nbd/server.c

   1 /*
   2  *  Copyright (C) 2005  Anthony Liguori <anthony@codemonkey.ws>
   3  *
   4  *  Network Block Device Server Side
   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 "qemu/osdep.h"
  20 #include "qapi/error.h"
  21 #include "nbd-internal.h"
  22
  23 static int system_errno_to_nbd_errno(int err)
  24 {
  25     switch (err) {
  26     case 0:
  27         return NBD_SUCCESS;
  28     case EPERM:
  29     case EROFS:
  30         return NBD_EPERM;
  31     case EIO:
  32         return NBD_EIO;
  33     case ENOMEM:
  34         return NBD_ENOMEM;
  35 #ifdef EDQUOT
  36     case EDQUOT:
  37 #endif
  38     case EFBIG:
  39     case ENOSPC:
  40         return NBD_ENOSPC;
  41     case EINVAL:
  42     default:
  43         return NBD_EINVAL;
  44     }
  45 }
  46
  47 /* Definitions for opaque data types */
  48
  49 typedef struct NBDRequest NBDRequest;
  50
  51 struct NBDRequest {
  52     QSIMPLEQ_ENTRY(NBDRequest) entry;
  53     NBDClient *client;
  54     uint8_t *data;
  55     bool complete;
  56 };
  57
  58 struct NBDExport {
  59     int refcount;
  60     void (*close)(NBDExport *exp);
  61
  62     BlockBackend *blk;
  63     char *name;
  64     off_t dev_offset;
  65     off_t size;
  66     uint32_t nbdflags;
  67     QTAILQ_HEAD(, NBDClient) clients;
  68     QTAILQ_ENTRY(NBDExport) next;
  69
  70     AioContext *ctx;
  71
  72     Notifier eject_notifier;
  73 };
  74
  75 static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports);
  76
  77 struct NBDClient {
  78     int refcount;
  79     void (*close)(NBDClient *client);
  80
  81     NBDExport *exp;
  82     QCryptoTLSCreds *tlscreds;
  83     char *tlsaclname;
  84     QIOChannelSocket *sioc; /* The underlying data channel */
  85     QIOChannel *ioc; /* The current I/O channel which may differ (eg TLS) */
  86
  87     Coroutine *recv_coroutine;
  88
  89     CoMutex send_lock;
  90     Coroutine *send_coroutine;
  91
  92     bool can_read;
  93
  94     QTAILQ_ENTRY(NBDClient) next;
  95     int nb_requests;
  96     bool closing;
  97 };
  98
  99 /* That's all folks */
 100
 101 static void nbd_set_handlers(NBDClient *client);
 102 static void nbd_unset_handlers(NBDClient *client);
 103 static void nbd_update_can_read(NBDClient *client);
 104
 105 static gboolean nbd_negotiate_continue(QIOChannel *ioc,
 106                                        GIOCondition condition,
 107                                        void *opaque)
 108 {
 109     qemu_coroutine_enter(opaque, NULL);
 110     return TRUE;
 111 }
 112
 113 static ssize_t nbd_negotiate_read(QIOChannel *ioc, void *buffer, size_t size)
 114 {
 115     ssize_t ret;
 116     guint watch;
 117
 118     assert(qemu_in_coroutine());
 119     /* Negotiation are always in main loop. */
 120     watch = qio_channel_add_watch(ioc,
 121                                   G_IO_IN,
 122                                   nbd_negotiate_continue,
 123                                   qemu_coroutine_self(),
 124                                   NULL);
 125     ret = read_sync(ioc, buffer, size);
 126     g_source_remove(watch);
 127     return ret;
 128
 129 }
 130
 131 static ssize_t nbd_negotiate_write(QIOChannel *ioc, void *buffer, size_t size)
 132 {
 133     ssize_t ret;
 134     guint watch;
 135
 136     assert(qemu_in_coroutine());
 137     /* Negotiation are always in main loop. */
 138     watch = qio_channel_add_watch(ioc,
 139                                   G_IO_OUT,
 140                                   nbd_negotiate_continue,
 141                                   qemu_coroutine_self(),
 142                                   NULL);
 143     ret = write_sync(ioc, buffer, size);
 144     g_source_remove(watch);
 145     return ret;
 146 }
 147
 148 static ssize_t nbd_negotiate_drop_sync(QIOChannel *ioc, size_t size)
 149 {
 150     ssize_t ret, dropped = size;
 151     uint8_t *buffer = g_malloc(MIN(65536, size));
 152
 153     while (size > 0) {
 154         ret = nbd_negotiate_read(ioc, buffer, MIN(65536, size));
 155         if (ret < 0) {
 156             g_free(buffer);
 157             return ret;
 158         }
 159
 160         assert(ret <= size);
 161         size -= ret;
 162     }
 163
 164     g_free(buffer);
 165     return dropped;
 166 }
 167
 168 /* Basic flow for negotiation
 169
 170    Server         Client
 171    Negotiate
 172
 173    or
 174
 175    Server         Client
 176    Negotiate #1
 177                   Option
 178    Negotiate #2
 179
 180    ----
 181
 182    followed by
 183
 184    Server         Client
 185                   Request
 186    Response
 187                   Request
 188    Response
 189                   ...
