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=%" PRIx32 " type=%" PRIx32, 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("Advertising 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("Client 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%" PRIx32 " 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%" PRIx32, clientflags);
 435         if (client->tlscreds &&
 436             client->ioc == (QIOChannel *)client->sioc) {
 437             QIOChannel *tioc;
 438             if (!fixedNewstyle) {
 439                 TRACE("Unsupported option 0x%" PRIx32, 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%" PRIx32 " not permitted before TLS",
 459                       clientflags);
 460                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 461                     return -EIO;
 462                 }
 463                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_TLS_REQD,
 464                                        clientflags);
 465                 break;
 466             }
 467         } else if (fixedNewstyle) {
 468             switch (clientflags) {
 469             case NBD_OPT_LIST:
 470                 ret = nbd_negotiate_handle_list(client, length);
 471                 if (ret < 0) {
 472                     return ret;
 473                 }
 474                 break;
 475
 476             case NBD_OPT_ABORT:
 477                 return -EINVAL;
 478
 479             case NBD_OPT_EXPORT_NAME:
 480                 return nbd_negotiate_handle_export_name(client, length);
 481
 482             case NBD_OPT_STARTTLS:
 483                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 484                     return -EIO;
 485                 }
 486                 if (client->tlscreds) {
 487                     TRACE("TLS already enabled");
 488                     nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_INVALID,
 489                                            clientflags);
 490                 } else {
 491                     TRACE("TLS not configured");
 492                     nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_POLICY,
 493                                            clientflags);
 494                 }
 495                 break;
 496             default:
 497                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 498                 if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 499                     return -EIO;
 500                 }
 501                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_UNSUP,
 502                                        clientflags);
 503                 break;
 504             }
 505         } else {
 506             /*
 507              * If broken new-style we should drop the connection
 508              * for anything except NBD_OPT_EXPORT_NAME
 509              */
 510             switch (clientflags) {
 511             case NBD_OPT_EXPORT_NAME:
 512                 return nbd_negotiate_handle_export_name(client, length);
 513
 514             default:
 515                 TRACE("Unsupported option 0x%" PRIx32, clientflags);
 516                 return -EINVAL;
 517             }
 518         }
 519     }
 520 }
 521
 522 typedef struct {
 523     NBDClient *client;
 524     Coroutine *co;
 525 } NBDClientNewData;
 526
 527 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 528 {
 529     NBDClient *client = data->client;
 530     char buf[8 + 8 + 8 + 128];
 531     int rc;
 532     const int myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 533                          NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 534     bool oldStyle;
 535
 536     /* Old style negotiation header without options
 537         [ 0 ..   7]   passwd       ("NBDMAGIC")
 538         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 539         [16 ..  23]   size
 540         [24 ..  25]   server flags (0)
 541         [26 ..  27]   export flags
 542         [28 .. 151]   reserved     (0)
 543
 544        New style negotiation header with options
 545         [ 0 ..   7]   passwd       ("NBDMAGIC")
 546         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 547         [16 ..  17]   server flags (0)
 548         ....options sent....
 549         [18 ..  25]   size
 550         [26 ..  27]   export flags
 551         [28 .. 151]   reserved     (0)
 552      */
 553
 554     qio_channel_set_blocking(client->ioc, false, NULL);
 555     rc = -EINVAL;
 556
 557     TRACE("Beginning negotiation.");
 558     memset(buf, 0, sizeof(buf));
 559     memcpy(buf, "NBDMAGIC", 8);
 560
 561     oldStyle = client->exp != NULL && !client->tlscreds;
 562     if (oldStyle) {
 563         assert ((client->exp->nbdflags & ~65535) == 0);
