qemu/qemu-nbd.c
<<
>>
Prefs
   1/*
   2 *  Copyright (C) 2005  Anthony Liguori <anthony@codemonkey.ws>
   3 *
   4 *  Network Block Device
   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 <getopt.h>
  21#include <libgen.h>
  22#include <pthread.h>
  23
  24#include "qapi/error.h"
  25#include "qemu/cutils.h"
  26#include "sysemu/block-backend.h"
  27#include "block/block_int.h"
  28#include "block/nbd.h"
  29#include "qemu/main-loop.h"
  30#include "qemu/option.h"
  31#include "qemu/error-report.h"
  32#include "qemu/config-file.h"
  33#include "qemu/bswap.h"
  34#include "qemu/log.h"
  35#include "qemu/systemd.h"
  36#include "block/snapshot.h"
  37#include "qapi/qmp/qdict.h"
  38#include "qapi/qmp/qstring.h"
  39#include "qom/object_interfaces.h"
  40#include "io/channel-socket.h"
  41#include "io/net-listener.h"
  42#include "crypto/init.h"
  43#include "trace/control.h"
  44#include "qemu-version.h"
  45
  46#ifdef __linux__
  47#define HAVE_NBD_DEVICE 1
  48#else
  49#define HAVE_NBD_DEVICE 0
  50#endif
  51
  52#define SOCKET_PATH                "/var/lock/qemu-nbd-%s"
  53#define QEMU_NBD_OPT_CACHE         256
  54#define QEMU_NBD_OPT_AIO           257
  55#define QEMU_NBD_OPT_DISCARD       258
  56#define QEMU_NBD_OPT_DETECT_ZEROES 259
  57#define QEMU_NBD_OPT_OBJECT        260
  58#define QEMU_NBD_OPT_TLSCREDS      261
  59#define QEMU_NBD_OPT_IMAGE_OPTS    262
  60#define QEMU_NBD_OPT_FORK          263
  61#define QEMU_NBD_OPT_TLSAUTHZ      264
  62
  63#define MBR_SIZE 512
  64
  65static NBDExport *export;
  66static int verbose;
  67static char *srcpath;
  68static SocketAddress *saddr;
  69static int persistent = 0;
  70static enum { RUNNING, TERMINATE, TERMINATING, TERMINATED } state;
  71static int shared = 1;
  72static int nb_fds;
  73static QIONetListener *server;
  74static QCryptoTLSCreds *tlscreds;
  75static const char *tlsauthz;
  76
  77static void usage(const char *name)
  78{
  79    (printf) (
  80"Usage: %s [OPTIONS] FILE\n"
  81"  or:  %s -L [OPTIONS]\n"
  82"QEMU Disk Network Block Device Utility\n"
  83"\n"
  84"  -h, --help                display this help and exit\n"
  85"  -V, --version             output version information and exit\n"
  86"\n"
  87"Connection properties:\n"
  88"  -p, --port=PORT           port to listen on (default `%d')\n"
  89"  -b, --bind=IFACE          interface to bind to (default `0.0.0.0')\n"
  90"  -k, --socket=PATH         path to the unix socket\n"
  91"                            (default '"SOCKET_PATH"')\n"
  92"  -e, --shared=NUM          device can be shared by NUM clients (default '1')\n"
  93"  -t, --persistent          don't exit on the last connection\n"
  94"  -v, --verbose             display extra debugging information\n"
  95"  -x, --export-name=NAME    expose export by name (default is empty string)\n"
  96"  -D, --description=TEXT    export a human-readable description\n"
  97"\n"
  98"Exposing part of the image:\n"
  99"  -o, --offset=OFFSET       offset into the image\n"
 100"  -P, --partition=NUM       only expose partition NUM\n"
 101"  -B, --bitmap=NAME         expose a persistent dirty bitmap\n"
 102"\n"
 103"General purpose options:\n"
 104"  -L, --list                list exports available from another NBD server\n"
 105"  --object type,id=ID,...   define an object such as 'secret' for providing\n"
 106"                            passwords and/or encryption keys\n"
 107"  --tls-creds=ID            use id of an earlier --object to provide TLS\n"
 108"  --tls-authz=ID            use id of an earlier --object to provide\n"
 109"                            authorization\n"
 110"  -T, --trace [[enable=]<pattern>][,events=<file>][,file=<file>]\n"
 111"                            specify tracing options\n"
 112"  --fork                    fork off the server process and exit the parent\n"
 113"                            once the server is running\n"
 114#if HAVE_NBD_DEVICE
 115"\n"
 116"Kernel NBD client support:\n"
 117"  -c, --connect=DEV         connect FILE to the local NBD device DEV\n"
 118"  -d, --disconnect          disconnect the specified device\n"
 119#endif
 120"\n"
 121"Block device options:\n"
 122"  -f, --format=FORMAT       set image format (raw, qcow2, ...)\n"
 123"  -r, --read-only           export read-only\n"
 124"  -s, --snapshot            use FILE as an external snapshot, create a temporary\n"
 125"                            file with backing_file=FILE, redirect the write to\n"
 126"                            the temporary one\n"
 127"  -l, --load-snapshot=SNAPSHOT_PARAM\n"
 128"                            load an internal snapshot inside FILE and export it\n"
 129"                            as an read-only device, SNAPSHOT_PARAM format is\n"
 130"                            'snapshot.id=[ID],snapshot.name=[NAME]', or\n"
 131"                            '[ID_OR_NAME]'\n"
 132"  -n, --nocache             disable host cache\n"
 133"      --cache=MODE          set cache mode (none, writeback, ...)\n"
 134"      --aio=MODE            set AIO mode (native or threads)\n"
 135"      --discard=MODE        set discard mode (ignore, unmap)\n"
 136"      --detect-zeroes=MODE  set detect-zeroes mode (off, on, unmap)\n"