 190    ...
 191                   Request (type == 2)
 192
 193 */
 194
 195 static int nbd_negotiate_send_rep(QIOChannel *ioc, uint32_t type, uint32_t opt)
 196 {
 197     uint64_t magic;
 198     uint32_t len;
 199
 200     TRACE("Reply opt=%" PRIx32 " type=%" PRIx32, type, opt);
 201
 202     magic = cpu_to_be64(NBD_REP_MAGIC);
 203     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 204         LOG("write failed (rep magic)");
 205         return -EINVAL;
 206     }
 207     opt = cpu_to_be32(opt);
 208     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 209         LOG("write failed (rep opt)");
 210         return -EINVAL;
 211     }
 212     type = cpu_to_be32(type);
 213     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 214         LOG("write failed (rep type)");
 215         return -EINVAL;
 216     }
 217     len = cpu_to_be32(0);
 218     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 219         LOG("write failed (rep data length)");
 220         return -EINVAL;
 221     }
 222     return 0;
 223 }
 224
 225 static int nbd_negotiate_send_rep_list(QIOChannel *ioc, NBDExport *exp)
 226 {
 227     uint64_t magic, name_len;
 228     uint32_t opt, type, len;
 229
 230     TRACE("Advertising export name '%s'", exp->name ? exp->name : "");
 231     name_len = strlen(exp->name);
 232     magic = cpu_to_be64(NBD_REP_MAGIC);
 233     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 234         LOG("write failed (magic)");
 235         return -EINVAL;
 236      }
 237     opt = cpu_to_be32(NBD_OPT_LIST);
 238     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 239         LOG("write failed (opt)");
 240         return -EINVAL;
 241     }
 242     type = cpu_to_be32(NBD_REP_SERVER);
 243     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 244         LOG("write failed (reply type)");
 245         return -EINVAL;
 246     }
 247     len = cpu_to_be32(name_len + sizeof(len));
 248     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 249         LOG("write failed (length)");
 250         return -EINVAL;
 251     }
 252     len = cpu_to_be32(name_len);
 253     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 254         LOG("write failed (length)");
 255         return -EINVAL;
 256     }
 257     if (nbd_negotiate_write(ioc, exp->name, name_len) != name_len) {
 258         LOG("write failed (buffer)");
 259         return -EINVAL;
 260     }
 261     return 0;
 262 }
 263
 264 static int nbd_negotiate_handle_list(NBDClient *client, uint32_t length)
 265 {
 266     NBDExport *exp;
 267
 268     if (length) {
 269         if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 270             return -EIO;
 271         }
 272         return nbd_negotiate_send_rep(client->ioc,
 273                                       NBD_REP_ERR_INVALID, NBD_OPT_LIST);
 274     }
 275
 276     /* For each export, send a NBD_REP_SERVER reply. */
 277     QTAILQ_FOREACH(exp, &exports, next) {
 278         if (nbd_negotiate_send_rep_list(client->ioc, exp)) {
 279             return -EINVAL;
 280         }
 281     }
 282     /* Finish with a NBD_REP_ACK. */
 283     return nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_LIST);
 284 }
 285
 286 static int nbd_negotiate_handle_export_name(NBDClient *client, uint32_t length)
 287 {
 288     int rc = -EINVAL;
 289     char name[256];
 290
 291     /* Client sends:
 292         [20 ..  xx]   export name (length bytes)
 293      */
 294     TRACE("Checking length");
 295     if (length > 255) {
 296         LOG("Bad length received");
 297         goto fail;
 298     }
 299     if (nbd_negotiate_read(client->ioc, name, length) != length) {
 300         LOG("read failed");
 301         goto fail;
 302     }
 303     name[length] = '\0';
 304
 305     TRACE("Client requested export '%s'", name);
 306
 307     client->exp = nbd_export_find(name);
 308     if (!client->exp) {
 309         LOG("export not found");
 310         goto fail;
 311     }
 312
 313     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 314     nbd_export_get(client->exp);
 315     rc = 0;
 316 fail:
 317     return rc;
 318 }
 319
 320
 321 static QIOChannel *nbd_negotiate_handle_starttls(NBDClient *client,
 322                                                  uint32_t length)
 323 {
 324     QIOChannel *ioc;
 325     QIOChannelTLS *tioc;
 326     struct NBDTLSHandshakeData data = { 0 };
 327
 328     TRACE("Setting up TLS");
 329     ioc = client->ioc;
 330     if (length) {
 331         if (nbd_negotiate_drop_sync(ioc, length) != length) {
 332             return NULL;
 333         }
 334         nbd_negotiate_send_rep(ioc, NBD_REP_ERR_INVALID, NBD_OPT_STARTTLS);
 335         return NULL;
 336     }
 337
 338     if (nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK,
 339                                NBD_OPT_STARTTLS) < 0) {
 340         return NULL;
 341     }
 342
 343     tioc = qio_channel_tls_new_server(ioc,
 344                                       client->tlscreds,
 345                                       client->tlsaclname,
 346                                       NULL);
 347     if (!tioc) {
 348         return NULL;
 349     }
 350
 351     TRACE("Starting TLS handshake");
 352     data.loop = g_main_loop_new(g_main_context_default(), FALSE);
 353     qio_channel_tls_handshake(tioc,
 354                               nbd_tls_handshake,
 355                               &data,
 356                               NULL);
 357
 358     if (!data.complete) {
 359         g_main_loop_run(data.loop);
 360     }
 361     g_main_loop_unref(data.loop);
 362     if (data.error) {
 363         object_unref(OBJECT(tioc));
 364         error_free(data.error);
 365         return NULL;
 366     }
 367
 368     return QIO_CHANNEL(tioc);
 369 }
 370
 371
 372 static int nbd_negotiate_options(NBDClient *client)
 373 {
 374     uint32_t flags;
 375     bool fixedNewstyle = false;
 376
 377     /* Client sends:
 378         [ 0 ..   3]   client flags
 379
 380         [ 0 ..   7]   NBD_OPTS_MAGIC
 381         [ 8 ..  11]   NBD option
 382         [12 ..  15]   Data length
 383         ...           Rest of request
 384
 385         [ 0 ..   7]   NBD_OPTS_MAGIC
 386         [ 8 ..  11]   Second NBD option
 387         [12 ..  15]   Data length
 388         ...           Rest of request
 389     */
 390
 391     if (nbd_negotiate_read(client->ioc, &flags, sizeof(flags)) !=
 392         sizeof(flags)) {
 393         LOG("read failed");
 394         return -EIO;
 395     }
 396     TRACE("Checking client flags");
