nbd/server.c

   1 /*
   2  *  Copyright (C) 2005  Anthony Liguori <anthony@codemonkey.ws>
   3  *
   4  *  Network Block Device Server Side
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; under version 2 of the License.
   9  *
  10  *  This program is distributed in the hope that it will be useful,
  11  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  *  GNU General Public License for more details.
  14  *
  15  *  You should have received a copy of the GNU General Public License
  16  *  along with this program; if not, see <http://www.gnu.org/licenses/>.
  17  */
  18
  19 #include "qemu/osdep.h"
  20 #include "qapi/error.h"
  21 #include "nbd-internal.h"
  22
  23 static int system_errno_to_nbd_errno(int err)
  24 {
  25     switch (err) {
  26     case 0:
  27         return NBD_SUCCESS;
  28     case EPERM:
  29     case EROFS:
  30         return NBD_EPERM;
  31     case EIO:
  32         return NBD_EIO;
  33     case ENOMEM:
  34         return NBD_ENOMEM;
  35 #ifdef EDQUOT
  36     case EDQUOT:
  37 #endif
  38     case EFBIG:
  39     case ENOSPC:
  40         return NBD_ENOSPC;
  41     case EINVAL:
  42     default:
  43         return NBD_EINVAL;
  44     }
  45 }
  46
  47 /* Definitions for opaque data types */
  48
  49 typedef struct NBDRequest NBDRequest;
  50
  51 struct NBDRequest {
  52     QSIMPLEQ_ENTRY(NBDRequest) entry;
  53     NBDClient *client;
  54     uint8_t *data;
  55 };
  56
  57 struct NBDExport {
  58     int refcount;
  59     void (*close)(NBDExport *exp);
  60
  61     BlockBackend *blk;
  62     char *name;
  63     off_t dev_offset;
  64     off_t size;
  65     uint32_t nbdflags;
  66     QTAILQ_HEAD(, NBDClient) clients;
  67     QTAILQ_ENTRY(NBDExport) next;
  68
  69     AioContext *ctx;
  70
  71     Notifier eject_notifier;
  72 };
  73
  74 static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports);
  75
  76 struct NBDClient {
  77     int refcount;
  78     void (*close)(NBDClient *client);
  79
  80     NBDExport *exp;
  81     QCryptoTLSCreds *tlscreds;
  82     char *tlsaclname;
  83     QIOChannelSocket *sioc; /* The underlying data channel */
  84     QIOChannel *ioc; /* The current I/O channel which may differ (eg TLS) */
  85
  86     Coroutine *recv_coroutine;
  87
  88     CoMutex send_lock;
  89     Coroutine *send_coroutine;
  90
  91     bool can_read;
  92
  93     QTAILQ_ENTRY(NBDClient) next;
  94     int nb_requests;
  95     bool closing;
  96 };
  97
  98 /* That's all folks */
  99
 100 static void nbd_set_handlers(NBDClient *client);
 101 static void nbd_unset_handlers(NBDClient *client);
 102 static void nbd_update_can_read(NBDClient *client);
 103
 104 static gboolean nbd_negotiate_continue(QIOChannel *ioc,
 105                                        GIOCondition condition,
 106                                        void *opaque)
 107 {
 108     qemu_coroutine_enter(opaque, NULL);
 109     return TRUE;
 110 }
 111
 112 static ssize_t nbd_negotiate_read(QIOChannel *ioc, void *buffer, size_t size)
 113 {
 114     ssize_t ret;
 115     guint watch;
 116
 117     assert(qemu_in_coroutine());
 118     /* Negotiation are always in main loop. */
 119     watch = qio_channel_add_watch(ioc,
 120                                   G_IO_IN,
 121                                   nbd_negotiate_continue,
 122                                   qemu_coroutine_self(),
 123                                   NULL);
 124     ret = read_sync(ioc, buffer, size);
 125     g_source_remove(watch);
 126     return ret;
 127
 128 }
 129
 130 static ssize_t nbd_negotiate_write(QIOChannel *ioc, void *buffer, size_t size)
 131 {
 132     ssize_t ret;
 133     guint watch;
 134
 135     assert(qemu_in_coroutine());
 136     /* Negotiation are always in main loop. */
 137     watch = qio_channel_add_watch(ioc,
 138                                   G_IO_OUT,
 139                                   nbd_negotiate_continue,
 140                                   qemu_coroutine_self(),
 141                                   NULL);
 142     ret = write_sync(ioc, buffer, size);
 143     g_source_remove(watch);
 144     return ret;
 145 }
 146
 147 static ssize_t nbd_negotiate_drop_sync(QIOChannel *ioc, size_t size)
 148 {
 149     ssize_t ret, dropped = size;
 150     uint8_t *buffer = g_malloc(MIN(65536, size));
 151
 152     while (size > 0) {
 153         ret = nbd_negotiate_read(ioc, buffer, MIN(65536, size));
 154         if (ret < 0) {
 155             g_free(buffer);
 156             return ret;
 157         }
 158
 159         assert(ret <= size);
 160         size -= ret;
 161     }
 162
 163     g_free(buffer);
 164     return dropped;
 165 }
 166
 167 /* Basic flow for negotiation
 168
 169    Server         Client
 170    Negotiate
 171
 172    or
 173
 174    Server         Client
 175    Negotiate #1
 176                   Option
 177    Negotiate #2
 178
 179    ----
 180
 181    followed by
 182
 183    Server         Client
 184                   Request
 185    Response
 186                   Request
 187    Response
 188                   ...
