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 #ifdef WITH_SDP
 100 #include <sdp_inet.h>
 101 #endif
 102
 103 /** Default position of the config file */
 104 #ifndef SYSCONFDIR
 105 #define SYSCONFDIR "/etc"
 106 #endif
 107 #define CFILE SYSCONFDIR "/nbd-server/config"
 108
 109 /** Where our config file actually is */
 110 gchar* config_file_pos;
 111
 112 /** What user we're running as */
 113 gchar* runuser=NULL;
 114 /** What group we're running as */
 115 gchar* rungroup=NULL;
 116 /** whether to export using the old negotiation protocol (port-based) */
 117 gboolean do_oldstyle=FALSE;
 118
 119 /** Logging macros, now nothing goes to syslog unless you say ISSERVER */
 120 #ifdef ISSERVER
 121 #define msg2(a,b) syslog(a,b)
 122 #define msg3(a,b,c) syslog(a,b,c)
 123 #define msg4(a,b,c,d) syslog(a,b,c,d)
 124 #else
 125 #define msg2(a,b) g_message(b)
 126 #define msg3(a,b,c) g_message(b,c)
 127 #define msg4(a,b,c,d) g_message(b,c,d)
 128 #endif
 129
 130 /* Debugging macros */
 131 //#define DODBG
 132 #ifdef DODBG
 133 #define DEBUG( a ) printf( a )
 134 #define DEBUG2( a,b ) printf( a,b )
 135 #define DEBUG3( a,b,c ) printf( a,b,c )
 136 #define DEBUG4( a,b,c,d ) printf( a,b,c,d )
 137 #else
 138 #define DEBUG( a )
 139 #define DEBUG2( a,b )
 140 #define DEBUG3( a,b,c )
 141 #define DEBUG4( a,b,c,d )
 142 #endif
 143 #ifndef PACKAGE_VERSION
 144 #define PACKAGE_VERSION ""
 145 #endif
 146 /**
 147  * The highest value a variable of type off_t can reach. This is a signed
 148  * integer, so set all bits except for the leftmost one.
 149  **/
 150 #define OFFT_MAX ~((off_t)1<<(sizeof(off_t)*8-1))
 151 #define LINELEN 256       /**< Size of static buffer used to read the
 152                                authorization file (yuck) */
 153 #define BUFSIZE (1024*1024) /**< Size of buffer that can hold requests */
 154 #define DIFFPAGESIZE 4096 /**< diff file uses those chunks */
 155 #define F_READONLY 1      /**< flag to tell us a file is readonly */
 156 #define F_MULTIFILE 2     /**< flag to tell us a file is exported using -m */
 157 #define F_COPYONWRITE 4   /**< flag to tell us a file is exported using
 158                             copyonwrite */
 159 #define F_AUTOREADONLY 8  /**< flag to tell us a file is set to autoreadonly */
 160 #define F_SPARSE 16       /**< flag to tell us copyronwrite should use a sparse file */
 161 #define F_SDP 32          /**< flag to tell us the export should be done using the Socket Direct Protocol for RDMA */
 162 #define F_SYNC 64         /**< Whether to fsync() after a write */
 163 GHashTable *children;
 164 char pidfname[256]; /**< name of our PID file */
 165 char pidftemplate[256]; /**< template to be used for the filename of the PID file */
 166 char default_authname[] = SYSCONFDIR "/nbd-server/allow"; /**< default name of allow file */
 167
 168 int modernsock=0;         /**< Socket for the modern handler. Not used
 169                                if a client was only specified on the
 170                                command line; only port used if
 171                                oldstyle is set to false (and then the
 172                                command-line client isn't used, gna gna) */
 173 char* modern_listen;      /**< listenaddr value for modernsock */
 174
 175 /**
 176  * Types of virtuatlization
 177  **/
 178 typedef enum {
 179         VIRT_NONE=0,    /**< No virtualization */
 180         VIRT_IPLIT,     /**< Literal IP address as part of the filename */
 181         VIRT_IPHASH,    /**< Replacing all dots in an ip address by a / before
 182                              doing the same as in IPLIT */
 183         VIRT_CIDR,      /**< Every subnet in its own directory */
 184 } VIRT_STYLE;
 185
 186 /**
 187  * Variables associated with a server.
 188  **/
 189 typedef struct {
 190         gchar* exportname;    /**< (unprocessed) filename of the file we're exporting */
 191         off_t expected_size; /**< size of the exported file as it was told to
 192                                us through configuration */
 193         gchar* listenaddr;   /**< The IP address we're listening on */
 194         unsigned int port;   /**< port we're exporting this file at */
 195         char* authname;      /**< filename of the authorization file */
 196         int flags;           /**< flags associated with this exported file */
 197         int socket;          /**< The socket of this server. */
 198         int socket_family;   /**< family of the socket */
 199         VIRT_STYLE virtstyle;/**< The style of virtualization, if any */
 200         uint8_t cidrlen;     /**< The length of the mask when we use
 201                                   CIDR-style virtualization */
 202         gchar* prerun;       /**< command to be ran after connecting a client,
 203                                   but before starting to serve */
 204         gchar* postrun;      /**< command that will be ran after the client
 205                                   disconnects */
 206         gchar* servename;    /**< name of the export as selected by nbd-client */
 207         int max_connections; /**< maximum number of opened connections */
 208 } SERVER;
 209
 210 /**
 211  * Variables associated with a client socket.
 212  **/
 213 typedef struct {
 214         int fhandle;      /**< file descriptor */
 215         off_t startoff;   /**< starting offset of this file */
 216 } FILE_INFO;
 217
 218 typedef struct {
 219         off_t exportsize;    /**< size of the file we're exporting */
 220         char *clientname;    /**< peer */
 221         char *exportname;    /**< (processed) filename of the file we're exporting */
 222         GArray *export;    /**< array of FILE_INFO of exported files;
 223                                array size is always 1 unless we're
 224                                doing the multiple file option */
 225         int net;             /**< The actual client socket */
 226         SERVER *server;      /**< The server this client is getting data from */
 227         char* difffilename;  /**< filename of the copy-on-write file, if any */
 228         int difffile;        /**< filedescriptor of copyonwrite file. @todo
 229                                shouldn't this be an array too? (cfr export) Or
 230                                make -m and -c mutually exclusive */
 231         u32 difffilelen;     /**< number of pages in difffile */
 232         u32 *difmap;         /**< see comment on the global difmap for this one */
 233         gboolean modern;     /**< client was negotiated using modern negotiation protocol */
 234 } CLIENT;
 235
 236 /**
 237  * Type of configuration file values
 238  **/
 239 typedef enum {
 240         PARAM_INT,              /**< This parameter is an integer */
 241         PARAM_STRING,           /**< This parameter is a string */
 242         PARAM_BOOL,             /**< This parameter is a boolean */
 243 } PARAM_TYPE;
 244
 245 /**
 246  * Configuration file values
 247  **/
 248 typedef struct {
 249         gchar *paramname;       /**< Name of the parameter, as it appears in
 250                                   the config file */
 251         gboolean required;      /**< Whether this is a required (as opposed to
 252                                   optional) parameter */
 253         PARAM_TYPE ptype;       /**< Type of the parameter. */
 254         gpointer target;        /**< Pointer to where the data of this
 255                                   parameter should be written. If ptype is
 256                                   PARAM_BOOL, the data is or'ed rather than
 257                                   overwritten. */
 258         gint flagval;           /**< Flag mask for this parameter in case ptype
 259                                   is PARAM_BOOL. */
 260 } PARAM;
 261
 262 /**
 263  * Check whether a client is allowed to connect. Works with an authorization
 264  * file which contains one line per machine, no wildcards.
 265  *
 266  * @param opts The client who's trying to connect.
