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)(NBDClient *client);
  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 static gboolean nbd_negotiate_continue(QIOChannel *ioc,
 108                                        GIOCondition condition,
 109                                        void *opaque)
 110 {
 111     qemu_coroutine_enter(opaque);
 112     return TRUE;
 113 }
 114
 115 static ssize_t nbd_negotiate_read(QIOChannel *ioc, void *buffer, size_t size)
 116 {
 117     ssize_t ret;
 118     guint watch;
 119
 120     assert(qemu_in_coroutine());
 121     /* Negotiation are always in main loop. */
 122     watch = qio_channel_add_watch(ioc,
 123                                   G_IO_IN,
 124                                   nbd_negotiate_continue,
 125                                   qemu_coroutine_self(),
 126                                   NULL);
 127     ret = read_sync(ioc, buffer, size);
 128     g_source_remove(watch);
 129     return ret;
 130
 131 }
 132
 133 static ssize_t nbd_negotiate_write(QIOChannel *ioc, const void *buffer,
 134                                    size_t size)
 135 {
 136     ssize_t ret;
 137     guint watch;
 138
 139     assert(qemu_in_coroutine());
 140     /* Negotiation are always in main loop. */
 141     watch = qio_channel_add_watch(ioc,
 142                                   G_IO_OUT,
 143                                   nbd_negotiate_continue,
 144                                   qemu_coroutine_self(),
 145                                   NULL);
 146     ret = write_sync(ioc, buffer, size);
 147     g_source_remove(watch);
 148     return ret;
 149 }
 150
 151 static ssize_t nbd_negotiate_drop_sync(QIOChannel *ioc, size_t size)
 152 {
 153     ssize_t ret, dropped = size;
 154     uint8_t *buffer = g_malloc(MIN(65536, size));
 155
 156     while (size > 0) {
 157         ret = nbd_negotiate_read(ioc, buffer, MIN(65536, size));
 158         if (ret < 0) {
 159             g_free(buffer);
 160             return ret;
 161         }
 162
 163         assert(ret <= size);
 164         size -= ret;
 165     }
 166
 167     g_free(buffer);
 168     return dropped;
 169 }
 170
 171 /* Basic flow for negotiation
 172
 173    Server         Client
 174    Negotiate
 175
 176    or
 177
 178    Server         Client
 179    Negotiate #1
 180                   Option
 181    Negotiate #2
 182
 183    ----
 184
 185    followed by
 186
 187    Server         Client
 188                   Request
 189    Response
 190                   Request
 191    Response
 192                   ...
 193    ...
 194                   Request (type == 2)
 195
 196 */
 197
 198 /* Send a reply header, including length, but no payload.
 199  * Return -errno on error, 0 on success. */
 200 static int nbd_negotiate_send_rep_len(QIOChannel *ioc, uint32_t type,
 201                                       uint32_t opt, uint32_t len)
 202 {
 203     uint64_t magic;
 204
 205     TRACE("Reply opt=%" PRIx32 " type=%" PRIx32 " len=%" PRIu32,
 206           type, opt, len);
 207
 208     magic = cpu_to_be64(NBD_REP_MAGIC);
 209     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 210         LOG("write failed (rep magic)");
 211         return -EINVAL;
 212     }
 213     opt = cpu_to_be32(opt);
 214     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 215         LOG("write failed (rep opt)");
 216         return -EINVAL;
 217     }
 218     type = cpu_to_be32(type);
 219     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 220         LOG("write failed (rep type)");
 221         return -EINVAL;
 222     }
 223     len = cpu_to_be32(len);
 224     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 225         LOG("write failed (rep data length)");
 226         return -EINVAL;
 227     }
 228     return 0;
 229 }
 230
 231 /* Send a reply header with default 0 length.
 232  * Return -errno on error, 0 on success. */
 233 static int nbd_negotiate_send_rep(QIOChannel *ioc, uint32_t type, uint32_t opt)
 234 {
 235     return nbd_negotiate_send_rep_len(ioc, type, opt, 0);
 236 }
 237
 238 /* Send an error reply.
 239  * Return -errno on error, 0 on success. */
 240 static int GCC_FMT_ATTR(4, 5)
 241 nbd_negotiate_send_rep_err(QIOChannel *ioc, uint32_t type,
 242                            uint32_t opt, const char *fmt, ...)
 243 {
 244     va_list va;
 245     char *msg;
 246     int ret;
 247     size_t len;
 248
 249     va_start(va, fmt);
 250     msg = g_strdup_vprintf(fmt, va);
 251     va_end(va);
 252     len = strlen(msg);
 253     assert(len < 4096);
 254     TRACE("sending error message \"%s\"", msg);
 255     ret = nbd_negotiate_send_rep_len(ioc, type, opt, len);
 256     if (ret < 0) {
 257         goto out;
 258     }
 259     if (nbd_negotiate_write(ioc, msg, len) != len) {
 260         LOG("write failed (error message)");
 261         ret = -EIO;
 262     } else {
 263         ret = 0;
 264     }
 265 out:
 266     g_free(msg);
 267     return ret;
 268 }
 269
 270 /* Send a single NBD_REP_SERVER reply to NBD_OPT_LIST, including payload.
