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 <sys/un.h>
  29 #include <netinet/in.h>
  30 #include <netinet/tcp.h>
  31 #include <arpa/inet.h>
  32 #include <netdb.h>
  33
  34 #if defined(QEMU_NBD)
  35 extern int verbose;
  36 #else
  37 static int verbose = 0;
  38 #endif
  39
  40 #define TRACE(msg, ...) do { \
  41     if (verbose) LOG(msg, ## __VA_ARGS__); \
  42 } while(0)
  43
  44 #define LOG(msg, ...) do { \
  45     fprintf(stderr, "%s:%s():L%d: " msg "\n", \
  46             __FILE__, __FUNCTION__, __LINE__, ## __VA_ARGS__); \
  47 } while(0)
  48
  49 /* This is all part of the "official" NBD API */
  50
  51 #define NBD_REQUEST_MAGIC       0x25609513
  52 #define NBD_REPLY_MAGIC         0x67446698
  53
  54 #define NBD_SET_SOCK            _IO(0xab, 0)
  55 #define NBD_SET_BLKSIZE         _IO(0xab, 1)
  56 #define NBD_SET_SIZE            _IO(0xab, 2)
  57 #define NBD_DO_IT               _IO(0xab, 3)
  58 #define NBD_CLEAR_SOCK          _IO(0xab, 4)
  59 #define NBD_CLEAR_QUE           _IO(0xab, 5)
  60 #define NBD_PRINT_DEBUG         _IO(0xab, 6)
  61 #define NBD_SET_SIZE_BLOCKS     _IO(0xab, 7)
  62 #define NBD_DISCONNECT          _IO(0xab, 8)
  63
  64 /* That's all folks */
  65
  66 #define read_sync(fd, buffer, size) nbd_wr_sync(fd, buffer, size, true)
  67 #define write_sync(fd, buffer, size) nbd_wr_sync(fd, buffer, size, false)
  68
  69 size_t nbd_wr_sync(int fd, void *buffer, size_t size, bool do_read)
  70 {
  71     size_t offset = 0;
  72
  73     while (offset < size) {
  74         ssize_t len;
  75
  76         if (do_read) {
  77             len = read(fd, buffer + offset, size - offset);
  78         } else {
  79             len = write(fd, buffer + offset, size - offset);
  80         }
  81
  82         /* recoverable error */
  83         if (len == -1 && (errno == EAGAIN || errno == EINTR)) {
  84             continue;
  85         }
  86
  87         /* eof */
  88         if (len == 0) {
  89             break;
  90         }
  91
  92         /* unrecoverable error */
  93         if (len == -1) {
  94             return 0;
  95         }
  96
  97         offset += len;
  98     }
  99
 100     return offset;
 101 }
 102
 103 int tcp_socket_outgoing(const char *address, uint16_t port)
 104 {
 105     int s;
 106     struct in_addr in;
 107     struct sockaddr_in addr;
 108     int serrno;
 109
 110     s = socket(PF_INET, SOCK_STREAM, 0);
 111     if (s == -1) {
 112         return -1;
 113     }
 114
 115     if (inet_aton(address, &in) == 0) {
 116         struct hostent *ent;
 117
 118         ent = gethostbyname(address);
 119         if (ent == NULL) {
 120             goto error;
 121         }
 122
 123         memcpy(&in, ent->h_addr, sizeof(in));
 124     }
 125
 126     addr.sin_family = AF_INET;
 127     addr.sin_port = htons(port);
 128     memcpy(&addr.sin_addr.s_addr, &in, sizeof(in));
 129
 130     if (connect(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 131         goto error;
 132     }
 133
 134     return s;
 135 error:
 136     serrno = errno;
 137     close(s);
 138     errno = serrno;
 139     return -1;
 140 }
 141
 142 int tcp_socket_incoming(const char *address, uint16_t port)
 143 {
 144     int s;
 145     struct in_addr in;
 146     struct sockaddr_in addr;
 147     int serrno;
 148     int opt;
 149
 150     s = socket(PF_INET, SOCK_STREAM, 0);
 151     if (s == -1) {
 152         return -1;
 153     }
 154
 155     if (inet_aton(address, &in) == 0) {
 156         struct hostent *ent;
 157
 158         ent = gethostbyname(address);
 159         if (ent == NULL) {
 160             goto error;
 161         }
 162
 163         memcpy(&in, ent->h_addr, sizeof(in));
 164     }
 165
 166     addr.sin_family = AF_INET;
 167     addr.sin_port = htons(port);
 168     memcpy(&addr.sin_addr.s_addr, &in, sizeof(in));
 169
 170     opt = 1;
 171     if (setsockopt(s, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) == -1) {
 172         goto error;
 173     }
 174
 175     if (bind(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 176         goto error;
 177     }
 178
 179     if (listen(s, 128) == -1) {
 180         goto error;
 181     }
 182
 183     return s;
 184 error:
 185     serrno = errno;
 186     close(s);
 187     errno = serrno;
 188     return -1;
 189 }
 190
 191 int unix_socket_incoming(const char *path)
 192 {
 193     int s;
 194     struct sockaddr_un addr;
 195     int serrno;
 196
 197     s = socket(PF_UNIX, SOCK_STREAM, 0);
 198     if (s == -1) {
 199         return -1;
 200     }
 201
 202     memset(&addr, 0, sizeof(addr));
 203     addr.sun_family = AF_UNIX;
 204     pstrcpy(addr.sun_path, sizeof(addr.sun_path), path);
