| 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 "hw/virtio/vhost.h" |
| 15 | #include "hw/virtio/vhost-user.h" |
| 16 | #include "standard-headers/linux/virtio_net.h" |
| 17 | #include "chardev/char-fe.h" |
| 18 | #include "qapi/error.h" |
| 19 | #include "qapi/qapi-commands-net.h" |
| 20 | #include "qapi/qapi-events-net.h" |
| 21 | #include "qemu/config-file.h" |
| 22 | #include "qemu/error-report.h" |
| 23 | #include "qemu/option.h" |
| 24 | #include "trace.h" |
| 25 | |
| 26 | static const int user_feature_bits[] = { |
| 27 | VIRTIO_F_NOTIFY_ON_EMPTY, |
| 28 | VIRTIO_F_NOTIFICATION_DATA, |
| 29 | VIRTIO_RING_F_INDIRECT_DESC, |
| 30 | VIRTIO_RING_F_EVENT_IDX, |
| 31 | |
| 32 | VIRTIO_F_ANY_LAYOUT, |
| 33 | VIRTIO_F_VERSION_1, |
| 34 | VIRTIO_NET_F_CSUM, |
| 35 | VIRTIO_NET_F_GUEST_CSUM, |
| 36 | VIRTIO_NET_F_GSO, |
| 37 | VIRTIO_NET_F_GUEST_TSO4, |
| 38 | VIRTIO_NET_F_GUEST_TSO6, |
| 39 | VIRTIO_NET_F_GUEST_ECN, |
| 40 | VIRTIO_NET_F_GUEST_UFO, |
| 41 | VIRTIO_NET_F_HOST_TSO4, |
| 42 | VIRTIO_NET_F_HOST_TSO6, |
| 43 | VIRTIO_NET_F_HOST_ECN, |
| 44 | VIRTIO_NET_F_HOST_UFO, |
| 45 | VIRTIO_NET_F_MRG_RXBUF, |
| 46 | VIRTIO_NET_F_MTU, |
| 47 | VIRTIO_F_IOMMU_PLATFORM, |
| 48 | VIRTIO_F_RING_PACKED, |
| 49 | VIRTIO_F_RING_RESET, |
| 50 | VIRTIO_F_IN_ORDER, |
| 51 | VIRTIO_NET_F_RSS, |
| 52 | VIRTIO_NET_F_RSC_EXT, |
| 53 | VIRTIO_NET_F_HASH_REPORT, |
| 54 | VIRTIO_NET_F_GUEST_USO4, |
| 55 | VIRTIO_NET_F_GUEST_USO6, |
| 56 | VIRTIO_NET_F_HOST_USO, |
| 57 | |
| 58 | /* This bit implies RARP isn't sent by QEMU out of band */ |
| 59 | VIRTIO_NET_F_GUEST_ANNOUNCE, |
| 60 | |
| 61 | VIRTIO_NET_F_MQ, |
| 62 | |
| 63 | VHOST_INVALID_FEATURE_BIT |
| 64 | }; |
| 65 | |
| 66 | typedef struct NetVhostUserState { |
| 67 | NetClientState nc; |
| 68 | CharFrontend chr; /* only queue index 0 */ |
| 69 | VhostUserState *vhost_user; |
| 70 | VHostNetState *vhost_net; |
| 71 | guint watch; |
| 72 | uint64_t acked_features; |
| 73 | bool started; |
| 74 | } NetVhostUserState; |
| 75 | |
| 76 | static struct vhost_net *vhost_user_get_vhost_net(NetClientState *nc) |
| 77 | { |
| 78 | NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc); |
| 79 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER); |
| 80 | return s->vhost_net; |
| 81 | } |
| 82 | |
| 83 | static uint64_t vhost_user_get_acked_features(NetClientState *nc) |
| 84 | { |
| 85 | NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc); |
| 86 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER); |
| 87 | return s->acked_features; |
| 88 | } |
| 89 | |
| 90 | static void vhost_user_save_acked_features(NetClientState *nc) |
| 91 | { |
| 92 | NetVhostUserState *s; |
| 93 | |
| 94 | s = DO_UPCAST(NetVhostUserState, nc, nc); |
| 95 | if (s->vhost_net) { |
| 96 | uint64_t features = vhost_net_get_acked_features(s->vhost_net); |
| 97 | if (features) { |
| 98 | s->acked_features = features; |
| 99 | } |
| 100 | } |
| 101 | } |
| 102 | |
| 103 | static void vhost_user_stop(int queues, NetClientState *ncs[]) |
| 104 | { |
| 105 | int i; |