 564         TRACE("advertising size %" PRIu64 " and flags %x",
 565               client->exp->size, client->exp->nbdflags | myflags);
 566         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 567         stq_be_p(buf + 16, client->exp->size);
 568         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 569     } else {
 570         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 571         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE);
 572     }
 573
 574     if (oldStyle) {
 575         if (client->tlscreds) {
 576             TRACE("TLS cannot be enabled with oldstyle protocol");
 577             goto fail;
 578         }
 579         if (nbd_negotiate_write(client->ioc, buf, sizeof(buf)) != sizeof(buf)) {
 580             LOG("write failed");
 581             goto fail;
 582         }
 583     } else {
 584         if (nbd_negotiate_write(client->ioc, buf, 18) != 18) {
 585             LOG("write failed");
 586             goto fail;
 587         }
 588         rc = nbd_negotiate_options(client);
 589         if (rc != 0) {
 590             LOG("option negotiation failed");
 591             goto fail;
 592         }
 593
 594         assert ((client->exp->nbdflags & ~65535) == 0);
 595         TRACE("advertising size %" PRIu64 " and flags %x",
 596               client->exp->size, client->exp->nbdflags | myflags);
 597         stq_be_p(buf + 18, client->exp->size);
 598         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 599         if (nbd_negotiate_write(client->ioc, buf + 18, sizeof(buf) - 18) !=
 600             sizeof(buf) - 18) {
 601             LOG("write failed");
 602             goto fail;
 603         }
 604     }
 605
 606     TRACE("Negotiation succeeded.");
 607     rc = 0;
 608 fail:
 609     return rc;
 610 }
 611
 612 #ifdef __linux__
 613
 614 int nbd_disconnect(int fd)
 615 {
 616     ioctl(fd, NBD_CLEAR_QUE);
 617     ioctl(fd, NBD_DISCONNECT);
 618     ioctl(fd, NBD_CLEAR_SOCK);
 619     return 0;
 620 }
 621
 622 #else
 623
 624 int nbd_disconnect(int fd)
 625 {
 626     return -ENOTSUP;
 627 }
 628 #endif
 629
 630 static ssize_t nbd_receive_request(QIOChannel *ioc, struct nbd_request *request)
 631 {
 632     uint8_t buf[NBD_REQUEST_SIZE];
 633     uint32_t magic;
 634     ssize_t ret;
 635
 636     ret = read_sync(ioc, buf, sizeof(buf));
 637     if (ret < 0) {
 638         return ret;
 639     }
 640
 641     if (ret != sizeof(buf)) {
 642         LOG("read failed");
 643         return -EINVAL;
 644     }
 645
 646     /* Request
 647        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 648        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 649        [ 8 .. 15]   handle
 650        [16 .. 23]   from
 651        [24 .. 27]   len
 652      */
 653
 654     magic = be32_to_cpup((uint32_t*)buf);
 655     request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 656     request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 657     request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 658     request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 659
 660     TRACE("Got request: { magic = 0x%" PRIx32 ", .type = %" PRIx32
 661           ", from = %" PRIu64 " , len = %" PRIu32 " }",
 662           magic, request->type, request->from, request->len);
 663
 664     if (magic != NBD_REQUEST_MAGIC) {
 665         LOG("invalid magic (got 0x%" PRIx32 ")", magic);
 666         return -EINVAL;
 667     }
 668     return 0;
 669 }
 670
 671 static ssize_t nbd_send_reply(QIOChannel *ioc, struct nbd_reply *reply)
 672 {
 673     uint8_t buf[NBD_REPLY_SIZE];
 674     ssize_t ret;
 675
 676     reply->error = system_errno_to_nbd_errno(reply->error);
 677
 678     TRACE("Sending response to client: { .error = %" PRId32
 679           ", handle = %" PRIu64 " }",
 680           reply->error, reply->handle);
 681
 682     /* Reply
 683        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 684        [ 4 ..  7]    error   (0 == no error)
 685        [ 7 .. 15]    handle
 686      */
 687     stl_be_p(buf, NBD_REPLY_MAGIC);
 688     stl_be_p(buf + 4, reply->error);
 689     stq_be_p(buf + 8, reply->handle);
 690
 691     ret = write_sync(ioc, buf, sizeof(buf));
 692     if (ret < 0) {
 693         return ret;
 694     }
 695
 696     if (ret != sizeof(buf)) {
 697         LOG("writing to socket failed");
 698         return -EINVAL;
 699     }
 700     return 0;
 701 }
 702
 703 #define MAX_NBD_REQUESTS 16
 704
 705 void nbd_client_get(NBDClient *client)
 706 {
 707     client->refcount++;
 708 }
 709
 710 void nbd_client_put(NBDClient *client)
 711 {
 712     if (--client->refcount == 0) {
 713         /* The last reference should be dropped by client->close,
 714          * which is called by client_close.
