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         return NBD_EPERM;
  30     case EIO:
  31         return NBD_EIO;
  32     case ENOMEM:
  33         return NBD_ENOMEM;
  34 #ifdef EDQUOT
  35     case EDQUOT:
  36 #endif
  37     case EFBIG:
  38     case ENOSPC:
  39         return NBD_ENOSPC;
  40     case EINVAL:
  41     default:
  42         return NBD_EINVAL;
  43     }
  44 }
  45
  46 /* Definitions for opaque data types */
  47
  48 typedef struct NBDRequest NBDRequest;
  49
  50 struct NBDRequest {
  51     QSIMPLEQ_ENTRY(NBDRequest) entry;
  52     NBDClient *client;
  53     uint8_t *data;
  54 };
  55
  56 struct NBDExport {
  57     int refcount;
  58     void (*close)(NBDExport *exp);
  59
  60     BlockBackend *blk;
  61     char *name;
  62     off_t dev_offset;
  63     off_t size;
  64     uint32_t nbdflags;
  65     QTAILQ_HEAD(, NBDClient) clients;
  66     QTAILQ_ENTRY(NBDExport) next;
  67
  68     AioContext *ctx;
  69
  70     Notifier eject_notifier;
  71 };
  72
  73 static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports);
  74
  75 struct NBDClient {
  76     int refcount;
  77     void (*close)(NBDClient *client);
  78
  79     NBDExport *exp;
  80     QCryptoTLSCreds *tlscreds;
  81     char *tlsaclname;
  82     QIOChannelSocket *sioc; /* The underlying data channel */
  83     QIOChannel *ioc; /* The current I/O channel which may differ (eg TLS) */
  84
  85     Coroutine *recv_coroutine;
  86
  87     CoMutex send_lock;
  88     Coroutine *send_coroutine;
  89
  90     bool can_read;
  91
  92     QTAILQ_ENTRY(NBDClient) next;
  93     int nb_requests;
  94     bool closing;
  95 };
  96
  97 /* That's all folks */
  98
  99 static void nbd_set_handlers(NBDClient *client);
 100 static void nbd_unset_handlers(NBDClient *client);
 101 static void nbd_update_can_read(NBDClient *client);
 102
 103 static gboolean nbd_negotiate_continue(QIOChannel *ioc,
 104                                        GIOCondition condition,
 105                                        void *opaque)
 106 {
 107     qemu_coroutine_enter(opaque, NULL);
 108     return TRUE;
 109 }
 110
 111 static ssize_t nbd_negotiate_read(QIOChannel *ioc, void *buffer, size_t size)
 112 {
 113     ssize_t ret;
 114     guint watch;
 115
 116     assert(qemu_in_coroutine());
 117     /* Negotiation are always in main loop. */
 118     watch = qio_channel_add_watch(ioc,
 119                                   G_IO_IN,
 120                                   nbd_negotiate_continue,
 121                                   qemu_coroutine_self(),
 122                                   NULL);
 123     ret = read_sync(ioc, buffer, size);
 124     g_source_remove(watch);
 125     return ret;
 126
 127 }
 128
 129 static ssize_t nbd_negotiate_write(QIOChannel *ioc, void *buffer, size_t size)
 130 {
 131     ssize_t ret;
 132     guint watch;
 133
 134     assert(qemu_in_coroutine());
 135     /* Negotiation are always in main loop. */
 136     watch = qio_channel_add_watch(ioc,
 137                                   G_IO_OUT,
 138                                   nbd_negotiate_continue,
 139                                   qemu_coroutine_self(),
 140                                   NULL);
 141     ret = write_sync(ioc, buffer, size);
 142     g_source_remove(watch);
 143     return ret;
 144 }
 145
 146 static ssize_t nbd_negotiate_drop_sync(QIOChannel *ioc, size_t size)
 147 {
 148     ssize_t ret, dropped = size;
 149     uint8_t *buffer = g_malloc(MIN(65536, size));
 150
 151     while (size > 0) {
 152         ret = nbd_negotiate_read(ioc, buffer, MIN(65536, size));
 153         if (ret < 0) {
 154             g_free(buffer);
 155             return ret;
 156         }
 157
 158         assert(ret <= size);
 159         size -= ret;
 160     }
 161
 162     g_free(buffer);
 163     return dropped;
 164 }
 165
 166 /* Basic flow for negotiation
 167
 168    Server         Client
 169    Negotiate
 170
 171    or
 172
 173    Server         Client
 174    Negotiate #1
 175                   Option
 176    Negotiate #2
 177
 178    ----
 179
 180    followed by
 181
 182    Server         Client
 183                   Request
 184    Response
 185                   Request
 186    Response
 187                   ...
 188    ...