 271  * Return -errno on error, 0 on success. */
 272 static int nbd_negotiate_send_rep_list(QIOChannel *ioc, NBDExport *exp)
 273 {
 274     size_t name_len, desc_len;
 275     uint32_t len;
 276     const char *name = exp->name ? exp->name : "";
 277     const char *desc = exp->description ? exp->description : "";
 278     int rc;
 279
 280     TRACE("Advertising export name '%s' description '%s'", name, desc);
 281     name_len = strlen(name);
 282     desc_len = strlen(desc);
 283     len = name_len + desc_len + sizeof(len);
 284     rc = nbd_negotiate_send_rep_len(ioc, NBD_REP_SERVER, NBD_OPT_LIST, len);
 285     if (rc < 0) {
 286         return rc;
 287     }
 288
 289     len = cpu_to_be32(name_len);
 290     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 291         LOG("write failed (name length)");
 292         return -EINVAL;
 293     }
 294     if (nbd_negotiate_write(ioc, name, name_len) != name_len) {
 295         LOG("write failed (name buffer)");
 296         return -EINVAL;
 297     }
 298     if (nbd_negotiate_write(ioc, desc, desc_len) != desc_len) {
 299         LOG("write failed (description buffer)");
 300         return -EINVAL;
 301     }
 302     return 0;
 303 }
 304
 305 /* Process the NBD_OPT_LIST command, with a potential series of replies.
 306  * Return -errno on error, 0 on success. */
 307 static int nbd_negotiate_handle_list(NBDClient *client, uint32_t length)
 308 {
 309     NBDExport *exp;
 310
 311     if (length) {
 312         if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 313             return -EIO;
 314         }
 315         return nbd_negotiate_send_rep_err(client->ioc,
 316                                           NBD_REP_ERR_INVALID, NBD_OPT_LIST,
 317                                           "OPT_LIST should not have length");
 318     }
 319
 320     /* For each export, send a NBD_REP_SERVER reply. */
 321     QTAILQ_FOREACH(exp, &exports, next) {
 322         if (nbd_negotiate_send_rep_list(client->ioc, exp)) {
 323             return -EINVAL;
 324         }
 325     }
 326     /* Finish with a NBD_REP_ACK. */
 327     return nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_LIST);
 328 }
 329
 330 static int nbd_negotiate_handle_export_name(NBDClient *client, uint32_t length)
 331 {
 332     int rc = -EINVAL;
 333     char name[NBD_MAX_NAME_SIZE + 1];
 334
 335     /* Client sends:
 336         [20 ..  xx]   export name (length bytes)
 337      */
 338     TRACE("Checking length");
 339     if (length >= sizeof(name)) {
 340         LOG("Bad length received");
 341         goto fail;
 342     }
 343     if (nbd_negotiate_read(client->ioc, name, length) != length) {
 344         LOG("read failed");
 345         goto fail;
 346     }
 347     name[length] = '\0';
 348
 349     TRACE("Client requested export '%s'", name);
 350
 351     client->exp = nbd_export_find(name);
 352     if (!client->exp) {
 353         LOG("export not found");
 354         goto fail;
 355     }
 356
 357     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 358     nbd_export_get(client->exp);
 359     rc = 0;
 360 fail:
 361     return rc;
 362 }
 363
 364 /* Handle NBD_OPT_STARTTLS. Return NULL to drop connection, or else the
 365  * new channel for all further (now-encrypted) communication. */
 366 static QIOChannel *nbd_negotiate_handle_starttls(NBDClient *client,
 367                                                  uint32_t length)
 368 {
 369     QIOChannel *ioc;
 370     QIOChannelTLS *tioc;
 371     struct NBDTLSHandshakeData data = { 0 };
 372
 373     TRACE("Setting up TLS");
 374     ioc = client->ioc;
 375     if (length) {
 376         if (nbd_negotiate_drop_sync(ioc, length) != length) {
 377             return NULL;
 378         }
 379         nbd_negotiate_send_rep_err(ioc, NBD_REP_ERR_INVALID, NBD_OPT_STARTTLS,
 380                                    "OPT_STARTTLS should not have length");
 381         return NULL;
 382     }
 383
 384     if (nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK,
 385                                NBD_OPT_STARTTLS) < 0) {
 386         return NULL;
 387     }
 388
 389     tioc = qio_channel_tls_new_server(ioc,
 390                                       client->tlscreds,
 391                                       client->tlsaclname,
 392                                       NULL);
 393     if (!tioc) {
 394         return NULL;
 395     }
 396
 397     qio_channel_set_name(QIO_CHANNEL(tioc), "nbd-server-tls");
 398     TRACE("Starting TLS handshake");
 399     data.loop = g_main_loop_new(g_main_context_default(), FALSE);
 400     qio_channel_tls_handshake(tioc,
 401                               nbd_tls_handshake,
 402                               &data,
 403                               NULL);
 404
 405     if (!data.complete) {
 406         g_main_loop_run(data.loop);
 407     }
 408     g_main_loop_unref(data.loop);
 409     if (data.error) {
 410         object_unref(OBJECT(tioc));
 411         error_free(data.error);
 412         return NULL;
 413     }
 414
 415     return QIO_CHANNEL(tioc);
 416 }
 417
 418
 419 /* Process all NBD_OPT_* client option commands.
