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     uint16_t nbdflags;
  67     QTAILQ_HEAD(, NBDClient) clients;
  68     QTAILQ_ENTRY(NBDExport) next;
  69
  70     AioContext *ctx;
  71
  72     BlockBackend *eject_notifier_blk;
  73     Notifier eject_notifier;
  74 };
  75
  76 static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports);
  77
  78 struct NBDClient {
  79     int refcount;
  80     void (*close)(NBDClient *client);
  81
  82     NBDExport *exp;
  83     QCryptoTLSCreds *tlscreds;
  84     char *tlsaclname;
  85     QIOChannelSocket *sioc; /* The underlying data channel */
  86     QIOChannel *ioc; /* The current I/O channel which may differ (eg TLS) */
  87
  88     Coroutine *recv_coroutine;
  89
  90     CoMutex send_lock;
  91     Coroutine *send_coroutine;
  92
  93     bool can_read;
  94
  95     QTAILQ_ENTRY(NBDClient) next;
  96     int nb_requests;
  97     bool closing;
  98 };
  99
 100 /* That's all folks */
 101
 102 static void nbd_set_handlers(NBDClient *client);
 103 static void nbd_unset_handlers(NBDClient *client);
 104 static void nbd_update_can_read(NBDClient *client);
 105
 106 static gboolean nbd_negotiate_continue(QIOChannel *ioc,
 107                                        GIOCondition condition,
 108                                        void *opaque)
 109 {
 110     qemu_coroutine_enter(opaque);
 111     return TRUE;
 112 }
 113
 114 static ssize_t nbd_negotiate_read(QIOChannel *ioc, void *buffer, size_t size)
 115 {
 116     ssize_t ret;
 117     guint watch;
 118
 119     assert(qemu_in_coroutine());
 120     /* Negotiation are always in main loop. */
 121     watch = qio_channel_add_watch(ioc,
 122                                   G_IO_IN,
 123                                   nbd_negotiate_continue,
 124                                   qemu_coroutine_self(),
 125                                   NULL);
 126     ret = read_sync(ioc, buffer, size);
 127     g_source_remove(watch);
 128     return ret;
 129
 130 }
 131
 132 static ssize_t nbd_negotiate_write(QIOChannel *ioc, void *buffer, size_t size)
 133 {
 134     ssize_t ret;
 135     guint watch;
 136
 137     assert(qemu_in_coroutine());
 138     /* Negotiation are always in main loop. */
 139     watch = qio_channel_add_watch(ioc,
 140                                   G_IO_OUT,
 141                                   nbd_negotiate_continue,
 142                                   qemu_coroutine_self(),
 143                                   NULL);
 144     ret = write_sync(ioc, buffer, size);
 145     g_source_remove(watch);
 146     return ret;
 147 }
 148
 149 static ssize_t nbd_negotiate_drop_sync(QIOChannel *ioc, size_t size)
 150 {
 151     ssize_t ret, dropped = size;
 152     uint8_t *buffer = g_malloc(MIN(65536, size));
 153
 154     while (size > 0) {
 155         ret = nbd_negotiate_read(ioc, buffer, MIN(65536, size));
 156         if (ret < 0) {
 157             g_free(buffer);
 158             return ret;
 159         }
 160
 161         assert(ret <= size);
 162         size -= ret;
 163     }
 164
 165     g_free(buffer);
 166     return dropped;
 167 }
 168
 169 /* Basic flow for negotiation
 170
 171    Server         Client
 172    Negotiate
 173
 174    or
 175
 176    Server         Client
 177    Negotiate #1
 178                   Option
 179    Negotiate #2
 180
 181    ----
 182
 183    followed by
 184
 185    Server         Client
 186                   Request
 187    Response
 188                   Request
 189    Response
 190                   ...
 191    ...
