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 "block/nbd.h"
  20 #include "block/block.h"
  21
  22 #include "block/coroutine.h"
  23
  24 #include <errno.h>
  25 #include <string.h>
  26 #ifndef _WIN32
  27 #include <sys/ioctl.h>
  28 #endif
  29 #if defined(__sun__) || defined(__HAIKU__)
  30 #include <sys/ioccom.h>
  31 #endif
  32 #include <ctype.h>
  33 #include <inttypes.h>
  34
  35 #ifdef __linux__
  36 #include <linux/fs.h>
  37 #endif
  38
  39 #include "qemu/sockets.h"
  40 #include "qemu/queue.h"
  41 #include "qemu/main-loop.h"
  42
  43 //#define DEBUG_NBD
  44
  45 #ifdef DEBUG_NBD
  46 #define TRACE(msg, ...) do { \
  47     LOG(msg, ## __VA_ARGS__); \
  48 } while(0)
  49 #else
  50 #define TRACE(msg, ...) \
  51     do { } while (0)
  52 #endif
  53
  54 #define LOG(msg, ...) do { \
  55     fprintf(stderr, "%s:%s():L%d: " msg "\n", \
  56             __FILE__, __FUNCTION__, __LINE__, ## __VA_ARGS__); \
  57 } while(0)
  58
  59 /* This is all part of the "official" NBD API */
  60
  61 #define NBD_REQUEST_SIZE        (4 + 4 + 8 + 8 + 4)
  62 #define NBD_REPLY_SIZE          (4 + 4 + 8)
  63 #define NBD_REQUEST_MAGIC       0x25609513
  64 #define NBD_REPLY_MAGIC         0x67446698
  65 #define NBD_OPTS_MAGIC          0x49484156454F5054LL
  66 #define NBD_CLIENT_MAGIC        0x0000420281861253LL
  67
  68 #define NBD_SET_SOCK            _IO(0xab, 0)
  69 #define NBD_SET_BLKSIZE         _IO(0xab, 1)
  70 #define NBD_SET_SIZE            _IO(0xab, 2)
  71 #define NBD_DO_IT               _IO(0xab, 3)
  72 #define NBD_CLEAR_SOCK          _IO(0xab, 4)
  73 #define NBD_CLEAR_QUE           _IO(0xab, 5)
  74 #define NBD_PRINT_DEBUG         _IO(0xab, 6)
  75 #define NBD_SET_SIZE_BLOCKS     _IO(0xab, 7)
  76 #define NBD_DISCONNECT          _IO(0xab, 8)
  77 #define NBD_SET_TIMEOUT         _IO(0xab, 9)
  78 #define NBD_SET_FLAGS           _IO(0xab, 10)
  79
  80 #define NBD_OPT_EXPORT_NAME     (1 << 0)
  81
  82 /* Definitions for opaque data types */
  83
  84 typedef struct NBDRequest NBDRequest;
  85
  86 struct NBDRequest {
  87     QSIMPLEQ_ENTRY(NBDRequest) entry;
  88     NBDClient *client;
  89     uint8_t *data;
  90 };
  91
  92 struct NBDExport {
  93     int refcount;
  94     void (*close)(NBDExport *exp);
  95
  96     BlockDriverState *bs;
  97     char *name;
  98     off_t dev_offset;
  99     off_t size;
 100     uint32_t nbdflags;
 101     QTAILQ_HEAD(, NBDClient) clients;
 102     QTAILQ_ENTRY(NBDExport) next;
 103 };
 104
 105 static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports);
 106
 107 struct NBDClient {
 108     int refcount;
 109     void (*close)(NBDClient *client);
 110
 111     NBDExport *exp;
 112     int sock;
 113
 114     Coroutine *recv_coroutine;
 115
 116     CoMutex send_lock;
 117     Coroutine *send_coroutine;
 118
 119     QTAILQ_ENTRY(NBDClient) next;
 120     int nb_requests;
 121     bool closing;
 122 };
 123
 124 /* That's all folks */
 125
 126 ssize_t nbd_wr_sync(int fd, void *buffer, size_t size, bool do_read)
 127 {
 128     size_t offset = 0;
 129     int err;
 130
 131     if (qemu_in_coroutine()) {
 132         if (do_read) {
 133             return qemu_co_recv(fd, buffer, size);
 134         } else {
 135             return qemu_co_send(fd, buffer, size);
 136         }
 137     }
 138
 139     while (offset < size) {
 140         ssize_t len;
 141
 142         if (do_read) {
 143             len = qemu_recv(fd, buffer + offset, size - offset, 0);
 144         } else {
 145             len = send(fd, buffer + offset, size - offset, 0);
 146         }
 147
 148         if (len < 0) {
 149             err = socket_error();
 150
 151             /* recoverable error */
 152             if (err == EINTR || (offset > 0 && err == EAGAIN)) {
 153                 continue;
 154             }
 155
 156             /* unrecoverable error */
 157             return -err;
 158         }
 159
 160         /* eof */
 161         if (len == 0) {
 162             break;
 163         }
 164
 165         offset += len;
 166     }
 167
 168     return offset;
 169 }
 170
 171 static ssize_t read_sync(int fd, void *buffer, size_t size)
 172 {
 173     /* Sockets are kept in blocking mode in the negotiation phase.  After
 174      * that, a non-readable socket simply means that another thread stole
 175      * our request/reply.  Synchronization is done with recv_coroutine, so
 176      * that this is coroutine-safe.
