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
  42 //#define DEBUG_NBD
  43
  44 #ifdef DEBUG_NBD
  45 #define TRACE(msg, ...) do { \
  46     LOG(msg, ## __VA_ARGS__); \
  47 } while(0)
  48 #else
  49 #define TRACE(msg, ...) \
  50     do { } while (0)
  51 #endif
  52
  53 #define LOG(msg, ...) do { \
  54     fprintf(stderr, "%s:%s():L%d: " msg "\n", \
  55             __FILE__, __FUNCTION__, __LINE__, ## __VA_ARGS__); \
  56 } while(0)
  57
  58 /* This is all part of the "official" NBD API */
  59
  60 #define NBD_REQUEST_SIZE        (4 + 4 + 8 + 8 + 4)
  61 #define NBD_REPLY_SIZE          (4 + 4 + 8)
  62 #define NBD_REQUEST_MAGIC       0x25609513
  63 #define NBD_REPLY_MAGIC         0x67446698
  64 #define NBD_OPTS_MAGIC          0x49484156454F5054LL
  65 #define NBD_CLIENT_MAGIC        0x0000420281861253LL
  66
  67 #define NBD_SET_SOCK            _IO(0xab, 0)
  68 #define NBD_SET_BLKSIZE         _IO(0xab, 1)
  69 #define NBD_SET_SIZE            _IO(0xab, 2)
  70 #define NBD_DO_IT               _IO(0xab, 3)
  71 #define NBD_CLEAR_SOCK          _IO(0xab, 4)
  72 #define NBD_CLEAR_QUE           _IO(0xab, 5)
  73 #define NBD_PRINT_DEBUG         _IO(0xab, 6)
  74 #define NBD_SET_SIZE_BLOCKS     _IO(0xab, 7)
  75 #define NBD_DISCONNECT          _IO(0xab, 8)
  76 #define NBD_SET_TIMEOUT         _IO(0xab, 9)
  77 #define NBD_SET_FLAGS           _IO(0xab, 10)
  78
  79 #define NBD_OPT_EXPORT_NAME     (1 << 0)
  80
  81 /* Definitions for opaque data types */
  82
  83 typedef struct NBDRequest NBDRequest;
  84
  85 struct NBDRequest {
  86     QSIMPLEQ_ENTRY(NBDRequest) entry;
  87     NBDClient *client;
  88     uint8_t *data;
  89 };
  90
  91 struct NBDExport {
  92     int refcount;
  93     void (*close)(NBDExport *exp);
  94
  95     BlockDriverState *bs;
  96     char *name;
  97     off_t dev_offset;
  98     off_t size;
  99     uint32_t nbdflags;
 100     QTAILQ_HEAD(, NBDClient) clients;
 101     QSIMPLEQ_HEAD(, NBDRequest) requests;
 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 static void combine_addr(char *buf, size_t len, const char* address,
 192                          uint16_t port)
 193 {
 194     /* If the address-part contains a colon, it's an IPv6 IP so needs [] */
 195     if (strstr(address, ":")) {
 196         snprintf(buf, len, "[%s]:%u", address, port);
 197     } else {
 198         snprintf(buf, len, "%s:%u", address, port);
 199     }
 200 }
 201
 202 int tcp_socket_outgoing_opts(QemuOpts *opts)
 203 {
 204     Error *local_err = NULL;
 205     int fd = inet_connect_opts(opts, &local_err, NULL, NULL);
 206     if (local_err != NULL) {
 207         qerror_report_err(local_err);
 208         error_free(local_err);
 209     }
 210
 211     return fd;
 212 }
 213
 214 int tcp_socket_incoming(const char *address, uint16_t port)
 215 {
 216     char address_and_port[128];
 217     combine_addr(address_and_port, 128, address, port);
 218     return tcp_socket_incoming_spec(address_and_port);
 219 }
 220
 221 int tcp_socket_incoming_spec(const char *address_and_port)
 222 {
 223     Error *local_err = NULL;
 224     int fd = inet_listen(address_and_port, NULL, 0, SOCK_STREAM, 0, &local_err);
 225
 226     if (local_err != NULL) {
 227         qerror_report_err(local_err);
 228         error_free(local_err);
 229     }
 230     return fd;
 231 }
 232
 233 int unix_socket_incoming(const char *path)
 234 {
 235     Error *local_err = NULL;
 236     int fd = unix_listen(path, NULL, 0, &local_err);
 237
 238     if (local_err != NULL) {
 239         qerror_report_err(local_err);
 240         error_free(local_err);
 241     }
 242     return fd;
 243 }
 244
 245 int unix_socket_outgoing(const char *path)
 246 {
 247     Error *local_err = NULL;
 248     int fd = unix_connect(path, &local_err);
 249
 250     if (local_err != NULL) {
 251         qerror_report_err(local_err);
 252         error_free(local_err);
 253     }
 254     return fd;
 255 }
 256
 257 /* Basic flow for negotiation
 258
 259    Server         Client
 260    Negotiate
 261
 262    or
 263
 264    Server         Client
 265    Negotiate #1
 266                   Option
 267    Negotiate #2
 268
 269    ----
 270
 271    followed by
 272
 273    Server         Client
 274                   Request
 275    Response
 276                   Request
 277    Response
 278                   ...