 192                   Request (type == 2)
 193
 194 */
 195
 196 static int nbd_negotiate_send_rep(QIOChannel *ioc, uint32_t type, uint32_t opt)
 197 {
 198     uint64_t magic;
 199     uint32_t len;
 200
 201     TRACE("Reply opt=%" PRIx32 " type=%" PRIx32, type, opt);
 202
 203     magic = cpu_to_be64(NBD_REP_MAGIC);
 204     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 205         LOG("write failed (rep magic)");
 206         return -EINVAL;
 207     }
 208     opt = cpu_to_be32(opt);
 209     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 210         LOG("write failed (rep opt)");
 211         return -EINVAL;
 212     }
 213     type = cpu_to_be32(type);
 214     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 215         LOG("write failed (rep type)");
 216         return -EINVAL;
 217     }
 218     len = cpu_to_be32(0);
 219     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 220         LOG("write failed (rep data length)");
 221         return -EINVAL;
 222     }
 223     return 0;
 224 }
 225
 226 static int nbd_negotiate_send_rep_list(QIOChannel *ioc, NBDExport *exp)
 227 {
 228     uint64_t magic, name_len;
 229     uint32_t opt, type, len;
 230
 231     TRACE("Advertising export name '%s'", exp->name ? exp->name : "");
 232     name_len = strlen(exp->name);
 233     magic = cpu_to_be64(NBD_REP_MAGIC);
 234     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 235         LOG("write failed (magic)");
 236         return -EINVAL;
 237      }
 238     opt = cpu_to_be32(NBD_OPT_LIST);
 239     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 240         LOG("write failed (opt)");
 241         return -EINVAL;
 242     }
 243     type = cpu_to_be32(NBD_REP_SERVER);
 244     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 245         LOG("write failed (reply type)");
 246         return -EINVAL;
 247     }
 248     len = cpu_to_be32(name_len + sizeof(len));
 249     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 250         LOG("write failed (length)");
 251         return -EINVAL;
 252     }
 253     len = cpu_to_be32(name_len);
 254     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 255         LOG("write failed (length)");
 256         return -EINVAL;
 257     }
 258     if (nbd_negotiate_write(ioc, exp->name, name_len) != name_len) {
 259         LOG("write failed (buffer)");
 260         return -EINVAL;
 261     }
 262     return 0;
 263 }
 264
 265 static int nbd_negotiate_handle_list(NBDClient *client, uint32_t length)
 266 {
 267     NBDExport *exp;
 268
 269     if (length) {
 270         if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 271             return -EIO;
 272         }
 273         return nbd_negotiate_send_rep(client->ioc,
 274                                       NBD_REP_ERR_INVALID, NBD_OPT_LIST);
 275     }
 276
 277     /* For each export, send a NBD_REP_SERVER reply. */
 278     QTAILQ_FOREACH(exp, &exports, next) {
 279         if (nbd_negotiate_send_rep_list(client->ioc, exp)) {
 280             return -EINVAL;
 281         }
 282     }
 283     /* Finish with a NBD_REP_ACK. */
 284     return nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_LIST);
 285 }
 286
 287 static int nbd_negotiate_handle_export_name(NBDClient *client, uint32_t length)
 288 {
 289     int rc = -EINVAL;
 290     char name[NBD_MAX_NAME_SIZE + 1];
 291
 292     /* Client sends:
 293         [20 ..  xx]   export name (length bytes)
 294      */
 295     TRACE("Checking length");
 296     if (length >= sizeof(name)) {
 297         LOG("Bad length received");
 298         goto fail;
 299     }
 300     if (nbd_negotiate_read(client->ioc, name, length) != length) {
 301         LOG("read failed");
 302         goto fail;
 303     }
 304     name[length] = '\0';
 305
 306     TRACE("Client requested export '%s'", name);
 307
 308     client->exp = nbd_export_find(name);
 309     if (!client->exp) {
 310         LOG("export not found");
 311         goto fail;
 312     }
 313
 314     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 315     nbd_export_get(client->exp);
 316     rc = 0;
 317 fail:
 318     return rc;
 319 }
 320
 321
 322 static QIOChannel *nbd_negotiate_handle_starttls(NBDClient *client,
 323                                                  uint32_t length)
 324 {
 325     QIOChannel *ioc;
 326     QIOChannelTLS *tioc;
 327     struct NBDTLSHandshakeData data = { 0 };
 328
 329     TRACE("Setting up TLS");
 330     ioc = client->ioc;
 331     if (length) {
 332         if (nbd_negotiate_drop_sync(ioc, length) != length) {
 333             return NULL;
 334         }
 335         nbd_negotiate_send_rep(ioc, NBD_REP_ERR_INVALID, NBD_OPT_STARTTLS);
 336         return NULL;
 337     }
 338
 339     if (nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK,
 340                                NBD_OPT_STARTTLS) < 0) {
 341         return NULL;
 342     }
 343
 344     tioc = qio_channel_tls_new_server(ioc,
 345                                       client->tlscreds,
 346                                       client->tlsaclname,
 347                                       NULL);
 348     if (!tioc) {
 349         return NULL;
 350     }
 351
 352     qio_channel_set_name(QIO_CHANNEL(tioc), "nbd-server-tls");
 353     TRACE("Starting TLS handshake");
 354     data.loop = g_main_loop_new(g_main_context_default(), FALSE);
 355     qio_channel_tls_handshake(tioc,
 356                               nbd_tls_handshake,
 357                               &data,
 358                               NULL);
 359
 360     if (!data.complete) {
 361         g_main_loop_run(data.loop);
 362     }
 363     g_main_loop_unref(data.loop);
 364     if (data.error) {
 365         object_unref(OBJECT(tioc));
 366         error_free(data.error);
 367         return NULL;
 368     }
 369
 370     return QIO_CHANNEL(tioc);
 371 }
 372
 373
 374 static int nbd_negotiate_options(NBDClient *client)
 375 {
 376     uint32_t flags;
 377     bool fixedNewstyle = false;
 378
 379     /* Client sends:
 380         [ 0 ..   3]   client flags
 381
 382         [ 0 ..   7]   NBD_OPTS_MAGIC
 383         [ 8 ..  11]   NBD option
 384         [12 ..  15]   Data length
 385         ...           Rest of request
 386
 387         [ 0 ..   7]   NBD_OPTS_MAGIC
 388         [ 8 ..  11]   Second NBD option
 389         [12 ..  15]   Data length
 390         ...           Rest of request
 391     */
 392
 393     if (nbd_negotiate_read(client->ioc, &flags, sizeof(flags)) !=
 394         sizeof(flags)) {
 395         LOG("read failed");
 396         return -EIO;
 397     }
