nbd/server.c

   1 /*
   2  *  Copyright (C) 2005  Anthony Liguori <[email protected]>
   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             case NBD_OPT_EXPORT_NAME:
 453                 /* No way to return an error to client, so drop connection */
 454                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 455                 return -EINVAL;
 456
 457             default:
 458                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 459                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 460                     return -EIO;
 461                 }
 462                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_TLS_REQD,
 463                                        clientflags);
 464                 break;
 465             }
 466         } else if (fixedNewstyle) {
 467             switch (clientflags) {
 468             case NBD_OPT_LIST:
 469                 ret = nbd_negotiate_handle_list(client, length);
 470                 if (ret < 0) {
 471                     return ret;
 472                 }
 473                 break;
 474
 475             case NBD_OPT_ABORT:
 476                 return -EINVAL;
 477
 478             case NBD_OPT_EXPORT_NAME:
 479                 return nbd_negotiate_handle_export_name(client, length);
 480
 481             case NBD_OPT_STARTTLS:
 482                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 483                     return -EIO;
 484                 }
 485                 if (client->tlscreds) {
 486                     TRACE("TLS already enabled");
 487                     nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_INVALID,
 488                                            clientflags);
 489                 } else {
 490                     TRACE("TLS not configured");
 491                     nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_POLICY,
 492                                            clientflags);
 493                 }
 494                 break;
 495             default:
 496                 TRACE("Unsupported option 0x%x", clientflags);
 497                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 498                     return -EIO;
 499                 }
 500                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_UNSUP,
 501                                        clientflags);
 502                 break;
 503             }
 504         } else {
 505             /*
 506              * If broken new-style we should drop the connection
 507              * for anything except NBD_OPT_EXPORT_NAME
 508              */
 509             switch (clientflags) {
 510             case NBD_OPT_EXPORT_NAME:
 511                 return nbd_negotiate_handle_export_name(client, length);
 512
 513             default:
 514                 TRACE("Unsupported option 0x%x", clientflags);
 515                 return -EINVAL;
 516             }
 517         }
 518     }
 519 }
 520
 521 typedef struct {
 522     NBDClient *client;
 523     Coroutine *co;
 524 } NBDClientNewData;
 525
 526 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 527 {
 528     NBDClient *client = data->client;
 529     char buf[8 + 8 + 8 + 128];
 530     int rc;
 531     const int myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 532                          NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 533     bool oldStyle;
 534
 535     /* Old style negotiation header without options
 536         [ 0 ..   7]   passwd       ("NBDMAGIC")
 537         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 538         [16 ..  23]   size
 539         [24 ..  25]   server flags (0)
 540         [26 ..  27]   export flags
 541         [28 .. 151]   reserved     (0)
 542
 543        New style negotiation header with options
 544         [ 0 ..   7]   passwd       ("NBDMAGIC")
 545         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 546         [16 ..  17]   server flags (0)
 547         ....options sent....
 548         [18 ..  25]   size
 549         [26 ..  27]   export flags
 550         [28 .. 151]   reserved     (0)
 551      */
 552
 553     qio_channel_set_blocking(client->ioc, false, NULL);
 554     rc = -EINVAL;
 555
 556     TRACE("Beginning negotiation.");
 557     memset(buf, 0, sizeof(buf));
 558     memcpy(buf, "NBDMAGIC", 8);
 559
 560     oldStyle = client->exp != NULL && !client->tlscreds;
 561     if (oldStyle) {