 279    ...
 280                   Request (type == 2)
 281
 282 */
 283
 284 static int nbd_receive_options(NBDClient *client)
 285 {
 286     int csock = client->sock;
 287     char name[256];
 288     uint32_t tmp, length;
 289     uint64_t magic;
 290     int rc;
 291
 292     /* Client sends:
 293         [ 0 ..   3]   reserved (0)
 294         [ 4 ..  11]   NBD_OPTS_MAGIC
 295         [12 ..  15]   NBD_OPT_EXPORT_NAME
 296         [16 ..  19]   length
 297         [20 ..  xx]   export name (length bytes)
 298      */
 299
 300     rc = -EINVAL;
 301     if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 302         LOG("read failed");
 303         goto fail;
 304     }
 305     TRACE("Checking reserved");
 306     if (tmp != 0) {
 307         LOG("Bad reserved received");
 308         goto fail;
 309     }
 310
 311     if (read_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 312         LOG("read failed");
 313         goto fail;
 314     }
 315     TRACE("Checking reserved");
 316     if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 317         LOG("Bad magic received");
 318         goto fail;
 319     }
 320
 321     if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 322         LOG("read failed");
 323         goto fail;
 324     }
 325     TRACE("Checking option");
 326     if (tmp != be32_to_cpu(NBD_OPT_EXPORT_NAME)) {
 327         LOG("Bad option received");
 328         goto fail;
 329     }
 330
 331     if (read_sync(csock, &length, sizeof(length)) != sizeof(length)) {
 332         LOG("read failed");
 333         goto fail;
 334     }
 335     TRACE("Checking length");
 336     length = be32_to_cpu(length);
 337     if (length > 255) {
 338         LOG("Bad length received");
 339         goto fail;
 340     }
 341     if (read_sync(csock, name, length) != length) {
 342         LOG("read failed");
 343         goto fail;
 344     }
 345     name[length] = '\0';
 346
 347     client->exp = nbd_export_find(name);
 348     if (!client->exp) {
 349         LOG("export not found");
 350         goto fail;
 351     }
 352
 353     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 354     nbd_export_get(client->exp);
 355
 356     TRACE("Option negotiation succeeded.");
 357     rc = 0;
 358 fail:
 359     return rc;
 360 }
 361
 362 static int nbd_send_negotiate(NBDClient *client)
 363 {
 364     int csock = client->sock;
 365     char buf[8 + 8 + 8 + 128];
 366     int rc;
 367     const int myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 368                          NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 369
 370     /* Negotiation header without options:
 371         [ 0 ..   7]   passwd       ("NBDMAGIC")
 372         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 373         [16 ..  23]   size
 374         [24 ..  25]   server flags (0)
 375         [24 ..  27]   export flags
 376         [28 .. 151]   reserved     (0)
 377
 378        Negotiation header with options, part 1:
 379         [ 0 ..   7]   passwd       ("NBDMAGIC")
 380         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 381         [16 ..  17]   server flags (0)
 382
 383        part 2 (after options are sent):
 384         [18 ..  25]   size
 385         [26 ..  27]   export flags
 386         [28 .. 151]   reserved     (0)
 387      */
 388
 389     qemu_set_block(csock);
 390     rc = -EINVAL;
 391
 392     TRACE("Beginning negotiation.");
 393     memset(buf, 0, sizeof(buf));
 394     memcpy(buf, "NBDMAGIC", 8);
 395     if (client->exp) {
 396         assert ((client->exp->nbdflags & ~65535) == 0);
 397         cpu_to_be64w((uint64_t*)(buf + 8), NBD_CLIENT_MAGIC);
 398         cpu_to_be64w((uint64_t*)(buf + 16), client->exp->size);
