net/vhost-user.c

   1 /*
   2  * vhost-user.c
   3  *
   4  * Copyright (c) 2013 Virtual Open Systems Sarl.
   5  *
   6  * This work is licensed under the terms of the GNU GPL, version 2 or later.
   7  * See the COPYING file in the top-level directory.
   8  *
   9  */
  10
  11 #include "qemu/osdep.h"
  12 #include "clients.h"
  13 #include "net/vhost_net.h"
  14 #include "net/vhost-user.h"
  15 #include "chardev/char-fe.h"
  16 #include "qapi/error.h"
  17 #include "qemu/config-file.h"
  18 #include "qemu/error-report.h"
  19 #include "qemu/option.h"
  20 #include "qmp-commands.h"
  21 #include "trace.h"
  22
  23 typedef struct VhostUserState {
  24     NetClientState nc;
  25     CharBackend chr; /* only queue index 0 */
  26     VHostNetState *vhost_net;
  27     guint watch;
  28     uint64_t acked_features;
  29     bool started;
  30 } VhostUserState;
  31
  32 VHostNetState *vhost_user_get_vhost_net(NetClientState *nc)
  33 {
  34     VhostUserState *s = DO_UPCAST(VhostUserState, nc, nc);
  35     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
  36     return s->vhost_net;
  37 }
  38
  39 uint64_t vhost_user_get_acked_features(NetClientState *nc)
  40 {
  41     VhostUserState *s = DO_UPCAST(VhostUserState, nc, nc);
  42     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
  43     return s->acked_features;
  44 }
  45
  46 static void vhost_user_stop(int queues, NetClientState *ncs[])
  47 {
  48     VhostUserState *s;
  49     int i;
  50
  51     for (i = 0; i < queues; i++) {
  52         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
  53
  54         s = DO_UPCAST(VhostUserState, nc, ncs[i]);
  55
  56         if (s->vhost_net) {
  57             /* save acked features */
  58             uint64_t features = vhost_net_get_acked_features(s->vhost_net);
  59             if (features) {
  60                 s->acked_features = features;
  61             }
  62             vhost_net_cleanup(s->vhost_net);
  63         }
  64     }
  65 }
  66
  67 static int vhost_user_start(int queues, NetClientState *ncs[], CharBackend *be)
  68 {
  69     VhostNetOptions options;
  70     struct vhost_net *net = NULL;
  71     VhostUserState *s;
  72     int max_queues;
  73     int i;
  74
  75     options.backend_type = VHOST_BACKEND_TYPE_USER;
  76
  77     for (i = 0; i < queues; i++) {
  78         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
  79
  80         s = DO_UPCAST(VhostUserState, nc, ncs[i]);
  81
  82         options.net_backend = ncs[i];
  83         options.opaque      = be;
  84         options.busyloop_timeout = 0;
  85         net = vhost_net_init(&options);
  86         if (!net) {
  87             error_report("failed to init vhost_net for queue %d", i);
  88             goto err;
  89         }
  90
  91         if (i == 0) {
  92             max_queues = vhost_net_get_max_queues(net);
  93             if (queues > max_queues) {
  94                 error_report("you are asking more queues than supported: %d",
  95                              max_queues);
  96                 goto err;
  97             }
  98         }
  99
 100         if (s->vhost_net) {
 101             vhost_net_cleanup(s->vhost_net);
 102             g_free(s->vhost_net);
 103         }
 104         s->vhost_net = net;
 105     }
 106
 107     return 0;
 108
 109 err:
 110     if (net) {
 111         vhost_net_cleanup(net);
 112         g_free(net);
 113     }
 114     vhost_user_stop(i, ncs);
 115     return -1;
 116 }
 117
 118 static ssize_t vhost_user_receive(NetClientState *nc, const uint8_t *buf,
 119                                   size_t size)
 120 {
 121     /* In case of RARP (message size is 60) notify backup to send a fake RARP.
 122        This fake RARP will be sent by backend only for guest
 123        without GUEST_ANNOUNCE capability.
