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     TRACE("Starting TLS handshake");
 353     data.loop = g_main_loop_new(g_main_context_default(), FALSE);
 354     qio_channel_tls_handshake(tioc,
 355                               nbd_tls_handshake,
 356                               &data,
 357                               NULL);
 358
 359     if (!data.complete) {
 360         g_main_loop_run(data.loop);
 361     }
 362     g_main_loop_unref(data.loop);
 363     if (data.error) {
 364         object_unref(OBJECT(tioc));
 365         error_free(data.error);
 366         return NULL;
 367     }
 368
 369     return QIO_CHANNEL(tioc);
 370 }
 371
 372
 373 static int nbd_negotiate_options(NBDClient *client)
 374 {
 375     uint32_t flags;
 376     bool fixedNewstyle = false;
 377
 378     /* Client sends:
 379         [ 0 ..   3]   client flags
 380
 381         [ 0 ..   7]   NBD_OPTS_MAGIC
 382         [ 8 ..  11]   NBD option
 383         [12 ..  15]   Data length
 384         ...           Rest of request
 385
 386         [ 0 ..   7]   NBD_OPTS_MAGIC
 387         [ 8 ..  11]   Second NBD option
 388         [12 ..  15]   Data length
 389         ...           Rest of request
 390     */
 391
 392     if (nbd_negotiate_read(client->ioc, &flags, sizeof(flags)) !=
 393         sizeof(flags)) {
 394         LOG("read failed");
 395         return -EIO;
 396     }
 397     TRACE("Checking client flags");
 398     be32_to_cpus(&flags);
 399     if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) {
 400         TRACE("Client supports fixed newstyle handshake");
 401         fixedNewstyle = true;
 402         flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE;
 403     }
 404     if (flags != 0) {
 405         TRACE("Unknown client flags 0x%" PRIx32 " received", flags);
 406         return -EIO;
 407     }
 408
 409     while (1) {
 410         int ret;
 411         uint32_t clientflags, length;
 412         uint64_t magic;
 413
 414         if (nbd_negotiate_read(client->ioc, &magic, sizeof(magic)) !=
 415             sizeof(magic)) {
 416             LOG("read failed");
 417             return -EINVAL;
 418         }
 419         TRACE("Checking opts magic");
 420         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 421             LOG("Bad magic received");
 422             return -EINVAL;
 423         }
 424
 425         if (nbd_negotiate_read(client->ioc, &clientflags,
 426                                sizeof(clientflags)) != sizeof(clientflags)) {
 427             LOG("read failed");
 428             return -EINVAL;
 429         }
 430         clientflags = be32_to_cpu(clientflags);
 431
 432         if (nbd_negotiate_read(client->ioc, &length, sizeof(length)) !=
 433             sizeof(length)) {
 434             LOG("read failed");
 435             return -EINVAL;
 436         }
 437         length = be32_to_cpu(length);
 438
 439         TRACE("Checking option 0x%" PRIx32, clientflags);
 440         if (client->tlscreds &&
 441             client->ioc == (QIOChannel *)client->sioc) {
 442             QIOChannel *tioc;
 443             if (!fixedNewstyle) {
 444                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 445                 return -EINVAL;
 446             }
 447             switch (clientflags) {
 448             case NBD_OPT_STARTTLS:
 449                 tioc = nbd_negotiate_handle_starttls(client, length);
 450                 if (!tioc) {
 451                     return -EIO;
 452                 }
 453                 object_unref(OBJECT(client->ioc));
 454                 client->ioc = QIO_CHANNEL(tioc);
 455                 break;
 456
 457             case NBD_OPT_EXPORT_NAME:
 458                 /* No way to return an error to client, so drop connection */
 459                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 460                 return -EINVAL;
 461
 462             default:
 463                 TRACE("Option 0x%" PRIx32 " not permitted before TLS",
 464                       clientflags);
 465                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 466                     return -EIO;
 467                 }
