nbd/server.c

   1 /*
   2  *  Copyright (C) 2016 Red Hat, Inc.
   3  *  Copyright (C) 2005  Anthony Liguori <anthony@codemonkey.ws>
   4  *
   5  *  Network Block Device Server Side
   6  *
   7  *  This program is free software; you can redistribute it and/or modify
   8  *  it under the terms of the GNU General Public License as published by
   9  *  the Free Software Foundation; under version 2 of the License.
  10  *
  11  *  This program is distributed in the hope that it will be useful,
  12  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  *  GNU General Public License for more details.
  15  *
  16  *  You should have received a copy of the GNU General Public License
  17  *  along with this program; if not, see <http://www.gnu.org/licenses/>.
  18  */
  19
  20 #include "qemu/osdep.h"
  21 #include "qapi/error.h"
  22 #include "nbd-internal.h"
  23
  24 static int system_errno_to_nbd_errno(int err)
  25 {
  26     switch (err) {
  27     case 0:
  28         return NBD_SUCCESS;
  29     case EPERM:
  30     case EROFS:
  31         return NBD_EPERM;
  32     case EIO:
  33         return NBD_EIO;
  34     case ENOMEM:
  35         return NBD_ENOMEM;
  36 #ifdef EDQUOT
  37     case EDQUOT:
  38 #endif
  39     case EFBIG:
  40     case ENOSPC:
  41         return NBD_ENOSPC;
  42     case ESHUTDOWN:
  43         return NBD_ESHUTDOWN;
  44     case EINVAL:
  45     default:
  46         return NBD_EINVAL;
  47     }
  48 }
  49
  50 /* Definitions for opaque data types */
  51
  52 typedef struct NBDRequestData NBDRequestData;
  53
  54 struct NBDRequestData {
  55     QSIMPLEQ_ENTRY(NBDRequestData) entry;
  56     NBDClient *client;
  57     uint8_t *data;
  58     bool complete;
  59 };
  60
  61 struct NBDExport {
  62     int refcount;
  63     void (*close)(NBDExport *exp);
  64
  65     BlockBackend *blk;
  66     char *name;
  67     char *description;
  68     off_t dev_offset;
  69     off_t size;
  70     uint16_t nbdflags;
  71     QTAILQ_HEAD(, NBDClient) clients;
  72     QTAILQ_ENTRY(NBDExport) next;
  73
  74     AioContext *ctx;
  75
  76     BlockBackend *eject_notifier_blk;
  77     Notifier eject_notifier;
  78 };
  79
  80 static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports);
  81
  82 struct NBDClient {
  83     int refcount;
  84     void (*close_fn)(NBDClient *client, bool negotiated);
  85
  86     bool no_zeroes;
  87     NBDExport *exp;
  88     QCryptoTLSCreds *tlscreds;
  89     char *tlsaclname;
  90     QIOChannelSocket *sioc; /* The underlying data channel */
  91     QIOChannel *ioc; /* The current I/O channel which may differ (eg TLS) */
  92
  93     Coroutine *recv_coroutine;
  94
  95     CoMutex send_lock;
  96     Coroutine *send_coroutine;
  97
  98     QTAILQ_ENTRY(NBDClient) next;
  99     int nb_requests;
 100     bool closing;
 101 };
 102
 103 /* That's all folks */
 104
 105 static void nbd_client_receive_next_request(NBDClient *client);
 106
 107 /* Basic flow for negotiation
 108
 109    Server         Client
 110    Negotiate
 111
 112    or
 113
 114    Server         Client
 115    Negotiate #1
 116                   Option
 117    Negotiate #2
 118
 119    ----
 120
 121    followed by
 122
 123    Server         Client
 124                   Request
 125    Response
 126                   Request
 127    Response
 128                   ...
 129    ...
 130                   Request (type == 2)
 131
 132 */
 133
 134 /* Send a reply header, including length, but no payload.
 135  * Return -errno on error, 0 on success. */
 136 static int nbd_negotiate_send_rep_len(QIOChannel *ioc, uint32_t type,
 137                                       uint32_t opt, uint32_t len, Error **errp)
 138 {
 139     uint64_t magic;
 140
 141     TRACE("Reply opt=%" PRIx32 " type=%" PRIx32 " len=%" PRIu32,
 142           opt, type, len);
 143
 144     magic = cpu_to_be64(NBD_REP_MAGIC);
 145     if (nbd_write(ioc, &magic, sizeof(magic), errp) < 0) {
 146         error_prepend(errp, "write failed (rep magic): ");
 147         return -EINVAL;
 148     }
 149
 150     opt = cpu_to_be32(opt);
 151     if (nbd_write(ioc, &opt, sizeof(opt), errp) < 0) {
 152         error_prepend(errp, "write failed (rep opt): ");
 153         return -EINVAL;
 154     }
 155
 156     type = cpu_to_be32(type);
 157     if (nbd_write(ioc, &type, sizeof(type), errp) < 0) {
 158         error_prepend(errp, "write failed (rep type): ");
 159         return -EINVAL;
 160     }
 161
 162     len = cpu_to_be32(len);
 163     if (nbd_write(ioc, &len, sizeof(len), errp) < 0) {
 164         error_prepend(errp, "write failed (rep data length): ");
 165         return -EINVAL;
 166     }
 167     return 0;
 168 }
 169
 170 /* Send a reply header with default 0 length.
 171  * Return -errno on error, 0 on success. */
 172 static int nbd_negotiate_send_rep(QIOChannel *ioc, uint32_t type, uint32_t opt,
 173                                   Error **errp)
 174 {
 175     return nbd_negotiate_send_rep_len(ioc, type, opt, 0, errp);
 176 }
 177
 178 /* Send an error reply.
