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 EINVAL:
  43     default:
  44         return NBD_EINVAL;
  45     }
  46 }
  47
  48 /* Definitions for opaque data types */
  49
  50 typedef struct NBDRequestData NBDRequestData;
  51
  52 struct NBDRequestData {
  53     QSIMPLEQ_ENTRY(NBDRequestData) entry;
  54     NBDClient *client;
  55     uint8_t *data;
  56     bool complete;
  57 };
  58
  59 struct NBDExport {
  60     int refcount;
  61     void (*close)(NBDExport *exp);
  62
  63     BlockBackend *blk;
  64     char *name;
  65     char *description;
  66     off_t dev_offset;
  67     off_t size;
  68     uint16_t nbdflags;
  69     QTAILQ_HEAD(, NBDClient) clients;
  70     QTAILQ_ENTRY(NBDExport) next;
  71
  72     AioContext *ctx;
  73
  74     BlockBackend *eject_notifier_blk;
  75     Notifier eject_notifier;
  76 };
  77
  78 static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports);
  79
  80 struct NBDClient {
  81     int refcount;
  82     void (*close)(NBDClient *client);
  83
  84     NBDExport *exp;
  85     QCryptoTLSCreds *tlscreds;
  86     char *tlsaclname;
  87     QIOChannelSocket *sioc; /* The underlying data channel */
  88     QIOChannel *ioc; /* The current I/O channel which may differ (eg TLS) */
  89
  90     Coroutine *recv_coroutine;
  91
  92     CoMutex send_lock;
  93     Coroutine *send_coroutine;
  94
  95     bool can_read;
  96
  97     QTAILQ_ENTRY(NBDClient) next;
  98     int nb_requests;
  99     bool closing;
 100 };
 101
 102 /* That's all folks */
 103
 104 static void nbd_set_handlers(NBDClient *client);
 105 static void nbd_unset_handlers(NBDClient *client);
 106 static void nbd_update_can_read(NBDClient *client);
 107
 108 static gboolean nbd_negotiate_continue(QIOChannel *ioc,
 109                                        GIOCondition condition,
 110                                        void *opaque)
 111 {
 112     qemu_coroutine_enter(opaque);
 113     return TRUE;
 114 }
 115
 116 static ssize_t nbd_negotiate_read(QIOChannel *ioc, void *buffer, size_t size)
 117 {
 118     ssize_t ret;
 119     guint watch;
 120
 121     assert(qemu_in_coroutine());
 122     /* Negotiation are always in main loop. */
 123     watch = qio_channel_add_watch(ioc,
 124                                   G_IO_IN,
 125                                   nbd_negotiate_continue,
 126                                   qemu_coroutine_self(),
 127                                   NULL);
 128     ret = read_sync(ioc, buffer, size);
 129     g_source_remove(watch);
 130     return ret;
 131
 132 }
 133
 134 static ssize_t nbd_negotiate_write(QIOChannel *ioc, const void *buffer,
 135                                    size_t size)
 136 {
 137     ssize_t ret;
 138     guint watch;
 139
 140     assert(qemu_in_coroutine());
 141     /* Negotiation are always in main loop. */
 142     watch = qio_channel_add_watch(ioc,
 143                                   G_IO_OUT,
 144                                   nbd_negotiate_continue,
 145                                   qemu_coroutine_self(),
 146                                   NULL);
 147     ret = write_sync(ioc, buffer, size);
 148     g_source_remove(watch);
 149     return ret;
 150 }
 151
 152 static ssize_t nbd_negotiate_drop_sync(QIOChannel *ioc, size_t size)
 153 {
 154     ssize_t ret, dropped = size;
 155     uint8_t *buffer = g_malloc(MIN(65536, size));
 156
 157     while (size > 0) {
 158         ret = nbd_negotiate_read(ioc, buffer, MIN(65536, size));
 159         if (ret < 0) {
 160             g_free(buffer);
 161             return ret;
 162         }
 163
 164         assert(ret <= size);
 165         size -= ret;
 166     }
 167
 168     g_free(buffer);
 169     return dropped;
 170 }
 171
 172 /* Basic flow for negotiation
 173
 174    Server         Client
 175    Negotiate
 176
 177    or
 178
 179    Server         Client
 180    Negotiate #1
 181                   Option
 182    Negotiate #2
 183
 184    ----
 185
 186    followed by
 187
 188    Server         Client
 189                   Request
 190    Response
 191                   Request
 192    Response
 193                   ...
 194    ...
 195                   Request (type == 2)
 196
 197 */
 198
 199 /* Send a reply header, including length, but no payload.
 200  * Return -errno on error, 0 on success. */
 201 static int nbd_negotiate_send_rep_len(QIOChannel *ioc, uint32_t type,
 202                                       uint32_t opt, uint32_t len)
 203 {
 204     uint64_t magic;
 205
 206     TRACE("Reply opt=%" PRIx32 " type=%" PRIx32 " len=%" PRIu32,
 207           type, opt, len);
 208
 209     magic = cpu_to_be64(NBD_REP_MAGIC);
 210     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 211         LOG("write failed (rep magic)");
 212         return -EINVAL;
 213     }
 214     opt = cpu_to_be32(opt);
 215     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 216         LOG("write failed (rep opt)");
 217         return -EINVAL;
 218     }
 219     type = cpu_to_be32(type);
 220     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 221         LOG("write failed (rep type)");
 222         return -EINVAL;
 223     }
 224     len = cpu_to_be32(len);
 225     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 226         LOG("write failed (rep data length)");
 227         return -EINVAL;
 228     }
 229     return 0;
 230 }
 231
 232 /* Send a reply header with default 0 length.