 177      */
 178     return nbd_wr_sync(fd, buffer, size, true);
 179 }
 180
 181 static ssize_t write_sync(int fd, void *buffer, size_t size)
 182 {
 183     int ret;
 184     do {
 185         /* For writes, we do expect the socket to be writable.  */
 186         ret = nbd_wr_sync(fd, buffer, size, false);
 187     } while (ret == -EAGAIN);
 188     return ret;
 189 }
 190
 191 /* Basic flow for negotiation
 192
 193    Server         Client
 194    Negotiate
 195
 196    or
 197
 198    Server         Client
 199    Negotiate #1
 200                   Option
 201    Negotiate #2
 202
 203    ----
 204
 205    followed by
 206
 207    Server         Client
 208                   Request
 209    Response
 210                   Request
 211    Response
 212                   ...
 213    ...
 214                   Request (type == 2)
 215
 216 */
 217
 218 static int nbd_receive_options(NBDClient *client)
 219 {
 220     int csock = client->sock;
 221     char name[256];
 222     uint32_t tmp, length;
 223     uint64_t magic;
 224     int rc;
 225
 226     /* Client sends:
 227         [ 0 ..   3]   reserved (0)
 228         [ 4 ..  11]   NBD_OPTS_MAGIC
 229         [12 ..  15]   NBD_OPT_EXPORT_NAME
 230         [16 ..  19]   length
 231         [20 ..  xx]   export name (length bytes)
 232      */
 233
 234     rc = -EINVAL;
 235     if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 236         LOG("read failed");
 237         goto fail;
 238     }
 239     TRACE("Checking reserved");
 240     if (tmp != 0) {
 241         LOG("Bad reserved received");
 242         goto fail;
 243     }
 244
 245     if (read_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 246         LOG("read failed");
 247         goto fail;
 248     }
 249     TRACE("Checking reserved");
 250     if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 251         LOG("Bad magic received");
 252         goto fail;
 253     }
 254
 255     if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 256         LOG("read failed");
 257         goto fail;
 258     }
 259     TRACE("Checking option");
 260     if (tmp != be32_to_cpu(NBD_OPT_EXPORT_NAME)) {
 261         LOG("Bad option received");
 262         goto fail;
 263     }
 264
 265     if (read_sync(csock, &length, sizeof(length)) != sizeof(length)) {
 266         LOG("read failed");
 267         goto fail;
 268     }
 269     TRACE("Checking length");
 270     length = be32_to_cpu(length);
 271     if (length > 255) {
 272         LOG("Bad length received");
 273         goto fail;
 274     }
 275     if (read_sync(csock, name, length) != length) {
 276         LOG("read failed");
 277         goto fail;
 278     }
 279     name[length] = '\0';
 280
 281     client->exp = nbd_export_find(name);
 282     if (!client->exp) {
 283         LOG("export not found");
 284         goto fail;
 285     }
 286
 287     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 288     nbd_export_get(client->exp);
 289
 290     TRACE("Option negotiation succeeded.");
 291     rc = 0;
 292 fail:
 293     return rc;
 294 }
 295
 296 static int nbd_send_negotiate(NBDClient *client)
 297 {
 298     int csock = client->sock;
 299     char buf[8 + 8 + 8 + 128];
 300     int rc;
 301     const int myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 302                          NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 303
 304     /* Negotiation header without options:
 305         [ 0 ..   7]   passwd       ("NBDMAGIC")
 306         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 307         [16 ..  23]   size
 308         [24 ..  25]   server flags (0)
 309         [24 ..  27]   export flags
 310         [28 .. 151]   reserved     (0)
 311
 312        Negotiation header with options, part 1:
 313         [ 0 ..   7]   passwd       ("NBDMAGIC")
 314         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 315         [16 ..  17]   server flags (0)
 316
 317        part 2 (after options are sent):