 179  * Return -errno on error, 0 on success. */
 180 static int GCC_FMT_ATTR(5, 6)
 181 nbd_negotiate_send_rep_err(QIOChannel *ioc, uint32_t type,
 182                            uint32_t opt, Error **errp, const char *fmt, ...)
 183 {
 184     va_list va;
 185     char *msg;
 186     int ret;
 187     size_t len;
 188
 189     va_start(va, fmt);
 190     msg = g_strdup_vprintf(fmt, va);
 191     va_end(va);
 192     len = strlen(msg);
 193     assert(len < 4096);
 194     TRACE("sending error message \"%s\"", msg);
 195     ret = nbd_negotiate_send_rep_len(ioc, type, opt, len, errp);
 196     if (ret < 0) {
 197         goto out;
 198     }
 199     if (nbd_write(ioc, msg, len, errp) < 0) {
 200         error_prepend(errp, "write failed (error message): ");
 201         ret = -EIO;
 202     } else {
 203         ret = 0;
 204     }
 205
 206 out:
 207     g_free(msg);
 208     return ret;
 209 }
 210
 211 /* Send a single NBD_REP_SERVER reply to NBD_OPT_LIST, including payload.
 212  * Return -errno on error, 0 on success. */
 213 static int nbd_negotiate_send_rep_list(QIOChannel *ioc, NBDExport *exp,
 214                                        Error **errp)
 215 {
 216     size_t name_len, desc_len;
 217     uint32_t len;
 218     const char *name = exp->name ? exp->name : "";
 219     const char *desc = exp->description ? exp->description : "";
 220     int ret;
 221
 222     TRACE("Advertising export name '%s' description '%s'", name, desc);
 223     name_len = strlen(name);
 224     desc_len = strlen(desc);
 225     len = name_len + desc_len + sizeof(len);
 226     ret = nbd_negotiate_send_rep_len(ioc, NBD_REP_SERVER, NBD_OPT_LIST, len,
 227                                      errp);
 228     if (ret < 0) {
 229         return ret;
 230     }
 231
 232     len = cpu_to_be32(name_len);
 233     if (nbd_write(ioc, &len, sizeof(len), errp) < 0) {
 234         error_prepend(errp, "write failed (name length): ");
 235         return -EINVAL;
 236     }
 237
 238     if (nbd_write(ioc, name, name_len, errp) < 0) {
 239         error_prepend(errp, "write failed (name buffer): ");
 240         return -EINVAL;
 241     }
 242
 243     if (nbd_write(ioc, desc, desc_len, errp) < 0) {
 244         error_prepend(errp, "write failed (description buffer): ");
 245         return -EINVAL;
 246     }
 247
 248     return 0;
 249 }
 250
 251 /* Process the NBD_OPT_LIST command, with a potential series of replies.
 252  * Return -errno on error, 0 on success. */
 253 static int nbd_negotiate_handle_list(NBDClient *client, uint32_t length,
 254                                      Error **errp)
 255 {
 256     NBDExport *exp;
 257
 258     if (length) {
 259         if (nbd_drop(client->ioc, length, errp) < 0) {
 260             return -EIO;
 261         }
 262         return nbd_negotiate_send_rep_err(client->ioc,
 263                                           NBD_REP_ERR_INVALID, NBD_OPT_LIST,
 264                                           errp,
 265                                           "OPT_LIST should not have length");
 266     }
 267
 268     /* For each export, send a NBD_REP_SERVER reply. */
 269     QTAILQ_FOREACH(exp, &exports, next) {
 270         if (nbd_negotiate_send_rep_list(client->ioc, exp, errp)) {
 271             return -EINVAL;
 272         }
 273     }
 274     /* Finish with a NBD_REP_ACK. */
 275     return nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_LIST, errp);
 276 }
 277
 278 static int nbd_negotiate_handle_export_name(NBDClient *client, uint32_t length,
 279                                             Error **errp)
 280 {
 281     char name[NBD_MAX_NAME_SIZE + 1];
 282
 283     /* Client sends:
 284         [20 ..  xx]   export name (length bytes)
 285      */
 286     TRACE("Checking length");
 287     if (length >= sizeof(name)) {
 288         error_setg(errp, "Bad length received");
 289         return -EINVAL;
 290     }
 291     if (nbd_read(client->ioc, name, length, errp) < 0) {
 292         error_prepend(errp, "read failed: ");
 293         return -EINVAL;
 294     }
 295     name[length] = '\0';
 296
 297     TRACE("Client requested export '%s'", name);
 298
 299     client->exp = nbd_export_find(name);
 300     if (!client->exp) {
 301         error_setg(errp, "export not found");
 302         return -EINVAL;
 303     }
 304
 305     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 306     nbd_export_get(client->exp);
 307
 308     return 0;
 309 }
 310
 311 /* Handle NBD_OPT_STARTTLS. Return NULL to drop connection, or else the
 312  * new channel for all further (now-encrypted) communication. */
 313 static QIOChannel *nbd_negotiate_handle_starttls(NBDClient *client,
 314                                                  uint32_t length,
 315                                                  Error **errp)
 316 {
 317     QIOChannel *ioc;
 318     QIOChannelTLS *tioc;
 319     struct NBDTLSHandshakeData data = { 0 };
 320
 321     TRACE("Setting up TLS");
 322     ioc = client->ioc;
 323     if (length) {
 324         if (nbd_drop(ioc, length, errp) < 0) {
 325             return NULL;
 326         }
 327         nbd_negotiate_send_rep_err(ioc, NBD_REP_ERR_INVALID, NBD_OPT_STARTTLS,
 328                                    errp,
 329                                    "OPT_STARTTLS should not have length");
 330         return NULL;
 331     }
 332
 333     if (nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK,
 334                                NBD_OPT_STARTTLS, errp) < 0) {
 335         return NULL;
 336     }
 337
 338     tioc = qio_channel_tls_new_server(ioc,
 339                                       client->tlscreds,
 340                                       client->tlsaclname,
 341                                       errp);
 342     if (!tioc) {
 343         return NULL;
 344     }
 345
 346     qio_channel_set_name(QIO_CHANNEL(tioc), "nbd-server-tls");
 347     TRACE("Starting TLS handshake");
 348     data.loop = g_main_loop_new(g_main_context_default(), FALSE);
 349     qio_channel_tls_handshake(tioc,
 350                               nbd_tls_handshake,
 351                               &data,
 352                               NULL);
 353
 354     if (!data.complete) {
 355         g_main_loop_run(data.loop);
 356     }
 357     g_main_loop_unref(data.loop);
 358     if (data.error) {
 359         object_unref(OBJECT(tioc));
 360         error_propagate(errp, data.error);
 361         return NULL;
 362     }
 363
 364     return QIO_CHANNEL(tioc);
 365 }
 366
 367 /* nbd_negotiate_options
 368  * Process all NBD_OPT_* client option commands.
