lib/librte_vhost/socket.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <limits.h>
   8 #include <stdlib.h>
   9 #include <unistd.h>
  10 #include <string.h>
  11 #include <sys/types.h>
  12 #include <sys/socket.h>
  13 #include <sys/un.h>
  14 #include <sys/queue.h>
  15 #include <errno.h>
  16 #include <fcntl.h>
  17 #include <pthread.h>
  18
  19 #include <rte_log.h>
  20
  21 #include "fd_man.h"
  22 #include "vhost.h"
  23 #include "vhost_user.h"
  24
  25
  26 TAILQ_HEAD(vhost_user_connection_list, vhost_user_connection);
  27
  28 /*
  29  * Every time rte_vhost_driver_register() is invoked, an associated
  30  * vhost_user_socket struct will be created.
  31  */
  32 struct vhost_user_socket {
  33         struct vhost_user_connection_list conn_list;
  34         pthread_mutex_t conn_mutex;
  35         char *path;
  36         int socket_fd;
  37         struct sockaddr_un un;
  38         bool is_server;
  39         bool reconnect;
  40         bool dequeue_zero_copy;
  41         bool iommu_support;
  42         bool use_builtin_virtio_net;
  43
  44         /*
  45          * The "supported_features" indicates the feature bits the
  46          * vhost driver supports. The "features" indicates the feature
  47          * bits after the rte_vhost_driver_features_disable/enable().
  48          * It is also the final feature bits used for vhost-user
  49          * features negotiation.
  50          */
  51         uint64_t supported_features;
  52         uint64_t features;
  53
  54         struct vhost_device_ops const *notify_ops;
  55 };
  56
  57 struct vhost_user_connection {
  58         struct vhost_user_socket *vsocket;
  59         int connfd;
  60         int vid;
  61
  62         TAILQ_ENTRY(vhost_user_connection) next;
  63 };
  64
  65 #define MAX_VHOST_SOCKET 1024
  66 struct vhost_user {
  67         struct vhost_user_socket *vsockets[MAX_VHOST_SOCKET];
  68         struct fdset fdset;
  69         int vsocket_cnt;
  70         pthread_mutex_t mutex;
  71 };
  72
  73 #define MAX_VIRTIO_BACKLOG 128
  74
  75 static void vhost_user_server_new_connection(int fd, void *data, int *remove);
  76 static void vhost_user_read_cb(int fd, void *dat, int *remove);
  77 static int create_unix_socket(struct vhost_user_socket *vsocket);
  78 static int vhost_user_start_client(struct vhost_user_socket *vsocket);
  79
  80 static struct vhost_user vhost_user = {
  81         .fdset = {
  82                 .fd = { [0 ... MAX_FDS - 1] = {-1, NULL, NULL, NULL, 0} },
  83                 .fd_mutex = PTHREAD_MUTEX_INITIALIZER,
  84                 .num = 0
  85         },
  86         .vsocket_cnt = 0,
  87         .mutex = PTHREAD_MUTEX_INITIALIZER,
  88 };
  89
  90 /* return bytes# of read on success or negative val on failure. */
  91 int
  92 read_fd_message(int sockfd, char *buf, int buflen, int *fds, int fd_num)
  93 {
  94         struct iovec iov;
  95         struct msghdr msgh;
  96         size_t fdsize = fd_num * sizeof(int);
  97         char control[CMSG_SPACE(fdsize)];
  98         struct cmsghdr *cmsg;
  99         int got_fds = 0;
 100         int ret;
 101
 102         memset(&msgh, 0, sizeof(msgh));
 103         iov.iov_base = buf;
 104         iov.iov_len  = buflen;
 105
 106         msgh.msg_iov = &iov;
 107         msgh.msg_iovlen = 1;
 108         msgh.msg_control = control;
 109         msgh.msg_controllen = sizeof(control);
 110
 111         ret = recvmsg(sockfd, &msgh, 0);
 112         if (ret <= 0) {
 113                 RTE_LOG(ERR, VHOST_CONFIG, "recvmsg failed\n");
 114                 return ret;
 115         }
 116
 117         if (msgh.msg_flags & (MSG_TRUNC | MSG_CTRUNC)) {
 118                 RTE_LOG(ERR, VHOST_CONFIG, "truncted msg\n");
 119                 return -1;
 120         }
 121
 122         for (cmsg = CMSG_FIRSTHDR(&msgh); cmsg != NULL;