 399         cpu_to_be16w((uint16_t*)(buf + 26), client->exp->nbdflags | myflags);
 400     } else {
 401         cpu_to_be64w((uint64_t*)(buf + 8), NBD_OPTS_MAGIC);
 402     }
 403
 404     if (client->exp) {
 405         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 406             LOG("write failed");
 407             goto fail;
 408         }
 409     } else {
 410         if (write_sync(csock, buf, 18) != 18) {
 411             LOG("write failed");
 412             goto fail;
 413         }
 414         rc = nbd_receive_options(client);
 415         if (rc < 0) {
 416             LOG("option negotiation failed");
 417             goto fail;
 418         }
 419
 420         assert ((client->exp->nbdflags & ~65535) == 0);
 421         cpu_to_be64w((uint64_t*)(buf + 18), client->exp->size);
 422         cpu_to_be16w((uint16_t*)(buf + 26), client->exp->nbdflags | myflags);
 423         if (write_sync(csock, buf + 18, sizeof(buf) - 18) != sizeof(buf) - 18) {
 424             LOG("write failed");
 425             goto fail;
 426         }
 427     }
 428
 429     TRACE("Negotiation succeeded.");
 430     rc = 0;
 431 fail:
 432     qemu_set_nonblock(csock);
 433     return rc;
 434 }
 435
 436 int nbd_receive_negotiate(int csock, const char *name, uint32_t *flags,
 437                           off_t *size, size_t *blocksize)
 438 {
 439     char buf[256];
 440     uint64_t magic, s;
 441     uint16_t tmp;
 442     int rc;
 443
 444     TRACE("Receiving negotiation.");
 445
 446     qemu_set_block(csock);
 447     rc = -EINVAL;
 448
 449     if (read_sync(csock, buf, 8) != 8) {
 450         LOG("read failed");
 451         goto fail;
 452     }
 453
 454     buf[8] = '\0';
 455     if (strlen(buf) == 0) {
 456         LOG("server connection closed");
 457         goto fail;
 458     }
 459
 460     TRACE("Magic is %c%c%c%c%c%c%c%c",
 461           qemu_isprint(buf[0]) ? buf[0] : '.',
 462           qemu_isprint(buf[1]) ? buf[1] : '.',
 463           qemu_isprint(buf[2]) ? buf[2] : '.',
 464           qemu_isprint(buf[3]) ? buf[3] : '.',
 465           qemu_isprint(buf[4]) ? buf[4] : '.',
 466           qemu_isprint(buf[5]) ? buf[5] : '.',
 467           qemu_isprint(buf[6]) ? buf[6] : '.',
 468           qemu_isprint(buf[7]) ? buf[7] : '.');
 469
 470     if (memcmp(buf, "NBDMAGIC", 8) != 0) {
 471         LOG("Invalid magic received");
 472         goto fail;
 473     }
 474
 475     if (read_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 476         LOG("read failed");
 477         goto fail;
 478     }
 479     magic = be64_to_cpu(magic);
 480     TRACE("Magic is 0x%" PRIx64, magic);
 481
 482     if (name) {
 483         uint32_t reserved = 0;
 484         uint32_t opt;
 485         uint32_t namesize;
 486
 487         TRACE("Checking magic (opts_magic)");
 488         if (magic != NBD_OPTS_MAGIC) {
 489             LOG("Bad magic received");
 490             goto fail;
 491         }
 492         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 493             LOG("flags read failed");
 494             goto fail;
 495         }
 496         *flags = be16_to_cpu(tmp) << 16;
 497         /* reserved for future use */
 498         if (write_sync(csock, &reserved, sizeof(reserved)) !=
 499             sizeof(reserved)) {
 500             LOG("write failed (reserved)");
 501             goto fail;
 502         }
 503         /* write the export name */
 504         magic = cpu_to_be64(magic);
 505         if (write_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 506             LOG("write failed (magic)");
 507             goto fail;
 508         }
 509         opt = cpu_to_be32(NBD_OPT_EXPORT_NAME);
 510         if (write_sync(csock, &opt, sizeof(opt)) != sizeof(opt)) {