 189                   Request (type == 2)
 190
 191 */
 192
 193 static int nbd_negotiate_send_rep(QIOChannel *ioc, uint32_t type, uint32_t opt)
 194 {
 195     uint64_t magic;
 196     uint32_t len;
 197
 198     TRACE("Reply opt=%x type=%x", type, opt);
 199
 200     magic = cpu_to_be64(NBD_REP_MAGIC);
 201     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 202         LOG("write failed (rep magic)");
 203         return -EINVAL;
 204     }
 205     opt = cpu_to_be32(opt);
 206     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 207         LOG("write failed (rep opt)");
 208         return -EINVAL;
 209     }
 210     type = cpu_to_be32(type);
 211     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 212         LOG("write failed (rep type)");
 213         return -EINVAL;
 214     }
 215     len = cpu_to_be32(0);
 216     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 217         LOG("write failed (rep data length)");
 218         return -EINVAL;
 219     }
 220     return 0;
 221 }
 222
 223 static int nbd_negotiate_send_rep_list(QIOChannel *ioc, NBDExport *exp)
 224 {
 225     uint64_t magic, name_len;
 226     uint32_t opt, type, len;
 227
 228     TRACE("Advertizing export name '%s'", exp->name ? exp->name : "");
 229     name_len = strlen(exp->name);
 230     magic = cpu_to_be64(NBD_REP_MAGIC);
 231     if (nbd_negotiate_write(ioc, &magic, sizeof(magic)) != sizeof(magic)) {
 232         LOG("write failed (magic)");
 233         return -EINVAL;
 234      }
 235     opt = cpu_to_be32(NBD_OPT_LIST);
 236     if (nbd_negotiate_write(ioc, &opt, sizeof(opt)) != sizeof(opt)) {
 237         LOG("write failed (opt)");
 238         return -EINVAL;
 239     }
 240     type = cpu_to_be32(NBD_REP_SERVER);
 241     if (nbd_negotiate_write(ioc, &type, sizeof(type)) != sizeof(type)) {
 242         LOG("write failed (reply type)");
 243         return -EINVAL;
 244     }
 245     len = cpu_to_be32(name_len + sizeof(len));
 246     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 247         LOG("write failed (length)");
 248         return -EINVAL;
 249     }
 250     len = cpu_to_be32(name_len);
 251     if (nbd_negotiate_write(ioc, &len, sizeof(len)) != sizeof(len)) {
 252         LOG("write failed (length)");
 253         return -EINVAL;
 254     }
 255     if (nbd_negotiate_write(ioc, exp->name, name_len) != name_len) {
 256         LOG("write failed (buffer)");
 257         return -EINVAL;
 258     }
 259     return 0;
 260 }
 261
 262 static int nbd_negotiate_handle_list(NBDClient *client, uint32_t length)
 263 {
 264     NBDExport *exp;
 265
 266     if (length) {
 267         if (nbd_negotiate_drop_sync(client->ioc, length) != length) {
 268             return -EIO;
 269         }
 270         return nbd_negotiate_send_rep(client->ioc,
 271                                       NBD_REP_ERR_INVALID, NBD_OPT_LIST);
 272     }
 273
 274     /* For each export, send a NBD_REP_SERVER reply. */
 275     QTAILQ_FOREACH(exp, &exports, next) {
 276         if (nbd_negotiate_send_rep_list(client->ioc, exp)) {
 277             return -EINVAL;
 278         }
 279     }
 280     /* Finish with a NBD_REP_ACK. */
 281     return nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_LIST);
 282 }
 283
 284 static int nbd_negotiate_handle_export_name(NBDClient *client, uint32_t length)
 285 {
 286     int rc = -EINVAL;
 287     char name[256];
 288
 289     /* Client sends:
 290         [20 ..  xx]   export name (length bytes)
 291      */
 292     TRACE("Checking length");
 293     if (length > 255) {
 294         LOG("Bad length received");
 295         goto fail;
 296     }
 297     if (nbd_negotiate_read(client->ioc, name, length) != length) {
 298         LOG("read failed");
 299         goto fail;
 300     }
 301     name[length] = '\0';
 302
 303     TRACE("Client requested export '%s'", name);
 304
 305     client->exp = nbd_export_find(name);
 306     if (!client->exp) {
 307         LOG("export not found");
 308         goto fail;
 309     }
 310
 311     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 312     nbd_export_get(client->exp);
 313     rc = 0;
 314 fail:
 315     return rc;
 316 }
 317
 318
 319 static QIOChannel *nbd_negotiate_handle_starttls(NBDClient *client,
 320                                                  uint32_t length)
 321 {
 322     QIOChannel *ioc;
 323     QIOChannelTLS *tioc;
 324     struct NBDTLSHandshakeData data = { 0 };
 325
 326     TRACE("Setting up TLS");
 327     ioc = client->ioc;
 328     if (length) {
 329         if (nbd_negotiate_drop_sync(ioc, length) != length) {
 330             return NULL;
 331         }
 332         nbd_negotiate_send_rep(ioc, NBD_REP_ERR_INVALID, NBD_OPT_STARTTLS);
 333         return NULL;
 334     }
 335
 336     nbd_negotiate_send_rep(client->ioc, NBD_REP_ACK, NBD_OPT_STARTTLS);
 337
 338     tioc = qio_channel_tls_new_server(ioc,
 339                                       client->tlscreds,
 340                                       client->tlsaclname,
 341                                       NULL);
 342     if (!tioc) {
 343         return NULL;
 344     }
 345
 346     TRACE("Starting TLS handshake");