| 106 | NetVhostUserState *s; |
| 107 | |
| 108 | for (i = 0; i < queues; i++) { |
| 109 | assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER); |
| 110 | |
| 111 | s = DO_UPCAST(NetVhostUserState, nc, ncs[i]); |
| 112 | |
| 113 | if (s->vhost_net) { |
| 114 | vhost_user_save_acked_features(ncs[i]); |
| 115 | vhost_net_cleanup(s->vhost_net); |
| 116 | } |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | static int vhost_user_start(int queues, NetClientState *ncs[], |
| 121 | VhostUserState *be) |
| 122 | { |
| 123 | VhostNetOptions options; |
| 124 | struct vhost_net *net = NULL; |
| 125 | NetVhostUserState *s; |
| 126 | int max_queues; |
| 127 | int i; |
| 128 | |
| 129 | options.backend_type = VHOST_BACKEND_TYPE_USER; |
| 130 | |
| 131 | for (i = 0; i < queues; i++) { |
| 132 | assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER); |
| 133 | |
| 134 | s = DO_UPCAST(NetVhostUserState, nc, ncs[i]); |
| 135 | |
| 136 | options.net_backend = ncs[i]; |
| 137 | options.opaque = be; |
| 138 | options.busyloop_timeout = 0; |
| 139 | options.nvqs = 2; |
| 140 | options.feature_bits = user_feature_bits; |
| 141 | options.max_tx_queue_size = VIRTQUEUE_MAX_SIZE; |
| 142 | options.get_acked_features = vhost_user_get_acked_features; |
| 143 | options.save_acked_features = vhost_user_save_acked_features; |
| 144 | options.is_vhost_user = true; |
| 145 | |
| 146 | net = vhost_net_init(&options); |
| 147 | if (!net) { |
| 148 | error_report("failed to init vhost_net for queue %d", i); |
| 149 | goto err; |
| 150 | } |
| 151 | |
| 152 | if (i == 0) { |
| 153 | max_queues = vhost_net_get_max_queues(net); |
| 154 | if (queues > max_queues) { |
| 155 | error_report("you are asking more queues than supported: %d", |
| 156 | max_queues); |
| 157 | goto err; |
| 158 | } |
| 159 | } |
| 160 | |
| 161 | if (s->vhost_net) { |
| 162 | vhost_net_cleanup(s->vhost_net); |
| 163 | g_free(s->vhost_net); |
| 164 | } |
| 165 | s->vhost_net = net; |
| 166 | } |
| 167 | |
| 168 | return 0; |
| 169 | |
| 170 | err: |
| 171 | if (net) { |
| 172 | vhost_net_cleanup(net); |
| 173 | g_free(net); |
| 174 | } |
| 175 | vhost_user_stop(i, ncs); |
| 176 | return -1; |
| 177 | } |
| 178 | |
| 179 | static ssize_t vhost_user_receive(NetClientState *nc, const uint8_t *buf, |
| 180 | size_t size) |
| 181 | { |
| 182 | /* In case of RARP (message size is 60) notify backup to send a fake RARP. |
| 183 | This fake RARP will be sent by backend only for guest |
| 184 | without GUEST_ANNOUNCE capability. |
| 185 | */ |
| 186 | if (size == 60) { |
| 187 | NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc); |
| 188 | int r; |
| 189 | static int display_rarp_failure = 1; |
| 190 | char mac_addr[6]; |
| 191 | |
| 192 | /* extract guest mac address from the RARP message */ |
| 193 | memcpy(mac_addr, &buf[6], 6); |
| 194 | |
| 195 | r = vhost_net_notify_migration_done(s->vhost_net, mac_addr); |
| 196 | |
| 197 | if ((r != 0) && (display_rarp_failure)) { |
| 198 | fprintf(stderr, |
| 199 | "Vhost user backend fails to broadcast fake RARP\n"); |