 468                 ret = nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_TLS_REQD,
 469                                              clientflags);
 470                 if (ret < 0) {
 471                     return ret;
 472                 }
 473                 break;
 474             }
 475         } else if (fixedNewstyle) {
 476             switch (clientflags) {
 477             case NBD_OPT_LIST:
 478                 ret = nbd_negotiate_handle_list(client, length);
 479                 if (ret < 0) {
 480                     return ret;
 481                 }
 482                 break;
 483
 484             case NBD_OPT_ABORT:
 485                 return -EINVAL;
 486
 487             case NBD_OPT_EXPORT_NAME:
 488                 return nbd_negotiate_handle_export_name(client, length);
 489
 490             case NBD_OPT_STARTTLS:
 491                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 492                     return -EIO;
 493                 }
 494                 if (client->tlscreds) {
 495                     TRACE("TLS already enabled");
 496                     ret = nbd_negotiate_send_rep(client->ioc,
 497                                                  NBD_REP_ERR_INVALID,
 498                                                  clientflags);
 499                 } else {
 500                     TRACE("TLS not configured");
 501                     ret = nbd_negotiate_send_rep(client->ioc,
 502                                                  NBD_REP_ERR_POLICY,
 503                                                  clientflags);
 504                 }
 505                 if (ret < 0) {
 506                     return ret;
 507                 }
 508                 break;
 509             default:
 510                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 511                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 512                     return -EIO;
 513                 }
 514                 ret = nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_UNSUP,
 515                                              clientflags);
 516                 if (ret < 0) {
 517                     return ret;
 518                 }
 519                 break;
 520             }
 521         } else {
 522             /*
 523              * If broken new-style we should drop the connection
 524              * for anything except NBD_OPT_EXPORT_NAME
 525              */
 526             switch (clientflags) {
 527             case NBD_OPT_EXPORT_NAME:
 528                 return nbd_negotiate_handle_export_name(client, length);
 529
 530             default:
 531                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 532                 return -EINVAL;
 533             }
 534         }
 535     }
 536 }
 537
 538 typedef struct {
 539     NBDClient *client;
 540     Coroutine *co;
 541 } NBDClientNewData;
 542
 543 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 544 {
 545     NBDClient *client = data->client;
 546     char buf[8 + 8 + 8 + 128];
 547     int rc;
 548     const uint16_t myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 549                               NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 550     bool oldStyle;
 551
 552     /* Old style negotiation header without options
 553         [ 0 ..   7]   passwd       ("NBDMAGIC")
 554         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 555         [16 ..  23]   size
 556         [24 ..  25]   server flags (0)
 557         [26 ..  27]   export flags
 558         [28 .. 151]   reserved     (0)
 559
 560        New style negotiation header with options
 561         [ 0 ..   7]   passwd       ("NBDMAGIC")
 562         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 563         [16 ..  17]   server flags (0)
 564         ....options sent....
 565         [18 ..  25]   size
 566         [26 ..  27]   export flags
 567         [28 .. 151]   reserved     (0)
 568      */
 569
 570     qio_channel_set_blocking(client->ioc, false, NULL);
 571     rc = -EINVAL;
 572
 573     TRACE("Beginning negotiation.");
 574     memset(buf, 0, sizeof(buf));