 347     data.loop = g_main_loop_new(g_main_context_default(), FALSE);
 348     qio_channel_tls_handshake(tioc,
 349                               nbd_tls_handshake,
 350                               &data,
 351                               NULL);
 352
 353     if (!data.complete) {
 354         g_main_loop_run(data.loop);
 355     }
 356     g_main_loop_unref(data.loop);
 357     if (data.error) {
 358         object_unref(OBJECT(tioc));
 359         error_free(data.error);
 360         return NULL;
 361     }
 362
 363     return QIO_CHANNEL(tioc);
 364 }
 365
 366
 367 static int nbd_negotiate_options(NBDClient *client)
 368 {
 369     uint32_t flags;
 370     bool fixedNewstyle = false;
 371
 372     /* Client sends:
 373         [ 0 ..   3]   client flags
 374
 375         [ 0 ..   7]   NBD_OPTS_MAGIC
 376         [ 8 ..  11]   NBD option
 377         [12 ..  15]   Data length
 378         ...           Rest of request
 379
 380         [ 0 ..   7]   NBD_OPTS_MAGIC
 381         [ 8 ..  11]   Second NBD option
 382         [12 ..  15]   Data length
 383         ...           Rest of request
 384     */
 385
 386     if (nbd_negotiate_read(client->ioc, &flags, sizeof(flags)) !=
 387         sizeof(flags)) {
 388         LOG("read failed");
 389         return -EIO;
 390     }
 391     TRACE("Checking client flags");
 392     be32_to_cpus(&flags);
 393     if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) {
 394         TRACE("Support supports fixed newstyle handshake");
 395         fixedNewstyle = true;
 396         flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE;
 397     }
 398     if (flags != 0) {
 399         TRACE("Unknown client flags 0x%x received", flags);
 400         return -EIO;
 401     }
 402
 403     while (1) {
 404         int ret;
 405         uint32_t clientflags, length;
 406         uint64_t magic;
 407
 408         if (nbd_negotiate_read(client->ioc, &magic, sizeof(magic)) !=
 409             sizeof(magic)) {
 410             LOG("read failed");
 411             return -EINVAL;
 412         }
 413         TRACE("Checking opts magic");
 414         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 415             LOG("Bad magic received");
 416             return -EINVAL;
 417         }
 418
 419         if (nbd_negotiate_read(client->ioc, &clientflags,
 420                                sizeof(clientflags)) != sizeof(clientflags)) {
 421             LOG("read failed");
 422             return -EINVAL;
 423         }
 424         clientflags = be32_to_cpu(clientflags);
 425
 426         if (nbd_negotiate_read(client->ioc, &length, sizeof(length)) !=
 427             sizeof(length)) {
 428             LOG("read failed");
 429             return -EINVAL;
 430         }
 431         length = be32_to_cpu(length);
 432
 433         TRACE("Checking option 0x%x", clientflags);
 434         if (client->tlscreds &&
 435             client->ioc == (QIOChannel *)client->sioc) {
 436             QIOChannel *tioc;
 437             if (!fixedNewstyle) {
 438                 TRACE("Unsupported option 0x%x", clientflags);
 439                 return -EINVAL;
 440             }
 441             switch (clientflags) {
 442             case NBD_OPT_STARTTLS:
 443                 tioc = nbd_negotiate_handle_starttls(client, length);
 444                 if (!tioc) {
 445                     return -EIO;
 446                 }
 447                 object_unref(OBJECT(client->ioc));
 448                 client->ioc = QIO_CHANNEL(tioc);
 449                 break;
 450
 451             default:
 452                 TRACE("Option 0x%x not permitted before TLS", clientflags);
 453                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_TLS_REQD,
 454                                        clientflags);
 455                 return -EINVAL;
 456             }
 457         } else if (fixedNewstyle) {
 458             switch (clientflags) {
 459             case NBD_OPT_LIST:
 460                 ret = nbd_negotiate_handle_list(client, length);
 461                 if (ret < 0) {
 462                     return ret;
 463                 }
 464                 break;
 465
 466             case NBD_OPT_ABORT:
 467                 return -EINVAL;
 468
 469             case NBD_OPT_EXPORT_NAME:
 470                 return nbd_negotiate_handle_export_name(client, length);
 471
 472             case NBD_OPT_STARTTLS:
 473                 if (client->tlscreds) {
 474                     TRACE("TLS already enabled");
 475                     nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_INVALID,
 476                                            clientflags);
 477                 } else {
 478                     TRACE("TLS not configured");
 479                     nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_POLICY,
 480                                            clientflags);
 481                 }
 482                 return -EINVAL;
 483             default:
 484                 TRACE("Unsupported option 0x%x", clientflags);
 485                 nbd_negotiate_send_rep(client->ioc, NBD_REP_ERR_UNSUP,
 486                                        clientflags);
 487                 return -EINVAL;