 511             LOG("write failed (opt)");
 512             goto fail;
 513         }
 514         namesize = cpu_to_be32(strlen(name));
 515         if (write_sync(csock, &namesize, sizeof(namesize)) !=
 516             sizeof(namesize)) {
 517             LOG("write failed (namesize)");
 518             goto fail;
 519         }
 520         if (write_sync(csock, (char*)name, strlen(name)) != strlen(name)) {
 521             LOG("write failed (name)");
 522             goto fail;
 523         }
 524     } else {
 525         TRACE("Checking magic (cli_magic)");
 526
 527         if (magic != NBD_CLIENT_MAGIC) {
 528             LOG("Bad magic received");
 529             goto fail;
 530         }
 531     }
 532
 533     if (read_sync(csock, &s, sizeof(s)) != sizeof(s)) {
 534         LOG("read failed");
 535         goto fail;
 536     }
 537     *size = be64_to_cpu(s);
 538     *blocksize = 1024;
 539     TRACE("Size is %" PRIu64, *size);
 540
 541     if (!name) {
 542         if (read_sync(csock, flags, sizeof(*flags)) != sizeof(*flags)) {
 543             LOG("read failed (flags)");
 544             goto fail;
 545         }
 546         *flags = be32_to_cpup(flags);
 547     } else {
 548         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 549             LOG("read failed (tmp)");
 550             goto fail;
 551         }
 552         *flags |= be32_to_cpu(tmp);
 553     }
 554     if (read_sync(csock, &buf, 124) != 124) {
 555         LOG("read failed (buf)");
 556         goto fail;
 557     }
 558     rc = 0;
 559
 560 fail:
 561     qemu_set_nonblock(csock);
 562     return rc;
 563 }
 564
 565 #ifdef __linux__
 566 int nbd_init(int fd, int csock, uint32_t flags, off_t size, size_t blocksize)
 567 {
 568     TRACE("Setting NBD socket");
 569
 570     if (ioctl(fd, NBD_SET_SOCK, csock) < 0) {
 571         int serrno = errno;
 572         LOG("Failed to set NBD socket");
 573         return -serrno;
 574     }
 575
 576     TRACE("Setting block size to %lu", (unsigned long)blocksize);
 577
 578     if (ioctl(fd, NBD_SET_BLKSIZE, blocksize) < 0) {
 579         int serrno = errno;
 580         LOG("Failed setting NBD block size");
 581         return -serrno;
 582     }
 583
 584         TRACE("Setting size to %zd block(s)", (size_t)(size / blocksize));
 585
 586     if (ioctl(fd, NBD_SET_SIZE_BLOCKS, size / blocksize) < 0) {
 587         int serrno = errno;
 588         LOG("Failed setting size (in blocks)");
 589         return -serrno;
 590     }
 591
 592     if (ioctl(fd, NBD_SET_FLAGS, flags) < 0) {
 593         if (errno == ENOTTY) {
 594             int read_only = (flags & NBD_FLAG_READ_ONLY) != 0;
 595             TRACE("Setting readonly attribute");
 596
 597             if (ioctl(fd, BLKROSET, (unsigned long) &read_only) < 0) {
 598                 int serrno = errno;
 599                 LOG("Failed setting read-only attribute");
 600                 return -serrno;
 601             }
 602         } else {
 603             int serrno = errno;
 604             LOG("Failed setting flags");
 605             return -serrno;
 606         }
 607     }
 608
 609     TRACE("Negotiation ended");
 610
 611     return 0;
 612 }
 613
 614 int nbd_disconnect(int fd)
 615 {
 616     ioctl(fd, NBD_CLEAR_QUE);
 617     ioctl(fd, NBD_DISCONNECT);
 618     ioctl(fd, NBD_CLEAR_SOCK);
 619     return 0;
 620 }
 621
 622 int nbd_client(int fd)
 623 {
 624     int ret;
 625     int serrno;
 626
 627     TRACE("Doing NBD loop");
 628
 629     ret = ioctl(fd, NBD_DO_IT);
 630     if (ret < 0 && errno == EPIPE) {
 631         /* NBD_DO_IT normally returns EPIPE when someone has disconnected
 632          * the socket via NBD_DISCONNECT.  We do not want to return 1 in
 633          * that case.