 420  * Return -errno on error, 0 on success. */
 421 static int nbd_negotiate_options(NBDClient *client)
 422 {
 423     uint32_t flags;
 424     bool fixedNewstyle = false;
 425
 426     /* Client sends:
 427         [ 0 ..   3]   client flags
 428
 429         [ 0 ..   7]   NBD_OPTS_MAGIC
 430         [ 8 ..  11]   NBD option
 431         [12 ..  15]   Data length
 432         ...           Rest of request
 433
 434         [ 0 ..   7]   NBD_OPTS_MAGIC
 435         [ 8 ..  11]   Second NBD option
 436         [12 ..  15]   Data length
 437         ...           Rest of request
 438     */
 439
 440     if (nbd_negotiate_read(client->ioc, &flags, sizeof(flags)) !=
 441         sizeof(flags)) {
 442         LOG("read failed");
 443         return -EIO;
 444     }
 445     TRACE("Checking client flags");
 446     be32_to_cpus(&flags);
 447     if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) {
 448         TRACE("Client supports fixed newstyle handshake");
 449         fixedNewstyle = true;
 450         flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE;
 451     }
 452     if (flags & NBD_FLAG_C_NO_ZEROES) {
 453         TRACE("Client supports no zeroes at handshake end");
 454         client->no_zeroes = true;
 455         flags &= ~NBD_FLAG_C_NO_ZEROES;
 456     }
 457     if (flags != 0) {
 458         TRACE("Unknown client flags 0x%" PRIx32 " received", flags);
 459         return -EIO;
 460     }
 461
 462     while (1) {
 463         int ret;
 464         uint32_t clientflags, length;
 465         uint64_t magic;
 466
 467         if (nbd_negotiate_read(client->ioc, &magic, sizeof(magic)) !=
 468             sizeof(magic)) {
 469             LOG("read failed");
 470             return -EINVAL;
 471         }
 472         TRACE("Checking opts magic");
 473         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 474             LOG("Bad magic received");
 475             return -EINVAL;
 476         }
 477
 478         if (nbd_negotiate_read(client->ioc, &clientflags,
 479                                sizeof(clientflags)) != sizeof(clientflags)) {
 480             LOG("read failed");
 481             return -EINVAL;
 482         }
 483         clientflags = be32_to_cpu(clientflags);
 484
 485         if (nbd_negotiate_read(client->ioc, &length, sizeof(length)) !=
 486             sizeof(length)) {
 487             LOG("read failed");
 488             return -EINVAL;
 489         }
 490         length = be32_to_cpu(length);
 491
 492         TRACE("Checking option 0x%" PRIx32, clientflags);
 493         if (client->tlscreds &&
 494             client->ioc == (QIOChannel *)client->sioc) {
 495             QIOChannel *tioc;
 496             if (!fixedNewstyle) {
 497                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 498                 return -EINVAL;
 499             }
 500             switch (clientflags) {
 501             case NBD_OPT_STARTTLS:
 502                 tioc = nbd_negotiate_handle_starttls(client, length);
 503                 if (!tioc) {
 504                     return -EIO;
 505                 }
 506                 object_unref(OBJECT(client->ioc));
 507                 client->ioc = QIO_CHANNEL(tioc);
 508                 break;
 509
 510             case NBD_OPT_EXPORT_NAME:
 511                 /* No way to return an error to client, so drop connection */
 512                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 513                 return -EINVAL;
 514
 515             default:
 516                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 517                     return -EIO;
 518                 }
 519                 ret = nbd_negotiate_send_rep_err(client->ioc,
 520                                                  NBD_REP_ERR_TLS_REQD,
 521                                                  clientflags,
 522                                                  "Option 0x%" PRIx32
 523                                                  "not permitted before TLS",
 524                                                  clientflags);
 525                 if (ret < 0) {
 526                     return ret;
 527                 }
 528                 /* Let the client keep trying, unless they asked to quit */
 529                 if (clientflags == NBD_OPT_ABORT) {
 530                     return -EINVAL;
 531                 }
 532                 break;
 533             }
 534         } else if (fixedNewstyle) {
 535             switch (clientflags) {
 536             case NBD_OPT_LIST:
 537                 ret = nbd_negotiate_handle_list(client, length);
 538                 if (ret < 0) {
 539                     return ret;
 540                 }
 541                 break;
 542
 543             case NBD_OPT_ABORT:
 544                 /* NBD spec says we must try to reply before
 545                  * disconnecting, but that we must also tolerate
 546                  * guests that don't wait for our reply. */
 547                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, clientflags);
 548                 return -EINVAL;
 549
 550             case NBD_OPT_EXPORT_NAME:
 551                 return nbd_negotiate_handle_export_name(client, length);
 552
 553             case NBD_OPT_STARTTLS:
 554                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 555                     return -EIO;
 556                 }
 557                 if (client->tlscreds) {
 558                     ret = nbd_negotiate_send_rep_err(client->ioc,
 559                                                      NBD_REP_ERR_INVALID,
 560                                                      clientflags,
 561                                                      "TLS already enabled");
 562                 } else {
 563                     ret = nbd_negotiate_send_rep_err(client->ioc,
 564                                                      NBD_REP_ERR_POLICY,
 565                                                      clientflags,
 566                                                      "TLS not configured");
 567                 }
 568                 if (ret < 0) {
 569                     return ret;
 570                 }
 571                 break;
 572             default:
 573                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 574                     return -EIO;
 575                 }
 576                 ret = nbd_negotiate_send_rep_err(client->ioc,
 577                                                  NBD_REP_ERR_UNSUP,
 578                                                  clientflags,
 579                                                  "Unsupported option 0x%"
 580                                                  PRIx32,
 581                                                  clientflags);
 582                 if (ret < 0) {
 583                     return ret;
 584                 }
 585                 break;
 586             }
 587         } else {
 588             /*
 589              * If broken new-style we should drop the connection
 590              * for anything except NBD_OPT_EXPORT_NAME
 591              */
 592             switch (clientflags) {
 593             case NBD_OPT_EXPORT_NAME:
 594                 return nbd_negotiate_handle_export_name(client, length);
 595
 596             default:
 597                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 598                 return -EINVAL;
 599             }
 600         }
 601     }
 602 }
 603
 604 typedef struct {
 605     NBDClient *client;
 606     Coroutine *co;
 607 } NBDClientNewData;
 608
 609 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 610 {
 611     NBDClient *client = data->client;
 612     char buf[8 + 8 + 8 + 128];
 613     int rc;
 614     const uint16_t myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 615                               NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA |
 616                               NBD_FLAG_SEND_WRITE_ZEROES);
 617     bool oldStyle;
 618     size_t len;
 619
 620     /* Old style negotiation header without options
 621         [ 0 ..   7]   passwd       ("NBDMAGIC")
 622         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 623         [16 ..  23]   size
 624         [24 ..  25]   server flags (0)
 625         [26 ..  27]   export flags
 626         [28 .. 151]   reserved     (0)
 627
 628        New style negotiation header with options
 629         [ 0 ..   7]   passwd       ("NBDMAGIC")
 630         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 631         [16 ..  17]   server flags (0)
 632         ....options sent....