 398     TRACE("Checking client flags");
 399     be32_to_cpus(&flags);
 400     if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) {
 401         TRACE("Client supports fixed newstyle handshake");
 402         fixedNewstyle = true;
 403         flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE;
 404     }
 405     if (flags != 0) {
 406         TRACE("Unknown client flags 0x%" PRIx32 " received", flags);
 407         return -EIO;
 408     }
 409
 410     while (1) {
 411         int ret;
 412         uint32_t clientflags, length;
 413         uint64_t magic;
 414
 415         if (nbd_negotiate_read(client->ioc, &magic, sizeof(magic)) !=
 416             sizeof(magic)) {
 417             LOG("read failed");
 418             return -EINVAL;
 419         }
 420         TRACE("Checking opts magic");
 421         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 422             LOG("Bad magic received");
 423             return -EINVAL;
 424         }
 425
 426         if (nbd_negotiate_read(client->ioc, &clientflags,
 427                                sizeof(clientflags)) != sizeof(clientflags)) {
 428             LOG("read failed");
 429             return -EINVAL;
 430         }
 431         clientflags = be32_to_cpu(clientflags);
 432
 433         if (nbd_negotiate_read(client->ioc, &length, sizeof(length)) !=
 434             sizeof(length)) {
 435             LOG("read failed");
 436             return -EINVAL;
 437         }
 438         length = be32_to_cpu(length);
 439
 440         TRACE("Checking option 0x%" PRIx32, clientflags);
 441         if (client->tlscreds &&
 442             client->ioc == (QIOChannel *)client->sioc) {
 443             QIOChannel *tioc;
 444             if (!fixedNewstyle) {
 445                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 446                 return -EINVAL;
 447             }
 448             switch (clientflags) {
 449             case NBD_OPT_STARTTLS:
 450                 tioc = nbd_negotiate_handle_starttls(client, length);
 451                 if (!tioc) {
 452                     return -EIO;
 453                 }
 454                 object_unref(OBJECT(client->ioc));
 455                 client->ioc = QIO_CHANNEL(tioc);
 456                 break;
 457
 458             case NBD_OPT_EXPORT_NAME:
 459                 /* No way to return an error to client, so drop connection */
 460                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 461                 return -EINVAL;
 462
 463             default:
 464                 TRACE("Option 0x%" PRIx32 " not permitted before TLS",
 465                       clientflags);
 466                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 467                     return -EIO;
 468                 }
 469                 ret = nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_TLS_REQD,
 470                                              clientflags);
 471                 if (ret < 0) {
 472                     return ret;
 473                 }
 474                 break;
 475             }
 476         } else if (fixedNewstyle) {
 477             switch (clientflags) {
 478             case NBD_OPT_LIST:
 479                 ret = nbd_negotiate_handle_list(client, length);
 480                 if (ret < 0) {
 481                     return ret;
 482                 }
 483                 break;
 484
 485             case NBD_OPT_ABORT:
 486                 return -EINVAL;
 487
 488             case NBD_OPT_EXPORT_NAME:
 489                 return nbd_negotiate_handle_export_name(client, length);
 490
 491             case NBD_OPT_STARTTLS:
 492                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 493                     return -EIO;
 494                 }
 495                 if (client->tlscreds) {
 496                     TRACE("TLS already enabled");
 497                     ret = nbd_negotiate_send_rep(client->ioc,
 498                                                  NBD_REP_ERR_INVALID,
 499                                                  clientflags);
 500                 } else {
 501                     TRACE("TLS not configured");
 502                     ret = nbd_negotiate_send_rep(client->ioc,
 503                                                  NBD_REP_ERR_POLICY,
 504                                                  clientflags);
 505                 }
 506                 if (ret < 0) {
 507                     return ret;
 508                 }
 509                 break;
 510             default:
 511                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 512                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 513                     return -EIO;
 514                 }
 515                 ret = nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_UNSUP,
 516                                              clientflags);
 517                 if (ret < 0) {
 518                     return ret;
 519                 }
 520                 break;
 521             }
 522         } else {
 523             /*
 524              * If broken new-style we should drop the connection
 525              * for anything except NBD_OPT_EXPORT_NAME
 526              */
 527             switch (clientflags) {
 528             case NBD_OPT_EXPORT_NAME:
 529                 return nbd_negotiate_handle_export_name(client, length);
 530
 531             default:
 532                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 533                 return -EINVAL;
 534             }
 535         }
 536     }
 537 }
 538
 539 typedef struct {
 540     NBDClient *client;
 541     Coroutine *co;
 542 } NBDClientNewData;
 543
 544 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 545 {
 546     NBDClient *client = data->client;
 547     char buf[8 + 8 + 8 + 128];
 548     int rc;
 549     const uint16_t myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 550                               NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 551     bool oldStyle;
 552
 553     /* Old style negotiation header without options
 554         [ 0 ..   7]   passwd       ("NBDMAGIC")
 555         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 556         [16 ..  23]   size
 557         [24 ..  25]   server flags (0)
 558         [26 ..  27]   export flags
 559         [28 .. 151]   reserved     (0)
 560
 561        New style negotiation header with options
 562         [ 0 ..   7]   passwd       ("NBDMAGIC")
 563         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 564         [16 ..  17]   server flags (0)
 565         ....options sent....