 575     memcpy(buf, "NBDMAGIC", 8);
 576
 577     oldStyle = client->exp != NULL && !client->tlscreds;
 578     if (oldStyle) {
 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         TRACE("advertising size %" PRIu64 " and flags %x",
 610               client->exp->size, client->exp->nbdflags | myflags);
 611         stq_be_p(buf + 18, client->exp->size);
 612         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 613         if (nbd_negotiate_write(client->ioc, buf + 18, sizeof(buf) - 18) !=
 614             sizeof(buf) - 18) {
 615             LOG("write failed");
 616             goto fail;
 617         }
 618     }
 619
 620     TRACE("Negotiation succeeded.");
 621     rc = 0;
 622 fail:
 623     return rc;
 624 }
 625
 626 static ssize_t nbd_receive_request(QIOChannel *ioc, struct nbd_request *request)
 627 {
 628     uint8_t buf[NBD_REQUEST_SIZE];
 629     uint32_t magic;
 630     ssize_t ret;
 631
 632     ret = read_sync(ioc, buf, sizeof(buf));
 633     if (ret < 0) {
 634         return ret;
 635     }
 636
 637     if (ret != sizeof(buf)) {
 638         LOG("read failed");
 639         return -EINVAL;
 640     }
 641
 642     /* Request
 643        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 644        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 645        [ 8 .. 15]   handle
 646        [16 .. 23]   from
 647        [24 .. 27]   len
 648      */
 649
 650     magic = ldl_be_p(buf);
 651     request->type   = ldl_be_p(buf + 4);
 652     request->handle = ldq_be_p(buf + 8);
 653     request->from   = ldq_be_p(buf + 16);
 654     request->len    = ldl_be_p(buf + 24);
 655
 656     TRACE("Got request: { magic = 0x%" PRIx32 ", .type = %" PRIx32
 657           ", from = %" PRIu64 " , len = %" PRIu32 " }",
 658           magic, request->type, request->from, request->len);
 659
 660     if (magic != NBD_REQUEST_MAGIC) {
 661         LOG("invalid magic (got 0x%" PRIx32 ")", magic);
 662         return -EINVAL;
 663     }
 664     return 0;
 665 }
 666
 667 static ssize_t nbd_send_reply(QIOChannel *ioc, struct nbd_reply *reply)
 668 {
 669     uint8_t buf[NBD_REPLY_SIZE];
 670     ssize_t ret;
 671
 672     reply->error = system_errno_to_nbd_errno(reply->error);
 673
 674     TRACE("Sending response to client: { .error = %" PRId32
 675           ", handle = %" PRIu64 " }",
 676           reply->error, reply->handle);
 677
 678     /* Reply
 679        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 680        [ 4 ..  7]    error   (0 == no error)
 681        [ 7 .. 15]    handle
 682      */
 683     stl_be_p(buf, NBD_REPLY_MAGIC);
 684     stl_be_p(buf + 4, reply->error);
 685     stq_be_p(buf + 8, reply->handle);
 686
 687     ret = write_sync(ioc, buf, sizeof(buf));
 688     if (ret < 0) {
 689         return ret;
 690     }
 691
 692     if (ret != sizeof(buf)) {
 693         LOG("writing to socket failed");
 694         return -EINVAL;
 695     }
 696     return 0;
 697 }
 698
 699 #define MAX_NBD_REQUESTS 16
 700
 701 void nbd_client_get(NBDClient *client)
 702 {
 703     client->refcount++;
 704 }
 705
 706 void nbd_client_put(NBDClient *client)
 707 {
 708     if (--client->refcount == 0) {
 709         /* The last reference should be dropped by client->close,
 710          * which is called by client_close.
 711          */
 712         assert(client->closing);
 713
 714         nbd_unset_handlers(client);
 715         object_unref(OBJECT(client->sioc));
 716         object_unref(OBJECT(client->ioc));
 717         if (client->tlscreds) {
 718             object_unref(OBJECT(client->tlscreds));
 719         }
 720         g_free(client->tlsaclname);
 721         if (client->exp) {
 722             QTAILQ_REMOVE(&client->exp->clients, client, next);
 723             nbd_export_put(client->exp);
 724         }
 725         g_free(client);
 726     }
 727 }
 728
 729 static void client_close(NBDClient *client)
 730 {
 731     if (client->closing) {
 732         return;
 733     }
 734
 735     client->closing = true;
 736
 737     /* Force requests to finish.  They will drop their own references,
 738      * then we'll close the socket and free the NBDClient.