 488             }
 489         } else {
 490             /*
 491              * If broken new-style we should drop the connection
 492              * for anything except NBD_OPT_EXPORT_NAME
 493              */
 494             switch (clientflags) {
 495             case NBD_OPT_EXPORT_NAME:
 496                 return nbd_negotiate_handle_export_name(client, length);
 497
 498             default:
 499                 TRACE("Unsupported option 0x%x", clientflags);
 500                 return -EINVAL;
 501             }
 502         }
 503     }
 504 }
 505
 506 typedef struct {
 507     NBDClient *client;
 508     Coroutine *co;
 509 } NBDClientNewData;
 510
 511 static coroutine_fn int nbd_negotiate(NBDClientNewData *data)
 512 {
 513     NBDClient *client = data->client;
 514     char buf[8 + 8 + 8 + 128];
 515     int rc;
 516     const int myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 517                          NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 518     bool oldStyle;
 519
 520     /* Old style negotiation header without options
 521         [ 0 ..   7]   passwd       ("NBDMAGIC")
 522         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 523         [16 ..  23]   size
 524         [24 ..  25]   server flags (0)
 525         [26 ..  27]   export flags
 526         [28 .. 151]   reserved     (0)
 527
 528        New style negotiation header with options
 529         [ 0 ..   7]   passwd       ("NBDMAGIC")
 530         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 531         [16 ..  17]   server flags (0)
 532         ....options sent....
 533         [18 ..  25]   size
 534         [26 ..  27]   export flags
 535         [28 .. 151]   reserved     (0)
 536      */
 537
 538     qio_channel_set_blocking(client->ioc, false, NULL);
 539     rc = -EINVAL;
 540
 541     TRACE("Beginning negotiation.");
 542     memset(buf, 0, sizeof(buf));
 543     memcpy(buf, "NBDMAGIC", 8);
 544
 545     oldStyle = client->exp != NULL && !client->tlscreds;
 546     if (oldStyle) {
 547         assert ((client->exp->nbdflags & ~65535) == 0);
 548         stq_be_p(buf + 8, NBD_CLIENT_MAGIC);
 549         stq_be_p(buf + 16, client->exp->size);
 550         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 551     } else {
 552         stq_be_p(buf + 8, NBD_OPTS_MAGIC);
 553         stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE);
 554     }
 555
 556     if (oldStyle) {
 557         if (client->tlscreds) {
 558             TRACE("TLS cannot be enabled with oldstyle protocol");
 559             goto fail;
 560         }
 561         if (nbd_negotiate_write(client->ioc, buf, sizeof(buf)) != sizeof(buf)) {
 562             LOG("write failed");
 563             goto fail;
 564         }
 565     } else {
 566         if (nbd_negotiate_write(client->ioc, buf, 18) != 18) {
 567             LOG("write failed");
 568             goto fail;
 569         }
 570         rc = nbd_negotiate_options(client);
 571         if (rc != 0) {
 572             LOG("option negotiation failed");
 573             goto fail;
 574         }
 575
 576         assert ((client->exp->nbdflags & ~65535) == 0);
 577         stq_be_p(buf + 18, client->exp->size);
 578         stw_be_p(buf + 26, client->exp->nbdflags | myflags);
 579         if (nbd_negotiate_write(client->ioc, buf + 18, sizeof(buf) - 18) !=
 580             sizeof(buf) - 18) {
 581             LOG("write failed");
 582             goto fail;
 583         }
 584     }
 585
 586     TRACE("Negotiation succeeded.");
 587     rc = 0;
 588 fail:
 589     return rc;
 590 }
 591
 592 #ifdef __linux__
 593
 594 int nbd_disconnect(int fd)
 595 {
 596     ioctl(fd, NBD_CLEAR_QUE);
 597     ioctl(fd, NBD_DISCONNECT);
 598     ioctl(fd, NBD_CLEAR_SOCK);
 599     return 0;
 600 }
 601
 602 #else
 603
 604 int nbd_disconnect(int fd)
 605 {
 606     return -ENOTSUP;
 607 }
 608 #endif
 609
 610 static ssize_t nbd_receive_request(QIOChannel *ioc, struct nbd_request *request)
 611 {
 612     uint8_t buf[NBD_REQUEST_SIZE];
 613     uint32_t magic;
 614     ssize_t ret;
 615
 616     ret = read_sync(ioc, buf, sizeof(buf));
 617     if (ret < 0) {
 618         return ret;
 619     }
 620
 621     if (ret != sizeof(buf)) {
 622         LOG("read failed");
 623         return -EINVAL;
 624     }
 625
 626     /* Request
 627        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 628        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 629        [ 8 .. 15]   handle
 630        [16 .. 23]   from
 631        [24 .. 27]   len
 632      */
 633
 634     magic = be32_to_cpup((uint32_t*)buf);
 635     request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 636     request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 637     request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 638     request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 639
 640     TRACE("Got request: "
 641           "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 642           magic, request->type, request->from, request->len);
 643
 644     if (magic != NBD_REQUEST_MAGIC) {
 645         LOG("invalid magic (got 0x%x)", magic);
 646         return -EINVAL;
