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 #include "block.h"
  21
  22 #include <errno.h>
  23 #include <string.h>
  24 #ifndef _WIN32
  25 #include <sys/ioctl.h>
  26 #endif
  27 #if defined(__sun__) || defined(__HAIKU__)
  28 #include <sys/ioccom.h>
  29 #endif
  30 #include <ctype.h>
  31 #include <inttypes.h>
  32
  33 #ifdef __linux__
  34 #include <linux/fs.h>
  35 #endif
  36
  37 #include "qemu_socket.h"
  38
  39 //#define DEBUG_NBD
  40
  41 #ifdef DEBUG_NBD
  42 #define TRACE(msg, ...) do { \
  43     LOG(msg, ## __VA_ARGS__); \
  44 } while(0)
  45 #else
  46 #define TRACE(msg, ...) \
  47     do { } while (0)
  48 #endif
  49
  50 #define LOG(msg, ...) do { \
  51     fprintf(stderr, "%s:%s():L%d: " msg "\n", \
  52             __FILE__, __FUNCTION__, __LINE__, ## __VA_ARGS__); \
  53 } while(0)
  54
  55 /* This is all part of the "official" NBD API */
  56
  57 #define NBD_REPLY_SIZE          (4 + 4 + 8)
  58 #define NBD_REQUEST_MAGIC       0x25609513
  59 #define NBD_REPLY_MAGIC         0x67446698
  60
  61 #define NBD_SET_SOCK            _IO(0xab, 0)
  62 #define NBD_SET_BLKSIZE         _IO(0xab, 1)
  63 #define NBD_SET_SIZE            _IO(0xab, 2)
  64 #define NBD_DO_IT               _IO(0xab, 3)
  65 #define NBD_CLEAR_SOCK          _IO(0xab, 4)
  66 #define NBD_CLEAR_QUE           _IO(0xab, 5)
  67 #define NBD_PRINT_DEBUG         _IO(0xab, 6)
  68 #define NBD_SET_SIZE_BLOCKS     _IO(0xab, 7)
  69 #define NBD_DISCONNECT          _IO(0xab, 8)
  70 #define NBD_SET_TIMEOUT         _IO(0xab, 9)
  71 #define NBD_SET_FLAGS           _IO(0xab, 10)
  72
  73 #define NBD_OPT_EXPORT_NAME     (1 << 0)
  74
  75 /* That's all folks */
  76
  77 #define read_sync(fd, buffer, size) nbd_wr_sync(fd, buffer, size, true)
  78 #define write_sync(fd, buffer, size) nbd_wr_sync(fd, buffer, size, false)
  79
  80 size_t nbd_wr_sync(int fd, void *buffer, size_t size, bool do_read)
  81 {
  82     size_t offset = 0;
  83
  84     while (offset < size) {
  85         ssize_t len;
  86
  87         if (do_read) {
  88             len = qemu_recv(fd, buffer + offset, size - offset, 0);
  89         } else {
  90             len = send(fd, buffer + offset, size - offset, 0);
  91         }
  92
  93         if (len == -1)
  94             errno = socket_error();
  95
  96         /* recoverable error */
  97         if (len == -1 && (errno == EAGAIN || errno == EINTR)) {
  98             continue;
  99         }
 100
 101         /* eof */
 102         if (len == 0) {
 103             break;
 104         }
 105
 106         /* unrecoverable error */
 107         if (len == -1) {
 108             return 0;
 109         }
 110
 111         offset += len;
 112     }
 113
 114     return offset;
 115 }
 116
 117 static void combine_addr(char *buf, size_t len, const char* address,
 118                          uint16_t port)
 119 {
 120     /* If the address-part contains a colon, it's an IPv6 IP so needs [] */
 121     if (strstr(address, ":")) {
 122         snprintf(buf, len, "[%s]:%u", address, port);
 123     } else {
 124         snprintf(buf, len, "%s:%u", address, port);
 125     }
 126 }
 127
 128 int tcp_socket_outgoing(const char *address, uint16_t port)
 129 {
 130     char address_and_port[128];
 131     combine_addr(address_and_port, 128, address, port);
 132     return tcp_socket_outgoing_spec(address_and_port);
 133 }
 134
 135 int tcp_socket_outgoing_spec(const char *address_and_port)
 136 {
 137     return inet_connect(address_and_port, SOCK_STREAM);
 138 }
 139
 140 int tcp_socket_incoming(const char *address, uint16_t port)
 141 {
 142     char address_and_port[128];
 143     combine_addr(address_and_port, 128, address, port);
 144     return tcp_socket_incoming_spec(address_and_port);
 145 }
 146
 147 int tcp_socket_incoming_spec(const char *address_and_port)
 148 {
 149     char *ostr  = NULL;
 150     int olen = 0;
 151     return inet_listen(address_and_port, ostr, olen, SOCK_STREAM, 0);
 152 }
 153
 154 int unix_socket_incoming(const char *path)
 155 {
 156     char *ostr = NULL;
 157     int olen = 0;
 158
 159     return unix_listen(path, ostr, olen);
 160 }
 161
 162 int unix_socket_outgoing(const char *path)
 163 {
 164     return unix_connect(path);
 165 }
 166
 167 /* Basic flow
 168
 169    Server         Client
 170
 171    Negotiate
 172                   Request
 173    Response
 174                   Request
 175    Response
 176                   ...
