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 "qapi/qapi-events-net.h" 20 #include "qemu/config-file.h" 21 #include "qemu/error-report.h" 22 #include "qemu/option.h" 23 #include "trace.h" 24 25 typedef struct NetVhostUserState { 26 NetClientState nc; 27 CharBackend chr; /* only queue index 0 */ 28 VhostUserState *vhost_user; 29 VHostNetState *vhost_net; 30 guint watch; 31 uint64_t acked_features; 32 bool started; 33 } NetVhostUserState; 34 35 static struct vhost_net *vhost_user_get_vhost_net(NetClientState *nc) 36 { 37 NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc); 38 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER); 39 return s->vhost_net; 40 } 41 42 uint64_t vhost_user_get_acked_features(NetClientState *nc) 43 { 44 NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc); 45 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER); 46 return s->acked_features; 47 } 48 49 void vhost_user_save_acked_features(NetClientState *nc) 50 { 51 NetVhostUserState *s; 52 53 s = DO_UPCAST(NetVhostUserState, nc, nc); 54 if (s->vhost_net) { 55 uint64_t features = vhost_net_get_acked_features(s->vhost_net); 56 if (features) { 57 s->acked_features = features; 58 } 59 } 60 } 61 62 static void vhost_user_stop(int queues, NetClientState *ncs[]) 63 { 64 int i; 65 NetVhostUserState *s; 66 67 for (i = 0; i < queues; i++) { 68 assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER); 69 70 s = DO_UPCAST(NetVhostUserState, nc, ncs[i]); 71 72 if (s->vhost_net) { 73 vhost_user_save_acked_features(ncs[i]); 74 vhost_net_cleanup(s->vhost_net); 75 } 76 } 77 } 78 79 static int vhost_user_start(int queues, NetClientState *ncs[], 80 VhostUserState *be) 81 { 82 VhostNetOptions options; 83 struct vhost_net *net = NULL; 84 NetVhostUserState *s; 85 int max_queues; 86 int i; 87 88 options.backend_type = VHOST_BACKEND_TYPE_USER; 89 90 for (i = 0; i < queues; i++) { 91 assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER); 92 93 s = DO_UPCAST(NetVhostUserState, nc, ncs[i]); 94 95 options.net_backend = ncs[i]; 96 options.opaque = be; 97 options.busyloop_timeout = 0; 98 options.nvqs = 2; 99 net = vhost_net_init(&options); 100 if (!net) { 101 error_report("failed to init vhost_net for queue %d", i); 102 goto err; 103 } 104 105 if (i == 0) { 106 max_queues = vhost_net_get_max_queues(net); 107 if (queues > max_queues) { 108 error_report("you are asking more queues than supported: %d", 109 max_queues); 110 goto err; 111 } 112 } 113 114 if (s->vhost_net) { 115 vhost_net_cleanup(s->vhost_net); 116 g_free(s->vhost_net); 117 } 118 s->vhost_net = net; 119 } 120 121 return 0; 122 123 err: 124 if (net) { 125 vhost_net_cleanup(net); 126 g_free(net); 127 } 128 vhost_user_stop(i, ncs); 129 return -1; 130 } 131 132 static ssize_t vhost_user_receive(NetClientState *nc, const uint8_t *buf, 133 size_t size) 134 { 135 /* In case of RARP (message size is 60) notify backup to send a fake RARP. 136 This fake RARP will be sent by backend only for guest 137 without GUEST_ANNOUNCE capability. 138 */ 139 if (size == 60) { 140 NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc); 141 int r; 142 static int display_rarp_failure = 1; 143 char mac_addr[6]; 144 145 /* extract guest mac address from the RARP message */ 146 memcpy(mac_addr, &buf[6], 6); 147 148 r = vhost_net_notify_migration_done(s->vhost_net, mac_addr); 149 150 if ((r != 0) && (display_rarp_failure)) { 151 fprintf(stderr, 152 "Vhost user backend fails to broadcast fake RARP\n"); 153 fflush(stderr); 154 display_rarp_failure = 0; 155 } 156 } 157 158 return size; 159 } 160 161 static void net_vhost_user_cleanup(NetClientState *nc) 162 { 163 NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc); 164 