 318         [18 ..  25]   size
 319         [26 ..  27]   export flags
 320         [28 .. 151]   reserved     (0)
 321      */
 322
 323     qemu_set_block(csock);
 324     rc = -EINVAL;
 325
 326     TRACE("Beginning negotiation.");
 327     memset(buf, 0, sizeof(buf));
 328     memcpy(buf, "NBDMAGIC", 8);
 329     if (client->exp) {
 330         assert ((client->exp->nbdflags & ~65535) == 0);
 331         cpu_to_be64w((uint64_t*)(buf + 8), NBD_CLIENT_MAGIC);
 332         cpu_to_be64w((uint64_t*)(buf + 16), client->exp->size);
 333         cpu_to_be16w((uint16_t*)(buf + 26), client->exp->nbdflags | myflags);
 334     } else {
 335         cpu_to_be64w((uint64_t*)(buf + 8), NBD_OPTS_MAGIC);
 336     }
 337
 338     if (client->exp) {
 339         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 340             LOG("write failed");
 341             goto fail;
 342         }
 343     } else {
 344         if (write_sync(csock, buf, 18) != 18) {
 345             LOG("write failed");
 346             goto fail;
 347         }
 348         rc = nbd_receive_options(client);
 349         if (rc < 0) {
 350             LOG("option negotiation failed");
 351             goto fail;
 352         }
 353
 354         assert ((client->exp->nbdflags & ~65535) == 0);
 355         cpu_to_be64w((uint64_t*)(buf + 18), client->exp->size);
 356         cpu_to_be16w((uint16_t*)(buf + 26), client->exp->nbdflags | myflags);
 357         if (write_sync(csock, buf + 18, sizeof(buf) - 18) != sizeof(buf) - 18) {
 358             LOG("write failed");
 359             goto fail;
 360         }
 361     }
 362
 363     TRACE("Negotiation succeeded.");
 364     rc = 0;
 365 fail:
 366     qemu_set_nonblock(csock);
 367     return rc;
 368 }
 369
 370 int nbd_receive_negotiate(int csock, const char *name, uint32_t *flags,
 371                           off_t *size, size_t *blocksize)
 372 {
 373     char buf[256];
 374     uint64_t magic, s;
 375     uint16_t tmp;
 376     int rc;
 377
 378     TRACE("Receiving negotiation.");
 379
 380     rc = -EINVAL;
 381
 382     if (read_sync(csock, buf, 8) != 8) {
 383         LOG("read failed");
 384         goto fail;
 385     }
 386
 387     buf[8] = '\0';
 388     if (strlen(buf) == 0) {
 389         LOG("server connection closed");
 390         goto fail;
 391     }
 392
 393     TRACE("Magic is %c%c%c%c%c%c%c%c",
 394           qemu_isprint(buf[0]) ? buf[0] : '.',
 395           qemu_isprint(buf[1]) ? buf[1] : '.',
 396           qemu_isprint(buf[2]) ? buf[2] : '.',
 397           qemu_isprint(buf[3]) ? buf[3] : '.',
 398           qemu_isprint(buf[4]) ? buf[4] : '.',
 399           qemu_isprint(buf[5]) ? buf[5] : '.',
 400           qemu_isprint(buf[6]) ? buf[6] : '.',
 401           qemu_isprint(buf[7]) ? buf[7] : '.');
 402
 403     if (memcmp(buf, "NBDMAGIC", 8) != 0) {
 404         LOG("Invalid magic received");
 405         goto fail;
 406     }
 407
 408     if (read_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 409         LOG("read failed");
 410         goto fail;
 411     }
 412     magic = be64_to_cpu(magic);
 413     TRACE("Magic is 0x%" PRIx64, magic);
 414
 415     if (name) {
 416         uint32_t reserved = 0;
 417         uint32_t opt;
 418         uint32_t namesize;
 419
 420         TRACE("Checking magic (opts_magic)");
 421         if (magic != NBD_OPTS_MAGIC) {
 422             LOG("Bad magic received");
 423             goto fail;
 424         }
 425         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 426             LOG("flags read failed");
 427             goto fail;
 428         }
 429         *flags = be16_to_cpu(tmp) << 16;
 430         /* reserved for future use */
 431         if (write_sync(csock, &reserved, sizeof(reserved)) !=
 432             sizeof(reserved)) {
 433             LOG("write failed (reserved)");
 434             goto fail;
 435         }
 436         /* write the export name */
 437         magic = cpu_to_be64(magic);