 397     be32_to_cpus(&flags);
 398     if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) {
 399         TRACE("Client supports fixed newstyle handshake");
 400         fixedNewstyle = true;
 401         flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE;
 402     }
 403     if (flags != 0) {
 404         TRACE("Unknown client flags 0x%" PRIx32 " received", flags);
 405         return -EIO;
 406     }
 407
 408     while (1) {
 409         int ret;
 410         uint32_t clientflags, length;
 411         uint64_t magic;
 412
 413         if (nbd_negotiate_read(client->ioc, &magic, sizeof(magic)) !=
 414             sizeof(magic)) {
 415             LOG("read failed");
 416             return -EINVAL;
 417         }
 418         TRACE("Checking opts magic");
 419         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 420             LOG("Bad magic received");
 421             return -EINVAL;
 422         }
 423
 424         if (nbd_negotiate_read(client->ioc, &clientflags,
 425                                sizeof(clientflags)) != sizeof(clientflags)) {
 426             LOG("read failed");
 427             return -EINVAL;
 428         }
 429         clientflags = be32_to_cpu(clientflags);
 430
 431         if (nbd_negotiate_read(client->ioc, &length, sizeof(length)) !=
 432             sizeof(length)) {
 433             LOG("read failed");
 434             return -EINVAL;
 435         }
 436         length = be32_to_cpu(length);
 437
 438         TRACE("Checking option 0x%" PRIx32, clientflags);
 439         if (client->tlscreds &&
 440             client->ioc == (QIOChannel *)client->sioc) {
 441             QIOChannel *tioc;
 442             if (!fixedNewstyle) {
 443                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 444                 return -EINVAL;
 445             }
 446             switch (clientflags) {
 447             case NBD_OPT_STARTTLS:
 448                 tioc = nbd_negotiate_handle_starttls(client, length);
 449                 if (!tioc) {
 450                     return -EIO;
 451                 }
 452                 object_unref(OBJECT(client->ioc));
 453                 client->ioc = QIO_CHANNEL(tioc);
 454                 break;
 455
 456             case NBD_OPT_EXPORT_NAME:
 457                 /* No way to return an error to client, so drop connection */
 458                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 459                 return -EINVAL;
 460
 461             default:
 462                 TRACE("Option 0x%" PRIx32 " not permitted before TLS",
 463                       clientflags);
 464                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 465                     return -EIO;
 466                 }
 467                 ret = nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_TLS_REQD,
 468                                              clientflags);
 469                 if (ret < 0) {
 470                     return ret;
 471                 }
 472                 break;
 473             }
 474         } else if (fixedNewstyle) {
 475             switch (clientflags) {
 476             case NBD_OPT_LIST:
 477                 ret = nbd_negotiate_handle_list(client, length);
 478                 if (ret < 0) {
 479                     return ret;
 480                 }
 481                 break;
 482
 483             case NBD_OPT_ABORT:
 484                 return -EINVAL;
 485
 486             case NBD_OPT_EXPORT_NAME:
 487                 return nbd_negotiate_handle_export_name(client, length);
 488
 489             case NBD_OPT_STARTTLS:
 490                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 491                     return -EIO;
 492                 }
 493                 if (client->tlscreds) {
 494                     TRACE("TLS already enabled");
 495                     ret = nbd_negotiate_send_rep(client->ioc,
 496                                                  NBD_REP_ERR_INVALID,
 497                                                  clientflags);
 498                 } else {
 499                     TRACE("TLS not configured");
 500                     ret = nbd_negotiate_send_rep(client->ioc,
 501                                                  NBD_REP_ERR_POLICY,
 502                                                  clientflags);
 503                 }
 504                 if (ret < 0) {
 505                     return ret;
 506                 }
 507                 break;
 508             default:
 509                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 510                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 511                     return -EIO;
 512                 }
 513                 ret = nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_UNSUP,
 514                                              clientflags);
 515                 if (ret < 0) {
 516                     return ret;
 517                 }
 518                 break;
 519             }
 520         } else {
 521             /*
 522              * If broken new-style we should drop the connection
 523              * for anything except NBD_OPT_EXPORT_NAME
 524              */
 525             switch (clientflags) {
 526             case NBD_OPT_EXPORT_NAME:
 527                 return nbd_negotiate_handle_export_name(client, length);
 528
 529             default:
 530                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 531                 return -EINVAL;
 532             }
 533         }
 534     }
 535 }
 536
 537 typedef struct {
 538     NBDClient *client;
 539     Coroutine *co;
 540 } NBDClientNewData;
 541
 542 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 543 {
 544     NBDClient *client = data->client;
 545     char buf[8 + 8 + 8 + 128];
 546     int rc;
 547     const int myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 548                          NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 549     bool oldStyle;
 550
 551     /* Old style negotiation header without options
 552         [ 0 ..   7]   passwd       ("NBDMAGIC")
 553         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 554         [16 ..  23]   size
 555         [24 ..  25]   server flags (0)
 556         [26 ..  27]   export flags
 557         [28 .. 151]   reserved     (0)
 558
 559        New style negotiation header with options
 560         [ 0 ..   7]   passwd       ("NBDMAGIC")
 561         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 562         [16 ..  17]   server flags (0)
 563         ....options sent....