 205
 206     if (bind(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 207         goto error;
 208     }
 209
 210     if (listen(s, 128) == -1) {
 211         goto error;
 212     }
 213
 214     return s;
 215 error:
 216     serrno = errno;
 217     close(s);
 218     errno = serrno;
 219     return -1;
 220 }
 221
 222 int unix_socket_outgoing(const char *path)
 223 {
 224     int s;
 225     struct sockaddr_un addr;
 226     int serrno;
 227
 228     s = socket(PF_UNIX, SOCK_STREAM, 0);
 229     if (s == -1) {
 230         return -1;
 231     }
 232
 233     memset(&addr, 0, sizeof(addr));
 234     addr.sun_family = AF_UNIX;
 235     pstrcpy(addr.sun_path, sizeof(addr.sun_path), path);
 236
 237     if (connect(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 238         goto error;
 239     }
 240
 241     return s;
 242 error:
 243     serrno = errno;
 244     close(s);
 245     errno = serrno;
 246     return -1;
 247 }
 248
 249
 250 /* Basic flow
 251
 252    Server         Client
 253
 254    Negotiate
 255                   Request
 256    Response
 257                   Request
 258    Response
 259                   ...
 260    ...
 261                   Request (type == 2)
 262 */
 263
 264 int nbd_negotiate(BlockDriverState *bs, int csock, off_t size)
 265 {
 266         char buf[8 + 8 + 8 + 128];
 267
 268         /* Negotiate
 269            [ 0 ..   7]   passwd   ("NBDMAGIC")
 270            [ 8 ..  15]   magic    (0x00420281861253)
 271            [16 ..  23]   size
 272            [24 .. 151]   reserved (0)
 273          */
 274
 275         TRACE("Beginning negotiation.");
 276         memcpy(buf, "NBDMAGIC", 8);
 277         cpu_to_be64w((uint64_t*)(buf + 8), 0x00420281861253LL);
 278         cpu_to_be64w((uint64_t*)(buf + 16), size);
 279         memset(buf + 24, 0, 128);
 280
 281         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 282                 LOG("write failed");
 283                 errno = EINVAL;
 284                 return -1;
 285         }
 286
 287         TRACE("Negotation succeeded.");
 288
 289         return 0;
 290 }
 291
 292 int nbd_receive_negotiate(int csock, off_t *size, size_t *blocksize)
 293 {
 294         char buf[8 + 8 + 8 + 128];
 295         uint64_t magic;
 296
 297         TRACE("Receiving negotation.");
 298
 299         if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 300                 LOG("read failed");
 301                 errno = EINVAL;
 302                 return -1;
 303         }
 304
 305         magic = be64_to_cpup((uint64_t*)(buf + 8));
 306         *size = be64_to_cpup((uint64_t*)(buf + 16));
 307         *blocksize = 1024;
 308
 309         TRACE("Magic is %c%c%c%c%c%c%c%c",
 310               isprint(buf[0]) ? buf[0] : '.',
 311               isprint(buf[1]) ? buf[1] : '.',
 312               isprint(buf[2]) ? buf[2] : '.',
 313               isprint(buf[3]) ? buf[3] : '.',
 314               isprint(buf[4]) ? buf[4] : '.',
 315               isprint(buf[5]) ? buf[5] : '.',
 316               isprint(buf[6]) ? buf[6] : '.',
 317               isprint(buf[7]) ? buf[7] : '.');
 318         TRACE("Magic is 0x%" PRIx64, magic);
 319         TRACE("Size is %" PRIu64, *size);
 320
 321         if (memcmp(buf, "NBDMAGIC", 8) != 0) {
 322                 LOG("Invalid magic received");
 323                 errno = EINVAL;
 324                 return -1;
 325         }
 326
 327         TRACE("Checking magic");
 328
 329         if (magic != 0x00420281861253LL) {
 330                 LOG("Bad magic received");
 331                 errno = EINVAL;
 332                 return -1;
 333         }
 334         return 0;
 335 }
 336
 337 int nbd_init(int fd, int csock, off_t size, size_t blocksize)
 338 {
 339         TRACE("Setting block size to %lu", (unsigned long)blocksize);
 340
 341         if (ioctl(fd, NBD_SET_BLKSIZE, blocksize) == -1) {
 342                 int serrno = errno;
 343                 LOG("Failed setting NBD block size");
 344                 errno = serrno;
 345                 return -1;
 346         }
 347
 348         TRACE("Setting size to %llu block(s)",
 349               (unsigned long long)(size / blocksize));
 350
 351         if (ioctl(fd, NBD_SET_SIZE_BLOCKS, size / blocksize) == -1) {
 352                 int serrno = errno;
 353                 LOG("Failed setting size (in blocks)");
 354                 errno = serrno;
 355                 return -1;
 356         }
 357
 358         TRACE("Clearing NBD socket");
 359
 360         if (ioctl(fd, NBD_CLEAR_SOCK) == -1) {
 361                 int serrno = errno;
 362                 LOG("Failed clearing NBD socket");