 739      */
 740     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 741                          NULL);
 742
 743     /* Also tell the client, so that they release their reference.  */
 744     if (client->close) {
 745         client->close(client);
 746     }
 747 }
 748
 749 static NBDRequest *nbd_request_get(NBDClient *client)
 750 {
 751     NBDRequest *req;
 752
 753     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 754     client->nb_requests++;
 755     nbd_update_can_read(client);
 756
 757     req = g_new0(NBDRequest, 1);
 758     nbd_client_get(client);
 759     req->client = client;
 760     return req;
 761 }
 762
 763 static void nbd_request_put(NBDRequest *req)
 764 {
 765     NBDClient *client = req->client;
 766
 767     if (req->data) {
 768         qemu_vfree(req->data);
 769     }
 770     g_free(req);
 771
 772     client->nb_requests--;
 773     nbd_update_can_read(client);
 774     nbd_client_put(client);
 775 }
 776
 777 static void blk_aio_attached(AioContext *ctx, void *opaque)
 778 {
 779     NBDExport *exp = opaque;
 780     NBDClient *client;
 781
 782     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 783
 784     exp->ctx = ctx;
 785
 786     QTAILQ_FOREACH(client, &exp->clients, next) {
 787         nbd_set_handlers(client);
 788     }
 789 }
 790
 791 static void blk_aio_detach(void *opaque)
 792 {
 793     NBDExport *exp = opaque;
 794     NBDClient *client;
 795
 796     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 797
 798     QTAILQ_FOREACH(client, &exp->clients, next) {
 799         nbd_unset_handlers(client);
 800     }
 801
 802     exp->ctx = NULL;
 803 }
 804
 805 static void nbd_eject_notifier(Notifier *n, void *data)
 806 {
 807     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 808     nbd_export_close(exp);
 809 }
 810
 811 NBDExport *nbd_export_new(BlockDriverState *bs, off_t dev_offset, off_t size,
 812                           uint16_t nbdflags, void (*close)(NBDExport *),
 813                           bool writethrough, BlockBackend *on_eject_blk,
 814                           Error **errp)
 815 {
 816     BlockBackend *blk;
 817     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 818
 819     blk = blk_new();
 820     blk_insert_bs(blk, bs);
 821     blk_set_enable_write_cache(blk, !writethrough);
 822
 823     exp->refcount = 1;
 824     QTAILQ_INIT(&exp->clients);
 825     exp->blk = blk;
 826     exp->dev_offset = dev_offset;
 827     exp->nbdflags = nbdflags;
 828     exp->size = size < 0 ? blk_getlength(blk) : size;
 829     if (exp->size < 0) {
 830         error_setg_errno(errp, -exp->size,
 831                          "Failed to determine the NBD export's length");
 832         goto fail;
 833     }
 834     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 835
 836     exp->close = close;
 837     exp->ctx = blk_get_aio_context(blk);
 838     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 839
 840     if (on_eject_blk) {
 841         blk_ref(on_eject_blk);
 842         exp->eject_notifier_blk = on_eject_blk;
 843         exp->eject_notifier.notify = nbd_eject_notifier;
 844         blk_add_remove_bs_notifier(on_eject_blk, &exp->eject_notifier);
 845     }
 846
 847     /*
 848      * NBD exports are used for non-shared storage migration.  Make sure
 849      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 850      * access since the export could be available before migration handover.
 851      */
 852     aio_context_acquire(exp->ctx);
 853     blk_invalidate_cache(blk, NULL);
 854     aio_context_release(exp->ctx);
 855     return exp;
 856
 857 fail:
 858     blk_unref(blk);
 859     g_free(exp);
 860     return NULL;
 861 }
 862
 863 NBDExport *nbd_export_find(const char *name)
 864 {
 865     NBDExport *exp;
 866     QTAILQ_FOREACH(exp, &exports, next) {
 867         if (strcmp(name, exp->name) == 0) {
 868             return exp;
 869         }
 870     }
 871
 872     return NULL;
 873 }
 874
 875 void nbd_export_set_name(NBDExport *exp, const char *name)
 876 {
 877     if (exp->name == name) {
 878         return;
 879     }
 880
 881     nbd_export_get(exp);
 882     if (exp->name != NULL) {
 883         g_free(exp->name);
 884         exp->name = NULL;
 885         QTAILQ_REMOVE(&exports, exp, next);
 886         nbd_export_put(exp);
 887     }
 888     if (name != NULL) {
 889         nbd_export_get(exp);
 890         exp->name = g_strdup(name);
 891         QTAILQ_INSERT_TAIL(&exports, exp, next);
 892     }
 893     nbd_export_put(exp);
 894 }
 895
 896 void nbd_export_close(NBDExport *exp)
 897 {
 898     NBDClient *client, *next;
 899
 900     nbd_export_get(exp);
 901     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 902         client_close(client);
 903     }
 904     nbd_export_set_name(exp, NULL);
 905     nbd_export_put(exp);
 906 }
 907