 369  * Return:
 370  * -errno  on error, errp is set
 371  * 0       on successful negotiation, errp is not set
 372  * 1       if client sent NBD_OPT_ABORT, i.e. on valid disconnect,
 373  *         errp is not set
 374  */
 375 static int nbd_negotiate_options(NBDClient *client, Error **errp)
 376 {
 377     uint32_t flags;
 378     bool fixedNewstyle = false;
 379     Error *local_err = NULL;
 380
 381     /* Client sends:
 382         [ 0 ..   3]   client flags
 383
 384         [ 0 ..   7]   NBD_OPTS_MAGIC
 385         [ 8 ..  11]   NBD option
 386         [12 ..  15]   Data length
 387         ...           Rest of request
 388
 389         [ 0 ..   7]   NBD_OPTS_MAGIC
 390         [ 8 ..  11]   Second NBD option
 391         [12 ..  15]   Data length
 392         ...           Rest of request
 393     */
 394
 395     if (nbd_read(client->ioc, &flags, sizeof(flags), errp) < 0) {
 396         error_prepend(errp, "read failed: ");
 397         return -EIO;
 398     }
 399     TRACE("Checking client flags");
 400     be32_to_cpus(&flags);
 401     if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) {
 402         TRACE("Client supports fixed newstyle handshake");
 403         fixedNewstyle = true;
 404         flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE;
 405     }
 406     if (flags & NBD_FLAG_C_NO_ZEROES) {
 407         TRACE("Client supports no zeroes at handshake end");
 408         client->no_zeroes = true;
 409         flags &= ~NBD_FLAG_C_NO_ZEROES;
 410     }
 411     if (flags != 0) {
 412         error_setg(errp, "Unknown client flags 0x%" PRIx32 " received", flags);
 413         return -EIO;
 414     }
 415
 416     while (1) {
 417         int ret;
 418         uint32_t clientflags, length;
 419         uint64_t magic;
 420
 421         if (nbd_read(client->ioc, &magic, sizeof(magic), errp) < 0) {
 422             error_prepend(errp, "read failed: ");
 423             return -EINVAL;
 424         }
 425         TRACE("Checking opts magic");
 426         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 427             error_setg(errp, "Bad magic received");
 428             return -EINVAL;
 429         }
 430
 431         if (nbd_read(client->ioc, &clientflags,
 432                      sizeof(clientflags), errp) < 0) {
 433             error_prepend(errp, "read failed: ");
 434             return -EINVAL;
 435         }
 436         clientflags = be32_to_cpu(clientflags);
 437
 438         if (nbd_read(client->ioc, &length, sizeof(length), errp) < 0) {
 439             error_prepend(errp, "read failed: ");
 440             return -EINVAL;
 441         }
 442         length = be32_to_cpu(length);
 443
 444         TRACE("Checking option 0x%" PRIx32, clientflags);
 445         if (client->tlscreds &&
 446             client->ioc == (QIOChannel *)client->sioc) {
 447             QIOChannel *tioc;
 448             if (!fixedNewstyle) {
 449                 error_setg(errp, "Unsupported option 0x%" PRIx32, clientflags);
 450                 return -EINVAL;
 451             }
 452             switch (clientflags) {
 453             case NBD_OPT_STARTTLS:
 454                 tioc = nbd_negotiate_handle_starttls(client, length, errp);
 455                 if (!tioc) {
 456                     return -EIO;
 457                 }
 458                 object_unref(OBJECT(client->ioc));
 459                 client->ioc = QIO_CHANNEL(tioc);
 460                 break;
 461
 462             case NBD_OPT_EXPORT_NAME:
 463                 /* No way to return an error to client, so drop connection */
 464                 error_setg(errp, "Option 0x%x not permitted before TLS",
 465                            clientflags);
 466                 return -EINVAL;
 467
 468             default:
 469                 if (nbd_drop(client->ioc, length, errp) < 0) {
 470                     return -EIO;
 471                 }
 472                 ret = nbd_negotiate_send_rep_err(client->ioc,
 473                                                  NBD_REP_ERR_TLS_REQD,
 474                                                  clientflags, errp,
 475                                                  "Option 0x%" PRIx32
 476                                                  "not permitted before TLS",
 477                                                  clientflags);
 478                 if (ret < 0) {
 479                     return ret;
 480                 }
 481                 /* Let the client keep trying, unless they asked to quit */
 482                 if (clientflags == NBD_OPT_ABORT) {
 483                     return 1;
 484                 }
 485                 break;
 486             }
 487         } else if (fixedNewstyle) {
 488             switch (clientflags) {
 489             case NBD_OPT_LIST:
 490                 ret = nbd_negotiate_handle_list(client, length, errp);
 491                 if (ret < 0) {
 492                     return ret;
 493                 }
 494                 break;
 495
 496             case NBD_OPT_ABORT:
 497                 /* NBD spec says we must try to reply before
 498                  * disconnecting, but that we must also tolerate
 499                  * guests that don't wait for our reply. */
 500                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, clientflags,
 501                                        &local_err);
 502
 503                 if (local_err != NULL) {
 504                     TRACE("Reply to NBD_OPT_ABORT request failed: %s",
 505                           error_get_pretty(local_err));
 506                     error_free(local_err);
 507                 }
 508
 509                 return 1;
 510
 511             case NBD_OPT_EXPORT_NAME:
 512                 return nbd_negotiate_handle_export_name(client, length, errp);
 513
 514             case NBD_OPT_STARTTLS:
 515                 if (nbd_drop(client->ioc, length, errp) < 0) {
 516                     return -EIO;
 517                 }