 137"      --image-opts          treat FILE as a full set of image options\n"
 138"\n"
 139QEMU_HELP_BOTTOM "\n"
 140    , name, name, NBD_DEFAULT_PORT, "DEVICE");
 141}
 142
 143static void version(const char *name)
 144{
 145    printf(
 146"%s " QEMU_FULL_VERSION "\n"
 147"Written by Anthony Liguori.\n"
 148"\n"
 149QEMU_COPYRIGHT "\n"
 150"This is free software; see the source for copying conditions.  There is NO\n"
 151"warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n"
 152    , name);
 153}
 154
 155struct partition_record
 156{
 157    uint8_t bootable;
 158    uint8_t start_head;
 159    uint32_t start_cylinder;
 160    uint8_t start_sector;
 161    uint8_t system;
 162    uint8_t end_head;
 163    uint8_t end_cylinder;
 164    uint8_t end_sector;
 165    uint32_t start_sector_abs;
 166    uint32_t nb_sectors_abs;
 167};
 168
 169static void read_partition(uint8_t *p, struct partition_record *r)
 170{
 171    r->bootable = p[0];
 172    r->start_head = p[1];
 173    r->start_cylinder = p[3] | ((p[2] << 2) & 0x0300);
 174    r->start_sector = p[2] & 0x3f;
 175    r->system = p[4];
 176    r->end_head = p[5];
 177    r->end_cylinder = p[7] | ((p[6] << 2) & 0x300);
 178    r->end_sector = p[6] & 0x3f;
 179
 180    r->start_sector_abs = ldl_le_p(p + 8);
 181    r->nb_sectors_abs   = ldl_le_p(p + 12);
 182}
 183
 184static int find_partition(BlockBackend *blk, int partition,
 185                          uint64_t *offset, uint64_t *size)
 186{
 187    struct partition_record mbr[4];
 188    uint8_t data[MBR_SIZE];
 189    int i;
 190    int ext_partnum = 4;
 191    int ret;
 192
 193    ret = blk_pread(blk, 0, data, sizeof(data));
 194    if (ret < 0) {
 195        error_report("error while reading: %s", strerror(-ret));
 196        exit(EXIT_FAILURE);
 197    }
 198
 199    if (data[510] != 0x55 || data[511] != 0xaa) {
 200        return -EINVAL;
 201    }
 202
 203    for (i = 0; i < 4; i++) {
 204        read_partition(&data[446 + 16 * i], &mbr[i]);
 205
 206        if (!mbr[i].system || !mbr[i].nb_sectors_abs) {
 207            continue;
 208        }
 209
 210        if (mbr[i].system == 0xF || mbr[i].system == 0x5) {
 211            struct partition_record ext[4];
 212            uint8_t data1[MBR_SIZE];
 213            int j;
 214
 215            ret = blk_pread(blk, mbr[i].start_sector_abs * MBR_SIZE,
 216                            data1, sizeof(data1));
 217            if (ret < 0) {
 218                error_report("error while reading: %s", strerror(-ret));
 219                exit(EXIT_FAILURE);
 220            }
 221
 222            for (j = 0; j < 4; j++) {
 223                read_partition(&data1[446 + 16 * j], &ext[j]);
 224                if (!ext[j].system || !ext[j].nb_sectors_abs) {
 225                    continue;
 226                }
 227
 228                if ((ext_partnum + j + 1) == partition) {
 229                    *offset = (uint64_t)ext[j].start_sector_abs << 9;
 230                    *size = (uint64_t)ext[j].nb_sectors_abs << 9;
 231                    return 0;
 232                }
 233            }
 234            ext_partnum += 4;
 235        } else if ((i + 1) == partition) {
 236            *offset = (uint64_t)mbr[i].start_sector_abs << 9;
 237            *size = (uint64_t)mbr[i].nb_sectors_abs << 9;
 238            return 0;
 239        }
 240    }
 241
 242    return -ENOENT;
 243}
 244
 245static void termsig_handler(int signum)
 246{
 247    atomic_cmpxchg(&state, RUNNING, TERMINATE);
 248    qemu_notify_event();
 249}
 250
 251
 252static int qemu_nbd_client_list(SocketAddress *saddr, QCryptoTLSCreds *tls,
 253                                const char *hostname)
 254{
 255    int ret = EXIT_FAILURE;
 256    int rc;
 257    Error *err = NULL;
 258    QIOChannelSocket *sioc;
 259    NBDExportInfo *list;
 260    int i, j;
 261
 262    sioc = qio_channel_socket_new();
 263    if (qio_channel_socket_connect_sync(sioc, saddr, &err) < 0) {
 264        error_report_err(err);
 265        return EXIT_FAILURE;
 266    }
 267    rc = nbd_receive_export_list(QIO_CHANNEL(sioc), tls, hostname, &list,
 268                                 &err);
 269    if (rc < 0) {
 270        if (err) {
 271            error_report_err(err);
 272        }
 273        goto out;
 274    }
 275    printf("exports available: %d\n", rc);
 276    for (i = 0; i < rc; i++) {
 277        printf(" export: '%s'\n", list[i].name);
 278        if (list[i].description && *list[i].description) {
 279            printf("  description: %s\n", list[i].description);
 280        }
 281        if (list[i].flags & NBD_FLAG_HAS_FLAGS) {
 282            printf("  size:  %" PRIu64 "\n", list[i].size);
 283            printf("  flags: 0x%x (", list[i].flags);
 284            if (list[i].flags & NBD_FLAG_READ_ONLY) {
 285                printf(" readonly");
 286            }
 287            if (list[i].flags & NBD_FLAG_SEND_FLUSH) {
 288                printf(" flush");
 289            }
 290            if (list[i].flags & NBD_FLAG_SEND_FUA) {
 291                printf(" fua");
 292            }
 293            if (list[i].flags & NBD_FLAG_ROTATIONAL) {
 294                printf(" rotational");
 295            }
 296            if (list[i].flags & NBD_FLAG_SEND_TRIM) {