 189    ...
 190                   Request (type == 2)
 191
 192 */
 193
 194 static int nbd_negotiate_send_rep(QIOChannel *ioc, uint32_t type, uint32_t opt)
 195 {
 196     uint64_t magic;
 197     uint32_t len;
 198
 199     TRACE("Reply opt=%x type=%x", type, opt);
 200
 201     magic = cpu_to_be64(NBD_REP_MAGIC);
 202     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 203         LOG("write failed (rep magic)");
 204         return -EINVAL;
 205     }
 206     opt = cpu_to_be32(opt);
 207     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 208         LOG("write failed (rep opt)");
 209         return -EINVAL;
 210     }
 211     type = cpu_to_be32(type);
 212     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 213         LOG("write failed (rep type)");
 214         return -EINVAL;
 215     }
 216     len = cpu_to_be32(0);
 217     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 218         LOG("write failed (rep data length)");
 219         return -EINVAL;
 220     }
 221     return 0;
 222 }
 223
 224 static int nbd_negotiate_send_rep_list(QIOChannel *ioc, NBDExport *exp)
 225 {
 226     uint64_t magic, name_len;
 227     uint32_t opt, type, len;
 228
 229     TRACE("Advertizing export name '%s'", exp->name ? exp->name : "");
 230     name_len = strlen(exp->name);
 231     magic = cpu_to_be64(NBD_REP_MAGIC);
 232     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 233         LOG("write failed (magic)");
 234         return -EINVAL;
 235      }
 236     opt = cpu_to_be32(NBD_OPT_LIST);
 237     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 238         LOG("write failed (opt)");
 239         return -EINVAL;
 240     }
 241     type = cpu_to_be32(NBD_REP_SERVER);
 242     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 243         LOG("write failed (reply type)");
 244         return -EINVAL;
 245     }
 246     len = cpu_to_be32(name_len + sizeof(len));
 247     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 248         LOG("write failed (length)");
 249         return -EINVAL;
 250     }
 251     len = cpu_to_be32(name_len);
 252     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 253         LOG("write failed (length)");
 254         return -EINVAL;
 255     }
 256     if (nbd_negotiate_write(ioc, exp->name, name_len) != name_len) {
 257         LOG("write failed (buffer)");
 258         return -EINVAL;
 259     }
 260     return 0;
 261 }
 262
 263 static int nbd_negotiate_handle_list(NBDClient *client, uint32_t length)
 264 {
 265     NBDExport *exp;
 266
 267     if (length) {
 268         if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 269             return -EIO;
 270         }
 271         return nbd_negotiate_send_rep(client->ioc,
 272                                       NBD_REP_ERR_INVALID, NBD_OPT_LIST);
 273     }
 274
 275     /* For each export, send a NBD_REP_SERVER reply. */
 276     QTAILQ_FOREACH(exp, &exports, next) {
 277         if (nbd_negotiate_send_rep_list(client->ioc, exp)) {
 278             return -EINVAL;
 279         }
 280     }
 281     /* Finish with a NBD_REP_ACK. */
 282     return nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_LIST);
 283 }
 284
 285 static int nbd_negotiate_handle_export_name(NBDClient *client, uint32_t length)
 286 {
 287     int rc = -EINVAL;
 288     char name[256];
 289
 290     /* Client sends:
 291         [20 ..  xx]   export name (length bytes)
 292      */
 293     TRACE("Checking length");
 294     if (length > 255) {
 295         LOG("Bad length received");
 296         goto fail;
 297     }
 298     if (nbd_negotiate_read(client->ioc, name, length) != length) {
 299         LOG("read failed");
 300         goto fail;
 301     }
 302     name[length] = '\0';
 303
 304     TRACE("Client requested export '%s'", name);
 305
 306     client->exp = nbd_export_find(name);
 307     if (!client->exp) {
 308         LOG("export not found");
 309         goto fail;
 310     }
 311
 312     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 313     nbd_export_get(client->exp);
 314     rc = 0;
 315 fail:
 316     return rc;
 317 }
 318
 319
 320 static QIOChannel *nbd_negotiate_handle_starttls(NBDClient *client,
 321                                                  uint32_t length)
 322 {
 323     QIOChannel *ioc;
 324     QIOChannelTLS *tioc;
 325     struct NBDTLSHandshakeData data = { 0 };
 326
 327     TRACE("Setting up TLS");
 328     ioc = client->ioc;
 329     if (length) {
 330         if (nbd_negotiate_drop_sync(ioc, length) != length) {
 331             return NULL;
 332         }
 333         nbd_negotiate_send_rep(ioc, NBD_REP_ERR_INVALID, NBD_OPT_STARTTLS);
 334         return NULL;
 335     }
 336
 337     nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_STARTTLS);
 338
 339     tioc = qio_channel_tls_new_server(ioc,
 340                                       client->tlscreds,
 341                                       client->tlsaclname,
 342                                       NULL);
 343     if (!tioc) {
 344         return NULL;
 345     }