 123                 cmsg = CMSG_NXTHDR(&msgh, cmsg)) {
 124                 if ((cmsg->cmsg_level == SOL_SOCKET) &&
 125                         (cmsg->cmsg_type == SCM_RIGHTS)) {
 126                         got_fds = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int);
 127                         memcpy(fds, CMSG_DATA(cmsg), got_fds * sizeof(int));
 128                         break;
 129                 }
 130         }
 131
 132         /* Clear out unused file descriptors */
 133         while (got_fds < fd_num)
 134                 fds[got_fds++] = -1;
 135
 136         return ret;
 137 }
 138
 139 int
 140 send_fd_message(int sockfd, char *buf, int buflen, int *fds, int fd_num)
 141 {
 142
 143         struct iovec iov;
 144         struct msghdr msgh;
 145         size_t fdsize = fd_num * sizeof(int);
 146         char control[CMSG_SPACE(fdsize)];
 147         struct cmsghdr *cmsg;
 148         int ret;
 149
 150         memset(&msgh, 0, sizeof(msgh));
 151         iov.iov_base = buf;
 152         iov.iov_len = buflen;
 153
 154         msgh.msg_iov = &iov;
 155         msgh.msg_iovlen = 1;
 156
 157         if (fds && fd_num > 0) {
 158                 msgh.msg_control = control;
 159                 msgh.msg_controllen = sizeof(control);
 160                 cmsg = CMSG_FIRSTHDR(&msgh);
 161                 if (cmsg == NULL) {
 162                         RTE_LOG(ERR, VHOST_CONFIG, "cmsg == NULL\n");
 163                         errno = EINVAL;
 164                         return -1;
 165                 }
 166                 cmsg->cmsg_len = CMSG_LEN(fdsize);
 167                 cmsg->cmsg_level = SOL_SOCKET;
 168                 cmsg->cmsg_type = SCM_RIGHTS;
 169                 memcpy(CMSG_DATA(cmsg), fds, fdsize);
 170         } else {
 171                 msgh.msg_control = NULL;
 172                 msgh.msg_controllen = 0;
 173         }
 174
 175         do {
 176                 ret = sendmsg(sockfd, &msgh, MSG_NOSIGNAL);
 177         } while (ret < 0 && errno == EINTR);
 178
 179         if (ret < 0) {
 180                 RTE_LOG(ERR, VHOST_CONFIG,  "sendmsg error\n");
 181                 return ret;
 182         }
 183
 184         return ret;
 185 }
 186
 187 static void
 188 vhost_user_add_connection(int fd, struct vhost_user_socket *vsocket)
 189 {
 190         int vid;
 191         size_t size;
 192         struct vhost_user_connection *conn;
 193         int ret;
 194
 195         conn = malloc(sizeof(*conn));
 196         if (conn == NULL) {
 197                 close(fd);
 198                 return;
 199         }
 200
 201         vid = vhost_new_device();
 202         if (vid == -1) {
 203                 goto err;
 204         }
 205
 206         size = strnlen(vsocket->path, PATH_MAX);
 207         vhost_set_ifname(vid, vsocket->path, size);
 208
 209         vhost_set_builtin_virtio_net(vid, vsocket->use_builtin_virtio_net);
 210
 211         if (vsocket->dequeue_zero_copy)
 212                 vhost_enable_dequeue_zero_copy(vid);
 213
 214         RTE_LOG(INFO, VHOST_CONFIG, "new device, handle is %d\n", vid);
 215
 216         if (vsocket->notify_ops->new_connection) {
 217                 ret = vsocket->notify_ops->new_connection(vid);
 218                 if (ret < 0) {
 219                         RTE_LOG(ERR, VHOST_CONFIG,
 220                                 "failed to add vhost user connection with fd %d\n",
 221                                 fd);
 222                         goto err;
 223                 }
 224         }
 225
 226         conn->connfd = fd;
 227         conn->vsocket = vsocket;
 228         conn->vid = vid;
 229         ret = fdset_add(&vhost_user.fdset, fd, vhost_user_read_cb,
 230                         NULL, conn);
 231         if (ret < 0) {
 232                 RTE_LOG(ERR, VHOST_CONFIG,