 634          */
 635         ret = 0;
 636     }
 637     serrno = errno;
 638
 639     TRACE("NBD loop returned %d: %s", ret, strerror(serrno));
 640
 641     TRACE("Clearing NBD queue");
 642     ioctl(fd, NBD_CLEAR_QUE);
 643
 644     TRACE("Clearing NBD socket");
 645     ioctl(fd, NBD_CLEAR_SOCK);
 646
 647     errno = serrno;
 648     return ret;
 649 }
 650 #else
 651 int nbd_init(int fd, int csock, uint32_t flags, off_t size, size_t blocksize)
 652 {
 653     return -ENOTSUP;
 654 }
 655
 656 int nbd_disconnect(int fd)
 657 {
 658     return -ENOTSUP;
 659 }
 660
 661 int nbd_client(int fd)
 662 {
 663     return -ENOTSUP;
 664 }
 665 #endif
 666
 667 ssize_t nbd_send_request(int csock, struct nbd_request *request)
 668 {
 669     uint8_t buf[NBD_REQUEST_SIZE];
 670     ssize_t ret;
 671
 672     cpu_to_be32w((uint32_t*)buf, NBD_REQUEST_MAGIC);
 673     cpu_to_be32w((uint32_t*)(buf + 4), request->type);
 674     cpu_to_be64w((uint64_t*)(buf + 8), request->handle);
 675     cpu_to_be64w((uint64_t*)(buf + 16), request->from);
 676     cpu_to_be32w((uint32_t*)(buf + 24), request->len);
 677
 678     TRACE("Sending request to client: "
 679           "{ .from = %" PRIu64", .len = %u, .handle = %" PRIu64", .type=%i}",
 680           request->from, request->len, request->handle, request->type);
 681
 682     ret = write_sync(csock, buf, sizeof(buf));
 683     if (ret < 0) {
 684         return ret;
 685     }
 686
 687     if (ret != sizeof(buf)) {
 688         LOG("writing to socket failed");
 689         return -EINVAL;
 690     }
 691     return 0;
 692 }
 693
 694 static ssize_t nbd_receive_request(int csock, struct nbd_request *request)
 695 {
 696     uint8_t buf[NBD_REQUEST_SIZE];
 697     uint32_t magic;
 698     ssize_t ret;
 699
 700     ret = read_sync(csock, buf, sizeof(buf));
 701     if (ret < 0) {
 702         return ret;
 703     }
 704
 705     if (ret != sizeof(buf)) {
 706         LOG("read failed");
 707         return -EINVAL;
 708     }
 709
 710     /* Request
 711        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 712        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 713        [ 8 .. 15]   handle
 714        [16 .. 23]   from
 715        [24 .. 27]   len
 716      */
 717
 718     magic = be32_to_cpup((uint32_t*)buf);
 719     request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 720     request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 721     request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 722     request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 723
 724     TRACE("Got request: "
 725           "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 726           magic, request->type, request->from, request->len);
 727
 728     if (magic != NBD_REQUEST_MAGIC) {
 729         LOG("invalid magic (got 0x%x)", magic);
 730         return -EINVAL;
 731     }
 732     return 0;
 733 }
 734
 735 ssize_t nbd_receive_reply(int csock, struct nbd_reply *reply)
 736 {
 737     uint8_t buf[NBD_REPLY_SIZE];
 738     uint32_t magic;
 739     ssize_t ret;
 740
 741     ret = read_sync(csock, buf, sizeof(buf));
 742     if (ret < 0) {
 743         return ret;
 744     }
 745
 746     if (ret != sizeof(buf)) {
 747         LOG("read failed");
 748         return -EINVAL;
 749     }
 750
 751     /* Reply
 752        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 753        [ 4 ..  7]    error   (0 == no error)
 754        [ 7 .. 15]    handle
 755      */
 756
 757     magic = be32_to_cpup((uint32_t*)buf);
 758     reply->error  = be32_to_cpup((uint32_t*)(buf + 4));
 759     reply->handle = be64_to_cpup((uint64_t*)(buf + 8));
 760
 761     TRACE("Got reply: "
 762           "{ magic = 0x%x, .error = %d, handle = %" PRIu64" }",
 763           magic, reply->error, reply->handle);
 764
 765     if (magic != NBD_REPLY_MAGIC) {
 766         LOG("invalid magic (got 0x%x)", magic);
 767         return -EINVAL;
 768     }
 769     return 0;
 770 }
 771
 772 static ssize_t nbd_send_reply(int csock, struct nbd_reply *reply)
 773 {
 774     uint8_t buf[NBD_REPLY_SIZE];
 775     ssize_t ret;
 776
 777     /* Reply
 778        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 779        [ 4 ..  7]    error   (0 == no error)
 780        [ 7 .. 15]    handle
 781      */
 782     cpu_to_be32w((uint32_t*)buf, NBD_REPLY_MAGIC);
 783     cpu_to_be32w((uint32_t*)(buf + 4), reply->error);
 784     cpu_to_be64w((uint64_t*)(buf + 8), reply->handle);
 785
 786     TRACE("Sending response to client");
 787
 788     ret = write_sync(csock, buf, sizeof(buf));
 789     if (ret < 0) {
 790         return ret;
 791     }
 792
 793     if (ret != sizeof(buf)) {
 794         LOG("writing to socket failed");
 795         return -EINVAL;
 796     }
 797     return 0;
 798 }
 799
 800 #define MAX_NBD_REQUESTS 16
 801
 802 void nbd_client_get(NBDClient *client)
 803 {
 804     client->refcount++;
 805 }
 806
 807 void nbd_client_put(NBDClient *client)
 808 {
 809     if (--client->refcount == 0) {
 810         /* The last reference should be dropped by client->close,
 811          * which is called by nbd_client_close.