 633         [18 ..  25]   size
 634         [26 ..  27]   export flags
 635         [28 .. 151]   reserved     (0, omit if no_zeroes)
 636      */
 637
 638     qio_channel_set_blocking(client->ioc, false, NULL);
 639     rc = -EINVAL;
 640
 641     TRACE("Beginning negotiation.");
 642     memset(buf, 0, sizeof(buf));
 643     memcpy(buf, "NBDMAGIC", 8);
 644
 645     oldStyle = client->exp != NULL && !client->tlscreds;
 646     if (oldStyle) {
 647         TRACE("advertising size %" PRIu64 " and flags %x",
 648               client->exp->size, client->exp->nbdflags | myflags);
 649         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 650         stq_be_p(buf + 16, client->exp->size);
 651         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 652     } else {
 653         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 654         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE | NBD_FLAG_NO_ZEROES);
 655     }
 656
 657     if (oldStyle) {
 658         if (client->tlscreds) {
 659             TRACE("TLS cannot be enabled with oldstyle protocol");
 660             goto fail;
 661         }
 662         if (nbd_negotiate_write(client->ioc, buf, sizeof(buf)) != sizeof(buf)) {
 663             LOG("write failed");
 664             goto fail;
 665         }
 666     } else {
 667         if (nbd_negotiate_write(client->ioc, buf, 18) != 18) {
 668             LOG("write failed");
 669             goto fail;
 670         }
 671         rc = nbd_negotiate_options(client);
 672         if (rc != 0) {
 673             LOG("option negotiation failed");
 674             goto fail;
 675         }
 676
 677         TRACE("advertising size %" PRIu64 " and flags %x",
 678               client->exp->size, client->exp->nbdflags | myflags);
 679         stq_be_p(buf + 18, client->exp->size);
 680         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 681         len = client->no_zeroes ? 10 : sizeof(buf) - 18;
 682         if (nbd_negotiate_write(client->ioc, buf + 18, len) != len) {
 683             LOG("write failed");
 684             goto fail;
 685         }
 686     }
 687
 688     TRACE("Negotiation succeeded.");
 689     rc = 0;
 690 fail:
 691     return rc;
 692 }
 693
 694 static ssize_t nbd_receive_request(QIOChannel *ioc, NBDRequest *request)
 695 {
 696     uint8_t buf[NBD_REQUEST_SIZE];
 697     uint32_t magic;
 698     ssize_t ret;
 699
 700     ret = read_sync(ioc, buf, sizeof(buf));
 701     if (ret < 0) {
 702         return ret;
 703     }
 704
 705     if (ret != sizeof(buf)) {
 706         LOG("read failed");
 707         return -EINVAL;
 708     }
 709
 710     /* Request
 711        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 712        [ 4 ..  5]   flags   (NBD_CMD_FLAG_FUA, ...)
 713        [ 6 ..  7]   type    (NBD_CMD_READ, ...)
