nbd.c

   1 /*\
   2  *  Copyright (C) 2005  Anthony Liguori <anthony@codemonkey.ws>
   3  *
   4  *  Network Block Device
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; under version 2 of the License.
   9  *
  10  *  This program is distributed in the hope that it will be useful,
  11  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  *  GNU General Public License for more details.
  14  *
  15  *  You should have received a copy of the GNU General Public License
  16  *  along with this program; if not, write to the Free Software
  17  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  18 \*/
  19
  20 #include "nbd.h"
  21
  22 #include <errno.h>
  23 #include <string.h>
  24 #include <sys/ioctl.h>
  25 #include <ctype.h>
  26 #include <inttypes.h>
  27 #include <sys/socket.h>
  28 #include <netinet/in.h>
  29 #include <netinet/tcp.h>
  30 #include <arpa/inet.h>
  31 #include <netdb.h>
  32
  33 extern int verbose;
  34
  35 #define LOG(msg, ...) do { \
  36     fprintf(stderr, "%s:%s():L%d: " msg "\n", \
  37             __FILE__, __FUNCTION__, __LINE__, ## __VA_ARGS__); \
  38 } while(0)
  39
  40 #define TRACE(msg, ...) do { \
  41     if (verbose) LOG(msg, ## __VA_ARGS__); \
  42 } while(0)
  43
  44 /* This is all part of the "official" NBD API */
  45
  46 #define NBD_REQUEST_MAGIC       0x25609513
  47 #define NBD_REPLY_MAGIC         0x67446698
  48
  49 #define NBD_SET_SOCK            _IO(0xab, 0)
  50 #define NBD_SET_BLKSIZE         _IO(0xab, 1)
  51 #define NBD_SET_SIZE            _IO(0xab, 2)
  52 #define NBD_DO_IT               _IO(0xab, 3)
  53 #define NBD_CLEAR_SOCK          _IO(0xab, 4)
  54 #define NBD_CLEAR_QUE           _IO(0xab, 5)
  55 #define NBD_PRINT_DEBUG         _IO(0xab, 6)
  56 #define NBD_SET_SIZE_BLOCKS     _IO(0xab, 7)
  57 #define NBD_DISCONNECT          _IO(0xab, 8)
  58
  59 /* That's all folks */
  60
  61 #define read_sync(fd, buffer, size) wr_sync(fd, buffer, size, true)
  62 #define write_sync(fd, buffer, size) wr_sync(fd, buffer, size, false)
  63
  64 static size_t wr_sync(int fd, void *buffer, size_t size, bool do_read)
  65 {
  66     size_t offset = 0;
  67
  68     while (offset < size) {
  69         ssize_t len;
  70
  71         if (do_read) {
  72             len = read(fd, buffer + offset, size - offset);
  73         } else {
  74             len = write(fd, buffer + offset, size - offset);
  75         }
  76
  77         /* recoverable error */
  78         if (len == -1 && errno == EAGAIN) {
  79             continue;
  80         }
  81
  82         /* eof */
  83         if (len == 0) {
  84             break;
  85         }
  86
  87         /* unrecoverable error */
  88         if (len == -1) {
  89             return 0;
  90         }
  91
  92         offset += len;
  93     }
  94
  95     return offset;
  96 }
  97
  98 static int tcp_socket_outgoing(const char *address, uint16_t port)
  99 {
 100     int s;
 101     struct in_addr in;
 102     struct sockaddr_in addr;
 103     int serrno;
 104
 105     s = socket(PF_INET, SOCK_STREAM, 0);
 106     if (s == -1) {
 107         return -1;
 108     }
 109
 110     if (inet_aton(address, &in) == 0) {
 111         struct hostent *ent;
 112
 113         ent = gethostbyname(address);
 114         if (ent == NULL) {
 115             goto error;
 116         }
 117
 118         memcpy(&in, ent->h_addr, sizeof(in));
 119     }
 120
 121     addr.sin_family = AF_INET;
 122     addr.sin_port = htons(port);
 123     memcpy(&addr.sin_addr.s_addr, &in, sizeof(in));
 124
 125     if (connect(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 126         goto error;
 127     }
 128
 129     return s;
 130 error:
 131     serrno = errno;
 132     close(s);
 133     errno = serrno;
 134     return -1;
 135 }
 136
 137 int tcp_socket_incoming(const char *address, uint16_t port)