 346
 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_free(data.error);
 361         return NULL;
 362     }
 363
 364     return QIO_CHANNEL(tioc);
 365 }
 366
 367
 368 static int nbd_negotiate_options(NBDClient *client)
 369 {
 370     uint32_t flags;
 371     bool fixedNewstyle = false;
 372
 373     /* Client sends:
 374         [ 0 ..   3]   client flags
 375
 376         [ 0 ..   7]   NBD_OPTS_MAGIC
 377         [ 8 ..  11]   NBD option
 378         [12 ..  15]   Data length
 379         ...           Rest of request
 380
 381         [ 0 ..   7]   NBD_OPTS_MAGIC
 382         [ 8 ..  11]   Second NBD option
 383         [12 ..  15]   Data length
 384         ...           Rest of request
 385     */
 386
 387     if (nbd_negotiate_read(client->ioc, &flags, sizeof(flags)) !=
 388         sizeof(flags)) {
 389         LOG("read failed");
 390         return -EIO;
 391     }
 392     TRACE("Checking client flags");
 393     be32_to_cpus(&flags);
 394     if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) {
 395         TRACE("Support supports fixed newstyle handshake");
 396         fixedNewstyle = true;
 397         flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE;
 398     }
 399     if (flags != 0) {
 400         TRACE("Unknown client flags 0x%x received", flags);
 401         return -EIO;
 402     }
 403
 404     while (1) {
 405         int ret;
 406         uint32_t clientflags, length;
 407         uint64_t magic;
 408
 409         if (nbd_negotiate_read(client->ioc, &magic, sizeof(magic)) !=
 410             sizeof(magic)) {
 411             LOG("read failed");
 412             return -EINVAL;
 413         }
 414         TRACE("Checking opts magic");
 415         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 416             LOG("Bad magic received");
 417             return -EINVAL;
 418         }
 419
 420         if (nbd_negotiate_read(client->ioc, &clientflags,
 421                                sizeof(clientflags)) != sizeof(clientflags)) {
 422             LOG("read failed");
 423             return -EINVAL;
 424         }
 425         clientflags = be32_to_cpu(clientflags);
 426
 427         if (nbd_negotiate_read(client->ioc, &length, sizeof(length)) !=
 428             sizeof(length)) {
 429             LOG("read failed");
 430             return -EINVAL;
 431         }
 432         length = be32_to_cpu(length);
 433
 434         TRACE("Checking option 0x%x", clientflags);
 435         if (client->tlscreds &&
 436             client->ioc == (QIOChannel *)client->sioc) {
 437             QIOChannel *tioc;
 438             if (!fixedNewstyle) {
 439                 TRACE("Unsupported option 0x%x", clientflags);
 440                 return -EINVAL;
 441             }
 442             switch (clientflags) {
 443             case NBD_OPT_STARTTLS:
 444                 tioc = nbd_negotiate_handle_starttls(client, length);
 445                 if (!tioc) {
 446                     return -EIO;
 447                 }
 448                 object_unref(OBJECT(client->ioc));
 449                 client->ioc = QIO_CHANNEL(tioc);
 450                 break;
 451
 452             default:
 453                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 454                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_TLS_REQD,
 455                                        clientflags);
 456                 return -EINVAL;
 457             }
 458         } else if (fixedNewstyle) {
 459             switch (clientflags) {
 460             case NBD_OPT_LIST:
 461                 ret = nbd_negotiate_handle_list(client, length);
 462                 if (ret < 0) {
 463                     return ret;
 464                 }
 465                 break;
 466
 467             case NBD_OPT_ABORT:
 468                 return -EINVAL;
 469
 470             case NBD_OPT_EXPORT_NAME:
 471                 return nbd_negotiate_handle_export_name(client, length);
 472
 473             case NBD_OPT_STARTTLS:
 474                 if (client->tlscreds) {
 475                     TRACE("TLS already enabled");
 476                     nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_INVALID,
 477                                            clientflags);
 478                 } else {
 479                     TRACE("TLS not configured");
 480                     nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_POLICY,
 481                                            clientflags);
 482                 }
 483                 return -EINVAL;
 484             default:
 485                 TRACE("Unsupported option 0x%x", clientflags);
 486                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_UNSUP,
 487                                        clientflags);
 488                 return -EINVAL;
 489             }
 490         } else {
 491             /*
 492              * If broken new-style we should drop the connection
 493              * for anything except NBD_OPT_EXPORT_NAME
 494              */
 495             switch (clientflags) {