 564         [18 ..  25]   size
 565         [26 ..  27]   export flags
 566         [28 .. 151]   reserved     (0)
 567      */
 568
 569     qio_channel_set_blocking(client->ioc, false, NULL);
 570     rc = -EINVAL;
 571
 572     TRACE("Beginning negotiation.");
 573     memset(buf, 0, sizeof(buf));
 574     memcpy(buf, "NBDMAGIC", 8);
 575
 576     oldStyle = client->exp != NULL && !client->tlscreds;
 577     if (oldStyle) {
 578         assert ((client->exp->nbdflags & ~65535) == 0);
 579         TRACE("advertising size %" PRIu64 " and flags %x",
 580               client->exp->size, client->exp->nbdflags | myflags);
 581         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 582         stq_be_p(buf + 16, client->exp->size);
 583         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 584     } else {
 585         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 586         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE);
 587     }
 588
 589     if (oldStyle) {
 590         if (client->tlscreds) {
 591             TRACE("TLS cannot be enabled with oldstyle protocol");
 592             goto fail;
 593         }
 594         if (nbd_negotiate_write(client->ioc, buf, sizeof(buf)) != sizeof(buf)) {
 595             LOG("write failed");
 596             goto fail;
 597         }
 598     } else {
 599         if (nbd_negotiate_write(client->ioc, buf, 18) != 18) {
 600             LOG("write failed");
 601             goto fail;
 602         }
 603         rc = nbd_negotiate_options(client);
 604         if (rc != 0) {
 605             LOG("option negotiation failed");
 606             goto fail;
 607         }
 608
 609         assert ((client->exp->nbdflags & ~65535) == 0);
 610         TRACE("advertising size %" PRIu64 " and flags %x",
 611               client->exp->size, client->exp->nbdflags | myflags);
 612         stq_be_p(buf + 18, client->exp->size);
 613         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 614         if (nbd_negotiate_write(client->ioc, buf + 18, sizeof(buf) - 18) !=
 615             sizeof(buf) - 18) {
 616             LOG("write failed");
 617             goto fail;
 618         }
 619     }
 620
 621     TRACE("Negotiation succeeded.");
 622     rc = 0;
 623 fail:
 624     return rc;
 625 }
 626
 627 #ifdef __linux__
 628
 629 int nbd_disconnect(int fd)
 630 {
 631     ioctl(fd, NBD_CLEAR_QUE);
 632     ioctl(fd, NBD_DISCONNECT);
 633     ioctl(fd, NBD_CLEAR_SOCK);
 634     return 0;
 635 }
 636
 637 #else
 638
 639 int nbd_disconnect(int fd)
 640 {
 641     return -ENOTSUP;
 642 }
 643 #endif
 644
 645 static ssize_t nbd_receive_request(QIOChannel *ioc, struct nbd_request *request)
 646 {
 647     uint8_t buf[NBD_REQUEST_SIZE];
 648     uint32_t magic;
 649     ssize_t ret;
 650
 651     ret = read_sync(ioc, buf, sizeof(buf));
 652     if (ret < 0) {
 653         return ret;
 654     }
 655
 656     if (ret != sizeof(buf)) {
 657         LOG("read failed");
 658         return -EINVAL;
 659     }
 660
 661     /* Request
 662        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 663        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 664        [ 8 .. 15]   handle
 665        [16 .. 23]   from
 666        [24 .. 27]   len
 667      */
 668
 669     magic = ldl_be_p(buf);
 670     request->type   = ldl_be_p(buf + 4);
 671     request->handle = ldq_be_p(buf + 8);
 672     request->from   = ldq_be_p(buf + 16);
 673     request->len    = ldl_be_p(buf + 24);
 674
 675     TRACE("Got request: { magic = 0x%" PRIx32 ", .type = %" PRIx32
 676           ", from = %" PRIu64 " , len = %" PRIu32 " }",
 677           magic, request->type, request->from, request->len);
 678
 679     if (magic != NBD_REQUEST_MAGIC) {
 680         LOG("invalid magic (got 0x%" PRIx32 ")", magic);
 681         return -EINVAL;
 682     }
 683     return 0;
 684 }
 685
 686 static ssize_t nbd_send_reply(QIOChannel *ioc, struct nbd_reply *reply)
 687 {
 688     uint8_t buf[NBD_REPLY_SIZE];
 689     ssize_t ret;
 690
 691     reply->error = system_errno_to_nbd_errno(reply->error);
 692
 693     TRACE("Sending response to client: { .error = %" PRId32
 694           ", handle = %" PRIu64 " }",
 695           reply->error, reply->handle);
 696
 697     /* Reply
 698        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 699        [ 4 ..  7]    error   (0 == no error)
 700        [ 7 .. 15]    handle
 701      */
 702     stl_be_p(buf, NBD_REPLY_MAGIC);
 703     stl_be_p(buf + 4, reply->error);
 704     stq_be_p(buf + 8, reply->handle);
 705