 177    ...
 178                   Request (type == 2)
 179 */
 180
 181 int nbd_negotiate(int csock, off_t size, uint32_t flags)
 182 {
 183     char buf[8 + 8 + 8 + 128];
 184
 185     /* Negotiate
 186         [ 0 ..   7]   passwd   ("NBDMAGIC")
 187         [ 8 ..  15]   magic    (0x00420281861253)
 188         [16 ..  23]   size
 189         [24 ..  27]   flags
 190         [28 .. 151]   reserved (0)
 191      */
 192
 193     TRACE("Beginning negotiation.");
 194     memcpy(buf, "NBDMAGIC", 8);
 195     cpu_to_be64w((uint64_t*)(buf + 8), 0x00420281861253LL);
 196     cpu_to_be64w((uint64_t*)(buf + 16), size);
 197     cpu_to_be32w((uint32_t*)(buf + 24), flags | NBD_FLAG_HAS_FLAGS);
 198     memset(buf + 28, 0, 124);
 199
 200     if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 201         LOG("write failed");
 202         errno = EINVAL;
 203         return -1;
 204     }
 205
 206     TRACE("Negotiation succeeded.");
 207
 208     return 0;
 209 }
 210
 211 int nbd_receive_negotiate(int csock, const char *name, uint32_t *flags,
 212                           off_t *size, size_t *blocksize)
 213 {
 214     char buf[256];
 215     uint64_t magic, s;
 216     uint16_t tmp;
 217
 218     TRACE("Receiving negotiation.");
 219
 220     if (read_sync(csock, buf, 8) != 8) {
 221         LOG("read failed");
 222         errno = EINVAL;
 223         return -1;
 224     }
 225
 226     buf[8] = '\0';
 227     if (strlen(buf) == 0) {
 228         LOG("server connection closed");
 229         errno = EINVAL;
 230         return -1;
 231     }
 232
 233     TRACE("Magic is %c%c%c%c%c%c%c%c",
 234           qemu_isprint(buf[0]) ? buf[0] : '.',
 235           qemu_isprint(buf[1]) ? buf[1] : '.',
 236           qemu_isprint(buf[2]) ? buf[2] : '.',
 237           qemu_isprint(buf[3]) ? buf[3] : '.',
 238           qemu_isprint(buf[4]) ? buf[4] : '.',
 239           qemu_isprint(buf[5]) ? buf[5] : '.',
 240           qemu_isprint(buf[6]) ? buf[6] : '.',
 241           qemu_isprint(buf[7]) ? buf[7] : '.');
 242
 243     if (memcmp(buf, "NBDMAGIC", 8) != 0) {
 244         LOG("Invalid magic received");
 245         errno = EINVAL;
 246         return -1;
 247     }
 248
 249     if (read_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 250         LOG("read failed");
 251         errno = EINVAL;
 252         return -1;
 253     }
 254     magic = be64_to_cpu(magic);
 255     TRACE("Magic is 0x%" PRIx64, magic);
 256
 257     if (name) {
 258         uint32_t reserved = 0;
 259         uint32_t opt;
 260         uint32_t namesize;
 261
 262         TRACE("Checking magic (opts_magic)");
 263         if (magic != 0x49484156454F5054LL) {
 264             LOG("Bad magic received");
 265             errno = EINVAL;
 266             return -1;
 267         }
 268         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 269             LOG("flags read failed");
 270             errno = EINVAL;
 271             return -1;
 272         }
 273         *flags = be16_to_cpu(tmp) << 16;
 274         /* reserved for future use */
 275         if (write_sync(csock, &reserved, sizeof(reserved)) !=
 276             sizeof(reserved)) {
 277             LOG("write failed (reserved)");
 278             errno = EINVAL;
 279             return -1;
 280         }
 281         /* write the export name */
 282         magic = cpu_to_be64(magic);
 283         if (write_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 284             LOG("write failed (magic)");
 285             errno = EINVAL;
 286             return -1;
 287         }
 288         opt = cpu_to_be32(NBD_OPT_EXPORT_NAME);
 289         if (write_sync(csock, &opt, sizeof(opt)) != sizeof(opt)) {
 290             LOG("write failed (opt)");
 291             errno = EINVAL;
 292             return -1;
 293         }
 294         namesize = cpu_to_be32(strlen(name));
 295         if (write_sync(csock, &namesize, sizeof(namesize)) !=
 296             sizeof(namesize)) {