 714        [ 8 .. 15]   handle
 715        [16 .. 23]   from
 716        [24 .. 27]   len
 717      */
 718
 719     magic = ldl_be_p(buf);
 720     request->flags  = lduw_be_p(buf + 4);
 721     request->type   = lduw_be_p(buf + 6);
 722     request->handle = ldq_be_p(buf + 8);
 723     request->from   = ldq_be_p(buf + 16);
 724     request->len    = ldl_be_p(buf + 24);
 725
 726     TRACE("Got request: { magic = 0x%" PRIx32 ", .flags = %" PRIx16
 727           ", .type = %" PRIx16 ", from = %" PRIu64 ", len = %" PRIu32 " }",
 728           magic, request->flags, request->type, request->from, request->len);
 729
 730     if (magic != NBD_REQUEST_MAGIC) {
 731         LOG("invalid magic (got 0x%" PRIx32 ")", magic);
 732         return -EINVAL;
 733     }
 734     return 0;
 735 }
 736
 737 static ssize_t nbd_send_reply(QIOChannel *ioc, NBDReply *reply)
 738 {
 739     uint8_t buf[NBD_REPLY_SIZE];
 740     ssize_t ret;
 741
 742     reply->error = system_errno_to_nbd_errno(reply->error);
 743
 744     TRACE("Sending response to client: { .error = %" PRId32
 745           ", handle = %" PRIu64 " }",
 746           reply->error, reply->handle);
 747
 748     /* Reply
 749        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 750        [ 4 ..  7]    error   (0 == no error)
 751        [ 7 .. 15]    handle
 752      */
 753     stl_be_p(buf, NBD_REPLY_MAGIC);
 754     stl_be_p(buf + 4, reply->error);
 755     stq_be_p(buf + 8, reply->handle);
 756
 757     ret = write_sync(ioc, buf, sizeof(buf));
 758     if (ret < 0) {
 759         return ret;
 760     }
 761
 762     if (ret != sizeof(buf)) {
 763         LOG("writing to socket failed");
 764         return -EINVAL;
 765     }
 766     return 0;
 767 }
 768
 769 #define MAX_NBD_REQUESTS 16
 770
 771 void nbd_client_get(NBDClient *client)
 772 {
 773     client->refcount++;
 774 }
 775
 776 void nbd_client_put(NBDClient *client)
 777 {
 778     if (--client->refcount == 0) {
 779         /* The last reference should be dropped by client->close,
 780          * which is called by client_close.
 781          */
 782         assert(client->closing);
 783
 784         qio_channel_detach_aio_context(client->ioc);
 785         object_unref(OBJECT(client->sioc));
 786         object_unref(OBJECT(client->ioc));
 787         if (client->tlscreds) {
 788             object_unref(OBJECT(client->tlscreds));
 789         }
 790         g_free(client->tlsaclname);
 791         if (client->exp) {
 792             QTAILQ_REMOVE(&client->exp->clients, client, next);
 793             nbd_export_put(client->exp);
 794         }
 795         g_free(client);
 796     }
 797 }
 798
 799 static void client_close(NBDClient *client)
 800 {
 801     if (client->closing) {
 802         return;
 803     }
 804
 805     client->closing = true;
 806
 807     /* Force requests to finish.  They will drop their own references,
 808      * then we'll close the socket and free the NBDClient.
 809      */
 810     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 811                          NULL);
 812
 813     /* Also tell the client, so that they release their reference.  */
 814     if (client->close) {
 815         client->close(client);
 816     }
 817 }
 818
 819 static NBDRequestData *nbd_request_get(NBDClient *client)
 820 {
 821     NBDRequestData *req;
 822
 823     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 824     client->nb_requests++;
 825
 826     req = g_new0(NBDRequestData, 1);
 827     nbd_client_get(client);
 828     req->client = client;
 829     return req;
 830 }
 831
 832 static void nbd_request_put(NBDRequestData *req)
 833 {
 834     NBDClient *client = req->client;
 835
 836     if (req->data) {
 837         qemu_vfree(req->data);
 838     }
 839     g_free(req);
 840
 841     client->nb_requests--;
 842     nbd_client_receive_next_request(client);
 843
 844     nbd_client_put(client);
 845 }
 846
 847 static void blk_aio_attached(AioContext *ctx, void *opaque)
 848 {
 849     NBDExport *exp = opaque;
 850     NBDClient *client;
 851
 852     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 853
 854     exp->ctx = ctx;
 855
 856     QTAILQ_FOREACH(client, &exp->clients, next) {
 857         qio_channel_attach_aio_context(client->ioc, ctx);
 858         if (client->recv_coroutine) {
 859             aio_co_schedule(ctx, client->recv_coroutine);
 860         }
 861         if (client->send_coroutine) {
 862             aio_co_schedule(ctx, client->send_coroutine);
 863         }
 864     }
 865 }
 866
 867 static void blk_aio_detach(void *opaque)
 868 {
 869     NBDExport *exp = opaque;
 870     NBDClient *client;
 871
 872     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 873
 874     QTAILQ_FOREACH(client, &exp->clients, next) {
 875         qio_channel_detach_aio_context(client->ioc);
 876     }
 877
 878     exp->ctx = NULL;
 879 }
 880
 881 static void nbd_eject_notifier(Notifier *n, void *data)
 882 {
 883     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 884     nbd_export_close(exp);
 885 }
 886
 887 NBDExport *nbd_export_new(BlockDriverState *bs, off_t dev_offset, off_t size,
 888                           uint16_t nbdflags, void (*close)(NBDExport *),
 889                           bool writethrough, BlockBackend *on_eject_blk,
 890                           Error **errp)
 891 {
 892     BlockBackend *blk;
 893     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 894
 895     /* FIXME Use real permissions */
 896     blk = blk_new(0, BLK_PERM_ALL);
 897     blk_insert_bs(blk, bs);
 898     blk_set_enable_write_cache(blk, !writethrough);
 899
 900     exp->refcount = 1;
 901     QTAILQ_INIT(&exp->clients);
 902     exp->blk = blk;
 903     exp->dev_offset = dev_offset;
 904     exp->nbdflags = nbdflags;
 905     exp->size = size < 0 ? blk_getlength(blk) : size;
 906     if (exp->size < 0) {
 907         error_setg_errno(errp, -exp->size,
 908                          "Failed to determine the NBD export's length");
 909         goto fail;
 910     }
 911     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 912
 913     exp->close = close;
 914     exp->ctx = blk_get_aio_context(blk);
 915     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 916
 917     if (on_eject_blk) {
 918         blk_ref(on_eject_blk);
 919         exp->eject_notifier_blk = on_eject_blk;
 920         exp->eject_notifier.notify = nbd_eject_notifier;
 921         blk_add_remove_bs_notifier(on_eject_blk, &exp->eject_notifier);
 922     }
 923
 924     /*
 925      * NBD exports are used for non-shared storage migration.  Make sure
 926      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 927      * access since the export could be available before migration handover.