 138 {
 139     int s;
 140     struct in_addr in;
 141     struct sockaddr_in addr;
 142     int serrno;
 143     int opt;
 144
 145     s = socket(PF_INET, SOCK_STREAM, 0);
 146     if (s == -1) {
 147         return -1;
 148     }
 149
 150     if (inet_aton(address, &in) == 0) {
 151         struct hostent *ent;
 152
 153         ent = gethostbyname(address);
 154         if (ent == NULL) {
 155             goto error;
 156         }
 157
 158         memcpy(&in, ent->h_addr, sizeof(in));
 159     }
 160
 161     addr.sin_family = AF_INET;
 162     addr.sin_port = htons(port);
 163     memcpy(&addr.sin_addr.s_addr, &in, sizeof(in));
 164
 165     opt = 1;
 166     if (setsockopt(s, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) == -1) {
 167         goto error;
 168     }
 169
 170     if (bind(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 171         goto error;
 172     }
 173
 174     if (listen(s, 128) == -1) {
 175         goto error;
 176     }
 177
 178     return s;
 179 error:
 180     serrno = errno;
 181     close(s);
 182     errno = serrno;
 183     return -1;
 184 }
 185
 186 /* Basic flow
 187
 188    Server         Client
 189
 190    Negotiate
 191                   Request
 192    Response
 193                   Request
 194    Response
 195                   ...
 196    ...
 197                   Request (type == 2)
 198 */
 199
 200 int nbd_negotiate(BlockDriverState *bs, int csock, off_t size)
 201 {
 202         char buf[8 + 8 + 8 + 128];
 203
 204         /* Negotiate
 205            [ 0 ..   7]   passwd   ("NBDMAGIC")
 206            [ 8 ..  15]   magic    (0x00420281861253)
 207            [16 ..  23]   size
 208            [24 .. 151]   reserved (0)
 209          */
 210
 211         TRACE("Beginning negotiation.");
 212         memcpy(buf, "NBDMAGIC", 8);
 213         cpu_to_be64w((uint64_t*)(buf + 8), 0x00420281861253LL);
 214         cpu_to_be64w((uint64_t*)(buf + 16), size);
 215         memset(buf + 24, 0, 128);
 216
 217         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 218                 LOG("write failed");
 219                 errno = EINVAL;
 220                 return -1;
 221         }
 222
 223         TRACE("Negotation succeeded.");
 224
 225         return 0;
 226 }
 227
 228 int nbd_receive_negotiate(int fd, int csock)
 229 {
 230         char buf[8 + 8 + 8 + 128];
 231         uint64_t magic;
 232         off_t size;
 233         size_t blocksize;
 234
 235         TRACE("Receiving negotation.");
 236
 237         if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 238                 LOG("read failed");
 239                 errno = EINVAL;
 240                 return -1;
 241         }
 242
 243         magic = be64_to_cpup((uint64_t*)(buf + 8));
 244         size = be64_to_cpup((uint64_t*)(buf + 16));
 245         blocksize = 1024;
 246
 247         TRACE("Magic is %c%c%c%c%c%c%c%c",
 248               isprint(buf[0]) ? buf[0] : '.',
 249               isprint(buf[1]) ? buf[1] : '.',
 250               isprint(buf[2]) ? buf[2] : '.',
 251               isprint(buf[3]) ? buf[3] : '.',
 252               isprint(buf[4]) ? buf[4] : '.',
 253               isprint(buf[5]) ? buf[5] : '.',
 254               isprint(buf[6]) ? buf[6] : '.',
 255               isprint(buf[7]) ? buf[7] : '.');
 256         TRACE("Magic is 0x%" PRIx64, magic);
 257         TRACE("Size is %" PRIu64, size);
 258
 259         if (memcmp(buf, "NBDMAGIC", 8) != 0) {
 260                 LOG("Invalid magic received");
 261                 errno = EINVAL;
 262                 return -1;
 263         }
 264
 265         TRACE("Checking magic");
 266
 267         if (magic != 0x00420281861253LL) {
 268                 LOG("Bad magic received");
 269                 errno = EINVAL;
 270                 return -1;
 271         }
 272