 715          */
 716         assert(client->closing);
 717
 718         nbd_unset_handlers(client);
 719         object_unref(OBJECT(client->sioc));
 720         object_unref(OBJECT(client->ioc));
 721         if (client->tlscreds) {
 722             object_unref(OBJECT(client->tlscreds));
 723         }
 724         g_free(client->tlsaclname);
 725         if (client->exp) {
 726             QTAILQ_REMOVE(&client->exp->clients, client, next);
 727             nbd_export_put(client->exp);
 728         }
 729         g_free(client);
 730     }
 731 }
 732
 733 static void client_close(NBDClient *client)
 734 {
 735     if (client->closing) {
 736         return;
 737     }
 738
 739     client->closing = true;
 740
 741     /* Force requests to finish.  They will drop their own references,
 742      * then we'll close the socket and free the NBDClient.
 743      */
 744     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 745                          NULL);
 746
 747     /* Also tell the client, so that they release their reference.  */
 748     if (client->close) {
 749         client->close(client);
 750     }
 751 }
 752
 753 static NBDRequest *nbd_request_get(NBDClient *client)
 754 {
 755     NBDRequest *req;
 756
 757     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 758     client->nb_requests++;
 759     nbd_update_can_read(client);
 760
 761     req = g_new0(NBDRequest, 1);
 762     nbd_client_get(client);
 763     req->client = client;
 764     return req;
 765 }
 766
 767 static void nbd_request_put(NBDRequest *req)
 768 {
 769     NBDClient *client = req->client;
 770
 771     if (req->data) {
 772         qemu_vfree(req->data);
 773     }
 774     g_free(req);
 775
 776     client->nb_requests--;
 777     nbd_update_can_read(client);
 778     nbd_client_put(client);
 779 }
 780
 781 static void blk_aio_attached(AioContext *ctx, void *opaque)
 782 {
 783     NBDExport *exp = opaque;
 784     NBDClient *client;
 785
 786     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 787
 788     exp->ctx = ctx;
 789
 790     QTAILQ_FOREACH(client, &exp->clients, next) {
 791         nbd_set_handlers(client);
 792     }
 793 }
 794
 795 static void blk_aio_detach(void *opaque)
 796 {
 797     NBDExport *exp = opaque;
 798     NBDClient *client;
 799
 800     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 801
 802     QTAILQ_FOREACH(client, &exp->clients, next) {
 803         nbd_unset_handlers(client);
 804     }
 805
 806     exp->ctx = NULL;
 807 }
 808
 809 static void nbd_eject_notifier(Notifier *n, void *data)
 810 {
 811     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 812     nbd_export_close(exp);
 813 }
 814
 815 NBDExport *nbd_export_new(BlockBackend *blk, off_t dev_offset, off_t size,
 816                           uint32_t nbdflags, void (*close)(NBDExport *),
 817                           Error **errp)
 818 {
 819     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 820     exp->refcount = 1;
 821     QTAILQ_INIT(&exp->clients);
 822     exp->blk = blk;
 823     exp->dev_offset = dev_offset;
 824     exp->nbdflags = nbdflags;
 825     exp->size = size < 0 ? blk_getlength(blk) : size;
 826     if (exp->size < 0) {
 827         error_setg_errno(errp, -exp->size,
 828                          "Failed to determine the NBD export's length");
 829         goto fail;
 830     }
 831     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 832
 833     exp->close = close;
 834     exp->ctx = blk_get_aio_context(blk);
 835     blk_ref(blk);
 836     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 837
 838     exp->eject_notifier.notify = nbd_eject_notifier;
 839     blk_add_remove_bs_notifier(blk, &exp->eject_notifier);
 840
 841     /*
 842      * NBD exports are used for non-shared storage migration.  Make sure
 843      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 844      * access since the export could be available before migration handover.