 267  * @return 0 - authorization refused, 1 - OK
 268  **/
 269 int authorized_client(CLIENT *opts) {
 270         const char *ERRMSG="Invalid entry '%s' in authfile '%s', so, refusing all connections.";
 271         FILE *f ;
 272         char line[LINELEN];
 273         char *tmp;
 274         struct in_addr addr;
 275         struct in_addr client;
 276         struct in_addr cltemp;
 277         int len;
 278
 279         if ((f=fopen(opts->server->authname,"r"))==NULL) {
 280                 msg4(LOG_INFO,"Can't open authorization file %s (%s).",
 281                      opts->server->authname,strerror(errno)) ;
 282                 return 1 ;
 283         }
 284
 285         inet_aton(opts->clientname, &client);
 286         while (fgets(line,LINELEN,f)!=NULL) {
 287                 if((tmp=index(line, '/'))) {
 288                         if(strlen(line)<=tmp-line) {
 289                                 msg4(LOG_CRIT, ERRMSG, line, opts->server->authname);
 290                                 return 0;
 291                         }
 292                         *(tmp++)=0;
 293                         if(!inet_aton(line,&addr)) {
 294                                 msg4(LOG_CRIT, ERRMSG, line, opts->server->authname);
 295                                 return 0;
 296                         }
 297                         len=strtol(tmp, NULL, 0);
 298                         addr.s_addr>>=32-len;
 299                         addr.s_addr<<=32-len;
 300                         memcpy(&cltemp,&client,sizeof(client));
 301                         cltemp.s_addr>>=32-len;
 302                         cltemp.s_addr<<=32-len;
 303                         if(addr.s_addr == cltemp.s_addr) {
 304                                 return 1;
 305                         }
 306                 }
 307                 if (strncmp(line,opts->clientname,strlen(opts->clientname))==0) {
 308                         fclose(f);
 309                         return 1;
 310                 }
 311         }
 312         fclose(f);
 313         return 0;
 314 }
 315
 316 /**
 317  * Read data from a file descriptor into a buffer
 318  *
 319  * @param f a file descriptor
 320  * @param buf a buffer
 321  * @param len the number of bytes to be read
 322  **/
 323 inline void readit(int f, void *buf, size_t len) {
 324         ssize_t res;
 325         while (len > 0) {
 326                 DEBUG("*");
 327                 if ((res = read(f, buf, len)) <= 0) {
 328                         if(errno != EAGAIN) {
 329                                 err("Read failed: %m");
 330                         }
 331                 } else {
 332                         len -= res;
 333                         buf += res;
 334                 }
 335         }
 336 }
 337
 338 /**
 339  * Write data from a buffer into a filedescriptor
 340  *
 341  * @param f a file descriptor
 342  * @param buf a buffer containing data
 343  * @param len the number of bytes to be written
 344  **/
 345 inline void writeit(int f, void *buf, size_t len) {
 346         ssize_t res;
 347         while (len > 0) {
 348                 DEBUG("+");
 349                 if ((res = write(f, buf, len)) <= 0)
 350                         err("Send failed: %m");
 351                 len -= res;
 352                 buf += res;
 353         }
 354 }
 355
 356 /**
 357  * Print out a message about how to use nbd-server. Split out to a separate
 358  * function so that we can call it from multiple places
 359  */
 360 void usage() {
 361         printf("This is nbd-server version " VERSION "\n");
 362         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] [-M max connections]\n"
 363                "\t-r|--read-only\t\tread only\n"
 364                "\t-m|--multi-file\t\tmultiple file\n"
 365                "\t-c|--copy-on-write\tcopy on write\n"
 366                "\t-C|--config-file\tspecify an alternate configuration file\n"
 367                "\t-l|--authorize-file\tfile with list of hosts that are allowed to\n\t\t\t\tconnect.\n"
 368                "\t-p|--pid-file\t\tspecify a filename to write our PID to\n"
 369                "\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"
 370                "\t-M|--max-connections\tspecify the maximum number of opened connections\n\n"
 371                "\tif port is set to 0, stdin is used (for running from inetd)\n"
 372                "\tif file_to_export contains '%%s', it is substituted with the IP\n"
 373                "\t\taddress of the machine trying to connect\n"
 374                "\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");
 375         printf("Using configuration file %s\n", CFILE);
 376 }
 377
 378 /* Dumps a config file section of the given SERVER*, and exits. */
 379 void dump_section(SERVER* serve, gchar* section_header) {
 380         printf("[%s]\n", section_header);
 381         printf("\texportname = %s\n", serve->exportname);
 382         printf("\tlistenaddr = %s\n", serve->listenaddr);
 383         printf("\tport = %d\n", serve->port);
 384         if(serve->flags & F_READONLY) {
 385                 printf("\treadonly = true\n");
 386         }
 387         if(serve->flags & F_MULTIFILE) {
 388                 printf("\tmultifile = true\n");
 389         }
 390         if(serve->flags & F_COPYONWRITE) {
 391                 printf("\tcopyonwrite = true\n");
 392         }
 393         if(serve->expected_size) {
 394                 printf("\tfilesize = %lld\n", (long long int)serve->expected_size);
 395         }
 396         if(serve->authname) {
 397                 printf("\tauthfile = %s\n", serve->authname);
 398         }
 399         exit(EXIT_SUCCESS);
 400 }
 401
 402 /**
 403  * Parse the command line.
 404  *
 405  * @param argc the argc argument to main()
 406  * @param argv the argv argument to main()
 407  **/
 408 SERVER* cmdline(int argc, char *argv[]) {
 409         int i=0;
 410         int nonspecial=0;
 411         int c;
 412         struct option long_options[] = {
 413                 {"read-only", no_argument, NULL, 'r'},
 414                 {"multi-file", no_argument, NULL, 'm'},
 415                 {"copy-on-write", no_argument, NULL, 'c'},
 416                 {"authorize-file", required_argument, NULL, 'l'},
 417                 {"config-file", required_argument, NULL, 'C'},
 418                 {"pid-file", required_argument, NULL, 'p'},
 419                 {"output-config", required_argument, NULL, 'o'},
 420                 {"max-connection", required_argument, NULL, 'M'},
 421                 {0,0,0,0}
 422         };
 423         SERVER *serve;
 424         off_t es;
 425         size_t last;
 426         char suffix;
 427         gboolean do_output=FALSE;
 428         gchar* section_header="";
 429         gchar** addr_port;
 430
 431         if(argc==1) {
 432                 return NULL;
 433         }
 434         serve=g_new0(SERVER, 1);
 435         serve->authname = g_strdup(default_authname);
 436         serve->virtstyle=VIRT_IPLIT;
 437         while((c=getopt_long(argc, argv, "-C:cl:mo:rp:M:", long_options, &i))>=0) {
 438                 switch (c) {
 439                 case 1:
 440                         /* non-option argument */
 441                         switch(nonspecial++) {
 442                         case 0:
 443                                 if(strchr(optarg, ':') == strrchr(optarg, ':')) {
 444                                         addr_port=g_strsplit(optarg, ":", 2);
 445
 446                                         /* Check for "@" - maybe user using this separator
 447                                                  for IPv4 address */
 448                                         if(!addr_port[1]) {
 449                                                 g_strfreev(addr_port);
 450                                                 addr_port=g_strsplit(optarg, "@", 2);
 451                                         }
 452                                 } else {
 453                                         addr_port=g_strsplit(optarg, "@", 2);
 454                                 }
 455
 456                                 if(addr_port[1]) {
 457                                         serve->port=strtol(addr_port[1], NULL, 0);
 458                                         serve->listenaddr=g_strdup(addr_port[0]);
 459                                 } else {
 460                                         serve->listenaddr=NULL;
 461                                         serve->port=strtol(addr_port[0], NULL, 0);
 462                                 }
 463                                 g_strfreev(addr_port);
 464                                 break;
 465                         case 1:
 466                                 serve->exportname = g_strdup(optarg);
 467                                 if(serve->exportname[0] != '/') {
 468                                         fprintf(stderr, "E: The to be exported file needs to be an absolute filename!\n");
 469                                         exit(EXIT_FAILURE);
 470                                 }
 471                                 break;
 472                         case 2:
 473                                 last=strlen(optarg)-1;
 474                                 suffix=optarg[last];
 475                                 if (suffix == 'k' || suffix == 'K' ||
 476                                     suffix == 'm' || suffix == 'M')
 477                                         optarg[last] = '\0';
 478                                 es = (off_t)atoll(optarg);
 479                                 switch (suffix) {
 480                                         case 'm':
 481                                         case 'M':  es <<= 10;
 482                                         case 'k':
 483                                         case 'K':  es <<= 10;
 484                                         default :  break;
 485                                 }
 486                                 serve->expected_size = es;
 487                                 break;
 488                         }
 489                         break;
 490                 case 'r':
 491                         serve->flags |= F_READONLY;
 492                         break;
 493                 case 'm':
 494                         serve->flags |= F_MULTIFILE;
 495                         break;
 496                 case 'o':
 497                         do_output = TRUE;
 498                         section_header = g_strdup(optarg);
 499                         break;
 500                 case 'p':
 501                         strncpy(pidftemplate, optarg, 256);
 502                         break;
 503                 case 'c':
 504                         serve->flags |=F_COPYONWRITE;
 505                         break;
 506                 case 'C':
 507                         g_free(config_file_pos);
 508                         config_file_pos=g_strdup(optarg);
 509                         break;
 510                 case 'l':
 511                         g_free(serve->authname);
 512                         serve->authname=g_strdup(optarg);
 513                         break;
 514                 case 'M':
 515                         serve->max_connections = strtol(optarg, NULL, 0);
 516                         break;
 517                 default:
 518                         usage();
 519                         exit(EXIT_FAILURE);
 520                         break;
 521                 }
 522         }
 523         /* What's left: the port to export, the name of the to be exported
 524          * file, and, optionally, the size of the file, in that order. */
 525         if(nonspecial<2) {
 526                 g_free(serve);
 527                 serve=NULL;
 528         } else {
 529                 do_oldstyle = TRUE;
 530         }
 531         if(do_output) {
 532                 if(!serve) {
 533                         g_critical("Need a complete configuration on the command line to output a config file section!");
 534                         exit(EXIT_FAILURE);
 535                 }
 536                 dump_section(serve, section_header);
 537         }
 538         return serve;
 539 }
 540
 541 /**
 542  * Error codes for config file parsing
 543  **/
 544 typedef enum {
 545         CFILE_NOTFOUND,         /**< The configuration file is not found */
 546         CFILE_MISSING_GENERIC,  /**< The (required) group "generic" is missing */
 547         CFILE_KEY_MISSING,      /**< A (required) key is missing */
 548         CFILE_VALUE_INVALID,    /**< A value is syntactically invalid */
 549         CFILE_VALUE_UNSUPPORTED,/**< A value is not supported in this build */
 550         CFILE_PROGERR,          /**< Programmer error */
 551         CFILE_NO_EXPORTS,       /**< A config file was specified that does not
 552                                      define any exports */
 553         CFILE_INCORRECT_PORT,   /**< The reserved port was specified for an
 554                                      old-style export. */
 555 } CFILE_ERRORS;
 556
 557 /**
 558  * Remove a SERVER from memory. Used from the hash table
 559  **/
 560 void remove_server(gpointer s) {
 561         SERVER *server;
 562
 563         server=(SERVER*)s;
 564         g_free(server->exportname);
 565         if(server->authname)
 566                 g_free(server->authname);
 567         if(server->listenaddr)
 568                 g_free(server->listenaddr);
 569         if(server->prerun)
 570                 g_free(server->prerun);
 571         if(server->postrun)
 572                 g_free(server->postrun);
 573         g_free(server);
 574 }
 575
 576 /**
 577  * duplicate server
 578  * @param s the old server we want to duplicate
 579  * @return new duplicated server
 580  **/
 581 SERVER* dup_serve(SERVER *s) {
 582         SERVER *serve = NULL;
 583
 584         serve=g_new0(SERVER, 1);
 585         if(serve == NULL)
 586                 return NULL;
 587
 588         if(s->exportname)
 589                 serve->exportname = g_strdup(s->exportname);
 590
 591         serve->expected_size = s->expected_size;
 592
 593         if(s->listenaddr)
 594                 serve->listenaddr = g_strdup(s->listenaddr);
 595
 596         serve->port = s->port;
 597
 598         if(s->authname)
 599                 serve->authname = strdup(s->authname);
 600
 601         serve->flags = s->flags;
 602         serve->socket = serve->socket;
 603         serve->socket_family = serve->socket_family;
 604         serve->cidrlen = s->cidrlen;
 605
 606         if(s->prerun)
 607                 serve->prerun = g_strdup(s->prerun);
 608
 609         if(s->postrun)
 610                 serve->postrun = g_strdup(s->postrun);
 611
 612         if(s->servename)
 613                 serve->servename = g_strdup(s->servename);
 614
 615         serve->max_connections = s->max_connections;
 616
 617         return serve;
 618 }
 619
 620 /**
 621  * append new server to array
 622  * @param s server
 623  * @param a server array
 624  * @return 0 success, -1 error
 625  */
 626 int append_serve(SERVER *s, GArray *a) {
 627         SERVER *ns = NULL;
 628         struct addrinfo hints;
 629         struct addrinfo *ai = NULL;
 630         struct addrinfo *rp = NULL;
 631         char   host[NI_MAXHOST];
 632         gchar  *port = NULL;
 633         int e;
 634         int ret;
 635
 636         if(!s) {
 637                 err("Invalid parsing server");
 638                 return -1;
 639         }
 640
 641         port = g_strdup_printf("%d", s->port);
 642
 643         memset(&hints,'\0',sizeof(hints));
 644         hints.ai_family = AF_UNSPEC;
 645         hints.ai_socktype = SOCK_STREAM;
 646         hints.ai_flags = AI_ADDRCONFIG | AI_PASSIVE;
 647         hints.ai_protocol = IPPROTO_TCP;
 648
 649         e = getaddrinfo(s->listenaddr, port, &hints, &ai);
 650
 651         if (port)
 652                 g_free(port);
 653
 654         if(e == 0) {
 655                 for (rp = ai; rp != NULL; rp = rp->ai_next) {
 656                         e = getnameinfo(rp->ai_addr, rp->ai_addrlen, host, sizeof(host), NULL, 0, NI_NUMERICHOST);
 657
 658                         if (e != 0) { // error
 659                                 fprintf(stderr, "getnameinfo: %s\n", gai_strerror(e));
 660                                 continue;
 661                         }
 662
 663                         // duplicate server and set listenaddr to resolved IP address
 664                         ns = dup_serve (s);
 665                         if (ns) {
 666                                 ns->listenaddr = g_strdup(host);
 667                                 ns->socket_family = rp->ai_family;
 668                                 g_array_append_val(a, *ns);
 669                                 free(ns);
 670                                 ns = NULL;
 671                         }
 672                 }
 673
 674                 ret = 0;
 675         } else {
 676                 fprintf(stderr, "getaddrinfo failed on listen host/address: %s (%s)\n", s->listenaddr ? s->listenaddr : "any", gai_strerror(e));
 677                 ret = -1;
 678         }
 679
 680         if (ai)
 681                 freeaddrinfo(ai);
 682
 683         return ret;
 684 }
 685
 686 /**
 687  * Parse the config file.
