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 "hw/virtio/vhost-user.h"
  16 #include "chardev/char-fe.h"
  17 #include "qapi/error.h"
  18 #include "qapi/qapi-commands-net.h"
  19 #include "qemu/config-file.h"
  20 #include "qemu/error-report.h"
  21 #include "qemu/option.h"
  22 #include "trace.h"
  23
  24 typedef struct NetVhostUserState {
  25     NetClientState nc;
  26     CharBackend chr; /* only queue index 0 */
  27     VhostUserState *vhost_user;
  28     VHostNetState *vhost_net;
  29     guint watch;
  30     uint64_t acked_features;
  31     bool started;
  32 } NetVhostUserState;
  33
  34 VHostNetState *vhost_user_get_vhost_net(NetClientState *nc)
  35 {
  36     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
  37     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
  38     return s->vhost_net;
  39 }
  40
  41 uint64_t vhost_user_get_acked_features(NetClientState *nc)
  42 {
  43     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
  44     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
  45     return s->acked_features;
  46 }
  47
  48 void vhost_user_save_acked_features(NetClientState *nc)
  49 {
  50     NetVhostUserState *s;
  51
  52     s = DO_UPCAST(NetVhostUserState, nc, nc);
  53     if (s->vhost_net) {
  54         uint64_t features = vhost_net_get_acked_features(s->vhost_net);
  55         if (features) {
  56             s->acked_features = features;
  57         }
  58     }
  59 }
  60
  61 static void vhost_user_stop(int queues, NetClientState *ncs[])
  62 {
  63     int i;
  64     NetVhostUserState *s;
  65
  66     for (i = 0; i < queues; i++) {
  67         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
  68
  69         s = DO_UPCAST(NetVhostUserState, nc, ncs[i]);
  70
  71         if (s->vhost_net) {
  72             vhost_user_save_acked_features(ncs[i]);
  73             vhost_net_cleanup(s->vhost_net);
  74         }
  75     }
  76 }
  77
  78 static int vhost_user_start(int queues, NetClientState *ncs[],
  79                             VhostUserState *be)
  80 {
  81     VhostNetOptions options;
  82     struct vhost_net *net = NULL;
  83     NetVhostUserState *s;
  84     int max_queues;
  85     int i;
  86
  87     options.backend_type = VHOST_BACKEND_TYPE_USER;
  88
  89     for (i = 0; i < queues; i++) {
  90         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
  91
  92         s = DO_UPCAST(NetVhostUserState, nc, ncs[i]);
  93
  94         options.net_backend = ncs[i];
  95         options.opaque      = be;
  96         options.busyloop_timeout = 0;
  97         options.nvqs = 2;
  98         net = vhost_net_init(&options);
  99         if (!net) {
 100             error_report("failed to init vhost_net for queue %d", i);
 101             goto err;
 102         }
 103
 104         if (i == 0) {
 105             max_queues = vhost_net_get_max_queues(net);
 106             if (queues > max_queues) {
 107                 error_report("you are asking more queues than supported: %d",
 108                              max_queues);
 109                 goto err;
 110             }
 111         }
 112
 113         if (s->vhost_net) {
 114             vhost_net_cleanup(s->vhost_net);
 115             g_free(s->vhost_net);
 116         }
 117         s->vhost_net = net;
 118     }
 119
 120     return 0;
 121
 122 err:
 123     if (net) {
 124         vhost_net_cleanup(net);
 125         g_free(net);
 126     }
 127     vhost_user_stop(i, ncs);
 128     return -1;
 129 }
 130
 131 static ssize_t vhost_user_receive(NetClientState *nc, const uint8_t *buf,
 132                                   size_t size)
 133 {
 134     /* In case of RARP (message size is 60) notify backup to send a fake RARP.
 135        This fake RARP will be sent by backend only for guest
 136        without GUEST_ANNOUNCE capability.
 137      */
 138     if (size == 60) {
 139         NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
 140         int r;
 141         static int display_rarp_failure = 1;
 142         char mac_addr[6];
 143
 144         /* extract guest mac address from the RARP message */
 145         memcpy(mac_addr, &buf[6], 6);
 146
 147         r = vhost_net_notify_migration_done(s->vhost_net, mac_addr);
 148
 149         if ((r != 0) && (display_rarp_failure)) {
 150             fprintf(stderr,
 151                     "Vhost user backend fails to broadcast fake RARP\n");
 152             fflush(stderr);
 153             display_rarp_failure = 0;
 154         }
 155     }
 156
 157     return size;
 158 }
 159
 160 static void net_vhost_user_cleanup(NetClientState *nc)
 161 {
 162     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
 163
 164     if (s->vhost_net) {
 165         vhost_net_cleanup(s->vhost_net);
 166         g_free(s->vhost_net);
 167         s->vhost_net = NULL;
 168     }
 169     if (nc->queue_index == 0) {
 170         if (s->watch) {
 171             g_source_remove(s->watch);
 172             s->watch = 0;
 173         }
 174         qemu_chr_fe_deinit(&s->chr, true);
 175         if (s->vhost_user) {
 176             vhost_user_cleanup(s->vhost_user);
 177             g_free(s->vhost_user);
 178             s->vhost_user = NULL;
 179         }
 180     }
 181
 182     qemu_purge_queued_packets(nc);
 183 }
 184
 185 static int vhost_user_set_vnet_endianness(NetClientState *nc,
 186                                           bool enable)
 187 {
 188     /* Nothing to do.  If the server supports
 189      * VHOST_USER_PROTOCOL_F_CROSS_ENDIAN, it will get the
 190      * vnet header endianness from there.  If it doesn't, negotiation
 191      * fails.