 297                printf(" trim");
 298            }
 299            if (list[i].flags & NBD_FLAG_SEND_WRITE_ZEROES) {
 300                printf(" zeroes");
 301            }
 302            if (list[i].flags & NBD_FLAG_SEND_DF) {
 303                printf(" df");
 304            }
 305            if (list[i].flags & NBD_FLAG_CAN_MULTI_CONN) {
 306                printf(" multi");
 307            }
 308            if (list[i].flags & NBD_FLAG_SEND_RESIZE) {
 309                printf(" resize");
 310            }
 311            if (list[i].flags & NBD_FLAG_SEND_CACHE) {
 312                printf(" cache");
 313            }
 314            printf(" )\n");
 315        }
 316        if (list[i].min_block) {
 317            printf("  min block: %u\n", list[i].min_block);
 318            printf("  opt block: %u\n", list[i].opt_block);
 319            printf("  max block: %u\n", list[i].max_block);
 320        }
 321        if (list[i].n_contexts) {
 322            printf("  available meta contexts: %d\n", list[i].n_contexts);
 323            for (j = 0; j < list[i].n_contexts; j++) {
 324                printf("   %s\n", list[i].contexts[j]);
 325            }
 326        }
 327    }
 328    nbd_free_export_list(list, rc);
 329
 330    ret = EXIT_SUCCESS;
 331 out:
 332    object_unref(OBJECT(sioc));
 333    return ret;
 334}
 335
 336
 337#if HAVE_NBD_DEVICE
 338static void *show_parts(void *arg)
 339{
 340    char *device = arg;
 341    int nbd;
 342
 343    /* linux just needs an open() to trigger
 344     * the partition table update
 345     * but remember to load the module with max_part != 0 :
 346     *     modprobe nbd max_part=63
 347     */
 348    nbd = open(device, O_RDWR);
 349    if (nbd >= 0) {
 350        close(nbd);
 351    }
 352    return NULL;
 353}
 354
 355static void *nbd_client_thread(void *arg)
 356{
 357    char *device = arg;
 358    NBDExportInfo info = { .request_sizes = false, .name = g_strdup("") };
 359    QIOChannelSocket *sioc;
 360    int fd;
 361    int ret;
 362    pthread_t show_parts_thread;
 363    Error *local_error = NULL;
 364
 365    sioc = qio_channel_socket_new();
 366    if (qio_channel_socket_connect_sync(sioc,
 367                                        saddr,
 368                                        &local_error) < 0) {
 369        error_report_err(local_error);
 370        goto out;
 371    }
 372
 373    ret = nbd_receive_negotiate(QIO_CHANNEL(sioc),
 374                                NULL, NULL, NULL, &info, &local_error);
 375    if (ret < 0) {
 376        if (local_error) {
 377            error_report_err(local_error);
 378        }
 379        goto out_socket;
 380    }
 381
 382    fd = open(device, O_RDWR);
 383    if (fd < 0) {
 384        /* Linux-only, we can use %m in printf.  */
 385        error_report("Failed to open %s: %m", device);
 386        goto out_socket;
 387    }
 388
 389    ret = nbd_init(fd, sioc, &info, &local_error);
 390    if (ret < 0) {
 391        error_report_err(local_error);
 392        goto out_fd;
 393    }
 394
 395    /* update partition table */
 396    pthread_create(&show_parts_thread, NULL, show_parts, device);
 397
 398    if (verbose) {
 399        fprintf(stderr, "NBD device %s is now connected to %s\n",
 400                device, srcpath);
 401    } else {
 402        /* Close stderr so that the qemu-nbd process exits.  */
 403        dup2(STDOUT_FILENO, STDERR_FILENO);
 404    }
 405
 406    ret = nbd_client(fd);
 407    if (ret) {
 408        goto out_fd;
 409    }
 410    close(fd);
 411    object_unref(OBJECT(sioc));
 412    g_free(info.name);
 413    kill(getpid(), SIGTERM);
 414    return (void *) EXIT_SUCCESS;
 415
 416out_fd:
 417    close(fd);
 418out_socket:
 419    object_unref(OBJECT(sioc));
 420out:
 421    g_free(info.name);
 422    kill(getpid(), SIGTERM);
 423    return (void *) EXIT_FAILURE;
 424}
 425#endif /* HAVE_NBD_DEVICE */
 426
 427static int nbd_can_accept(void)
 428{
 429    return state == RUNNING && nb_fds < shared;
 430}
 431
 432static void nbd_export_closed(NBDExport *export)
 433{
 434    assert(state == TERMINATING);
 435    state = TERMINATED;
 436}
 437
 438static void nbd_update_server_watch(void);
 439
 440static void nbd_client_closed(NBDClient *client, bool negotiated)
 441{
 442    nb_fds--;
 443    if (negotiated && nb_fds == 0 && !persistent && state == RUNNING) {
 444        state = TERMINATE;
 445    }
 446    nbd_update_server_watch();
 447    nbd_client_put(client);
 448}
 449
 450static void nbd_accept(QIONetListener *listener, QIOChannelSocket *cioc,
 451                       gpointer opaque)
 452{
 453    if (state >= TERMINATE) {
 454        return;
 455    }
 456
 457    nb_fds++;
 458    nbd_update_server_watch();
 459    nbd_client_new(cioc, tlscreds, tlsauthz, nbd_client_closed);
 460}
 461
 462static void nbd_update_server_watch(void)
 463{
 464    if (nbd_can_accept()) {
 465        qio_net_listener_set_client_func(server, nbd_accept, NULL, NULL);
 466    } else {
 467        qio_net_listener_set_client_func(server, NULL, NULL, NULL);
 468    }
 469}
 470
 471
 472static SocketAddress *nbd_build_socket_address(const char *sockpath,