 688  *
 689  * @param f the name of the config file
 690  * @param e a GError. @see CFILE_ERRORS for what error values this function can
 691  *      return.
 692  * @return a Array of SERVER* pointers, If the config file is empty or does not
 693  *      exist, returns an empty GHashTable; if the config file contains an
 694  *      error, returns NULL, and e is set appropriately
 695  **/
 696 GArray* parse_cfile(gchar* f, GError** e) {
 697         const char* DEFAULT_ERROR = "Could not parse %s in group %s: %s";
 698         const char* MISSING_REQUIRED_ERROR = "Could not find required value %s in group %s: %s";
 699         SERVER s;
 700         gchar *virtstyle=NULL;
 701         PARAM lp[] = {
 702                 { "exportname", TRUE,   PARAM_STRING,   NULL, 0 },
 703                 { "port",       TRUE,   PARAM_INT,      NULL, 0 },
 704                 { "authfile",   FALSE,  PARAM_STRING,   NULL, 0 },
 705                 { "filesize",   FALSE,  PARAM_INT,      NULL, 0 },
 706                 { "virtstyle",  FALSE,  PARAM_STRING,   NULL, 0 },
 707                 { "prerun",     FALSE,  PARAM_STRING,   NULL, 0 },
 708                 { "postrun",    FALSE,  PARAM_STRING,   NULL, 0 },
 709                 { "readonly",   FALSE,  PARAM_BOOL,     NULL, F_READONLY },
 710                 { "multifile",  FALSE,  PARAM_BOOL,     NULL, F_MULTIFILE },
 711                 { "copyonwrite", FALSE, PARAM_BOOL,     NULL, F_COPYONWRITE },
 712                 { "sparse_cow", FALSE,  PARAM_BOOL,     NULL, F_SPARSE },
 713                 { "sdp",        FALSE,  PARAM_BOOL,     NULL, F_SDP },
 714                 { "sync",       FALSE,  PARAM_BOOL,     NULL, F_SYNC },
 715                 { "listenaddr", FALSE,  PARAM_STRING,   NULL, 0 },
 716                 { "maxconnections", FALSE, PARAM_INT,   NULL, 0 },
 717         };
 718         const int lp_size=sizeof(lp)/sizeof(PARAM);
 719         PARAM gp[] = {
 720                 { "user",       FALSE, PARAM_STRING,    &runuser,       0 },
 721                 { "group",      FALSE, PARAM_STRING,    &rungroup,      0 },
 722                 { "oldstyle",   FALSE, PARAM_BOOL,      &do_oldstyle,   1 },
 723                 { "listenaddr", FALSE, PARAM_STRING,    &modern_listen, 0 },
 724         };
 725         PARAM* p=gp;
 726         int p_size=sizeof(gp)/sizeof(PARAM);
 727         GKeyFile *cfile;
 728         GError *err = NULL;
 729         const char *err_msg=NULL;
 730         GQuark errdomain;
 731         GArray *retval=NULL;
 732         gchar **groups;
 733         gboolean value;
 734         gchar* startgroup;
 735         gint i;
 736         gint j;
 737
 738         errdomain = g_quark_from_string("parse_cfile");
 739         cfile = g_key_file_new();
 740         retval = g_array_new(FALSE, TRUE, sizeof(SERVER));
 741         if(!g_key_file_load_from_file(cfile, f, G_KEY_FILE_KEEP_COMMENTS |
 742                         G_KEY_FILE_KEEP_TRANSLATIONS, &err)) {
 743                 g_set_error(e, errdomain, CFILE_NOTFOUND, "Could not open config file %s.", f);
 744                 g_key_file_free(cfile);
 745                 return retval;
 746         }
 747         startgroup = g_key_file_get_start_group(cfile);
 748         if(!startgroup || strcmp(startgroup, "generic")) {
 749                 g_set_error(e, errdomain, CFILE_MISSING_GENERIC, "Config file does not contain the [generic] group!");
 750                 g_key_file_free(cfile);
 751                 return NULL;
 752         }
 753         groups = g_key_file_get_groups(cfile, NULL);
 754         for(i=0;groups[i];i++) {
 755                 memset(&s, '\0', sizeof(SERVER));
 756                 lp[0].target=&(s.exportname);
 757                 lp[1].target=&(s.port);
 758                 lp[2].target=&(s.authname);
 759                 lp[3].target=&(s.expected_size);
 760                 lp[4].target=&(virtstyle);
 761                 lp[5].target=&(s.prerun);
 762                 lp[6].target=&(s.postrun);
 763                 lp[7].target=lp[8].target=lp[9].target=
 764                                 lp[10].target=lp[11].target=
 765                                 lp[12].target=&(s.flags);
 766                 lp[13].target=&(s.listenaddr);
 767                 lp[14].target=&(s.max_connections);
 768
 769                 /* After the [generic] group, start parsing exports */
 770                 if(i==1) {
 771                         p=lp;
 772                         p_size=lp_size;
 773                 }
 774                 for(j=0;j<p_size;j++) {
 775                         g_assert(p[j].target != NULL);
 776                         g_assert(p[j].ptype==PARAM_INT||p[j].ptype==PARAM_STRING||p[j].ptype==PARAM_BOOL);
 777                         switch(p[j].ptype) {
 778                                 case PARAM_INT:
 779                                         *((gint*)p[j].target) =
 780                                                 g_key_file_get_integer(cfile,
 781                                                                 groups[i],
 782                                                                 p[j].paramname,
 783                                                                 &err);
 784                                         break;
 785                                 case PARAM_STRING:
 786                                         *((gchar**)p[j].target) =
 787                                                 g_key_file_get_string(cfile,
 788                                                                 groups[i],
 789                                                                 p[j].paramname,
 790                                                                 &err);
 791                                         break;
 792                                 case PARAM_BOOL:
 793                                         value = g_key_file_get_boolean(cfile,
 794                                                         groups[i],
 795                                                         p[j].paramname, &err);
 796                                         if(!err) {
 797                                                 if(value) {
 798                                                         *((gint*)p[j].target) |= p[j].flagval;
 799                                                 } else {
 800                                                         *((gint*)p[j].target) &= ~(p[j].flagval);
 801                                                 }
 802                                         }
 803                                         break;
 804                         }
 805                         if(!strcmp(p[j].paramname, "port") && !strcmp(p[j].target, NBD_DEFAULT_PORT)) {
 806                                 g_set_error(e, errdomain, CFILE_INCORRECT_PORT, "Config file specifies default port for oldstyle export");
 807                                 g_key_file_free(cfile);
 808                                 return NULL;
 809                         }
 810                         if(err) {
 811                                 if(err->code == G_KEY_FILE_ERROR_KEY_NOT_FOUND) {
 812                                         if(!p[j].required) {
 813                                                 /* Ignore not-found error for optional values */
 814                                                 g_clear_error(&err);
 815                                                 continue;
 816                                         } else {
 817                                                 err_msg = MISSING_REQUIRED_ERROR;
 818                                         }
 819                                 } else {
 820                                         err_msg = DEFAULT_ERROR;
 821                                 }
 822                                 g_set_error(e, errdomain, CFILE_VALUE_INVALID, err_msg, p[j].paramname, groups[i], err->message);
 823                                 g_array_free(retval, TRUE);
 824                                 g_error_free(err);
 825                                 g_key_file_free(cfile);
 826                                 return NULL;
 827                         }
 828                 }
 829                 if(virtstyle) {
 830                         if(!strncmp(virtstyle, "none", 4)) {
 831                                 s.virtstyle=VIRT_NONE;
 832                         } else if(!strncmp(virtstyle, "ipliteral", 9)) {
 833                                 s.virtstyle=VIRT_IPLIT;
 834                         } else if(!strncmp(virtstyle, "iphash", 6)) {
 835                                 s.virtstyle=VIRT_IPHASH;
 836                         } else if(!strncmp(virtstyle, "cidrhash", 8)) {
 837                                 s.virtstyle=VIRT_CIDR;
 838                                 if(strlen(virtstyle)<10) {
 839                                         g_set_error(e, errdomain, CFILE_VALUE_INVALID, "Invalid value %s for parameter virtstyle in group %s: missing length", virtstyle, groups[i]);
 840                                         g_array_free(retval, TRUE);
 841                                         g_key_file_free(cfile);
 842                                         return NULL;
 843                                 }
 844                                 s.cidrlen=strtol(virtstyle+8, NULL, 0);
 845                         } else {
 846                                 g_set_error(e, errdomain, CFILE_VALUE_INVALID, "Invalid value %s for parameter virtstyle in group %s", virtstyle, groups[i]);