 297             LOG("write failed (namesize)");
 298             errno = EINVAL;
 299             return -1;
 300         }
 301         if (write_sync(csock, (char*)name, strlen(name)) != strlen(name)) {
 302             LOG("write failed (name)");
 303             errno = EINVAL;
 304             return -1;
 305         }
 306     } else {
 307         TRACE("Checking magic (cli_magic)");
 308
 309         if (magic != 0x00420281861253LL) {
 310             LOG("Bad magic received");
 311             errno = EINVAL;
 312             return -1;
 313         }
 314     }
 315
 316     if (read_sync(csock, &s, sizeof(s)) != sizeof(s)) {
 317         LOG("read failed");
 318         errno = EINVAL;
 319         return -1;
 320     }
 321     *size = be64_to_cpu(s);
 322     *blocksize = 1024;
 323     TRACE("Size is %" PRIu64, *size);
 324
 325     if (!name) {
 326         if (read_sync(csock, flags, sizeof(*flags)) != sizeof(*flags)) {
 327             LOG("read failed (flags)");
 328             errno = EINVAL;
 329             return -1;
 330         }
 331         *flags = be32_to_cpup(flags);
 332     } else {
 333         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 334             LOG("read failed (tmp)");
 335             errno = EINVAL;
 336             return -1;
 337         }
 338         *flags |= be32_to_cpu(tmp);
 339     }
 340     if (read_sync(csock, &buf, 124) != 124) {
 341         LOG("read failed (buf)");
 342         errno = EINVAL;
 343         return -1;
 344     }
 345         return 0;
 346 }
 347
 348 #ifdef __linux__
 349 int nbd_init(int fd, int csock, uint32_t flags, off_t size, size_t blocksize)
 350 {
 351     TRACE("Setting block size to %lu", (unsigned long)blocksize);
 352
 353     if (ioctl(fd, NBD_SET_BLKSIZE, blocksize) == -1) {
 354         int serrno = errno;
 355         LOG("Failed setting NBD block size");
 356         errno = serrno;
 357         return -1;
 358     }
 359
 360         TRACE("Setting size to %zd block(s)", (size_t)(size / blocksize));
 361
 362     if (ioctl(fd, NBD_SET_SIZE_BLOCKS, size / blocksize) == -1) {
 363         int serrno = errno;
 364         LOG("Failed setting size (in blocks)");
 365         errno = serrno;
 366         return -1;
 367     }
 368
 369     if (flags & NBD_FLAG_READ_ONLY) {
 370         int read_only = 1;
 371         TRACE("Setting readonly attribute");
 372
 373         if (ioctl(fd, BLKROSET, (unsigned long) &read_only) < 0) {
 374             int serrno = errno;
 375             LOG("Failed setting read-only attribute");
 376             errno = serrno;
 377             return -1;
 378         }
 379     }
 380
 381     if (ioctl(fd, NBD_SET_FLAGS, flags) < 0
 382         && errno != ENOTTY) {
 383         int serrno = errno;
 384         LOG("Failed setting flags");
 385         errno = serrno;
 386         return -1;
 387     }
 388
 389     TRACE("Clearing NBD socket");
 390
 391     if (ioctl(fd, NBD_CLEAR_SOCK) == -1) {
 392         int serrno = errno;
 393         LOG("Failed clearing NBD socket");
 394         errno = serrno;
 395         return -1;
 396     }
 397
 398     TRACE("Setting NBD socket");
 399
 400     if (ioctl(fd, NBD_SET_SOCK, csock) == -1) {
 401         int serrno = errno;
 402         LOG("Failed to set NBD socket");
 403         errno = serrno;
 404         return -1;
 405     }
 406
 407     TRACE("Negotiation ended");
 408
 409     return 0;
 410 }
 411
 412 int nbd_disconnect(int fd)
 413 {
 414     ioctl(fd, NBD_CLEAR_QUE);
 415     ioctl(fd, NBD_DISCONNECT);
 416     ioctl(fd, NBD_CLEAR_SOCK);
 417     return 0;
 418 }
 419
 420 int nbd_client(int fd)
 421 {
 422     int ret;
 423     int serrno;
 424
 425     TRACE("Doing NBD loop");
 426
 427     ret = ioctl(fd, NBD_DO_IT);
 428     if (ret == -1 && errno == EPIPE) {
 429         /* NBD_DO_IT normally returns EPIPE when someone has disconnected
 430          * the socket via NBD_DISCONNECT.  We do not want to return 1 in
 431          * that case.