 233                         "failed to add fd %d into vhost server fdset\n",
 234                         fd);
 235
 236                 if (vsocket->notify_ops->destroy_connection)
 237                         vsocket->notify_ops->destroy_connection(conn->vid);
 238
 239                 goto err;
 240         }
 241
 242         pthread_mutex_lock(&vsocket->conn_mutex);
 243         TAILQ_INSERT_TAIL(&vsocket->conn_list, conn, next);
 244         pthread_mutex_unlock(&vsocket->conn_mutex);
 245
 246         fdset_pipe_notify(&vhost_user.fdset);
 247         return;
 248
 249 err:
 250         free(conn);
 251         close(fd);
 252 }
 253
 254 /* call back when there is new vhost-user connection from client  */
 255 static void
 256 vhost_user_server_new_connection(int fd, void *dat, int *remove __rte_unused)
 257 {
 258         struct vhost_user_socket *vsocket = dat;
 259
 260         fd = accept(fd, NULL, NULL);
 261         if (fd < 0)
 262                 return;
 263
 264         RTE_LOG(INFO, VHOST_CONFIG, "new vhost user connection is %d\n", fd);
 265         vhost_user_add_connection(fd, vsocket);
 266 }
 267
 268 static void
 269 vhost_user_read_cb(int connfd, void *dat, int *remove)
 270 {
 271         struct vhost_user_connection *conn = dat;
 272         struct vhost_user_socket *vsocket = conn->vsocket;
 273         int ret;
 274
 275         ret = vhost_user_msg_handler(conn->vid, connfd);
 276         if (ret < 0) {
 277                 close(connfd);
 278                 *remove = 1;
 279                 vhost_destroy_device(conn->vid);
 280
 281                 if (vsocket->notify_ops->destroy_connection)
 282                         vsocket->notify_ops->destroy_connection(conn->vid);
 283
 284                 pthread_mutex_lock(&vsocket->conn_mutex);
 285                 TAILQ_REMOVE(&vsocket->conn_list, conn, next);
 286                 pthread_mutex_unlock(&vsocket->conn_mutex);
 287
 288                 free(conn);
 289
 290                 if (vsocket->reconnect) {
 291                         create_unix_socket(vsocket);
 292                         vhost_user_start_client(vsocket);
 293                 }
 294         }
 295 }
 296
 297 static int
 298 create_unix_socket(struct vhost_user_socket *vsocket)
 299 {
 300         int fd;
 301         struct sockaddr_un *un = &vsocket->un;
 302
 303         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 304         if (fd < 0)
 305                 return -1;
 306         RTE_LOG(INFO, VHOST_CONFIG, "vhost-user %s: socket created, fd: %d\n",
 307                 vsocket->is_server ? "server" : "client", fd);
 308
 309         if (!vsocket->is_server && fcntl(fd, F_SETFL, O_NONBLOCK)) {
 310                 RTE_LOG(ERR, VHOST_CONFIG,
 311                         "vhost-user: can't set nonblocking mode for socket, fd: "
 312                         "%d (%s)\n", fd, strerror(errno));
 313                 close(fd);
 314                 return -1;
 315         }
 316
 317         memset(un, 0, sizeof(*un));
 318         un->sun_family = AF_UNIX;
 319         strncpy(un->sun_path, vsocket->path, sizeof(un->sun_path));
 320         un->sun_path[sizeof(un->sun_path) - 1] = '\0';
 321
 322         vsocket->socket_fd = fd;
 323         return 0;
 324 }
 325
 326 static int
 327 vhost_user_start_server(struct vhost_user_socket *vsocket)
 328 {
 329         int ret;
 330         int fd = vsocket->socket_fd;
 331         const char *path = vsocket->path;
 332
 333         /*
 334          * bind () may fail if the socket file with the same name already
 335          * exists. But the library obviously should not delete the file
 336          * provided by the user, since we can not be sure that it is not
 337          * being used by other applications. Moreover, many applications form
 338          * socket names based on user input, which is prone to errors.