 496             case NBD_OPT_EXPORT_NAME:
 497                 return nbd_negotiate_handle_export_name(client, length);
 498
 499             default:
 500                 TRACE("Unsupported option 0x%x", clientflags);
 501                 return -EINVAL;
 502             }
 503         }
 504     }
 505 }
 506
 507 typedef struct {
 508     NBDClient *client;
 509     Coroutine *co;
 510 } NBDClientNewData;
 511
 512 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 513 {
 514     NBDClient *client = data->client;
 515     char buf[8 + 8 + 8 + 128];
 516     int rc;
 517     const int myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 518                          NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 519     bool oldStyle;
 520
 521     /* Old style negotiation header without options
 522         [ 0 ..   7]   passwd       ("NBDMAGIC")
 523         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 524         [16 ..  23]   size
 525         [24 ..  25]   server flags (0)
 526         [26 ..  27]   export flags
 527         [28 .. 151]   reserved     (0)
 528
 529        New style negotiation header with options
 530         [ 0 ..   7]   passwd       ("NBDMAGIC")
 531         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 532         [16 ..  17]   server flags (0)
 533         ....options sent....
 534         [18 ..  25]   size
 535         [26 ..  27]   export flags
 536         [28 .. 151]   reserved     (0)
 537      */
 538
 539     qio_channel_set_blocking(client->ioc, false, NULL);
 540     rc = -EINVAL;
 541
 542     TRACE("Beginning negotiation.");
 543     memset(buf, 0, sizeof(buf));
 544     memcpy(buf, "NBDMAGIC", 8);
 545
 546     oldStyle = client->exp != NULL && !client->tlscreds;
 547     if (oldStyle) {
 548         assert ((client->exp->nbdflags & ~65535) == 0);
 549         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 550         stq_be_p(buf + 16, client->exp->size);
 551         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 552     } else {
 553         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 554         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE);
 555     }
 556
 557     if (oldStyle) {
 558         if (client->tlscreds) {
 559             TRACE("TLS cannot be enabled with oldstyle protocol");
 560             goto fail;
 561         }
 562         if (nbd_negotiate_write(client->ioc, buf, sizeof(buf)) != sizeof(buf)) {
 563             LOG("write failed");
 564             goto fail;
 565         }
 566     } else {
 567         if (nbd_negotiate_write(client->ioc, buf, 18) != 18) {
 568             LOG("write failed");
 569             goto fail;
 570         }
 571         rc = nbd_negotiate_options(client);
 572         if (rc != 0) {
 573             LOG("option negotiation failed");
 574             goto fail;
 575         }
 576
 577         assert ((client->exp->nbdflags & ~65535) == 0);
 578         stq_be_p(buf + 18, client->exp->size);
 579         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 580         if (nbd_negotiate_write(client->ioc, buf + 18, sizeof(buf) - 18) !=
 581             sizeof(buf) - 18) {
 582             LOG("write failed");
 583             goto fail;
 584         }
 585     }
 586
 587     TRACE("Negotiation succeeded.");
 588     rc = 0;
 589 fail:
 590     return rc;
 591 }
 592
 593 #ifdef __linux__
 594
 595 int nbd_disconnect(int fd)
 596 {
 597     ioctl(fd, NBD_CLEAR_QUE);
 598     ioctl(fd, NBD_DISCONNECT);
 599     ioctl(fd, NBD_CLEAR_SOCK);
 600     return 0;
 601 }
 602
 603 #else
 604
 605 int nbd_disconnect(int fd)
 606 {
 607     return -ENOTSUP;
 608 }
 609 #endif
 610
 611 static ssize_t nbd_receive_request(QIOChannel *ioc, struct nbd_request *request)
 612 {
 613     uint8_t buf[NBD_REQUEST_SIZE];
 614     uint32_t magic;
 615     ssize_t ret;
 616
 617     ret = read_sync(ioc, buf, sizeof(buf));
 618     if (ret < 0) {
 619         return ret;
 620     }
 621
 622     if (ret != sizeof(buf)) {
 623         LOG("read failed");
 624         return -EINVAL;
 625     }
 626
 627     /* Request
 628        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 629        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 630        [ 8 .. 15]   handle
 631        [16 .. 23]   from
 632        [24 .. 27]   len
 633      */
 634
 635     magic = be32_to_cpup((uint32_t*)buf);
 636     request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 637     request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 638     request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 639     request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 640
 641     TRACE("Got request: "
 642           "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 643           magic, request->type, request->from, request->len);
 644
 645     if (magic != NBD_REQUEST_MAGIC) {
 646         LOG("invalid magic (got 0x%x)", magic);
 647         return -EINVAL;
 648     }
 649     return 0;
 650 }
 651
 652 static ssize_t nbd_send_reply(QIOChannel *ioc, struct nbd_reply *reply)