 432          */
 433         ret = 0;
 434     }
 435     serrno = errno;
 436
 437     TRACE("NBD loop returned %d: %s", ret, strerror(serrno));
 438
 439     TRACE("Clearing NBD queue");
 440     ioctl(fd, NBD_CLEAR_QUE);
 441
 442     TRACE("Clearing NBD socket");
 443     ioctl(fd, NBD_CLEAR_SOCK);
 444
 445     errno = serrno;
 446     return ret;
 447 }
 448 #else
 449 int nbd_init(int fd, int csock, uint32_t flags, off_t size, size_t blocksize)
 450 {
 451     errno = ENOTSUP;
 452     return -1;
 453 }
 454
 455 int nbd_disconnect(int fd)
 456 {
 457     errno = ENOTSUP;
 458     return -1;
 459 }
 460
 461 int nbd_client(int fd)
 462 {
 463     errno = ENOTSUP;
 464     return -1;
 465 }
 466 #endif
 467
 468 int nbd_send_request(int csock, struct nbd_request *request)
 469 {
 470     uint8_t buf[4 + 4 + 8 + 8 + 4];
 471
 472     cpu_to_be32w((uint32_t*)buf, NBD_REQUEST_MAGIC);
 473     cpu_to_be32w((uint32_t*)(buf + 4), request->type);
 474     cpu_to_be64w((uint64_t*)(buf + 8), request->handle);
 475     cpu_to_be64w((uint64_t*)(buf + 16), request->from);
 476     cpu_to_be32w((uint32_t*)(buf + 24), request->len);
 477
 478     TRACE("Sending request to client: "
 479           "{ .from = %" PRIu64", .len = %u, .handle = %" PRIu64", .type=%i}",
 480           request->from, request->len, request->handle, request->type);
 481
 482     if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 483         LOG("writing to socket failed");
 484         errno = EINVAL;
 485         return -1;
 486     }
 487     return 0;
 488 }
 489
 490 static int nbd_receive_request(int csock, struct nbd_request *request)
 491 {
 492     uint8_t buf[4 + 4 + 8 + 8 + 4];
 493     uint32_t magic;
 494
 495     if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 496         LOG("read failed");
 497         errno = EINVAL;
 498         return -1;
 499     }
 500
 501     /* Request
 502        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 503        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 504        [ 8 .. 15]   handle
 505        [16 .. 23]   from
 506        [24 .. 27]   len
 507      */
 508
 509     magic = be32_to_cpup((uint32_t*)buf);
 510     request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 511     request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 512     request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 513     request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 514
 515     TRACE("Got request: "
 516           "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 517           magic, request->type, request->from, request->len);
 518
 519     if (magic != NBD_REQUEST_MAGIC) {
 520         LOG("invalid magic (got 0x%x)", magic);
 521         errno = EINVAL;
 522         return -1;
 523     }
 524     return 0;
 525 }
 526
 527 int nbd_receive_reply(int csock, struct nbd_reply *reply)
 528 {
 529     uint8_t buf[NBD_REPLY_SIZE];
 530     uint32_t magic;
 531
 532     memset(buf, 0xAA, sizeof(buf));
 533
 534     if (read_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 535         LOG("read failed");
 536         errno = EINVAL;
 537         return -1;
 538     }
 539
 540     /* Reply
 541        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 542        [ 4 ..  7]    error   (0 == no error)
 543        [ 7 .. 15]    handle
 544      */
 545
 546     magic = be32_to_cpup((uint32_t*)buf);
 547     reply->error  = be32_to_cpup((uint32_t*)(buf + 4));
 548     reply->handle = be64_to_cpup((uint64_t*)(buf + 8));
 549
 550     TRACE("Got reply: "
 551           "{ magic = 0x%x, .error = %d, handle = %" PRIu64" }",
 552           magic, reply->error, reply->handle);
 553
 554     if (magic != NBD_REPLY_MAGIC) {
 555         LOG("invalid magic (got 0x%x)", magic);
 556         errno = EINVAL;
 557         return -1;
 558     }
 559     return 0;
 560 }
 561
 562 static int nbd_send_reply(int csock, struct nbd_reply *reply)
 563 {
 564     uint8_t buf[4 + 4 + 8];
 565
 566     /* Reply
 567        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 568        [ 4 ..  7]    error   (0 == no error)
 569        [ 7 .. 15]    handle
 570      */
 571     cpu_to_be32w((uint32_t*)buf, NBD_REPLY_MAGIC);
 572     cpu_to_be32w((uint32_t*)(buf + 4), reply->error);