 647     }
 648     return 0;
 649 }
 650
 651 static ssize_t nbd_send_reply(QIOChannel *ioc, struct nbd_reply *reply)
 652 {
 653     uint8_t buf[NBD_REPLY_SIZE];
 654     ssize_t ret;
 655
 656     reply->error = system_errno_to_nbd_errno(reply->error);
 657
 658     /* Reply
 659        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 660        [ 4 ..  7]    error   (0 == no error)
 661        [ 7 .. 15]    handle
 662      */
 663     stl_be_p(buf, NBD_REPLY_MAGIC);
 664     stl_be_p(buf + 4, reply->error);
 665     stq_be_p(buf + 8, reply->handle);
 666
 667     TRACE("Sending response to client");
 668
 669     ret = write_sync(ioc, buf, sizeof(buf));
 670     if (ret < 0) {
 671         return ret;
 672     }
 673
 674     if (ret != sizeof(buf)) {
 675         LOG("writing to socket failed");
 676         return -EINVAL;
 677     }
 678     return 0;
 679 }
 680
 681 #define MAX_NBD_REQUESTS 16
 682
 683 void nbd_client_get(NBDClient *client)
 684 {
 685     client->refcount++;
 686 }
 687
 688 void nbd_client_put(NBDClient *client)
 689 {
 690     if (--client->refcount == 0) {
 691         /* The last reference should be dropped by client->close,
 692          * which is called by client_close.
 693          */
 694         assert(client->closing);
 695
 696         nbd_unset_handlers(client);
 697         object_unref(OBJECT(client->sioc));
 698         object_unref(OBJECT(client->ioc));
 699         if (client->tlscreds) {
 700             object_unref(OBJECT(client->tlscreds));
 701         }
 702         g_free(client->tlsaclname);
 703         if (client->exp) {
 704             QTAILQ_REMOVE(&client->exp->clients, client, next);
 705             nbd_export_put(client->exp);
 706         }
 707         g_free(client);
 708     }
 709 }
 710
 711 static void client_close(NBDClient *client)
 712 {
 713     if (client->closing) {
 714         return;
 715     }
 716
 717     client->closing = true;
 718
 719     /* Force requests to finish.  They will drop their own references,
 720      * then we'll close the socket and free the NBDClient.
 721      */
 722     qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH,
 723                          NULL);
 724
 725     /* Also tell the client, so that they release their reference.  */
 726     if (client->close) {
 727         client->close(client);
 728     }
 729 }
 730
 731 static NBDRequest *nbd_request_get(NBDClient *client)
 732 {
 733     NBDRequest *req;
 734
 735     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 736     client->nb_requests++;
 737     nbd_update_can_read(client);
 738
 739     req = g_new0(NBDRequest, 1);
 740     nbd_client_get(client);
 741     req->client = client;
 742     return req;
 743 }
 744
 745 static void nbd_request_put(NBDRequest *req)
 746 {
 747     NBDClient *client = req->client;
 748
 749     if (req->data) {
 750         qemu_vfree(req->data);
 751     }
 752     g_free(req);
 753
 754     client->nb_requests--;
 755     nbd_update_can_read(client);
 756     nbd_client_put(client);
 757 }
 758
 759 static void blk_aio_attached(AioContext *ctx, void *opaque)
 760 {
 761     NBDExport *exp = opaque;
 762     NBDClient *client;
 763
 764     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 765
 766     exp->ctx = ctx;
 767
 768     QTAILQ_FOREACH(client, &exp->clients, next) {
 769         nbd_set_handlers(client);
 770     }
 771 }
 772
 773 static void blk_aio_detach(void *opaque)
 774 {
 775     NBDExport *exp = opaque;
 776     NBDClient *client;
 777
 778     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 779
 780     QTAILQ_FOREACH(client, &exp->clients, next) {
 781         nbd_unset_handlers(client);
 782     }
 783
 784     exp->ctx = NULL;
 785 }
 786
 787 static void nbd_eject_notifier(Notifier *n, void *data)
 788 {
 789     NBDExport *exp = container_of(n, NBDExport, eject_notifier);
 790     nbd_export_close(exp);
 791 }
 792
 793 NBDExport *nbd_export_new(BlockBackend *blk, off_t dev_offset, off_t size,
 794                           uint32_t nbdflags, void (*close)(NBDExport *),
 795                           Error **errp)
 796 {
 797     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 798     exp->refcount = 1;
 799     QTAILQ_INIT(&exp->clients);
 800     exp->blk = blk;
 801     exp->dev_offset = dev_offset;
 802     exp->nbdflags = nbdflags;
 803     exp->size = size < 0 ? blk_getlength(blk) : size;
 804     if (exp->size < 0) {
 805         error_setg_errno(errp, -exp->size,
 806                          "Failed to determine the NBD export's length");
 807         goto fail;
 808     }
 809     exp->size -= exp->size % BDRV_SECTOR_SIZE;
 810
 811     exp->close = close;
 812     exp->ctx = blk_get_aio_context(blk);
 813     blk_ref(blk);
 814     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 815
 816     exp->eject_notifier.notify = nbd_eject_notifier;
 817     blk_add_remove_bs_notifier(blk, &exp->eject_notifier);
 818
 819     /*
 820      * NBD exports are used for non-shared storage migration.  Make sure
 821      * that BDRV_O_INACTIVE is cleared and the image is ready for write
 822      * access since the export could be available before migration handover.
