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 "qemu-common.h"
  25#include "qapi/error.h"
  26#include "qemu/cutils.h"
  27#include "sysemu/block-backend.h"
  28#include "block/block_int.h"
  29#include "block/nbd.h"
  30#include "qemu/main-loop.h"
  31#include "qemu/module.h"
  32#include "qemu/option.h"
  33#include "qemu/error-report.h"
  34#include "qemu/config-file.h"
  35#include "qemu/bswap.h"
  36#include "qemu/log.h"
  37#include "qemu/systemd.h"
  38#include "block/snapshot.h"
  39#include "qapi/qmp/qdict.h"
  40#include "qapi/qmp/qstring.h"
  41#include "qom/object_interfaces.h"
  42#include "io/channel-socket.h"
  43#include "io/net-listener.h"
  44#include "crypto/init.h"
  45#include "trace/control.h"
  46#include "qemu-version.h"
  47
  48#ifdef __linux__
  49#define HAVE_NBD_DEVICE 1
  50#else
  51#define HAVE_NBD_DEVICE 0
  52#endif
  53
  54#define SOCKET_PATH                "/var/lock/qemu-nbd-%s"
  55#define QEMU_NBD_OPT_CACHE         256
  56#define QEMU_NBD_OPT_AIO           257
  57#define QEMU_NBD_OPT_DISCARD       258
  58#define QEMU_NBD_OPT_DETECT_ZEROES 259
  59#define QEMU_NBD_OPT_OBJECT        260
  60#define QEMU_NBD_OPT_TLSCREDS      261
  61#define QEMU_NBD_OPT_IMAGE_OPTS    262
  62#define QEMU_NBD_OPT_FORK          263
  63#define QEMU_NBD_OPT_TLSAUTHZ      264
  64#define QEMU_NBD_OPT_PID_FILE      265
  65
  66#define MBR_SIZE 512
  67
  68static NBDExport *export;
  69static int verbose;
  70static char *srcpath;
  71static SocketAddress *saddr;
  72static int persistent = 0;
  73static enum { RUNNING, TERMINATE, TERMINATING, TERMINATED } state;
  74static int shared = 1;
  75static int nb_fds;
  76static QIONetListener *server;
  77static QCryptoTLSCreds *tlscreds;
  78static const char *tlsauthz;
  79
  80static void usage(const char *name)
  81{
  82    (printf) (
  83"Usage: %s [OPTIONS] FILE\n"
  84"  or:  %s -L [OPTIONS]\n"
  85"QEMU Disk Network Block Device Utility\n"
  86"\n"
  87"  -h, --help                display this help and exit\n"
  88"  -V, --version             output version information and exit\n"
  89"\n"
  90"Connection properties:\n"
  91"  -p, --port=PORT           port to listen on (default `%d')\n"
  92"  -b, --bind=IFACE          interface to bind to (default `0.0.0.0')\n"
  93"  -k, --socket=PATH         path to the unix socket\n"
  94"                            (default '"SOCKET_PATH"')\n"
  95"  -e, --shared=NUM          device can be shared by NUM clients (default '1')\n"
  96"  -t, --persistent          don't exit on the last connection\n"
  97"  -v, --verbose             display extra debugging information\n"
  98"  -x, --export-name=NAME    expose export by name (default is empty string)\n"
  99"  -D, --description=TEXT    export a human-readable description\n"
 100"\n"
 101"Exposing part of the image:\n"
 102"  -o, --offset=OFFSET       offset into the image\n"
 103"  -P, --partition=NUM       only expose partition NUM\n"
 104"  -B, --bitmap=NAME         expose a persistent dirty bitmap\n"
 105"\n"
 106"General purpose options:\n"
 107"  -L, --list                list exports available from another NBD server\n"
 108"  --object type,id=ID,...   define an object such as 'secret' for providing\n"
 109"                            passwords and/or encryption keys\n"
 110"  --tls-creds=ID            use id of an earlier --object to provide TLS\n"
 111"  --tls-authz=ID            use id of an earlier --object to provide\n"
 112"                            authorization\n"
 113"  -T, --trace [[enable=]<pattern>][,events=<file>][,file=<file>]\n"
 114"                            specify tracing options\n"
 115"  --fork                    fork off the server process and exit the parent\n"
 116"                            once the server is running\n"
 117"  --pid-file=PATH           store the server's process ID in the given file\n"
 118#if HAVE_NBD_DEVICE
 119"\n"
 120"Kernel NBD client support:\n"
 121"  -c, --connect=DEV         connect FILE to the local NBD device DEV\n"
 122"  -d, --disconnect          disconnect the specified device\n"
 123#endif
 124"\n"
 125"Block device options:\n"
 126"  -f, --format=FORMAT       set image format (raw, qcow2, ...)\n"
 127"  -r, --read-only           export read-only\n"
 128"  -s, --snapshot            use FILE as an external snapshot, create a temporary\n"
 129"                            file with backing_file=FILE, redirect the write to\n"
 130"                            the temporary one\n"
 131"  -l, --load-snapshot=SNAPSHOT_PARAM\n"
 132"                            load an internal snapshot inside FILE and export it\n"