 653 {
 654     uint8_t buf[NBD_REPLY_SIZE];
 655     ssize_t ret;
 656
 657     reply->error = system_errno_to_nbd_errno(reply->error);
 658
 659     /* Reply
 660        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 661        [ 4 ..  7]    error   (0 == no error)
 662        [ 7 .. 15]    handle
 663      */
 664     stl_be_p(buf, NBD_REPLY_MAGIC);
 665     stl_be_p(buf + 4, reply->error);
 666     stq_be_p(buf + 8, reply->handle);
 667
 668     TRACE("Sending response to client");
 669
 670     ret = write_sync(ioc, buf, sizeof(buf));
 671     if (ret < 0) {
 672         return ret;
 673     }
 674
 675     if (ret != sizeof(buf)) {
 676         LOG("writing to socket failed");
 677         return -EINVAL;
 678     }
 679     return 0;
 680 }
 681
 682 #define MAX_NBD_REQUESTS 16
 683
 684 void nbd_client_get(NBDClient *client)
 685 {
 686     client->refcount++;
 687 }
 688
 689 void nbd_client_put(NBDClient *client)
 690 {
 691     if (--client->refcount == 0) {
 692         /* The last reference should be dropped by client->close,
 693          * which is called by client_close.
 694          */
 695         assert(client->closing);
 696
 697         nbd_unset_handlers(client);
 698         object_unref(OBJECT(client->sioc));
 699         object_unref(OBJECT(client->ioc));
 700         if (client->tlscreds) {
 701             object_unref(OBJECT(client->tlscreds));
 702         }
 703         g_free(client->tlsaclname);
 704         if (client->exp) {
 705             QTAILQ_REMOVE(&client->exp->clients, client, next);
 706             nbd_export_put(client->exp);
 707         }
 708         g_free(client);
 709     }
 710 }
 711
 712 static void client_close(NBDClient *client)
 713 {
 714     if (client->closing) {
 715         return;
 716     }
 717
 718     client->closing = true;
 719
 720     /* Force requests to finish.  They will drop their own references,
 721      * then we'll close the socket and free the NBDClient.
 722      */
 723     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 724                          NULL);
 725
 726     /* Also tell the client, so that they release their reference.  */
 727     if (client->close) {
 728         client->close(client);
 729     }
 730 }
 731
 732 static NBDRequest *nbd_request_get(NBDClient *client)
 733 {
 734     NBDRequest *req;
 735
 736     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 737     client->nb_requests++;
 738     nbd_update_can_read(client);
 739
 740     req = g_new0(NBDRequest, 1);
 741     nbd_client_get(client);
 742     req->client = client;
 743     return req;
 744 }
 745
 746 static void nbd_request_put(NBDRequest *req)
 747 {
 748     NBDClient *client = req->client;
 749
 750     if (req->data) {
 751         qemu_vfree(req->data);
 752     }
 753     g_free(req);
 754
 755     client->nb_requests--;
 756     nbd_update_can_read(client);
 757     nbd_client_put(client);
 758 }
 759
 760 static void blk_aio_attached(AioContext *ctx, void *opaque)
 761 {
 762     NBDExport *exp = opaque;
 763     NBDClient *client;
 764
 765     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 766
 767     exp->ctx = ctx;
 768
 769     QTAILQ_FOREACH(client, &exp->clients, next) {
 770         nbd_set_handlers(client);
 771     }
 772 }
 773
 774 static void blk_aio_detach(void *opaque)
 775 {
 776     NBDExport *exp = opaque;
 777     NBDClient *client;
 778
 779     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 780
 781     QTAILQ_FOREACH(client, &exp->clients, next) {
 782         nbd_unset_handlers(client);
 783     }
 784
 785     exp->ctx = NULL;
 786 }
 787
 788 static void nbd_eject_notifier(Notifier *n, void *data)
 789 {
 790     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 791     nbd_export_close(exp);
 792 }
 793
 794 NBDExport *nbd_export_new(BlockBackend *blk, off_t dev_offset, off_t size,
 795                           uint32_t nbdflags, void (*close)(NBDExport *),
 796                           Error **errp)
 797 {
 798     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 799     exp->refcount = 1;
 800     QTAILQ_INIT(&exp->clients);
 801     exp->blk = blk;
 802     exp->dev_offset = dev_offset;
 803     exp->nbdflags = nbdflags;
 804     exp->size = size < 0 ? blk_getlength(blk) : size;
 805     if (exp->size < 0) {
 806         error_setg_errno(errp, -exp->size,
 807                          "Failed to determine the NBD export's length");
 808         goto fail;
 809     }
 810     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 811
 812     exp->close = close;
 813     exp->ctx = blk_get_aio_context(blk);
 814     blk_ref(blk);
 815     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 816
 817     exp->eject_notifier.notify = nbd_eject_notifier;
 818     blk_add_remove_bs_notifier(blk, &exp->eject_notifier);
 819
 820     /*
 821      * NBD exports are used for non-shared storage migration.  Make sure
 822      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 823      * access since the export could be available before migration handover.