 233  * Return -errno on error, 0 on success. */
 234 static int nbd_negotiate_send_rep(QIOChannel *ioc, uint32_t type, uint32_t opt)
 235 {
 236     return nbd_negotiate_send_rep_len(ioc, type, opt, 0);
 237 }
 238
 239 /* Send an error reply.
 240  * Return -errno on error, 0 on success. */
 241 static int GCC_FMT_ATTR(4, 5)
 242 nbd_negotiate_send_rep_err(QIOChannel *ioc, uint32_t type,
 243                            uint32_t opt, const char *fmt, ...)
 244 {
 245     va_list va;
 246     char *msg;
 247     int ret;
 248     size_t len;
 249
 250     va_start(va, fmt);
 251     msg = g_strdup_vprintf(fmt, va);
 252     va_end(va);
 253     len = strlen(msg);
 254     assert(len < 4096);
 255     TRACE("sending error message \"%s\"", msg);
 256     ret = nbd_negotiate_send_rep_len(ioc, type, opt, len);
 257     if (ret < 0) {
 258         goto out;
 259     }
 260     if (nbd_negotiate_write(ioc, msg, len) != len) {
 261         LOG("write failed (error message)");
 262         ret = -EIO;
 263     } else {
 264         ret = 0;
 265     }
 266 out:
 267     g_free(msg);
 268     return ret;
 269 }
 270
 271 /* Send a single NBD_REP_SERVER reply to NBD_OPT_LIST, including payload.
 272  * Return -errno on error, 0 on success. */
 273 static int nbd_negotiate_send_rep_list(QIOChannel *ioc, NBDExport *exp)
 274 {
 275     size_t name_len, desc_len;
 276     uint32_t len;
 277     const char *name = exp->name ? exp->name : "";
 278     const char *desc = exp->description ? exp->description : "";
 279     int rc;
 280
 281     TRACE("Advertising export name '%s' description '%s'", name, desc);
 282     name_len = strlen(name);
 283     desc_len = strlen(desc);
 284     len = name_len + desc_len + sizeof(len);
 285     rc = nbd_negotiate_send_rep_len(ioc, NBD_REP_SERVER, NBD_OPT_LIST, len);
 286     if (rc < 0) {
 287         return rc;
 288     }
 289
 290     len = cpu_to_be32(name_len);
 291     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 292         LOG("write failed (name length)");
 293         return -EINVAL;
 294     }
 295     if (nbd_negotiate_write(ioc, name, name_len) != name_len) {
 296         LOG("write failed (name buffer)");
 297         return -EINVAL;
 298     }
 299     if (nbd_negotiate_write(ioc, desc, desc_len) != desc_len) {
 300         LOG("write failed (description buffer)");
 301         return -EINVAL;
 302     }
 303     return 0;
 304 }
 305
 306 /* Process the NBD_OPT_LIST command, with a potential series of replies.
 307  * Return -errno on error, 0 on success. */
 308 static int nbd_negotiate_handle_list(NBDClient *client, uint32_t length)
 309 {
 310     NBDExport *exp;
 311
 312     if (length) {
 313         if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 314             return -EIO;
 315         }
 316         return nbd_negotiate_send_rep_err(client->ioc,
 317                                           NBD_REP_ERR_INVALID, NBD_OPT_LIST,
 318                                           "OPT_LIST should not have length");
 319     }
 320
 321     /* For each export, send a NBD_REP_SERVER reply. */
 322     QTAILQ_FOREACH(exp, &exports, next) {
 323         if (nbd_negotiate_send_rep_list(client->ioc, exp)) {
 324             return -EINVAL;
 325         }
 326     }
 327     /* Finish with a NBD_REP_ACK. */
 328     return nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_LIST);
 329 }
 330
 331 static int nbd_negotiate_handle_export_name(NBDClient *client, uint32_t length)
 332 {
 333     int rc = -EINVAL;
 334     char name[NBD_MAX_NAME_SIZE + 1];
 335
 336     /* Client sends:
 337         [20 ..  xx]   export name (length bytes)
 338      */
 339     TRACE("Checking length");
 340     if (length >= sizeof(name)) {
 341         LOG("Bad length received");
 342         goto fail;
 343     }
 344     if (nbd_negotiate_read(client->ioc, name, length) != length) {
 345         LOG("read failed");
 346         goto fail;
 347     }
 348     name[length] = '\0';
 349
 350     TRACE("Client requested export '%s'", name);
 351
 352     client->exp = nbd_export_find(name);
 353     if (!client->exp) {
 354         LOG("export not found");
 355         goto fail;
 356     }
 357
 358     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 359     nbd_export_get(client->exp);
 360     rc = 0;
 361 fail:
 362     return rc;
 363 }
 364
 365 /* Handle NBD_OPT_STARTTLS. Return NULL to drop connection, or else the
 366  * new channel for all further (now-encrypted) communication. */
 367 static QIOChannel *nbd_negotiate_handle_starttls(NBDClient *client,
 368                                                  uint32_t length)
 369 {
 370     QIOChannel *ioc;
 371     QIOChannelTLS *tioc;
 372     struct NBDTLSHandshakeData data = { 0 };
 373
 374     TRACE("Setting up TLS");
 375     ioc = client->ioc;
 376     if (length) {
 377         if (nbd_negotiate_drop_sync(ioc, length) != length) {
 378             return NULL;
 379         }
 380         nbd_negotiate_send_rep_err(ioc, NBD_REP_ERR_INVALID, NBD_OPT_STARTTLS,
 381                                    "OPT_STARTTLS should not have length");
 382         return NULL;
 383     }
 384
 385     if (nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK,
 386                                NBD_OPT_STARTTLS) < 0) {
 387         return NULL;
 388     }
 389
 390     tioc = qio_channel_tls_new_server(ioc,
 391                                       client->tlscreds,
 392                                       client->tlsaclname,
 393                                       NULL);
 394     if (!tioc) {
 395         return NULL;
 396     }
 397
 398     qio_channel_set_name(QIO_CHANNEL(tioc), "nbd-server-tls");
 399     TRACE("Starting TLS handshake");
 400     data.loop = g_main_loop_new(g_main_context_default(), FALSE);
 401     qio_channel_tls_handshake(tioc,
 402                               nbd_tls_handshake,
 403                               &data,
 404                               NULL);
 405
 406     if (!data.complete) {
 407         g_main_loop_run(data.loop);
 408     }
 409     g_main_loop_unref(data.loop);
 410     if (data.error) {
 411         object_unref(OBJECT(tioc));
 412         error_free(data.error);
 413         return NULL;
 414     }
 415
 416     return QIO_CHANNEL(tioc);
 417 }
 418
 419
 420 /* Process all NBD_OPT_* client option commands.