| 200 | fflush(stderr); |
| 201 | display_rarp_failure = 0; |
| 202 | } |
| 203 | } |
| 204 | |
| 205 | return size; |
| 206 | } |
| 207 | |
| 208 | static void net_vhost_user_cleanup(NetClientState *nc) |
| 209 | { |
| 210 | NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc); |
| 211 | |
| 212 | if (s->vhost_net) { |
| 213 | vhost_net_cleanup(s->vhost_net); |
| 214 | g_free(s->vhost_net); |
| 215 | s->vhost_net = NULL; |
| 216 | } |
| 217 | if (nc->queue_index == 0) { |
| 218 | g_clear_handle_id(&s->watch, g_source_remove); |
| 219 | qemu_chr_fe_deinit(&s->chr, true); |
| 220 | if (s->vhost_user) { |
| 221 | vhost_user_cleanup(s->vhost_user); |
| 222 | g_free(s->vhost_user); |
| 223 | s->vhost_user = NULL; |
| 224 | } |
| 225 | } |
| 226 | |
| 227 | qemu_purge_queued_packets(nc); |
| 228 | } |
| 229 | |
| 230 | static int vhost_user_set_vnet_endianness(NetClientState *nc, |
| 231 | bool enable) |
| 232 | { |
| 233 | /* Nothing to do. If the server supports |
| 234 | * VHOST_USER_PROTOCOL_F_CROSS_ENDIAN, it will get the |
| 235 | * vnet header endianness from there. If it doesn't, negotiation |
| 236 | * fails. |
| 237 | */ |
| 238 | return 0; |
| 239 | } |
| 240 | |
| 241 | static bool vhost_user_has_vnet_hdr(NetClientState *nc) |
| 242 | { |
| 243 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER); |
| 244 | |
| 245 | return true; |
| 246 | } |
| 247 | |
| 248 | static bool vhost_user_has_ufo(NetClientState *nc) |
| 249 | { |
| 250 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER); |
| 251 | |
| 252 | return true; |
| 253 | } |
| 254 | |
| 255 | static bool vhost_user_check_peer_type(NetClientState *nc, ObjectClass *oc, |
| 256 | Error **errp) |
| 257 | { |
| 258 | const char *driver = object_class_get_name(oc); |
| 259 | |
| 260 | if (!g_str_has_prefix(driver, "virtio-net-")) { |
| 261 | error_setg(errp, "vhost-user requires frontend driver virtio-net-*"); |
| 262 | return false; |
| 263 | } |
| 264 | |
| 265 | return true; |
| 266 | } |
| 267 | |
| 268 | static NetClientInfo net_vhost_user_info = { |
| 269 | .type = NET_CLIENT_DRIVER_VHOST_USER, |
| 270 | .size = sizeof(NetVhostUserState), |
| 271 | .receive = vhost_user_receive, |
| 272 | .cleanup = net_vhost_user_cleanup, |
| 273 | .has_vnet_hdr = vhost_user_has_vnet_hdr, |
| 274 | .has_ufo = vhost_user_has_ufo, |
| 275 | .set_vnet_be = vhost_user_set_vnet_endianness, |
| 276 | .set_vnet_le = vhost_user_set_vnet_endianness, |
| 277 | .check_peer_type = vhost_user_check_peer_type, |
| 278 | .get_vhost_net = vhost_user_get_vhost_net, |
| 279 | }; |
| 280 | |
| 281 | static gboolean net_vhost_user_watch(void *do_not_use, GIOCondition cond, |
| 282 | void *opaque) |
| 283 | { |
| 284 | NetVhostUserState *s = opaque; |
| 285 | |
| 286 | qemu_chr_fe_disconnect(&s->chr); |
| 287 | |
| 288 | return G_SOURCE_CONTINUE; |
| 289 | } |
| 290 | |
| 291 | static void net_vhost_user_event(void *opaque, QEMUChrEvent event); |
| 292 | |
| 293 | static void chr_closed_bh(void *opaque) |
| 294 | { |
| 295 | const char *name = opaque; |