 824      */
 825     aio_context_acquire(exp->ctx);
 826     blk_invalidate_cache(blk, NULL);
 827     aio_context_release(exp->ctx);
 828     return exp;
 829
 830 fail:
 831     g_free(exp);
 832     return NULL;
 833 }
 834
 835 NBDExport *nbd_export_find(const char *name)
 836 {
 837     NBDExport *exp;
 838     QTAILQ_FOREACH(exp, &exports, next) {
 839         if (strcmp(name, exp->name) == 0) {
 840             return exp;
 841         }
 842     }
 843
 844     return NULL;
 845 }
 846
 847 void nbd_export_set_name(NBDExport *exp, const char *name)
 848 {
 849     if (exp->name == name) {
 850         return;
 851     }
 852
 853     nbd_export_get(exp);
 854     if (exp->name != NULL) {
 855         g_free(exp->name);
 856         exp->name = NULL;
 857         QTAILQ_REMOVE(&exports, exp, next);
 858         nbd_export_put(exp);
 859     }
 860     if (name != NULL) {
 861         nbd_export_get(exp);
 862         exp->name = g_strdup(name);
 863         QTAILQ_INSERT_TAIL(&exports, exp, next);
 864     }
 865     nbd_export_put(exp);
 866 }
 867
 868 void nbd_export_close(NBDExport *exp)
 869 {
 870     NBDClient *client, *next;
 871
 872     nbd_export_get(exp);
 873     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 874         client_close(client);
 875     }
 876     nbd_export_set_name(exp, NULL);
 877     nbd_export_put(exp);
 878 }
 879
 880 void nbd_export_get(NBDExport *exp)
 881 {
 882     assert(exp->refcount > 0);
 883     exp->refcount++;
 884 }
 885
 886 void nbd_export_put(NBDExport *exp)
 887 {
 888     assert(exp->refcount > 0);
 889     if (exp->refcount == 1) {
 890         nbd_export_close(exp);
 891     }
 892
 893     if (--exp->refcount == 0) {
 894         assert(exp->name == NULL);
 895
 896         if (exp->close) {
 897             exp->close(exp);
 898         }
 899
 900         if (exp->blk) {
 901             notifier_remove(&exp->eject_notifier);
 902             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 903                                             blk_aio_detach, exp);
 904             blk_unref(exp->blk);
 905             exp->blk = NULL;
 906         }
 907
 908         g_free(exp);
 909     }
 910 }
 911
 912 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
 913 {
 914     return exp->blk;
 915 }
 916
 917 void nbd_export_close_all(void)
 918 {
 919     NBDExport *exp, *next;
 920
 921     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 922         nbd_export_close(exp);
 923     }
 924 }
 925
 926 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 927                                  int len)
 928 {
 929     NBDClient *client = req->client;
 930     ssize_t rc, ret;
 931
 932     g_assert(qemu_in_coroutine());
 933     qemu_co_mutex_lock(&client->send_lock);
 934     client->send_coroutine = qemu_coroutine_self();
 935     nbd_set_handlers(client);
 936
 937     if (!len) {
 938         rc = nbd_send_reply(client->ioc, reply);
 939     } else {
 940         qio_channel_set_cork(client->ioc, true);
 941         rc = nbd_send_reply(client->ioc, reply);
 942         if (rc >= 0) {
 943             ret = write_sync(client->ioc, req->data, len);
 944             if (ret != len) {
 945                 rc = -EIO;
 946             }
 947         }
 948         qio_channel_set_cork(client->ioc, false);
 949     }
 950
 951     client->send_coroutine = NULL;
 952     nbd_set_handlers(client);
 953     qemu_co_mutex_unlock(&client->send_lock);
 954     return rc;
 955 }
 956
 957 static ssize_t nbd_co_receive_request(NBDRequest *req, struct nbd_request *request)
 958 {
 959     NBDClient *client = req->client;
 960     uint32_t command;
 961     ssize_t rc;
 962
 963     g_assert(qemu_in_coroutine());
 964     client->recv_coroutine = qemu_coroutine_self();
 965     nbd_update_can_read(client);
 966
 967     rc = nbd_receive_request(client->ioc, request);
 968     if (rc < 0) {
 969         if (rc != -EAGAIN) {
 970             rc = -EIO;
 971         }
 972         goto out;
 973     }
 974
 975     if ((request->from + request->len) < request->from) {
 976         LOG("integer overflow detected! "
 977             "you're probably being attacked");
 978         rc = -EINVAL;
 979         goto out;
 980     }
 981
 982     TRACE("Decoding type");
 983
 984     command = request->type & NBD_CMD_MASK_COMMAND;
 985     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
 986         if (request->len > NBD_MAX_BUFFER_SIZE) {
 987             LOG("len (%u) is larger than max len (%u)",