 363                 errno = serrno;
 364                 return -1;
 365         }
 366
 367         TRACE("Setting NBD socket");
 368
 369         if (ioctl(fd, NBD_SET_SOCK, csock) == -1) {
 370                 int serrno = errno;
 371                 LOG("Failed to set NBD socket");
 372                 errno = serrno;
 373                 return -1;
 374         }
 375
 376         TRACE("Negotiation ended");
 377
 378         return 0;
 379 }
 380
 381 int nbd_disconnect(int fd)
 382 {
 383         ioctl(fd, NBD_CLEAR_QUE);
 384         ioctl(fd, NBD_DISCONNECT);
 385         ioctl(fd, NBD_CLEAR_SOCK);
 386         return 0;
 387 }
 388
 389 int nbd_client(int fd, int csock)
 390 {
 391         int ret;
 392         int serrno;
 393
 394         TRACE("Doing NBD loop");
 395
 396         ret = ioctl(fd, NBD_DO_IT);
 397         serrno = errno;
 398
 399         TRACE("NBD loop returned %d: %s", ret, strerror(serrno));
 400
 401         TRACE("Clearing NBD queue");
 402         ioctl(fd, NBD_CLEAR_QUE);
 403
 404         TRACE("Clearing NBD socket");
 405         ioctl(fd, NBD_CLEAR_SOCK);
 406
 407         errno = serrno;
 408         return ret;
 409 }
 410
 411 int nbd_send_request(int csock, struct nbd_request *request)
 412 {
 413         uint8_t buf[4 + 4 + 8 + 8 + 4];
 414
 415         cpu_to_be32w((uint32_t*)buf, NBD_REQUEST_MAGIC);
 416         cpu_to_be32w((uint32_t*)(buf + 4), request->type);
 417         cpu_to_be64w((uint64_t*)(buf + 8), request->handle);
 418         cpu_to_be64w((uint64_t*)(buf + 16), request->from);
 419         cpu_to_be32w((uint32_t*)(buf + 24), request->len);
 420
 421         TRACE("Sending request to client");
 422
 423         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 424                 LOG("writing to socket failed");
 425                 errno = EINVAL;
 426                 return -1;
 427         }
 428         return 0;
 429 }
 430
 431
 432 static int nbd_receive_request(int csock, struct nbd_request *request)
 433 {
 434         uint8_t buf[4 + 4 + 8 + 8 + 4];
 435         uint32_t magic;
 436
 437         if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 438                 LOG("read failed");
 439                 errno = EINVAL;
 440                 return -1;
 441         }
 442
 443         /* Request
 444            [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 445            [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 446            [ 8 .. 15]   handle
 447            [16 .. 23]   from
 448            [24 .. 27]   len
 449          */
 450
 451         magic = be32_to_cpup((uint32_t*)buf);
 452         request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 453         request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 454         request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 455         request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 456
 457         TRACE("Got request: "
 458               "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 459               magic, request->type, request->from, request->len);
 460
 461         if (magic != NBD_REQUEST_MAGIC) {
 462                 LOG("invalid magic (got 0x%x)", magic);
 463                 errno = EINVAL;
 464                 return -1;
 465         }
 466         return 0;
 467 }
 468
 469 int nbd_receive_reply(int csock, struct nbd_reply *reply)
 470 {
 471         uint8_t buf[4 + 4 + 8];
 472         uint32_t magic;
 473
 474         memset(buf, 0xAA, sizeof(buf));
 475
 476         if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 477                 LOG("read failed");
 478                 errno = EINVAL;
 479                 return -1;
 480         }
 481
 482         /* Reply
 483            [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 484            [ 4 ..  7]    error   (0 == no error)
 485            [ 7 .. 15]    handle
 486          */
 487
 488         magic = be32_to_cpup((uint32_t*)buf);
 489         reply->error  = be32_to_cpup((uint32_t*)(buf + 4));
 490         reply->handle = be64_to_cpup((uint64_t*)(buf + 8));
 491
 492         TRACE("Got reply: "
 493               "{ magic = 0x%x, .error = %d, handle = %" PRIu64" }",
 494               magic, reply->error, reply->handle);
 495
 496         if (magic != NBD_REPLY_MAGIC) {
 497                 LOG("invalid magic (got 0x%x)", magic);
 498                 errno = EINVAL;
 499                 return -1;
 500         }
 501         return 0;
 502 }
 503
 504 static int nbd_send_reply(int csock, struct nbd_reply *reply)
 505 {
 506         uint8_t buf[4 + 4 + 8];
 507
 508         /* Reply
 509            [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 510            [ 4 ..  7]    error   (0 == no error)