 566         [18 ..  25]   size
 567         [26 ..  27]   export flags
 568         [28 .. 151]   reserved     (0)
 569      */
 570
 571     qio_channel_set_blocking(client->ioc, false, NULL);
 572     rc = -EINVAL;
 573
 574     TRACE("Beginning negotiation.");
 575     memset(buf, 0, sizeof(buf));
 576     memcpy(buf, "NBDMAGIC", 8);
 577
 578     oldStyle = client->exp != NULL && !client->tlscreds;
 579     if (oldStyle) {
 580         TRACE("advertising size %" PRIu64 " and flags %x",
 581               client->exp->size, client->exp->nbdflags | myflags);
 582         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 583         stq_be_p(buf + 16, client->exp->size);
 584         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 585     } else {
 586         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 587         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE);
 588     }
 589
 590     if (oldStyle) {
 591         if (client->tlscreds) {
 592             TRACE("TLS cannot be enabled with oldstyle protocol");
 593             goto fail;
 594         }
 595         if (nbd_negotiate_write(client->ioc, buf, sizeof(buf)) != sizeof(buf)) {
 596             LOG("write failed");
 597             goto fail;
 598         }
 599     } else {
 600         if (nbd_negotiate_write(client->ioc, buf, 18) != 18) {
 601             LOG("write failed");
 602             goto fail;
 603         }
 604         rc = nbd_negotiate_options(client);
 605         if (rc != 0) {
 606             LOG("option negotiation failed");
 607             goto fail;
 608         }
 609
 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 static ssize_t nbd_receive_request(QIOChannel *ioc, struct nbd_request *request)
 628 {
 629     uint8_t buf[NBD_REQUEST_SIZE];
 630     uint32_t magic;
 631     ssize_t ret;
 632
 633     ret = read_sync(ioc, buf, sizeof(buf));
 634     if (ret < 0) {
 635         return ret;
 636     }
 637
 638     if (ret != sizeof(buf)) {
 639         LOG("read failed");
 640         return -EINVAL;
 641     }
 642
 643     /* Request
 644        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 645        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 646        [ 8 .. 15]   handle
 647        [16 .. 23]   from
 648        [24 .. 27]   len
 649      */
 650
 651     magic = ldl_be_p(buf);
 652     request->type   = ldl_be_p(buf + 4);
 653     request->handle = ldq_be_p(buf + 8);
 654     request->from   = ldq_be_p(buf + 16);
 655     request->len    = ldl_be_p(buf + 24);
 656
 657     TRACE("Got request: { magic = 0x%" PRIx32 ", .type = %" PRIx32
 658           ", from = %" PRIu64 " , len = %" PRIu32 " }",
 659           magic, request->type, request->from, request->len);
 660
 661     if (magic != NBD_REQUEST_MAGIC) {
 662         LOG("invalid magic (got 0x%" PRIx32 ")", magic);
 663         return -EINVAL;
 664     }
 665     return 0;
 666 }
 667
 668 static ssize_t nbd_send_reply(QIOChannel *ioc, struct nbd_reply *reply)
 669 {
 670     uint8_t buf[NBD_REPLY_SIZE];
 671     ssize_t ret;
 672
 673     reply->error = system_errno_to_nbd_errno(reply->error);
 674
 675     TRACE("Sending response to client: { .error = %" PRId32
 676           ", handle = %" PRIu64 " }",
 677           reply->error, reply->handle);
 678
 679     /* Reply
 680        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 681        [ 4 ..  7]    error   (0 == no error)
 682        [ 7 .. 15]    handle
 683      */
 684     stl_be_p(buf, NBD_REPLY_MAGIC);
 685     stl_be_p(buf + 4, reply->error);
 686     stq_be_p(buf + 8, reply->handle);
 687
 688     ret = write_sync(ioc, buf, sizeof(buf));
 689     if (ret < 0) {
 690         return ret;
 691     }
 692
 693     if (ret != sizeof(buf)) {
 694         LOG("writing to socket failed");
 695         return -EINVAL;
 696     }
 697     return 0;
 698 }
 699
 700 #define MAX_NBD_REQUESTS 16
 701
 702 void nbd_client_get(NBDClient *client)
 703 {
 704     client->refcount++;
 705 }
 706
 707 void nbd_client_put(NBDClient *client)
 708 {
 709     if (--client->refcount == 0) {
 710         /* The last reference should be dropped by client->close,
 711          * which is called by client_close.