 845      */
 846     aio_context_acquire(exp->ctx);
 847     blk_invalidate_cache(blk, NULL);
 848     aio_context_release(exp->ctx);
 849     return exp;
 850
 851 fail:
 852     g_free(exp);
 853     return NULL;
 854 }
 855
 856 NBDExport *nbd_export_find(const char *name)
 857 {
 858     NBDExport *exp;
 859     QTAILQ_FOREACH(exp, &exports, next) {
 860         if (strcmp(name, exp->name) == 0) {
 861             return exp;
 862         }
 863     }
 864
 865     return NULL;
 866 }
 867
 868 void nbd_export_set_name(NBDExport *exp, const char *name)
 869 {
 870     if (exp->name == name) {
 871         return;
 872     }
 873
 874     nbd_export_get(exp);
 875     if (exp->name != NULL) {
 876         g_free(exp->name);
 877         exp->name = NULL;
 878         QTAILQ_REMOVE(&exports, exp, next);
 879         nbd_export_put(exp);
 880     }
 881     if (name != NULL) {
 882         nbd_export_get(exp);
 883         exp->name = g_strdup(name);
 884         QTAILQ_INSERT_TAIL(&exports, exp, next);
 885     }
 886     nbd_export_put(exp);
 887 }
 888
 889 void nbd_export_close(NBDExport *exp)
 890 {
 891     NBDClient *client, *next;
 892
 893     nbd_export_get(exp);
 894     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 895         client_close(client);
 896     }
 897     nbd_export_set_name(exp, NULL);
 898     nbd_export_put(exp);
 899 }
 900
 901 void nbd_export_get(NBDExport *exp)
 902 {
 903     assert(exp->refcount > 0);
 904     exp->refcount++;
 905 }
 906
 907 void nbd_export_put(NBDExport *exp)
 908 {
 909     assert(exp->refcount > 0);
 910     if (exp->refcount == 1) {
 911         nbd_export_close(exp);
 912     }
 913
 914     if (--exp->refcount == 0) {
 915         assert(exp->name == NULL);
 916
 917         if (exp->close) {
 918             exp->close(exp);
 919         }
 920
 921         if (exp->blk) {
 922             notifier_remove(&exp->eject_notifier);
 923             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 924                                             blk_aio_detach, exp);
 925             blk_unref(exp->blk);
 926             exp->blk = NULL;
 927         }
 928
 929         g_free(exp);
 930     }
 931 }
 932
 933 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
 934 {
 935     return exp->blk;
 936 }
 937
 938 void nbd_export_close_all(void)
 939 {
 940     NBDExport *exp, *next;
 941
 942     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 943         nbd_export_close(exp);
 944     }
 945 }
 946
 947 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 948                                  int len)
 949 {
 950     NBDClient *client = req->client;
 951     ssize_t rc, ret;
 952
 953     g_assert(qemu_in_coroutine());
 954     qemu_co_mutex_lock(&client->send_lock);
 955     client->send_coroutine = qemu_coroutine_self();
 956     nbd_set_handlers(client);
 957
 958     if (!len) {
 959         rc = nbd_send_reply(client->ioc, reply);
 960     } else {
 961         qio_channel_set_cork(client->ioc, true);
 962         rc = nbd_send_reply(client->ioc, reply);
 963         if (rc >= 0) {
 964             ret = write_sync(client->ioc, req->data, len);
 965             if (ret != len) {
 966                 rc = -EIO;
 967             }
 968         }
 969         qio_channel_set_cork(client->ioc, false);
 970     }
 971
 972     client->send_coroutine = NULL;
 973     nbd_set_handlers(client);
 974     qemu_co_mutex_unlock(&client->send_lock);
 975     return rc;
 976 }
 977
 978 static ssize_t nbd_co_receive_request(NBDRequest *req, struct nbd_request *request)
 979 {
 980     NBDClient *client = req->client;
 981     uint32_t command;
 982     ssize_t rc;
 983
 984     g_assert(qemu_in_coroutine());
 985     client->recv_coroutine = qemu_coroutine_self();
 986     nbd_update_can_read(client);
 987
 988     rc = nbd_receive_request(client->ioc, request);
 989     if (rc < 0) {
 990         if (rc != -EAGAIN) {
 991             rc = -EIO;
 992         }
 993         goto out;
 994     }
 995
 996     if ((request->from + request->len) < request->from) {
 997         LOG("integer overflow detected! "
 998             "you're probably being attacked");
 999         rc = -EINVAL;
1000         goto out;
1001     }
1002
1003     TRACE("Decoding type");
1004
1005     command = request->type & NBD_CMD_MASK_COMMAND;
1006     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
1007         if (request->len > NBD_MAX_BUFFER_SIZE) {
1008             LOG("len (%" PRIu32" ) is larger than max len (%u)",
1009                 request->len, NBD_MAX_BUFFER_SIZE);
1010             rc = -EINVAL;