 908 void nbd_export_get(NBDExport *exp)
 909 {
 910     assert(exp->refcount > 0);
 911     exp->refcount++;
 912 }
 913
 914 void nbd_export_put(NBDExport *exp)
 915 {
 916     assert(exp->refcount > 0);
 917     if (exp->refcount == 1) {
 918         nbd_export_close(exp);
 919     }
 920
 921     if (--exp->refcount == 0) {
 922         assert(exp->name == NULL);
 923
 924         if (exp->close) {
 925             exp->close(exp);
 926         }
 927
 928         if (exp->blk) {
 929             if (exp->eject_notifier_blk) {
 930                 notifier_remove(&exp->eject_notifier);
 931                 blk_unref(exp->eject_notifier_blk);
 932             }
 933             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 934                                             blk_aio_detach, exp);
 935             blk_unref(exp->blk);
 936             exp->blk = NULL;
 937         }
 938
 939         g_free(exp);
 940     }
 941 }
 942
 943 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
 944 {
 945     return exp->blk;
 946 }
 947
 948 void nbd_export_close_all(void)
 949 {
 950     NBDExport *exp, *next;
 951
 952     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 953         nbd_export_close(exp);
 954     }
 955 }
 956
 957 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 958                                  int len)
 959 {
 960     NBDClient *client = req->client;
 961     ssize_t rc, ret;
 962
 963     g_assert(qemu_in_coroutine());
 964     qemu_co_mutex_lock(&client->send_lock);
 965     client->send_coroutine = qemu_coroutine_self();
 966     nbd_set_handlers(client);
 967
 968     if (!len) {
 969         rc = nbd_send_reply(client->ioc, reply);
 970     } else {
 971         qio_channel_set_cork(client->ioc, true);
 972         rc = nbd_send_reply(client->ioc, reply);
 973         if (rc >= 0) {
 974             ret = write_sync(client->ioc, req->data, len);
 975             if (ret != len) {
 976                 rc = -EIO;
 977             }
 978         }
 979         qio_channel_set_cork(client->ioc, false);
 980     }
 981
 982     client->send_coroutine = NULL;
 983     nbd_set_handlers(client);
 984     qemu_co_mutex_unlock(&client->send_lock);
 985     return rc;
 986 }
 987
 988 /* Collect a client request.  Return 0 if request looks valid, -EAGAIN
 989  * to keep trying the collection, -EIO to drop connection right away,
 990  * and any other negative value to report an error to the client
 991  * (although the caller may still need to disconnect after reporting
 992  * the error).  */
 993 static ssize_t nbd_co_receive_request(NBDRequest *req,
 994                                       struct nbd_request *request)
 995 {
 996     NBDClient *client = req->client;
 997     uint32_t command;
 998     ssize_t rc;
 999
1000     g_assert(qemu_in_coroutine());
1001     client->recv_coroutine = qemu_coroutine_self();
1002     nbd_update_can_read(client);
1003
1004     rc = nbd_receive_request(client->ioc, request);
1005     if (rc < 0) {
1006         if (rc != -EAGAIN) {
1007             rc = -EIO;
1008         }
1009         goto out;
1010     }
1011
1012     TRACE("Decoding type");
1013
1014     command = request->type & NBD_CMD_MASK_COMMAND;
1015     if (command != NBD_CMD_WRITE) {
1016         /* No payload, we are ready to read the next request.  */
1017         req->complete = true;
1018     }
1019
1020     if (command == NBD_CMD_DISC) {
1021         /* Special case: we're going to disconnect without a reply,
1022          * whether or not flags, from, or len are bogus */
1023         TRACE("Request type is DISCONNECT");
1024         rc = -EIO;
1025         goto out;
1026     }
1027
1028     /* Check for sanity in the parameters, part 1.  Defer as many
1029      * checks as possible until after reading any NBD_CMD_WRITE
1030      * payload, so we can try and keep the connection alive.  */
1031     if ((request->from + request->len) < request->from) {
1032         LOG("integer overflow detected, you're probably being attacked");
1033         rc = -EINVAL;
1034         goto out;
1035     }
1036
1037     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
1038         if (request->len > NBD_MAX_BUFFER_SIZE) {
1039             LOG("len (%" PRIu32" ) is larger than max len (%u)",
1040                 request->len, NBD_MAX_BUFFER_SIZE);
1041             rc = -EINVAL;
1042             goto out;
1043         }
1044
1045         req->data = blk_try_blockalign(client->exp->blk, request->len);
1046         if (req->data == NULL) {
1047             rc = -ENOMEM;
1048             goto out;
1049         }
1050     }
1051     if (command == NBD_CMD_WRITE) {
1052         TRACE("Reading %" PRIu32 " byte(s)", request->len);
1053
1054         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1055             LOG("reading from socket failed");
1056             rc = -EIO;
1057             goto out;