 823      */
 824     aio_context_acquire(exp->ctx);
 825     blk_invalidate_cache(blk, NULL);
 826     aio_context_release(exp->ctx);
 827     return exp;
 828
 829 fail:
 830     g_free(exp);
 831     return NULL;
 832 }
 833
 834 NBDExport *nbd_export_find(const char *name)
 835 {
 836     NBDExport *exp;
 837     QTAILQ_FOREACH(exp, &exports, next) {
 838         if (strcmp(name, exp->name) == 0) {
 839             return exp;
 840         }
 841     }
 842
 843     return NULL;
 844 }
 845
 846 void nbd_export_set_name(NBDExport *exp, const char *name)
 847 {
 848     if (exp->name == name) {
 849         return;
 850     }
 851
 852     nbd_export_get(exp);
 853     if (exp->name != NULL) {
 854         g_free(exp->name);
 855         exp->name = NULL;
 856         QTAILQ_REMOVE(&exports, exp, next);
 857         nbd_export_put(exp);
 858     }
 859     if (name != NULL) {
 860         nbd_export_get(exp);
 861         exp->name = g_strdup(name);
 862         QTAILQ_INSERT_TAIL(&exports, exp, next);
 863     }
 864     nbd_export_put(exp);
 865 }
 866
 867 void nbd_export_close(NBDExport *exp)
 868 {
 869     NBDClient *client, *next;
 870
 871     nbd_export_get(exp);
 872     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
 873         client_close(client);
 874     }
 875     nbd_export_set_name(exp, NULL);
 876     nbd_export_put(exp);
 877 }
 878
 879 void nbd_export_get(NBDExport *exp)
 880 {
 881     assert(exp->refcount > 0);
 882     exp->refcount++;
 883 }
 884
 885 void nbd_export_put(NBDExport *exp)
 886 {
 887     assert(exp->refcount > 0);
 888     if (exp->refcount == 1) {
 889         nbd_export_close(exp);
 890     }
 891
 892     if (--exp->refcount == 0) {
 893         assert(exp->name == NULL);
 894
 895         if (exp->close) {
 896             exp->close(exp);
 897         }
 898
 899         if (exp->blk) {
 900             notifier_remove(&exp->eject_notifier);
 901             blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
 902                                             blk_aio_detach, exp);
 903             blk_unref(exp->blk);
 904             exp->blk = NULL;
 905         }
 906
 907         g_free(exp);
 908     }
 909 }
 910
 911 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
 912 {
 913     return exp->blk;
 914 }
 915
 916 void nbd_export_close_all(void)
 917 {
 918     NBDExport *exp, *next;
 919
 920     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
 921         nbd_export_close(exp);
 922     }
 923 }
 924
 925 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
 926                                  int len)
 927 {
 928     NBDClient *client = req->client;
 929     ssize_t rc, ret;
 930
 931     g_assert(qemu_in_coroutine());
 932     qemu_co_mutex_lock(&client->send_lock);
 933     client->send_coroutine = qemu_coroutine_self();
 934     nbd_set_handlers(client);
 935
 936     if (!len) {
 937         rc = nbd_send_reply(client->ioc, reply);
 938     } else {
 939         qio_channel_set_cork(client->ioc, true);
 940         rc = nbd_send_reply(client->ioc, reply);
 941         if (rc >= 0) {
 942             ret = write_sync(client->ioc, req->data, len);
 943             if (ret != len) {
 944                 rc = -EIO;
 945             }
 946         }
 947         qio_channel_set_cork(client->ioc, false);
 948     }
 949
 950     client->send_coroutine = NULL;
 951     nbd_set_handlers(client);
 952     qemu_co_mutex_unlock(&client->send_lock);
 953     return rc;
 954 }
 955
 956 static ssize_t nbd_co_receive_request(NBDRequest *req, struct nbd_request *request)
 957 {
 958     NBDClient *client = req->client;
 959     uint32_t command;
 960     ssize_t rc;
 961
 962     g_assert(qemu_in_coroutine());
 963     client->recv_coroutine = qemu_coroutine_self();
 964     nbd_update_can_read(client);
 965
 966     rc = nbd_receive_request(client->ioc, request);
 967     if (rc < 0) {
 968         if (rc != -EAGAIN) {
 969             rc = -EIO;
 970         }
 971         goto out;
 972     }
 973
 974     if ((request->from + request->len) < request->from) {
 975         LOG("integer overflow detected! "
 976             "you're probably being attacked");
 977         rc = -EINVAL;
 978         goto out;
 979     }
 980
 981     TRACE("Decoding type");
 982
 983     command = request->type & NBD_CMD_MASK_COMMAND;
 984     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
 985         if (request->len > NBD_MAX_BUFFER_SIZE) {
 986             LOG("len (%u) is larger than max len (%u)",