| 296 | NetClientState *ncs[MAX_QUEUE_NUM]; |
| 297 | NetVhostUserState *s; |
| 298 | int queues, i; |
| 299 | |
| 300 | queues = qemu_find_net_clients_except(name, ncs, |
| 301 | NET_CLIENT_DRIVER_NIC, |
| 302 | MAX_QUEUE_NUM); |
| 303 | assert(queues < MAX_QUEUE_NUM); |
| 304 | |
| 305 | s = DO_UPCAST(NetVhostUserState, nc, ncs[0]); |
| 306 | |
| 307 | for (i = queues -1; i >= 0; i--) { |
| 308 | vhost_user_save_acked_features(ncs[i]); |
| 309 | } |
| 310 | |
| 311 | net_client_set_link(ncs, queues, false); |
| 312 | |
| 313 | qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, net_vhost_user_event, |
| 314 | NULL, opaque, NULL, true); |
| 315 | |
| 316 | qapi_event_send_netdev_vhost_user_disconnected(name); |
| 317 | } |
| 318 | |
| 319 | static void net_vhost_user_event(void *opaque, QEMUChrEvent event) |
| 320 | { |
| 321 | const char *name = opaque; |
| 322 | NetClientState *ncs[MAX_QUEUE_NUM]; |
| 323 | NetVhostUserState *s; |
| 324 | Chardev *chr; |
| 325 | int queues; |
| 326 | |
| 327 | queues = qemu_find_net_clients_except(name, ncs, |
| 328 | NET_CLIENT_DRIVER_NIC, |
| 329 | MAX_QUEUE_NUM); |
| 330 | assert(queues < MAX_QUEUE_NUM); |
| 331 | |
| 332 | s = DO_UPCAST(NetVhostUserState, nc, ncs[0]); |
| 333 | chr = qemu_chr_fe_get_driver(&s->chr); |
| 334 | trace_vhost_user_event(chr->label, event); |
| 335 | switch (event) { |
| 336 | case CHR_EVENT_OPENED: |
| 337 | if (vhost_user_start(queues, ncs, s->vhost_user) < 0) { |
| 338 | qemu_chr_fe_disconnect(&s->chr); |
| 339 | return; |
| 340 | } |
| 341 | s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP, |
| 342 | net_vhost_user_watch, s); |
| 343 | net_client_set_link(ncs, queues, true); |
| 344 | s->started = true; |
| 345 | qapi_event_send_netdev_vhost_user_connected(name, chr->label); |
| 346 | break; |
| 347 | case CHR_EVENT_CLOSED: |
| 348 | /* a close event may happen during a read/write, but vhost |
| 349 | * code assumes the vhost_dev remains setup, so delay the |
| 350 | * stop & clear to idle. |
| 351 | * FIXME: better handle failure in vhost code, remove bh |
| 352 | */ |
| 353 | if (s->watch) { |
| 354 | AioContext *ctx = qemu_get_current_aio_context(); |
| 355 | |
| 356 | g_clear_handle_id(&s->watch, g_source_remove); |
| 357 | qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, NULL, NULL, |
| 358 | NULL, NULL, false); |
| 359 | |
| 360 | aio_bh_schedule_oneshot(ctx, chr_closed_bh, opaque); |
| 361 | } |
| 362 | break; |
| 363 | case CHR_EVENT_BREAK: |
| 364 | case CHR_EVENT_MUX_IN: |
| 365 | case CHR_EVENT_MUX_OUT: |
| 366 | /* Ignore */ |
| 367 | break; |
| 368 | } |
| 369 | } |
| 370 | |
| 371 | static int net_vhost_user_init(NetClientState *peer, const char *device, |
| 372 | const char *name, Chardev *chr, |
| 373 | int queues) |
| 374 | { |
| 375 | Error *err = NULL; |
| 376 | NetClientState *nc, *nc0 = NULL; |
| 377 | NetVhostUserState *s = NULL; |
| 378 | VhostUserState *user; |
| 379 | int i; |
| 380 | |
| 381 | assert(name); |
| 382 | assert(queues > 0); |
| 383 | |