 573     cpu_to_be64w((uint64_t*)(buf + 8), reply->handle);
 574
 575     TRACE("Sending response to client");
 576
 577     if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 578         LOG("writing to socket failed");
 579         errno = EINVAL;
 580         return -1;
 581     }
 582     return 0;
 583 }
 584
 585 int nbd_trip(BlockDriverState *bs, int csock, off_t size, uint64_t dev_offset,
 586              off_t *offset, uint32_t nbdflags, uint8_t *data, int data_size)
 587 {
 588     struct nbd_request request;
 589     struct nbd_reply reply;
 590
 591     TRACE("Reading request.");
 592
 593     if (nbd_receive_request(csock, &request) == -1)
 594         return -1;
 595
 596     if (request.len + NBD_REPLY_SIZE > data_size) {
 597         LOG("len (%u) is larger than max len (%u)",
 598             request.len + NBD_REPLY_SIZE, data_size);
 599         errno = EINVAL;
 600         return -1;
 601     }
 602
 603     if ((request.from + request.len) < request.from) {
 604         LOG("integer overflow detected! "
 605             "you're probably being attacked");
 606         errno = EINVAL;
 607         return -1;
 608     }
 609
 610     if ((request.from + request.len) > size) {
 611             LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
 612             ", Offset: %" PRIu64 "\n",
 613                     request.from, request.len, (uint64_t)size, dev_offset);
 614         LOG("requested operation past EOF--bad client?");
 615         errno = EINVAL;
 616         return -1;
 617     }
 618
 619     TRACE("Decoding type");
 620
 621     reply.handle = request.handle;
 622     reply.error = 0;
 623
 624     switch (request.type) {
 625     case NBD_CMD_READ:
 626         TRACE("Request type is READ");
 627
 628         if (bdrv_read(bs, (request.from + dev_offset) / 512,
 629                   data + NBD_REPLY_SIZE,
 630                   request.len / 512) == -1) {
 631             LOG("reading from file failed");
 632             errno = EINVAL;
 633             return -1;
 634         }
 635         *offset += request.len;
 636
 637         TRACE("Read %u byte(s)", request.len);
 638
 639         /* Reply
 640            [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 641            [ 4 ..  7]    error   (0 == no error)
 642            [ 7 .. 15]    handle
 643          */
 644
 645         cpu_to_be32w((uint32_t*)data, NBD_REPLY_MAGIC);
 646         cpu_to_be32w((uint32_t*)(data + 4), reply.error);
 647         cpu_to_be64w((uint64_t*)(data + 8), reply.handle);
 648
 649         TRACE("Sending data to client");
 650
 651         if (write_sync(csock, data,
 652                    request.len + NBD_REPLY_SIZE) !=
 653                    request.len + NBD_REPLY_SIZE) {
 654             LOG("writing to socket failed");
 655             errno = EINVAL;
 656             return -1;
 657         }
 658         break;
 659     case NBD_CMD_WRITE:
 660         TRACE("Request type is WRITE");
 661
 662         TRACE("Reading %u byte(s)", request.len);
 663
 664         if (read_sync(csock, data, request.len) != request.len) {
 665             LOG("reading from socket failed");
 666             errno = EINVAL;
 667             return -1;
 668         }
 669
 670         if (nbdflags & NBD_FLAG_READ_ONLY) {
 671             TRACE("Server is read-only, return error");
 672             reply.error = 1;
 673         } else {
 674             TRACE("Writing to device");
 675
 676             if (bdrv_write(bs, (request.from + dev_offset) / 512,
 677                        data, request.len / 512) == -1) {
 678                 LOG("writing to file failed");
 679                 errno = EINVAL;
 680                 return -1;
 681             }
 682
 683             *offset += request.len;
 684         }
 685
 686         if (nbd_send_reply(csock, &reply) == -1)
 687             return -1;
 688         break;
 689     case NBD_CMD_DISC:
 690         TRACE("Request type is DISCONNECT");
 691         errno = 0;
 692         return 1;
 693     default:
 694         LOG("invalid request type (%u) received", request.type);
 695         errno = EINVAL;
 696         return -1;
 697     }
 698
 699     TRACE("Request/Reply complete");
 700
 701     return 0;
 702 }