1011             goto out;
1012         }
1013
1014         req->data = blk_try_blockalign(client->exp->blk, request->len);
1015         if (req->data == NULL) {
1016             rc = -ENOMEM;
1017             goto out;
1018         }
1019     }
1020     if (command == NBD_CMD_WRITE) {
1021         TRACE("Reading %" PRIu32 " byte(s)", request->len);
1022
1023         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1024             LOG("reading from socket failed");
1025             rc = -EIO;
1026             goto out;
1027         }
1028     }
1029     rc = 0;
1030
1031 out:
1032     client->recv_coroutine = NULL;
1033     nbd_update_can_read(client);
1034
1035     return rc;
1036 }
1037
1038 static void nbd_trip(void *opaque)
1039 {
1040     NBDClient *client = opaque;
1041     NBDExport *exp = client->exp;
1042     NBDRequest *req;
1043     struct nbd_request request;
1044     struct nbd_reply reply;
1045     ssize_t ret;
1046     uint32_t command;
1047
1048     TRACE("Reading request.");
1049     if (client->closing) {
1050         return;
1051     }
1052
1053     req = nbd_request_get(client);
1054     ret = nbd_co_receive_request(req, &request);
1055     if (ret == -EAGAIN) {
1056         goto done;
1057     }
1058     if (ret == -EIO) {
1059         goto out;
1060     }
1061
1062     reply.handle = request.handle;
1063     reply.error = 0;
1064
1065     if (ret < 0) {
1066         reply.error = -ret;
1067         goto error_reply;
1068     }
1069     command = request.type & NBD_CMD_MASK_COMMAND;
1070     if (command != NBD_CMD_DISC && (request.from + request.len) > exp->size) {
1071             LOG("From: %" PRIu64 ", Len: %" PRIu32", Size: %" PRIu64
1072                 ", Offset: %" PRIu64 "\n",
1073                 request.from, request.len,
1074                 (uint64_t)exp->size, (uint64_t)exp->dev_offset);
1075         LOG("requested operation past EOF--bad client?");
1076         goto invalid_request;
1077     }
1078
1079     if (client->closing) {
1080         /*
1081          * The client may be closed when we are blocked in
1082          * nbd_co_receive_request()
1083          */
1084         goto done;
1085     }
1086
1087     switch (command) {
1088     case NBD_CMD_READ:
1089         TRACE("Request type is READ");
1090
1091         if (request.type & NBD_CMD_FLAG_FUA) {
1092             ret = blk_co_flush(exp->blk);
1093             if (ret < 0) {
1094                 LOG("flush failed");
1095                 reply.error = -ret;
1096                 goto error_reply;
1097             }
1098         }
1099
1100         ret = blk_pread(exp->blk, request.from + exp->dev_offset,
1101                         req->data, request.len);
1102         if (ret < 0) {
1103             LOG("reading from file failed");
1104             reply.error = -ret;
1105             goto error_reply;
1106         }
1107
1108         TRACE("Read %" PRIu32" byte(s)", request.len);
1109         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1110             goto out;
1111         break;
1112     case NBD_CMD_WRITE:
1113         TRACE("Request type is WRITE");
1114
1115         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1116             TRACE("Server is read-only, return error");
1117             reply.error = EROFS;
1118             goto error_reply;
1119         }
1120
1121         TRACE("Writing to device");
1122
1123         ret = blk_pwrite(exp->blk, request.from + exp->dev_offset,
1124                         req->data, request.len);
1125         if (ret < 0) {
1126             LOG("writing to file failed");
1127             reply.error = -ret;
1128             goto error_reply;
1129         }
1130
1131         if (request.type & NBD_CMD_FLAG_FUA) {
1132             ret = blk_co_flush(exp->blk);
1133             if (ret < 0) {
1134                 LOG("flush failed");
1135                 reply.error = -ret;
1136                 goto error_reply;
1137             }
1138         }
1139
1140         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1141             goto out;
1142         }
1143         break;
1144     case NBD_CMD_DISC:
1145         TRACE("Request type is DISCONNECT");
1146         errno = 0;
1147         goto out;
1148     case NBD_CMD_FLUSH:
1149         TRACE("Request type is FLUSH");
1150
1151         ret = blk_co_flush(exp->blk);
1152         if (ret < 0) {
1153             LOG("flush failed");
1154             reply.error = -ret;
1155         }
1156         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1157             goto out;
1158         }
1159         break;
1160     case NBD_CMD_TRIM:
1161         TRACE("Request type is TRIM");