 847                                 g_array_free(retval, TRUE);
 848                                 g_key_file_free(cfile);
 849                                 return NULL;
 850                         }
 851                         if(s.port && !do_oldstyle) {
 852                                 g_warning("A port was specified, but oldstyle exports were not requested. This may not do what you expect.");
 853                                 g_warning("Please read 'man 5 nbd-server' and search for oldstyle for more info");
 854                         }
 855                 } else {
 856                         s.virtstyle=VIRT_IPLIT;
 857                 }
 858                 /* Don't need to free this, it's not our string */
 859                 virtstyle=NULL;
 860                 /* Don't append values for the [generic] group */
 861                 if(i>0) {
 862                         s.socket_family = AF_UNSPEC;
 863                         s.servename = groups[i];
 864
 865                         append_serve(&s, retval);
 866                 } else {
 867                         if(!do_oldstyle) {
 868                                 lp[1].required = 0;
 869                         }
 870                 }
 871 #ifndef WITH_SDP
 872                 if(s.flags & F_SDP) {
 873                         g_set_error(e, errdomain, CFILE_VALUE_UNSUPPORTED, "This nbd-server was built without support for SDP, yet group %s uses it", groups[i]);
 874                         g_array_free(retval, TRUE);
 875                         g_key_file_free(cfile);
 876                         return NULL;
 877                 }
 878 #endif
 879         }
 880         if(i==1) {
 881                 g_set_error(e, errdomain, CFILE_NO_EXPORTS, "The config file does not specify any exports");
 882         }
 883         g_key_file_free(cfile);
 884         return retval;
 885 }
 886
 887 /**
 888  * Signal handler for SIGCHLD
 889  * @param s the signal we're handling (must be SIGCHLD, or something
 890  * is severely wrong)
 891  **/
 892 void sigchld_handler(int s) {
 893         int status;
 894         int* i;
 895         pid_t pid;
 896
 897         while((pid=waitpid(-1, &status, WNOHANG)) > 0) {
 898                 if(WIFEXITED(status)) {
 899                         msg3(LOG_INFO, "Child exited with %d", WEXITSTATUS(status));
 900                 }
 901                 i=g_hash_table_lookup(children, &pid);
 902                 if(!i) {
 903                         msg3(LOG_INFO, "SIGCHLD received for an unknown child with PID %ld", (long)pid);
 904                 } else {
 905                         DEBUG2("Removing %d from the list of children", pid);
 906                         g_hash_table_remove(children, &pid);
 907                 }
 908         }
 909 }
 910
 911 /**
 912  * Kill a child. Called from sigterm_handler::g_hash_table_foreach.
 913  *
 914  * @param key the key
 915  * @param value the value corresponding to the above key
 916  * @param user_data a pointer which we always set to 1, so that we know what
 917  * will happen next.
 918  **/
 919 void killchild(gpointer key, gpointer value, gpointer user_data) {
 920         pid_t *pid=value;
 921         int *parent=user_data;
 922
 923         kill(*pid, SIGTERM);
 924         *parent=1;
 925 }
 926
 927 /**
 928  * Handle SIGTERM and dispatch it to our children
 929  * @param s the signal we're handling (must be SIGTERM, or something
 930  * is severely wrong).
 931  **/
 932 void sigterm_handler(int s) {
 933         int parent=0;
 934
 935         g_hash_table_foreach(children, killchild, &parent);
 936
 937         if(parent) {
 938                 unlink(pidfname);
 939         }
 940
 941         exit(EXIT_SUCCESS);
 942 }
 943
 944 /**
 945  * Detect the size of a file.
 946  *
 947  * @param fhandle An open filedescriptor
 948  * @return the size of the file, or OFFT_MAX if detection was
 949  * impossible.
 950  **/
 951 off_t size_autodetect(int fhandle) {
 952         off_t es;
 953         u64 bytes;
 954         struct stat stat_buf;
 955         int error;
 956
 957 #ifdef HAVE_SYS_MOUNT_H
 958 #ifdef HAVE_SYS_IOCTL_H
 959 #ifdef BLKGETSIZE64
 960         DEBUG("looking for export size with ioctl BLKGETSIZE64\n");
 961         if (!ioctl(fhandle, BLKGETSIZE64, &bytes) && bytes) {
 962                 return (off_t)bytes;
 963         }
 964 #endif /* BLKGETSIZE64 */
 965 #endif /* HAVE_SYS_IOCTL_H */
 966 #endif /* HAVE_SYS_MOUNT_H */
 967
 968         DEBUG("looking for fhandle size with fstat\n");
 969         stat_buf.st_size = 0;
 970         error = fstat(fhandle, &stat_buf);
 971         if (!error) {
 972                 if(stat_buf.st_size > 0)
 973                         return (off_t)stat_buf.st_size;
 974         } else {
 975                 err("fstat failed: %m");
 976         }
 977
 978         DEBUG("looking for fhandle size with lseek SEEK_END\n");
 979         es = lseek(fhandle, (off_t)0, SEEK_END);
 980         if (es > ((off_t)0)) {
 981                 return es;
 982         } else {
 983                 DEBUG2("lseek failed: %d", errno==EBADF?1:(errno==ESPIPE?2:(errno==EINVAL?3:4)));
 984         }
 985
 986         err("Could not find size of exported block device: %m");
 987         return OFFT_MAX;
 988 }
 989
 990 /**
 991  * Get the file handle and offset, given an export offset.
 992  *
 993  * @param export An array of export files
 994  * @param a The offset to get corresponding file/offset for
 995  * @param fhandle [out] File descriptor
 996  * @param foffset [out] Offset into fhandle
 997  * @param maxbytes [out] Tells how many bytes can be read/written
 998  * from fhandle starting at foffset (0 if there is no limit)
 999  * @return 0 on success, -1 on failure
1000  **/
1001 int get_filepos(GArray* export, off_t a, int* fhandle, off_t* foffset, size_t* maxbytes ) {
1002         /* Negative offset not allowed */
1003         if(a < 0)
1004                 return -1;
1005
1006         /* Binary search for last file with starting offset <= a */
1007         FILE_INFO fi;
1008         int start = 0;
1009         int end = export->len - 1;
1010         while( start <= end ) {
1011                 int mid = (start + end) / 2;
1012                 fi = g_array_index(export, FILE_INFO, mid);
1013                 if( fi.startoff < a ) {
1014                         start = mid + 1;
1015                 } else if( fi.startoff > a ) {
1016                         end = mid - 1;
1017                 } else {
1018                         start = end = mid;
1019                         break;
1020                 }
1021         }
1022
1023         /* end should never go negative, since first startoff is 0 and a >= 0 */
1024         g_assert(end >= 0);
1025
1026         fi = g_array_index(export, FILE_INFO, end);
1027         *fhandle = fi.fhandle;
1028         *foffset = a - fi.startoff;
1029         *maxbytes = 0;
1030         if( end+1 < export->len ) {
1031                 FILE_INFO fi_next = g_array_index(export, FILE_INFO, end+1);
1032                 *maxbytes = fi_next.startoff - a;
1033         }
1034
1035         return 0;
1036 }
1037
1038 /**
1039  * seek to a position in a file, with error handling.
1040  * @param handle a filedescriptor
1041  * @param a position to seek to
1042  * @todo get rid of this; lastpoint is a global variable right now, but it
1043  * shouldn't be. If we pass it on as a parameter, that makes things a *lot*
1044  * easier.
1045  **/
1046 void myseek(int handle,off_t a) {
1047         if (lseek(handle, a, SEEK_SET) < 0) {
1048                 err("Can not seek locally!\n");
1049         }
1050 }
1051
1052 /**
1053  * Write an amount of bytes at a given offset to the right file. This
1054  * abstracts the write-side of the multiple file option.
1055  *
1056  * @param a The offset where the write should start
1057  * @param buf The buffer to write from
1058  * @param len The length of buf
1059  * @param client The client we're serving for
1060  * @return The number of bytes actually written, or -1 in case of an error
1061  **/
1062 ssize_t rawexpwrite(off_t a, char *buf, size_t len, CLIENT *client) {
1063         int fhandle;
1064         off_t foffset;
1065         size_t maxbytes;
1066         ssize_t retval;
1067
1068         if(get_filepos(client->export, a, &fhandle, &foffset, &maxbytes))
1069                 return -1;
1070         if(maxbytes && len > maxbytes)
1071                 len = maxbytes;
1072
1073         DEBUG4("(WRITE to fd %d offset %llu len %u), ", fhandle, foffset, len);
1074
1075         myseek(fhandle, foffset);
1076         retval = write(fhandle, buf, len);
1077         if(client->server->flags & F_SYNC) {
1078                 fsync(fhandle);
1079         }
1080         return retval;
1081 }
1082
1083 /**
1084  * Call rawexpwrite repeatedly until all data has been written.
1085  * @return 0 on success, nonzero on failure
1086  **/
1087 int rawexpwrite_fully(off_t a, char *buf, size_t len, CLIENT *client) {
1088         ssize_t ret=0;
1089
1090         while(len > 0 && (ret=rawexpwrite(a, buf, len, client)) > 0 ) {
1091                 a += ret;
1092                 buf += ret;
1093                 len -= ret;
1094         }
1095         return (ret < 0 || len != 0);
1096 }
1097
1098 /**
1099  * Read an amount of bytes at a given offset from the right file. This
1100  * abstracts the read-side of the multiple files option.