 133"                            as an read-only device, SNAPSHOT_PARAM format is\n"
 134"                            'snapshot.id=[ID],snapshot.name=[NAME]', or\n"
 135"                            '[ID_OR_NAME]'\n"
 136"  -n, --nocache             disable host cache\n"
 137"      --cache=MODE          set cache mode (none, writeback, ...)\n"
 138"      --aio=MODE            set AIO mode (native or threads)\n"
 139"      --discard=MODE        set discard mode (ignore, unmap)\n"
 140"      --detect-zeroes=MODE  set detect-zeroes mode (off, on, unmap)\n"
 141"      --image-opts          treat FILE as a full set of image options\n"
 142"\n"
 143QEMU_HELP_BOTTOM "\n"
 144    , name, name, NBD_DEFAULT_PORT, "DEVICE");
 145}
 146
 147static void version(const char *name)
 148{
 149    printf(
 150"%s " QEMU_FULL_VERSION "\n"
 151"Written by Anthony Liguori.\n"
 152"\n"
 153QEMU_COPYRIGHT "\n"
 154"This is free software; see the source for copying conditions.  There is NO\n"
 155"warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n"
 156    , name);
 157}
 158
 159struct partition_record
 160{
 161    uint8_t bootable;
 162    uint8_t start_head;
 163    uint32_t start_cylinder;
 164    uint8_t start_sector;
 165    uint8_t system;
 166    uint8_t end_head;
 167    uint8_t end_cylinder;
 168    uint8_t end_sector;
 169    uint32_t start_sector_abs;
 170    uint32_t nb_sectors_abs;
 171};
 172
 173static void read_partition(uint8_t *p, struct partition_record *r)
 174{
 175    r->bootable = p[0];
 176    r->start_head = p[1];
 177    r->start_cylinder = p[3] | ((p[2] << 2) & 0x0300);
 178    r->start_sector = p[2] & 0x3f;
 179    r->system = p[4];
 180    r->end_head = p[5];
 181    r->end_cylinder = p[7] | ((p[6] << 2) & 0x300);
 182    r->end_sector = p[6] & 0x3f;
 183
 184    r->start_sector_abs = ldl_le_p(p + 8);
 185    r->nb_sectors_abs   = ldl_le_p(p + 12);
 186}
 187
 188static int find_partition(BlockBackend *blk, int partition,
 189                          uint64_t *offset, uint64_t *size)
 190{
 191    struct partition_record mbr[4];
 192    uint8_t data[MBR_SIZE];
 193    int i;
 194    int ext_partnum = 4;
 195    int ret;
 196
 197    ret = blk_pread(blk, 0, data, sizeof(data));
 198    if (ret < 0) {
 199        error_report("error while reading: %s", strerror(-ret));
 200        exit(EXIT_FAILURE);
 201    }
 202
 203    if (data[510] != 0x55 || data[511] != 0xaa) {
 204        return -EINVAL;
 205    }
 206
 207    for (i = 0; i < 4; i++) {
 208        read_partition(&data[446 + 16 * i], &mbr[i]);
 209
 210        if (!mbr[i].system || !mbr[i].nb_sectors_abs) {
 211            continue;
 212        }
 213
 214        if (mbr[i].system == 0xF || mbr[i].system == 0x5) {
 215            struct partition_record ext[4];
 216            uint8_t data1[MBR_SIZE];
 217            int j;
 218
 219            ret = blk_pread(blk, mbr[i].start_sector_abs * MBR_SIZE,
 220                            data1, sizeof(data1));
 221            if (ret < 0) {
 222                error_report("error while reading: %s", strerror(-ret));
 223                exit(EXIT_FAILURE);
 224            }
 225
 226            for (j = 0; j < 4; j++) {
 227                read_partition(&data1[446 + 16 * j], &ext[j]);
 228                if (!ext[j].system || !ext[j].nb_sectors_abs) {
 229                    continue;
 230                }
 231
 232                if ((ext_partnum + j + 1) == partition) {
 233                    *offset = (uint64_t)ext[j].start_sector_abs << 9;
 234                    *size = (uint64_t)ext[j].nb_sectors_abs << 9;
 235                    return 0;
 236                }
 237            }
 238            ext_partnum += 4;
 239        } else if ((i + 1) == partition) {
 240            *offset = (uint64_t)mbr[i].start_sector_abs << 9;
 241            *size = (uint64_t)mbr[i].nb_sectors_abs << 9;
 242            return 0;
 243        }
 244    }
 245
 246    return -ENOENT;
 247}
 248
 249static void termsig_handler(int signum)
 250{
 251    atomic_cmpxchg(&state, RUNNING, TERMINATE);
 252    qemu_notify_event();
 253}
 254
 255
 256static int qemu_nbd_client_list(SocketAddress *saddr, QCryptoTLSCreds *tls,
 257                                const char *hostname)
 258{
 259    int ret = EXIT_FAILURE;
 260    int rc;
 261    Error *err = NULL;
 262    QIOChannelSocket *sioc;
 263    NBDExportInfo *list;
 264    int i, j;
 265
 266    sioc = qio_channel_socket_new();
 267    if (qio_channel_socket_connect_sync(sioc, saddr, &err) < 0) {
 268        error_report_err(err);
