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, see <http://www.gnu.org/licenses/>.
  17  */
  18
  19 #include "nbd.h"
  20
  21 #include <errno.h>
  22 #include <string.h>
  23 #ifndef _WIN32
  24 #include <sys/ioctl.h>
  25 #endif
  26 #ifdef __sun__
  27 #include <sys/ioccom.h>
  28 #endif
  29 #include <ctype.h>
  30 #include <inttypes.h>
  31
  32 #include "qemu_socket.h"
  33
  34 //#define DEBUG_NBD
  35
  36 #ifdef DEBUG_NBD
  37 #define TRACE(msg, ...) do { \
  38     LOG(msg, ## __VA_ARGS__); \
  39 } while(0)
  40 #else
  41 #define TRACE(msg, ...) \
  42     do { } while (0)
  43 #endif
  44
  45 #define LOG(msg, ...) do { \
  46     fprintf(stderr, "%s:%s():L%d: " msg "\n", \
  47             __FILE__, __FUNCTION__, __LINE__, ## __VA_ARGS__); \
  48 } while(0)
  49
  50 /* This is all part of the "official" NBD API */
  51
  52 #define NBD_REQUEST_MAGIC       0x25609513
  53 #define NBD_REPLY_MAGIC         0x67446698
  54
  55 #define NBD_SET_SOCK            _IO(0xab, 0)
  56 #define NBD_SET_BLKSIZE         _IO(0xab, 1)
  57 #define NBD_SET_SIZE            _IO(0xab, 2)
  58 #define NBD_DO_IT               _IO(0xab, 3)
  59 #define NBD_CLEAR_SOCK          _IO(0xab, 4)
  60 #define NBD_CLEAR_QUE           _IO(0xab, 5)
  61 #define NBD_PRINT_DEBUG         _IO(0xab, 6)
  62 #define NBD_SET_SIZE_BLOCKS     _IO(0xab, 7)
  63 #define NBD_DISCONNECT          _IO(0xab, 8)
  64
  65 /* That's all folks */
  66
  67 #define read_sync(fd, buffer, size) nbd_wr_sync(fd, buffer, size, true)
  68 #define write_sync(fd, buffer, size) nbd_wr_sync(fd, buffer, size, false)
  69
  70 size_t nbd_wr_sync(int fd, void *buffer, size_t size, bool do_read)
  71 {
  72     size_t offset = 0;
  73
  74     while (offset < size) {
  75         ssize_t len;
  76
  77         if (do_read) {
  78             len = recv(fd, buffer + offset, size - offset, 0);
  79         } else {
  80             len = send(fd, buffer + offset, size - offset, 0);
  81         }
  82
  83         if (len == -1)
  84             errno = socket_error();
  85
  86         /* recoverable error */
  87         if (len == -1 && (errno == EAGAIN || errno == EINTR)) {
  88             continue;
  89         }
  90
  91         /* eof */
  92         if (len == 0) {
  93             break;
  94         }
  95
  96         /* unrecoverable error */
  97         if (len == -1) {
  98             return 0;
  99         }
 100
 101         offset += len;
 102     }
 103
 104     return offset;
 105 }
 106
 107 int tcp_socket_outgoing(const char *address, uint16_t port)
 108 {
 109     int s;
 110     struct in_addr in;
 111     struct sockaddr_in addr;
 112
 113     s = socket(PF_INET, SOCK_STREAM, 0);
 114     if (s == -1) {
 115         return -1;
 116     }
 117
 118     if (inet_aton(address, &in) == 0) {
 119         struct hostent *ent;
 120
 121         ent = gethostbyname(address);
 122         if (ent == NULL) {
 123             goto error;
 124         }
 125
 126         memcpy(&in, ent->h_addr, sizeof(in));
 127     }
 128
 129     addr.sin_family = AF_INET;
 130     addr.sin_port = htons(port);
 131     memcpy(&addr.sin_addr.s_addr, &in, sizeof(in));
 132
 133     if (connect(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 134         goto error;
 135     }
 136
 137     return s;
 138 error:
 139     closesocket(s);
 140     return -1;
 141 }
 142
 143 int tcp_socket_incoming(const char *address, uint16_t port)
 144 {
 145     int s;
 146     struct in_addr in;
 147     struct sockaddr_in addr;
 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,
 172                    (const void *) &opt, sizeof(opt)) == -1) {
 173         goto error;
 174     }
 175
 176     if (bind(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 177         goto error;