 421  * Return -errno on error, 0 on success. */
 422 static int nbd_negotiate_options(NBDClient *client)
 423 {
 424     uint32_t flags;
 425     bool fixedNewstyle = false;
 426
 427     /* Client sends:
 428         [ 0 ..   3]   client flags
 429
 430         [ 0 ..   7]   NBD_OPTS_MAGIC
 431         [ 8 ..  11]   NBD option
 432         [12 ..  15]   Data length
 433         ...           Rest of request
 434
 435         [ 0 ..   7]   NBD_OPTS_MAGIC
 436         [ 8 ..  11]   Second NBD option
 437         [12 ..  15]   Data length
 438         ...           Rest of request
 439     */
 440
 441     if (nbd_negotiate_read(client->ioc, &flags, sizeof(flags)) !=
 442         sizeof(flags)) {
 443         LOG("read failed");
 444         return -EIO;
 445     }
 446     TRACE("Checking client flags");
 447     be32_to_cpus(&flags);
 448     if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) {
 449         TRACE("Client supports fixed newstyle handshake");
 450         fixedNewstyle = true;
 451         flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE;
 452     }
 453     if (flags != 0) {
 454         TRACE("Unknown client flags 0x%" PRIx32 " received", flags);
 455         return -EIO;
 456     }
 457
 458     while (1) {
 459         int ret;
 460         uint32_t clientflags, length;
 461         uint64_t magic;
 462
 463         if (nbd_negotiate_read(client->ioc, &magic, sizeof(magic)) !=
 464             sizeof(magic)) {
 465             LOG("read failed");
 466             return -EINVAL;
 467         }
 468         TRACE("Checking opts magic");
 469         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 470             LOG("Bad magic received");
 471             return -EINVAL;
 472         }
 473
 474         if (nbd_negotiate_read(client->ioc, &clientflags,
 475                                sizeof(clientflags)) != sizeof(clientflags)) {
 476             LOG("read failed");
 477             return -EINVAL;
 478         }
 479         clientflags = be32_to_cpu(clientflags);
 480
 481         if (nbd_negotiate_read(client->ioc, &length, sizeof(length)) !=
 482             sizeof(length)) {
 483             LOG("read failed");
 484             return -EINVAL;
 485         }
 486         length = be32_to_cpu(length);
 487
 488         TRACE("Checking option 0x%" PRIx32, clientflags);
 489         if (client->tlscreds &&
 490             client->ioc == (QIOChannel *)client->sioc) {
 491             QIOChannel *tioc;
 492             if (!fixedNewstyle) {
 493                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 494                 return -EINVAL;
 495             }
 496             switch (clientflags) {
 497             case NBD_OPT_STARTTLS:
 498                 tioc = nbd_negotiate_handle_starttls(client, length);
 499                 if (!tioc) {
 500                     return -EIO;
 501                 }
 502                 object_unref(OBJECT(client->ioc));
 503                 client->ioc = QIO_CHANNEL(tioc);
 504                 break;
 505
 506             case NBD_OPT_EXPORT_NAME:
 507                 /* No way to return an error to client, so drop connection */
 508                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 509                 return -EINVAL;
 510
 511             default:
 512                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 513                     return -EIO;
 514                 }
 515                 ret = nbd_negotiate_send_rep_err(client->ioc,
 516                                                  NBD_REP_ERR_TLS_REQD,
 517                                                  clientflags,
 518                                                  "Option 0x%" PRIx32
 519                                                  "not permitted before TLS",
 520                                                  clientflags);
 521                 if (ret < 0) {
 522                     return ret;
 523                 }
 524                 break;
 525             }
 526         } else if (fixedNewstyle) {
 527             switch (clientflags) {
 528             case NBD_OPT_LIST:
 529                 ret = nbd_negotiate_handle_list(client, length);
 530                 if (ret < 0) {
 531                     return ret;
 532                 }
 533                 break;
 534
 535             case NBD_OPT_ABORT:
 536                 return -EINVAL;
 537
 538             case NBD_OPT_EXPORT_NAME:
 539                 return nbd_negotiate_handle_export_name(client, length);
 540
 541             case NBD_OPT_STARTTLS:
 542                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 543                     return -EIO;
 544                 }
 545                 if (client->tlscreds) {
 546                     ret = nbd_negotiate_send_rep_err(client->ioc,
 547                                                      NBD_REP_ERR_INVALID,
 548                                                      clientflags,
 549                                                      "TLS already enabled");
 550                 } else {
 551                     ret = nbd_negotiate_send_rep_err(client->ioc,
 552                                                      NBD_REP_ERR_POLICY,
 553                                                      clientflags,