 269        return EXIT_FAILURE;
 270    }
 271    rc = nbd_receive_export_list(QIO_CHANNEL(sioc), tls, hostname, &list,
 272                                 &err);
 273    if (rc < 0) {
 274        if (err) {
 275            error_report_err(err);
 276        }
 277        goto out;
 278    }
 279    printf("exports available: %d\n", rc);
 280    for (i = 0; i < rc; i++) {
 281        printf(" export: '%s'\n", list[i].name);
 282        if (list[i].description && *list[i].description) {
 283            printf("  description: %s\n", list[i].description);
 284        }
 285        if (list[i].flags & NBD_FLAG_HAS_FLAGS) {
 286            static const char *const flag_names[] = {
 287                [NBD_FLAG_READ_ONLY_BIT]            = "readonly",
 288                [NBD_FLAG_SEND_FLUSH_BIT]           = "flush",
 289                [NBD_FLAG_SEND_FUA_BIT]             = "fua",
 290                [NBD_FLAG_ROTATIONAL_BIT]           = "rotational",
 291                [NBD_FLAG_SEND_TRIM_BIT]            = "trim",
 292                [NBD_FLAG_SEND_WRITE_ZEROES_BIT]    = "zeroes",
 293                [NBD_FLAG_SEND_DF_BIT]              = "df",
 294                [NBD_FLAG_CAN_MULTI_CONN_BIT]       = "multi",
 295                [NBD_FLAG_SEND_RESIZE_BIT]          = "resize",
 296                [NBD_FLAG_SEND_CACHE_BIT]           = "cache",
 297                [NBD_FLAG_SEND_FAST_ZERO_BIT]       = "fast-zero",
 298            };
 299
 300            printf("  size:  %" PRIu64 "\n", list[i].size);
 301            printf("  flags: 0x%x (", list[i].flags);
 302            for (size_t bit = 0; bit < ARRAY_SIZE(flag_names); bit++) {
 303                if (flag_names[bit] && (list[i].flags & (1 << bit))) {
 304                    printf(" %s", flag_names[bit]);
 305                }
 306            }
 307            printf(" )\n");
 308        }
 309        if (list[i].min_block) {
 310            printf("  min block: %u\n", list[i].min_block);
 311            printf("  opt block: %u\n", list[i].opt_block);
 312            printf("  max block: %u\n", list[i].max_block);
 313        }
 314        if (list[i].n_contexts) {
 315            printf("  available meta contexts: %d\n", list[i].n_contexts);
 316            for (j = 0; j < list[i].n_contexts; j++) {
 317                printf("   %s\n", list[i].contexts[j]);
 318            }
 319        }
 320    }
 321    nbd_free_export_list(list, rc);
 322
 323    ret = EXIT_SUCCESS;
 324 out:
 325    object_unref(OBJECT(sioc));
 326    return ret;
 327}
 328
 329
 330#if HAVE_NBD_DEVICE
 331static void *show_parts(void *arg)
 332{
 333    char *device = arg;
 334    int nbd;
 335
 336    /* linux just needs an open() to trigger
 337     * the partition table update
 338     * but remember to load the module with max_part != 0 :
 339     *     modprobe nbd max_part=63
 340     */
 341    nbd = open(device, O_RDWR);
 342    if (nbd >= 0) {
 343        close(nbd);
 344    }
 345    return NULL;
 346}
 347
 348static void *nbd_client_thread(void *arg)
 349{
 350    char *device = arg;
 351    NBDExportInfo info = { .request_sizes = false, .name = g_strdup("") };
 352    QIOChannelSocket *sioc;
 353    int fd;
 354    int ret;
 355    pthread_t show_parts_thread;
 356    Error *local_error = NULL;
 357
 358    sioc = qio_channel_socket_new();
 359    if (qio_channel_socket_connect_sync(sioc,
 360                                        saddr,
 361                                        &local_error) < 0) {
 362        error_report_err(local_error);
 363        goto out;
 364    }
 365
 366    ret = nbd_receive_negotiate(NULL, QIO_CHANNEL(sioc),
 367                                NULL, NULL, NULL, &info, &local_error);
 368    if (ret < 0) {
 369        if (local_error) {
 370            error_report_err(local_error);
 371        }
 372        goto out_socket;
 373    }
 374
 375    fd = open(device, O_RDWR);
 376    if (fd < 0) {
 377        /* Linux-only, we can use %m in printf.  */
 378        error_report("Failed to open %s: %m", device);
 379        goto out_socket;
 380    }
 381
 382    ret = nbd_init(fd, sioc, &info, &local_error);
 383    if (ret < 0) {
 384        error_report_err(local_error);
 385        goto out_fd;
 386    }
 387
 388    /* update partition table */
 389    pthread_create(&show_parts_thread, NULL, show_parts, device);
 390
 391    if (verbose) {
 392        fprintf(stderr, "NBD device %s is now connected to %s\n",
 393                device, srcpath);
 394    } else {
 395        /* Close stderr so that the qemu-nbd process exits.  */