 473                                               const char *bindto,
 474                                               const char *port)
 475{
 476    SocketAddress *saddr;
 477
 478    saddr = g_new0(SocketAddress, 1);
 479    if (sockpath) {
 480        saddr->type = SOCKET_ADDRESS_TYPE_UNIX;
 481        saddr->u.q_unix.path = g_strdup(sockpath);
 482    } else {
 483        InetSocketAddress *inet;
 484        saddr->type = SOCKET_ADDRESS_TYPE_INET;
 485        inet = &saddr->u.inet;
 486        inet->host = g_strdup(bindto);
 487        if (port) {
 488            inet->port = g_strdup(port);
 489        } else  {
 490            inet->port = g_strdup_printf("%d", NBD_DEFAULT_PORT);
 491        }
 492    }
 493
 494    return saddr;
 495}
 496
 497
 498static QemuOptsList file_opts = {
 499    .name = "file",
 500    .implied_opt_name = "file",
 501    .head = QTAILQ_HEAD_INITIALIZER(file_opts.head),
 502    .desc = {
 503        /* no elements => accept any params */
 504        { /* end of list */ }
 505    },
 506};
 507
 508static QemuOptsList qemu_object_opts = {
 509    .name = "object",
 510    .implied_opt_name = "qom-type",
 511    .head = QTAILQ_HEAD_INITIALIZER(qemu_object_opts.head),
 512    .desc = {
 513        { }
 514    },
 515};
 516
 517
 518
 519static QCryptoTLSCreds *nbd_get_tls_creds(const char *id, bool list,
 520                                          Error **errp)
 521{
 522    Object *obj;
 523    QCryptoTLSCreds *creds;
 524
 525    obj = object_resolve_path_component(
 526        object_get_objects_root(), id);
 527    if (!obj) {
 528        error_setg(errp, "No TLS credentials with id '%s'",
 529                   id);
 530        return NULL;
 531    }
 532    creds = (QCryptoTLSCreds *)
 533        object_dynamic_cast(obj, TYPE_QCRYPTO_TLS_CREDS);
 534    if (!creds) {
 535        error_setg(errp, "Object with id '%s' is not TLS credentials",
 536                   id);
 537        return NULL;
 538    }
 539
 540    if (list) {
 541        if (creds->endpoint != QCRYPTO_TLS_CREDS_ENDPOINT_CLIENT) {
 542            error_setg(errp,
 543                       "Expecting TLS credentials with a client endpoint");
 544            return NULL;
 545        }
 546    } else {
 547        if (creds->endpoint != QCRYPTO_TLS_CREDS_ENDPOINT_SERVER) {
 548            error_setg(errp,
 549                       "Expecting TLS credentials with a server endpoint");
 550            return NULL;
 551        }
 552    }
 553    object_ref(obj);
 554    return creds;
 555}
 556
 557static void setup_address_and_port(const char **address, const char **port)
 558{
 559    if (*address == NULL) {
 560        *address = "0.0.0.0";
 561    }
 562
 563    if (*port == NULL) {
 564        *port = stringify(NBD_DEFAULT_PORT);
 565    }
 566}
 567
 568/*
 569 * Check socket parameters compatibility when socket activation is used.
 570 */
 571static const char *socket_activation_validate_opts(const char *device,
 572                                                   const char *sockpath,
 573                                                   const char *address,
 574                                                   const char *port,
 575                                                   bool list)
 576{
 577    if (device != NULL) {
 578        return "NBD device can't be set when using socket activation";
 579    }
 580
 581    if (sockpath != NULL) {
 582        return "Unix socket can't be set when using socket activation";
 583    }
 584
 585    if (address != NULL) {
 586        return "The interface can't be set when using socket activation";
 587    }
 588
 589    if (port != NULL) {
 590        return "TCP port number can't be set when using socket activation";
 591    }
 592
 593    if (list) {
 594        return "List mode is incompatible with socket activation";
 595    }
 596
 597    return NULL;
 598}
 599
 600static void qemu_nbd_shutdown(void)
 601{
 602    job_cancel_sync_all();
 603    bdrv_close_all();
 604}
 605
 606int main(int argc, char **argv)
 607{
 608    BlockBackend *blk;
 609    BlockDriverState *bs;
 610    uint64_t dev_offset = 0;
 611    uint16_t nbdflags = 0;
 612    bool disconnect = false;
 613    const char *bindto = NULL;
 614    const char *port = NULL;
 615    char *sockpath = NULL;
 616    char *device = NULL;
 617    int64_t fd_size;
 618    QemuOpts *sn_opts = NULL;
 619    const char *sn_id_or_name = NULL;
 620    const char *sopt = "hVb:o:p:rsnP:c:dvk:e:f:tl:x:T:D:B:L";
 621    struct option lopt[] = {
 622        { "help", no_argument, NULL, 'h' },
 623        { "version", no_argument, NULL, 'V' },
 624        { "bind", required_argument, NULL, 'b' },
 625        { "port", required_argument, NULL, 'p' },
 626        { "socket", required_argument, NULL, 'k' },
 627        { "offset", required_argument, NULL, 'o' },
 628        { "read-only", no_argument, NULL, 'r' },
 629        { "partition", required_argument, NULL, 'P' },
 630        { "bitmap", required_argument, NULL, 'B' },
 631        { "connect", required_argument, NULL, 'c' },
 632        { "disconnect", no_argument, NULL, 'd' },
 633        { "list", no_argument, NULL, 'L' },
 634        { "snapshot", no_argument, NULL, 's' },