 554                                                      "TLS not configured");
 555                 }
 556                 if (ret < 0) {
 557                     return ret;
 558                 }
 559                 break;
 560             default:
 561                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 562                     return -EIO;
 563                 }
 564                 ret = nbd_negotiate_send_rep_err(client->ioc,
 565                                                  NBD_REP_ERR_UNSUP,
 566                                                  clientflags,
 567                                                  "Unsupported option 0x%"
 568                                                  PRIx32,
 569                                                  clientflags);
 570                 if (ret < 0) {
 571                     return ret;
 572                 }
 573                 break;
 574             }
 575         } else {
 576             /*
 577              * If broken new-style we should drop the connection
 578              * for anything except NBD_OPT_EXPORT_NAME
 579              */
 580             switch (clientflags) {
 581             case NBD_OPT_EXPORT_NAME:
 582                 return nbd_negotiate_handle_export_name(client, length);
 583
 584             default:
 585                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 586                 return -EINVAL;
 587             }
 588         }
 589     }
 590 }
 591
 592 typedef struct {
 593     NBDClient *client;
 594     Coroutine *co;
 595 } NBDClientNewData;
 596
 597 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 598 {
 599     NBDClient *client = data->client;
 600     char buf[8 + 8 + 8 + 128];
 601     int rc;
 602     const uint16_t myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 603                               NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 604     bool oldStyle;
 605
 606     /* Old style negotiation header without options
 607         [ 0 ..   7]   passwd       ("NBDMAGIC")
 608         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 609         [16 ..  23]   size
 610         [24 ..  25]   server flags (0)
 611         [26 ..  27]   export flags
 612         [28 .. 151]   reserved     (0)
 613
 614        New style negotiation header with options
 615         [ 0 ..   7]   passwd       ("NBDMAGIC")
 616         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 617         [16 ..  17]   server flags (0)
 618         ....options sent....
 619         [18 ..  25]   size
 620         [26 ..  27]   export flags
 621         [28 .. 151]   reserved     (0)
 622      */
 623
 624     qio_channel_set_blocking(client->ioc, false, NULL);
 625     rc = -EINVAL;
 626
 627     TRACE("Beginning negotiation.");
 628     memset(buf, 0, sizeof(buf));
 629     memcpy(buf, "NBDMAGIC", 8);
 630
 631     oldStyle = client->exp != NULL && !client->tlscreds;
 632     if (oldStyle) {
 633         TRACE("advertising size %" PRIu64 " and flags %x",
 634               client->exp->size, client->exp->nbdflags | myflags);
 635         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 636         stq_be_p(buf + 16, client->exp->size);
 637         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 638     } else {
 639         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 640         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE);
 641     }
 642
 643     if (oldStyle) {
 644         if (client->tlscreds) {
 645             TRACE("TLS cannot be enabled with oldstyle protocol");
 646             goto fail;
 647         }
 648         if (nbd_negotiate_write(client->ioc, buf, sizeof(buf)) != sizeof(buf)) {
 649             LOG("write failed");
 650             goto fail;
 651         }
 652     } else {
 653         if (nbd_negotiate_write(client->ioc, buf, 18) != 18) {
 654             LOG("write failed");
 655             goto fail;
 656         }
 657         rc = nbd_negotiate_options(client);
 658         if (rc != 0) {
 659             LOG("option negotiation failed");
 660             goto fail;
 661         }
 662
 663         TRACE("advertising size %" PRIu64 " and flags %x",
 664               client->exp->size, client->exp->nbdflags | myflags);
 665         stq_be_p(buf + 18, client->exp->size);
 666         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 667         if (nbd_negotiate_write(client->ioc, buf + 18, sizeof(buf) - 18) !=
 668             sizeof(buf) - 18) {
 669             LOG("write failed");
 670             goto fail;
 671         }
 672     }
 673
 674     TRACE("Negotiation succeeded.");
 675     rc = 0;
 676 fail:
 677     return rc;
 678 }
 679
 680 static ssize_t nbd_receive_request(QIOChannel *ioc, NBDRequest *request)
 681 {
 682     uint8_t buf[NBD_REQUEST_SIZE];
 683     uint32_t magic;
 684     ssize_t ret;
 685
 686     ret = read_sync(ioc, buf, sizeof(buf));
 687     if (ret < 0) {
 688         return ret;
 689     }
 690
 691     if (ret != sizeof(buf)) {
 692         LOG("read failed");
 693         return -EINVAL;
 694     }
 695
 696     /* Request
 697        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 698        [ 4 ..  5]   flags   (NBD_CMD_FLAG_FUA, ...)
 699        [ 6 ..  7]   type    (NBD_CMD_READ, ...)