 438         if (write_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 439             LOG("write failed (magic)");
 440             goto fail;
 441         }
 442         opt = cpu_to_be32(NBD_OPT_EXPORT_NAME);
 443         if (write_sync(csock, &opt, sizeof(opt)) != sizeof(opt)) {
 444             LOG("write failed (opt)");
 445             goto fail;
 446         }
 447         namesize = cpu_to_be32(strlen(name));
 448         if (write_sync(csock, &namesize, sizeof(namesize)) !=
 449             sizeof(namesize)) {
 450             LOG("write failed (namesize)");
 451             goto fail;
 452         }
 453         if (write_sync(csock, (char*)name, strlen(name)) != strlen(name)) {
 454             LOG("write failed (name)");
 455             goto fail;
 456         }
 457     } else {
 458         TRACE("Checking magic (cli_magic)");
 459
 460         if (magic != NBD_CLIENT_MAGIC) {
 461             LOG("Bad magic received");
 462             goto fail;
 463         }
 464     }
 465
 466     if (read_sync(csock, &s, sizeof(s)) != sizeof(s)) {
 467         LOG("read failed");
 468         goto fail;
 469     }
 470     *size = be64_to_cpu(s);
 471     *blocksize = 1024;
 472     TRACE("Size is %" PRIu64, *size);
 473
 474     if (!name) {
 475         if (read_sync(csock, flags, sizeof(*flags)) != sizeof(*flags)) {
 476             LOG("read failed (flags)");
 477             goto fail;
 478         }
 479         *flags = be32_to_cpup(flags);
 480     } else {
 481         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 482             LOG("read failed (tmp)");
 483             goto fail;
 484         }
 485         *flags |= be32_to_cpu(tmp);
 486     }
 487     if (read_sync(csock, &buf, 124) != 124) {
 488         LOG("read failed (buf)");
 489         goto fail;
 490     }
 491     rc = 0;
 492
 493 fail:
 494     return rc;
 495 }
 496
 497 #ifdef __linux__
 498 int nbd_init(int fd, int csock, uint32_t flags, off_t size, size_t blocksize)
 499 {
 500     TRACE("Setting NBD socket");
 501
 502     if (ioctl(fd, NBD_SET_SOCK, csock) < 0) {
 503         int serrno = errno;
 504         LOG("Failed to set NBD socket");
 505         return -serrno;
 506     }
 507
 508     TRACE("Setting block size to %lu", (unsigned long)blocksize);
 509
 510     if (ioctl(fd, NBD_SET_BLKSIZE, blocksize) < 0) {
 511         int serrno = errno;
 512         LOG("Failed setting NBD block size");
 513         return -serrno;
 514     }
 515
 516         TRACE("Setting size to %zd block(s)", (size_t)(size / blocksize));
 517
 518     if (ioctl(fd, NBD_SET_SIZE_BLOCKS, size / blocksize) < 0) {
 519         int serrno = errno;
 520         LOG("Failed setting size (in blocks)");
 521         return -serrno;
 522     }
 523
 524     if (ioctl(fd, NBD_SET_FLAGS, flags) < 0) {
 525         if (errno == ENOTTY) {
 526             int read_only = (flags & NBD_FLAG_READ_ONLY) != 0;
 527             TRACE("Setting readonly attribute");
 528
 529             if (ioctl(fd, BLKROSET, (unsigned long) &read_only) < 0) {
 530                 int serrno = errno;
 531                 LOG("Failed setting read-only attribute");
 532                 return -serrno;
 533             }
 534         } else {
 535             int serrno = errno;
 536             LOG("Failed setting flags");
 537             return -serrno;
 538         }
 539     }
 540
 541     TRACE("Negotiation ended");
 542
 543     return 0;
 544 }
 545
 546 int nbd_disconnect(int fd)
 547 {
 548     ioctl(fd, NBD_CLEAR_QUE);
 549     ioctl(fd, NBD_DISCONNECT);
 550     ioctl(fd, NBD_CLEAR_SOCK);
 551     return 0;
 552 }
 553
 554 int nbd_client(int fd)
 555 {
 556     int ret;
 557     int serrno;
 558
 559     TRACE("Doing NBD loop");
 560
 561     ret = ioctl(fd, NBD_DO_IT);
 562     if (ret < 0 && errno == EPIPE) {
 563         /* NBD_DO_IT normally returns EPIPE when someone has disconnected
 564          * the socket via NBD_DISCONNECT.  We do not want to return 1 in
 565          * that case.