1101  *
1102  * @param a The offset where the read should start
1103  * @param buf A buffer to read into
1104  * @param len The size of buf
1105  * @param client The client we're serving for
1106  * @return The number of bytes actually read, or -1 in case of an
1107  * error.
1108  **/
1109 ssize_t rawexpread(off_t a, char *buf, size_t len, CLIENT *client) {
1110         int fhandle;
1111         off_t foffset;
1112         size_t maxbytes;
1113
1114         if(get_filepos(client->export, a, &fhandle, &foffset, &maxbytes))
1115                 return -1;
1116         if(maxbytes && len > maxbytes)
1117                 len = maxbytes;
1118
1119         DEBUG4("(READ from fd %d offset %llu len %u), ", fhandle, foffset, len);
1120
1121         myseek(fhandle, foffset);
1122         return read(fhandle, buf, len);
1123 }
1124
1125 /**
1126  * Call rawexpread repeatedly until all data has been read.
1127  * @return 0 on success, nonzero on failure
1128  **/
1129 int rawexpread_fully(off_t a, char *buf, size_t len, CLIENT *client) {
1130         ssize_t ret=0;
1131
1132         while(len > 0 && (ret=rawexpread(a, buf, len, client)) > 0 ) {
1133                 a += ret;
1134                 buf += ret;
1135                 len -= ret;
1136         }
1137         return (ret < 0 || len != 0);
1138 }
1139
1140 /**
1141  * Read an amount of bytes at a given offset from the right file. This
1142  * abstracts the read-side of the copyonwrite stuff, and calls
1143  * rawexpread() with the right parameters to do the actual work.
1144  * @param a The offset where the read should start
1145  * @param buf A buffer to read into
1146  * @param len The size of buf
1147  * @param client The client we're going to read for
1148  * @return 0 on success, nonzero on failure
1149  **/
1150 int expread(off_t a, char *buf, size_t len, CLIENT *client) {
1151         off_t rdlen, offset;
1152         off_t mapcnt, mapl, maph, pagestart;
1153
1154         if (!(client->server->flags & F_COPYONWRITE))
1155                 return(rawexpread_fully(a, buf, len, client));
1156         DEBUG3("Asked to read %d bytes at %llu.\n", len, (unsigned long long)a);
1157
1158         mapl=a/DIFFPAGESIZE; maph=(a+len-1)/DIFFPAGESIZE;
1159
1160         for (mapcnt=mapl;mapcnt<=maph;mapcnt++) {
1161                 pagestart=mapcnt*DIFFPAGESIZE;
1162                 offset=a-pagestart;
1163                 rdlen=(0<DIFFPAGESIZE-offset && len<(size_t)(DIFFPAGESIZE-offset)) ?
1164                         len : (size_t)DIFFPAGESIZE-offset;
1165                 if (client->difmap[mapcnt]!=(u32)(-1)) { /* the block is already there */
1166                         DEBUG3("Page %llu is at %lu\n", (unsigned long long)mapcnt,
1167                                (unsigned long)(client->difmap[mapcnt]));
1168                         myseek(client->difffile, client->difmap[mapcnt]*DIFFPAGESIZE+offset);
1169                         if (read(client->difffile, buf, rdlen) != rdlen) return -1;
1170                 } else { /* the block is not there */
1171                         DEBUG2("Page %llu is not here, we read the original one\n",
1172                                (unsigned long long)mapcnt);
1173                         if(rawexpread_fully(a, buf, rdlen, client)) return -1;
1174                 }
1175                 len-=rdlen; a+=rdlen; buf+=rdlen;
1176         }
1177         return 0;
1178 }
1179
1180 /**
1181  * Write an amount of bytes at a given offset to the right file. This
1182  * abstracts the write-side of the copyonwrite option, and calls
1183  * rawexpwrite() with the right parameters to do the actual work.
1184  *
1185  * @param a The offset where the write should start
1186  * @param buf The buffer to write from
1187  * @param len The length of buf
1188  * @param client The client we're going to write for.
1189  * @return 0 on success, nonzero on failure
1190  **/
1191 int expwrite(off_t a, char *buf, size_t len, CLIENT *client) {
1192         char pagebuf[DIFFPAGESIZE];
1193         off_t mapcnt,mapl,maph;
1194         off_t wrlen,rdlen;
1195         off_t pagestart;
1196         off_t offset;
1197
1198         if (!(client->server->flags & F_COPYONWRITE))
1199                 return(rawexpwrite_fully(a, buf, len, client));
1200         DEBUG3("Asked to write %d bytes at %llu.\n", len, (unsigned long long)a);
1201
1202         mapl=a/DIFFPAGESIZE ; maph=(a+len-1)/DIFFPAGESIZE ;
1203
1204         for (mapcnt=mapl;mapcnt<=maph;mapcnt++) {
1205                 pagestart=mapcnt*DIFFPAGESIZE ;
1206                 offset=a-pagestart ;
1207                 wrlen=(0<DIFFPAGESIZE-offset && len<(size_t)(DIFFPAGESIZE-offset)) ?
1208                         len : (size_t)DIFFPAGESIZE-offset;
1209
1210                 if (client->difmap[mapcnt]!=(u32)(-1)) { /* the block is already there */
1211                         DEBUG3("Page %llu is at %lu\n", (unsigned long long)mapcnt,
1212                                (unsigned long)(client->difmap[mapcnt])) ;
1213                         myseek(client->difffile,
1214                                         client->difmap[mapcnt]*DIFFPAGESIZE+offset);
1215                         if (write(client->difffile, buf, wrlen) != wrlen) return -1 ;
1216                 } else { /* the block is not there */
1217                         myseek(client->difffile,client->difffilelen*DIFFPAGESIZE) ;
1218                         client->difmap[mapcnt]=(client->server->flags&F_SPARSE)?mapcnt:client->difffilelen++;
1219                         DEBUG3("Page %llu is not here, we put it at %lu\n",
1220                                (unsigned long long)mapcnt,
1221                                (unsigned long)(client->difmap[mapcnt]));
1222                         rdlen=DIFFPAGESIZE ;
1223                         if (rawexpread_fully(pagestart, pagebuf, rdlen, client))
1224                                 return -1;
1225                         memcpy(pagebuf+offset,buf,wrlen) ;
1226                         if (write(client->difffile, pagebuf, DIFFPAGESIZE) !=
1227                                         DIFFPAGESIZE)
1228                                 return -1;
1229                 }
1230                 len-=wrlen ; a+=wrlen ; buf+=wrlen ;
1231         }
1232         return 0;
1233 }
1234
1235 /**
1236  * Do the initial negotiation.
1237  *
1238  * @param client The client we're negotiating with.
1239  **/
1240 CLIENT* negotiate(int net, CLIENT *client, GArray* servers) {
1241         char zeros[128];
1242         uint64_t size_host;
1243         uint32_t flags = NBD_FLAG_HAS_FLAGS;
1244         uint16_t smallflags = 0;
1245         uint64_t magic;
1246
1247         memset(zeros, '\0', sizeof(zeros));
1248         if(!client || !client->modern) {
1249                 /* common */
1250                 if (write(net, INIT_PASSWD, 8) < 0) {
1251                         err_nonfatal("Negotiation failed: %m");
1252                         if(client)
1253                                 exit(EXIT_FAILURE);
1254                 }
1255                 if(!client || client->modern) {
1256                         /* modern */
1257                         magic = htonll(opts_magic);
1258                 } else {
1259                         /* oldstyle */
1260                         magic = htonll(cliserv_magic);
1261                 }
1262                 if (write(net, &magic, sizeof(magic)) < 0) {
1263                         err_nonfatal("Negotiation failed: %m");
1264                         if(client)
1265                                 exit(EXIT_FAILURE);
1266                 }
1267         }
1268         if(!client) {
1269                 /* modern */
1270                 uint32_t reserved;
1271                 uint32_t opt;
1272                 uint32_t namelen;
1273                 char* name;
1274                 int i;
1275
1276                 if(!servers)
1277                         err("programmer error");
1278                 if (write(net, &smallflags, sizeof(uint16_t)) < 0)
1279                         err("Negotiation failed: %m");
1280                 if (read(net, &reserved, sizeof(reserved)) < 0)
1281                         err("Negotiation failed: %m");
1282                 if (read(net, &magic, sizeof(magic)) < 0)
1283                         err("Negotiation failed: %m");
1284                 magic = ntohll(magic);
1285                 if(magic != opts_magic) {
1286                         close(net);
1287                         return NULL;
1288                 }
1289                 if (read(net, &opt, sizeof(opt)) < 0)
1290                         err("Negotiation failed: %m");
1291                 opt = ntohl(opt);
1292                 if(opt != NBD_OPT_EXPORT_NAME) {
1293                         close(net);
1294                         return NULL;
1295                 }
1296                 if (read(net, &namelen, sizeof(namelen)) < 0)
1297                         err("Negotiation failed: %m");
1298                 namelen = ntohl(namelen);
1299                 name = malloc(namelen+1);
1300                 name[namelen]=0;
1301                 if (read(net, name, namelen) < 0)
1302                         err("Negotiation failed: %m");
1303                 for(i=0; i<servers->len; i++) {
1304                         SERVER* serve = &(g_array_index(servers, SERVER, i));
1305                         if(!strcmp(serve->servename, name)) {
1306                                 CLIENT* client = g_new0(CLIENT, 1);
1307                                 client->server = serve;
1308                                 client->exportsize = OFFT_MAX;
1309                                 client->net = net;
1310                                 client->modern = TRUE;
1311                                 return client;
1312                         }
1313                 }
1314                 return NULL;
1315         }
1316         /* common */
1317         size_host = htonll((u64)(client->exportsize));
1318         if (write(net, &size_host, 8) < 0)
1319                 err("Negotiation failed: %m");
1320         if (client->server->flags & F_READONLY)
1321                 flags |= NBD_FLAG_READ_ONLY;
1322         if (!client->modern) {
1323                 /* oldstyle */
1324                 flags = htonl(flags);
1325                 if (write(client->net, &flags, 4) < 0)
1326                         err("Negotiation failed: %m");
1327         } else {
1328                 /* modern */
1329                 smallflags = (uint16_t)(flags & ~((uint16_t)0));
1330                 smallflags = htons(smallflags);
1331                 if (write(client->net, &smallflags, sizeof(smallflags)) < 0) {
1332                         err("Negotiation failed: %m");
1333                 }
1334         }
1335         /* common */
1336         if (write(client->net, zeros, 124) < 0)
1337                 err("Negotiation failed: %m");
1338         return NULL;
1339 }
1340
1341 /** sending macro. */
1342 #define SEND(net,reply) writeit( net, &reply, sizeof( reply ));
1343 /** error macro. */
1344 #define ERROR(client,reply,errcode) { reply.error = htonl(errcode); SEND(client->net,reply); reply.error = 0; }
1345 /**
1346  * Serve a file to a single client.