 518                 if (client->tlscreds) {
 519                     ret = nbd_negotiate_send_rep_err(client->ioc,
 520                                                      NBD_REP_ERR_INVALID,
 521                                                      clientflags, errp,
 522                                                      "TLS already enabled");
 523                 } else {
 524                     ret = nbd_negotiate_send_rep_err(client->ioc,
 525                                                      NBD_REP_ERR_POLICY,
 526                                                      clientflags, errp,
 527                                                      "TLS not configured");
 528                 }
 529                 if (ret < 0) {
 530                     return ret;
 531                 }
 532                 break;
 533             default:
 534                 if (nbd_drop(client->ioc, length, errp) < 0) {
 535                     return -EIO;
 536                 }
 537                 ret = nbd_negotiate_send_rep_err(client->ioc,
 538                                                  NBD_REP_ERR_UNSUP,
 539                                                  clientflags, errp,
 540                                                  "Unsupported option 0x%"
 541                                                  PRIx32,
 542                                                  clientflags);
 543                 if (ret < 0) {
 544                     return ret;
 545                 }
 546                 break;
 547             }
 548         } else {
 549             /*
 550              * If broken new-style we should drop the connection
 551              * for anything except NBD_OPT_EXPORT_NAME
 552              */
 553             switch (clientflags) {
 554             case NBD_OPT_EXPORT_NAME:
 555                 return nbd_negotiate_handle_export_name(client, length, errp);
 556
 557             default:
 558                 error_setg(errp, "Unsupported option 0x%" PRIx32, clientflags);
 559                 return -EINVAL;
 560             }
 561         }
 562     }
 563 }
 564
 565 /* nbd_negotiate
 566  * Return:
 567  * -errno  on error, errp is set
 568  * 0       on successful negotiation, errp is not set
 569  * 1       if client sent NBD_OPT_ABORT, i.e. on valid disconnect,
 570  *         errp is not set
 571  */
 572 static coroutine_fn int nbd_negotiate(NBDClient *client, Error **errp)
 573 {
 574     char buf[8 + 8 + 8 + 128];
 575     int ret;
 576     const uint16_t myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 577                               NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA |
 578                               NBD_FLAG_SEND_WRITE_ZEROES);
 579     bool oldStyle;
 580     size_t len;
 581
 582     /* Old style negotiation header without options
 583         [ 0 ..   7]   passwd       ("NBDMAGIC")
 584         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 585         [16 ..  23]   size
 586         [24 ..  25]   server flags (0)
 587         [26 ..  27]   export flags
 588         [28 .. 151]   reserved     (0)
 589
 590        New style negotiation header with options
 591         [ 0 ..   7]   passwd       ("NBDMAGIC")
 592         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 593         [16 ..  17]   server flags (0)
 594         ....options sent....
 595         [18 ..  25]   size
 596         [26 ..  27]   export flags
 597         [28 .. 151]   reserved     (0, omit if no_zeroes)
 598      */
 599
 600     qio_channel_set_blocking(client->ioc, false, NULL);
 601
 602     TRACE("Beginning negotiation.");
 603     memset(buf, 0, sizeof(buf));
 604     memcpy(buf, "NBDMAGIC", 8);
 605
 606     oldStyle = client->exp != NULL && !client->tlscreds;
 607     if (oldStyle) {
 608         TRACE("advertising size %" PRIu64 " and flags %x",
 609               client->exp->size, client->exp->nbdflags | myflags);
 610         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 611         stq_be_p(buf + 16, client->exp->size);
 612         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 613
 614         if (nbd_write(client->ioc, buf, sizeof(buf), errp) < 0) {
 615             error_prepend(errp, "write failed: ");
 616             return -EINVAL;
 617         }
 618     } else {
 619         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 620         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE | NBD_FLAG_NO_ZEROES);
 621
 622         if (nbd_write(client->ioc, buf, 18, errp) < 0) {
 623             error_prepend(errp, "write failed: ");
 624             return -EINVAL;
 625         }
 626         ret = nbd_negotiate_options(client, errp);
 627         if (ret != 0) {
 628             if (ret < 0) {
 629                 error_prepend(errp, "option negotiation failed: ");
 630             }
 631             return ret;
 632         }
 633
 634         TRACE("advertising size %" PRIu64 " and flags %x",
 635               client->exp->size, client->exp->nbdflags | myflags);
 636         stq_be_p(buf + 18, client->exp->size);
 637         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 638         len = client->no_zeroes ? 10 : sizeof(buf) - 18;
 639         ret = nbd_write(client->ioc, buf + 18, len, errp);
 640         if (ret < 0) {
 641             error_prepend(errp, "write failed: ");
 642             return ret;
 643         }
 644     }
 645
 646     TRACE("Negotiation succeeded.");
 647
 648     return 0;
 649 }
 650
 651 static int nbd_receive_request(QIOChannel *ioc, NBDRequest *request,
 652                                Error **errp)
 653 {
 654     uint8_t buf[NBD_REQUEST_SIZE];
 655     uint32_t magic;
 656     int ret;
 657
 658     ret = nbd_read(ioc, buf, sizeof(buf), errp);
 659     if (ret < 0) {
 660         return ret;
 661     }
 662
 663     /* Request
 664        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 665        [ 4 ..  5]   flags   (NBD_CMD_FLAG_FUA, ...)