 178     }
 179
 180     if (listen(s, 128) == -1) {
 181         goto error;
 182     }
 183
 184     return s;
 185 error:
 186     closesocket(s);
 187     return -1;
 188 }
 189
 190 #ifndef _WIN32
 191 int unix_socket_incoming(const char *path)
 192 {
 193     int s;
 194     struct sockaddr_un addr;
 195
 196     s = socket(PF_UNIX, SOCK_STREAM, 0);
 197     if (s == -1) {
 198         return -1;
 199     }
 200
 201     memset(&addr, 0, sizeof(addr));
 202     addr.sun_family = AF_UNIX;
 203     pstrcpy(addr.sun_path, sizeof(addr.sun_path), path);
 204
 205     if (bind(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 206         goto error;
 207     }
 208
 209     if (listen(s, 128) == -1) {
 210         goto error;
 211     }
 212
 213     return s;
 214 error:
 215     closesocket(s);
 216     return -1;
 217 }
 218
 219 int unix_socket_outgoing(const char *path)
 220 {
 221     int s;
 222     struct sockaddr_un addr;
 223
 224     s = socket(PF_UNIX, SOCK_STREAM, 0);
 225     if (s == -1) {
 226         return -1;
 227     }
 228
 229     memset(&addr, 0, sizeof(addr));
 230     addr.sun_family = AF_UNIX;
 231     pstrcpy(addr.sun_path, sizeof(addr.sun_path), path);
 232
 233     if (connect(s, (struct sockaddr *)&addr, sizeof(addr)) == -1) {
 234         goto error;
 235     }
 236
 237     return s;
 238 error:
 239     closesocket(s);
 240     return -1;
 241 }
 242 #else
 243 int unix_socket_incoming(const char *path)
 244 {
 245     errno = ENOTSUP;
 246     return -1;
 247 }
 248
 249 int unix_socket_outgoing(const char *path)
 250 {
 251     errno = ENOTSUP;
 252     return -1;
 253 }
 254 #endif
 255
 256
 257 /* Basic flow
 258
 259    Server         Client
 260
 261    Negotiate
 262                   Request
 263    Response
 264                   Request
 265    Response
 266                   ...
 267    ...
 268                   Request (type == 2)
 269 */
 270
 271 int nbd_negotiate(int csock, off_t size)
 272 {
 273         char buf[8 + 8 + 8 + 128];
 274
 275         /* Negotiate
 276            [ 0 ..   7]   passwd   ("NBDMAGIC")
 277            [ 8 ..  15]   magic    (0x00420281861253)
 278            [16 ..  23]   size
 279            [24 .. 151]   reserved (0)
 280          */
 281
 282         TRACE("Beginning negotiation.");
 283         memcpy(buf, "NBDMAGIC", 8);
 284         cpu_to_be64w((uint64_t*)(buf + 8), 0x00420281861253LL);
 285         cpu_to_be64w((uint64_t*)(buf + 16), size);
 286         memset(buf + 24, 0, 128);
 287
 288         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 289                 LOG("write failed");
 290                 errno = EINVAL;
 291                 return -1;
 292         }
 293
 294         TRACE("Negotation succeeded.");
 295
 296         return 0;
 297 }
 298
 299 int nbd_receive_negotiate(int csock, off_t *size, size_t *blocksize)
 300 {
 301         char buf[8 + 8 + 8 + 128];
 302         uint64_t magic;
 303
 304         TRACE("Receiving negotation.");
 305
 306         if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 307                 LOG("read failed");
 308                 errno = EINVAL;
 309                 return -1;
 310         }
 311
 312         magic = be64_to_cpup((uint64_t*)(buf + 8));
 313         *size = be64_to_cpup((uint64_t*)(buf + 16));
 314         *blocksize = 1024;
 315
 316         TRACE("Magic is %c%c%c%c%c%c%c%c",
 317               qemu_isprint(buf[0]) ? buf[0] : '.',
 318               qemu_isprint(buf[1]) ? buf[1] : '.',
 319               qemu_isprint(buf[2]) ? buf[2] : '.',
 320               qemu_isprint(buf[3]) ? buf[3] : '.',
 321               qemu_isprint(buf[4]) ? buf[4] : '.',
 322               qemu_isprint(buf[5]) ? buf[5] : '.',
 323               qemu_isprint(buf[6]) ? buf[6] : '.',