 635        { "load-snapshot", required_argument, NULL, 'l' },
 636        { "nocache", no_argument, NULL, 'n' },
 637        { "cache", required_argument, NULL, QEMU_NBD_OPT_CACHE },
 638        { "aio", required_argument, NULL, QEMU_NBD_OPT_AIO },
 639        { "discard", required_argument, NULL, QEMU_NBD_OPT_DISCARD },
 640        { "detect-zeroes", required_argument, NULL,
 641          QEMU_NBD_OPT_DETECT_ZEROES },
 642        { "shared", required_argument, NULL, 'e' },
 643        { "format", required_argument, NULL, 'f' },
 644        { "persistent", no_argument, NULL, 't' },
 645        { "verbose", no_argument, NULL, 'v' },
 646        { "object", required_argument, NULL, QEMU_NBD_OPT_OBJECT },
 647        { "export-name", required_argument, NULL, 'x' },
 648        { "description", required_argument, NULL, 'D' },
 649        { "tls-creds", required_argument, NULL, QEMU_NBD_OPT_TLSCREDS },
 650        { "tls-authz", required_argument, NULL, QEMU_NBD_OPT_TLSAUTHZ },
 651        { "image-opts", no_argument, NULL, QEMU_NBD_OPT_IMAGE_OPTS },
 652        { "trace", required_argument, NULL, 'T' },
 653        { "fork", no_argument, NULL, QEMU_NBD_OPT_FORK },
 654        { NULL, 0, NULL, 0 }
 655    };
 656    int ch;
 657    int opt_ind = 0;
 658    int flags = BDRV_O_RDWR;
 659    int partition = 0;
 660    int ret = 0;
 661    bool seen_cache = false;
 662    bool seen_discard = false;
 663    bool seen_aio = false;
 664    pthread_t client_thread;
 665    const char *fmt = NULL;
 666    Error *local_err = NULL;
 667    BlockdevDetectZeroesOptions detect_zeroes = BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF;
 668    QDict *options = NULL;
 669    const char *export_name = NULL; /* defaults to "" later for server mode */
 670    const char *export_description = NULL;
 671    const char *bitmap = NULL;
 672    const char *tlscredsid = NULL;
 673    bool imageOpts = false;
 674    bool writethrough = true;
 675    char *trace_file = NULL;
 676    bool fork_process = false;
 677    bool list = false;
 678    int old_stderr = -1;
 679    unsigned socket_activation;
 680
 681    /* The client thread uses SIGTERM to interrupt the server.  A signal
 682     * handler ensures that "qemu-nbd -v -c" exits with a nice status code.
 683     */
 684    struct sigaction sa_sigterm;
 685    memset(&sa_sigterm, 0, sizeof(sa_sigterm));
 686    sa_sigterm.sa_handler = termsig_handler;
 687    sigaction(SIGTERM, &sa_sigterm, NULL);
 688
 689#ifdef CONFIG_POSIX
 690    signal(SIGPIPE, SIG_IGN);
 691#endif
 692
 693    module_call_init(MODULE_INIT_TRACE);
 694    error_set_progname(argv[0]);
 695    qcrypto_init(&error_fatal);
 696
 697    module_call_init(MODULE_INIT_QOM);
 698    qemu_add_opts(&qemu_object_opts);
 699    qemu_add_opts(&qemu_trace_opts);
 700    qemu_init_exec_dir(argv[0]);
 701
 702    while ((ch = getopt_long(argc, argv, sopt, lopt, &opt_ind)) != -1) {
 703        switch (ch) {
 704        case 's':
 705            flags |= BDRV_O_SNAPSHOT;
 706            break;
 707        case 'n':
 708            optarg = (char *) "none";
 709            /* fallthrough */
 710        case QEMU_NBD_OPT_CACHE:
 711            if (seen_cache) {
 712                error_report("-n and --cache can only be specified once");
 713                exit(EXIT_FAILURE);
 714            }
 715            seen_cache = true;
 716            if (bdrv_parse_cache_mode(optarg, &flags, &writethrough) == -1) {
 717                error_report("Invalid cache mode `%s'", optarg);
 718                exit(EXIT_FAILURE);
 719            }
 720            break;
 721        case QEMU_NBD_OPT_AIO:
 722            if (seen_aio) {
 723                error_report("--aio can only be specified once");
 724                exit(EXIT_FAILURE);
 725            }
 726            seen_aio = true;
 727            if (!strcmp(optarg, "native")) {
 728                flags |= BDRV_O_NATIVE_AIO;
 729            } else if (!strcmp(optarg, "threads")) {
 730                /* this is the default */
 731            } else {
 732               error_report("invalid aio mode `%s'", optarg);
 733               exit(EXIT_FAILURE);
 734            }
 735            break;
 736        case QEMU_NBD_OPT_DISCARD:
 737            if (seen_discard) {
 738                error_report("--discard can only be specified once");
 739                exit(EXIT_FAILURE);
 740            }
 741            seen_discard = true;
 742            if (bdrv_parse_discard_flags(optarg, &flags) == -1) {
 743                error_report("Invalid discard mode `%s'", optarg);
 744                exit(EXIT_FAILURE);
 745            }
 746            break;
 747        case QEMU_NBD_OPT_DETECT_ZEROES:
 748            detect_zeroes =
 749                qapi_enum_parse(&BlockdevDetectZeroesOptions_lookup,
 750                                optarg,
 751                                BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF,
 752                                &local_err);
 753            if (local_err) {
 754                error_reportf_err(local_err,
 755                                  "Failed to parse detect_zeroes mode: ");
 756                exit(EXIT_FAILURE);
 757            }