 928      */
 929     aio_context_acquire(exp->ctx);
 930     blk_invalidate_cache(blk, NULL);
 931     aio_context_release(exp->ctx);
 932     return exp;
 933
 934 fail:
 935     blk_unref(blk);
 936     g_free(exp);
 937     return NULL;
 938 }
 939
 940 NBDExport *nbd_export_find(const char *name)
 941 {
 942     NBDExport *exp;
 943     QTAILQ_FOREACH(exp, &exports, next) {
 944         if (strcmp(name, exp->name) == 0) {
 945             return exp;
 946         }
 947     }
 948
 949     return NULL;
 950 }
 951
 952 void nbd_export_set_name(NBDExport *exp, const char *name)
 953 {
 954     if (exp->name == name) {
 955         return;
 956     }
 957
 958     nbd_export_get(exp);
 959     if (exp->name != NULL) {
 960         g_free(exp->name);
 961         exp->name = NULL;
 962         QTAILQ_REMOVE(&exports, exp, next);
 963         nbd_export_put(exp);
 964     }
 965     if (name != NULL) {
 966         nbd_export_get(exp);
 967         exp->name = g_strdup(name);
 968         QTAILQ_INSERT_TAIL(&exports, exp, next);
 969     }
 970     nbd_export_put(exp);
 971 }
 972
 973 void nbd_export_set_description(NBDExport *exp, const char *description)
 974 {
 975     g_free(exp->description);
 976     exp->description = g_strdup(description);
 977 }
 978
 979 void nbd_export_close(NBDExport *exp)
 980 {
 981     NBDClient *client, *next;
 982
 983     nbd_export_get(exp);
 984     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 985         client_close(client);
 986     }
 987     nbd_export_set_name(exp, NULL);
 988     nbd_export_set_description(exp, NULL);
 989     nbd_export_put(exp);
 990 }
 991
 992 void nbd_export_get(NBDExport *exp)
 993 {
 994     assert(exp->refcount > 0);
 995     exp->refcount++;
 996 }
 997
 998 void nbd_export_put(NBDExport *exp)
 999 {
1000     assert(exp->refcount > 0);
1001     if (exp->refcount == 1) {
1002         nbd_export_close(exp);
1003     }
1004
1005     if (--exp->refcount == 0) {
1006         assert(exp->name == NULL);
1007         assert(exp->description == NULL);
1008
1009         if (exp->close) {
1010             exp->close(exp);
1011         }
1012
1013         if (exp->blk) {
1014             if (exp->eject_notifier_blk) {
1015                 notifier_remove(&exp->eject_notifier);
1016                 blk_unref(exp->eject_notifier_blk);
1017             }
1018             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
1019                                             blk_aio_detach, exp);
1020             blk_unref(exp->blk);
1021             exp->blk = NULL;
1022         }
1023
1024         g_free(exp);
1025     }
1026 }
1027
1028 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
1029 {
1030     return exp->blk;
1031 }
1032
1033 void nbd_export_close_all(void)
1034 {
1035     NBDExport *exp, *next;
1036
1037     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
1038         nbd_export_close(exp);
1039     }
1040 }
1041
1042 static ssize_t nbd_co_send_reply(NBDRequestData *req, NBDReply *reply,
1043                                  int len)
1044 {
1045     NBDClient *client = req->client;
1046     ssize_t rc, ret;
1047
1048     g_assert(qemu_in_coroutine());
1049     qemu_co_mutex_lock(&client->send_lock);
1050     client->send_coroutine = qemu_coroutine_self();
1051
1052     if (!len) {
1053         rc = nbd_send_reply(client->ioc, reply);
1054     } else {
1055         qio_channel_set_cork(client->ioc, true);
1056         rc = nbd_send_reply(client->ioc, reply);
1057         if (rc >= 0) {
1058             ret = write_sync(client->ioc, req->data, len);
1059             if (ret != len) {
1060                 rc = -EIO;
1061             }
1062         }
1063         qio_channel_set_cork(client->ioc, false);
1064     }
1065
1066     client->send_coroutine = NULL;
1067     qemu_co_mutex_unlock(&client->send_lock);
1068     return rc;
1069 }
1070
1071 /* Collect a client request.  Return 0 if request looks valid, -EAGAIN
1072  * to keep trying the collection, -EIO to drop connection right away,
1073  * and any other negative value to report an error to the client
1074  * (although the caller may still need to disconnect after reporting
1075  * the error).  */
1076 static ssize_t nbd_co_receive_request(NBDRequestData *req,
1077                                       NBDRequest *request)
1078 {
1079     NBDClient *client = req->client;
1080     ssize_t rc;
1081
1082     g_assert(qemu_in_coroutine());
1083     assert(client->recv_coroutine == qemu_coroutine_self());
1084     rc = nbd_receive_request(client->ioc, request);
1085     if (rc < 0) {
1086         if (rc != -EAGAIN) {
1087             rc = -EIO;
1088         }
1089         goto out;
1090     }
1091
1092     TRACE("Decoding type");
1093
1094     if (request->type != NBD_CMD_WRITE) {
1095         /* No payload, we are ready to read the next request.  */
1096         req->complete = true;
1097     }
1098
1099     if (request->type == NBD_CMD_DISC) {
1100         /* Special case: we're going to disconnect without a reply,