 396        dup2(STDOUT_FILENO, STDERR_FILENO);
 397    }
 398
 399    ret = nbd_client(fd);
 400    if (ret) {
 401        goto out_fd;
 402    }
 403    close(fd);
 404    object_unref(OBJECT(sioc));
 405    g_free(info.name);
 406    kill(getpid(), SIGTERM);
 407    return (void *) EXIT_SUCCESS;
 408
 409out_fd:
 410    close(fd);
 411out_socket:
 412    object_unref(OBJECT(sioc));
 413out:
 414    g_free(info.name);
 415    kill(getpid(), SIGTERM);
 416    return (void *) EXIT_FAILURE;
 417}
 418#endif /* HAVE_NBD_DEVICE */
 419
 420static int nbd_can_accept(void)
 421{
 422    return state == RUNNING && nb_fds < shared;
 423}
 424
 425static void nbd_export_closed(NBDExport *export)
 426{
 427    assert(state == TERMINATING);
 428    state = TERMINATED;
 429}
 430
 431static void nbd_update_server_watch(void);
 432
 433static void nbd_client_closed(NBDClient *client, bool negotiated)
 434{
 435    nb_fds--;
 436    if (negotiated && nb_fds == 0 && !persistent && state == RUNNING) {
 437        state = TERMINATE;
 438    }
 439    nbd_update_server_watch();
 440    nbd_client_put(client);
 441}
 442
 443static void nbd_accept(QIONetListener *listener, QIOChannelSocket *cioc,
 444                       gpointer opaque)
 445{
 446    if (state >= TERMINATE) {
 447        return;
 448    }
 449
 450    nb_fds++;
 451    nbd_update_server_watch();
 452    nbd_client_new(cioc, tlscreds, tlsauthz, nbd_client_closed);
 453}
 454
 455static void nbd_update_server_watch(void)
 456{
 457    if (nbd_can_accept()) {
 458        qio_net_listener_set_client_func(server, nbd_accept, NULL, NULL);
 459    } else {
 460        qio_net_listener_set_client_func(server, NULL, NULL, NULL);
 461    }
 462}
 463
 464
 465static SocketAddress *nbd_build_socket_address(const char *sockpath,
 466                                               const char *bindto,
 467                                               const char *port)
 468{
 469    SocketAddress *saddr;
 470
 471    saddr = g_new0(SocketAddress, 1);
 472    if (sockpath) {
 473        saddr->type = SOCKET_ADDRESS_TYPE_UNIX;
 474        saddr->u.q_unix.path = g_strdup(sockpath);
 475    } else {
 476        InetSocketAddress *inet;
 477        saddr->type = SOCKET_ADDRESS_TYPE_INET;
 478        inet = &saddr->u.inet;
 479        inet->host = g_strdup(bindto);
 480        if (port) {
 481            inet->port = g_strdup(port);
 482        } else  {
 483            inet->port = g_strdup_printf("%d", NBD_DEFAULT_PORT);
 484        }
 485    }
 486
 487    return saddr;
 488}
 489
 490
 491static QemuOptsList file_opts = {
 492    .name = "file",
 493    .implied_opt_name = "file",
 494    .head = QTAILQ_HEAD_INITIALIZER(file_opts.head),
 495    .desc = {
 496        /* no elements => accept any params */
 497        { /* end of list */ }
 498    },
 499};
 500
 501static QemuOptsList qemu_object_opts = {
 502    .name = "object",
 503    .implied_opt_name = "qom-type",
 504    .head = QTAILQ_HEAD_INITIALIZER(qemu_object_opts.head),
 505    .desc = {
 506        { }
 507    },
 508};
 509
 510static bool qemu_nbd_object_print_help(const char *type, QemuOpts *opts)
 511{
 512    if (user_creatable_print_help(type, opts)) {
 513        exit(0);
 514    }
 515    return true;
 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    bool readonly = false;
 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        { "pid-file", required_argument, NULL, QEMU_NBD_OPT_PID_FILE },
 655        { NULL, 0, NULL, 0 }
 656    };
 657    int ch;
 658    int opt_ind = 0;
 659    int flags = BDRV_O_RDWR;
 660    int partition = 0;
 661    int ret = 0;
 662    bool seen_cache = false;
 663    bool seen_discard = false;
 664    bool seen_aio = false;
 665    pthread_t client_thread;
 666    const char *fmt = NULL;
 667    Error *local_err = NULL;
 668    BlockdevDetectZeroesOptions detect_zeroes = BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF;
 669    QDict *options = NULL;
 670    const char *export_name = NULL; /* defaults to "" later for server mode */
 671    const char *export_description = NULL;
 672    const char *bitmap = NULL;
 673    const char *tlscredsid = NULL;
 674    bool imageOpts = false;
 675    bool writethrough = true;
 676    char *trace_file = NULL;
 677    bool fork_process = false;
 678    bool list = false;
 679    int old_stderr = -1;
 680    unsigned socket_activation;
 681    const char *pid_file_name = NULL;
 682
 683    /* The client thread uses SIGTERM to interrupt the server.  A signal
 684     * handler ensures that "qemu-nbd -v -c" exits with a nice status code.