 562         assert ((client->exp->nbdflags & ~65535) == 0);
 563         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 564         stq_be_p(buf + 16, client->exp->size);
 565         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 566     } else {
 567         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 568         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE);
 569     }
 570
 571     if (oldStyle) {
 572         if (client->tlscreds) {
 573             TRACE("TLS cannot be enabled with oldstyle protocol");
 574             goto fail;
 575         }
 576         if (nbd_negotiate_write(client->ioc, buf, sizeof(buf)) != sizeof(buf)) {
 577             LOG("write failed");
 578             goto fail;
 579         }
 580     } else {
 581         if (nbd_negotiate_write(client->ioc, buf, 18) != 18) {
 582             LOG("write failed");
 583             goto fail;
 584         }
 585         rc = nbd_negotiate_options(client);
 586         if (rc != 0) {
 587             LOG("option negotiation failed");
 588             goto fail;
 589         }
 590
 591         assert ((client->exp->nbdflags & ~65535) == 0);
 592         stq_be_p(buf + 18, client->exp->size);
 593         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 594         if (nbd_negotiate_write(client->ioc, buf + 18, sizeof(buf) - 18) !=
 595             sizeof(buf) - 18) {
 596             LOG("write failed");
 597             goto fail;
 598         }
 599     }
 600
 601     TRACE("Negotiation succeeded.");
 602     rc = 0;
 603 fail:
 604     return rc;
 605 }
 606
 607 #ifdef __linux__
 608
 609 int nbd_disconnect(int fd)
 610 {
 611     ioctl(fd, NBD_CLEAR_QUE);
 612     ioctl(fd, NBD_DISCONNECT);
 613     ioctl(fd, NBD_CLEAR_SOCK);
 614     return 0;
 615 }
 616
 617 #else
 618
 619 int nbd_disconnect(int fd)
 620 {
 621     return -ENOTSUP;
 622 }
 623 #endif
 624
 625 static ssize_t nbd_receive_request(QIOChannel *ioc, struct nbd_request *request)
 626 {
 627     uint8_t buf[NBD_REQUEST_SIZE];
 628     uint32_t magic;
 629     ssize_t ret;
 630
 631     ret = read_sync(ioc, buf, sizeof(buf));
 632     if (ret < 0) {
 633         return ret;
 634     }
 635
 636     if (ret != sizeof(buf)) {
 637         LOG("read failed");
 638         return -EINVAL;
 639     }
 640
 641     /* Request
 642        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 643        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 644        [ 8 .. 15]   handle
 645        [16 .. 23]   from
 646        [24 .. 27]   len
 647      */
 648
 649     magic = be32_to_cpup((uint32_t*)buf);
 650     request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 651     request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 652     request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 653     request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 654
 655     TRACE("Got request: "
 656           "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 657           magic, request->type, request->from, request->len);
 658
 659     if (magic != NBD_REQUEST_MAGIC) {
 660         LOG("invalid magic (got 0x%x)", magic);
 661         return -EINVAL;
 662     }
 663     return 0;
 664 }
 665
 666 static ssize_t nbd_send_reply(QIOChannel *ioc, struct nbd_reply *reply)
 667 {
 668     uint8_t buf[NBD_REPLY_SIZE];
 669     ssize_t ret;
 670
 671     reply->error = system_errno_to_nbd_errno(reply->error);
 672
 673     TRACE("Sending response to client: { .error = %d, handle = %" PRIu64 " }",
 674           reply->error, reply->handle);
 675
 676     /* Reply
 677        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 678        [ 4 ..  7]    error   (0 == no error)
 679        [ 7 .. 15]    handle
 680      */
 681     stl_be_p(buf, NBD_REPLY_MAGIC);
 682     stl_be_p(buf + 4, reply->error);
 683     stq_be_p(buf + 8, reply->handle);
 684
 685     ret = write_sync(ioc, buf, sizeof(buf));
 686     if (ret < 0) {
 687         return ret;
 688     }
 689
 690     if (ret != sizeof(buf)) {
 691         LOG("writing to socket failed");
 692         return -EINVAL;
 693     }
 694     return 0;
 695 }
 696
 697 #define MAX_NBD_REQUESTS 16
 698
 699 void nbd_client_get(NBDClient *client)
 700 {
 701     client->refcount++;
 702 }
 703
 704 void nbd_client_put(NBDClient *client)
 705 {
 706     if (--client->refcount == 0) {
 707         /* The last reference should be dropped by client->close,
 708          * which is called by client_close.