 706     ret = write_sync(ioc, buf, sizeof(buf));
 707     if (ret < 0) {
 708         return ret;
 709     }
 710
 711     if (ret != sizeof(buf)) {
 712         LOG("writing to socket failed");
 713         return -EINVAL;
 714     }
 715     return 0;
 716 }
 717
 718 #define MAX_NBD_REQUESTS 16
 719
 720 void nbd_client_get(NBDClient *client)
 721 {
 722     client->refcount++;
 723 }
 724
 725 void nbd_client_put(NBDClient *client)
 726 {
 727     if (--client->refcount == 0) {
 728         /* The last reference should be dropped by client->close,
 729          * which is called by client_close.
 730          */
 731         assert(client->closing);
 732
 733         nbd_unset_handlers(client);
 734         object_unref(OBJECT(client->sioc));
 735         object_unref(OBJECT(client->ioc));
 736         if (client->tlscreds) {
 737             object_unref(OBJECT(client->tlscreds));
 738         }
 739         g_free(client->tlsaclname);
 740         if (client->exp) {
 741             QTAILQ_REMOVE(&client->exp->clients, client, next);
 742             nbd_export_put(client->exp);
 743         }
 744         g_free(client);
 745     }
 746 }
 747
 748 static void client_close(NBDClient *client)
 749 {
 750     if (client->closing) {
 751         return;
 752     }
 753
 754     client->closing = true;
 755
 756     /* Force requests to finish.  They will drop their own references,
 757      * then we'll close the socket and free the NBDClient.
 758      */
 759     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 760                          NULL);
 761
 762     /* Also tell the client, so that they release their reference.  */
 763     if (client->close) {
 764         client->close(client);
 765     }
 766 }
 767
 768 static NBDRequest *nbd_request_get(NBDClient *client)
 769 {
 770     NBDRequest *req;
 771
 772     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 773     client->nb_requests++;
 774     nbd_update_can_read(client);
 775
 776     req = g_new0(NBDRequest, 1);
 777     nbd_client_get(client);
 778     req->client = client;
 779     return req;
 780 }
 781
 782 static void nbd_request_put(NBDRequest *req)
 783 {
 784     NBDClient *client = req->client;
 785
 786     if (req->data) {
 787         qemu_vfree(req->data);
 788     }
 789     g_free(req);
 790
 791     client->nb_requests--;
 792     nbd_update_can_read(client);
 793     nbd_client_put(client);
 794 }
 795
 796 static void blk_aio_attached(AioContext *ctx, void *opaque)
 797 {
 798     NBDExport *exp = opaque;
 799     NBDClient *client;
 800
 801     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 802
 803     exp->ctx = ctx;
 804
 805     QTAILQ_FOREACH(client, &exp->clients, next) {
 806         nbd_set_handlers(client);
 807     }
 808 }
 809
 810 static void blk_aio_detach(void *opaque)
 811 {
 812     NBDExport *exp = opaque;
 813     NBDClient *client;
 814
 815     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 816
 817     QTAILQ_FOREACH(client, &exp->clients, next) {
 818         nbd_unset_handlers(client);
 819     }
 820
 821     exp->ctx = NULL;
 822 }
 823
 824 static void nbd_eject_notifier(Notifier *n, void *data)
 825 {
 826     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 827     nbd_export_close(exp);
 828 }
 829
 830 NBDExport *nbd_export_new(BlockBackend *blk, off_t dev_offset, off_t size,
 831                           uint32_t nbdflags, void (*close)(NBDExport *),
 832                           Error **errp)
 833 {
 834     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 835     exp->refcount = 1;
 836     QTAILQ_INIT(&exp->clients);
 837     exp->blk = blk;
 838     exp->dev_offset = dev_offset;
 839     exp->nbdflags = nbdflags;
 840     exp->size = size < 0 ? blk_getlength(blk) : size;
 841     if (exp->size < 0) {
 842         error_setg_errno(errp, -exp->size,
 843                          "Failed to determine the NBD export's length");
 844         goto fail;
 845     }
 846     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 847
 848     exp->close = close;
 849     exp->ctx = blk_get_aio_context(blk);
 850     blk_ref(blk);
 851     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 852
 853     exp->eject_notifier.notify = nbd_eject_notifier;
 854     blk_add_remove_bs_notifier(blk, &exp->eject_notifier);
 855
 856     /*
 857      * NBD exports are used for non-shared storage migration.  Make sure
 858      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 859      * access since the export could be available before migration handover.