 685     */
 686    struct sigaction sa_sigterm;
 687    memset(&sa_sigterm, 0, sizeof(sa_sigterm));
 688    sa_sigterm.sa_handler = termsig_handler;
 689    sigaction(SIGTERM, &sa_sigterm, NULL);
 690
 691#ifdef CONFIG_POSIX
 692    signal(SIGPIPE, SIG_IGN);
 693#endif
 694
 695    error_init(argv[0]);
 696    module_call_init(MODULE_INIT_TRACE);
 697    qcrypto_init(&error_fatal);
 698
 699    module_call_init(MODULE_INIT_QOM);
 700    qemu_add_opts(&qemu_object_opts);
 701    qemu_add_opts(&qemu_trace_opts);
 702    qemu_init_exec_dir(argv[0]);
 703
 704    while ((ch = getopt_long(argc, argv, sopt, lopt, &opt_ind)) != -1) {
 705        switch (ch) {
 706        case 's':
 707            flags |= BDRV_O_SNAPSHOT;
 708            break;
 709        case 'n':
 710            optarg = (char *) "none";
 711            /* fallthrough */
 712        case QEMU_NBD_OPT_CACHE:
 713            if (seen_cache) {
 714                error_report("-n and --cache can only be specified once");
 715                exit(EXIT_FAILURE);
 716            }
 717            seen_cache = true;
 718            if (bdrv_parse_cache_mode(optarg, &flags, &writethrough) == -1) {
 719                error_report("Invalid cache mode `%s'", optarg);
 720                exit(EXIT_FAILURE);
 721            }
 722            break;
 723        case QEMU_NBD_OPT_AIO:
 724            if (seen_aio) {
 725                error_report("--aio can only be specified once");
 726                exit(EXIT_FAILURE);
 727            }
 728            seen_aio = true;
 729            if (!strcmp(optarg, "native")) {
 730                flags |= BDRV_O_NATIVE_AIO;
 731            } else if (!strcmp(optarg, "threads")) {
 732                /* this is the default */
 733            } else {
 734               error_report("invalid aio mode `%s'", optarg);
 735               exit(EXIT_FAILURE);
 736            }
 737            break;
 738        case QEMU_NBD_OPT_DISCARD:
 739            if (seen_discard) {
 740                error_report("--discard can only be specified once");
 741                exit(EXIT_FAILURE);
 742            }
 743            seen_discard = true;
 744            if (bdrv_parse_discard_flags(optarg, &flags) == -1) {
 745                error_report("Invalid discard mode `%s'", optarg);
 746                exit(EXIT_FAILURE);
 747            }
 748            break;
 749        case QEMU_NBD_OPT_DETECT_ZEROES:
 750            detect_zeroes =
 751                qapi_enum_parse(&BlockdevDetectZeroesOptions_lookup,
 752                                optarg,
 753                                BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF,
 754                                &local_err);
 755            if (local_err) {
 756                error_reportf_err(local_err,
 757                                  "Failed to parse detect_zeroes mode: ");
 758                exit(EXIT_FAILURE);
 759            }
 760            if (detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP &&
 761                !(flags & BDRV_O_UNMAP)) {
 762                error_report("setting detect-zeroes to unmap is not allowed "
 763                             "without setting discard operation to unmap");
 764                exit(EXIT_FAILURE);
 765            }
 766            break;
 767        case 'b':
 768            bindto = optarg;
 769            break;
 770        case 'p':
 771            port = optarg;
 772            break;
 773        case 'o':
 774            if (qemu_strtou64(optarg, NULL, 0, &dev_offset) < 0) {
 775                error_report("Invalid offset '%s'", optarg);
 776                exit(EXIT_FAILURE);
 777            }
 778            break;
 779        case 'l':
 780            if (strstart(optarg, SNAPSHOT_OPT_BASE, NULL)) {
 781                sn_opts = qemu_opts_parse_noisily(&internal_snapshot_opts,
 782                                                  optarg, false);
 783                if (!sn_opts) {
 784                    error_report("Failed in parsing snapshot param `%s'",
 785                                 optarg);
 786                    exit(EXIT_FAILURE);
 787                }
 788            } else {
 789                sn_id_or_name = optarg;
 790            }
 791            /* fall through */
 792        case 'r':
 793            readonly = true;
 794            flags &= ~BDRV_O_RDWR;
 795            break;
 796        case 'P':
 797            warn_report("The '-P' option is deprecated; use --image-opts with "
 798                        "a raw device wrapper for subset exports instead");
 799            if (qemu_strtoi(optarg, NULL, 0, &partition) < 0 ||
 800                partition < 1 || partition > 8) {
 801                error_report("Invalid partition '%s'", optarg);
 802                exit(EXIT_FAILURE);
 803            }
 804            break;
 805        case 'B':
 806            bitmap = optarg;
 807            break;
 808        case 'k':
 809            sockpath = optarg;
 810            if (sockpath[0] != '/') {
 811                error_report("socket path must be absolute");
 812                exit(EXIT_FAILURE);
 813            }
 814            break;
 815        case 'd':
 816            disconnect = true;
 817            break;
 818        case 'c':
 819            device = optarg;
 820            break;
 821        case 'e':
 822            if (qemu_strtoi(optarg, NULL, 0, &shared) < 0 ||
 823                shared < 1) {
 824                error_report("Invalid shared device number '%s'", optarg);
 825                exit(EXIT_FAILURE);
 826            }
 827            break;
 828        case 'f':
 829            fmt = optarg;
 830            break;
 831        case 't':
 832            persistent = 1;
 833            break;
 834        case 'x':
 835            export_name = optarg;
 836            if (strlen(export_name) > NBD_MAX_STRING_SIZE) {
 837                error_report("export name '%s' too long", export_name);
 838                exit(EXIT_FAILURE);
 839            }
 840            break;
 841        case 'D':
 842            export_description = optarg;
 843            if (strlen(export_description) > NBD_MAX_STRING_SIZE) {