 709          */
 710         assert(client->closing);
 711
 712         nbd_unset_handlers(client);
 713         object_unref(OBJECT(client->sioc));
 714         object_unref(OBJECT(client->ioc));
 715         if (client->tlscreds) {
 716             object_unref(OBJECT(client->tlscreds));
 717         }
 718         g_free(client->tlsaclname);
 719         if (client->exp) {
 720             QTAILQ_REMOVE(&client->exp->clients, client, next);
 721             nbd_export_put(client->exp);
 722         }
 723         g_free(client);
 724     }
 725 }
 726
 727 static void client_close(NBDClient *client)
 728 {
 729     if (client->closing) {
 730         return;
 731     }
 732
 733     client->closing = true;
 734
 735     /* Force requests to finish.  They will drop their own references,
 736      * then we'll close the socket and free the NBDClient.
 737      */
 738     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 739                          NULL);
 740
 741     /* Also tell the client, so that they release their reference.  */
 742     if (client->close) {
 743         client->close(client);
 744     }
 745 }
 746
 747 static NBDRequest *nbd_request_get(NBDClient *client)
 748 {
 749     NBDRequest *req;
 750
 751     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 752     client->nb_requests++;
 753     nbd_update_can_read(client);
 754
 755     req = g_new0(NBDRequest, 1);
 756     nbd_client_get(client);
 757     req->client = client;
 758     return req;
 759 }
 760
 761 static void nbd_request_put(NBDRequest *req)
 762 {
 763     NBDClient *client = req->client;
 764
 765     if (req->data) {
 766         qemu_vfree(req->data);
 767     }
 768     g_free(req);
 769
 770     client->nb_requests--;
 771     nbd_update_can_read(client);
 772     nbd_client_put(client);
 773 }
 774
 775 static void blk_aio_attached(AioContext *ctx, void *opaque)
 776 {
 777     NBDExport *exp = opaque;
 778     NBDClient *client;
 779
 780     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 781
 782     exp->ctx = ctx;
 783
 784     QTAILQ_FOREACH(client, &exp->clients, next) {
 785         nbd_set_handlers(client);
 786     }
 787 }
 788
 789 static void blk_aio_detach(void *opaque)
 790 {
 791     NBDExport *exp = opaque;
 792     NBDClient *client;
 793
 794     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 795
 796     QTAILQ_FOREACH(client, &exp->clients, next) {
 797         nbd_unset_handlers(client);
 798     }
 799
 800     exp->ctx = NULL;
 801 }
 802
 803 static void nbd_eject_notifier(Notifier *n, void *data)
 804 {
 805     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 806     nbd_export_close(exp);
 807 }
 808
 809 NBDExport *nbd_export_new(BlockBackend *blk, off_t dev_offset, off_t size,
 810                           uint32_t nbdflags, void (*close)(NBDExport *),
 811                           Error **errp)
 812 {
 813     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 814     exp->refcount = 1;
 815     QTAILQ_INIT(&exp->clients);
 816     exp->blk = blk;
 817     exp->dev_offset = dev_offset;
 818     exp->nbdflags = nbdflags;
 819     exp->size = size < 0 ? blk_getlength(blk) : size;
 820     if (exp->size < 0) {
 821         error_setg_errno(errp, -exp->size,
 822                          "Failed to determine the NBD export's length");
 823         goto fail;
 824     }
 825     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 826
 827     exp->close = close;
 828     exp->ctx = blk_get_aio_context(blk);
 829     blk_ref(blk);
 830     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 831
 832     exp->eject_notifier.notify = nbd_eject_notifier;
 833     blk_add_remove_bs_notifier(blk, &exp->eject_notifier);
 834
 835     /*
 836      * NBD exports are used for non-shared storage migration.  Make sure
 837      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 838      * access since the export could be available before migration handover.