 566          */
 567         ret = 0;
 568     }
 569     serrno = errno;
 570
 571     TRACE("NBD loop returned %d: %s", ret, strerror(serrno));
 572
 573     TRACE("Clearing NBD queue");
 574     ioctl(fd, NBD_CLEAR_QUE);
 575
 576     TRACE("Clearing NBD socket");
 577     ioctl(fd, NBD_CLEAR_SOCK);
 578
 579     errno = serrno;
 580     return ret;
 581 }
 582 #else
 583 int nbd_init(int fd, int csock, uint32_t flags, off_t size, size_t blocksize)
 584 {
 585     return -ENOTSUP;
 586 }
 587
 588 int nbd_disconnect(int fd)
 589 {
 590     return -ENOTSUP;
 591 }
 592
 593 int nbd_client(int fd)
 594 {
 595     return -ENOTSUP;
 596 }
 597 #endif
 598
 599 ssize_t nbd_send_request(int csock, struct nbd_request *request)
 600 {
 601     uint8_t buf[NBD_REQUEST_SIZE];
 602     ssize_t ret;
 603
 604     cpu_to_be32w((uint32_t*)buf, NBD_REQUEST_MAGIC);
 605     cpu_to_be32w((uint32_t*)(buf + 4), request->type);
 606     cpu_to_be64w((uint64_t*)(buf + 8), request->handle);
 607     cpu_to_be64w((uint64_t*)(buf + 16), request->from);
 608     cpu_to_be32w((uint32_t*)(buf + 24), request->len);
 609
 610     TRACE("Sending request to client: "
 611           "{ .from = %" PRIu64", .len = %u, .handle = %" PRIu64", .type=%i}",
 612           request->from, request->len, request->handle, request->type);
 613
 614     ret = write_sync(csock, buf, sizeof(buf));
 615     if (ret < 0) {
 616         return ret;
 617     }
 618
 619     if (ret != sizeof(buf)) {
 620         LOG("writing to socket failed");
 621         return -EINVAL;
 622     }
 623     return 0;
 624 }
 625
 626 static ssize_t nbd_receive_request(int csock, struct nbd_request *request)
 627 {
 628     uint8_t buf[NBD_REQUEST_SIZE];
 629     uint32_t magic;
 630     ssize_t ret;
 631
 632     ret = read_sync(csock, buf, sizeof(buf));
 633     if (ret < 0) {
 634         return ret;
 635     }
 636
 637     if (ret != sizeof(buf)) {
 638         LOG("read failed");
 639         return -EINVAL;
 640     }
 641
 642     /* Request
 643        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 644        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 645        [ 8 .. 15]   handle
 646        [16 .. 23]   from
 647        [24 .. 27]   len
 648      */
 649
 650     magic = be32_to_cpup((uint32_t*)buf);
 651     request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 652     request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 653     request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 654     request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 655
 656     TRACE("Got request: "
 657           "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 658           magic, request->type, request->from, request->len);
 659
 660     if (magic != NBD_REQUEST_MAGIC) {
 661         LOG("invalid magic (got 0x%x)", magic);
 662         return -EINVAL;
 663     }
 664     return 0;
 665 }
 666
 667 ssize_t nbd_receive_reply(int csock, struct nbd_reply *reply)
 668 {
 669     uint8_t buf[NBD_REPLY_SIZE];
 670     uint32_t magic;
 671     ssize_t ret;
 672
 673     ret = read_sync(csock, buf, sizeof(buf));
 674     if (ret < 0) {
 675         return ret;
 676     }
 677
 678     if (ret != sizeof(buf)) {
 679         LOG("read failed");
 680         return -EINVAL;
 681     }
 682
 683     /* Reply
 684        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 685        [ 4 ..  7]    error   (0 == no error)
 686        [ 7 .. 15]    handle
 687      */
 688
 689     magic = be32_to_cpup((uint32_t*)buf);
 690     reply->error  = be32_to_cpup((uint32_t*)(buf + 4));
 691     reply->handle = be64_to_cpup((uint64_t*)(buf + 8));
 692
 693     TRACE("Got reply: "
 694           "{ magic = 0x%x, .error = %d, handle = %" PRIu64" }",
 695           magic, reply->error, reply->handle);
 696
 697     if (magic != NBD_REPLY_MAGIC) {
 698         LOG("invalid magic (got 0x%x)", magic);
 699         return -EINVAL;
 700     }
 701     return 0;
 702 }
 703
 704 static ssize_t nbd_send_reply(int csock, struct nbd_reply *reply)
 705 {
 706     uint8_t buf[NBD_REPLY_SIZE];
 707     ssize_t ret;
 708
 709     /* Reply
 710        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 711        [ 4 ..  7]    error   (0 == no error)
 712        [ 7 .. 15]    handle
 713      */
 714     cpu_to_be32w((uint32_t*)buf, NBD_REPLY_MAGIC);
 715     cpu_to_be32w((uint32_t*)(buf + 4), reply->error);
 716     cpu_to_be64w((uint64_t*)(buf + 8), reply->handle);
 717
 718     TRACE("Sending response to client");
 719
 720     ret = write_sync(csock, buf, sizeof(buf));
 721     if (ret < 0) {
 722         return ret;
 723     }
 724
 725     if (ret != sizeof(buf)) {
 726         LOG("writing to socket failed");
 727         return -EINVAL;
 728     }
 729     return 0;
 730 }
 731
 732 #define MAX_NBD_REQUESTS 16
 733
 734 void nbd_client_get(NBDClient *client)
 735 {
 736     client->refcount++;
 737 }
 738
 739 void nbd_client_put(NBDClient *client)
 740 {
 741     if (--client->refcount == 0) {
 742         /* The last reference should be dropped by client->close,
 743          * which is called by nbd_client_close.