 758            if (detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP &&
 759                !(flags & BDRV_O_UNMAP)) {
 760                error_report("setting detect-zeroes to unmap is not allowed "
 761                             "without setting discard operation to unmap");
 762                exit(EXIT_FAILURE);
 763            }
 764            break;
 765        case 'b':
 766            bindto = optarg;
 767            break;
 768        case 'p':
 769            port = optarg;
 770            break;
 771        case 'o':
 772            if (qemu_strtou64(optarg, NULL, 0, &dev_offset) < 0) {
 773                error_report("Invalid offset '%s'", optarg);
 774                exit(EXIT_FAILURE);
 775            }
 776            break;
 777        case 'l':
 778            if (strstart(optarg, SNAPSHOT_OPT_BASE, NULL)) {
 779                sn_opts = qemu_opts_parse_noisily(&internal_snapshot_opts,
 780                                                  optarg, false);
 781                if (!sn_opts) {
 782                    error_report("Failed in parsing snapshot param `%s'",
 783                                 optarg);
 784                    exit(EXIT_FAILURE);
 785                }
 786            } else {
 787                sn_id_or_name = optarg;
 788            }
 789            /* fall through */
 790        case 'r':
 791            nbdflags |= NBD_FLAG_READ_ONLY;
 792            flags &= ~BDRV_O_RDWR;
 793            break;
 794        case 'P':
 795            warn_report("The '-P' option is deprecated; use --image-opts with "
 796                        "a raw device wrapper for subset exports instead");
 797            if (qemu_strtoi(optarg, NULL, 0, &partition) < 0 ||
 798                partition < 1 || partition > 8) {
 799                error_report("Invalid partition '%s'", optarg);
 800                exit(EXIT_FAILURE);
 801            }
 802            break;
 803        case 'B':
 804            bitmap = optarg;
 805            break;
 806        case 'k':
 807            sockpath = optarg;
 808            if (sockpath[0] != '/') {
 809                error_report("socket path must be absolute");
 810                exit(EXIT_FAILURE);
 811            }
 812            break;
 813        case 'd':
 814            disconnect = true;
 815            break;
 816        case 'c':
 817            device = optarg;
 818            break;
 819        case 'e':
 820            if (qemu_strtoi(optarg, NULL, 0, &shared) < 0 ||
 821                shared < 1) {
 822                error_report("Invalid shared device number '%s'", optarg);
 823                exit(EXIT_FAILURE);
 824            }
 825            break;
 826        case 'f':
 827            fmt = optarg;
 828            break;
 829        case 't':
 830            persistent = 1;
 831            break;
 832        case 'x':
 833            export_name = optarg;
 834            break;
 835        case 'D':
 836            export_description = optarg;
 837            break;
 838        case 'v':
 839            verbose = 1;
 840            break;
 841        case 'V':
 842            version(argv[0]);
 843            exit(0);
 844            break;
 845        case 'h':
 846            usage(argv[0]);
 847            exit(0);
 848            break;
 849        case '?':
 850            error_report("Try `%s --help' for more information.", argv[0]);
 851            exit(EXIT_FAILURE);
 852        case QEMU_NBD_OPT_OBJECT: {
 853            QemuOpts *opts;
 854            opts = qemu_opts_parse_noisily(&qemu_object_opts,
 855                                           optarg, true);
 856            if (!opts) {
 857                exit(EXIT_FAILURE);
 858            }
 859        }   break;
 860        case QEMU_NBD_OPT_TLSCREDS:
 861            tlscredsid = optarg;
 862            break;
 863        case QEMU_NBD_OPT_IMAGE_OPTS:
 864            imageOpts = true;
 865            break;
 866        case 'T':
 867            g_free(trace_file);
 868            trace_file = trace_opt_parse(optarg);
 869            break;
 870        case QEMU_NBD_OPT_TLSAUTHZ:
 871            tlsauthz = optarg;
 872            break;
 873        case QEMU_NBD_OPT_FORK:
 874            fork_process = true;
 875            break;
 876        case 'L':
 877            list = true;
 878            break;
 879        }
 880    }
 881
 882    if (list) {
 883        if (argc != optind) {
 884            error_report("List mode is incompatible with a file name");
 885            exit(EXIT_FAILURE);
 886        }
 887        if (export_name || export_description || dev_offset || partition ||
 888            device || disconnect || fmt || sn_id_or_name || bitmap ||
 889            seen_aio || seen_discard || seen_cache) {
 890            error_report("List mode is incompatible with per-device settings");
 891            exit(EXIT_FAILURE);
 892        }
 893        if (fork_process) {
 894            error_report("List mode is incompatible with forking");
 895            exit(EXIT_FAILURE);
 896        }
 897    } else if ((argc - optind) != 1) {
 898        error_report("Invalid number of arguments");
 899        error_printf("Try `%s --help' for more information.\n", argv[0]);
 900        exit(EXIT_FAILURE);