 192      */
 193     return 0;
 194 }
 195
 196 static bool vhost_user_has_vnet_hdr(NetClientState *nc)
 197 {
 198     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
 199
 200     return true;
 201 }
 202
 203 static bool vhost_user_has_ufo(NetClientState *nc)
 204 {
 205     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
 206
 207     return true;
 208 }
 209
 210 static bool vhost_user_check_peer_type(NetClientState *nc, ObjectClass *oc,
 211                                        Error **errp)
 212 {
 213     const char *driver = object_class_get_name(oc);
 214
 215     if (!g_str_has_prefix(driver, "virtio-net-")) {
 216         error_setg(errp, "vhost-user requires frontend driver virtio-net-*");
 217         return false;
 218     }
 219
 220     return true;
 221 }
 222
 223 static NetClientInfo net_vhost_user_info = {
 224         .type = NET_CLIENT_DRIVER_VHOST_USER,
 225         .size = sizeof(NetVhostUserState),
 226         .receive = vhost_user_receive,
 227         .cleanup = net_vhost_user_cleanup,
 228         .has_vnet_hdr = vhost_user_has_vnet_hdr,
 229         .has_ufo = vhost_user_has_ufo,
 230         .set_vnet_be = vhost_user_set_vnet_endianness,
 231         .set_vnet_le = vhost_user_set_vnet_endianness,
 232         .check_peer_type = vhost_user_check_peer_type,
 233 };
 234
 235 static gboolean net_vhost_user_watch(void *do_not_use, GIOCondition cond,
 236                                      void *opaque)
 237 {
 238     NetVhostUserState *s = opaque;
 239
 240     qemu_chr_fe_disconnect(&s->chr);
 241
 242     return TRUE;
 243 }
 244
 245 static void net_vhost_user_event(void *opaque, QEMUChrEvent event);
 246
 247 static void chr_closed_bh(void *opaque)
 248 {
 249     const char *name = opaque;
 250     NetClientState *ncs[MAX_QUEUE_NUM];
 251     NetVhostUserState *s;
 252     Error *err = NULL;
 253     int queues, i;
 254
 255     queues = qemu_find_net_clients_except(name, ncs,
 256                                           NET_CLIENT_DRIVER_NIC,
 257                                           MAX_QUEUE_NUM);
 258     assert(queues < MAX_QUEUE_NUM);
 259
 260     s = DO_UPCAST(NetVhostUserState, nc, ncs[0]);
 261
 262     for (i = queues -1; i >= 0; i--) {
 263         vhost_user_save_acked_features(ncs[i]);
 264     }
 265
 266     qmp_set_link(name, false, &err);
 267
 268     qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, net_vhost_user_event,
 269                              NULL, opaque, NULL, true);
 270
 271     if (err) {
 272         error_report_err(err);
 273     }
 274 }
 275
 276 static void net_vhost_user_event(void *opaque, QEMUChrEvent event)
 277 {
 278     const char *name = opaque;
 279     NetClientState *ncs[MAX_QUEUE_NUM];
 280     NetVhostUserState *s;
 281     Chardev *chr;
 282     Error *err = NULL;
 283     int queues;
 284
 285     queues = qemu_find_net_clients_except(name, ncs,
 286                                           NET_CLIENT_DRIVER_NIC,
 287                                           MAX_QUEUE_NUM);
 288     assert(queues < MAX_QUEUE_NUM);
 289
 290     s = DO_UPCAST(NetVhostUserState, nc, ncs[0]);
 291     chr = qemu_chr_fe_get_driver(&s->chr);
 292     trace_vhost_user_event(chr->label, event);
 293     switch (event) {
 294     case CHR_EVENT_OPENED:
 295         if (vhost_user_start(queues, ncs, s->vhost_user) < 0) {
 296             qemu_chr_fe_disconnect(&s->chr);
 297             return;
 298         }
 299         s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP,
 300                                          net_vhost_user_watch, s);
 301         qmp_set_link(name, true, &err);
 302         s->started = true;
 303         break;
 304     case CHR_EVENT_CLOSED:
 305         /* a close event may happen during a read/write, but vhost
 306          * code assumes the vhost_dev remains setup, so delay the
 307          * stop & clear to idle.
