1 /* 2 * Sharing QEMU devices via vhost-user protocol 3 * 4 * Copyright (c) Coiby Xu <coiby.xu@gmail.com>. 5 * Copyright (c) 2020 Red Hat, Inc. 6 * 7 * This work is licensed under the terms of the GNU GPL, version 2 or 8 * later. See the COPYING file in the top-level directory. 9 */ 10 #include "qemu/osdep.h" 11 #include "qemu/error-report.h" 12 #include "qemu/main-loop.h" 13 #include "qemu/vhost-user-server.h" 14 #include "block/aio-wait.h" 15 16 /* 17 * Theory of operation: 18 * 19 * VuServer is started and stopped by vhost_user_server_start() and 20 * vhost_user_server_stop() from the main loop thread. Starting the server 21 * opens a vhost-user UNIX domain socket and listens for incoming connections. 22 * Only one connection is allowed at a time. 23 * 24 * The connection is handled by the vu_client_trip() coroutine in the 25 * VuServer->ctx AioContext. The coroutine consists of a vu_dispatch() loop 26 * where libvhost-user calls vu_message_read() to receive the next vhost-user 27 * protocol messages over the UNIX domain socket. 28 * 29 * When virtqueues are set up libvhost-user calls set_watch() to monitor kick 30 * fds. These fds are also handled in the VuServer->ctx AioContext. 31 * 32 * Both vu_client_trip() and kick fd monitoring can be stopped by shutting down 33 * the socket connection. Shutting down the socket connection causes 34 * vu_message_read() to fail since no more data can be received from the socket. 35 * After vu_dispatch() fails, vu_client_trip() calls vu_deinit() to stop 36 * libvhost-user before terminating the coroutine. vu_deinit() calls 37 * remove_watch() to stop monitoring kick fds and this stops virtqueue 38 * processing. 39 * 40 * When vu_client_trip() has finished cleaning up it schedules a BH in the main 41 * loop thread to accept the next client connection. 42 * 43 * When libvhost-user detects an error it calls panic_cb() and sets the 44 * dev->broken flag. Both vu_client_trip() and kick fd processing stop when 45 * the dev->broken flag is set. 46 * 47 * It is possible to switch AioContexts using 48 * vhost_user_server_detach_aio_context() and 49 * vhost_user_server_attach_aio_context(). They stop monitoring fds in the old 50 * AioContext and resume monitoring in the new AioContext. The vu_client_trip() 51 * coroutine remains in a yielded state during the switch. This is made 52 * possible by QIOChannel's support for spurious coroutine re-entry in 53 * qio_channel_yield(). The coroutine will restart I/O when re-entered from the 54 * new AioContext. 55 */ 56 57 static void vmsg_close_fds(VhostUserMsg *vmsg) 58 { 59 int i; 60 for (i = 0; i < vmsg->fd_num; i++) { 61 close(vmsg->fds[i]); 62 } 63 } 64 65 static void vmsg_unblock_fds(VhostUserMsg *vmsg) 66 { 67 int i; 68 69 /* 70 * These messages carry fd used to map memory, not to send/receive messages, 71 * so this operation is useless. In addition, in some systems this 72 * operation can fail (e.g. in macOS setting an fd returned by shm_open() 73 * non-blocking fails with errno = ENOTTY) 74 */ 75 if (vmsg->request == VHOST_USER_ADD_MEM_REG || 76 vmsg->request == VHOST_USER_SET_MEM_TABLE) { 77 return; 78 } 79 80 for (i = 0; i < vmsg->fd_num; i++) { 81 qemu_socket_set_nonblock(vmsg->fds[i]); 82 } 83 } 84 85 static void panic_cb(VuDev *vu_dev, const char *buf) 86 { 87 error_report("vu_panic: %s", buf); 88 } 89 90 void vhost_user_server_inc_in_flight(VuServer *server) 91 { 92 assert(!server->wait_idle); 93 qatomic_inc(&server->in_flight); 94 } 95 96 void vhost_user_server_dec_in_flight(VuServer *server) 97 { 98 if (qatomic_fetch_dec(&server->in_flight) == 1) { 99 if (server->wait_idle) { 100 aio_co_wake(server->co_trip); 101 } 102 } 103 } 104 105 bool vhost_user_server_has_in_flight(VuServer *server) 106 { 107 return qatomic_load_acquire(&server->in_flight) > 0; 108 } 109 110 static bool coroutine_fn 111 vu_message_read(VuDev *vu_dev, int conn_fd, VhostUserMsg *vmsg) 112 { 113 struct iovec iov = { 114 .iov_base = (char *)vmsg, 115 .iov_len = VHOST_USER_HDR_SIZE, 116 }; 117 int rc, read_bytes = 0; 118 Error *local_err = NULL; 119 const size_t max_fds = G_N_ELEMENTS(vmsg->fds); 120 VuServer *server = container_of(vu_dev, VuServer, vu_dev); 121 QIOChannel *ioc = server->ioc; 122 123 vmsg->fd_num = 0; 124 if (!ioc) { 125 error_report_err(local_err); 126 goto fail; 127 } 128 129 assert(qemu_in_coroutine()); 130 do { 131 size_t nfds = 0; 132 int *fds = NULL; 133 134 /* 135 * qio_channel_readv_full may have short reads, keeping calling it 136 * until getting VHOST_USER_HDR_SIZE or 0 bytes in total 137 */ 138 rc = qio_channel_readv_full(ioc, &iov, 1, &fds, &nfds, 0, &local_err); 139 if (rc < 0) { 140 if (rc == QIO_CHANNEL_ERR_BLOCK) { 141 assert(local_err == NULL); 142 if (server->ctx) { 143 server->in_qio_channel_yield = true; 144 qio_channel_yield(ioc, G_IO_IN); 145 server->in_qio_channel_yield = false; 146 } else { 147 return false; 148 } 149 continue; 150 } else { 151 error_report_err(local_err); 152 goto fail; 153 } 154 } 155 156 if (nfds > 0) { 157 if (vmsg->fd_num + nfds > max_fds) { 158 error_report("A maximum of %zu fds are allowed, " 159 "however got %zu fds now", 160 max_fds, vmsg->fd_num + nfds); 161 g_free(fds); 162 goto fail; 163 } 164 memcpy(vmsg->fds + vmsg->fd_num, fds, nfds * sizeof(vmsg->fds[0])); 165 vmsg->fd_num += nfds; 166 g_free(fds); 167 } 168 169 if (rc == 0) { /* socket closed */ 170 goto fail; 171 } 172 173 iov.iov_base += rc; 174 iov.iov_len -= rc; 175 read_bytes += rc; 176 } while (read_bytes != VHOST_USER_HDR_SIZE); 177 178 /* qio_channel_readv_full will make socket fds blocking, unblock them */ 179 vmsg_unblock_fds(vmsg); 180 if (vmsg->size > sizeof(vmsg->payload)) { 181 error_report("Error: too big message request: %d, " 182 "size: vmsg->size: %u, " 183 "while sizeof(vmsg->payload) = %zu", 184 vmsg->request, vmsg->size, sizeof(vmsg->payload)); 185 goto fail; 186 } 187 188 struct iovec iov_payload = { 189 .iov_base = (char *)&vmsg->payload, 190 .iov_len = vmsg->size, 191 }; 192 if (vmsg->size) { 193 rc = qio_channel_readv_all_eof(ioc, &iov_payload, 1, &local_err); 194 if (rc != 1) { 195 if (local_err) { 196 error_report_err(local_err); 197 } 198 goto fail; 199 } 200 } 201 202 return true; 203 204 fail: 205 vmsg_close_fds(vmsg); 206 207 return false; 208 } 209 210 static coroutine_fn void vu_client_trip(void *opaque) 211 { 212 VuServer *server = opaque; 213 VuDev *vu_dev = &server->vu_dev; 214 215 while (!vu_dev->broken) { 216 if (server->quiescing) { 217 server->co_trip = NULL; 218 aio_wait_kick(); 219 return; 220 } 221 /* vu_dispatch() returns false if server->ctx went away */ 222 if (!vu_dispatch(vu_dev) && server->ctx) { 223 break; 224 } 225 } 226 227 if (vhost_user_server_has_in_flight(server)) { 228 /* Wait for requests to complete before we can unmap the memory */ 229 server->wait_idle = true; 230 qemu_coroutine_yield(); 231 server->wait_idle = false; 232 } 233 assert(!vhost_user_server_has_in_flight(server)); 234 235 vu_deinit(vu_dev); 236 237 /* vu_deinit() should have called remove_watch() */ 238 assert(QTAILQ_EMPTY(&server->vu_fd_watches)); 239 240 object_unref(OBJECT(server->sioc)); 241 server->sioc = NULL; 242 243 object_unref(OBJECT(server->ioc)); 244 server->ioc = NULL; 245 246 server->co_trip = NULL; 247 if (server->restart_listener_bh) { 248 qemu_bh_schedule(server->restart_listener_bh); 249 } 250 