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