 273         TRACE("Setting block size to %lu", (unsigned long)blocksize);
 274
 275         if (ioctl(fd, NBD_SET_BLKSIZE, blocksize) == -1) {
 276                 int serrno = errno;
 277                 LOG("Failed setting NBD block size");
 278                 errno = serrno;
 279                 return -1;
 280         }
 281
 282         TRACE("Setting size to %llu block(s)",
 283               (unsigned long long)(size / blocksize));
 284
 285         if (ioctl(fd, NBD_SET_SIZE_BLOCKS, size / blocksize) == -1) {
 286                 int serrno = errno;
 287                 LOG("Failed setting size (in blocks)");
 288                 errno = serrno;
 289                 return -1;
 290         }
 291
 292         TRACE("Clearing NBD socket");
 293
 294         if (ioctl(fd, NBD_CLEAR_SOCK) == -1) {
 295                 int serrno = errno;
 296                 LOG("Failed clearing NBD socket");
 297                 errno = serrno;
 298                 return -1;
 299         }
 300
 301         TRACE("Setting NBD socket");
 302
 303         if (ioctl(fd, NBD_SET_SOCK, csock) == -1) {
 304                 int serrno = errno;
 305                 LOG("Failed to set NBD socket");
 306                 errno = serrno;
 307                 return -1;
 308         }
 309
 310         TRACE("Negotiation ended");
 311
 312         return 0;
 313 }
 314
 315 int nbd_disconnect(int fd)
 316 {
 317         ioctl(fd, NBD_CLEAR_QUE);
 318         ioctl(fd, NBD_DISCONNECT);
 319         ioctl(fd, NBD_CLEAR_SOCK);
 320         return 0;
 321 }
 322
 323 int nbd_client(int fd, int csock)
 324 {
 325         int ret;
 326         int serrno;
 327
 328         TRACE("Doing NBD loop");
 329
 330         ret = ioctl(fd, NBD_DO_IT);
 331         serrno = errno;
 332
 333         TRACE("NBD loop returned %d: %s", ret, strerror(serrno));
 334
 335         TRACE("Clearing NBD queue");
 336         ioctl(fd, NBD_CLEAR_QUE);
 337
 338         TRACE("Clearing NBD socket");
 339         ioctl(fd, NBD_CLEAR_SOCK);
 340
 341         errno = serrno;
 342         return ret;
 343 }
 344
 345 int nbd_trip(BlockDriverState *bs, int csock, off_t size, uint64_t dev_offset, off_t *offset, bool readonly)
 346 {
 347 #ifndef _REENTRANT
 348         static uint8_t data[1024 * 1024]; // keep this off of the stack
 349 #else
 350         uint8_t data[1024 * 1024];
 351 #endif
 352         uint8_t buf[4 + 4 + 8 + 8 + 4];
 353         uint32_t magic;
 354         uint32_t type;
 355         uint64_t from;
 356         uint32_t len;
 357
 358         TRACE("Reading request.");
 359
 360         if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 361                 LOG("read failed");
 362                 errno = EINVAL;
 363                 return -1;
 364         }
 365
 366         /* Request
 367           [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 368           [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 369           [ 8 .. 15]   handle
 370           [16 .. 23]   from
 371           [24 .. 27]   len
 372          */
 373
 374         magic = be32_to_cpup((uint32_t*)buf);
 375         type  = be32_to_cpup((uint32_t*)(buf + 4));
 376         from  = be64_to_cpup((uint64_t*)(buf + 16));
 377         len   = be32_to_cpup((uint32_t*)(buf + 24));
 378
 379         TRACE("Got request: "
 380               "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 381               magic, type, from, len);
 382
 383
 384         if (magic != NBD_REQUEST_MAGIC) {
 385                 LOG("invalid magic (got 0x%x)", magic);
 386                 errno = EINVAL;
 387                 return -1;
 388         }
 389
 390         if (len > sizeof(data)) {
 391                 LOG("len (%u) is larger than max len (%u)",
 392                     len, sizeof(data));
 393                 errno = EINVAL;