 812          */
 813         assert(client->closing);
 814
 815         qemu_set_fd_handler2(client->sock, NULL, NULL, NULL, NULL);
 816         close(client->sock);
 817         client->sock = -1;
 818         if (client->exp) {
 819             QTAILQ_REMOVE(&client->exp->clients, client, next);
 820             nbd_export_put(client->exp);
 821         }
 822         g_free(client);
 823     }
 824 }
 825
 826 void nbd_client_close(NBDClient *client)
 827 {
 828     if (client->closing) {
 829         return;
 830     }
 831
 832     client->closing = true;
 833
 834     /* Force requests to finish.  They will drop their own references,
 835      * then we'll close the socket and free the NBDClient.
 836      */
 837     shutdown(client->sock, 2);
 838
 839     /* Also tell the client, so that they release their reference.  */
 840     if (client->close) {
 841         client->close(client);
 842     }
 843 }
 844
 845 static NBDRequest *nbd_request_get(NBDClient *client)
 846 {
 847     NBDRequest *req;
 848     NBDExport *exp = client->exp;
 849
 850     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 851     client->nb_requests++;
 852
 853     if (QSIMPLEQ_EMPTY(&exp->requests)) {
 854         req = g_malloc0(sizeof(NBDRequest));
 855         req->data = qemu_blockalign(exp->bs, NBD_BUFFER_SIZE);
 856     } else {
 857         req = QSIMPLEQ_FIRST(&exp->requests);
 858         QSIMPLEQ_REMOVE_HEAD(&exp->requests, entry);
 859     }
 860     nbd_client_get(client);
 861     req->client = client;
 862     return req;
 863 }
 864
 865 static void nbd_request_put(NBDRequest *req)
 866 {
 867     NBDClient *client = req->client;
 868     QSIMPLEQ_INSERT_HEAD(&client->exp->requests, req, entry);
 869     if (client->nb_requests-- == MAX_NBD_REQUESTS) {
 870         qemu_notify_event();
 871     }
 872     nbd_client_put(client);
 873 }
 874
 875 NBDExport *nbd_export_new(BlockDriverState *bs, off_t dev_offset,
 876                           off_t size, uint32_t nbdflags,
 877                           void (*close)(NBDExport *))
 878 {
 879     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 880     QSIMPLEQ_INIT(&exp->requests);
 881     exp->refcount = 1;
 882     QTAILQ_INIT(&exp->clients);
 883     exp->bs = bs;
 884     exp->dev_offset = dev_offset;
 885     exp->nbdflags = nbdflags;
 886     exp->size = size == -1 ? bdrv_getlength(bs) : size;
 887     exp->close = close;
 888     return exp;
 889 }
 890
 891 NBDExport *nbd_export_find(const char *name)
 892 {
 893     NBDExport *exp;
 894     QTAILQ_FOREACH(exp, &exports, next) {
 895         if (strcmp(name, exp->name) == 0) {
 896             return exp;
 897         }
 898     }
 899
 900     return NULL;
 901 }
 902
 903 void nbd_export_set_name(NBDExport *exp, const char *name)
 904 {
 905     if (exp->name == name) {
 906         return;
 907     }
 908
 909     nbd_export_get(exp);
 910     if (exp->name != NULL) {
 911         g_free(exp->name);
 912         exp->name = NULL;
 913         QTAILQ_REMOVE(&exports, exp, next);
 914         nbd_export_put(exp);
 915     }
 916     if (name != NULL) {
 917         nbd_export_get(exp);
 918         exp->name = g_strdup(name);
 919         QTAILQ_INSERT_TAIL(&exports, exp, next);
 920     }
 921     nbd_export_put(exp);
 922 }
 923
 924 void nbd_export_close(NBDExport *exp)
 925 {
 926     NBDClient *client, *next;
 927
 928     nbd_export_get(exp);
 929     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 930         nbd_client_close(client);
 931     }
 932     nbd_export_set_name(exp, NULL);
 933     nbd_export_put(exp);
 934 }
 935
 936 void nbd_export_get(NBDExport *exp)
 937 {
 938     assert(exp->refcount > 0);
 939     exp->refcount++;
 940 }
 941
 942 void nbd_export_put(NBDExport *exp)
 943 {
 944     assert(exp->refcount > 0);
 945     if (exp->refcount == 1) {
 946         nbd_export_close(exp);
 947     }
 948
 949     if (--exp->refcount == 0) {
 950         assert(exp->name == NULL);
 951
 952         if (exp->close) {
 953             exp->close(exp);
 954         }
 955
 956         while (!QSIMPLEQ_EMPTY(&exp->requests)) {
 957             NBDRequest *first = QSIMPLEQ_FIRST(&exp->requests);
 958             QSIMPLEQ_REMOVE_HEAD(&exp->requests, entry);
 959             qemu_vfree(first->data);
 960             g_free(first);
 961         }
 962
 963         g_free(exp);
 964     }
 965 }
 966