 860      */
 861     aio_context_acquire(exp->ctx);
 862     blk_invalidate_cache(blk, NULL);
 863     aio_context_release(exp->ctx);
 864     return exp;
 865
 866 fail:
 867     g_free(exp);
 868     return NULL;
 869 }
 870
 871 NBDExport *nbd_export_find(const char *name)
 872 {
 873     NBDExport *exp;
 874     QTAILQ_FOREACH(exp, &exports, next) {
 875         if (strcmp(name, exp->name) == 0) {
 876             return exp;
 877         }
 878     }
 879
 880     return NULL;
 881 }
 882
 883 void nbd_export_set_name(NBDExport *exp, const char *name)
 884 {
 885     if (exp->name == name) {
 886         return;
 887     }
 888
 889     nbd_export_get(exp);
 890     if (exp->name != NULL) {
 891         g_free(exp->name);
 892         exp->name = NULL;
 893         QTAILQ_REMOVE(&exports, exp, next);
 894         nbd_export_put(exp);
 895     }
 896     if (name != NULL) {
 897         nbd_export_get(exp);
 898         exp->name = g_strdup(name);
 899         QTAILQ_INSERT_TAIL(&exports, exp, next);
 900     }
 901     nbd_export_put(exp);
 902 }
 903
 904 void nbd_export_close(NBDExport *exp)
 905 {
 906     NBDClient *client, *next;
 907
 908     nbd_export_get(exp);
 909     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 910         client_close(client);
 911     }
 912     nbd_export_set_name(exp, NULL);
 913     nbd_export_put(exp);
 914 }
 915
 916 void nbd_export_get(NBDExport *exp)
 917 {
 918     assert(exp->refcount > 0);
 919     exp->refcount++;
 920 }
 921
 922 void nbd_export_put(NBDExport *exp)
 923 {
 924     assert(exp->refcount > 0);
 925     if (exp->refcount == 1) {
 926         nbd_export_close(exp);
 927     }
 928
 929     if (--exp->refcount == 0) {
 930         assert(exp->name == NULL);
 931
 932         if (exp->close) {
 933             exp->close(exp);
 934         }
 935
 936         if (exp->blk) {
 937             notifier_remove(&exp->eject_notifier);
 938             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 939                                             blk_aio_detach, exp);
 940             blk_unref(exp->blk);
 941             exp->blk = NULL;
 942         }
 943
 944         g_free(exp);
 945     }
 946 }
 947
 948 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
 949 {
 950     return exp->blk;
 951 }
 952
 953 void nbd_export_close_all(void)
 954 {
 955     NBDExport *exp, *next;
 956
 957     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 958         nbd_export_close(exp);
 959     }
 960 }
 961
 962 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 963                                  int len)
 964 {
 965     NBDClient *client = req->client;
 966     ssize_t rc, ret;
 967
 968     g_assert(qemu_in_coroutine());
 969     qemu_co_mutex_lock(&client->send_lock);
 970     client->send_coroutine = qemu_coroutine_self();
 971     nbd_set_handlers(client);
 972
 973     if (!len) {
 974         rc = nbd_send_reply(client->ioc, reply);
 975     } else {
 976         qio_channel_set_cork(client->ioc, true);
 977         rc = nbd_send_reply(client->ioc, reply);
 978         if (rc >= 0) {
 979             ret = write_sync(client->ioc, req->data, len);
 980             if (ret != len) {
 981                 rc = -EIO;
 982             }
 983         }
 984         qio_channel_set_cork(client->ioc, false);
 985     }
 986
 987     client->send_coroutine = NULL;
 988     nbd_set_handlers(client);
 989     qemu_co_mutex_unlock(&client->send_lock);
 990     return rc;
 991 }
 992
 993 /* Collect a client request.  Return 0 if request looks valid, -EAGAIN
 994  * to keep trying the collection, -EIO to drop connection right away,
 995  * and any other negative value to report an error to the client
 996  * (although the caller may still need to disconnect after reporting
 997  * the error).  */
 998 static ssize_t nbd_co_receive_request(NBDRequest *req,
 999                                       struct nbd_request *request)
1000 {
1001     NBDClient *client = req->client;
1002     uint32_t command;
1003     ssize_t rc;
1004
1005     g_assert(qemu_in_coroutine());
1006     client->recv_coroutine = qemu_coroutine_self();
1007     nbd_update_can_read(client);
1008
1009     rc = nbd_receive_request(client->ioc, request);
1010     if (rc < 0) {
1011         if (rc != -EAGAIN) {
1012             rc = -EIO;
1013         }
1014         goto out;
1015     }
1016
1017     TRACE("Decoding type");
1018
1019     command = request->type & NBD_CMD_MASK_COMMAND;
1020     if (command != NBD_CMD_WRITE) {
1021         /* No payload, we are ready to read the next request.  */
1022         req->complete = true;
1023     }
1024
1025     if (command == NBD_CMD_DISC) {
1026         /* Special case: we're going to disconnect without a reply,
1027          * whether or not flags, from, or len are bogus */
1028         TRACE("Request type is DISCONNECT");
1029         rc = -EIO;
1030         goto out;
1031     }
1032
1033     /* Check for sanity in the parameters, part 1.  Defer as many
1034      * checks as possible until after reading any NBD_CMD_WRITE