 308          * FIXME: better handle failure in vhost code, remove bh
 309          */
 310         if (s->watch) {
 311             AioContext *ctx = qemu_get_current_aio_context();
 312
 313             g_source_remove(s->watch);
 314             s->watch = 0;
 315             qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, NULL, NULL,
 316                                      NULL, NULL, false);
 317
 318             aio_bh_schedule_oneshot(ctx, chr_closed_bh, opaque);
 319         }
 320         break;
 321     case CHR_EVENT_BREAK:
 322     case CHR_EVENT_MUX_IN:
 323     case CHR_EVENT_MUX_OUT:
 324         /* Ignore */
 325         break;
 326     }
 327
 328     if (err) {
 329         error_report_err(err);
 330     }
 331 }
 332
 333 static int net_vhost_user_init(NetClientState *peer, const char *device,
 334                                const char *name, Chardev *chr,
 335                                int queues)
 336 {
 337     Error *err = NULL;
 338     NetClientState *nc, *nc0 = NULL;
 339     NetVhostUserState *s = NULL;
 340     VhostUserState *user;
 341     int i;
 342
 343     assert(name);
 344     assert(queues > 0);
 345
 346     user = g_new0(struct VhostUserState, 1);
 347     for (i = 0; i < queues; i++) {
 348         nc = qemu_new_net_client(&net_vhost_user_info, peer, device, name);
 349         qemu_set_info_str(nc, "vhost-user%d to %s", i, chr->label);
 350         nc->queue_index = i;
 351         if (!nc0) {
 352             nc0 = nc;
 353             s = DO_UPCAST(NetVhostUserState, nc, nc);
 354             if (!qemu_chr_fe_init(&s->chr, chr, &err) ||
 355                 !vhost_user_init(user, &s->chr, &err)) {
 356                 error_report_err(err);
 357                 goto err;
 358             }
 359         }
 360         s = DO_UPCAST(NetVhostUserState, nc, nc);
 361         s->vhost_user = user;
 362     }
 363
 364     s = DO_UPCAST(NetVhostUserState, nc, nc0);
 365     do {
 366         if (qemu_chr_fe_wait_connected(&s->chr, &err) < 0) {
 367             error_report_err(err);
 368             goto err;
 369         }
 370         qemu_chr_fe_set_handlers(&s->chr, NULL, NULL,
 371                                  net_vhost_user_event, NULL, nc0->name, NULL,
 372                                  true);
 373     } while (!s->started);
 374
 375     assert(s->vhost_net);
 376
 377     return 0;
 378
 379 err:
 380     if (user) {
 381         vhost_user_cleanup(user);
 382         g_free(user);
 383         if (s) {
 384             s->vhost_user = NULL;
 385         }
 386     }
 387     if (nc0) {
 388         qemu_del_net_client(nc0);
 389     }
 390
 391     return -1;
 392 }
 393
 394 static Chardev *net_vhost_claim_chardev(
 395     const NetdevVhostUserOptions *opts, Error **errp)
 396 {
 397     Chardev *chr = qemu_chr_find(opts->chardev);
 398
 399     if (chr == NULL) {
 400         error_setg(errp, "chardev \"%s\" not found", opts->chardev);
 401         return NULL;
 402     }
 403
 404     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) {
 405         error_setg(errp, "chardev \"%s\" is not reconnectable",
 406                    opts->chardev);
 407         return NULL;
 408     }
 409     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_FD_PASS)) {
 410         error_setg(errp, "chardev \"%s\" does not support FD passing",
 411                    opts->chardev);
 412         return NULL;
 413     }
 414
 415     return chr;
 416 }
 417
 418 int net_init_vhost_user(const Netdev *netdev, const char *name,
 419                         NetClientState *peer, Error **errp)
 420 {
 421     int queues;
 422     const NetdevVhostUserOptions *vhost_user_opts;
 423     Chardev *chr;
 424
 425     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_USER);
 426     vhost_user_opts = &netdev->u.vhost_user;
 427
 428     chr = net_vhost_claim_chardev(vhost_user_opts, errp);
 429     if (!chr) {
 430         return -1;
 431     }
 432
 433     queues = vhost_user_opts->has_queues ? vhost_user_opts->queues : 1;
 434     if (queues < 1 || queues > MAX_QUEUE_NUM) {
 435         error_setg(errp,
 436                    "vhost-user number of queues must be in range [1, %d]",
 437                    MAX_QUEUE_NUM);
 438         return -1;
 439     }
 440
 441     return net_vhost_user_init(peer, "vhost_user", name, chr, queues);
 442 }