 844                error_report("export description '%s' too long",
 845                             export_description);
 846                exit(EXIT_FAILURE);
 847            }
 848            break;
 849        case 'v':
 850            verbose = 1;
 851            break;
 852        case 'V':
 853            version(argv[0]);
 854            exit(0);
 855            break;
 856        case 'h':
 857            usage(argv[0]);
 858            exit(0);
 859            break;
 860        case '?':
 861            error_report("Try `%s --help' for more information.", argv[0]);
 862            exit(EXIT_FAILURE);
 863        case QEMU_NBD_OPT_OBJECT: {
 864            QemuOpts *opts;
 865            opts = qemu_opts_parse_noisily(&qemu_object_opts,
 866                                           optarg, true);
 867            if (!opts) {
 868                exit(EXIT_FAILURE);
 869            }
 870        }   break;
 871        case QEMU_NBD_OPT_TLSCREDS:
 872            tlscredsid = optarg;
 873            break;
 874        case QEMU_NBD_OPT_IMAGE_OPTS:
 875            imageOpts = true;
 876            break;
 877        case 'T':
 878            g_free(trace_file);
 879            trace_file = trace_opt_parse(optarg);
 880            break;
 881        case QEMU_NBD_OPT_TLSAUTHZ:
 882            tlsauthz = optarg;
 883            break;
 884        case QEMU_NBD_OPT_FORK:
 885            fork_process = true;
 886            break;
 887        case 'L':
 888            list = true;
 889            break;
 890        case QEMU_NBD_OPT_PID_FILE:
 891            pid_file_name = optarg;
 892            break;
 893        }
 894    }
 895
 896    if (list) {
 897        if (argc != optind) {
 898            error_report("List mode is incompatible with a file name");
 899            exit(EXIT_FAILURE);
 900        }
 901        if (export_name || export_description || dev_offset || partition ||
 902            device || disconnect || fmt || sn_id_or_name || bitmap ||
 903            seen_aio || seen_discard || seen_cache) {
 904            error_report("List mode is incompatible with per-device settings");
 905            exit(EXIT_FAILURE);
 906        }
 907        if (fork_process) {
 908            error_report("List mode is incompatible with forking");
 909            exit(EXIT_FAILURE);
 910        }
 911    } else if ((argc - optind) != 1) {
 912        error_report("Invalid number of arguments");
 913        error_printf("Try `%s --help' for more information.\n", argv[0]);
 914        exit(EXIT_FAILURE);
 915    } else if (!export_name) {
 916        export_name = "";
 917    }
 918
 919    qemu_opts_foreach(&qemu_object_opts,
 920                      user_creatable_add_opts_foreach,
 921                      qemu_nbd_object_print_help, &error_fatal);
 922
 923    if (!trace_init_backends()) {
 924        exit(1);
 925    }
 926    trace_init_file(trace_file);
 927    qemu_set_log(LOG_TRACE);
 928
 929    socket_activation = check_socket_activation();
 930    if (socket_activation == 0) {
 931        setup_address_and_port(&bindto, &port);
 932    } else {
 933        /* Using socket activation - check user didn't use -p etc. */
 934        const char *err_msg = socket_activation_validate_opts(device, sockpath,
 935                                                              bindto, port,
 936                                                              list);
 937        if (err_msg != NULL) {
 938            error_report("%s", err_msg);
 939            exit(EXIT_FAILURE);
 940        }
 941
 942        /* qemu-nbd can only listen on a single socket.  */
 943        if (socket_activation > 1) {
 944            error_report("qemu-nbd does not support socket activation with %s > 1",
 945                         "LISTEN_FDS");
 946            exit(EXIT_FAILURE);
 947        }
 948    }
 949
 950    if (tlscredsid) {
 951        if (sockpath) {
 952            error_report("TLS is only supported with IPv4/IPv6");
 953            exit(EXIT_FAILURE);
 954        }
 955        if (device) {
 956            error_report("TLS is not supported with a host device");
 957            exit(EXIT_FAILURE);
 958        }
 959        if (tlsauthz && list) {
 960            error_report("TLS authorization is incompatible with export list");
 961            exit(EXIT_FAILURE);
 962        }
 963        tlscreds = nbd_get_tls_creds(tlscredsid, list, &local_err);
 964        if (local_err) {
 965            error_report("Failed to get TLS creds %s",
 966                         error_get_pretty(local_err));
 967            exit(EXIT_FAILURE);
 968        }
 969    } else {
 970        if (tlsauthz) {
 971            error_report("--tls-authz is not permitted without --tls-creds");
 972            exit(EXIT_FAILURE);
 973        }
 974    }
 975
 976    if (list) {
 977        saddr = nbd_build_socket_address(sockpath, bindto, port);
 978        return qemu_nbd_client_list(saddr, tlscreds, bindto);
 979    }
 980
 981#if !HAVE_NBD_DEVICE
 982    if (disconnect || device) {
 983        error_report("Kernel /dev/nbdN support not available");
 984        exit(EXIT_FAILURE);
 985    }
 986#else /* HAVE_NBD_DEVICE */
 987    if (disconnect) {
 988        int nbdfd = open(argv[optind], O_RDWR);
 989        if (nbdfd < 0) {
 990            error_report("Cannot open %s: %s", argv[optind],
 991                         strerror(errno));
 992            exit(EXIT_FAILURE);
 993        }
 994        nbd_disconnect(nbdfd);
 995
 996        close(nbdfd);
 997
 998        printf("%s disconnected\n", argv[optind]);
 999
1000        return 0;
1001    }
1002#endif
1003
1004    if ((device && !verbose) || fork_process) {
1005        int stderr_fd[2];
1006        pid_t pid;
1007        int ret;
1008
1009        if (qemu_pipe(stderr_fd) < 0) {
1010            error_report("Error setting up communication pipe: %s",
1011                         strerror(errno));
1012            exit(EXIT_FAILURE);
1013        }
1014
1015        /* Now daemonize, but keep a communication channel open to
1016         * print errors and exit with the proper status code.