 988                 request->len, NBD_MAX_BUFFER_SIZE);
 989             rc = -EINVAL;
 990             goto out;
 991         }
 992
 993         req->data = blk_try_blockalign(client->exp->blk, request->len);
 994         if (req->data == NULL) {
 995             rc = -ENOMEM;
 996             goto out;
 997         }
 998     }
 999     if (command == NBD_CMD_WRITE) {
1000         TRACE("Reading %u byte(s)", request->len);
1001
1002         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1003             LOG("reading from socket failed");
1004             rc = -EIO;
1005             goto out;
1006         }
1007     }
1008     rc = 0;
1009
1010 out:
1011     client->recv_coroutine = NULL;
1012     nbd_update_can_read(client);
1013
1014     return rc;
1015 }
1016
1017 static void nbd_trip(void *opaque)
1018 {
1019     NBDClient *client = opaque;
1020     NBDExport *exp = client->exp;
1021     NBDRequest *req;
1022     struct nbd_request request;
1023     struct nbd_reply reply;
1024     ssize_t ret;
1025     uint32_t command;
1026
1027     TRACE("Reading request.");
1028     if (client->closing) {
1029         return;
1030     }
1031
1032     req = nbd_request_get(client);
1033     ret = nbd_co_receive_request(req, &request);
1034     if (ret == -EAGAIN) {
1035         goto done;
1036     }
1037     if (ret == -EIO) {
1038         goto out;
1039     }
1040
1041     reply.handle = request.handle;
1042     reply.error = 0;
1043
1044     if (ret < 0) {
1045         reply.error = -ret;
1046         goto error_reply;
1047     }
1048     command = request.type & NBD_CMD_MASK_COMMAND;
1049     if (command != NBD_CMD_DISC && (request.from + request.len) > exp->size) {
1050             LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
1051             ", Offset: %" PRIu64 "\n",
1052                     request.from, request.len,
1053                     (uint64_t)exp->size, (uint64_t)exp->dev_offset);
1054         LOG("requested operation past EOF--bad client?");
1055         goto invalid_request;
1056     }
1057
1058     if (client->closing) {
1059         /*
1060          * The client may be closed when we are blocked in
1061          * nbd_co_receive_request()
1062          */
1063         goto done;
1064     }
1065
1066     switch (command) {
1067     case NBD_CMD_READ:
1068         TRACE("Request type is READ");
1069
1070         if (request.type & NBD_CMD_FLAG_FUA) {
1071             ret = blk_co_flush(exp->blk);
1072             if (ret < 0) {
1073                 LOG("flush failed");
1074                 reply.error = -ret;
1075                 goto error_reply;
1076             }
1077         }
1078
1079         ret = blk_read(exp->blk,
1080                        (request.from + exp->dev_offset) / BDRV_SECTOR_SIZE,
1081                        req->data, request.len / BDRV_SECTOR_SIZE);
1082         if (ret < 0) {
1083             LOG("reading from file failed");
1084             reply.error = -ret;
1085             goto error_reply;
1086         }
1087
1088         TRACE("Read %u byte(s)", request.len);
1089         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1090             goto out;
1091         break;
1092     case NBD_CMD_WRITE:
1093         TRACE("Request type is WRITE");
1094
1095         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1096             TRACE("Server is read-only, return error");
1097             reply.error = EROFS;
1098             goto error_reply;
1099         }
1100
1101         TRACE("Writing to device");
1102
1103         ret = blk_write(exp->blk,
1104                         (request.from + exp->dev_offset) / BDRV_SECTOR_SIZE,
1105                         req->data, request.len / BDRV_SECTOR_SIZE);
1106         if (ret < 0) {
1107             LOG("writing to file failed");
1108             reply.error = -ret;
1109             goto error_reply;
1110         }
1111
1112         if (request.type & NBD_CMD_FLAG_FUA) {
1113             ret = blk_co_flush(exp->blk);
1114             if (ret < 0) {
1115                 LOG("flush failed");
1116                 reply.error = -ret;
1117                 goto error_reply;
1118             }
1119         }
1120
1121         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1122             goto out;
1123         }
1124         break;
1125     case NBD_CMD_DISC:
1126         TRACE("Request type is DISCONNECT");
1127         errno = 0;
1128         goto out;
1129     case NBD_CMD_FLUSH:
1130         TRACE("Request type is FLUSH");
1131
1132         ret = blk_co_flush(exp->blk);