1347  *
1348  * @todo This beast needs to be split up in many tiny little manageable
1349  * pieces. Preferably with a chainsaw.
1350  *
1351  * @param client The client we're going to serve to.
1352  * @return when the client disconnects
1353  **/
1354 int mainloop(CLIENT *client) {
1355         struct nbd_request request;
1356         struct nbd_reply reply;
1357         gboolean go_on=TRUE;
1358 #ifdef DODBG
1359         int i = 0;
1360 #endif
1361         negotiate(client->net, client, NULL);
1362         DEBUG("Entering request loop!\n");
1363         reply.magic = htonl(NBD_REPLY_MAGIC);
1364         reply.error = 0;
1365         while (go_on) {
1366                 char buf[BUFSIZE];
1367                 size_t len;
1368 #ifdef DODBG
1369                 i++;
1370                 printf("%d: ", i);
1371 #endif
1372                 readit(client->net, &request, sizeof(request));
1373                 request.from = ntohll(request.from);
1374                 request.type = ntohl(request.type);
1375
1376                 if (request.type==NBD_CMD_DISC) {
1377                         msg2(LOG_INFO, "Disconnect request received.");
1378                         if (client->server->flags & F_COPYONWRITE) {
1379                                 if (client->difmap) g_free(client->difmap) ;
1380                                 close(client->difffile);
1381                                 unlink(client->difffilename);
1382                                 free(client->difffilename);
1383                         }
1384                         go_on=FALSE;
1385                         continue;
1386                 }
1387
1388                 len = ntohl(request.len);
1389
1390                 if (request.magic != htonl(NBD_REQUEST_MAGIC))
1391                         err("Not enough magic.");
1392                 if (len > BUFSIZE + sizeof(struct nbd_reply))
1393                         err("Request too big!");
1394 #ifdef DODBG
1395                 printf("%s from %llu (%llu) len %d, ", request.type ? "WRITE" :
1396                                 "READ", (unsigned long long)request.from,
1397                                 (unsigned long long)request.from / 512, len);
1398 #endif
1399                 memcpy(reply.handle, request.handle, sizeof(reply.handle));
1400                 if ((request.from + len) > (OFFT_MAX)) {
1401                         DEBUG("[Number too large!]");
1402                         ERROR(client, reply, EINVAL);
1403                         continue;
1404                 }
1405
1406                 if (((ssize_t)((off_t)request.from + len) > client->exportsize)) {
1407                         DEBUG("[RANGE!]");
1408                         ERROR(client, reply, EINVAL);
1409                         continue;
1410                 }
1411
1412                 if (request.type==NBD_CMD_WRITE) {
1413                         DEBUG("wr: net->buf, ");
1414                         readit(client->net, buf, len);
1415                         DEBUG("buf->exp, ");
1416                         if ((client->server->flags & F_READONLY) ||
1417                             (client->server->flags & F_AUTOREADONLY)) {
1418                                 DEBUG("[WRITE to READONLY!]");
1419                                 ERROR(client, reply, EPERM);
1420                                 continue;
1421                         }
1422                         if (expwrite(request.from, buf, len, client)) {
1423                                 DEBUG("Write failed: %m" );
1424                                 ERROR(client, reply, errno);
1425                                 continue;
1426                         }
1427                         SEND(client->net, reply);
1428                         DEBUG("OK!\n");
1429                         continue;
1430                 }
1431                 /* READ */
1432
1433                 DEBUG("exp->buf, ");
1434                 if (expread(request.from, buf + sizeof(struct nbd_reply), len, client)) {
1435                         DEBUG("Read failed: %m");
1436                         ERROR(client, reply, errno);
1437                         continue;
1438                 }
1439
1440                 DEBUG("buf->net, ");
1441                 memcpy(buf, &reply, sizeof(struct nbd_reply));
1442                 writeit(client->net, buf, len + sizeof(struct nbd_reply));
1443                 DEBUG("OK!\n");
1444         }
1445         return 0;
1446 }
1447
1448 /**
1449  * Set up client export array, which is an array of FILE_INFO.
1450  * Also, split a single exportfile into multiple ones, if that was asked.
1451  * @param client information on the client which we want to setup export for
1452  **/
1453 void setupexport(CLIENT* client) {
1454         int i;
1455         off_t laststartoff = 0, lastsize = 0;
1456         int multifile = (client->server->flags & F_MULTIFILE);
1457
1458         client->export = g_array_new(TRUE, TRUE, sizeof(FILE_INFO));
1459
1460         /* If multi-file, open as many files as we can.
1461          * If not, open exactly one file.
1462          * Calculate file sizes as we go to get total size. */
1463         for(i=0; ; i++) {
1464                 FILE_INFO fi;
1465                 gchar *tmpname;
1466                 gchar* error_string;
1467                 mode_t mode = (client->server->flags & F_READONLY) ? O_RDONLY : O_RDWR;
1468
1469                 if(multifile) {
1470                         tmpname=g_strdup_printf("%s.%d", client->exportname, i);
1471                 } else {
1472                         tmpname=g_strdup(client->exportname);
1473                 }
1474                 DEBUG2( "Opening %s\n", tmpname );
1475                 fi.fhandle = open(tmpname, mode);
1476                 if(fi.fhandle == -1 && mode == O_RDWR) {
1477                         /* Try again because maybe media was read-only */
1478                         fi.fhandle = open(tmpname, O_RDONLY);
1479                         if(fi.fhandle != -1) {
1480                                 /* Opening the base file in copyonwrite mode is
1481                                  * okay */
1482                                 if(!(client->server->flags & F_COPYONWRITE)) {
1483                                         client->server->flags |= F_AUTOREADONLY;
1484                                         client->server->flags |= F_READONLY;
1485                                 }
1486                         }
1487                 }
1488                 if(fi.fhandle == -1) {
1489                         if(multifile && i>0)
1490                                 break;
1491                         error_string=g_strdup_printf(
1492                                 "Could not open exported file %s: %%m",
1493                                 tmpname);
1494                         err(error_string);
1495                 }
1496                 fi.startoff = laststartoff + lastsize;
1497                 g_array_append_val(client->export, fi);
1498                 g_free(tmpname);
1499
1500                 /* Starting offset and size of this file will be used to
1501                  * calculate starting offset of next file */
1502                 laststartoff = fi.startoff;
1503                 lastsize = size_autodetect(fi.fhandle);
1504
1505                 if(!multifile)
1506                         break;
1507         }
1508
1509         /* Set export size to total calculated size */
1510         client->exportsize = laststartoff + lastsize;
1511
1512         /* Export size may be overridden */
1513         if(client->server->expected_size) {
1514                 /* desired size must be <= total calculated size */
1515                 if(client->server->expected_size > client->exportsize) {
1516                         err("Size of exported file is too big\n");
1517                 }
1518
1519                 client->exportsize = client->server->expected_size;
1520         }
1521
1522         msg3(LOG_INFO, "Size of exported file/device is %llu", (unsigned long long)client->exportsize);
1523         if(multifile) {
1524                 msg3(LOG_INFO, "Total number of files: %d", i);
1525         }
1526 }
1527
1528 int copyonwrite_prepare(CLIENT* client) {
1529         off_t i;
1530         if ((client->difffilename = malloc(1024))==NULL)
1531                 err("Failed to allocate string for diff file name");
1532         snprintf(client->difffilename, 1024, "%s-%s-%d.diff",client->exportname,client->clientname,
1533                 (int)getpid()) ;
1534         client->difffilename[1023]='\0';
1535         msg3(LOG_INFO,"About to create map and diff file %s",client->difffilename) ;
1536         client->difffile=open(client->difffilename,O_RDWR | O_CREAT | O_TRUNC,0600) ;
1537         if (client->difffile<0) err("Could not create diff file (%m)") ;
1538         if ((client->difmap=calloc(client->exportsize/DIFFPAGESIZE,sizeof(u32)))==NULL)
1539                 err("Could not allocate memory") ;
1540         for (i=0;i<client->exportsize/DIFFPAGESIZE;i++) client->difmap[i]=(u32)-1 ;
1541
1542         return 0;
1543 }
1544
1545 /**
1546  * Run a command. This is used for the ``prerun'' and ``postrun'' config file
1547  * options
1548  *
1549  * @param command the command to be ran. Read from the config file
1550  * @param file the file name we're about to export
1551  **/
1552 int do_run(gchar* command, gchar* file) {
1553         gchar* cmd;
1554         int retval=0;
1555
1556         if(command && *command) {
1557                 cmd = g_strdup_printf(command, file);
1558                 retval=system(cmd);
1559                 g_free(cmd);
1560         }
1561         return retval;
1562 }
1563
1564 /**
1565  * Serve a connection.
1566  *
1567  * @todo allow for multithreading, perhaps use libevent. Not just yet, though;
1568  * follow the road map.
1569  *
1570  * @param client a connected client
1571  **/
1572 void serveconnection(CLIENT *client) {
1573         if(do_run(client->server->prerun, client->exportname)) {
1574                 exit(EXIT_FAILURE);
1575         }
1576         setupexport(client);
1577
1578         if (client->server->flags & F_COPYONWRITE) {
1579                 copyonwrite_prepare(client);
1580         }
1581
1582         setmysockopt(client->net);
1583
1584         mainloop(client);
1585         do_run(client->server->postrun, client->exportname);
1586 }
1587
1588 /**
1589  * Find the name of the file we have to serve. This will use g_strdup_printf
1590  * to put the IP address of the client inside a filename containing
1591  * "%s" (in the form as specified by the "virtstyle" option). That name
1592  * is then written to client->exportname.