 324               qemu_isprint(buf[7]) ? buf[7] : '.');
 325         TRACE("Magic is 0x%" PRIx64, magic);
 326         TRACE("Size is %" PRIu64, *size);
 327
 328         if (memcmp(buf, "NBDMAGIC", 8) != 0) {
 329                 LOG("Invalid magic received");
 330                 errno = EINVAL;
 331                 return -1;
 332         }
 333
 334         TRACE("Checking magic");
 335
 336         if (magic != 0x00420281861253LL) {
 337                 LOG("Bad magic received");
 338                 errno = EINVAL;
 339                 return -1;
 340         }
 341         return 0;
 342 }
 343
 344 #ifndef _WIN32
 345 int nbd_init(int fd, int csock, off_t size, size_t blocksize)
 346 {
 347         TRACE("Setting block size to %lu", (unsigned long)blocksize);
 348
 349         if (ioctl(fd, NBD_SET_BLKSIZE, blocksize) == -1) {
 350                 int serrno = errno;
 351                 LOG("Failed setting NBD block size");
 352                 errno = serrno;
 353                 return -1;
 354         }
 355
 356         TRACE("Setting size to %zd block(s)", (size_t)(size / blocksize));
 357
 358         if (ioctl(fd, NBD_SET_SIZE_BLOCKS, size / blocksize) == -1) {
 359                 int serrno = errno;
 360                 LOG("Failed setting size (in blocks)");
 361                 errno = serrno;
 362                 return -1;
 363         }
 364
 365         TRACE("Clearing NBD socket");
 366
 367         if (ioctl(fd, NBD_CLEAR_SOCK) == -1) {
 368                 int serrno = errno;
 369                 LOG("Failed clearing NBD socket");
 370                 errno = serrno;
 371                 return -1;
 372         }
 373
 374         TRACE("Setting NBD socket");
 375
 376         if (ioctl(fd, NBD_SET_SOCK, csock) == -1) {
 377                 int serrno = errno;
 378                 LOG("Failed to set NBD socket");
 379                 errno = serrno;
 380                 return -1;
 381         }
 382
 383         TRACE("Negotiation ended");
 384
 385         return 0;
 386 }
 387
 388 int nbd_disconnect(int fd)
 389 {
 390         ioctl(fd, NBD_CLEAR_QUE);
 391         ioctl(fd, NBD_DISCONNECT);
 392         ioctl(fd, NBD_CLEAR_SOCK);
 393         return 0;
 394 }
 395
 396 int nbd_client(int fd, int csock)
 397 {
 398         int ret;
 399         int serrno;
 400
 401         TRACE("Doing NBD loop");
 402
 403         ret = ioctl(fd, NBD_DO_IT);
 404         serrno = errno;
 405
 406         TRACE("NBD loop returned %d: %s", ret, strerror(serrno));
 407
 408         TRACE("Clearing NBD queue");
 409         ioctl(fd, NBD_CLEAR_QUE);
 410
 411         TRACE("Clearing NBD socket");
 412         ioctl(fd, NBD_CLEAR_SOCK);
 413
 414         errno = serrno;
 415         return ret;
 416 }
 417 #else
 418 int nbd_init(int fd, int csock, off_t size, size_t blocksize)
 419 {
 420     errno = ENOTSUP;
 421     return -1;
 422 }
 423
 424 int nbd_disconnect(int fd)
 425 {
 426     errno = ENOTSUP;
 427     return -1;
 428 }
 429
 430 int nbd_client(int fd, int csock)
 431 {
 432     errno = ENOTSUP;
 433     return -1;
 434 }
 435 #endif
 436
 437 int nbd_send_request(int csock, struct nbd_request *request)
 438 {
 439         uint8_t buf[4 + 4 + 8 + 8 + 4];
 440
 441         cpu_to_be32w((uint32_t*)buf, NBD_REQUEST_MAGIC);
 442         cpu_to_be32w((uint32_t*)(buf + 4), request->type);
 443         cpu_to_be64w((uint64_t*)(buf + 8), request->handle);
 444         cpu_to_be64w((uint64_t*)(buf + 16), request->from);
 445         cpu_to_be32w((uint32_t*)(buf + 24), request->len);
 446
 447         TRACE("Sending request to client");
 448
 449         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 450                 LOG("writing to socket failed");
 451                 errno = EINVAL;
