++) {
1752                                         serve=&g_array_index(servers, SERVER, i);
1753                                         close(serve->socket);
1754                                 }
1755                                 /* FALSE does not free the
1756                                 actual data. This is required,
1757                                 because the client has a
1758                                 direct reference into that
1759                                 data, and otherwise we get a
1760                                 segfault... */
1761                                 g_array_free(servers, FALSE);
1762 #endif // NOFORK
1763                                 msg2(LOG_INFO,"Starting to serve");
1764                                 serveconnection(client);
1765                                 exit(EXIT_SUCCESS);
1766                         }
1767                 }
1768         }
1769 }
1770
1771 void dosockopts(int socket) {
1772 #ifndef sun
1773         int yes=1;
1774 #else
1775         char yes='1';
1776 #endif /* sun */
1777         int sock_flags;
1778
1779         /* lose the pesky "Address already in use" error message */
1780         if (setsockopt(socket,SOL_SOCKET,SO_REUSEADDR,&yes,sizeof(int)) == -1) {
1781                 err("setsockopt SO_REUSEADDR");
1782         }
1783         if (setsockopt(socket,SOL_SOCKET,SO_KEEPALIVE,&yes,sizeof(int)) == -1) {
1784                 err("setsockopt SO_KEEPALIVE");
1785         }
1786
1787         /* make the listening socket non-blocking */
1788         if ((sock_flags = fcntl(socket, F_GETFL, 0)) == -1) {
1789                 err("fcntl F_GETFL");
1790         }
1791         if (fcntl(socket, F_SETFL, sock_flags | O_NONBLOCK) == -1) {
1792                 err("fcntl F_SETFL O_NONBLOCK");
1793         }
1794 }
1795
1796 /**
1797  * Connect a server's socket.
1798  *
1799  * @param serve the server we want to connect.
1800  **/
1801 int setup_serve(SERVER *serve) {
1802         struct addrinfo hints;
1803         struct addrinfo *ai = NULL;
1804         gchar *port = NULL;
1805         int e;
1806
1807         if(!do_oldstyle) {
1808                 return serve->servename ? 1 : 0;
1809         }
1810         memset(&hints,'\0',sizeof(hints));
1811         hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG | AI_NUMERICSERV;
1812         hints.ai_socktype = SOCK_STREAM;
1813         hints.ai_family = serve->socket_family;
1814
1815         port = g_strdup_printf ("%d", serve->port);
1816         if (port == NULL)
1817                 return 0;
1818
1819         e = getaddrinfo(serve->listenaddr,port,&hints,&ai);
1820
1821         g_free(port);
1822
1823         if(e != 0) {
1824                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1825                 serve->socket = -1;
1826                 freeaddrinfo(ai);
1827                 exit(EXIT_FAILURE);
1828         }
1829
1830         if(serve->socket_family == AF_UNSPEC)
1831                 serve->socket_family = ai->ai_family;
1832
1833 #ifdef WITH_SDP
1834         if ((serve->flags) && F_SDP) {
1835                 if (ai->ai_family == AF_INET)
1836                         ai->ai_family = AF_INET_SDP;
1837                 else (ai->ai_family == AF_INET6)
1838                         ai->ai_family = AF_INET6_SDP;
1839         }
1840 #endif
1841         if ((serve->socket = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol)) < 0)
1842                 err("socket: %m");
1843
1844         dosockopts(serve->socket);
1845
1846         DEBUG("Waiting for connections... bind, ");
1847         e = bind(serve->socket, ai->ai_addr, ai->ai_addrlen);
1848         if (e != 0 && errno != EADDRINUSE)
1849                 err("bind: %m");
1850         DEBUG("listen, ");
1851         if (listen(serve->socket, 1) < 0)
1852                 err("listen: %m");
1853
1854         freeaddrinfo (ai);
1855         if(serve->servename) {
1856                 return 1;
1857         } else {
1858                 return 0;
1859         }
1860 }
1861
1862 void open_modern(void) {
1863         struct addrinfo hints;
1864         struct addrinfo* ai = NULL;
1865         struct sock_flags;
1866         int e;
1867
1868         memset(&hints, '\0', sizeof(hints));
1869         hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG;
1870         hints.ai_socktype = SOCK_STREAM;
1871         hints.ai_family = AF_UNSPEC;
1872         hints.ai_protocol = IPPROTO_TCP;
1873         e = getaddrinfo(modern_listen, NBD_DEFAULT_PORT, &hints, &ai);
1874         if(e != 0) {
1875                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1876                 exit(EXIT_FAILURE);
1877         }
1878         if((modernsock = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol))<0) {
1879                 err("socket: %m");
1880         }
1881
1882         dosockopts(modernsock);
1883
1884         if(bind(modernsock, ai->ai_addr, ai->ai_addrlen)) {
1885                 err("bind: %m");
1886         }
1887         if(listen(modernsock, 10) <0) {
1888                 err("listen: %m");
1889         }
1890
1891         freeaddrinfo(ai);
1892 }
1893
1894 /**
1895  * Connect our servers.