1593  *
1594  * @param net A socket connected to an nbd client
1595  * @param client information about the client. The IP address in human-readable
1596  * format will be written to a new char* buffer, the address of which will be
1597  * stored in client->clientname.
1598  **/
1599 void set_peername(int net, CLIENT *client) {
1600         struct sockaddr_storage addrin;
1601         struct sockaddr_storage netaddr;
1602         struct sockaddr_in  *netaddr4 = NULL;
1603         struct sockaddr_in6 *netaddr6 = NULL;
1604         size_t addrinlen = sizeof( addrin );
1605         struct addrinfo hints;
1606         struct addrinfo *ai = NULL;
1607         char peername[NI_MAXHOST];
1608         char netname[NI_MAXHOST];
1609         char *tmp = NULL;
1610         int i;
1611         int e;
1612         int shift;
1613
1614         if (getpeername(net, (struct sockaddr *) &addrin, (socklen_t *)&addrinlen) < 0)
1615                 err("getsockname failed: %m");
1616
1617         getnameinfo((struct sockaddr *)&addrin, (socklen_t)addrinlen,
1618                 peername, sizeof (peername), NULL, 0, NI_NUMERICHOST);
1619
1620         memset(&hints, '\0', sizeof (hints));
1621         hints.ai_flags = AI_ADDRCONFIG;
1622         e = getaddrinfo(peername, NULL, &hints, &ai);
1623
1624         if(e != 0) {
1625                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1626                 freeaddrinfo(ai);
1627                 return;
1628         }
1629
1630         switch(client->server->virtstyle) {
1631                 case VIRT_NONE:
1632                         client->exportname=g_strdup(client->server->exportname);
1633                         break;
1634                 case VIRT_IPHASH:
1635                         for(i=0;i<strlen(peername);i++) {
1636                                 if(peername[i]=='.') {
1637                                         peername[i]='/';
1638                                 }
1639                         }
1640                 case VIRT_IPLIT:
1641                         client->exportname=g_strdup_printf(client->server->exportname, peername);
1642                         break;
1643                 case VIRT_CIDR:
1644                         memcpy(&netaddr, &addrin, addrinlen);
1645                         if(ai->ai_family == AF_INET) {
1646                                 netaddr4 = (struct sockaddr_in *)&netaddr;
1647                                 (netaddr4->sin_addr).s_addr>>=32-(client->server->cidrlen);
1648                                 (netaddr4->sin_addr).s_addr<<=32-(client->server->cidrlen);
1649
1650                                 getnameinfo((struct sockaddr *) netaddr4, (socklen_t) addrinlen,
1651                                                         netname, sizeof (netname), NULL, 0, NI_NUMERICHOST);
1652                                 tmp=g_strdup_printf("%s/%s", netname, peername);
1653                         }else if(ai->ai_family == AF_INET6) {
1654                                 netaddr6 = (struct sockaddr_in6 *)&netaddr;
1655
1656                                 shift = 128-(client->server->cidrlen);
1657                                 i = 3;
1658                                 while(shift >= 32) {
1659                                         ((netaddr6->sin6_addr).s6_addr32[i])=0;
1660                                         shift-=32;
1661                                         i--;
1662                                 }
1663                                 (netaddr6->sin6_addr).s6_addr32[i]>>=shift;
1664                                 (netaddr6->sin6_addr).s6_addr32[i]<<=shift;
1665
1666                                 getnameinfo((struct sockaddr *)netaddr6, (socklen_t)addrinlen,
1667                                             netname, sizeof(netname), NULL, 0, NI_NUMERICHOST);
1668                                 tmp=g_strdup_printf("%s/%s", netname, peername);
1669                         }
1670
1671                         if(tmp != NULL)
1672                           client->exportname=g_strdup_printf(client->server->exportname, tmp);
1673
1674                         break;
1675         }
1676
1677         freeaddrinfo(ai);
1678         msg4(LOG_INFO, "connect from %s, assigned file is %s",
1679              peername, client->exportname);
1680         client->clientname=g_strdup(peername);
1681 }
1682
1683 /**
1684  * Destroy a pid_t*
1685  * @param data a pointer to pid_t which should be freed
1686  **/
1687 void destroy_pid_t(gpointer data) {
1688         g_free(data);
1689 }
1690
1691 /**
1692  * Loop through the available servers, and serve them. Never returns.
1693  **/
1694 int serveloop(GArray* servers) {
1695         struct sockaddr_storage addrin;
1696         socklen_t addrinlen=sizeof(addrin);
1697         int i;
1698         int max;
1699         int sock;
1700         fd_set mset;
1701         fd_set rset;
1702
1703         /*
1704          * Set up the master fd_set. The set of descriptors we need
1705          * to select() for never changes anyway and it buys us a *lot*
1706          * of time to only build this once. However, if we ever choose
1707          * to not fork() for clients anymore, we may have to revisit
1708          * this.
1709          */
1710         max=0;
1711         FD_ZERO(&mset);
1712         for(i=0;i<servers->len;i++) {
1713                 if((sock=(g_array_index(servers, SERVER, i)).socket)) {
1714                         FD_SET(sock, &mset);
1715                         max=sock>max?sock:max;
1716                 }
1717         }
1718         if(modernsock) {
1719                 FD_SET(modernsock, &mset);
1720                 max=modernsock>max?modernsock:max;
1721         }
1722         for(;;) {
1723                 CLIENT *client = NULL;
1724                 pid_t *pid;
1725
1726                 memcpy(&rset, &mset, sizeof(fd_set));
1727                 if(select(max+1, &rset, NULL, NULL, NULL)>0) {
1728                         int net = 0;
1729                         SERVER* serve;
1730
1731                         DEBUG("accept, ");
1732                         if(FD_ISSET(modernsock, &rset)) {
1733                                 if((net=accept(modernsock, (struct sockaddr *) &addrin, &addrinlen)) < 0)
1734                                         err("accept: %m");
1735                                 client = negotiate(net, NULL, servers);
1736                                 if(!client) {
1737                                         err_nonfatal("negotiation failed");
1738                                         close(net);
1739                                         net=0;
1740                                 }
1741                         }
1742                         for(i=0;i<servers->len && !net;i++) {
1743                                 serve=&(g_array_index(servers, SERVER, i));
1744                                 if(FD_ISSET(serve->socket, &rset)) {
1745                                         if ((net=accept(serve->socket, (struct sockaddr *) &addrin, &addrinlen)) < 0)
1746                                                 err("accept: %m");
1747                                 }
1748                         }
1749                         if(net) {
1750                                 int sock_flags;
1751
1752                                 if(serve->max_connections > 0 &&
1753                                    g_hash_table_size(children) >= serve->max_connections) {
1754                                         msg2(LOG_INFO, "Max connections reached");
1755                                         close(net);
1756                                         continue;
1757                                 }
1758                                 if((sock_flags = fcntl(net, F_GETFL, 0))==-1) {
1759                                         err("fcntl F_GETFL");
1760                                 }
1761                                 if(fcntl(net, F_SETFL, sock_flags &~O_NONBLOCK)==-1) {
1762                                         err("fcntl F_SETFL ~O_NONBLOCK");
1763                                 }
1764                                 if(!client) {
1765                                         client = g_new0(CLIENT, 1);
1766                                         client->server=serve;
1767                                         client->exportsize=OFFT_MAX;
1768                                         client->net=net;
1769                                 }
1770                                 set_peername(net, client);
1771                                 if (!authorized_client(client)) {
1772                                         msg2(LOG_INFO,"Unauthorized client") ;
1773                                         close(net);
1774                                         continue;
1775                                 }
1776                                 msg2(LOG_INFO,"Authorized client") ;
1777                                 pid=g_malloc(sizeof(pid_t));
1778 #ifndef NOFORK
1779                                 if ((*pid=fork())<0) {
1780                                         msg3(LOG_INFO,"Could not fork (%s)",strerror(errno)) ;
1781                                         close(net);
1782                                         continue;
1783                                 }
1784                                 if (*pid>0) { /* parent */
1785                                         close(net);
1786                                         g_hash_table_insert(children, pid, pid);
1787                                         continue;
1788                                 }
1789                                 /* child */
1790                                 g_hash_table_destroy(children);
1791                                 for(i=0;i<servers->len;i++) {
1792                                         serve=&g_array_index(servers, SERVER, i);
1793                                         close(serve->socket);
1794                                 }
1795                                 /* FALSE does not free the
1796                                 actual data. This is required,
1797                                 because the client has a
1798                                 direct reference into that
1799                                 data, and otherwise we get a
1800                                 segfault... */
1801                                 g_array_free(servers, FALSE);
1802 #endif // NOFORK
1803                                 msg2(LOG_INFO,"Starting to serve");
1804                                 serveconnection(client);
1805                                 exit(EXIT_SUCCESS);
1806                         }
1807                 }
1808         }
1809 }
1810
1811 void dosockopts(int socket) {
1812 #ifndef sun
1813         int yes=1;
1814 #else
1815         char yes='1';
1816 #endif /* sun */
1817         int sock_flags;
1818
1819         /* lose the pesky "Address already in use" error message */
1820         if (setsockopt(socket,SOL_SOCKET,SO_REUSEADDR,&yes,sizeof(int)) == -1) {
1821                 err("setsockopt SO_REUSEADDR");
1822         }
1823         if (setsockopt(socket,SOL_SOCKET,SO_KEEPALIVE,&yes,sizeof(int)) == -1) {
1824                 err("setsockopt SO_KEEPALIVE");
1825         }
1826
1827         /* make the listening socket non-blocking */
1828         if ((sock_flags = fcntl(socket, F_GETFL, 0)) == -1) {
1829                 err("fcntl F_GETFL");
1830         }
1831         if (fcntl(socket, F_SETFL, sock_flags | O_NONBLOCK) == -1) {
1832                 err("fcntl F_SETFL O_NONBLOCK");
1833         }
1834 }
1835
1836 /**
1837  * Connect a server's socket.
1838  *
1839  * @param serve the server we want to connect.