 666        [ 6 ..  7]   type    (NBD_CMD_READ, ...)
 667        [ 8 .. 15]   handle
 668        [16 .. 23]   from
 669        [24 .. 27]   len
 670      */
 671
 672     magic = ldl_be_p(buf);
 673     request->flags  = lduw_be_p(buf + 4);
 674     request->type   = lduw_be_p(buf + 6);
 675     request->handle = ldq_be_p(buf + 8);
 676     request->from   = ldq_be_p(buf + 16);
 677     request->len    = ldl_be_p(buf + 24);
 678
 679     TRACE("Got request: { magic = 0x%" PRIx32 ", .flags = %" PRIx16
 680           ", .type = %" PRIx16 ", from = %" PRIu64 ", len = %" PRIu32 " }",
 681           magic, request->flags, request->type, request->from, request->len);
 682
 683     if (magic != NBD_REQUEST_MAGIC) {
 684         error_setg(errp, "invalid magic (got 0x%" PRIx32 ")", magic);
 685         return -EINVAL;
 686     }
 687     return 0;
 688 }
 689
 690 static int nbd_send_reply(QIOChannel *ioc, NBDReply *reply, Error **errp)
 691 {
 692     uint8_t buf[NBD_REPLY_SIZE];
 693
 694     reply->error = system_errno_to_nbd_errno(reply->error);
 695
 696     TRACE("Sending response to client: { .error = %" PRId32
 697           ", handle = %" PRIu64 " }",
 698           reply->error, reply->handle);
 699
 700     /* Reply
 701        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 702        [ 4 ..  7]    error   (0 == no error)
 703        [ 7 .. 15]    handle
 704      */
 705     stl_be_p(buf, NBD_REPLY_MAGIC);
 706     stl_be_p(buf + 4, reply->error);
 707     stq_be_p(buf + 8, reply->handle);
 708
 709     return nbd_write(ioc, buf, sizeof(buf), errp);
 710 }
 711
 712 #define MAX_NBD_REQUESTS 16
 713
 714 void nbd_client_get(NBDClient *client)
 715 {
 716     client->refcount++;
 717 }
 718
 719 void nbd_client_put(NBDClient *client)
 720 {
 721     if (--client->refcount == 0) {
 722         /* The last reference should be dropped by client->close,
 723          * which is called by client_close.
 724          */
 725         assert(client->closing);
 726
 727         qio_channel_detach_aio_context(client->ioc);
 728         object_unref(OBJECT(client->sioc));
 729         object_unref(OBJECT(client->ioc));
 730         if (client->tlscreds) {
 731             object_unref(OBJECT(client->tlscreds));
 732         }
 733         g_free(client->tlsaclname);
 734         if (client->exp) {
 735             QTAILQ_REMOVE(&client->exp->clients, client, next);
 736             nbd_export_put(client->exp);
 737         }
 738         g_free(client);
 739     }
 740 }
 741
 742 static void client_close(NBDClient *client, bool negotiated)
 743 {
 744     if (client->closing) {
 745         return;
 746     }
 747
 748     client->closing = true;
 749
 750     /* Force requests to finish.  They will drop their own references,
 751      * then we'll close the socket and free the NBDClient.
 752      */
 753     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 754                          NULL);
 755
 756     /* Also tell the client, so that they release their reference.  */
 757     if (client->close_fn) {
 758         client->close_fn(client, negotiated);
 759     }
 760 }
 761
 762 static NBDRequestData *nbd_request_get(NBDClient *client)
 763 {
 764     NBDRequestData *req;
 765
 766     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 767     client->nb_requests++;
 768
 769     req = g_new0(NBDRequestData, 1);
 770     nbd_client_get(client);
 771     req->client = client;
 772     return req;
 773 }
 774
 775 static void nbd_request_put(NBDRequestData *req)
 776 {
 777     NBDClient *client = req->client;
 778
 779     if (req->data) {
 780         qemu_vfree(req->data);
 781     }
 782     g_free(req);
 783
 784     client->nb_requests--;
 785     nbd_client_receive_next_request(client);
 786
 787     nbd_client_put(client);
 788 }
 789
 790 static void blk_aio_attached(AioContext *ctx, void *opaque)
 791 {
 792     NBDExport *exp = opaque;
 793     NBDClient *client;
 794
 795     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 796
 797     exp->ctx = ctx;
 798
 799     QTAILQ_FOREACH(client, &exp->clients, next) {
 800         qio_channel_attach_aio_context(client->ioc, ctx);
 801         if (client->recv_coroutine) {
 802             aio_co_schedule(ctx, client->recv_coroutine);
 803         }
 804         if (client->send_coroutine) {
 805             aio_co_schedule(ctx, client->send_coroutine);
 806         }
 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         qio_channel_detach_aio_context(client->ioc);
 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(BlockDriverState *bs, off_t dev_offset, off_t size,
 831                           uint16_t nbdflags, void (*close)(NBDExport *),
 832                           bool writethrough, BlockBackend *on_eject_blk,
 833                           Error **errp)
 834 {
 835     BlockBackend *blk;
 836     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 837     uint64_t perm;
 838     int ret;
 839
 840     /* Don't allow resize while the NBD server is running, otherwise we don't
 841      * care what happens with the node. */
 842     perm = BLK_PERM_CONSISTENT_READ;
 843     if ((nbdflags & NBD_FLAG_READ_ONLY) == 0) {
 844         perm |= BLK_PERM_WRITE;
 845     }
 846     blk = blk_new(perm, BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
 847                         BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD);
 848     ret = blk_insert_bs(blk, bs, errp);
 849     if (ret < 0) {
 850         goto fail;
 851     }
 852     blk_set_enable_write_cache(blk, !writethrough);
 853
 854     exp->refcount = 1;
 855     QTAILQ_INIT(&exp->clients);
 856     exp->blk = blk;
 857     exp->dev_offset = dev_offset;
 858     exp->nbdflags = nbdflags;
 859     exp->size = size < 0 ? blk_getlength(blk) : size;
 860     if (exp->size < 0) {
 861         error_setg_errno(errp, -exp->size,
 862                          "Failed to determine the NBD export's length");
 863         goto fail;
 864     }
 865     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 866
 867     exp->close = close;
 868     exp->ctx = blk_get_aio_context(blk);
 869     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 870
 871     if (on_eject_blk) {
 872         blk_ref(on_eject_blk);
 873         exp->eject_notifier_blk = on_eject_blk;
 874         exp->eject_notifier.notify = nbd_eject_notifier;
 875         blk_add_remove_bs_notifier(on_eject_blk, &exp->eject_notifier);
 876     }
 877
 878     /*
 879      * NBD exports are used for non-shared storage migration.  Make sure
 880      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 881      * access since the export could be available before migration handover.