1035      * payload, so we can try and keep the connection alive.  */
1036     if ((request->from + request->len) < request->from) {
1037         LOG("integer overflow detected, you're probably being attacked");
1038         rc = -EINVAL;
1039         goto out;
1040     }
1041
1042     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
1043         if (request->len > NBD_MAX_BUFFER_SIZE) {
1044             LOG("len (%" PRIu32" ) is larger than max len (%u)",
1045                 request->len, NBD_MAX_BUFFER_SIZE);
1046             rc = -EINVAL;
1047             goto out;
1048         }
1049
1050         req->data = blk_try_blockalign(client->exp->blk, request->len);
1051         if (req->data == NULL) {
1052             rc = -ENOMEM;
1053             goto out;
1054         }
1055     }
1056     if (command == NBD_CMD_WRITE) {
1057         TRACE("Reading %" PRIu32 " byte(s)", request->len);
1058
1059         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1060             LOG("reading from socket failed");
1061             rc = -EIO;
1062             goto out;
1063         }
1064         req->complete = true;
1065     }
1066
1067     /* Sanity checks, part 2. */
1068     if (request->from + request->len > client->exp->size) {
1069         LOG("operation past EOF; From: %" PRIu64 ", Len: %" PRIu32
1070             ", Size: %" PRIu64, request->from, request->len,
1071             (uint64_t)client->exp->size);
1072         rc = command == NBD_CMD_WRITE ? -ENOSPC : -EINVAL;
1073         goto out;
1074     }
1075     if (request->type & ~NBD_CMD_MASK_COMMAND & ~NBD_CMD_FLAG_FUA) {
1076         LOG("unsupported flags (got 0x%x)",
1077             request->type & ~NBD_CMD_MASK_COMMAND);
1078         return -EINVAL;
1079     }
1080
1081     rc = 0;
1082
1083 out:
1084     client->recv_coroutine = NULL;
1085     nbd_update_can_read(client);
1086
1087     return rc;
1088 }
1089
1090 static void nbd_trip(void *opaque)
1091 {
1092     NBDClient *client = opaque;
1093     NBDExport *exp = client->exp;
1094     NBDRequest *req;
1095     struct nbd_request request;
1096     struct nbd_reply reply;
1097     ssize_t ret;
1098     uint32_t command;
1099     int flags;
1100
1101     TRACE("Reading request.");
1102     if (client->closing) {
1103         return;
1104     }
1105
1106     req = nbd_request_get(client);
1107     ret = nbd_co_receive_request(req, &request);
1108     if (ret == -EAGAIN) {
1109         goto done;
1110     }
1111     if (ret == -EIO) {
1112         goto out;
1113     }
1114
1115     reply.handle = request.handle;
1116     reply.error = 0;
1117
1118     if (ret < 0) {
1119         reply.error = -ret;
1120         goto error_reply;
1121     }
1122     command = request.type & NBD_CMD_MASK_COMMAND;
1123
1124     if (client->closing) {
1125         /*
1126          * The client may be closed when we are blocked in
1127          * nbd_co_receive_request()
1128          */
1129         goto done;
1130     }
1131
1132     switch (command) {
1133     case NBD_CMD_READ:
1134         TRACE("Request type is READ");
1135
1136         if (request.type & NBD_CMD_FLAG_FUA) {
1137             ret = blk_co_flush(exp->blk);
1138             if (ret < 0) {
1139                 LOG("flush failed");
1140                 reply.error = -ret;
1141                 goto error_reply;
1142             }
1143         }
1144
1145         ret = blk_pread(exp->blk, request.from + exp->dev_offset,
1146                         req->data, request.len);
1147         if (ret < 0) {
1148             LOG("reading from file failed");
1149             reply.error = -ret;
1150             goto error_reply;
1151         }
1152
1153         TRACE("Read %" PRIu32" byte(s)", request.len);
1154         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1155             goto out;
1156         break;
1157     case NBD_CMD_WRITE:
1158         TRACE("Request type is WRITE");
1159
1160         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1161             TRACE("Server is read-only, return error");
1162             reply.error = EROFS;
1163             goto error_reply;
1164         }
1165
1166         TRACE("Writing to device");
1167
1168         flags = 0;
1169         if (request.type & NBD_CMD_FLAG_FUA) {
1170             flags |= BDRV_REQ_FUA;
1171         }
1172         ret = blk_pwrite(exp->blk, request.from + exp->dev_offset,
1173                          req->data, request.len, flags);
1174         if (ret < 0) {
1175             LOG("writing to file failed");
1176             reply.error = -ret;
1177             goto error_reply;
1178         }
1179
1180         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1181             goto out;
1182         }
1183         break;
1184
1185     case NBD_CMD_DISC:
1186         /* unreachable, thanks to special case in nbd_co_receive_request() */
1187         abort();
1188
1189     case NBD_CMD_FLUSH:
1190         TRACE("Request type is FLUSH");
1191
1192         ret = blk_co_flush(exp->blk);