165 if (s->vhost_net) { 166 vhost_net_cleanup(s->vhost_net); 167 g_free(s->vhost_net); 168 s->vhost_net = NULL; 169 } 170 if (nc->queue_index == 0) { 171 if (s->watch) { 172 g_source_remove(s->watch); 173 s->watch = 0; 174 } 175 qemu_chr_fe_deinit(&s->chr, true); 176 if (s->vhost_user) { 177 vhost_user_cleanup(s->vhost_user); 178 g_free(s->vhost_user); 179 s->vhost_user = NULL; 180 } 181 } 182 183 qemu_purge_queued_packets(nc); 184 } 185 186 static int vhost_user_set_vnet_endianness(NetClientState *nc, 187 bool enable) 188 { 189 /* Nothing to do. If the server supports 190 * VHOST_USER_PROTOCOL_F_CROSS_ENDIAN, it will get the 191 * vnet header endianness from there. If it doesn't, negotiation 192 * fails. 193 */ 194 return 0; 195 } 196 197 static bool vhost_user_has_vnet_hdr(NetClientState *nc) 198 { 199 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER); 200 201 return true; 202 } 203 204 static bool vhost_user_has_ufo(NetClientState *nc) 205 { 206 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER); 207 208 return true; 209 } 210 211 static bool vhost_user_check_peer_type(NetClientState *nc, ObjectClass *oc, 212 Error **errp) 213 { 214 const char *driver = object_class_get_name(oc); 215 216 if (!g_str_has_prefix(driver, "virtio-net-")) { 217 error_setg(errp, "vhost-user requires frontend driver virtio-net-*"); 218 return false; 219 } 220 221 return true; 222 } 223 224 static NetClientInfo net_vhost_user_info = { 225 .type = NET_CLIENT_DRIVER_VHOST_USER, 226 .size = sizeof(NetVhostUserState), 227 .receive = vhost_user_receive, 228 .cleanup = net_vhost_user_cleanup, 229 .has_vnet_hdr = vhost_user_has_vnet_hdr, 230 .has_ufo = vhost_user_has_ufo, 231 .set_vnet_be = vhost_user_set_vnet_endianness, 232 .set_vnet_le = vhost_user_set_vnet_endianness, 233 .check_peer_type = vhost_user_check_peer_type, 234 .get_vhost_net = vhost_user_get_vhost_net, 235 }; 236 237 static gboolean net_vhost_user_watch(void *do_not_use, GIOCondition cond, 238 void *opaque) 239 { 240 NetVhostUserState *s = opaque; 241 242 qemu_chr_fe_disconnect(&s->chr); 243 244 return G_SOURCE_CONTINUE; 245 } 246 247 static void net_vhost_user_event(void *opaque, QEMUChrEvent event); 248 249 static void chr_closed_bh(void *opaque) 250 { 251 const char *name = opaque; 252 NetClientState *ncs[MAX_QUEUE_NUM]; 253 NetVhostUserState *s; 254 Error *err = NULL; 255 int queues, i; 256 257 queues = qemu_find_net_clients_except(name, ncs, 258 NET_CLIENT_DRIVER_NIC, 259 MAX_QUEUE_NUM); 260 assert(queues < MAX_QUEUE_NUM); 261 262 s = DO_UPCAST(NetVhostUserState, nc, ncs[0]); 263 264 for (i = queues -1; i >= 0; i--) { 265 vhost_user_save_acked_features(ncs[i]); 266 } 267 268 net_client_set_link(ncs, queues, false); 269 270 qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, net_vhost_user_event, 271 NULL, opaque, NULL, true); 272 273 if (err) { 274 error_report_err(err); 275 } 276 qapi_event_send_netdev_vhost_user_disconnected(name); 277 } 278 279 static void net_vhost_user_event(void *opaque, QEMUChrEvent event) 280 { 281 const char *name = opaque; 282 NetClientState *ncs[MAX_QUEUE_NUM]; 283 NetVhostUserState *s; 284 Chardev *chr; 285 Error *err = NULL; 286 int queues; 287 288 queues = qemu_find_net_clients_except(name, ncs, 289 NET_CLIENT_DRIVER_NIC, 290 MAX_QUEUE_NUM); 291 assert(queues < MAX_QUEUE_NUM); 292 293 s = DO_UPCAST(NetVhostUserState, nc, ncs[0]); 294 chr = qemu_chr_fe_get_driver(&s->chr); 295 trace_vhost_user_event(chr->label, event); 296 switch (event) { 297 case CHR_EVENT_OPENED: 298 if (vhost_user_start(queues, ncs, s->vhost_user) < 0) { 299 qemu_chr_fe_disconnect(&s->chr); 300 return; 301 } 302 s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP, 303 