 839      */
 840     aio_context_acquire(exp->ctx);
 841     blk_invalidate_cache(blk, NULL);
 842     aio_context_release(exp->ctx);
 843     return exp;
 844
 845 fail:
 846     g_free(exp);
 847     return NULL;
 848 }
 849
 850 NBDExport *nbd_export_find(const char *name)
 851 {
 852     NBDExport *exp;
 853     QTAILQ_FOREACH(exp, &exports, next) {
 854         if (strcmp(name, exp->name) == 0) {
 855             return exp;
 856         }
 857     }
 858
 859     return NULL;
 860 }
 861
 862 void nbd_export_set_name(NBDExport *exp, const char *name)
 863 {
 864     if (exp->name == name) {
 865         return;
 866     }
 867
 868     nbd_export_get(exp);
 869     if (exp->name != NULL) {
 870         g_free(exp->name);
 871         exp->name = NULL;
 872         QTAILQ_REMOVE(&exports, exp, next);
 873         nbd_export_put(exp);
 874     }
 875     if (name != NULL) {
 876         nbd_export_get(exp);
 877         exp->name = g_strdup(name);
 878         QTAILQ_INSERT_TAIL(&exports, exp, next);
 879     }
 880     nbd_export_put(exp);
 881 }
 882
 883 void nbd_export_close(NBDExport *exp)
 884 {
 885     NBDClient *client, *next;
 886
 887     nbd_export_get(exp);
 888     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 889         client_close(client);
 890     }
 891     nbd_export_set_name(exp, NULL);
 892     nbd_export_put(exp);
 893 }
 894
 895 void nbd_export_get(NBDExport *exp)
 896 {
 897     assert(exp->refcount > 0);
 898     exp->refcount++;
 899 }
 900
 901 void nbd_export_put(NBDExport *exp)
 902 {
 903     assert(exp->refcount > 0);
 904     if (exp->refcount == 1) {
 905         nbd_export_close(exp);
 906     }
 907
 908     if (--exp->refcount == 0) {
 909         assert(exp->name == NULL);
 910
 911         if (exp->close) {
 912             exp->close(exp);
 913         }
 914
 915         if (exp->blk) {
 916             notifier_remove(&exp->eject_notifier);
 917             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 918                                             blk_aio_detach, exp);
 919             blk_unref(exp->blk);
 920             exp->blk = NULL;
 921         }
 922
 923         g_free(exp);
 924     }
 925 }
 926
 927 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
 928 {
 929     return exp->blk;
 930 }
 931
 932 void nbd_export_close_all(void)
 933 {
 934     NBDExport *exp, *next;
 935
 936     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 937         nbd_export_close(exp);
 938     }
 939 }
 940
 941 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 942                                  int len)
 943 {
 944     NBDClient *client = req->client;
 945     ssize_t rc, ret;
 946
 947     g_assert(qemu_in_coroutine());
 948     qemu_co_mutex_lock(&client->send_lock);
 949     client->send_coroutine = qemu_coroutine_self();
 950     nbd_set_handlers(client);
 951
 952     if (!len) {
 953         rc = nbd_send_reply(client->ioc, reply);
 954     } else {
 955         qio_channel_set_cork(client->ioc, true);
 956         rc = nbd_send_reply(client->ioc, reply);
 957         if (rc >= 0) {
 958             ret = write_sync(client->ioc, req->data, len);
 959             if (ret != len) {
 960                 rc = -EIO;
 961             }
 962         }
 963         qio_channel_set_cork(client->ioc, false);
 964     }
 965
 966     client->send_coroutine = NULL;
 967     nbd_set_handlers(client);
 968     qemu_co_mutex_unlock(&client->send_lock);
 969     return rc;
 970 }
 971
 972 static ssize_t nbd_co_receive_request(NBDRequest *req, struct nbd_request *request)
 973 {
 974     NBDClient *client = req->client;
 975     uint32_t command;
 976     ssize_t rc;
 977
 978     g_assert(qemu_in_coroutine());
 979     client->recv_coroutine = qemu_coroutine_self();
 980     nbd_update_can_read(client);
 981
 982     rc = nbd_receive_request(client->ioc, request);
 983     if (rc < 0) {
 984         if (rc != -EAGAIN) {
 985             rc = -EIO;
 986         }
 987         goto out;
 988     }
 989
 990     if ((request->from + request->len) < request->from) {
 991         LOG("integer overflow detected! "
 992             "you're probably being attacked");
 993         rc = -EINVAL;
 994         goto out;
 995     }
 996
 997     TRACE("Decoding type");
 998
 999     command = request->type & NBD_CMD_MASK_COMMAND;
1000     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
1001         if (request->len > NBD_MAX_BUFFER_SIZE) {