 700        [ 8 .. 15]   handle
 701        [16 .. 23]   from
 702        [24 .. 27]   len
 703      */
 704
 705     magic = ldl_be_p(buf);
 706     request->flags  = lduw_be_p(buf + 4);
 707     request->type   = lduw_be_p(buf + 6);
 708     request->handle = ldq_be_p(buf + 8);
 709     request->from   = ldq_be_p(buf + 16);
 710     request->len    = ldl_be_p(buf + 24);
 711
 712     TRACE("Got request: { magic = 0x%" PRIx32 ", .flags = %" PRIx16
 713           ", .type = %" PRIx16 ", from = %" PRIu64 ", len = %" PRIu32 " }",
 714           magic, request->flags, request->type, request->from, request->len);
 715
 716     if (magic != NBD_REQUEST_MAGIC) {
 717         LOG("invalid magic (got 0x%" PRIx32 ")", magic);
 718         return -EINVAL;
 719     }
 720     return 0;
 721 }
 722
 723 static ssize_t nbd_send_reply(QIOChannel *ioc, NBDReply *reply)
 724 {
 725     uint8_t buf[NBD_REPLY_SIZE];
 726     ssize_t ret;
 727
 728     reply->error = system_errno_to_nbd_errno(reply->error);
 729
 730     TRACE("Sending response to client: { .error = %" PRId32
 731           ", handle = %" PRIu64 " }",
 732           reply->error, reply->handle);
 733
 734     /* Reply
 735        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 736        [ 4 ..  7]    error   (0 == no error)
 737        [ 7 .. 15]    handle
 738      */
 739     stl_be_p(buf, NBD_REPLY_MAGIC);
 740     stl_be_p(buf + 4, reply->error);
 741     stq_be_p(buf + 8, reply->handle);
 742
 743     ret = write_sync(ioc, buf, sizeof(buf));
 744     if (ret < 0) {
 745         return ret;
 746     }
 747
 748     if (ret != sizeof(buf)) {
 749         LOG("writing to socket failed");
 750         return -EINVAL;
 751     }
 752     return 0;
 753 }
 754
 755 #define MAX_NBD_REQUESTS 16
 756
 757 void nbd_client_get(NBDClient *client)
 758 {
 759     client->refcount++;
 760 }
 761
 762 void nbd_client_put(NBDClient *client)
 763 {
 764     if (--client->refcount == 0) {
 765         /* The last reference should be dropped by client->close,
 766          * which is called by client_close.
 767          */
 768         assert(client->closing);
 769
 770         nbd_unset_handlers(client);
 771         object_unref(OBJECT(client->sioc));
 772         object_unref(OBJECT(client->ioc));
 773         if (client->tlscreds) {
 774             object_unref(OBJECT(client->tlscreds));
 775         }
 776         g_free(client->tlsaclname);
 777         if (client->exp) {
 778             QTAILQ_REMOVE(&client->exp->clients, client, next);
 779             nbd_export_put(client->exp);
 780         }
 781         g_free(client);
 782     }
 783 }
 784
 785 static void client_close(NBDClient *client)
 786 {
 787     if (client->closing) {
 788         return;
 789     }
 790
 791     client->closing = true;
 792
 793     /* Force requests to finish.  They will drop their own references,
 794      * then we'll close the socket and free the NBDClient.
 795      */
 796     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 797                          NULL);
 798
 799     /* Also tell the client, so that they release their reference.  */
 800     if (client->close) {
 801         client->close(client);
 802     }
 803 }
 804
 805 static NBDRequestData *nbd_request_get(NBDClient *client)
 806 {
 807     NBDRequestData *req;
 808
 809     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 810     client->nb_requests++;
 811     nbd_update_can_read(client);
 812
 813     req = g_new0(NBDRequestData, 1);
 814     nbd_client_get(client);
 815     req->client = client;
 816     return req;
 817 }
 818
 819 static void nbd_request_put(NBDRequestData *req)
 820 {
 821     NBDClient *client = req->client;
 822
 823     if (req->data) {
 824         qemu_vfree(req->data);
 825     }
 826     g_free(req);
 827
 828     client->nb_requests--;
 829     nbd_update_can_read(client);
 830     nbd_client_put(client);
 831 }
 832
 833 static void blk_aio_attached(AioContext *ctx, void *opaque)
 834 {
 835     NBDExport *exp = opaque;
 836     NBDClient *client;
 837
 838     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 839
 840     exp->ctx = ctx;
 841
 842     QTAILQ_FOREACH(client, &exp->clients, next) {
 843         nbd_set_handlers(client);
 844     }
 845 }
 846
 847 static void blk_aio_detach(void *opaque)
 848 {
 849     NBDExport *exp = opaque;
 850     NBDClient *client;
 851
 852     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 853
 854     QTAILQ_FOREACH(client, &exp->clients, next) {
 855         nbd_unset_handlers(client);
 856     }
 857
 858     exp->ctx = NULL;
 859 }
 860
 861 static void nbd_eject_notifier(Notifier *n, void *data)
 862 {
 863     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 864     nbd_export_close(exp);
 865 }
 866
 867 NBDExport *nbd_export_new(BlockDriverState *bs, off_t dev_offset, off_t size,
 868                           uint16_t nbdflags, void (*close)(NBDExport *),
 869                           bool writethrough, BlockBackend *on_eject_blk,
 870                           Error **errp)
 871 {
 872     BlockBackend *blk;
 873     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 874
 875     blk = blk_new();
 876     blk_insert_bs(blk, bs);
 877     blk_set_enable_write_cache(blk, !writethrough);
 878
 879     exp->refcount = 1;
 880     QTAILQ_INIT(&exp->clients);
 881     exp->blk = blk;
 882     exp->dev_offset = dev_offset;
 883     exp->nbdflags = nbdflags;
 884     exp->size = size < 0 ? blk_getlength(blk) : size;
 885     if (exp->size < 0) {
 886         error_setg_errno(errp, -exp->size,
 887                          "Failed to determine the NBD export's length");
 888         goto fail;
 889     }
 890     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 891
 892     exp->close = close;
 893     exp->ctx = blk_get_aio_context(blk);
 894     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 895
 896     if (on_eject_blk) {
 897         blk_ref(on_eject_blk);
 898         exp->eject_notifier_blk = on_eject_blk;
 899         exp->eject_notifier.notify = nbd_eject_notifier;
 900         blk_add_remove_bs_notifier(on_eject_blk, &exp->eject_notifier);
 901     }
 902
 903     /*
 904      * NBD exports are used for non-shared storage migration.  Make sure
 905      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 906      * access since the export could be available before migration handover.