 967 BlockDriverState *nbd_export_get_blockdev(NBDExport *exp)
 968 {
 969     return exp->bs;
 970 }
 971
 972 void nbd_export_close_all(void)
 973 {
 974     NBDExport *exp, *next;
 975
 976     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 977         nbd_export_close(exp);
 978     }
 979 }
 980
 981 static int nbd_can_read(void *opaque);
 982 static void nbd_read(void *opaque);
 983 static void nbd_restart_write(void *opaque);
 984
 985 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 986                                  int len)
 987 {
 988     NBDClient *client = req->client;
 989     int csock = client->sock;
 990     ssize_t rc, ret;
 991
 992     qemu_co_mutex_lock(&client->send_lock);
 993     qemu_set_fd_handler2(csock, nbd_can_read, nbd_read,
 994                          nbd_restart_write, client);
 995     client->send_coroutine = qemu_coroutine_self();
 996
 997     if (!len) {
 998         rc = nbd_send_reply(csock, reply);
 999     } else {
1000         socket_set_cork(csock, 1);
1001         rc = nbd_send_reply(csock, reply);
1002         if (rc >= 0) {
1003             ret = qemu_co_send(csock, req->data, len);
1004             if (ret != len) {
1005                 rc = -EIO;
1006             }
1007         }
1008         socket_set_cork(csock, 0);
1009     }
1010
1011     client->send_coroutine = NULL;
1012     qemu_set_fd_handler2(csock, nbd_can_read, nbd_read, NULL, client);
1013     qemu_co_mutex_unlock(&client->send_lock);
1014     return rc;
1015 }
1016
1017 static ssize_t nbd_co_receive_request(NBDRequest *req, struct nbd_request *request)
1018 {
1019     NBDClient *client = req->client;
1020     int csock = client->sock;
1021     ssize_t rc;
1022
1023     client->recv_coroutine = qemu_coroutine_self();
1024     rc = nbd_receive_request(csock, request);
1025     if (rc < 0) {
1026         if (rc != -EAGAIN) {
1027             rc = -EIO;
1028         }
1029         goto out;
1030     }
1031
1032     if (request->len > NBD_BUFFER_SIZE) {
1033         LOG("len (%u) is larger than max len (%u)",
1034             request->len, NBD_BUFFER_SIZE);
1035         rc = -EINVAL;
1036         goto out;
1037     }
1038
1039     if ((request->from + request->len) < request->from) {
1040         LOG("integer overflow detected! "
1041             "you're probably being attacked");
1042         rc = -EINVAL;
1043         goto out;
1044     }
1045
1046     TRACE("Decoding type");
1047
1048     if ((request->type & NBD_CMD_MASK_COMMAND) == NBD_CMD_WRITE) {
1049         TRACE("Reading %u byte(s)", request->len);
1050
1051         if (qemu_co_recv(csock, req->data, request->len) != request->len) {
1052             LOG("reading from socket failed");
1053             rc = -EIO;
1054             goto out;
1055         }
1056     }
1057     rc = 0;
1058
1059 out:
1060     client->recv_coroutine = NULL;
1061     return rc;
1062 }
1063
1064 static void nbd_trip(void *opaque)
1065 {
1066     NBDClient *client = opaque;
1067     NBDExport *exp = client->exp;
1068     NBDRequest *req;
1069     struct nbd_request request;
1070     struct nbd_reply reply;
1071     ssize_t ret;
1072
1073     TRACE("Reading request.");
1074     if (client->closing) {
1075         return;
1076     }
1077
1078     req = nbd_request_get(client);
1079     ret = nbd_co_receive_request(req, &request);
1080     if (ret == -EAGAIN) {
1081         goto done;
1082     }
1083     if (ret == -EIO) {
1084         goto out;
1085     }
1086
1087     reply.handle = request.handle;
1088     reply.error = 0;
1089
1090     if (ret < 0) {
1091         reply.error = -ret;
1092         goto error_reply;
1093     }
1094
1095     if ((request.from + request.len) > exp->size) {
1096             LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
1097             ", Offset: %" PRIu64 "\n",
1098                     request.from, request.len,
1099                     (uint64_t)exp->size, (uint64_t)exp->dev_offset);
1100         LOG("requested operation past EOF--bad client?");
1101         goto invalid_request;
1102     }
1103
1104     switch (request.type & NBD_CMD_MASK_COMMAND) {
1105     case NBD_CMD_READ:
1106         TRACE("Request type is READ");
1107
1108         if (request.type & NBD_CMD_FLAG_FUA) {
1109             ret = bdrv_co_flush(exp->bs);
1110             if (ret < 0) {
1111                 LOG("flush failed");
1112                 reply.error = -ret;
1113                 goto error_reply;
1114             }