1002             LOG("len (%u) is larger than max len (%u)",
1003                 request->len, NBD_MAX_BUFFER_SIZE);
1004             rc = -EINVAL;
1005             goto out;
1006         }
1007
1008         req->data = blk_try_blockalign(client->exp->blk, request->len);
1009         if (req->data == NULL) {
1010             rc = -ENOMEM;
1011             goto out;
1012         }
1013     }
1014     if (command == NBD_CMD_WRITE) {
1015         TRACE("Reading %u byte(s)", request->len);
1016
1017         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1018             LOG("reading from socket failed");
1019             rc = -EIO;
1020             goto out;
1021         }
1022     }
1023     rc = 0;
1024
1025 out:
1026     client->recv_coroutine = NULL;
1027     nbd_update_can_read(client);
1028
1029     return rc;
1030 }
1031
1032 static void nbd_trip(void *opaque)
1033 {
1034     NBDClient *client = opaque;
1035     NBDExport *exp = client->exp;
1036     NBDRequest *req;
1037     struct nbd_request request;
1038     struct nbd_reply reply;
1039     ssize_t ret;
1040     uint32_t command;
1041
1042     TRACE("Reading request.");
1043     if (client->closing) {
1044         return;
1045     }
1046
1047     req = nbd_request_get(client);
1048     ret = nbd_co_receive_request(req, &request);
1049     if (ret == -EAGAIN) {
1050         goto done;
1051     }
1052     if (ret == -EIO) {
1053         goto out;
1054     }
1055
1056     reply.handle = request.handle;
1057     reply.error = 0;
1058
1059     if (ret < 0) {
1060         reply.error = -ret;
1061         goto error_reply;
1062     }
1063     command = request.type & NBD_CMD_MASK_COMMAND;
1064     if (command != NBD_CMD_DISC && (request.from + request.len) > exp->size) {
1065             LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
1066             ", Offset: %" PRIu64 "\n",
1067                     request.from, request.len,
1068                     (uint64_t)exp->size, (uint64_t)exp->dev_offset);
1069         LOG("requested operation past EOF--bad client?");
1070         goto invalid_request;
1071     }
1072
1073     if (client->closing) {
1074         /*
1075          * The client may be closed when we are blocked in
1076          * nbd_co_receive_request()
1077          */
1078         goto done;
1079     }
1080
1081     switch (command) {
1082     case NBD_CMD_READ:
1083         TRACE("Request type is READ");
1084
1085         if (request.type & NBD_CMD_FLAG_FUA) {
1086             ret = blk_co_flush(exp->blk);
1087             if (ret < 0) {
1088                 LOG("flush failed");
1089                 reply.error = -ret;
1090                 goto error_reply;
1091             }
1092         }
1093
1094         ret = blk_pread(exp->blk, request.from + exp->dev_offset,
1095                         req->data, request.len);
1096         if (ret < 0) {
1097             LOG("reading from file failed");
1098             reply.error = -ret;
1099             goto error_reply;
1100         }
1101
1102         TRACE("Read %u byte(s)", request.len);
1103         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1104             goto out;
1105         break;
1106     case NBD_CMD_WRITE:
1107         TRACE("Request type is WRITE");
1108
1109         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1110             TRACE("Server is read-only, return error");
1111             reply.error = EROFS;
1112             goto error_reply;
1113         }
1114
1115         TRACE("Writing to device");
1116
1117         ret = blk_pwrite(exp->blk, request.from + exp->dev_offset,
1118                          req->data, request.len, 0);
1119         if (ret < 0) {
1120             LOG("writing to file failed");
1121             reply.error = -ret;
1122             goto error_reply;
1123         }
1124
1125         if (request.type & NBD_CMD_FLAG_FUA) {
1126             ret = blk_co_flush(exp->blk);
1127             if (ret < 0) {
1128                 LOG("flush failed");
1129                 reply.error = -ret;
1130                 goto error_reply;
1131             }
1132         }
1133
1134         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1135             goto out;
1136         }
1137         break;
1138     case NBD_CMD_DISC:
1139         TRACE("Request type is DISCONNECT");
1140         errno = 0;
1141         goto out;
1142     case NBD_CMD_FLUSH:
1143         TRACE("Request type is FLUSH");
1144
1145         ret = blk_co_flush(exp->blk);
1146         if (ret < 0) {