1133         if (ret < 0) {
1134             LOG("flush failed");
1135             reply.error = -ret;
1136         }
1137         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1138             goto out;
1139         }
1140         break;
1141     case NBD_CMD_TRIM:
1142         TRACE("Request type is TRIM");
1143         ret = blk_co_discard(exp->blk, (request.from + exp->dev_offset)
1144                                        / BDRV_SECTOR_SIZE,
1145                              request.len / BDRV_SECTOR_SIZE);
1146         if (ret < 0) {
1147             LOG("discard failed");
1148             reply.error = -ret;
1149         }
1150         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1151             goto out;
1152         }
1153         break;
1154     default:
1155         LOG("invalid request type (%u) received", request.type);
1156     invalid_request:
1157         reply.error = EINVAL;
1158     error_reply:
1159         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1160             goto out;
1161         }
1162         break;
1163     }
1164
1165     TRACE("Request/Reply complete");
1166
1167 done:
1168     nbd_request_put(req);
1169     return;
1170
1171 out:
1172     nbd_request_put(req);
1173     client_close(client);
1174 }
1175
1176 static void nbd_read(void *opaque)
1177 {
1178     NBDClient *client = opaque;
1179
1180     if (client->recv_coroutine) {
1181         qemu_coroutine_enter(client->recv_coroutine, NULL);
1182     } else {
1183         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip), client);
1184     }
1185 }
1186
1187 static void nbd_restart_write(void *opaque)
1188 {
1189     NBDClient *client = opaque;
1190
1191     qemu_coroutine_enter(client->send_coroutine, NULL);
1192 }
1193
1194 static void nbd_set_handlers(NBDClient *client)
1195 {
1196     if (client->exp && client->exp->ctx) {
1197         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1198                            true,
1199                            client->can_read ? nbd_read : NULL,
1200                            client->send_coroutine ? nbd_restart_write : NULL,
1201                            client);
1202     }
1203 }
1204
1205 static void nbd_unset_handlers(NBDClient *client)
1206 {
1207     if (client->exp && client->exp->ctx) {
1208         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1209                            true, NULL, NULL, NULL);
1210     }
1211 }
1212
1213 static void nbd_update_can_read(NBDClient *client)
1214 {
1215     bool can_read = client->recv_coroutine ||
1216                     client->nb_requests < MAX_NBD_REQUESTS;
1217
1218     if (can_read != client->can_read) {
1219         client->can_read = can_read;
1220         nbd_set_handlers(client);
1221
1222         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1223          * in nbd_set_handlers() will have taken care of that */
1224     }
1225 }
1226
1227 static coroutine_fn void nbd_co_client_start(void *opaque)
1228 {
1229     NBDClientNewData *data = opaque;
1230     NBDClient *client = data->client;
1231     NBDExport *exp = client->exp;
1232
1233     if (exp) {
1234         nbd_export_get(exp);
1235     }
1236     if (nbd_negotiate(data)) {
1237         client_close(client);
1238         goto out;
1239     }
1240     qemu_co_mutex_init(&client->send_lock);
1241     nbd_set_handlers(client);
1242
1243     if (exp) {
1244         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1245     }
1246 out:
1247     g_free(data);
1248 }
1249
1250 void nbd_client_new(NBDExport *exp,
1251                     QIOChannelSocket *sioc,
1252                     QCryptoTLSCreds *tlscreds,
1253                     const char *tlsaclname,
1254                     void (*close_fn)(NBDClient *))
1255 {
1256     NBDClient *client;
1257     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1258
1259     client = g_malloc0(sizeof(NBDClient));
1260     client->refcount = 1;
1261     client->exp = exp;
1262     client->tlscreds = tlscreds;
1263     if (tlscreds) {
1264         object_ref(OBJECT(client->tlscreds));
1265     }
1266     client->tlsaclname = g_strdup(tlsaclname);
1267     client->sioc = sioc;
1268     object_ref(OBJECT(client->sioc));
1269     client->ioc = QIO_CHANNEL(sioc);
1270     object_ref(OBJECT(client->ioc));
1271     client->can_read = true;
1272     client->close = close_fn;
1273
1274     data->client = client;
1275     data->co = qemu_coroutine_create(nbd_co_client_start);
1276     qemu_coroutine_enter(data->co, data);
1277 }