 511            [ 7 .. 15]    handle
 512          */
 513         cpu_to_be32w((uint32_t*)buf, NBD_REPLY_MAGIC);
 514         cpu_to_be32w((uint32_t*)(buf + 4), reply->error);
 515         cpu_to_be64w((uint64_t*)(buf + 8), reply->handle);
 516
 517         TRACE("Sending response to client");
 518
 519         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 520                 LOG("writing to socket failed");
 521                 errno = EINVAL;
 522                 return -1;
 523         }
 524         return 0;
 525 }
 526
 527 int nbd_trip(BlockDriverState *bs, int csock, off_t size, uint64_t dev_offset,
 528              off_t *offset, bool readonly, uint8_t *data, int data_size)
 529 {
 530         struct nbd_request request;
 531         struct nbd_reply reply;
 532
 533         TRACE("Reading request.");
 534
 535         if (nbd_receive_request(csock, &request) == -1)
 536                 return -1;
 537
 538         if (request.len > data_size) {
 539                 LOG("len (%u) is larger than max len (%u)",
 540                     request.len, data_size);
 541                 errno = EINVAL;
 542                 return -1;
 543         }
 544
 545         if ((request.from + request.len) < request.from) {
 546                 LOG("integer overflow detected! "
 547                     "you're probably being attacked");
 548                 errno = EINVAL;
 549                 return -1;
 550         }
 551
 552         if ((request.from + request.len) > size) {
 553                 LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
 554                     ", Offset: %" PRIu64 "\n",
 555                      request.from, request.len, size, dev_offset);
 556                 LOG("requested operation past EOF--bad client?");
 557                 errno = EINVAL;
 558                 return -1;
 559         }
 560
 561         TRACE("Decoding type");
 562
 563         reply.handle = request.handle;
 564         reply.error = 0;
 565
 566         switch (request.type) {
 567         case NBD_CMD_READ:
 568                 TRACE("Request type is READ");
 569
 570                 if (bdrv_read(bs, (request.from + dev_offset) / 512, data,
 571                               request.len / 512) == -1) {
 572                         LOG("reading from file failed");
 573                         errno = EINVAL;
 574                         return -1;
 575                 }
 576                 *offset += request.len;
 577
 578                 TRACE("Read %u byte(s)", request.len);
 579
 580                 if (nbd_send_reply(csock, &reply) == -1)
 581                         return -1;
 582
 583                 TRACE("Sending data to client");
 584
 585                 if (write_sync(csock, data, request.len) != request.len) {
 586                         LOG("writing to socket failed");
 587                         errno = EINVAL;
 588                         return -1;
 589                 }
 590                 break;
 591         case NBD_CMD_WRITE:
 592                 TRACE("Request type is WRITE");
 593
 594                 TRACE("Reading %u byte(s)", request.len);
 595
 596                 if (read_sync(csock, data, request.len) != request.len) {
 597                         LOG("reading from socket failed");
 598                         errno = EINVAL;
 599                         return -1;
 600                 }
 601
 602                 if (readonly) {
 603                         TRACE("Server is read-only, return error");
 604                         reply.error = 1;
 605                 } else {
 606                         TRACE("Writing to device");
 607
 608                         if (bdrv_write(bs, (request.from + dev_offset) / 512,
 609                                        data, request.len / 512) == -1) {
 610                                 LOG("writing to file failed");
 611                                 errno = EINVAL;
 612                                 return -1;
 613                         }
 614
 615                         *offset += request.len;
 616                 }
 617
 618                 if (nbd_send_reply(csock, &reply) == -1)
 619                         return -1;
 620                 break;
 621         case NBD_CMD_DISC:
 622                 TRACE("Request type is DISCONNECT");
 623                 errno = 0;
 624                 return 1;
 625         default:
 626                 LOG("invalid request type (%u) received", request.type);
 627                 errno = EINVAL;
 628                 return -1;
 629         }
 630
 631         TRACE("Request/Reply complete");
 632
 633         return 0;
 634 }