1896  **/
1897 void setup_servers(GArray* servers) {
1898         int i;
1899         struct sigaction sa;
1900         int want_modern=0;
1901
1902         for(i=0;i<servers->len;i++) {
1903                 want_modern |= setup_serve(&(g_array_index(servers, SERVER, i)));
1904         }
1905         if(want_modern) {
1906                 open_modern();
1907         }
1908         children=g_hash_table_new_full(g_int_hash, g_int_equal, NULL, destroy_pid_t);
1909
1910         sa.sa_handler = sigchld_handler;
1911         sigemptyset(&sa.sa_mask);
1912         sa.sa_flags = SA_RESTART;
1913         if(sigaction(SIGCHLD, &sa, NULL) == -1)
1914                 err("sigaction: %m");
1915         sa.sa_handler = sigterm_handler;
1916         sigemptyset(&sa.sa_mask);
1917         sa.sa_flags = SA_RESTART;
1918         if(sigaction(SIGTERM, &sa, NULL) == -1)
1919                 err("sigaction: %m");
1920 }
1921
1922 /**
1923  * Go daemon (unless we specified at compile time that we didn't want this)
1924  * @param serve the first server of our configuration. If its port is zero,
1925  *      then do not daemonize, because we're doing inetd then. This parameter
1926  *      is only used to create a PID file of the form
1927  *      /var/run/nbd-server.&lt;port&gt;.pid; it's not modified in any way.
1928  **/
1929 #if !defined(NODAEMON) && !defined(NOFORK)
1930 void daemonize(SERVER* serve) {
1931         FILE*pidf;
1932
1933         if(serve && !(serve->port)) {
1934                 return;
1935         }
1936         if(daemon(0,0)<0) {
1937                 err("daemon");
1938         }
1939         if(!*pidftemplate) {
1940                 if(serve) {
1941                         strncpy(pidftemplate, "/var/run/nbd-server.%d.pid", 255);
1942                 } else {
1943                         strncpy(pidftemplate, "/var/run/nbd-server.pid", 255);
1944                 }
1945         }
1946         snprintf(pidfname, 255, pidftemplate, serve ? serve->port : 0);
1947         pidf=fopen(pidfname, "w");
1948         if(pidf) {
1949                 fprintf(pidf,"%d\n", (int)getpid());
1950                 fclose(pidf);
1951         } else {
1952                 perror("fopen");
1953                 fprintf(stderr, "Not fatal; continuing");
1954         }
1955 }
1956 #else
1957 #define daemonize(serve)
1958 #endif /* !defined(NODAEMON) && !defined(NOFORK) */
1959
1960 /*
1961  * Everything beyond this point (in the file) is run in non-daemon mode.
1962  * The stuff above daemonize() isn't.