 907      */
 908     aio_context_acquire(exp->ctx);
 909     blk_invalidate_cache(blk, NULL);
 910     aio_context_release(exp->ctx);
 911     return exp;
 912
 913 fail:
 914     blk_unref(blk);
 915     g_free(exp);
 916     return NULL;
 917 }
 918
 919 NBDExport *nbd_export_find(const char *name)
 920 {
 921     NBDExport *exp;
 922     QTAILQ_FOREACH(exp, &exports, next) {
 923         if (strcmp(name, exp->name) == 0) {
 924             return exp;
 925         }
 926     }
 927
 928     return NULL;
 929 }
 930
 931 void nbd_export_set_name(NBDExport *exp, const char *name)
 932 {
 933     if (exp->name == name) {
 934         return;
 935     }
 936
 937     nbd_export_get(exp);
 938     if (exp->name != NULL) {
 939         g_free(exp->name);
 940         exp->name = NULL;
 941         QTAILQ_REMOVE(&exports, exp, next);
 942         nbd_export_put(exp);
 943     }
 944     if (name != NULL) {
 945         nbd_export_get(exp);
 946         exp->name = g_strdup(name);
 947         QTAILQ_INSERT_TAIL(&exports, exp, next);
 948     }
 949     nbd_export_put(exp);
 950 }
 951
 952 void nbd_export_set_description(NBDExport *exp, const char *description)
 953 {
 954     g_free(exp->description);
 955     exp->description = g_strdup(description);
 956 }
 957
 958 void nbd_export_close(NBDExport *exp)
 959 {
 960     NBDClient *client, *next;
 961
 962     nbd_export_get(exp);
 963     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 964         client_close(client);
 965     }
 966     nbd_export_set_name(exp, NULL);
 967     nbd_export_set_description(exp, NULL);
 968     nbd_export_put(exp);
 969 }
 970
 971 void nbd_export_get(NBDExport *exp)
 972 {
 973     assert(exp->refcount > 0);
 974     exp->refcount++;
 975 }
 976
 977 void nbd_export_put(NBDExport *exp)
 978 {
 979     assert(exp->refcount > 0);
 980     if (exp->refcount == 1) {
 981         nbd_export_close(exp);
 982     }
 983
 984     if (--exp->refcount == 0) {
 985         assert(exp->name == NULL);
 986         assert(exp->description == NULL);
 987
 988         if (exp->close) {
 989             exp->close(exp);
 990         }
 991
 992         if (exp->blk) {
 993             if (exp->eject_notifier_blk) {
 994                 notifier_remove(&exp->eject_notifier);
 995                 blk_unref(exp->eject_notifier_blk);
 996             }
 997             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 998                                             blk_aio_detach, exp);
 999             blk_unref(exp->blk);
1000             exp->blk = NULL;
1001         }
1002
1003         g_free(exp);
1004     }
1005 }
1006
1007 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
1008 {
1009     return exp->blk;
1010 }
1011
1012 void nbd_export_close_all(void)
1013 {
1014     NBDExport *exp, *next;
1015
1016     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
1017         nbd_export_close(exp);
1018     }
1019 }
1020
1021 static ssize_t nbd_co_send_reply(NBDRequestData *req, NBDReply *reply,
1022                                  int len)
1023 {
1024     NBDClient *client = req->client;
1025     ssize_t rc, ret;
1026
1027     g_assert(qemu_in_coroutine());
1028     qemu_co_mutex_lock(&client->send_lock);
1029     client->send_coroutine = qemu_coroutine_self();
1030     nbd_set_handlers(client);
1031
1032     if (!len) {
1033         rc = nbd_send_reply(client->ioc, reply);
1034     } else {
1035         qio_channel_set_cork(client->ioc, true);
1036         rc = nbd_send_reply(client->ioc, reply);
1037         if (rc >= 0) {
1038             ret = write_sync(client->ioc, req->data, len);
1039             if (ret != len) {
1040                 rc = -EIO;
1041             }
1042         }
1043         qio_channel_set_cork(client->ioc, false);
1044     }
1045
1046     client->send_coroutine = NULL;
1047     nbd_set_handlers(client);
1048     qemu_co_mutex_unlock(&client->send_lock);
1049     return rc;
1050 }
1051
1052 /* Collect a client request.  Return 0 if request looks valid, -EAGAIN
1053  * to keep trying the collection, -EIO to drop connection right away,
1054  * and any other negative value to report an error to the client
1055  * (although the caller may still need to disconnect after reporting
1056  * the error).  */
1057 static ssize_t nbd_co_receive_request(NBDRequestData *req,
1058                                       NBDRequest *request)
1059 {
1060     NBDClient *client = req->client;
1061     ssize_t rc;
1062
1063     g_assert(qemu_in_coroutine());
1064     client->recv_coroutine = qemu_coroutine_self();
1065     nbd_update_can_read(client);
1066
1067     rc = nbd_receive_request(client->ioc, request);
1068     if (rc < 0) {
1069         if (rc != -EAGAIN) {
1070             rc = -EIO;
1071         }
1072         goto out;
1073     }
1074
1075     TRACE("Decoding type");
1076
1077     if (request->type != NBD_CMD_WRITE) {
1078         /* No payload, we are ready to read the next request.  */
1079         req->complete = true;