 744          */
 745         assert(client->closing);
 746
 747         qemu_set_fd_handler2(client->sock, NULL, NULL, NULL, NULL);
 748         close(client->sock);
 749         client->sock = -1;
 750         if (client->exp) {
 751             QTAILQ_REMOVE(&client->exp->clients, client, next);
 752             nbd_export_put(client->exp);
 753         }
 754         g_free(client);
 755     }
 756 }
 757
 758 void nbd_client_close(NBDClient *client)
 759 {
 760     if (client->closing) {
 761         return;
 762     }
 763
 764     client->closing = true;
 765
 766     /* Force requests to finish.  They will drop their own references,
 767      * then we'll close the socket and free the NBDClient.
 768      */
 769     shutdown(client->sock, 2);
 770
 771     /* Also tell the client, so that they release their reference.  */
 772     if (client->close) {
 773         client->close(client);
 774     }
 775 }
 776
 777 static NBDRequest *nbd_request_get(NBDClient *client)
 778 {
 779     NBDRequest *req;
 780
 781     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 782     client->nb_requests++;
 783
 784     req = g_slice_new0(NBDRequest);
 785     nbd_client_get(client);
 786     req->client = client;
 787     return req;
 788 }
 789
 790 static void nbd_request_put(NBDRequest *req)
 791 {
 792     NBDClient *client = req->client;
 793
 794     if (req->data) {
 795         qemu_vfree(req->data);
 796     }
 797     g_slice_free(NBDRequest, req);
 798
 799     if (client->nb_requests-- == MAX_NBD_REQUESTS) {
 800         qemu_notify_event();
 801     }
 802     nbd_client_put(client);
 803 }
 804
 805 NBDExport *nbd_export_new(BlockDriverState *bs, off_t dev_offset,
 806                           off_t size, uint32_t nbdflags,
 807                           void (*close)(NBDExport *))
 808 {
 809     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 810     exp->refcount = 1;
 811     QTAILQ_INIT(&exp->clients);
 812     exp->bs = bs;
 813     exp->dev_offset = dev_offset;
 814     exp->nbdflags = nbdflags;
 815     exp->size = size == -1 ? bdrv_getlength(bs) : size;
 816     exp->close = close;
 817     bdrv_ref(bs);
 818     return exp;
 819 }
 820
 821 NBDExport *nbd_export_find(const char *name)
 822 {
 823     NBDExport *exp;
 824     QTAILQ_FOREACH(exp, &exports, next) {
 825         if (strcmp(name, exp->name) == 0) {
 826             return exp;
 827         }
 828     }
 829
 830     return NULL;
 831 }
 832
 833 void nbd_export_set_name(NBDExport *exp, const char *name)
 834 {
 835     if (exp->name == name) {
 836         return;
 837     }
 838
 839     nbd_export_get(exp);
 840     if (exp->name != NULL) {
 841         g_free(exp->name);
 842         exp->name = NULL;
 843         QTAILQ_REMOVE(&exports, exp, next);
 844         nbd_export_put(exp);
 845     }
 846     if (name != NULL) {
 847         nbd_export_get(exp);
 848         exp->name = g_strdup(name);
 849         QTAILQ_INSERT_TAIL(&exports, exp, next);
 850     }
 851     nbd_export_put(exp);
 852 }
 853
 854 void nbd_export_close(NBDExport *exp)
 855 {
 856     NBDClient *client, *next;
 857
 858     nbd_export_get(exp);
 859     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 860         nbd_client_close(client);
 861     }
 862     nbd_export_set_name(exp, NULL);
 863     nbd_export_put(exp);
 864     if (exp->bs) {
 865         bdrv_unref(exp->bs);
 866         exp->bs = NULL;
 867     }
 868 }
 869
 870 void nbd_export_get(NBDExport *exp)
 871 {
 872     assert(exp->refcount > 0);
 873     exp->refcount++;
 874 }
 875
 876 void nbd_export_put(NBDExport *exp)
 877 {
 878     assert(exp->refcount > 0);
 879     if (exp->refcount == 1) {
 880         nbd_export_close(exp);
 881     }
 882
 883     if (--exp->refcount == 0) {
 884         assert(exp->name == NULL);
 885
 886         if (exp->close) {
 887             exp->close(exp);
 888         }
 889
 890         g_free(exp);
 891     }
 892 }
 893
 894 BlockDriverState *nbd_export_get_blockdev(NBDExport *exp)
 895 {
 896     return exp->bs;
 897 }
 898
 899 void nbd_export_close_all(void)
 900 {
 901     NBDExport *exp, *next;
 902
 903     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 904         nbd_export_close(exp);
 905     }
 906 }
 907
 908 static int nbd_can_read(void *opaque);
 909 static void nbd_read(void *opaque);