1840  **/
1841 int setup_serve(SERVER *serve) {
1842         struct addrinfo hints;
1843         struct addrinfo *ai = NULL;
1844         gchar *port = NULL;
1845         int e;
1846
1847         if(!do_oldstyle) {
1848                 return serve->servename ? 1 : 0;
1849         }
1850         memset(&hints,'\0',sizeof(hints));
1851         hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG | AI_NUMERICSERV;
1852         hints.ai_socktype = SOCK_STREAM;
1853         hints.ai_family = serve->socket_family;
1854
1855         port = g_strdup_printf ("%d", serve->port);
1856         if (port == NULL)
1857                 return 0;
1858
1859         e = getaddrinfo(serve->listenaddr,port,&hints,&ai);
1860
1861         g_free(port);
1862
1863         if(e != 0) {
1864                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1865                 serve->socket = -1;
1866                 freeaddrinfo(ai);
1867                 exit(EXIT_FAILURE);
1868         }
1869
1870         if(serve->socket_family == AF_UNSPEC)
1871                 serve->socket_family = ai->ai_family;
1872
1873 #ifdef WITH_SDP
1874         if ((serve->flags) && F_SDP) {
1875                 if (ai->ai_family == AF_INET)
1876                         ai->ai_family = AF_INET_SDP;
1877                 else (ai->ai_family == AF_INET6)
1878                         ai->ai_family = AF_INET6_SDP;
1879         }
1880 #endif
1881         if ((serve->socket = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol)) < 0)
1882                 err("socket: %m");
1883
1884         dosockopts(serve->socket);
1885
1886         DEBUG("Waiting for connections... bind, ");
1887         e = bind(serve->socket, ai->ai_addr, ai->ai_addrlen);
1888         if (e != 0 && errno != EADDRINUSE)
1889                 err("bind: %m");
1890         DEBUG("listen, ");
1891         if (listen(serve->socket, 1) < 0)
1892                 err("listen: %m");
1893
1894         freeaddrinfo (ai);
1895         if(serve->servename) {
1896                 return 1;
1897         } else {
1898                 return 0;
1899         }
1900 }
1901
1902 void open_modern(void) {
1903         struct addrinfo hints;
1904         struct addrinfo* ai = NULL;
1905         struct sock_flags;
1906         int e;
1907
1908         memset(&hints, '\0', sizeof(hints));
1909         hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG;
1910         hints.ai_socktype = SOCK_STREAM;
1911         hints.ai_family = AF_UNSPEC;
1912         hints.ai_protocol = IPPROTO_TCP;
1913         e = getaddrinfo(modern_listen, NBD_DEFAULT_PORT, &hints, &ai);
1914         if(e != 0) {
1915                 fprintf(stderr, "getaddrinfo failed: %s\n", gai_strerror(e));
1916                 exit(EXIT_FAILURE);
1917         }
1918         if((modernsock = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol))<0) {
1919                 err("socket: %m");
1920         }
1921
1922         dosockopts(modernsock);
1923
1924         if(bind(modernsock, ai->ai_addr, ai->ai_addrlen)) {
1925                 err("bind: %m");
1926         }
1927         if(listen(modernsock, 10) <0) {
1928                 err("listen: %m");
1929         }
1930
1931         freeaddrinfo(ai);
1932 }
1933
1934 /**
1935  * Connect our servers.
1936  **/
1937 void setup_servers(GArray* servers) {
1938         int i;
1939         struct sigaction sa;
1940         int want_modern=0;
1941
1942         for(i=0;i<servers->len;i++) {
1943                 want_modern |= setup_serve(&(g_array_index(servers, SERVER, i)));
1944         }
1945         if(want_modern) {
1946                 open_modern();
1947         }
1948         children=g_hash_table_new_full(g_int_hash, g_int_equal, NULL, destroy_pid_t);
1949
1950         sa.sa_handler = sigchld_handler;
1951         sigemptyset(&sa.sa_mask);
1952         sa.sa_flags = SA_RESTART;
1953         if(sigaction(SIGCHLD, &sa, NULL) == -1)
1954                 err("sigaction: %m");
1955         sa.sa_handler = sigterm_handler;
1956         sigemptyset(&sa.sa_mask);
1957         sa.sa_flags = SA_RESTART;
1958         if(sigaction(SIGTERM, &sa, NULL) == -1)
1959                 err("sigaction: %m");
1960 }
1961
1962 /**
1963  * Go daemon (unless we specified at compile time that we didn't want this)
1964  * @param serve the first server of our configuration. If its port is zero,
1965  *      then do not daemonize, because we're doing inetd then. This parameter
1966  *      is only used to create a PID file of the form
1967  *      /var/run/nbd-server.&lt;port&gt;.pid; it's not modified in any way.
1968  **/
1969 #if !defined(NODAEMON) && !defined(NOFORK)
1970 void daemonize(SERVER* serve) {
1971         FILE*pidf;
1972
1973         if(serve && !(serve->port)) {
1974                 return;
1975         }
1976         if(daemon(0,0)<0) {
1977                 err("daemon");
1978         }
1979         if(!*pidftemplate) {
1980                 if(serve) {
1981                         strncpy(pidftemplate, "/var/run/nbd-server.%d.pid", 255);
1982                 } else {
1983                         strncpy(pidftemplate, "/var/run/nbd-server.pid", 255);
1984                 }
1985         }
1986         snprintf(pidfname, 255, pidftemplate, serve ? serve->port : 0);
1987         pidf=fopen(pidfname, "w");
1988         if(pidf) {
1989                 fprintf(pidf,"%d\n", (int)getpid());
1990                 fclose(pidf);
1991         } else {
1992                 perror("fopen");
1993                 fprintf(stderr, "Not fatal; continuing");
1994         }
1995 }
1996 #else
1997 #define daemonize(serve)
1998 #endif /* !defined(NODAEMON) && !defined(NOFORK) */
1999
2000 /*
2001  * Everything beyond this point (in the file) is run in non-daemon mode.
2002  * The stuff above daemonize() isn't.
2003  */
2004
2005 void serve_err(SERVER* serve, const char* msg) G_GNUC_NORETURN;
2006
2007 void serve_err(SERVER* serve, const char* msg) {
2008         g_message("Export of %s on port %d failed:", serve->exportname,
2009                         serve->port);
2010         err(msg);
2011 }
2012
2013 /**
2014  * Set up user-ID and/or group-ID
2015  **/
2016 void dousers(void) {
2017         struct passwd *pw;
2018         struct group *gr;
2019         gchar* str;
2020         if(rungroup) {
2021                 gr=getgrnam(rungroup);
2022                 if(!gr) {
2023                         str = g_strdup_printf("Invalid group name: %s", rungroup);
2024                         err(str);
2025                 }
2026                 if(setgid(gr->gr_gid)<0) {
2027                         err("Could not set GID: %m");
2028                 }
2029         }
2030         if(runuser) {
2031                 pw=getpwnam(runuser);
2032                 if(!pw) {
2033                         str = g_strdup_printf("Invalid user name: %s", runuser);
2034                         err(str);
2035                 }
2036                 if(setuid(pw->pw_uid)<0) {
2037                         err("Could not set UID: %m");
2038                 }
2039         }
2040 }
2041
2042 #ifndef ISSERVER
2043 void glib_message_syslog_redirect(const gchar *log_domain,
2044                                   GLogLevelFlags log_level,
2045                                   const gchar *message,
2046                                   gpointer user_data)
2047 {
2048     int level=LOG_DEBUG;
2049
2050     switch( log_level )
2051     {
2052       case G_LOG_FLAG_FATAL:
2053       case G_LOG_LEVEL_CRITICAL:
2054       case G_LOG_LEVEL_ERROR:
2055         level=LOG_ERR;
2056         break;
2057       case G_LOG_LEVEL_WARNING:
2058         level=LOG_WARNING;
2059         break;
2060       case G_LOG_LEVEL_MESSAGE:
2061       case G_LOG_LEVEL_INFO:
2062         level=LOG_INFO;
2063         break;
2064       case G_LOG_LEVEL_DEBUG:
2065         level=LOG_DEBUG;
2066       default:
2067         level=LOG_ERR;
2068     }
2069     syslog(level, "%s", message);
2070 }
2071 #endif
2072
2073 /**
2074  * Main entry point...
2075  **/
2076 int main(int argc, char *argv[]) {
2077         SERVER *serve;
2078         GArray *servers;
2079         GError *err=NULL;
2080
2081         if (sizeof( struct nbd_request )!=28) {
2082                 fprintf(stderr,"Bad size of structure. Alignment problems?\n");
2083                 exit(EXIT_FAILURE) ;
2084         }
2085
2086         memset(pidftemplate, '\0', 256);
2087
2088         logging();
2089         config_file_pos = g_strdup(CFILE);
2090         serve=cmdline(argc, argv);
2091         servers = parse_cfile(config_file_pos, &err);
2092
2093         if(serve) {
2094                 serve->socket_family = AF_UNSPEC;
2095
2096                 append_serve(serve, servers);
2097
2098                 if (!(serve->port)) {
2099                         CLIENT *client;
2100 #ifndef ISSERVER
2101                         /* You really should define ISSERVER if you're going to use
2102                          * inetd mode, but if you don't, closing stdout and stderr
2103                          * (which inetd had connected to the client socket) will let it
2104                          * work. */
2105                         close(1);
2106                         close(2);
2107                         open("/dev/null", O_WRONLY);
2108                         open("/dev/null", O_WRONLY);
2109                         g_log_set_default_handler( glib_message_syslog_redirect, NULL );
2110 #endif
2111                         client=g_malloc(sizeof(CLIENT));
2112                         client->server=serve;
2113                         client->net=0;
2114                         client->exportsize=OFFT_MAX;
2115                         set_peername(0,client);
2116                         serveconnection(client);
2117                         return 0;
2118                 }
2119         }
2120
2121         if(!servers || !servers->len) {
2122                 if(err && !(err->domain == g_quark_from_string("parse_cfile")
2123                                 && err->code == CFILE_NOTFOUND)) {
2124                         g_warning("Could not parse config file: %s",
2125                                         err ? err->message : "Unknown error");
2126                 }
2127         }
2128         if(serve) {
2129                 g_warning("Specifying an export on the command line is deprecated.");
2130                 g_warning("Please use a configuration file instead.");
2131         }
2132
2133         if((!serve) && (!servers||!servers->len)) {
2134                 g_message("No configured exports; quitting.");
2135                 exit(EXIT_FAILURE);
2136         }
2137         daemonize(serve);
2138         setup_servers(servers);
2139         dousers();
2140         serveloop(servers);
2141         return 0 ;
2142 }