1080     }
1081
1082     if (request->type == NBD_CMD_DISC) {
1083         /* Special case: we're going to disconnect without a reply,
1084          * whether or not flags, from, or len are bogus */
1085         TRACE("Request type is DISCONNECT");
1086         rc = -EIO;
1087         goto out;
1088     }
1089
1090     /* Check for sanity in the parameters, part 1.  Defer as many
1091      * checks as possible until after reading any NBD_CMD_WRITE
1092      * payload, so we can try and keep the connection alive.  */
1093     if ((request->from + request->len) < request->from) {
1094         LOG("integer overflow detected, you're probably being attacked");
1095         rc = -EINVAL;
1096         goto out;
1097     }
1098
1099     if (request->type == NBD_CMD_READ || request->type == NBD_CMD_WRITE) {
1100         if (request->len > NBD_MAX_BUFFER_SIZE) {
1101             LOG("len (%" PRIu32" ) is larger than max len (%u)",
1102                 request->len, NBD_MAX_BUFFER_SIZE);
1103             rc = -EINVAL;
1104             goto out;
1105         }
1106
1107         req->data = blk_try_blockalign(client->exp->blk, request->len);
1108         if (req->data == NULL) {
1109             rc = -ENOMEM;
1110             goto out;
1111         }
1112     }
1113     if (request->type == NBD_CMD_WRITE) {
1114         TRACE("Reading %" PRIu32 " byte(s)", request->len);
1115
1116         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1117             LOG("reading from socket failed");
1118             rc = -EIO;
1119             goto out;
1120         }
1121         req->complete = true;
1122     }
1123
1124     /* Sanity checks, part 2. */
1125     if (request->from + request->len > client->exp->size) {
1126         LOG("operation past EOF; From: %" PRIu64 ", Len: %" PRIu32
1127             ", Size: %" PRIu64, request->from, request->len,
1128             (uint64_t)client->exp->size);
1129         rc = request->type == NBD_CMD_WRITE ? -ENOSPC : -EINVAL;
1130         goto out;
1131     }
1132     if (request->flags & ~NBD_CMD_FLAG_FUA) {
1133         LOG("unsupported flags (got 0x%x)", request->flags);
1134         rc = -EINVAL;
1135         goto out;
1136     }
1137
1138     rc = 0;
1139
1140 out:
1141     client->recv_coroutine = NULL;
1142     nbd_update_can_read(client);
1143
1144     return rc;
1145 }
1146
1147 static void nbd_trip(void *opaque)
1148 {
1149     NBDClient *client = opaque;
1150     NBDExport *exp = client->exp;
1151     NBDRequestData *req;
1152     NBDRequest request;
1153     NBDReply reply;
1154     ssize_t ret;
1155     int flags;
1156
1157     TRACE("Reading request.");
1158     if (client->closing) {
1159         return;
1160     }
1161
1162     req = nbd_request_get(client);
1163     ret = nbd_co_receive_request(req, &request);
1164     if (ret == -EAGAIN) {
1165         goto done;
1166     }
1167     if (ret == -EIO) {
1168         goto out;
1169     }
1170
1171     reply.handle = request.handle;
1172     reply.error = 0;
1173
1174     if (ret < 0) {
1175         reply.error = -ret;
1176         goto error_reply;
1177     }
1178
1179     if (client->closing) {
1180         /*
1181          * The client may be closed when we are blocked in
1182          * nbd_co_receive_request()
1183          */
1184         goto done;
1185     }
1186
1187     switch (request.type) {
1188     case NBD_CMD_READ:
1189         TRACE("Request type is READ");
1190
1191         /* XXX: NBD Protocol only documents use of FUA with WRITE */
1192         if (request.flags & NBD_CMD_FLAG_FUA) {
1193             ret = blk_co_flush(exp->blk);
1194             if (ret < 0) {
1195                 LOG("flush failed");
1196                 reply.error = -ret;
1197                 goto error_reply;
1198             }
1199         }
1200
1201         ret = blk_pread(exp->blk, request.from + exp->dev_offset,
1202                         req->data, request.len);
1203         if (ret < 0) {
1204             LOG("reading from file failed");
1205             reply.error = -ret;
1206             goto error_reply;
1207         }
1208
1209         TRACE("Read %" PRIu32" byte(s)", request.len);
1210         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1211             goto out;
1212         break;
1213     case NBD_CMD_WRITE:
1214         TRACE("Request type is WRITE");
1215
1216         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1217             TRACE("Server is read-only, return error");
1218             reply.error = EROFS;
1219             goto error_reply;
1220         }
1221
1222         TRACE("Writing to device");
1223
1224         flags = 0;
1225         if (request.flags & NBD_CMD_FLAG_FUA) {
1226             flags |= BDRV_REQ_FUA;
1227         }
1228         ret = blk_pwrite(exp->blk, request.from + exp->dev_offset,
1229                          req->data, request.len, flags);
1230         if (ret < 0) {
1231             LOG("writing to file failed");
1232             reply.error = -ret;