 901    } else if (!export_name) {
 902        export_name = "";
 903    }
 904
 905    qemu_opts_foreach(&qemu_object_opts,
 906                      user_creatable_add_opts_foreach,
 907                      NULL, &error_fatal);
 908
 909    if (!trace_init_backends()) {
 910        exit(1);
 911    }
 912    trace_init_file(trace_file);
 913    qemu_set_log(LOG_TRACE);
 914
 915    socket_activation = check_socket_activation();
 916    if (socket_activation == 0) {
 917        setup_address_and_port(&bindto, &port);
 918    } else {
 919        /* Using socket activation - check user didn't use -p etc. */
 920        const char *err_msg = socket_activation_validate_opts(device, sockpath,
 921                                                              bindto, port,
 922                                                              list);
 923        if (err_msg != NULL) {
 924            error_report("%s", err_msg);
 925            exit(EXIT_FAILURE);
 926        }
 927
 928        /* qemu-nbd can only listen on a single socket.  */
 929        if (socket_activation > 1) {
 930            error_report("qemu-nbd does not support socket activation with %s > 1",
 931                         "LISTEN_FDS");
 932            exit(EXIT_FAILURE);
 933        }
 934    }
 935
 936    if (tlscredsid) {
 937        if (sockpath) {
 938            error_report("TLS is only supported with IPv4/IPv6");
 939            exit(EXIT_FAILURE);
 940        }
 941        if (device) {
 942            error_report("TLS is not supported with a host device");
 943            exit(EXIT_FAILURE);
 944        }
 945        if (tlsauthz && list) {
 946            error_report("TLS authorization is incompatible with export list");
 947            exit(EXIT_FAILURE);
 948        }
 949        tlscreds = nbd_get_tls_creds(tlscredsid, list, &local_err);
 950        if (local_err) {
 951            error_report("Failed to get TLS creds %s",
 952                         error_get_pretty(local_err));
 953            exit(EXIT_FAILURE);
 954        }
 955    } else {
 956        if (tlsauthz) {
 957            error_report("--tls-authz is not permitted without --tls-creds");
 958            exit(EXIT_FAILURE);
 959        }
 960    }
 961
 962    if (list) {
 963        saddr = nbd_build_socket_address(sockpath, bindto, port);
 964        return qemu_nbd_client_list(saddr, tlscreds, bindto);
 965    }
 966
 967#if !HAVE_NBD_DEVICE
 968    if (disconnect || device) {
 969        error_report("Kernel /dev/nbdN support not available");
 970        exit(EXIT_FAILURE);
 971    }
 972#else /* HAVE_NBD_DEVICE */
 973    if (disconnect) {
 974        int nbdfd = open(argv[optind], O_RDWR);
 975        if (nbdfd < 0) {
 976            error_report("Cannot open %s: %s", argv[optind],
 977                         strerror(errno));
 978            exit(EXIT_FAILURE);
 979        }
 980        nbd_disconnect(nbdfd);
 981
 982        close(nbdfd);
 983
 984        printf("%s disconnected\n", argv[optind]);
 985
 986        return 0;
 987    }
 988#endif
 989
 990    if ((device && !verbose) || fork_process) {
 991        int stderr_fd[2];
 992        pid_t pid;
 993        int ret;
 994
 995        if (qemu_pipe(stderr_fd) < 0) {
 996            error_report("Error setting up communication pipe: %s",
 997                         strerror(errno));
 998            exit(EXIT_FAILURE);
 999        }
1000
1001        /* Now daemonize, but keep a communication channel open to
1002         * print errors and exit with the proper status code.
1003         */
1004        pid = fork();
1005        if (pid < 0) {
1006            error_report("Failed to fork: %s", strerror(errno));
1007            exit(EXIT_FAILURE);
1008        } else if (pid == 0) {
1009            close(stderr_fd[0]);
1010            ret = qemu_daemon(1, 0);
1011
1012            /* Temporarily redirect stderr to the parent's pipe...  */
1013            old_stderr = dup(STDERR_FILENO);
1014            dup2(stderr_fd[1], STDERR_FILENO);
1015            if (ret < 0) {
1016                error_report("Failed to daemonize: %s", strerror(errno));
1017                exit(EXIT_FAILURE);
1018            }
1019
1020            /* ... close the descriptor we inherited and go on.  */
1021            close(stderr_fd[1]);
1022        } else {
1023            bool errors = false;
1024            char *buf;
1025
1026            /* In the parent.  Print error messages from the child until
1027             * it closes the pipe.
1028             */
1029            close(stderr_fd[1]);
1030            buf = g_malloc(1024);
1031            while ((ret = read(stderr_fd[0], buf, 1024)) > 0) {
1032                errors = true;
1033                ret = qemu_write_full(STDERR_FILENO, buf, ret);
1034                if (ret < 0) {
1035                    exit(EXIT_FAILURE);
1036                }
1037            }
1038            if (ret < 0) {
1039                error_report("Cannot read from daemon: %s",
1040                             strerror(errno));
1041                exit(EXIT_FAILURE);
1042            }
1043
1044            /* Usually the daemon should not print any message.
1045             * Exit with zero status in that case.