1147             LOG("flush failed");
1148             reply.error = -ret;
1149         }
1150         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1151             goto out;
1152         }
1153         break;
1154     case NBD_CMD_TRIM:
1155         TRACE("Request type is TRIM");
1156         ret = blk_co_discard(exp->blk, (request.from + exp->dev_offset)
1157                                        / BDRV_SECTOR_SIZE,
1158                              request.len / BDRV_SECTOR_SIZE);
1159         if (ret < 0) {
1160             LOG("discard failed");
1161             reply.error = -ret;
1162         }
1163         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1164             goto out;
1165         }
1166         break;
1167     default:
1168         LOG("invalid request type (%u) received", request.type);
1169     invalid_request:
1170         reply.error = EINVAL;
1171     error_reply:
1172         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1173             goto out;
1174         }
1175         break;
1176     }
1177
1178     TRACE("Request/Reply complete");
1179
1180 done:
1181     nbd_request_put(req);
1182     return;
1183
1184 out:
1185     nbd_request_put(req);
1186     client_close(client);
1187 }
1188
1189 static void nbd_read(void *opaque)
1190 {
1191     NBDClient *client = opaque;
1192
1193     if (client->recv_coroutine) {
1194         qemu_coroutine_enter(client->recv_coroutine, NULL);
1195     } else {
1196         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip), client);
1197     }
1198 }
1199
1200 static void nbd_restart_write(void *opaque)
1201 {
1202     NBDClient *client = opaque;
1203
1204     qemu_coroutine_enter(client->send_coroutine, NULL);
1205 }
1206
1207 static void nbd_set_handlers(NBDClient *client)
1208 {
1209     if (client->exp && client->exp->ctx) {
1210         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1211                            true,
1212                            client->can_read ? nbd_read : NULL,
1213                            client->send_coroutine ? nbd_restart_write : NULL,
1214                            client);
1215     }
1216 }
1217
1218 static void nbd_unset_handlers(NBDClient *client)
1219 {
1220     if (client->exp && client->exp->ctx) {
1221         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1222                            true, NULL, NULL, NULL);
1223     }
1224 }
1225
1226 static void nbd_update_can_read(NBDClient *client)
1227 {
1228     bool can_read = client->recv_coroutine ||
1229                     client->nb_requests < MAX_NBD_REQUESTS;
1230
1231     if (can_read != client->can_read) {
1232         client->can_read = can_read;
1233         nbd_set_handlers(client);
1234
1235         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1236          * in nbd_set_handlers() will have taken care of that */
1237     }
1238 }
1239
1240 static coroutine_fn void nbd_co_client_start(void *opaque)
1241 {
1242     NBDClientNewData *data = opaque;
1243     NBDClient *client = data->client;
1244     NBDExport *exp = client->exp;
1245
1246     if (exp) {
1247         nbd_export_get(exp);
1248     }
1249     if (nbd_negotiate(data)) {
1250         client_close(client);
1251         goto out;
1252     }
1253     qemu_co_mutex_init(&client->send_lock);
1254     nbd_set_handlers(client);
1255
1256     if (exp) {
1257         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1258     }
1259 out:
1260     g_free(data);
1261 }
1262
1263 void nbd_client_new(NBDExport *exp,
1264                     QIOChannelSocket *sioc,
1265                     QCryptoTLSCreds *tlscreds,
1266                     const char *tlsaclname,
1267                     void (*close_fn)(NBDClient *))
1268 {
1269     NBDClient *client;
1270     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1271
1272     client = g_malloc0(sizeof(NBDClient));
1273     client->refcount = 1;
1274     client->exp = exp;
1275     client->tlscreds = tlscreds;
1276     if (tlscreds) {
1277         object_ref(OBJECT(client->tlscreds));
1278     }
1279     client->tlsaclname = g_strdup(tlsaclname);
1280     client->sioc = sioc;
1281     object_ref(OBJECT(client->sioc));
1282     client->ioc = QIO_CHANNEL(sioc);
1283     object_ref(OBJECT(client->ioc));
1284     client->can_read = true;
1285     client->close = close_fn;
1286
1287     data->client = client;
1288     data->co = qemu_coroutine_create(nbd_co_client_start);
1289     qemu_coroutine_enter(data->co, data);
1290 }