1233             goto error_reply;
1234         }
1235
1236         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1237             goto out;
1238         }
1239         break;
1240
1241     case NBD_CMD_DISC:
1242         /* unreachable, thanks to special case in nbd_co_receive_request() */
1243         abort();
1244
1245     case NBD_CMD_FLUSH:
1246         TRACE("Request type is FLUSH");
1247
1248         ret = blk_co_flush(exp->blk);
1249         if (ret < 0) {
1250             LOG("flush failed");
1251             reply.error = -ret;
1252         }
1253         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1254             goto out;
1255         }
1256         break;
1257     case NBD_CMD_TRIM:
1258         TRACE("Request type is TRIM");
1259         ret = blk_co_pdiscard(exp->blk, request.from + exp->dev_offset,
1260                               request.len);
1261         if (ret < 0) {
1262             LOG("discard failed");
1263             reply.error = -ret;
1264         }
1265         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1266             goto out;
1267         }
1268         break;
1269     default:
1270         LOG("invalid request type (%" PRIu32 ") received", request.type);
1271         reply.error = EINVAL;
1272     error_reply:
1273         /* We must disconnect after NBD_CMD_WRITE if we did not
1274          * read the payload.
1275          */
1276         if (nbd_co_send_reply(req, &reply, 0) < 0 || !req->complete) {
1277             goto out;
1278         }
1279         break;
1280     }
1281
1282     TRACE("Request/Reply complete");
1283
1284 done:
1285     nbd_request_put(req);
1286     return;
1287
1288 out:
1289     nbd_request_put(req);
1290     client_close(client);
1291 }
1292
1293 static void nbd_read(void *opaque)
1294 {
1295     NBDClient *client = opaque;
1296
1297     if (client->recv_coroutine) {
1298         qemu_coroutine_enter(client->recv_coroutine);
1299     } else {
1300         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip, client));
1301     }
1302 }
1303
1304 static void nbd_restart_write(void *opaque)
1305 {
1306     NBDClient *client = opaque;
1307
1308     qemu_coroutine_enter(client->send_coroutine);
1309 }
1310
1311 static void nbd_set_handlers(NBDClient *client)
1312 {
1313     if (client->exp && client->exp->ctx) {
1314         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1315                            true,
1316                            client->can_read ? nbd_read : NULL,
1317                            client->send_coroutine ? nbd_restart_write : NULL,
1318                            client);
1319     }
1320 }
1321
1322 static void nbd_unset_handlers(NBDClient *client)
1323 {
1324     if (client->exp && client->exp->ctx) {
1325         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1326                            true, NULL, NULL, NULL);
1327     }
1328 }
1329
1330 static void nbd_update_can_read(NBDClient *client)
1331 {
1332     bool can_read = client->recv_coroutine ||
1333                     client->nb_requests < MAX_NBD_REQUESTS;
1334
1335     if (can_read != client->can_read) {
1336         client->can_read = can_read;
1337         nbd_set_handlers(client);
1338
1339         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1340          * in nbd_set_handlers() will have taken care of that */
1341     }
1342 }
1343
1344 static coroutine_fn void nbd_co_client_start(void *opaque)
1345 {
1346     NBDClientNewData *data = opaque;
1347     NBDClient *client = data->client;
1348     NBDExport *exp = client->exp;
1349
1350     if (exp) {
1351         nbd_export_get(exp);
1352     }
1353     if (nbd_negotiate(data)) {
1354         client_close(client);
1355         goto out;
1356     }
1357     qemu_co_mutex_init(&client->send_lock);
1358     nbd_set_handlers(client);
1359
1360     if (exp) {
1361         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1362     }
1363 out:
1364     g_free(data);
1365 }
1366
1367 void nbd_client_new(NBDExport *exp,
1368                     QIOChannelSocket *sioc,
1369                     QCryptoTLSCreds *tlscreds,
1370                     const char *tlsaclname,
1371                     void (*close_fn)(NBDClient *))
1372 {
1373     NBDClient *client;
1374     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1375
1376     client = g_malloc0(sizeof(NBDClient));
1377     client->refcount = 1;
1378     client->exp = exp;
1379     client->tlscreds = tlscreds;
1380     if (tlscreds) {
1381         object_ref(OBJECT(client->tlscreds));
1382     }
1383     client->tlsaclname = g_strdup(tlsaclname);
1384     client->sioc = sioc;
1385     object_ref(OBJECT(client->sioc));
1386     client->ioc = QIO_CHANNEL(sioc);
1387     object_ref(OBJECT(client->ioc));
1388     client->can_read = true;
1389     client->close = close_fn;
1390
1391     data->client = client;
1392     data->co = qemu_coroutine_create(nbd_co_client_start, data);
1393     qemu_coroutine_enter(data->co);
1394 }