net_vhost_user_watch, s); 304 net_client_set_link(ncs, queues, true); 305 s->started = true; 306 qapi_event_send_netdev_vhost_user_connected(name, chr->label); 307 break; 308 case CHR_EVENT_CLOSED: 309 /* a close event may happen during a read/write, but vhost 310 * code assumes the vhost_dev remains setup, so delay the 311 * stop & clear to idle. 312 * FIXME: better handle failure in vhost code, remove bh 313 */ 314 if (s->watch) { 315 AioContext *ctx = qemu_get_current_aio_context(); 316 317 g_source_remove(s->watch); 318 s->watch = 0; 319 qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, NULL, NULL, 320 NULL, NULL, false); 321 322 aio_bh_schedule_oneshot(ctx, chr_closed_bh, opaque); 323 } 324 break; 325 case CHR_EVENT_BREAK: 326 case CHR_EVENT_MUX_IN: 327 case CHR_EVENT_MUX_OUT: 328 /* Ignore */ 329 break; 330 } 331 332 if (err) { 333 error_report_err(err); 334 } 335 } 336 337 static int net_vhost_user_init(NetClientState *peer, const char *device, 338 const char *name, Chardev *chr, 339 int queues) 340 { 341 Error *err = NULL; 342 NetClientState *nc, *nc0 = NULL; 343 NetVhostUserState *s = NULL; 344 VhostUserState *user; 345 int i; 346 347 assert(name); 348 assert(queues > 0); 349 350 user = g_new0(struct VhostUserState, 1); 351 for (i = 0; i < queues; i++) { 352 nc = qemu_new_net_client(&net_vhost_user_info, peer, device, name); 353 qemu_set_info_str(nc, "vhost-user%d to %s", i, chr->label); 354 nc->queue_index = i; 355 if (!nc0) { 356 nc0 = nc; 357 s = DO_UPCAST(NetVhostUserState, nc, nc); 358 if (!qemu_chr_fe_init(&s->chr, chr, &err) || 359 !vhost_user_init(user, &s->chr, &err)) { 360 error_report_err(err); 361 goto err; 362 } 363 } 364 s = DO_UPCAST(NetVhostUserState, nc, nc); 365 s->vhost_user = user; 366 } 367 368 s = DO_UPCAST(NetVhostUserState, nc, nc0); 369 do { 370 if (qemu_chr_fe_wait_connected(&s->chr, &err) < 0) { 371 error_report_err(err); 372 goto err; 373 } 374 qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, 375 net_vhost_user_event, NULL, nc0->name, NULL, 376 true); 377 } while (!s->started); 378 379 assert(s->vhost_net); 380 381 return 0; 382 383 err: 384 if (user) { 385 vhost_user_cleanup(user); 386 g_free(user); 387 if (s) { 388 s->vhost_user = NULL; 389 } 390 } 391 if (nc0) { 392 qemu_del_net_client(nc0); 393 } 394 395 return -1; 396 } 397 398 static Chardev *net_vhost_claim_chardev( 399 const NetdevVhostUserOptions *opts, Error **errp) 400 { 401 Chardev *chr = qemu_chr_find(opts->chardev); 402 403 if (chr == NULL) { 404 error_setg(errp, "chardev \"%s\" not found", opts->chardev); 405 return NULL; 406 } 407 408 if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) { 409 error_setg(errp, "chardev \"%s\" is not reconnectable", 410 opts->chardev); 411 return NULL; 412 } 413 if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_FD_PASS)) { 414 error_setg(errp, "chardev \"%s\" does not support FD passing", 415 opts->chardev); 416 return NULL; 417 } 418 419 return chr; 420 } 421 422 int net_init_vhost_user(const Netdev *netdev, const char *name, 423 NetClientState *peer, Error **errp) 424 { 425 int queues; 426 const NetdevVhostUserOptions *vhost_user_opts; 427 Chardev *chr; 428 429 assert(netdev->type == NET_CLIENT_DRIVER_VHOST_USER); 430 vhost_user_opts = &netdev->u.vhost_user; 431 432 chr = net_vhost_claim_chardev(vhost_user_opts, errp); 433 if (!chr) { 434 return -1; 435 } 436 437 queues = vhost_user_opts->has_queues ? vhost_user_opts->queues : 1; 438 if (queues < 1 || queues > MAX_QUEUE_NUM) { 439 error_setg(errp, 440 "vhost-user number of queues must be in range [1, %d]", 441 MAX_QUEUE_NUM); 442 return -1; 443 } 444 445 return net_vhost_user_init(peer, "vhost_user", name, chr, queues); 446 } 447