1017         */
1018        pid = fork();
1019        if (pid < 0) {
1020            error_report("Failed to fork: %s", strerror(errno));
1021            exit(EXIT_FAILURE);
1022        } else if (pid == 0) {
1023            close(stderr_fd[0]);
1024
1025            old_stderr = dup(STDERR_FILENO);
1026            ret = qemu_daemon(1, 0);
1027
1028            /* Temporarily redirect stderr to the parent's pipe...  */
1029            dup2(stderr_fd[1], STDERR_FILENO);
1030            if (ret < 0) {
1031                error_report("Failed to daemonize: %s", strerror(errno));
1032                exit(EXIT_FAILURE);
1033            }
1034
1035            /* ... close the descriptor we inherited and go on.  */
1036            close(stderr_fd[1]);
1037        } else {
1038            bool errors = false;
1039            char *buf;
1040
1041            /* In the parent.  Print error messages from the child until
1042             * it closes the pipe.
1043             */
1044            close(stderr_fd[1]);
1045            buf = g_malloc(1024);
1046            while ((ret = read(stderr_fd[0], buf, 1024)) > 0) {
1047                errors = true;
1048                ret = qemu_write_full(STDERR_FILENO, buf, ret);
1049                if (ret < 0) {
1050                    exit(EXIT_FAILURE);
1051                }
1052            }
1053            if (ret < 0) {
1054                error_report("Cannot read from daemon: %s",
1055                             strerror(errno));
1056                exit(EXIT_FAILURE);
1057            }
1058
1059            /* Usually the daemon should not print any message.
1060             * Exit with zero status in that case.
1061             */
1062            exit(errors);
1063        }
1064    }
1065
1066    if (device != NULL && sockpath == NULL) {
1067        sockpath = g_malloc(128);
1068        snprintf(sockpath, 128, SOCKET_PATH, basename(device));
1069    }
1070
1071    server = qio_net_listener_new();
1072    if (socket_activation == 0) {
1073        saddr = nbd_build_socket_address(sockpath, bindto, port);
1074        if (qio_net_listener_open_sync(server, saddr, 1, &local_err) < 0) {
1075            object_unref(OBJECT(server));
1076            error_report_err(local_err);
1077            exit(EXIT_FAILURE);
1078        }
1079    } else {
1080        size_t i;
1081        /* See comment in check_socket_activation above. */
1082        for (i = 0; i < socket_activation; i++) {
1083            QIOChannelSocket *sioc;
1084            sioc = qio_channel_socket_new_fd(FIRST_SOCKET_ACTIVATION_FD + i,
1085                                             &local_err);
1086            if (sioc == NULL) {
1087                object_unref(OBJECT(server));
1088                error_report("Failed to use socket activation: %s",
1089                             error_get_pretty(local_err));
1090                exit(EXIT_FAILURE);
1091            }
1092            qio_net_listener_add(server, sioc);
1093            object_unref(OBJECT(sioc));
1094        }
1095    }
1096
1097    if (qemu_init_main_loop(&local_err)) {
1098        error_report_err(local_err);
1099        exit(EXIT_FAILURE);
1100    }
1101    bdrv_init();
1102    atexit(qemu_nbd_shutdown);
1103
1104    srcpath = argv[optind];
1105    if (imageOpts) {
1106        QemuOpts *opts;
1107        if (fmt) {
1108            error_report("--image-opts and -f are mutually exclusive");
1109            exit(EXIT_FAILURE);
1110        }
1111        opts = qemu_opts_parse_noisily(&file_opts, srcpath, true);
1112        if (!opts) {
1113            qemu_opts_reset(&file_opts);
1114            exit(EXIT_FAILURE);
1115        }
1116        options = qemu_opts_to_qdict(opts, NULL);
1117        qemu_opts_reset(&file_opts);
1118        blk = blk_new_open(NULL, NULL, options, flags, &local_err);
1119    } else {
1120        if (fmt) {
1121            options = qdict_new();
1122            qdict_put_str(options, "driver", fmt);
1123        }
1124        blk = blk_new_open(srcpath, NULL, options, flags, &local_err);
1125    }
1126
1127    if (!blk) {
1128        error_reportf_err(local_err, "Failed to blk_new_open '%s': ",
1129                          argv[optind]);
1130        exit(EXIT_FAILURE);
1131    }