 882      */
 883     aio_context_acquire(exp->ctx);
 884     blk_invalidate_cache(blk, NULL);
 885     aio_context_release(exp->ctx);
 886     return exp;
 887
 888 fail:
 889     blk_unref(blk);
 890     g_free(exp);
 891     return NULL;
 892 }
 893
 894 NBDExport *nbd_export_find(const char *name)
 895 {
 896     NBDExport *exp;
 897     QTAILQ_FOREACH(exp, &exports, next) {
 898         if (strcmp(name, exp->name) == 0) {
 899             return exp;
 900         }
 901     }
 902
 903     return NULL;
 904 }
 905
 906 void nbd_export_set_name(NBDExport *exp, const char *name)
 907 {
 908     if (exp->name == name) {
 909         return;
 910     }
 911
 912     nbd_export_get(exp);
 913     if (exp->name != NULL) {
 914         g_free(exp->name);
 915         exp->name = NULL;
 916         QTAILQ_REMOVE(&exports, exp, next);
 917         nbd_export_put(exp);
 918     }
 919     if (name != NULL) {
 920         nbd_export_get(exp);
 921         exp->name = g_strdup(name);
 922         QTAILQ_INSERT_TAIL(&exports, exp, next);
 923     }
 924     nbd_export_put(exp);
 925 }
 926
 927 void nbd_export_set_description(NBDExport *exp, const char *description)
 928 {
 929     g_free(exp->description);
 930     exp->description = g_strdup(description);
 931 }
 932
 933 void nbd_export_close(NBDExport *exp)
 934 {
 935     NBDClient *client, *next;
 936
 937     nbd_export_get(exp);
 938     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 939         client_close(client, true);
 940     }
 941     nbd_export_set_name(exp, NULL);
 942     nbd_export_set_description(exp, NULL);
 943     nbd_export_put(exp);
 944 }
 945
 946 void nbd_export_get(NBDExport *exp)
 947 {
 948     assert(exp->refcount > 0);
 949     exp->refcount++;
 950 }
 951
 952 void nbd_export_put(NBDExport *exp)
 953 {
 954     assert(exp->refcount > 0);
 955     if (exp->refcount == 1) {
 956         nbd_export_close(exp);
 957     }
 958
 959     if (--exp->refcount == 0) {
 960         assert(exp->name == NULL);
 961         assert(exp->description == NULL);
 962
 963         if (exp->close) {
 964             exp->close(exp);
 965         }
 966
 967         if (exp->blk) {
 968             if (exp->eject_notifier_blk) {
 969                 notifier_remove(&exp->eject_notifier);
 970                 blk_unref(exp->eject_notifier_blk);
 971             }
 972             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 973                                             blk_aio_detach, exp);
 974             blk_unref(exp->blk);
 975             exp->blk = NULL;
 976         }
 977
 978         g_free(exp);
 979     }
 980 }
 981
 982 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
 983 {
 984     return exp->blk;
 985 }
 986
 987 void nbd_export_close_all(void)
 988 {
 989     NBDExport *exp, *next;
 990
 991     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 992         nbd_export_close(exp);
 993     }
 994 }
 995
 996 static int nbd_co_send_reply(NBDRequestData *req, NBDReply *reply, int len,
 997                              Error **errp)
 998 {
 999     NBDClient *client = req->client;
1000     int ret;
1001
1002     g_assert(qemu_in_coroutine());
1003     qemu_co_mutex_lock(&client->send_lock);
1004     client->send_coroutine = qemu_coroutine_self();
1005
1006     if (!len) {
1007         ret = nbd_send_reply(client->ioc, reply, errp);
1008     } else {
1009         qio_channel_set_cork(client->ioc, true);
1010         ret = nbd_send_reply(client->ioc, reply, errp);
1011         if (ret == 0) {
1012             ret = nbd_write(client->ioc, req->data, len, errp);
1013             if (ret < 0) {
1014                 ret = -EIO;
1015             }
1016         }
1017         qio_channel_set_cork(client->ioc, false);
1018     }
1019
1020     client->send_coroutine = NULL;
1021     qemu_co_mutex_unlock(&client->send_lock);
1022     return ret;
1023 }
1024
1025 /* nbd_co_receive_request
1026  * Collect a client request. Return 0 if request looks valid, -EIO to drop
1027  * connection right away, and any other negative value to report an error to
1028  * the client (although the caller may still need to disconnect after reporting
1029  * the error).