1046             */
1047            exit(errors);
1048        }
1049    }
1050
1051    if (device != NULL && sockpath == NULL) {
1052        sockpath = g_malloc(128);
1053        snprintf(sockpath, 128, SOCKET_PATH, basename(device));
1054    }
1055
1056    server = qio_net_listener_new();
1057    if (socket_activation == 0) {
1058        saddr = nbd_build_socket_address(sockpath, bindto, port);
1059        if (qio_net_listener_open_sync(server, saddr, &local_err) < 0) {
1060            object_unref(OBJECT(server));
1061            error_report_err(local_err);
1062            exit(EXIT_FAILURE);
1063        }
1064    } else {
1065        size_t i;
1066        /* See comment in check_socket_activation above. */
1067        for (i = 0; i < socket_activation; i++) {
1068            QIOChannelSocket *sioc;
1069            sioc = qio_channel_socket_new_fd(FIRST_SOCKET_ACTIVATION_FD + i,
1070                                             &local_err);
1071            if (sioc == NULL) {
1072                object_unref(OBJECT(server));
1073                error_report("Failed to use socket activation: %s",
1074                             error_get_pretty(local_err));
1075                exit(EXIT_FAILURE);
1076            }
1077            qio_net_listener_add(server, sioc);
1078            object_unref(OBJECT(sioc));
1079        }
1080    }
1081
1082    if (qemu_init_main_loop(&local_err)) {
1083        error_report_err(local_err);
1084        exit(EXIT_FAILURE);
1085    }
1086    bdrv_init();
1087    atexit(qemu_nbd_shutdown);
1088
1089    srcpath = argv[optind];
1090    if (imageOpts) {
1091        QemuOpts *opts;
1092        if (fmt) {
1093            error_report("--image-opts and -f are mutually exclusive");
1094            exit(EXIT_FAILURE);
1095        }
1096        opts = qemu_opts_parse_noisily(&file_opts, srcpath, true);
1097        if (!opts) {
1098            qemu_opts_reset(&file_opts);
1099            exit(EXIT_FAILURE);
1100        }
1101        options = qemu_opts_to_qdict(opts, NULL);
1102        qemu_opts_reset(&file_opts);
1103        blk = blk_new_open(NULL, NULL, options, flags, &local_err);
1104    } else {
1105        if (fmt) {
1106            options = qdict_new();
1107            qdict_put_str(options, "driver", fmt);
1108        }
1109        blk = blk_new_open(srcpath, NULL, options, flags, &local_err);
1110    }
1111
1112    if (!blk) {
1113        error_reportf_err(local_err, "Failed to blk_new_open '%s': ",
1114                          argv[optind]);
1115        exit(EXIT_FAILURE);
1116    }
1117    bs = blk_bs(blk);
1118
1119    blk_set_enable_write_cache(blk, !writethrough);
1120
1121    if (sn_opts) {
1122        ret = bdrv_snapshot_load_tmp(bs,
1123                                     qemu_opt_get(sn_opts, SNAPSHOT_OPT_ID),
1124                                     qemu_opt_get(sn_opts, SNAPSHOT_OPT_NAME),
1125                                     &local_err);
1126    } else if (sn_id_or_name) {
1127        ret = bdrv_snapshot_load_tmp_by_id_or_name(bs, sn_id_or_name,
1128                                                   &local_err);
1129    }
1130    if (ret < 0) {
1131        error_reportf_err(local_err, "Failed to load snapshot: ");
1132        exit(EXIT_FAILURE);
1133    }
1134
1135    bs->detect_zeroes = detect_zeroes;
1136    fd_size = blk_getlength(blk);
1137    if (fd_size < 0) {
1138        error_report("Failed to determine the image length: %s",
1139                     strerror(-fd_size));
1140        exit(EXIT_FAILURE);
1141    }
1142
1143    if (dev_offset >= fd_size) {
1144        error_report("Offset (%" PRIu64 ") has to be smaller than the image "
1145                     "size (%" PRId64 ")", dev_offset, fd_size);
1146        exit(EXIT_FAILURE);
1147    }
1148    fd_size -= dev_offset;
1149
1150    if (partition) {
1151        uint64_t limit;
1152
1153        if (dev_offset) {
1154            error_report("Cannot request partition and offset together");
1155            exit(EXIT_FAILURE);
1156        }
1157        ret = find_partition(blk, partition, &dev_offset, &limit);
1158        if (ret < 0) {
1159            error_report("Could not find partition %d: %s", partition,
1160                         strerror(-ret));
1161            exit(EXIT_FAILURE);
1162        }
1163        /*
1164         * MBR partition limits are (32-bit << 9); this assert lets
1165         * the compiler know that we can't overflow 64 bits.
1166         */
1167        assert(dev_offset + limit >= dev_offset);
1168        if (dev_offset + limit > fd_size) {
1169            error_report("Discovered partition %d at offset %" PRIu64
1170                         " size %" PRIu64 ", but size exceeds file length %"
1171                         PRId64, partition, dev_offset, limit, fd_size);
1172            exit(EXIT_FAILURE);
1173        }
1174        fd_size = limit;
1175    }
1176
1177    export = nbd_export_new(bs, dev_offset, fd_size, export_name,
1178                            export_description, bitmap, nbdflags,
1179                            nbd_export_closed, writethrough, NULL,
1180                            &error_fatal);
1181
1182    if (device) {
1183#if HAVE_NBD_DEVICE
1184        int ret;
1185
1186        ret = pthread_create(&client_thread, NULL, nbd_client_thread, device);
1187        if (ret != 0) {
1188            error_report("Failed to create client thread: %s", strerror(ret));
1189            exit(EXIT_FAILURE);
1190        }
1191#endif
1192    } else {
1193        /* Shut up GCC warnings.  */
1194        memset(&client_thread, 0, sizeof(client_thread));
1195    }
1196
1197    nbd_update_server_watch();
1198
1199    /* now when the initialization is (almost) complete, chdir("/")
1200     * to free any busy filesystems */
1201    if (chdir("/") < 0) {
1202        error_report("Could not chdir to root directory: %s",
1203                     strerror(errno));
1204        exit(EXIT_FAILURE);
1205    }
1206
1207    if (fork_process) {
1208        dup2(old_stderr, STDERR_FILENO);
1209        close(old_stderr);
1210    }
1211
1212    state = RUNNING;
1213    do {
1214        main_loop_wait(false);
1215        if (state == TERMINATE) {
1216            state = TERMINATING;
1217            nbd_export_close(export);
1218            nbd_export_put(export);
1219            export = NULL;
1220        }
1221    } while (state != TERMINATED);
1222
1223    blk_unref(blk);
1224    if (sockpath) {
1225        unlink(sockpath);
1226    }
1227
1228    qemu_opts_del(sn_opts);
1229
1230    if (device) {
1231        void *ret;
1232        pthread_join(client_thread, &ret);
1233        exit(ret != NULL);
1234    } else {
1235        exit(EXIT_SUCCESS);
1236    }
1237}
1238