1132    bs = blk_bs(blk);
1133
1134    blk_set_enable_write_cache(blk, !writethrough);
1135
1136    if (sn_opts) {
1137        ret = bdrv_snapshot_load_tmp(bs,
1138                                     qemu_opt_get(sn_opts, SNAPSHOT_OPT_ID),
1139                                     qemu_opt_get(sn_opts, SNAPSHOT_OPT_NAME),
1140                                     &local_err);
1141    } else if (sn_id_or_name) {
1142        ret = bdrv_snapshot_load_tmp_by_id_or_name(bs, sn_id_or_name,
1143                                                   &local_err);
1144    }
1145    if (ret < 0) {
1146        error_reportf_err(local_err, "Failed to load snapshot: ");
1147        exit(EXIT_FAILURE);
1148    }
1149
1150    bs->detect_zeroes = detect_zeroes;
1151    fd_size = blk_getlength(blk);
1152    if (fd_size < 0) {
1153        error_report("Failed to determine the image length: %s",
1154                     strerror(-fd_size));
1155        exit(EXIT_FAILURE);
1156    }
1157
1158    if (dev_offset >= fd_size) {
1159        error_report("Offset (%" PRIu64 ") has to be smaller than the image "
1160                     "size (%" PRId64 ")", dev_offset, fd_size);
1161        exit(EXIT_FAILURE);
1162    }
1163    fd_size -= dev_offset;
1164
1165    if (partition) {
1166        uint64_t limit;
1167
1168        if (dev_offset) {
1169            error_report("Cannot request partition and offset together");
1170            exit(EXIT_FAILURE);
1171        }
1172        ret = find_partition(blk, partition, &dev_offset, &limit);
1173        if (ret < 0) {
1174            error_report("Could not find partition %d: %s", partition,
1175                         strerror(-ret));
1176            exit(EXIT_FAILURE);
1177        }
1178        /*
1179         * MBR partition limits are (32-bit << 9); this assert lets
1180         * the compiler know that we can't overflow 64 bits.
1181         */
1182        assert(dev_offset + limit >= dev_offset);
1183        if (dev_offset + limit > fd_size) {
1184            error_report("Discovered partition %d at offset %" PRIu64
1185                         " size %" PRIu64 ", but size exceeds file length %"
1186                         PRId64, partition, dev_offset, limit, fd_size);
1187            exit(EXIT_FAILURE);
1188        }
1189        fd_size = limit;
1190    }
1191
1192    export = nbd_export_new(bs, dev_offset, fd_size, export_name,
1193                            export_description, bitmap, readonly, shared > 1,
1194                            nbd_export_closed, writethrough, NULL,
1195                            &error_fatal);
1196
1197    if (device) {
1198#if HAVE_NBD_DEVICE
1199        int ret;
1200
1201        ret = pthread_create(&client_thread, NULL, nbd_client_thread, device);
1202        if (ret != 0) {
1203            error_report("Failed to create client thread: %s", strerror(ret));
1204            exit(EXIT_FAILURE);
1205        }
1206#endif
1207    } else {
1208        /* Shut up GCC warnings.  */
1209        memset(&client_thread, 0, sizeof(client_thread));
1210    }
1211
1212    nbd_update_server_watch();
1213
1214    if (pid_file_name) {
1215        qemu_write_pidfile(pid_file_name, &error_fatal);
1216    }
1217
1218    /* now when the initialization is (almost) complete, chdir("/")
1219     * to free any busy filesystems */
1220    if (chdir("/") < 0) {
1221        error_report("Could not chdir to root directory: %s",
1222                     strerror(errno));
1223        exit(EXIT_FAILURE);
1224    }
1225
1226    if (fork_process) {
1227        dup2(old_stderr, STDERR_FILENO);
1228        close(old_stderr);
1229    }
1230
1231    state = RUNNING;
1232    do {
1233        main_loop_wait(false);
1234        if (state == TERMINATE) {
1235            state = TERMINATING;
1236            nbd_export_close(export);
1237            nbd_export_put(export);
1238            export = NULL;
1239        }
1240    } while (state != TERMINATED);
1241
1242    blk_unref(blk);
1243    if (sockpath) {
1244        unlink(sockpath);
1245    }
1246
1247    qemu_opts_del(sn_opts);
1248
1249    if (device) {
1250        void *ret;
1251        pthread_join(client_thread, &ret);
1252        exit(ret != NULL);
1253    } else {
1254        exit(EXIT_SUCCESS);
1255    }
1256}
1257