 712          */
 713         assert(client->closing);
 714
 715         nbd_unset_handlers(client);
 716         object_unref(OBJECT(client->sioc));
 717         object_unref(OBJECT(client->ioc));
 718         if (client->tlscreds) {
 719             object_unref(OBJECT(client->tlscreds));
 720         }
 721         g_free(client->tlsaclname);
 722         if (client->exp) {
 723             QTAILQ_REMOVE(&client->exp->clients, client, next);
 724             nbd_export_put(client->exp);
 725         }
 726         g_free(client);
 727     }
 728 }
 729
 730 static void client_close(NBDClient *client)
 731 {
 732     if (client->closing) {
 733         return;
 734     }
 735
 736     client->closing = true;
 737
 738     /* Force requests to finish.  They will drop their own references,
 739      * then we'll close the socket and free the NBDClient.
 740      */
 741     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 742                          NULL);
 743
 744     /* Also tell the client, so that they release their reference.  */
 745     if (client->close) {
 746         client->close(client);
 747     }
 748 }
 749
 750 static NBDRequest *nbd_request_get(NBDClient *client)
 751 {
 752     NBDRequest *req;
 753
 754     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 755     client->nb_requests++;
 756     nbd_update_can_read(client);
 757
 758     req = g_new0(NBDRequest, 1);
 759     nbd_client_get(client);
 760     req->client = client;
 761     return req;
 762 }
 763
 764 static void nbd_request_put(NBDRequest *req)
 765 {
 766     NBDClient *client = req->client;
 767
 768     if (req->data) {
 769         qemu_vfree(req->data);
 770     }
 771     g_free(req);
 772
 773     client->nb_requests--;
 774     nbd_update_can_read(client);
 775     nbd_client_put(client);
 776 }
 777
 778 static void blk_aio_attached(AioContext *ctx, void *opaque)
 779 {
 780     NBDExport *exp = opaque;
 781     NBDClient *client;
 782
 783     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 784
 785     exp->ctx = ctx;
 786
 787     QTAILQ_FOREACH(client, &exp->clients, next) {
 788         nbd_set_handlers(client);
 789     }
 790 }
 791
 792 static void blk_aio_detach(void *opaque)
 793 {
 794     NBDExport *exp = opaque;
 795     NBDClient *client;
 796
 797     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 798
 799     QTAILQ_FOREACH(client, &exp->clients, next) {
 800         nbd_unset_handlers(client);
 801     }
 802
 803     exp->ctx = NULL;
 804 }
 805
 806 static void nbd_eject_notifier(Notifier *n, void *data)
 807 {
 808     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 809     nbd_export_close(exp);
 810 }
 811
 812 NBDExport *nbd_export_new(BlockDriverState *bs, off_t dev_offset, off_t size,
 813                           uint16_t nbdflags, void (*close)(NBDExport *),
 814                           bool writethrough, BlockBackend *on_eject_blk,
 815                           Error **errp)
 816 {
 817     BlockBackend *blk;
 818     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 819
 820     blk = blk_new();
 821     blk_insert_bs(blk, bs);
 822     blk_set_enable_write_cache(blk, !writethrough);
 823
 824     exp->refcount = 1;
 825     QTAILQ_INIT(&exp->clients);
 826     exp->blk = blk;
 827     exp->dev_offset = dev_offset;
 828     exp->nbdflags = nbdflags;
 829     exp->size = size < 0 ? blk_getlength(blk) : size;
 830     if (exp->size < 0) {
 831         error_setg_errno(errp, -exp->size,
 832                          "Failed to determine the NBD export's length");
 833         goto fail;
 834     }
 835     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 836
 837     exp->close = close;
 838     exp->ctx = blk_get_aio_context(blk);
 839     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 840
 841     if (on_eject_blk) {
 842         blk_ref(on_eject_blk);
 843         exp->eject_notifier_blk = on_eject_blk;
 844         exp->eject_notifier.notify = nbd_eject_notifier;
 845         blk_add_remove_bs_notifier(on_eject_blk, &exp->eject_notifier);
 846     }
 847
 848     /*
 849      * NBD exports are used for non-shared storage migration.  Make sure
 850      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 851      * access since the export could be available before migration handover.