 124      */
 125     if (size == 60) {
 126         VhostUserState *s = DO_UPCAST(VhostUserState, nc, nc);
 127         int r;
 128         static int display_rarp_failure = 1;
 129         char mac_addr[6];
 130
 131         /* extract guest mac address from the RARP message */
 132         memcpy(mac_addr, &buf[6], 6);
 133
 134         r = vhost_net_notify_migration_done(s->vhost_net, mac_addr);
 135
 136         if ((r != 0) && (display_rarp_failure)) {
 137             fprintf(stderr,
 138                     "Vhost user backend fails to broadcast fake RARP\n");
 139             fflush(stderr);
 140             display_rarp_failure = 0;
 141         }
 142     }
 143
 144     return size;
 145 }
 146
 147 static void vhost_user_cleanup(NetClientState *nc)
 148 {
 149     VhostUserState *s = DO_UPCAST(VhostUserState, nc, nc);
 150
 151     if (s->vhost_net) {
 152         vhost_net_cleanup(s->vhost_net);
 153         g_free(s->vhost_net);
 154         s->vhost_net = NULL;
 155     }
 156     if (nc->queue_index == 0) {
 157         if (s->watch) {
 158             g_source_remove(s->watch);
 159             s->watch = 0;
 160         }
 161         qemu_chr_fe_deinit(&s->chr, true);
 162     }
 163
 164     qemu_purge_queued_packets(nc);
 165 }
 166
 167 static bool vhost_user_has_vnet_hdr(NetClientState *nc)
 168 {
 169     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
 170
 171     return true;
 172 }
 173
 174 static bool vhost_user_has_ufo(NetClientState *nc)
 175 {
 176     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
 177
 178     return true;
 179 }
 180
 181 static NetClientInfo net_vhost_user_info = {
 182         .type = NET_CLIENT_DRIVER_VHOST_USER,
 183         .size = sizeof(VhostUserState),
 184         .receive = vhost_user_receive,
 185         .cleanup = vhost_user_cleanup,
 186         .has_vnet_hdr = vhost_user_has_vnet_hdr,
 187         .has_ufo = vhost_user_has_ufo,
 188 };
 189
 190 static gboolean net_vhost_user_watch(GIOChannel *chan, GIOCondition cond,
 191                                            void *opaque)
 192 {
 193     VhostUserState *s = opaque;
 194
 195     qemu_chr_fe_disconnect(&s->chr);
 196
 197     return TRUE;
 198 }
 199
 200 static void net_vhost_user_event(void *opaque, int event);
 201
 202 static void chr_closed_bh(void *opaque)
 203 {
 204     const char *name = opaque;
 205     NetClientState *ncs[MAX_QUEUE_NUM];
 206     VhostUserState *s;
 207     Error *err = NULL;
 208     int queues;
 209
 210     queues = qemu_find_net_clients_except(name, ncs,
 211                                           NET_CLIENT_DRIVER_NIC,
 212                                           MAX_QUEUE_NUM);
 213     assert(queues < MAX_QUEUE_NUM);
 214
 215     s = DO_UPCAST(VhostUserState, nc, ncs[0]);
 216
 217     qmp_set_link(name, false, &err);
 218     vhost_user_stop(queues, ncs);
 219
 220     qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, net_vhost_user_event,
 221                              NULL, opaque, NULL, true);
 222
 223     if (err) {
 224         error_report_err(err);
 225     }
 226 }
 227
 228 static void net_vhost_user_event(void *opaque, int event)
 229 {
 230     const char *name = opaque;
 231     NetClientState *ncs[MAX_QUEUE_NUM];
 232     VhostUserState *s;
 233     Chardev *chr;
 234     Error *err = NULL;
 235     int queues;
 236
 237     queues = qemu_find_net_clients_except(name, ncs,
 238                                           NET_CLIENT_DRIVER_NIC,
 239                                           MAX_QUEUE_NUM);
 240     assert(queues < MAX_QUEUE_NUM);
 241
 242     s = DO_UPCAST(VhostUserState, nc, ncs[0]);
 243     chr = qemu_chr_fe_get_driver(&s->chr);
 244     trace_vhost_user_event(chr->label, event);
 245     switch (event) {
 246     case CHR_EVENT_OPENED:
 247         if (vhost_user_start(queues, ncs, &s->chr) < 0) {
 248             qemu_chr_fe_disconnect(&s->chr);
 249             return;
 250         }
 251         s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP,
 252                                          net_vhost_user_watch, s);
 253         qmp_set_link(name, true, &err);
 254         s->started = true;
 255         break;
 256     case CHR_EVENT_CLOSED:
 257         /* a close event may happen during a read/write, but vhost
 258          * code assumes the vhost_dev remains setup, so delay the
 259          * stop & clear to idle.