1963  */
1964
1965 void serve_err(SERVER* serve, const char* msg) G_GNUC_NORETURN;
1966
1967 void serve_err(SERVER* serve, const char* msg) {
1968         g_message("Export of %s on port %d failed:", serve->exportname,
1969                         serve->port);
1970         err(msg);
1971 }
1972
1973 /**
1974  * Set up user-ID and/or group-ID
1975  **/
1976 void dousers(void) {
1977         struct passwd *pw;
1978         struct group *gr;
1979         gchar* str;
1980         if(rungroup) {
1981                 gr=getgrnam(rungroup);
1982                 if(!gr) {
1983                         str = g_strdup_printf("Invalid group name: %s", rungroup);
1984                         err(str);
1985                 }
1986                 if(setgid(gr->gr_gid)<0) {
1987                         err("Could not set GID: %m");
1988                 }
1989         }
1990         if(runuser) {
1991                 pw=getpwnam(runuser);
1992                 if(!pw) {
1993                         str = g_strdup_printf("Invalid user name: %s", runuser);
1994                         err(str);
1995                 }
1996                 if(setuid(pw->pw_uid)<0) {
1997                         err("Could not set UID: %m");
1998                 }
1999         }
2000 }
2001
2002 #ifndef ISSERVER
2003 void glib_message_syslog_redirect(const gchar *log_domain,
2004                                   GLogLevelFlags log_level,
2005                                   const gchar *message,
2006                                   gpointer user_data)
2007 {
2008     int level=LOG_DEBUG;
2009
2010     switch( log_level )
2011     {
2012       case G_LOG_FLAG_FATAL:
2013       case G_LOG_LEVEL_CRITICAL:
2014       case G_LOG_LEVEL_ERROR:
2015         level=LOG_ERR;
2016         break;
2017       case G_LOG_LEVEL_WARNING:
2018         level=LOG_WARNING;
2019         break;
2020       case G_LOG_LEVEL_MESSAGE:
2021       case G_LOG_LEVEL_INFO:
2022         level=LOG_INFO;
2023         break;
2024       case G_LOG_LEVEL_DEBUG:
2025         level=LOG_DEBUG;
2026       default:
2027         level=LOG_ERR;
2028     }
2029     syslog(level, message);
2030 }
2031 #endif
2032
2033 /**
2034  * Main entry point...
2035  **/
2036 int main(int argc, char *argv[]) {
2037         SERVER *serve;
2038         GArray *servers;
2039         GError *err=NULL;
2040
2041         if (sizeof( struct nbd_request )!=28) {
2042                 fprintf(stderr,"Bad size of structure. Alignment problems?\n");
2043                 exit(EXIT_FAILURE) ;
2044         }
2045
2046         memset(pidftemplate, '\0', 256);
2047
2048         logging();
2049         config_file_pos = g_strdup(CFILE);
2050         serve=cmdline(argc, argv);
2051         servers = parse_cfile(config_file_pos, &err);
2052
2053         if(serve) {
2054                 serve->socket_family = AF_UNSPEC;
2055
2056                 append_serve(serve, servers);
2057
2058                 if (!(serve->port)) {
2059                         CLIENT *client;
2060 #ifndef ISSERVER
2061                         /* You really should define ISSERVER if you're going to use
2062                          * inetd mode, but if you don't, closing stdout and stderr
2063                          * (which inetd had connected to the client socket) will let it
2064                          * work. */
2065                         close(1);
2066                         close(2);
2067                         open("/dev/null", O_WRONLY);
2068                         open("/dev/null", O_WRONLY);
2069                         g_log_set_default_handler( glib_message_syslog_redirect, NULL );
2070 #endif
2071                         client=g_malloc(sizeof(CLIENT));
2072                         client->server=serve;
2073                         client->net=0;
2074                         client->exportsize=OFFT_MAX;
2075                         set_peername(0,client);
2076                         serveconnection(client);
2077                         return 0;
2078                 }
2079         }
2080
2081         if(!servers || !servers->len) {
2082                 g_warning("Could not parse config file: %s",
2083                                 err ? err->message : "Unknown error");
2084         }
2085         if(serve) {
2086                 g_warning("Specifying an export on the command line is deprecated.");
2087                 g_warning("Please use a configuration file instead.");
2088         }
2089
2090         if((!serve) && (!servers||!servers->len)) {
2091                 g_message("Nothing to do! Bye!");
2092                 exit(EXIT_FAILURE);
2093         }
2094         daemonize(serve);
2095         setup_servers(servers);
2096         dousers();
2097         serveloop(servers);
2098         return 0 ;
2099 }