1030  */
1031 static int nbd_co_receive_request(NBDRequestData *req, NBDRequest *request,
1032                                   Error **errp)
1033 {
1034     NBDClient *client = req->client;
1035
1036     g_assert(qemu_in_coroutine());
1037     assert(client->recv_coroutine == qemu_coroutine_self());
1038     if (nbd_receive_request(client->ioc, request, errp) < 0) {
1039         return -EIO;
1040     }
1041
1042     TRACE("Decoding type");
1043
1044     if (request->type != NBD_CMD_WRITE) {
1045         /* No payload, we are ready to read the next request.  */
1046         req->complete = true;
1047     }
1048
1049     if (request->type == NBD_CMD_DISC) {
1050         /* Special case: we're going to disconnect without a reply,
1051          * whether or not flags, from, or len are bogus */
1052         TRACE("Request type is DISCONNECT");
1053         return -EIO;
1054     }
1055
1056     /* Check for sanity in the parameters, part 1.  Defer as many
1057      * checks as possible until after reading any NBD_CMD_WRITE
1058      * payload, so we can try and keep the connection alive.  */
1059     if ((request->from + request->len) < request->from) {
1060         error_setg(errp,
1061                    "integer overflow detected, you're probably being attacked");
1062         return -EINVAL;
1063     }
1064
1065     if (request->type == NBD_CMD_READ || request->type == NBD_CMD_WRITE) {
1066         if (request->len > NBD_MAX_BUFFER_SIZE) {
1067             error_setg(errp, "len (%" PRIu32" ) is larger than max len (%u)",
1068                        request->len, NBD_MAX_BUFFER_SIZE);
1069             return -EINVAL;
1070         }
1071
1072         req->data = blk_try_blockalign(client->exp->blk, request->len);
1073         if (req->data == NULL) {
1074             error_setg(errp, "No memory");
1075             return -ENOMEM;
1076         }
1077     }
1078     if (request->type == NBD_CMD_WRITE) {
1079         TRACE("Reading %" PRIu32 " byte(s)", request->len);
1080
1081         if (nbd_read(client->ioc, req->data, request->len, errp) < 0) {
1082             error_prepend(errp, "reading from socket failed: ");
1083             return -EIO;
1084         }
1085         req->complete = true;
1086     }
1087
1088     /* Sanity checks, part 2. */
1089     if (request->from + request->len > client->exp->size) {
1090         error_setg(errp, "operation past EOF; From: %" PRIu64 ", Len: %" PRIu32
1091                    ", Size: %" PRIu64, request->from, request->len,
1092                    (uint64_t)client->exp->size);
1093         return request->type == NBD_CMD_WRITE ? -ENOSPC : -EINVAL;
1094     }
1095     if (request->flags & ~(NBD_CMD_FLAG_FUA | NBD_CMD_FLAG_NO_HOLE)) {
1096         error_setg(errp, "unsupported flags (got 0x%x)", request->flags);
1097         return -EINVAL;
1098     }
1099     if (request->type != NBD_CMD_WRITE_ZEROES &&
1100         (request->flags & NBD_CMD_FLAG_NO_HOLE)) {
1101         error_setg(errp, "unexpected flags (got 0x%x)", request->flags);
1102         return -EINVAL;
1103     }
1104
1105     return 0;
1106 }
1107
1108 /* Owns a reference to the NBDClient passed as opaque.  */
1109 static coroutine_fn void nbd_trip(void *opaque)
1110 {
1111     NBDClient *client = opaque;
1112     NBDExport *exp = client->exp;
1113     NBDRequestData *req;
1114     NBDRequest request = { 0 };    /* GCC thinks it can be used uninitialized */
1115     NBDReply reply;
1116     int ret;
1117     int flags;
1118     int reply_data_len = 0;
1119     Error *local_err = NULL;
1120
1121     TRACE("Reading request.");
1122     if (client->closing) {
1123         nbd_client_put(client);
1124         return;
1125     }
1126
1127     req = nbd_request_get(client);
1128     ret = nbd_co_receive_request(req, &request, &local_err);
1129     client->recv_coroutine = NULL;
1130     nbd_client_receive_next_request(client);
1131     if (ret == -EIO) {
1132         goto disconnect;
1133     }
1134
1135     reply.handle = request.handle;
1136     reply.error = 0;
1137
1138     if (ret < 0) {
1139         reply.error = -ret;
1140         goto reply;
1141     }
1142
1143     if (client->closing) {
1144         /*
1145          * The client may be closed when we are blocked in
1146          * nbd_co_receive_request()
1147          */
1148         goto done;
1149     }
1150
1151     switch (request.type) {
1152     case NBD_CMD_READ:
1153         TRACE("Request type is READ");
1154
1155         /* XXX: NBD Protocol only documents use of FUA with WRITE */
1156         if (request.flags & NBD_CMD_FLAG_FUA) {
1157             ret = blk_co_flush(exp->blk);
1158             if (ret < 0) {
1159                 error_setg_errno(&local_err, -ret, "flush failed");
1160                 reply.error = -ret;
1161                 break;
1162             }
1163         }
1164
1165         ret = blk_pread(exp->blk, request.from + exp->dev_offset,
1166                         req->data, request.len);
1167         if (ret < 0) {
1168             error_setg_errno(&local_err, -ret, "reading from file failed");
1169             reply.error = -ret;
1170             break;
1171         }
1172
1173         reply_data_len = request.len;
1174         TRACE("Read %" PRIu32" byte(s)", request.len);
1175
1176         break;
1177     case NBD_CMD_WRITE:
1178         TRACE("Request type is WRITE");
1179
1180         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1181             TRACE("Server is read-only, return error");
1182             reply.error = EROFS;
1183             break;
1184         }
1185
1186         TRACE("Writing to device");
1187
1188         flags = 0;