| 384 | user = g_new0(struct VhostUserState, 1); |
| 385 | for (i = 0; i < queues; i++) { |
| 386 | nc = qemu_new_net_client(&net_vhost_user_info, peer, device, name); |
| 387 | qemu_set_info_str(nc, "vhost-user%d to %s", i, chr->label); |
| 388 | nc->queue_index = i; |
| 389 | if (!nc0) { |
| 390 | nc0 = nc; |
| 391 | s = DO_UPCAST(NetVhostUserState, nc, nc); |
| 392 | if (!qemu_chr_fe_init(&s->chr, chr, &err) || |
| 393 | !vhost_user_init(user, &s->chr, &err)) { |
| 394 | error_report_err(err); |
| 395 | goto err; |
| 396 | } |
| 397 | } |
| 398 | s = DO_UPCAST(NetVhostUserState, nc, nc); |
| 399 | s->vhost_user = user; |
| 400 | } |
| 401 | |
| 402 | s = DO_UPCAST(NetVhostUserState, nc, nc0); |
| 403 | do { |
| 404 | if (qemu_chr_fe_wait_connected(&s->chr, &err) < 0) { |
| 405 | error_report_err(err); |
| 406 | goto err; |
| 407 | } |
| 408 | qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, |
| 409 | net_vhost_user_event, NULL, nc0->name, NULL, |
| 410 | true); |
| 411 | } while (!s->started); |
| 412 | |
| 413 | assert(s->vhost_net); |
| 414 | |
| 415 | return 0; |
| 416 | |
| 417 | err: |
| 418 | if (user) { |
| 419 | vhost_user_cleanup(user); |
| 420 | g_free(user); |
| 421 | if (s) { |
| 422 | s->vhost_user = NULL; |
| 423 | } |
| 424 | } |
| 425 | if (nc0) { |
| 426 | qemu_del_net_client(nc0); |
| 427 | } |
| 428 | |
| 429 | return -1; |
| 430 | } |
| 431 | |
| 432 | static Chardev *net_vhost_claim_chardev( |
| 433 | const NetdevVhostUserOptions *opts, Error **errp) |
| 434 | { |
| 435 | Chardev *chr = qemu_chr_find(opts->chardev); |
| 436 | |
| 437 | if (chr == NULL) { |
| 438 | error_setg(errp, "chardev \"%s\" not found", opts->chardev); |
| 439 | return NULL; |
| 440 | } |
| 441 | |
| 442 | if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) { |
| 443 | error_setg(errp, "chardev \"%s\" is not reconnectable", |
| 444 | opts->chardev); |
| 445 | return NULL; |
| 446 | } |
| 447 | if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_FD_PASS)) { |
| 448 | error_setg(errp, "chardev \"%s\" does not support FD passing", |
| 449 | opts->chardev); |
| 450 | return NULL; |
| 451 | } |
| 452 | |
| 453 | return chr; |
| 454 | } |
| 455 | |
| 456 | int net_init_vhost_user(const Netdev *netdev, const char *name, |
| 457 | NetClientState *peer, Error **errp) |
| 458 | { |
| 459 | int queues; |
| 460 | const NetdevVhostUserOptions *vhost_user_opts; |
| 461 | Chardev *chr; |
| 462 | |
| 463 | assert(netdev->type == NET_CLIENT_DRIVER_VHOST_USER); |
| 464 | vhost_user_opts = &netdev->u.vhost_user; |
| 465 | |
| 466 | chr = net_vhost_claim_chardev(vhost_user_opts, errp); |
| 467 | if (!chr) { |
| 468 | return -1; |
| 469 | } |
| 470 | |
| 471 | queues = vhost_user_opts->has_queues ? vhost_user_opts->queues : 1; |
| 472 | if (queues < 1 || queues > MAX_QUEUE_NUM) { |
| 473 | error_setg(errp, |
| 474 | "vhost-user number of queues must be in range [1, %d]", |
| 475 | MAX_QUEUE_NUM); |
| 476 | return -1; |
| 477 | } |
| 478 | |
| 479 | return net_vhost_user_init(peer, "vhost_user", name, chr, queues); |
| 480 | } |