aio_wait_kick(); 251 } 252 253 /* 254 * a wrapper for vu_kick_cb 255 * 256 * since aio_dispatch can only pass one user data pointer to the 257 * callback function, pack VuDev and pvt into a struct. Then unpack it 258 * and pass them to vu_kick_cb 259 */ 260 static void kick_handler(void *opaque) 261 { 262 VuFdWatch *vu_fd_watch = opaque; 263 VuDev *vu_dev = vu_fd_watch->vu_dev; 264 265 vu_fd_watch->cb(vu_dev, 0, vu_fd_watch->pvt); 266 267 /* Stop vu_client_trip() if an error occurred in vu_fd_watch->cb() */ 268 if (vu_dev->broken) { 269 VuServer *server = container_of(vu_dev, VuServer, vu_dev); 270 271 qio_channel_shutdown(server->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL); 272 } 273 } 274 275 static VuFdWatch *find_vu_fd_watch(VuServer *server, int fd) 276 { 277 278 VuFdWatch *vu_fd_watch, *next; 279 QTAILQ_FOREACH_SAFE(vu_fd_watch, &server->vu_fd_watches, next, next) { 280 if (vu_fd_watch->fd == fd) { 281 return vu_fd_watch; 282 } 283 } 284 return NULL; 285 } 286 287 static void 288 set_watch(VuDev *vu_dev, int fd, int vu_evt, 289 vu_watch_cb cb, void *pvt) 290 { 291 292 VuServer *server = container_of(vu_dev, VuServer, vu_dev); 293 g_assert(vu_dev); 294 g_assert(fd >= 0); 295 g_assert(cb); 296 297 VuFdWatch *vu_fd_watch = find_vu_fd_watch(server, fd); 298 299 if (!vu_fd_watch) { 300 vu_fd_watch = g_new0(VuFdWatch, 1); 301 302 QTAILQ_INSERT_TAIL(&server->vu_fd_watches, vu_fd_watch, next); 303 304 vu_fd_watch->fd = fd; 305 vu_fd_watch->cb = cb; 306 qemu_socket_set_nonblock(fd); 307 aio_set_fd_handler(server->ctx, fd, kick_handler, 308 NULL, NULL, NULL, vu_fd_watch); 309 vu_fd_watch->vu_dev = vu_dev; 310 vu_fd_watch->pvt = pvt; 311 } 312 } 313 314 315 static void remove_watch(VuDev *vu_dev, int fd) 316 { 317 VuServer *server; 318 g_assert(vu_dev); 319 g_assert(fd >= 0); 320 321 server = container_of(vu_dev, VuServer, vu_dev); 322 323 VuFdWatch *vu_fd_watch = find_vu_fd_watch(server, fd); 324 325 if (!vu_fd_watch) { 326 return; 327 } 328 aio_set_fd_handler(server->ctx, fd, NULL, NULL, NULL, NULL, NULL); 329 330 QTAILQ_REMOVE(&server->vu_fd_watches, vu_fd_watch, next); 331 g_free(vu_fd_watch); 332 } 333 334 335 static void vu_accept(QIONetListener *listener, QIOChannelSocket *sioc, 336 gpointer opaque) 337 { 338 VuServer *server = opaque; 339 340 if (server->sioc) { 341 warn_report("Only one vhost-user client is allowed to " 342 "connect the server one time"); 343 return; 344 } 345 346 if (!vu_init(&server->vu_dev, server->max_queues, sioc->fd, panic_cb, 347 vu_message_read, set_watch, remove_watch, server->vu_iface)) { 348 error_report("Failed to initialize libvhost-user"); 349 return; 350 } 351 352 /* 353 * Unset the callback function for network listener to make another 354 * vhost-user client keeping waiting until this client disconnects 355 */ 356 qio_net_listener_set_client_func(server->listener, 357 NULL, 358 NULL, 359 NULL); 360 server->sioc = sioc; 361 /* 362 * Increase the object reference, so sioc will not freed by 363 * qio_net_listener_channel_func which will call object_unref(OBJECT(sioc)) 364 */ 365 object_ref(OBJECT(server->sioc)); 366 qio_channel_set_name(QIO_CHANNEL(sioc), "vhost-user client"); 367 server->ioc = QIO_CHANNEL(sioc); 368 object_ref(OBJECT(server->ioc)); 369 370 /* TODO vu_message_write() spins if non-blocking! */ 371 qio_channel_set_blocking(server->ioc, false, NULL); 372 373 qio_channel_set_follow_coroutine_ctx(server->ioc, true); 374 375 vhost_user_server_attach_aio_context(server, server->ctx); 376 } 377 378 /* server->ctx acquired by caller */ 379 void vhost_user_server_stop(VuServer *server) 380 { 381 qemu_bh_delete(server->restart_listener_bh); 382 server->restart_listener_bh = NULL; 383 384 if (server->sioc) { 385 VuFdWatch *vu_fd_watch; 386 387 QTAILQ_FOREACH(vu_fd_watch, &server->vu_fd_watches, next) { 388 aio_set_fd_handler(server->ctx, vu_fd_watch->fd, 389 NULL, NULL, NULL, NULL, vu_fd_watch); 390 } 391 392 qio_channel_shutdown(server->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL); 393 394 AIO_WAIT_WHILE(server->ctx, server->co_trip); 395 } 396 397 if (server->listener) { 398 qio_net_listener_disconnect(server->listener); 399 object_unref(OBJECT(server->listener)); 400 } 401 } 402 403 /* 404 * Allow the next client to connect to the server. Called from a BH in the main 405 * loop. 406 */ 407 static void restart_listener_bh(void *opaque) 408 { 409 VuServer *server = opaque; 410 411 qio_net_listener_set_client_func(server->listener, vu_accept, server, 412 NULL); 413 } 414 415 /* Called with ctx acquired */ 416 void vhost_user_server_attach_aio_context(VuServer *server, AioContext *ctx) 417 { 418 VuFdWatch *vu_fd_watch; 419 420 server->ctx = ctx; 421 422 if (!server->sioc) { 423 return; 424 } 425 426 QTAILQ_FOREACH(vu_fd_watch, &server->vu_fd_watches, next) { 427 aio_set_fd_handler(ctx, vu_fd_watch->fd, kick_handler, NULL, 428 NULL, NULL, vu_fd_watch); 429 } 430 431 if (server->co_trip) { 432 /* 433 * The caller didn't fully shut down co_trip (this can happen on 434 * non-polling drains like in bdrv_graph_wrlock()). This is okay as long 435 * as it no longer tries to shut it down and we're guaranteed to still 436 * be in the same AioContext as before. 437 * 438 * co_ctx can still be NULL if we get multiple calls and only just 439 * scheduled a new coroutine in the else branch. 440 */ 441 AioContext *co_ctx = qemu_coroutine_get_aio_context(server->co_trip); 442 443 assert(!server->quiescing); 444 assert(!co_ctx || co_ctx == ctx); 445 } else { 446 server->co_trip = qemu_coroutine_create(vu_client_trip, server); 447 assert(!server->in_qio_channel_yield); 448 aio_co_schedule(ctx, server->co_trip); 449 } 450 } 451 452 /* Called with server->ctx acquired */ 453 void vhost_user_server_detach_aio_context(VuServer *server) 454 { 455 if (server->sioc) { 456 VuFdWatch *vu_fd_watch; 457 458 QTAILQ_FOREACH(vu_fd_watch, &server->vu_fd_watches, next) { 459 aio_set_fd_handler(server->ctx, vu_fd_watch->fd, 460 NULL, NULL, NULL, NULL, vu_fd_watch); 461 } 462 } 463 464 server->ctx = NULL; 465 466 if (server->ioc) { 467 if (server->in_qio_channel_yield) { 468 /* Stop receiving the next vhost-user message */ 469 qio_channel_wake_read(server->ioc); 470 } 471 } 472 } 473 474 bool vhost_user_server_start(VuServer *server, 475 SocketAddress *socket_addr, 476 AioContext *ctx, 477 uint16_t max_queues, 478 const VuDevIface *vu_iface, 479 Error **errp) 480 { 481 QEMUBH *bh; 482 QIONetListener *listener; 483 484 if (socket_addr->type != SOCKET_ADDRESS_TYPE_UNIX && 485 socket_addr->type != SOCKET_ADDRESS_TYPE_FD) { 486 error_setg(errp, "Only socket address types 'unix' and 'fd' are supported"); 487 return false; 488 } 489 490 listener = qio_net_listener_new(); 491 if (qio_net_listener_open_sync(listener, socket_addr, 1, 492 errp) < 0) { 493 object_unref(OBJECT(listener)); 494 return false; 495 } 496 497 bh = qemu_bh_new(restart_listener_bh, server); 498 499 /* zero out unspecified fields */ 500 *server = (VuServer) { 501 .listener = listener, 502 .restart_listener_bh = bh, 503 .vu_iface = vu_iface, 504 .max_queues = max_queues, 505 .ctx = ctx, 506 }; 507 508 qio_net_listener_set_name(server->listener, "vhost-user-backend-listener"); 509 510 qio_net_listener_set_client_func(server->listener, 511 vu_accept, 512 server, 513 NULL); 514 515 QTAILQ_INIT(&server->vu_fd_watches); 516 return true; 517 } 518