 987                 request->len, NBD_MAX_BUFFER_SIZE);
 988             rc = -EINVAL;
 989             goto out;
 990         }
 991
 992         req->data = blk_try_blockalign(client->exp->blk, request->len);
 993         if (req->data == NULL) {
 994             rc = -ENOMEM;
 995             goto out;
 996         }
 997     }
 998     if (command == NBD_CMD_WRITE) {
 999         TRACE("Reading %u byte(s)", request->len);
1000
1001         if (read_sync(client->ioc, req->data, request->len) != request->len) {
1002             LOG("reading from socket failed");
1003             rc = -EIO;
1004             goto out;
1005         }
1006     }
1007     rc = 0;
1008
1009 out:
1010     client->recv_coroutine = NULL;
1011     nbd_update_can_read(client);
1012
1013     return rc;
1014 }
1015
1016 static void nbd_trip(void *opaque)
1017 {
1018     NBDClient *client = opaque;
1019     NBDExport *exp = client->exp;
1020     NBDRequest *req;
1021     struct nbd_request request;
1022     struct nbd_reply reply;
1023     ssize_t ret;
1024     uint32_t command;
1025
1026     TRACE("Reading request.");
1027     if (client->closing) {
1028         return;
1029     }
1030
1031     req = nbd_request_get(client);
1032     ret = nbd_co_receive_request(req, &request);
1033     if (ret == -EAGAIN) {
1034         goto done;
1035     }
1036     if (ret == -EIO) {
1037         goto out;
1038     }
1039
1040     reply.handle = request.handle;
1041     reply.error = 0;
1042
1043     if (ret < 0) {
1044         reply.error = -ret;
1045         goto error_reply;
1046     }
1047     command = request.type & NBD_CMD_MASK_COMMAND;
1048     if (command != NBD_CMD_DISC && (request.from + request.len) > exp->size) {
1049             LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
1050             ", Offset: %" PRIu64 "\n",
1051                     request.from, request.len,
1052                     (uint64_t)exp->size, (uint64_t)exp->dev_offset);
1053         LOG("requested operation past EOF--bad client?");
1054         goto invalid_request;
1055     }
1056
1057     if (client->closing) {
1058         /*
1059          * The client may be closed when we are blocked in
1060          * nbd_co_receive_request()
1061          */
1062         goto done;
1063     }
1064
1065     switch (command) {
1066     case NBD_CMD_READ:
1067         TRACE("Request type is READ");
1068
1069         if (request.type & NBD_CMD_FLAG_FUA) {
1070             ret = blk_co_flush(exp->blk);
1071             if (ret < 0) {
1072                 LOG("flush failed");
1073                 reply.error = -ret;
1074                 goto error_reply;
1075             }
1076         }
1077
1078         ret = blk_read(exp->blk,
1079                        (request.from + exp->dev_offset) / BDRV_SECTOR_SIZE,
1080                        req->data, request.len / BDRV_SECTOR_SIZE);
1081         if (ret < 0) {
1082             LOG("reading from file failed");
1083             reply.error = -ret;
1084             goto error_reply;
1085         }
1086
1087         TRACE("Read %u byte(s)", request.len);
1088         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1089             goto out;
1090         break;
1091     case NBD_CMD_WRITE:
1092         TRACE("Request type is WRITE");
1093
1094         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1095             TRACE("Server is read-only, return error");
1096             reply.error = EROFS;
1097             goto error_reply;
1098         }
1099
1100         TRACE("Writing to device");
1101
1102         ret = blk_write(exp->blk,
1103                         (request.from + exp->dev_offset) / BDRV_SECTOR_SIZE,
1104                         req->data, request.len / BDRV_SECTOR_SIZE);
1105         if (ret < 0) {
1106             LOG("writing to file failed");
1107             reply.error = -ret;
1108             goto error_reply;
1109         }
1110
1111         if (request.type & NBD_CMD_FLAG_FUA) {
1112             ret = blk_co_flush(exp->blk);
1113             if (ret < 0) {
1114                 LOG("flush failed");
1115                 reply.error = -ret;
1116                 goto error_reply;
1117             }
1118         }
1119
1120         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1121             goto out;
1122         }
1123         break;
1124     case NBD_CMD_DISC:
1125         TRACE("Request type is DISCONNECT");
1126         errno = 0;
1127         goto out;
1128     case NBD_CMD_FLUSH:
1129         TRACE("Request type is FLUSH");
1130
1131         ret = blk_co_flush(exp->blk);