1058         }
1059         req->complete = true;
1060     }
1061
1062     /* Sanity checks, part 2. */
1063     if (request->from + request->len > client->exp->size) {
1064         LOG("operation past EOF; From: %" PRIu64 ", Len: %" PRIu32
1065             ", Size: %" PRIu64, request->from, request->len,
1066             (uint64_t)client->exp->size);
1067         rc = command == NBD_CMD_WRITE ? -ENOSPC : -EINVAL;
1068         goto out;
1069     }
1070     if (request->type & ~NBD_CMD_MASK_COMMAND & ~NBD_CMD_FLAG_FUA) {
1071         LOG("unsupported flags (got 0x%x)",
1072             request->type & ~NBD_CMD_MASK_COMMAND);
1073         rc = -EINVAL;
1074         goto out;
1075     }
1076
1077     rc = 0;
1078
1079 out:
1080     client->recv_coroutine = NULL;
1081     nbd_update_can_read(client);
1082
1083     return rc;
1084 }
1085
1086 static void nbd_trip(void *opaque)
1087 {
1088     NBDClient *client = opaque;
1089     NBDExport *exp = client->exp;
1090     NBDRequest *req;
1091     struct nbd_request request;
1092     struct nbd_reply reply;
1093     ssize_t ret;
1094     uint32_t command;
1095     int flags;
1096
1097     TRACE("Reading request.");
1098     if (client->closing) {
1099         return;
1100     }
1101
1102     req = nbd_request_get(client);
1103     ret = nbd_co_receive_request(req, &request);
1104     if (ret == -EAGAIN) {
1105         goto done;
1106     }
1107     if (ret == -EIO) {
1108         goto out;
1109     }
1110
1111     reply.handle = request.handle;
1112     reply.error = 0;
1113
1114     if (ret < 0) {
1115         reply.error = -ret;
1116         goto error_reply;
1117     }
1118     command = request.type & NBD_CMD_MASK_COMMAND;
1119
1120     if (client->closing) {
1121         /*
1122          * The client may be closed when we are blocked in
1123          * nbd_co_receive_request()
1124          */
1125         goto done;
1126     }
1127
1128     switch (command) {
1129     case NBD_CMD_READ:
1130         TRACE("Request type is READ");
1131
1132         if (request.type & NBD_CMD_FLAG_FUA) {
1133             ret = blk_co_flush(exp->blk);
1134             if (ret < 0) {
1135                 LOG("flush failed");
1136                 reply.error = -ret;
1137                 goto error_reply;
1138             }
1139         }
1140
1141         ret = blk_pread(exp->blk, request.from + exp->dev_offset,
1142                         req->data, request.len);
1143         if (ret < 0) {
1144             LOG("reading from file failed");
1145             reply.error = -ret;
1146             goto error_reply;
1147         }
1148
1149         TRACE("Read %" PRIu32" byte(s)", request.len);
1150         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1151             goto out;
1152         break;
1153     case NBD_CMD_WRITE:
1154         TRACE("Request type is WRITE");
1155
1156         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1157             TRACE("Server is read-only, return error");
1158             reply.error = EROFS;
1159             goto error_reply;
1160         }
1161
1162         TRACE("Writing to device");
1163
1164         flags = 0;
1165         if (request.type & NBD_CMD_FLAG_FUA) {
1166             flags |= BDRV_REQ_FUA;
1167         }
1168         ret = blk_pwrite(exp->blk, request.from + exp->dev_offset,
1169                          req->data, request.len, flags);
1170         if (ret < 0) {
1171             LOG("writing to file failed");
1172             reply.error = -ret;
1173             goto error_reply;
1174         }
1175
1176         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1177             goto out;
1178         }
1179         break;
1180
1181     case NBD_CMD_DISC:
1182         /* unreachable, thanks to special case in nbd_co_receive_request() */
1183         abort();
1184
1185     case NBD_CMD_FLUSH:
1186         TRACE("Request type is FLUSH");
1187
1188         ret = blk_co_flush(exp->blk);
1189         if (ret < 0) {
1190             LOG("flush failed");
1191             reply.error = -ret;
1192         }
1193         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1194             goto out;
1195         }
1196         break;
1197     case NBD_CMD_TRIM:
1198         TRACE("Request type is TRIM");
1199         ret = blk_co_pdiscard(exp->blk, request.from + exp->dev_offset,
1200                               request.len);
1201         if (ret < 0) {
1202             LOG("discard failed");
1203             reply.error = -ret;
1204         }
1205         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1206             goto out;
1207         }
1208         break;
1209     default:
1210         LOG("invalid request type (%" PRIu32 ") received", request.type);
1211         reply.error = EINVAL;
1212     error_reply:
1213         /* We must disconnect after NBD_CMD_WRITE if we did not
1214          * read the payload.