1193         if (ret < 0) {
1194             LOG("flush failed");
1195             reply.error = -ret;
1196         }
1197         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1198             goto out;
1199         }
1200         break;
1201     case NBD_CMD_TRIM:
1202         TRACE("Request type is TRIM");
1203         /* Ignore unaligned head or tail, until block layer adds byte
1204          * interface */
1205         if (request.len >= BDRV_SECTOR_SIZE) {
1206             request.len -= (request.from + request.len) % BDRV_SECTOR_SIZE;
1207             ret = blk_co_discard(exp->blk,
1208                                  DIV_ROUND_UP(request.from + exp->dev_offset,
1209                                               BDRV_SECTOR_SIZE),
1210                                  request.len / BDRV_SECTOR_SIZE);
1211             if (ret < 0) {
1212                 LOG("discard failed");
1213                 reply.error = -ret;
1214             }
1215         } else {
1216             TRACE("trim request too small, ignoring");
1217         }
1218         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1219             goto out;
1220         }
1221         break;
1222     default:
1223         LOG("invalid request type (%" PRIu32 ") received", request.type);
1224         reply.error = EINVAL;
1225     error_reply:
1226         /* We must disconnect after NBD_CMD_WRITE if we did not
1227          * read the payload.
1228          */
1229         if (nbd_co_send_reply(req, &reply, 0) < 0 || !req->complete) {
1230             goto out;
1231         }
1232         break;
1233     }
1234
1235     TRACE("Request/Reply complete");
1236
1237 done:
1238     nbd_request_put(req);
1239     return;
1240
1241 out:
1242     nbd_request_put(req);
1243     client_close(client);
1244 }
1245
1246 static void nbd_read(void *opaque)
1247 {
1248     NBDClient *client = opaque;
1249
1250     if (client->recv_coroutine) {
1251         qemu_coroutine_enter(client->recv_coroutine, NULL);
1252     } else {
1253         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip), client);
1254     }
1255 }
1256
1257 static void nbd_restart_write(void *opaque)
1258 {
1259     NBDClient *client = opaque;
1260
1261     qemu_coroutine_enter(client->send_coroutine, NULL);
1262 }
1263
1264 static void nbd_set_handlers(NBDClient *client)
1265 {
1266     if (client->exp && client->exp->ctx) {
1267         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1268                            true,
1269                            client->can_read ? nbd_read : NULL,
1270                            client->send_coroutine ? nbd_restart_write : NULL,
1271                            client);
1272     }
1273 }
1274
1275 static void nbd_unset_handlers(NBDClient *client)
1276 {
1277     if (client->exp && client->exp->ctx) {
1278         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1279                            true, NULL, NULL, NULL);
1280     }
1281 }
1282
1283 static void nbd_update_can_read(NBDClient *client)
1284 {
1285     bool can_read = client->recv_coroutine ||
1286                     client->nb_requests < MAX_NBD_REQUESTS;
1287
1288     if (can_read != client->can_read) {
1289         client->can_read = can_read;
1290         nbd_set_handlers(client);
1291
1292         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1293          * in nbd_set_handlers() will have taken care of that */
1294     }
1295 }
1296
1297 static coroutine_fn void nbd_co_client_start(void *opaque)
1298 {
1299     NBDClientNewData *data = opaque;
1300     NBDClient *client = data->client;
1301     NBDExport *exp = client->exp;
1302
1303     if (exp) {
1304         nbd_export_get(exp);
1305     }
1306     if (nbd_negotiate(data)) {
1307         client_close(client);
1308         goto out;
1309     }
1310     qemu_co_mutex_init(&client->send_lock);
1311     nbd_set_handlers(client);
1312
1313     if (exp) {
1314         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1315     }
1316 out:
1317     g_free(data);
1318 }
1319
1320 void nbd_client_new(NBDExport *exp,
1321                     QIOChannelSocket *sioc,
1322                     QCryptoTLSCreds *tlscreds,
1323                     const char *tlsaclname,
1324                     void (*close_fn)(NBDClient *))
1325 {
1326     NBDClient *client;
1327     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1328
1329     client = g_malloc0(sizeof(NBDClient));
1330     client->refcount = 1;
1331     client->exp = exp;
1332     client->tlscreds = tlscreds;
1333     if (tlscreds) {
1334         object_ref(OBJECT(client->tlscreds));
1335     }
1336     client->tlsaclname = g_strdup(tlsaclname);
1337     client->sioc = sioc;
1338     object_ref(OBJECT(client->sioc));
1339     client->ioc = QIO_CHANNEL(sioc);
1340     object_ref(OBJECT(client->ioc));
1341     client->can_read = true;
1342     client->close = close_fn;
1343
1344     data->client = client;
1345     data->co = qemu_coroutine_create(nbd_co_client_start);
1346     qemu_coroutine_enter(data->co, data);
1347 }