 852      */
 853     aio_context_acquire(exp->ctx);
 854     blk_invalidate_cache(blk, NULL);
 855     aio_context_release(exp->ctx);
 856     return exp;
 857
 858 fail:
 859     blk_unref(blk);
 860     g_free(exp);
 861     return NULL;
 862 }
 863
 864 NBDExport *nbd_export_find(const char *name)
 865 {
 866     NBDExport *exp;
 867     QTAILQ_FOREACH(exp, &exports, next) {
 868         if (strcmp(name, exp->name) == 0) {
 869             return exp;
 870         }
 871     }
 872
 873     return NULL;
 874 }
 875
 876 void nbd_export_set_name(NBDExport *exp, const char *name)
 877 {
 878     if (exp->name == name) {
 879         return;
 880     }
 881
 882     nbd_export_get(exp);
 883     if (exp->name != NULL) {
 884         g_free(exp->name);
 885         exp->name = NULL;
 886         QTAILQ_REMOVE(&exports, exp, next);
 887         nbd_export_put(exp);
 888     }
 889     if (name != NULL) {
 890         nbd_export_get(exp);
 891         exp->name = g_strdup(name);
 892         QTAILQ_INSERT_TAIL(&exports, exp, next);
 893     }
 894     nbd_export_put(exp);
 895 }
 896
 897 void nbd_export_close(NBDExport *exp)
 898 {
 899     NBDClient *client, *next;
 900
 901     nbd_export_get(exp);
 902     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 903         client_close(client);
 904     }
 905     nbd_export_set_name(exp, NULL);
 906     nbd_export_put(exp);
 907 }
 908
 909 void nbd_export_get(NBDExport *exp)
 910 {
 911     assert(exp->refcount > 0);
 912     exp->refcount++;
 913 }
 914
 915 void nbd_export_put(NBDExport *exp)
 916 {
 917     assert(exp->refcount > 0);
 918     if (exp->refcount == 1) {
 919         nbd_export_close(exp);
 920     }
 921
 922     if (--exp->refcount == 0) {
 923         assert(exp->name == NULL);
 924
 925         if (exp->close) {
 926             exp->close(exp);
 927         }
 928
 929         if (exp->blk) {
 930             if (exp->eject_notifier_blk) {
 931                 notifier_remove(&exp->eject_notifier);
 932                 blk_unref(exp->eject_notifier_blk);
 933             }
 934             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 935                                             blk_aio_detach, exp);
 936             blk_unref(exp->blk);
 937             exp->blk = NULL;
 938         }
 939
 940         g_free(exp);
 941     }
 942 }
 943
 944 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
 945 {
 946     return exp->blk;
 947 }
 948
 949 void nbd_export_close_all(void)
 950 {
 951     NBDExport *exp, *next;
 952
 953     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 954         nbd_export_close(exp);
 955     }
 956 }
 957
 958 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 959                                  int len)
 960 {
 961     NBDClient *client = req->client;
 962     ssize_t rc, ret;
 963
 964     g_assert(qemu_in_coroutine());
 965     qemu_co_mutex_lock(&client->send_lock);
 966     client->send_coroutine = qemu_coroutine_self();
 967     nbd_set_handlers(client);
 968
 969     if (!len) {
 970         rc = nbd_send_reply(client->ioc, reply);
 971     } else {
 972         qio_channel_set_cork(client->ioc, true);
 973         rc = nbd_send_reply(client->ioc, reply);
 974         if (rc >= 0) {
 975             ret = write_sync(client->ioc, req->data, len);
 976             if (ret != len) {
 977                 rc = -EIO;
 978             }
 979         }
 980         qio_channel_set_cork(client->ioc, false);
 981     }
 982
 983     client->send_coroutine = NULL;
 984     nbd_set_handlers(client);
 985     qemu_co_mutex_unlock(&client->send_lock);
 986     return rc;
 987 }
 988
 989 /* Collect a client request.  Return 0 if request looks valid, -EAGAIN
 990  * to keep trying the collection, -EIO to drop connection right away,
 991  * and any other negative value to report an error to the client
 992  * (although the caller may still need to disconnect after reporting
 993  * the error).  */
 994 static ssize_t nbd_co_receive_request(NBDRequest *req,
 995                                       struct nbd_request *request)
 996 {
 997     NBDClient *client = req->client;
 998     uint32_t command;
 999     ssize_t rc;
1000
1001     g_assert(qemu_in_coroutine());
1002     client->recv_coroutine = qemu_coroutine_self();
1003     nbd_update_can_read(client);
1004
1005     rc = nbd_receive_request(client->ioc, request);
1006     if (rc < 0) {
1007         if (rc != -EAGAIN) {
1008             rc = -EIO;
1009         }
1010         goto out;
1011     }
1012
1013     TRACE("Decoding type");
1014
1015     command = request->type & NBD_CMD_MASK_COMMAND;
1016     if (command != NBD_CMD_WRITE) {
1017         /* No payload, we are ready to read the next request.  */
1018         req->complete = true;
1019     }
1020
1021     if (command == NBD_CMD_DISC) {
1022         /* Special case: we're going to disconnect without a reply,
1023          * whether or not flags, from, or len are bogus */
1024         TRACE("Request type is DISCONNECT");
1025         rc = -EIO;
1026         goto out;
1027     }
1028
1029     /* Check for sanity in the parameters, part 1.  Defer as many
1030      * checks as possible until after reading any NBD_CMD_WRITE
1031      * payload, so we can try and keep the connection alive.  */
1032     if ((request->from + request->len) < request->from) {
1033         LOG("integer overflow detected, you're probably being attacked");
1034         rc = -EINVAL;
1035         goto out;
1036     }
1037
1038     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
1039         if (request->len > NBD_MAX_BUFFER_SIZE) {