1215          */
1216         if (nbd_co_send_reply(req, &reply, 0) < 0 || !req->complete) {
1217             goto out;
1218         }
1219         break;
1220     }
1221
1222     TRACE("Request/Reply complete");
1223
1224 done:
1225     nbd_request_put(req);
1226     return;
1227
1228 out:
1229     nbd_request_put(req);
1230     client_close(client);
1231 }
1232
1233 static void nbd_read(void *opaque)
1234 {
1235     NBDClient *client = opaque;
1236
1237     if (client->recv_coroutine) {
1238         qemu_coroutine_enter(client->recv_coroutine);
1239     } else {
1240         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip, client));
1241     }
1242 }
1243
1244 static void nbd_restart_write(void *opaque)
1245 {
1246     NBDClient *client = opaque;
1247
1248     qemu_coroutine_enter(client->send_coroutine);
1249 }
1250
1251 static void nbd_set_handlers(NBDClient *client)
1252 {
1253     if (client->exp && client->exp->ctx) {
1254         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1255                            true,
1256                            client->can_read ? nbd_read : NULL,
1257                            client->send_coroutine ? nbd_restart_write : NULL,
1258                            client);
1259     }
1260 }
1261
1262 static void nbd_unset_handlers(NBDClient *client)
1263 {
1264     if (client->exp && client->exp->ctx) {
1265         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1266                            true, NULL, NULL, NULL);
1267     }
1268 }
1269
1270 static void nbd_update_can_read(NBDClient *client)
1271 {
1272     bool can_read = client->recv_coroutine ||
1273                     client->nb_requests < MAX_NBD_REQUESTS;
1274
1275     if (can_read != client->can_read) {
1276         client->can_read = can_read;
1277         nbd_set_handlers(client);
1278
1279         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1280          * in nbd_set_handlers() will have taken care of that */
1281     }
1282 }
1283
1284 static coroutine_fn void nbd_co_client_start(void *opaque)
1285 {
1286     NBDClientNewData *data = opaque;
1287     NBDClient *client = data->client;
1288     NBDExport *exp = client->exp;
1289
1290     if (exp) {
1291         nbd_export_get(exp);
1292     }
1293     if (nbd_negotiate(data)) {
1294         client_close(client);
1295         goto out;
1296     }
1297     qemu_co_mutex_init(&client->send_lock);
1298     nbd_set_handlers(client);
1299
1300     if (exp) {
1301         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1302     }
1303 out:
1304     g_free(data);
1305 }
1306
1307 void nbd_client_new(NBDExport *exp,
1308                     QIOChannelSocket *sioc,
1309                     QCryptoTLSCreds *tlscreds,
1310                     const char *tlsaclname,
1311                     void (*close_fn)(NBDClient *))
1312 {
1313     NBDClient *client;
1314     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1315
1316     client = g_malloc0(sizeof(NBDClient));
1317     client->refcount = 1;
1318     client->exp = exp;
1319     client->tlscreds = tlscreds;
1320     if (tlscreds) {
1321         object_ref(OBJECT(client->tlscreds));
1322     }
1323     client->tlsaclname = g_strdup(tlsaclname);
1324     client->sioc = sioc;
1325     object_ref(OBJECT(client->sioc));
1326     client->ioc = QIO_CHANNEL(sioc);
1327     object_ref(OBJECT(client->ioc));
1328     client->can_read = true;
1329     client->close = close_fn;
1330
1331     data->client = client;
1332     data->co = qemu_coroutine_create(nbd_co_client_start, data);
1333     qemu_coroutine_enter(data->co);
1334 }