1115         }
1116
1117         ret = bdrv_read(exp->bs, (request.from + exp->dev_offset) / 512,
1118                         req->data, request.len / 512);
1119         if (ret < 0) {
1120             LOG("reading from file failed");
1121             reply.error = -ret;
1122             goto error_reply;
1123         }
1124
1125         TRACE("Read %u byte(s)", request.len);
1126         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1127             goto out;
1128         break;
1129     case NBD_CMD_WRITE:
1130         TRACE("Request type is WRITE");
1131
1132         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1133             TRACE("Server is read-only, return error");
1134             reply.error = EROFS;
1135             goto error_reply;
1136         }
1137
1138         TRACE("Writing to device");
1139
1140         ret = bdrv_write(exp->bs, (request.from + exp->dev_offset) / 512,
1141                          req->data, request.len / 512);
1142         if (ret < 0) {
1143             LOG("writing to file failed");
1144             reply.error = -ret;
1145             goto error_reply;
1146         }
1147
1148         if (request.type & NBD_CMD_FLAG_FUA) {
1149             ret = bdrv_co_flush(exp->bs);
1150             if (ret < 0) {
1151                 LOG("flush failed");
1152                 reply.error = -ret;
1153                 goto error_reply;
1154             }
1155         }
1156
1157         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1158             goto out;
1159         }
1160         break;
1161     case NBD_CMD_DISC:
1162         TRACE("Request type is DISCONNECT");
1163         errno = 0;
1164         goto out;
1165     case NBD_CMD_FLUSH:
1166         TRACE("Request type is FLUSH");
1167
1168         ret = bdrv_co_flush(exp->bs);
1169         if (ret < 0) {
1170             LOG("flush failed");
1171             reply.error = -ret;
1172         }
1173         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1174             goto out;
1175         }
1176         break;
1177     case NBD_CMD_TRIM:
1178         TRACE("Request type is TRIM");
1179         ret = bdrv_co_discard(exp->bs, (request.from + exp->dev_offset) / 512,
1180                               request.len / 512);
1181         if (ret < 0) {
1182             LOG("discard failed");
1183             reply.error = -ret;
1184         }
1185         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1186             goto out;
1187         }
1188         break;
1189     default:
1190         LOG("invalid request type (%u) received", request.type);
1191     invalid_request:
1192         reply.error = -EINVAL;
1193     error_reply:
1194         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1195             goto out;
1196         }
1197         break;
1198     }
1199
1200     TRACE("Request/Reply complete");
1201
1202 done:
1203     nbd_request_put(req);
1204     return;
1205
1206 out:
1207     nbd_request_put(req);
1208     nbd_client_close(client);
1209 }
1210
1211 static int nbd_can_read(void *opaque)
1212 {
1213     NBDClient *client = opaque;
1214
1215     return client->recv_coroutine || client->nb_requests < MAX_NBD_REQUESTS;
1216 }
1217
1218 static void nbd_read(void *opaque)
1219 {
1220     NBDClient *client = opaque;
1221
1222     if (client->recv_coroutine) {
1223         qemu_coroutine_enter(client->recv_coroutine, NULL);
1224     } else {
1225         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip), client);
1226     }
1227 }
1228
1229 static void nbd_restart_write(void *opaque)
1230 {
1231     NBDClient *client = opaque;
1232
1233     qemu_coroutine_enter(client->send_coroutine, NULL);
1234 }
1235
1236 NBDClient *nbd_client_new(NBDExport *exp, int csock,
1237                           void (*close)(NBDClient *))
1238 {
1239     NBDClient *client;
1240     client = g_malloc0(sizeof(NBDClient));
1241     client->refcount = 1;
1242     client->exp = exp;
1243     client->sock = csock;
1244     if (nbd_send_negotiate(client) < 0) {
1245         g_free(client);
1246         return NULL;
1247     }
1248     client->close = close;
1249     qemu_co_mutex_init(&client->send_lock);
1250     qemu_set_fd_handler2(csock, nbd_can_read, nbd_read, NULL, client);
1251
1252     if (exp) {
1253         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1254         nbd_export_get(exp);
1255     }
1256     return client;
1257 }