1101          * whether or not flags, from, or len are bogus */
1102         TRACE("Request type is DISCONNECT");
1103         rc = -EIO;
1104         goto out;
1105     }
1106
1107     /* Check for sanity in the parameters, part 1.  Defer as many
1108      * checks as possible until after reading any NBD_CMD_WRITE
1109      * payload, so we can try and keep the connection alive.  */
1110     if ((request->from + request->len) < request->from) {
1111         LOG("integer overflow detected, you're probably being attacked");
1112         rc = -EINVAL;
1113         goto out;
1114     }
1115
1116     if (request->type == NBD_CMD_READ || request->type == NBD_CMD_WRITE) {
1117         if (request->len > NBD_MAX_BUFFER_SIZE) {
1118             LOG("len (%" PRIu32" ) is larger than max len (%u)",
1119                 request->len, NBD_MAX_BUFFER_SIZE);
1120             rc = -EINVAL;
1121             goto out;
1122         }
1123
1124         req->data = blk_try_blockalign(client->exp->blk, request->len);
1125         if (req->data == NULL) {
1126             rc = -ENOMEM;
1127             goto out;
1128         }
1129     }
1130     if (request->type == NBD_CMD_WRITE) {
1131         TRACE("Reading %" PRIu32 " byte(s)", request->len);
1132
1133         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1134             LOG("reading from socket failed");
1135             rc = -EIO;
1136             goto out;
1137         }
1138         req->complete = true;
1139     }
1140
1141     /* Sanity checks, part 2. */
1142     if (request->from + request->len > client->exp->size) {
1143         LOG("operation past EOF; From: %" PRIu64 ", Len: %" PRIu32
1144             ", Size: %" PRIu64, request->from, request->len,
1145             (uint64_t)client->exp->size);
1146         rc = request->type == NBD_CMD_WRITE ? -ENOSPC : -EINVAL;
1147         goto out;
1148     }
1149     if (request->flags & ~(NBD_CMD_FLAG_FUA | NBD_CMD_FLAG_NO_HOLE)) {
1150         LOG("unsupported flags (got 0x%x)", request->flags);
1151         rc = -EINVAL;
1152         goto out;
1153     }
1154     if (request->type != NBD_CMD_WRITE_ZEROES &&
1155         (request->flags & NBD_CMD_FLAG_NO_HOLE)) {
1156         LOG("unexpected flags (got 0x%x)", request->flags);
1157         rc = -EINVAL;
1158         goto out;
1159     }
1160
1161     rc = 0;
1162
1163 out:
1164     client->recv_coroutine = NULL;
1165     nbd_client_receive_next_request(client);
1166
1167     return rc;
1168 }
1169
1170 /* Owns a reference to the NBDClient passed as opaque.  */
1171 static coroutine_fn void nbd_trip(void *opaque)
1172 {
1173     NBDClient *client = opaque;
1174     NBDExport *exp = client->exp;
1175     NBDRequestData *req;
1176     NBDRequest request = { 0 };    /* GCC thinks it can be used uninitialized */
1177     NBDReply reply;
1178     ssize_t ret;
1179     int flags;
1180
1181     TRACE("Reading request.");
1182     if (client->closing) {
1183         nbd_client_put(client);
1184         return;
1185     }
1186
1187     req = nbd_request_get(client);
1188     ret = nbd_co_receive_request(req, &request);
1189     if (ret == -EAGAIN) {
1190         goto done;
1191     }
1192     if (ret == -EIO) {
1193         goto out;
1194     }
1195
1196     reply.handle = request.handle;
1197     reply.error = 0;
1198
1199     if (ret < 0) {
1200         reply.error = -ret;
1201         goto error_reply;
1202     }
1203
1204     if (client->closing) {
1205         /*
1206          * The client may be closed when we are blocked in
1207          * nbd_co_receive_request()
1208          */
1209         goto done;
1210     }
1211
1212     switch (request.type) {
1213     case NBD_CMD_READ:
1214         TRACE("Request type is READ");
1215
1216         /* XXX: NBD Protocol only documents use of FUA with WRITE */
1217         if (request.flags & NBD_CMD_FLAG_FUA) {
1218             ret = blk_co_flush(exp->blk);
1219             if (ret < 0) {
1220                 LOG("flush failed");
1221                 reply.error = -ret;
1222                 goto error_reply;
1223             }
1224         }
1225
1226         ret = blk_pread(exp->blk, request.from + exp->dev_offset,
1227                         req->data, request.len);
1228         if (ret < 0) {
1229             LOG("reading from file failed");
1230             reply.error = -ret;
1231             goto error_reply;
1232         }
1233
1234         TRACE("Read %" PRIu32" byte(s)", request.len);
1235         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1236             goto out;
1237         break;
1238     case NBD_CMD_WRITE:
1239         TRACE("Request type is WRITE");
1240
1241         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1242             TRACE("Server is read-only, return error");
1243             reply.error = EROFS;
1244             goto error_reply;
1245         }
1246
1247         TRACE("Writing to device");
1248
1249         flags = 0;
1250         if (request.flags & NBD_CMD_FLAG_FUA) {
1251             flags |= BDRV_REQ_FUA;
1252         }