 910 static void nbd_restart_write(void *opaque);
 911
 912 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 913                                  int len)
 914 {
 915     NBDClient *client = req->client;
 916     int csock = client->sock;
 917     ssize_t rc, ret;
 918
 919     qemu_co_mutex_lock(&client->send_lock);
 920     qemu_set_fd_handler2(csock, nbd_can_read, nbd_read,
 921                          nbd_restart_write, client);
 922     client->send_coroutine = qemu_coroutine_self();
 923
 924     if (!len) {
 925         rc = nbd_send_reply(csock, reply);
 926     } else {
 927         socket_set_cork(csock, 1);
 928         rc = nbd_send_reply(csock, reply);
 929         if (rc >= 0) {
 930             ret = qemu_co_send(csock, req->data, len);
 931             if (ret != len) {
 932                 rc = -EIO;
 933             }
 934         }
 935         socket_set_cork(csock, 0);
 936     }
 937
 938     client->send_coroutine = NULL;
 939     qemu_set_fd_handler2(csock, nbd_can_read, nbd_read, NULL, client);
 940     qemu_co_mutex_unlock(&client->send_lock);
 941     return rc;
 942 }
 943
 944 static ssize_t nbd_co_receive_request(NBDRequest *req, struct nbd_request *request)
 945 {
 946     NBDClient *client = req->client;
 947     int csock = client->sock;
 948     uint32_t command;
 949     ssize_t rc;
 950
 951     client->recv_coroutine = qemu_coroutine_self();
 952     rc = nbd_receive_request(csock, request);
 953     if (rc < 0) {
 954         if (rc != -EAGAIN) {
 955             rc = -EIO;
 956         }
 957         goto out;
 958     }
 959
 960     if (request->len > NBD_MAX_BUFFER_SIZE) {
 961         LOG("len (%u) is larger than max len (%u)",
 962             request->len, NBD_MAX_BUFFER_SIZE);
 963         rc = -EINVAL;
 964         goto out;
 965     }
 966
 967     if ((request->from + request->len) < request->from) {
 968         LOG("integer overflow detected! "
 969             "you're probably being attacked");
 970         rc = -EINVAL;
 971         goto out;
 972     }
 973
 974     TRACE("Decoding type");
 975
 976     command = request->type & NBD_CMD_MASK_COMMAND;
 977     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
 978         req->data = qemu_blockalign(client->exp->bs, request->len);
 979     }
 980     if (command == NBD_CMD_WRITE) {
 981         TRACE("Reading %u byte(s)", request->len);
 982
 983         if (qemu_co_recv(csock, req->data, request->len) != request->len) {
 984             LOG("reading from socket failed");
 985             rc = -EIO;
 986             goto out;
 987         }
 988     }
 989     rc = 0;
 990
 991 out:
 992     client->recv_coroutine = NULL;
 993     return rc;
 994 }
 995
 996 static void nbd_trip(void *opaque)
 997 {
 998     NBDClient *client = opaque;
 999     NBDExport *exp = client->exp;
1000     NBDRequest *req;
1001     struct nbd_request request;
1002     struct nbd_reply reply;
1003     ssize_t ret;
1004
1005     TRACE("Reading request.");
1006     if (client->closing) {
1007         return;
1008     }
1009
1010     req = nbd_request_get(client);
1011     ret = nbd_co_receive_request(req, &request);
1012     if (ret == -EAGAIN) {
1013         goto done;
1014     }
1015     if (ret == -EIO) {
1016         goto out;
1017     }
1018
1019     reply.handle = request.handle;
1020     reply.error = 0;
1021
1022     if (ret < 0) {
1023         reply.error = -ret;
1024         goto error_reply;
1025     }
1026
1027     if ((request.from + request.len) > exp->size) {
1028             LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
1029             ", Offset: %" PRIu64 "\n",
1030                     request.from, request.len,
1031                     (uint64_t)exp->size, (uint64_t)exp->dev_offset);
1032         LOG("requested operation past EOF--bad client?");
1033         goto invalid_request;
1034     }
1035
1036     switch (request.type & NBD_CMD_MASK_COMMAND) {
1037     case NBD_CMD_READ:
1038         TRACE("Request type is READ");
1039
1040         if (request.type & NBD_CMD_FLAG_FUA) {
1041             ret = bdrv_co_flush(exp->bs);
1042             if (ret < 0) {
1043                 LOG("flush failed");
1044                 reply.error = -ret;
1045                 goto error_reply;
1046             }
1047         }
1048
1049         ret = bdrv_read(exp->bs, (request.from + exp->dev_offset) / 512,