 452                 return -1;
 453         }
 454         return 0;
 455 }
 456
 457
 458 static int nbd_receive_request(int csock, struct nbd_request *request)
 459 {
 460         uint8_t buf[4 + 4 + 8 + 8 + 4];
 461         uint32_t magic;
 462
 463         if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 464                 LOG("read failed");
 465                 errno = EINVAL;
 466                 return -1;
 467         }
 468
 469         /* Request
 470            [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 471            [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 472            [ 8 .. 15]   handle
 473            [16 .. 23]   from
 474            [24 .. 27]   len
 475          */
 476
 477         magic = be32_to_cpup((uint32_t*)buf);
 478         request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 479         request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 480         request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 481         request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 482
 483         TRACE("Got request: "
 484               "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 485               magic, request->type, request->from, request->len);
 486
 487         if (magic != NBD_REQUEST_MAGIC) {
 488                 LOG("invalid magic (got 0x%x)", magic);
 489                 errno = EINVAL;
 490                 return -1;
 491         }
 492         return 0;
 493 }
 494
 495 int nbd_receive_reply(int csock, struct nbd_reply *reply)
 496 {
 497         uint8_t buf[4 + 4 + 8];
 498         uint32_t magic;
 499
 500         memset(buf, 0xAA, sizeof(buf));
 501
 502         if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 503                 LOG("read failed");
 504                 errno = EINVAL;
 505                 return -1;
 506         }
 507
 508         /* Reply
 509            [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 510            [ 4 ..  7]    error   (0 == no error)
 511            [ 7 .. 15]    handle
 512          */
 513
 514         magic = be32_to_cpup((uint32_t*)buf);
 515         reply->error  = be32_to_cpup((uint32_t*)(buf + 4));
 516         reply->handle = be64_to_cpup((uint64_t*)(buf + 8));
 517
 518         TRACE("Got reply: "
 519               "{ magic = 0x%x, .error = %d, handle = %" PRIu64" }",
 520               magic, reply->error, reply->handle);
 521
 522         if (magic != NBD_REPLY_MAGIC) {
 523                 LOG("invalid magic (got 0x%x)", magic);
 524                 errno = EINVAL;
 525                 return -1;
 526         }
 527         return 0;
 528 }
 529
 530 static int nbd_send_reply(int csock, struct nbd_reply *reply)
 531 {
 532         uint8_t buf[4 + 4 + 8];
 533
 534         /* Reply
 535            [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 536            [ 4 ..  7]    error   (0 == no error)
 537            [ 7 .. 15]    handle
 538          */
 539         cpu_to_be32w((uint32_t*)buf, NBD_REPLY_MAGIC);
 540         cpu_to_be32w((uint32_t*)(buf + 4), reply->error);
 541         cpu_to_be64w((uint64_t*)(buf + 8), reply->handle);
 542
 543         TRACE("Sending response to client");
 544
 545         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 546                 LOG("writing to socket failed");
 547                 errno = EINVAL;
 548                 return -1;
 549         }
 550         return 0;
 551 }
 552
 553 int nbd_trip(BlockDriverState *bs, int csock, off_t size, uint64_t dev_offset,
 554              off_t *offset, bool readonly, uint8_t *data, int data_size)
 555 {
 556         struct nbd_request request;
 557         struct nbd_reply reply;
 558
 559         TRACE("Reading request.");
 560