 260          * FIXME: better handle failure in vhost code, remove bh
 261          */
 262         if (s->watch) {
 263             AioContext *ctx = qemu_get_current_aio_context();
 264
 265             g_source_remove(s->watch);
 266             s->watch = 0;
 267             qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, NULL, NULL,
 268                                      NULL, NULL, false);
 269
 270             aio_bh_schedule_oneshot(ctx, chr_closed_bh, opaque);
 271         }
 272         break;
 273     }
 274
 275     if (err) {
 276         error_report_err(err);
 277     }
 278 }
 279
 280 static int net_vhost_user_init(NetClientState *peer, const char *device,
 281                                const char *name, Chardev *chr,
 282                                int queues)
 283 {
 284     Error *err = NULL;
 285     NetClientState *nc, *nc0 = NULL;
 286     VhostUserState *s;
 287     int i;
 288
 289     assert(name);
 290     assert(queues > 0);
 291
 292     for (i = 0; i < queues; i++) {
 293         nc = qemu_new_net_client(&net_vhost_user_info, peer, device, name);
 294         snprintf(nc->info_str, sizeof(nc->info_str), "vhost-user%d to %s",
 295                  i, chr->label);
 296         nc->queue_index = i;
 297         if (!nc0) {
 298             nc0 = nc;
 299             s = DO_UPCAST(VhostUserState, nc, nc);
 300             if (!qemu_chr_fe_init(&s->chr, chr, &err)) {
 301                 error_report_err(err);
 302                 return -1;
 303             }
 304         }
 305
 306     }
 307
 308     s = DO_UPCAST(VhostUserState, nc, nc0);
 309     do {
 310         if (qemu_chr_fe_wait_connected(&s->chr, &err) < 0) {
 311             error_report_err(err);
 312             return -1;
 313         }
 314         qemu_chr_fe_set_handlers(&s->chr, NULL, NULL,
 315                                  net_vhost_user_event, NULL, nc0->name, NULL,
 316                                  true);
 317     } while (!s->started);
 318
 319     assert(s->vhost_net);
 320
 321     return 0;
 322 }
 323
 324 static Chardev *net_vhost_claim_chardev(
 325     const NetdevVhostUserOptions *opts, Error **errp)
 326 {
 327     Chardev *chr = qemu_chr_find(opts->chardev);
 328
 329     if (chr == NULL) {
 330         error_setg(errp, "chardev \"%s\" not found", opts->chardev);
 331         return NULL;
 332     }
 333
 334     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) {
 335         error_setg(errp, "chardev \"%s\" is not reconnectable",
 336                    opts->chardev);
 337         return NULL;
 338     }
 339     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_FD_PASS)) {
 340         error_setg(errp, "chardev \"%s\" does not support FD passing",
 341                    opts->chardev);
 342         return NULL;
 343     }
 344
 345     return chr;
 346 }
 347
 348 static int net_vhost_check_net(void *opaque, QemuOpts *opts, Error **errp)
 349 {
 350     const char *name = opaque;
 351     const char *driver, *netdev;
 352
 353     driver = qemu_opt_get(opts, "driver");
 354     netdev = qemu_opt_get(opts, "netdev");
 355
 356     if (!driver || !netdev) {
 357         return 0;
 358     }
 359
 360     if (strcmp(netdev, name) == 0 &&
 361         !g_str_has_prefix(driver, "virtio-net-")) {
 362         error_setg(errp, "vhost-user requires frontend driver virtio-net-*");
 363         return -1;
 364     }
 365
 366     return 0;
 367 }
 368
 369 int net_init_vhost_user(const Netdev *netdev, const char *name,
 370                         NetClientState *peer, Error **errp)
 371 {
 372     int queues;
 373     const NetdevVhostUserOptions *vhost_user_opts;
 374     Chardev *chr;
 375
 376     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_USER);
 377     vhost_user_opts = &netdev->u.vhost_user;
 378
 379     chr = net_vhost_claim_chardev(vhost_user_opts, errp);
 380     if (!chr) {
 381         return -1;
 382     }
 383
 384     /* verify net frontend */
 385     if (qemu_opts_foreach(qemu_find_opts("device"), net_vhost_check_net,
 386                           (char *)name, errp)) {
 387         return -1;
 388     }
 389
 390     queues = vhost_user_opts->has_queues ? vhost_user_opts->queues : 1;
 391     if (queues < 1 || queues > MAX_QUEUE_NUM) {
 392         error_setg(errp,
 393                    "vhost-user number of queues must be in range [1, %d]",
 394                    MAX_QUEUE_NUM);
 395         return -1;
 396     }
 397
 398     return net_vhost_user_init(peer, "vhost_user", name, chr, queues);
 399 }