 394                 return -1;
 395         }
 396
 397         if ((from + len) < from) {
 398                 LOG("integer overflow detected! "
 399                     "you're probably being attacked");
 400                 errno = EINVAL;
 401                 return -1;
 402         }
 403
 404         if ((from + len) > size) {
 405                 LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
 406                     ", Offset: %" PRIu64 "\n",
 407                      from, len, size, dev_offset);
 408                 LOG("requested operation past EOF--bad client?");
 409                 errno = EINVAL;
 410                 return -1;
 411         }
 412
 413         /* Reply
 414          [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 415          [ 4 ..  7]    error   (0 == no error)
 416          [ 7 .. 15]    handle
 417          */
 418         cpu_to_be32w((uint32_t*)buf, NBD_REPLY_MAGIC);
 419         cpu_to_be32w((uint32_t*)(buf + 4), 0);
 420
 421         TRACE("Decoding type");
 422
 423         switch (type) {
 424         case 0:
 425                 TRACE("Request type is READ");
 426
 427                 if (bdrv_read(bs, (from + dev_offset) / 512, data, len / 512) == -1) {
 428                         LOG("reading from file failed");
 429                         errno = EINVAL;
 430                         return -1;
 431                 }
 432                 *offset += len;
 433
 434                 TRACE("Read %u byte(s)", len);
 435
 436                 TRACE("Sending OK response");
 437
 438                 if (write_sync(csock, buf, 16) != 16) {
 439                         LOG("writing to socket failed");
 440                         errno = EINVAL;
 441                         return -1;
 442                 }
 443
 444                 TRACE("Sending data to client");
 445
 446                 if (write_sync(csock, data, len) != len) {
 447                         LOG("writing to socket failed");
 448                         errno = EINVAL;
 449                         return -1;
 450                 }
 451                 break;
 452         case 1:
 453                 TRACE("Request type is WRITE");
 454
 455                 TRACE("Reading %u byte(s)", len);
 456
 457                 if (read_sync(csock, data, len) != len) {
 458                         LOG("reading from socket failed");
 459                         errno = EINVAL;
 460                         return -1;
 461                 }
 462
 463                 if (readonly) {
 464                         TRACE("Server is read-only, return error");
 465
 466                         cpu_to_be32w((uint32_t*)(buf + 4), 1);
 467                 } else {
 468                         TRACE("Writing to device");
 469
 470                         if (bdrv_write(bs, (from + dev_offset) / 512, data, len / 512) == -1) {
 471                                 LOG("writing to file failed");
 472                                 errno = EINVAL;
 473                                 return -1;
 474                         }
 475
 476                         *offset += len;
 477                 }
 478
 479                 TRACE("Sending response to client");
 480
 481                 if (write_sync(csock, buf, 16) != 16) {
 482                         LOG("writing to socket failed");
 483                         errno = EINVAL;
 484                         return -1;
 485                 }
 486                 break;
 487         case 2:
 488                 TRACE("Request type is DISCONNECT");
 489                 errno = 0;
 490                 return 1;
 491         default:
 492                 LOG("invalid request type (%u) received", type);
 493                 errno = EINVAL;
 494                 return -1;
 495         }
 496
 497         TRACE("Request/Reply complete");
 498
 499         return 0;
 500 }