 561         if (nbd_receive_request(csock, &request) == -1)
 562                 return -1;
 563
 564         if (request.len > data_size) {
 565                 LOG("len (%u) is larger than max len (%u)",
 566                     request.len, data_size);
 567                 errno = EINVAL;
 568                 return -1;
 569         }
 570
 571         if ((request.from + request.len) < request.from) {
 572                 LOG("integer overflow detected! "
 573                     "you're probably being attacked");
 574                 errno = EINVAL;
 575                 return -1;
 576         }
 577
 578         if ((request.from + request.len) > size) {
 579                 LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
 580                     ", Offset: %" PRIu64 "\n",
 581                     request.from, request.len, (uint64_t)size, dev_offset);
 582                 LOG("requested operation past EOF--bad client?");
 583                 errno = EINVAL;
 584                 return -1;
 585         }
 586
 587         TRACE("Decoding type");
 588
 589         reply.handle = request.handle;
 590         reply.error = 0;
 591
 592         switch (request.type) {
 593         case NBD_CMD_READ:
 594                 TRACE("Request type is READ");
 595
 596                 if (bdrv_read(bs, (request.from + dev_offset) / 512, data,
 597                               request.len / 512) == -1) {
 598                         LOG("reading from file failed");
 599                         errno = EINVAL;
 600                         return -1;
 601                 }
 602                 *offset += request.len;
 603
 604                 TRACE("Read %u byte(s)", request.len);
 605
 606                 if (nbd_send_reply(csock, &reply) == -1)
 607                         return -1;
 608
 609                 TRACE("Sending data to client");
 610
 611                 if (write_sync(csock, data, request.len) != request.len) {
 612                         LOG("writing to socket failed");
 613                         errno = EINVAL;
 614                         return -1;
 615                 }
 616                 break;
 617         case NBD_CMD_WRITE:
 618                 TRACE("Request type is WRITE");
 619
 620                 TRACE("Reading %u byte(s)", request.len);
 621
 622                 if (read_sync(csock, data, request.len) != request.len) {
 623                         LOG("reading from socket failed");
 624                         errno = EINVAL;
 625                         return -1;
 626                 }
 627
 628                 if (readonly) {
 629                         TRACE("Server is read-only, return error");
 630                         reply.error = 1;
 631                 } else {
 632                         TRACE("Writing to device");
 633
 634                         if (bdrv_write(bs, (request.from + dev_offset) / 512,
 635                                        data, request.len / 512) == -1) {
 636                                 LOG("writing to file failed");
 637                                 errno = EINVAL;
 638                                 return -1;
 639                         }
 640
 641                         *offset += request.len;
 642                 }
 643
 644                 if (nbd_send_reply(csock, &reply) == -1)
 645                         return -1;
 646                 break;
 647         case NBD_CMD_DISC:
 648                 TRACE("Request type is DISCONNECT");
 649                 errno = 0;
 650                 return 1;
 651         default:
 652                 LOG("invalid request type (%u) received", request.type);
 653                 errno = EINVAL;
 654                 return -1;
 655         }
 656
 657         TRACE("Request/Reply complete");
 658
 659         return 0;
 660 }