1253         ret = blk_pwrite(exp->blk, request.from + exp->dev_offset,
1254                          req->data, request.len, flags);
1255         if (ret < 0) {
1256             LOG("writing to file failed");
1257             reply.error = -ret;
1258             goto error_reply;
1259         }
1260
1261         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1262             goto out;
1263         }
1264         break;
1265
1266     case NBD_CMD_WRITE_ZEROES:
1267         TRACE("Request type is WRITE_ZEROES");
1268
1269         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1270             TRACE("Server is read-only, return error");
1271             reply.error = EROFS;
1272             goto error_reply;
1273         }
1274
1275         TRACE("Writing to device");
1276
1277         flags = 0;
1278         if (request.flags & NBD_CMD_FLAG_FUA) {
1279             flags |= BDRV_REQ_FUA;
1280         }
1281         if (!(request.flags & NBD_CMD_FLAG_NO_HOLE)) {
1282             flags |= BDRV_REQ_MAY_UNMAP;
1283         }
1284         ret = blk_pwrite_zeroes(exp->blk, request.from + exp->dev_offset,
1285                                 request.len, flags);
1286         if (ret < 0) {
1287             LOG("writing to file failed");
1288             reply.error = -ret;
1289             goto error_reply;
1290         }
1291
1292         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1293             goto out;
1294         }
1295         break;
1296
1297     case NBD_CMD_DISC:
1298         /* unreachable, thanks to special case in nbd_co_receive_request() */
1299         abort();
1300
1301     case NBD_CMD_FLUSH:
1302         TRACE("Request type is FLUSH");
1303
1304         ret = blk_co_flush(exp->blk);
1305         if (ret < 0) {
1306             LOG("flush failed");
1307             reply.error = -ret;
1308         }
1309         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1310             goto out;
1311         }
1312         break;
1313     case NBD_CMD_TRIM:
1314         TRACE("Request type is TRIM");
1315         ret = blk_co_pdiscard(exp->blk, request.from + exp->dev_offset,
1316                               request.len);
1317         if (ret < 0) {
1318             LOG("discard failed");
1319             reply.error = -ret;
1320         }
1321         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1322             goto out;
1323         }
1324         break;
1325     default:
1326         LOG("invalid request type (%" PRIu32 ") received", request.type);
1327         reply.error = EINVAL;
1328     error_reply:
1329         /* We must disconnect after NBD_CMD_WRITE if we did not
1330          * read the payload.
1331          */
1332         if (nbd_co_send_reply(req, &reply, 0) < 0 || !req->complete) {
1333             goto out;
1334         }
1335         break;
1336     }
1337
1338     TRACE("Request/Reply complete");
1339
1340 done:
1341     nbd_request_put(req);
1342     nbd_client_put(client);
1343     return;
1344
1345 out:
1346     nbd_request_put(req);
1347     client_close(client);
1348     nbd_client_put(client);
1349 }
1350
1351 static void nbd_client_receive_next_request(NBDClient *client)
1352 {
1353     if (!client->recv_coroutine && client->nb_requests < MAX_NBD_REQUESTS) {
1354         nbd_client_get(client);
1355         client->recv_coroutine = qemu_coroutine_create(nbd_trip, client);
1356         aio_co_schedule(client->exp->ctx, client->recv_coroutine);
1357     }
1358 }
1359
1360 static coroutine_fn void nbd_co_client_start(void *opaque)
1361 {
1362     NBDClientNewData *data = opaque;
1363     NBDClient *client = data->client;
1364     NBDExport *exp = client->exp;
1365
1366     if (exp) {
1367         nbd_export_get(exp);
1368     }
1369     if (nbd_negotiate(data)) {
1370         client_close(client);
1371         goto out;
1372     }
1373     qemu_co_mutex_init(&client->send_lock);
1374
1375     if (exp) {
1376         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1377     }
1378
1379     nbd_client_receive_next_request(client);
1380
1381 out:
1382     g_free(data);
1383 }
1384
1385 void nbd_client_new(NBDExport *exp,
1386                     QIOChannelSocket *sioc,
1387                     QCryptoTLSCreds *tlscreds,
1388                     const char *tlsaclname,
1389                     void (*close_fn)(NBDClient *))
1390 {
1391     NBDClient *client;
1392     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1393
1394     client = g_malloc0(sizeof(NBDClient));
1395     client->refcount = 1;
1396     client->exp = exp;
1397     client->tlscreds = tlscreds;
1398     if (tlscreds) {
1399         object_ref(OBJECT(client->tlscreds));
1400     }
1401     client->tlsaclname = g_strdup(tlsaclname);
1402     client->sioc = sioc;
1403     object_ref(OBJECT(client->sioc));
1404     client->ioc = QIO_CHANNEL(sioc);
1405     object_ref(OBJECT(client->ioc));
1406     client->close = close_fn;
1407
1408     data->client = client;
1409     data->co = qemu_coroutine_create(nbd_co_client_start, data);
1410     qemu_coroutine_enter(data->co);
1411 }