1040             LOG("len (%" PRIu32" ) is larger than max len (%u)",
1041                 request->len, NBD_MAX_BUFFER_SIZE);
1042             rc = -EINVAL;
1043             goto out;
1044         }
1045
1046         req->data = blk_try_blockalign(client->exp->blk, request->len);
1047         if (req->data == NULL) {
1048             rc = -ENOMEM;
1049             goto out;
1050         }
1051     }
1052     if (command == NBD_CMD_WRITE) {
1053         TRACE("Reading %" PRIu32 " byte(s)", request->len);
1054
1055         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1056             LOG("reading from socket failed");
1057             rc = -EIO;
1058             goto out;
1059         }
1060         req->complete = true;
1061     }
1062
1063     /* Sanity checks, part 2. */
1064     if (request->from + request->len > client->exp->size) {
1065         LOG("operation past EOF; From: %" PRIu64 ", Len: %" PRIu32
1066             ", Size: %" PRIu64, request->from, request->len,
1067             (uint64_t)client->exp->size);
1068         rc = command == NBD_CMD_WRITE ? -ENOSPC : -EINVAL;
1069         goto out;
1070     }
1071     if (request->type & ~NBD_CMD_MASK_COMMAND & ~NBD_CMD_FLAG_FUA) {
1072         LOG("unsupported flags (got 0x%x)",
1073             request->type & ~NBD_CMD_MASK_COMMAND);
1074         rc = -EINVAL;
1075         goto out;
1076     }
1077
1078     rc = 0;
1079
1080 out:
1081     client->recv_coroutine = NULL;
1082     nbd_update_can_read(client);
1083
1084     return rc;
1085 }
1086
1087 static void nbd_trip(void *opaque)
1088 {
1089     NBDClient *client = opaque;
1090     NBDExport *exp = client->exp;
1091     NBDRequest *req;
1092     struct nbd_request request;
1093     struct nbd_reply reply;
1094     ssize_t ret;
1095     uint32_t command;
1096     int flags;
1097
1098     TRACE("Reading request.");
1099     if (client->closing) {
1100         return;
1101     }
1102
1103     req = nbd_request_get(client);
1104     ret = nbd_co_receive_request(req, &request);
1105     if (ret == -EAGAIN) {
1106         goto done;
1107     }
1108     if (ret == -EIO) {
1109         goto out;
1110     }
1111
1112     reply.handle = request.handle;
1113     reply.error = 0;
1114
1115     if (ret < 0) {
1116         reply.error = -ret;
1117         goto error_reply;
1118     }
1119     command = request.type & NBD_CMD_MASK_COMMAND;
1120
1121     if (client->closing) {
1122         /*
1123          * The client may be closed when we are blocked in
1124          * nbd_co_receive_request()
1125          */
1126         goto done;
1127     }
1128
1129     switch (command) {
1130     case NBD_CMD_READ:
1131         TRACE("Request type is READ");
1132
1133         if (request.type & NBD_CMD_FLAG_FUA) {
1134             ret = blk_co_flush(exp->blk);
1135             if (ret < 0) {
1136                 LOG("flush failed");
1137                 reply.error = -ret;
1138                 goto error_reply;
1139             }
1140         }
1141
1142         ret = blk_pread(exp->blk, request.from + exp->dev_offset,
1143                         req->data, request.len);
1144         if (ret < 0) {
1145             LOG("reading from file failed");
1146             reply.error = -ret;
1147             goto error_reply;
1148         }
1149
1150         TRACE("Read %" PRIu32" byte(s)", request.len);
1151         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1152             goto out;
1153         break;
1154     case NBD_CMD_WRITE:
1155         TRACE("Request type is WRITE");
1156
1157         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1158             TRACE("Server is read-only, return error");
1159             reply.error = EROFS;
1160             goto error_reply;
1161         }
1162
1163         TRACE("Writing to device");
1164
1165         flags = 0;
1166         if (request.type & NBD_CMD_FLAG_FUA) {
1167             flags |= BDRV_REQ_FUA;
1168         }
1169         ret = blk_pwrite(exp->blk, request.from + exp->dev_offset,
1170                          req->data, request.len, flags);
1171         if (ret < 0) {
1172             LOG("writing to file failed");
1173             reply.error = -ret;
1174             goto error_reply;
1175         }
1176
1177         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1178             goto out;
1179         }
1180         break;
1181
1182     case NBD_CMD_DISC:
1183         /* unreachable, thanks to special case in nbd_co_receive_request() */
1184         abort();
1185
1186     case NBD_CMD_FLUSH:
1187         TRACE("Request type is FLUSH");
1188
1189         ret = blk_co_flush(exp->blk);
1190         if (ret < 0) {
1191             LOG("flush failed");
1192             reply.error = -ret;
1193         }
1194         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1195             goto out;
1196         }
1197         break;
1198     case NBD_CMD_TRIM:
1199         TRACE("Request type is TRIM");
1200         ret = blk_co_pdiscard(exp->blk, request.from + exp->dev_offset,
1201                               request.len);
1202         if (ret < 0) {
1203             LOG("discard failed");
1204             reply.error = -ret;
1205         }
1206         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1207             goto out;
1208         }
1209         break;
1210     default:
1211         LOG("invalid request type (%" PRIu32 ") received", request.type);
1212         reply.error = EINVAL;
1213     error_reply:
1214         /* We must disconnect after NBD_CMD_WRITE if we did not
1215          * read the payload.