1162         /* Ignore unaligned head or tail, until block layer adds byte
1163          * interface */
1164         if (request.len >= BDRV_SECTOR_SIZE) {
1165             request.len -= (request.from + request.len) % BDRV_SECTOR_SIZE;
1166             ret = blk_co_discard(exp->blk,
1167                                  DIV_ROUND_UP(request.from + exp->dev_offset,
1168                                               BDRV_SECTOR_SIZE),
1169                                  request.len / BDRV_SECTOR_SIZE);
1170             if (ret < 0) {
1171                 LOG("discard failed");
1172                 reply.error = -ret;
1173             }
1174         } else {
1175             TRACE("trim request too small, ignoring");
1176         }
1177         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1178             goto out;
1179         }
1180         break;
1181     default:
1182         LOG("invalid request type (%" PRIu32 ") received", request.type);
1183     invalid_request:
1184         reply.error = EINVAL;
1185     error_reply:
1186         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1187             goto out;
1188         }
1189         break;
1190     }
1191
1192     TRACE("Request/Reply complete");
1193
1194 done:
1195     nbd_request_put(req);
1196     return;
1197
1198 out:
1199     nbd_request_put(req);
1200     client_close(client);
1201 }
1202
1203 static void nbd_read(void *opaque)
1204 {
1205     NBDClient *client = opaque;
1206
1207     if (client->recv_coroutine) {
1208         qemu_coroutine_enter(client->recv_coroutine, NULL);
1209     } else {
1210         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip), client);
1211     }
1212 }
1213
1214 static void nbd_restart_write(void *opaque)
1215 {
1216     NBDClient *client = opaque;
1217
1218     qemu_coroutine_enter(client->send_coroutine, NULL);
1219 }
1220
1221 static void nbd_set_handlers(NBDClient *client)
1222 {
1223     if (client->exp && client->exp->ctx) {
1224         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1225                            true,
1226                            client->can_read ? nbd_read : NULL,
1227                            client->send_coroutine ? nbd_restart_write : NULL,
1228                            client);
1229     }
1230 }
1231
1232 static void nbd_unset_handlers(NBDClient *client)
1233 {
1234     if (client->exp && client->exp->ctx) {
1235         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1236                            true, NULL, NULL, NULL);
1237     }
1238 }
1239
1240 static void nbd_update_can_read(NBDClient *client)
1241 {
1242     bool can_read = client->recv_coroutine ||
1243                     client->nb_requests < MAX_NBD_REQUESTS;
1244
1245     if (can_read != client->can_read) {
1246         client->can_read = can_read;
1247         nbd_set_handlers(client);
1248
1249         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1250          * in nbd_set_handlers() will have taken care of that */
1251     }
1252 }
1253
1254 static coroutine_fn void nbd_co_client_start(void *opaque)
1255 {
1256     NBDClientNewData *data = opaque;
1257     NBDClient *client = data->client;
1258     NBDExport *exp = client->exp;
1259
1260     if (exp) {
1261         nbd_export_get(exp);
1262     }
1263     if (nbd_negotiate(data)) {
1264         client_close(client);
1265         goto out;
1266     }
1267     qemu_co_mutex_init(&client->send_lock);
1268     nbd_set_handlers(client);
1269
1270     if (exp) {
1271         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1272     }
1273 out:
1274     g_free(data);
1275 }
1276
1277 void nbd_client_new(NBDExport *exp,
1278                     QIOChannelSocket *sioc,
1279                     QCryptoTLSCreds *tlscreds,
1280                     const char *tlsaclname,
1281                     void (*close_fn)(NBDClient *))
1282 {
1283     NBDClient *client;
1284     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1285
1286     client = g_malloc0(sizeof(NBDClient));
1287     client->refcount = 1;
1288     client->exp = exp;
1289     client->tlscreds = tlscreds;
1290     if (tlscreds) {
1291         object_ref(OBJECT(client->tlscreds));
1292     }
1293     client->tlsaclname = g_strdup(tlsaclname);
1294     client->sioc = sioc;
1295     object_ref(OBJECT(client->sioc));
1296     client->ioc = QIO_CHANNEL(sioc);
1297     object_ref(OBJECT(client->ioc));
1298     client->can_read = true;
1299     client->close = close_fn;
1300
1301     data->client = client;
1302     data->co = qemu_coroutine_create(nbd_co_client_start);
1303     qemu_coroutine_enter(data->co, data);
1304 }