1050                         req->data, request.len / 512);
1051         if (ret < 0) {
1052             LOG("reading from file failed");
1053             reply.error = -ret;
1054             goto error_reply;
1055         }
1056
1057         TRACE("Read %u byte(s)", request.len);
1058         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1059             goto out;
1060         break;
1061     case NBD_CMD_WRITE:
1062         TRACE("Request type is WRITE");
1063
1064         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1065             TRACE("Server is read-only, return error");
1066             reply.error = EROFS;
1067             goto error_reply;
1068         }
1069
1070         TRACE("Writing to device");
1071
1072         ret = bdrv_write(exp->bs, (request.from + exp->dev_offset) / 512,
1073                          req->data, request.len / 512);
1074         if (ret < 0) {
1075             LOG("writing to file failed");
1076             reply.error = -ret;
1077             goto error_reply;
1078         }
1079
1080         if (request.type & NBD_CMD_FLAG_FUA) {
1081             ret = bdrv_co_flush(exp->bs);
1082             if (ret < 0) {
1083                 LOG("flush failed");
1084                 reply.error = -ret;
1085                 goto error_reply;
1086             }
1087         }
1088
1089         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1090             goto out;
1091         }
1092         break;
1093     case NBD_CMD_DISC:
1094         TRACE("Request type is DISCONNECT");
1095         errno = 0;
1096         goto out;
1097     case NBD_CMD_FLUSH:
1098         TRACE("Request type is FLUSH");
1099
1100         ret = bdrv_co_flush(exp->bs);
1101         if (ret < 0) {
1102             LOG("flush failed");
1103             reply.error = -ret;
1104         }
1105         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1106             goto out;
1107         }
1108         break;
1109     case NBD_CMD_TRIM:
1110         TRACE("Request type is TRIM");
1111         ret = bdrv_co_discard(exp->bs, (request.from + exp->dev_offset) / 512,
1112                               request.len / 512);
1113         if (ret < 0) {
1114             LOG("discard failed");
1115             reply.error = -ret;
1116         }
1117         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1118             goto out;
1119         }
1120         break;
1121     default:
1122         LOG("invalid request type (%u) received", request.type);
1123     invalid_request:
1124         reply.error = -EINVAL;
1125     error_reply:
1126         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1127             goto out;
1128         }
1129         break;
1130     }
1131
1132     TRACE("Request/Reply complete");
1133
1134 done:
1135     nbd_request_put(req);
1136     return;
1137
1138 out:
1139     nbd_request_put(req);
1140     nbd_client_close(client);
1141 }
1142
1143 static int nbd_can_read(void *opaque)
1144 {
1145     NBDClient *client = opaque;
1146
1147     return client->recv_coroutine || client->nb_requests < MAX_NBD_REQUESTS;
1148 }
1149
1150 static void nbd_read(void *opaque)
1151 {
1152     NBDClient *client = opaque;
1153
1154     if (client->recv_coroutine) {
1155         qemu_coroutine_enter(client->recv_coroutine, NULL);
1156     } else {
1157         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip), client);
1158     }
1159 }
1160
1161 static void nbd_restart_write(void *opaque)
1162 {
1163     NBDClient *client = opaque;
1164
1165     qemu_coroutine_enter(client->send_coroutine, NULL);
1166 }
1167
1168 NBDClient *nbd_client_new(NBDExport *exp, int csock,
1169                           void (*close)(NBDClient *))
1170 {
1171     NBDClient *client;
1172     client = g_malloc0(sizeof(NBDClient));
1173     client->refcount = 1;
1174     client->exp = exp;
1175     client->sock = csock;
1176     if (nbd_send_negotiate(client) < 0) {
1177         g_free(client);
1178         return NULL;
1179     }
1180     client->close = close;
1181     qemu_co_mutex_init(&client->send_lock);
1182     qemu_set_fd_handler2(csock, nbd_can_read, nbd_read, NULL, client);
1183
1184     if (exp) {
1185         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1186         nbd_export_get(exp);
1187     }
1188     return client;
1189 }