1132         if (ret < 0) {
1133             LOG("flush failed");
1134             reply.error = -ret;
1135         }
1136         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1137             goto out;
1138         }
1139         break;
1140     case NBD_CMD_TRIM:
1141         TRACE("Request type is TRIM");
1142         ret = blk_co_discard(exp->blk, (request.from + exp->dev_offset)
1143                                        / BDRV_SECTOR_SIZE,
1144                              request.len / BDRV_SECTOR_SIZE);
1145         if (ret < 0) {
1146             LOG("discard failed");
1147             reply.error = -ret;
1148         }
1149         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1150             goto out;
1151         }
1152         break;
1153     default:
1154         LOG("invalid request type (%u) received", request.type);
1155     invalid_request:
1156         reply.error = EINVAL;
1157     error_reply:
1158         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1159             goto out;
1160         }
1161         break;
1162     }
1163
1164     TRACE("Request/Reply complete");
1165
1166 done:
1167     nbd_request_put(req);
1168     return;
1169
1170 out:
1171     nbd_request_put(req);
1172     client_close(client);
1173 }
1174
1175 static void nbd_read(void *opaque)
1176 {
1177     NBDClient *client = opaque;
1178
1179     if (client->recv_coroutine) {
1180         qemu_coroutine_enter(client->recv_coroutine, NULL);
1181     } else {
1182         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip), client);
1183     }
1184 }
1185
1186 static void nbd_restart_write(void *opaque)
1187 {
1188     NBDClient *client = opaque;
1189
1190     qemu_coroutine_enter(client->send_coroutine, NULL);
1191 }
1192
1193 static void nbd_set_handlers(NBDClient *client)
1194 {
1195     if (client->exp && client->exp->ctx) {
1196         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1197                            true,
1198                            client->can_read ? nbd_read : NULL,
1199                            client->send_coroutine ? nbd_restart_write : NULL,
1200                            client);
1201     }
1202 }
1203
1204 static void nbd_unset_handlers(NBDClient *client)
1205 {
1206     if (client->exp && client->exp->ctx) {
1207         aio_set_fd_handler(client->exp->ctx, client->sioc->fd,
1208                            true, NULL, NULL, NULL);
1209     }
1210 }
1211
1212 static void nbd_update_can_read(NBDClient *client)
1213 {
1214     bool can_read = client->recv_coroutine ||
1215                     client->nb_requests < MAX_NBD_REQUESTS;
1216
1217     if (can_read != client->can_read) {
1218         client->can_read = can_read;
1219         nbd_set_handlers(client);
1220
1221         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1222          * in nbd_set_handlers() will have taken care of that */
1223     }
1224 }
1225
1226 static coroutine_fn void nbd_co_client_start(void *opaque)
1227 {
1228     NBDClientNewData *data = opaque;
1229     NBDClient *client = data->client;
1230     NBDExport *exp = client->exp;
1231
1232     if (exp) {
1233         nbd_export_get(exp);
1234     }
1235     if (nbd_negotiate(data)) {
1236         client_close(client);
1237         goto out;
1238     }
1239     qemu_co_mutex_init(&client->send_lock);
1240     nbd_set_handlers(client);
1241
1242     if (exp) {
1243         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1244     }
1245 out:
1246     g_free(data);
1247 }
1248
1249 void nbd_client_new(NBDExport *exp,
1250                     QIOChannelSocket *sioc,
1251                     QCryptoTLSCreds *tlscreds,
1252                     const char *tlsaclname,
1253                     void (*close_fn)(NBDClient *))
1254 {
1255     NBDClient *client;
1256     NBDClientNewData *data = g_new(NBDClientNewData, 1);
1257
1258     client = g_malloc0(sizeof(NBDClient));
1259     client->refcount = 1;
1260     client->exp = exp;
1261     client->tlscreds = tlscreds;
1262     if (tlscreds) {
1263         object_ref(OBJECT(client->tlscreds));
1264     }
1265     client->tlsaclname = g_strdup(tlsaclname);
1266     client->sioc = sioc;
1267     object_ref(OBJECT(client->sioc));
1268     client->ioc = QIO_CHANNEL(sioc);
1269     object_ref(OBJECT(client->ioc));
1270     client->can_read = true;
1271     client->close = close_fn;
1272
1273     data->client = client;
1274     data->co = qemu_coroutine_create(nbd_co_client_start);
1275     qemu_coroutine_enter(data->co, data);
1276 }