1189         if (request.flags & NBD_CMD_FLAG_FUA) {
1190             flags |= BDRV_REQ_FUA;
1191         }
1192         ret = blk_pwrite(exp->blk, request.from + exp->dev_offset,
1193                          req->data, request.len, flags);
1194         if (ret < 0) {
1195             error_setg_errno(&local_err, -ret, "writing to file failed");
1196             reply.error = -ret;
1197         }
1198
1199         break;
1200     case NBD_CMD_WRITE_ZEROES:
1201         TRACE("Request type is WRITE_ZEROES");
1202
1203         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1204             error_setg(&local_err, "Server is read-only, return error");
1205             reply.error = EROFS;
1206             break;
1207         }
1208
1209         TRACE("Writing to device");
1210
1211         flags = 0;
1212         if (request.flags & NBD_CMD_FLAG_FUA) {
1213             flags |= BDRV_REQ_FUA;
1214         }
1215         if (!(request.flags & NBD_CMD_FLAG_NO_HOLE)) {
1216             flags |= BDRV_REQ_MAY_UNMAP;
1217         }
1218         ret = blk_pwrite_zeroes(exp->blk, request.from + exp->dev_offset,
1219                                 request.len, flags);
1220         if (ret < 0) {
1221             error_setg_errno(&local_err, -ret, "writing to file failed");
1222             reply.error = -ret;
1223         }
1224
1225         break;
1226     case NBD_CMD_DISC:
1227         /* unreachable, thanks to special case in nbd_co_receive_request() */
1228         abort();
1229
1230     case NBD_CMD_FLUSH:
1231         TRACE("Request type is FLUSH");
1232
1233         ret = blk_co_flush(exp->blk);
1234         if (ret < 0) {
1235             error_setg_errno(&local_err, -ret, "flush failed");
1236             reply.error = -ret;
1237         }
1238
1239         break;
1240     case NBD_CMD_TRIM:
1241         TRACE("Request type is TRIM");
1242         ret = blk_co_pdiscard(exp->blk, request.from + exp->dev_offset,
1243                               request.len);
1244         if (ret < 0) {
1245             error_setg_errno(&local_err, -ret, "discard failed");
1246             reply.error = -ret;
1247         }
1248
1249         break;
1250     default:
1251         error_setg(&local_err, "invalid request type (%" PRIu32 ") received",
1252                    request.type);
1253         reply.error = EINVAL;
1254     }
1255
1256 reply:
1257     if (local_err) {
1258         /* If we are here local_err is not fatal error, already stored in
1259          * reply.error */
1260         error_report_err(local_err);
1261         local_err = NULL;
1262     }
1263
1264     if (nbd_co_send_reply(req, &reply, reply_data_len, &local_err) < 0) {
1265         error_prepend(&local_err, "Failed to send reply: ");
1266         goto disconnect;
1267     }
1268
1269     /* We must disconnect after NBD_CMD_WRITE if we did not
1270      * read the payload.
1271      */
1272     if (!req->complete) {
1273         error_setg(&local_err, "Request handling failed in intermediate state");
1274         goto disconnect;
1275     }
1276
1277     TRACE("Request/Reply complete");
1278
1279 done:
1280     nbd_request_put(req);
1281     nbd_client_put(client);
1282     return;
1283
1284 disconnect:
1285     if (local_err) {
1286         error_reportf_err(local_err, "Disconnect client, due to: ");
1287     }
1288     nbd_request_put(req);
1289     client_close(client, true);
1290     nbd_client_put(client);
1291 }
1292
1293 static void nbd_client_receive_next_request(NBDClient *client)
1294 {
1295     if (!client->recv_coroutine && client->nb_requests < MAX_NBD_REQUESTS) {
1296         nbd_client_get(client);
1297         client->recv_coroutine = qemu_coroutine_create(nbd_trip, client);
1298         aio_co_schedule(client->exp->ctx, client->recv_coroutine);
1299     }
1300 }
1301
1302 static coroutine_fn void nbd_co_client_start(void *opaque)
1303 {
1304     NBDClient *client = opaque;
1305     NBDExport *exp = client->exp;
1306     Error *local_err = NULL;
1307
1308     if (exp) {
1309         nbd_export_get(exp);
1310         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1311     }
1312     qemu_co_mutex_init(&client->send_lock);
1313
1314     if (nbd_negotiate(client, &local_err)) {
1315         if (local_err) {
1316             error_report_err(local_err);
1317         }
1318         client_close(client, false);
1319         return;
1320     }
1321
1322     nbd_client_receive_next_request(client);
1323 }
1324
1325 /*
1326  * Create a new client listener on the given export @exp, using the
1327  * given channel @sioc.  Begin servicing it in a coroutine.  When the
1328  * connection closes, call @close_fn with an indication of whether the
1329  * client completed negotiation.
1330  */
1331 void nbd_client_new(NBDExport *exp,
1332                     QIOChannelSocket *sioc,
1333                     QCryptoTLSCreds *tlscreds,
1334                     const char *tlsaclname,
1335                     void (*close_fn)(NBDClient *, bool))
1336 {
1337     NBDClient *client;
1338     Coroutine *co;
1339
1340     client = g_malloc0(sizeof(NBDClient));
1341     client->refcount = 1;
1342     client->exp = exp;
1343     client->tlscreds = tlscreds;
1344     if (tlscreds) {
1345         object_ref(OBJECT(client->tlscreds));
1346     }
1347     client->tlsaclname = g_strdup(tlsaclname);
1348     client->sioc = sioc;
1349     object_ref(OBJECT(client->sioc));
1350     client->ioc = QIO_CHANNEL(sioc);
1351     object_ref(OBJECT(client->ioc));
1352     client->close_fn = close_fn;
1353
1354     co = qemu_coroutine_create(nbd_co_client_start, client);
1355     qemu_coroutine_enter(co);
1356 }