1216          */
1217         if (nbd_co_send_reply(req, &reply, 0) < 0 || !req->complete) {
1218             goto out;
1219         }
1220         break;
1221     }
1222
1223     TRACE("Request/Reply complete");
1224
1225 done:
1226     nbd_request_put(req);
1227     return;
1228
1229 out:
1230     nbd_request_put(req);
1231     client_close(client);
1232 }
1233
1234 static void nbd_read(void *opaque)
1235 {
1236     NBDClient *client = opaque;
1237
1238     if (client->recv_coroutine) {
1239         qemu_coroutine_enter(client->recv_coroutine);
1240     } else {
1241         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip, client));
1242     }
1243 }
1244
1245 static void nbd_restart_write(void *opaque)
1246 {
1247     NBDClient *client = opaque;
1248
1249     qemu_coroutine_enter(client->send_coroutine);
1250 }
1251
1252 static void nbd_set_handlers(NBDClient *client)
1253 {
1254     if (client->exp && client->exp->ctx) {
1255         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1256                            true,
1257                            client->can_read ? nbd_read : NULL,
1258                            client->send_coroutine ? nbd_restart_write : NULL,
1259                            client);
1260     }
1261 }
1262
1263 static void nbd_unset_handlers(NBDClient *client)
1264 {
1265     if (client->exp && client->exp->ctx) {
1266         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1267                            true, NULL, NULL, NULL);
1268     }
1269 }
1270
1271 static void nbd_update_can_read(NBDClient *client)
1272 {
1273     bool can_read = client->recv_coroutine ||
1274                     client->nb_requests < MAX_NBD_REQUESTS;
1275
1276     if (can_read != client->can_read) {
1277         client->can_read = can_read;
1278         nbd_set_handlers(client);
1279
1280         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1281          * in nbd_set_handlers() will have taken care of that */
1282     }
1283 }
1284
1285 static coroutine_fn void nbd_co_client_start(void *opaque)
1286 {
1287     NBDClientNewData *data = opaque;
1288     NBDClient *client = data->client;
1289     NBDExport *exp = client->exp;
1290
1291     if (exp) {
1292         nbd_export_get(exp);
1293     }
1294     if (nbd_negotiate(data)) {
1295         client_close(client);
1296         goto out;
1297     }
1298     qemu_co_mutex_init(&client->send_lock);
1299     nbd_set_handlers(client);
1300
1301     if (exp) {
1302         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1303     }
1304 out:
1305     g_free(data);
1306 }
1307
1308 void nbd_client_new(NBDExport *exp,
1309                     QIOChannelSocket *sioc,
1310                     QCryptoTLSCreds *tlscreds,
1311                     const char *tlsaclname,
1312                     void (*close_fn)(NBDClient *))
1313 {
1314     NBDClient *client;
1315     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1316
1317     client = g_malloc0(sizeof(NBDClient));
1318     client->refcount = 1;
1319     client->exp = exp;
1320     client->tlscreds = tlscreds;
1321     if (tlscreds) {
1322         object_ref(OBJECT(client->tlscreds));
1323     }
1324     client->tlsaclname = g_strdup(tlsaclname);
1325     client->sioc = sioc;
1326     object_ref(OBJECT(client->sioc));
1327     client->ioc = QIO_CHANNEL(sioc);
1328     object_ref(OBJECT(client->ioc));
1329     client->can_read = true;
1330     client->close = close_fn;
1331
1332     data->client = client;
1333     data->co = qemu_coroutine_create(nbd_co_client_start, data);
1334     qemu_coroutine_enter(data->co);
1335 }