 339          *
 340          * The user must ensure that the socket does not exist before
 341          * registering the vhost driver in server mode.
 342          */
 343         ret = bind(fd, (struct sockaddr *)&vsocket->un, sizeof(vsocket->un));
 344         if (ret < 0) {
 345                 RTE_LOG(ERR, VHOST_CONFIG,
 346                         "failed to bind to %s: %s; remove it and try again\n",
 347                         path, strerror(errno));
 348                 goto err;
 349         }
 350         RTE_LOG(INFO, VHOST_CONFIG, "bind to %s\n", path);
 351
 352         ret = listen(fd, MAX_VIRTIO_BACKLOG);
 353         if (ret < 0)
 354                 goto err;
 355
 356         ret = fdset_add(&vhost_user.fdset, fd, vhost_user_server_new_connection,
 357                   NULL, vsocket);
 358         if (ret < 0) {
 359                 RTE_LOG(ERR, VHOST_CONFIG,
 360                         "failed to add listen fd %d to vhost server fdset\n",
 361                         fd);
 362                 goto err;
 363         }
 364
 365         return 0;
 366
 367 err:
 368         close(fd);
 369         return -1;
 370 }
 371
 372 struct vhost_user_reconnect {
 373         struct sockaddr_un un;
 374         int fd;
 375         struct vhost_user_socket *vsocket;
 376
 377         TAILQ_ENTRY(vhost_user_reconnect) next;
 378 };
 379
 380 TAILQ_HEAD(vhost_user_reconnect_tailq_list, vhost_user_reconnect);
 381 struct vhost_user_reconnect_list {
 382         struct vhost_user_reconnect_tailq_list head;
 383         pthread_mutex_t mutex;
 384 };
 385
 386 static struct vhost_user_reconnect_list reconn_list;
 387 static pthread_t reconn_tid;
 388
 389 static int
 390 vhost_user_connect_nonblock(int fd, struct sockaddr *un, size_t sz)
 391 {
 392         int ret, flags;
 393
 394         ret = connect(fd, un, sz);
 395         if (ret < 0 && errno != EISCONN)
 396                 return -1;
 397
 398         flags = fcntl(fd, F_GETFL, 0);
 399         if (flags < 0) {
 400                 RTE_LOG(ERR, VHOST_CONFIG,
 401                         "can't get flags for connfd %d\n", fd);
 402                 return -2;
 403         }
 404         if ((flags & O_NONBLOCK) && fcntl(fd, F_SETFL, flags & ~O_NONBLOCK)) {
 405                 RTE_LOG(ERR, VHOST_CONFIG,
 406                                 "can't disable nonblocking on fd %d\n", fd);
 407                 return -2;
 408         }
 409         return 0;
 410 }
 411
 412 static void *
 413 vhost_user_client_reconnect(void *arg __rte_unused)
 414 {
 415         int ret;
 416         struct vhost_user_reconnect *reconn, *next;
 417
 418         while (1) {
 419                 pthread_mutex_lock(&reconn_list.mutex);
 420
 421                 /*
 422                  * An equal implementation of TAILQ_FOREACH_SAFE,
 423                  * which does not exist on all platforms.
 424                  */
 425                 for (reconn = TAILQ_FIRST(&reconn_list.head);
 426                      reconn != NULL; reconn = next) {
 427                         next = TAILQ_NEXT(reconn, next);
 428
 429                         ret = vhost_user_connect_nonblock(reconn->fd,
 430                                                 (struct sockaddr *)&reconn->un,
 431                                                 sizeof(reconn->un));
 432                         if (ret == -2) {
 433                                 close(reconn->fd);
 434                                 RTE_LOG(ERR, VHOST_CONFIG,
 435                                         "reconnection for fd %d failed\n",
 436                                         reconn->fd);
 437                                 goto remove_fd;
 438                         }
 439                         if (ret == -1)
 440                                 continue;
 441
 442                         RTE_LOG(INFO, VHOST_CONFIG,
 443                                 "%s: connected\n", reconn->vsocket->path);
 444                         vhost_user_add_connection(reconn->fd, reconn->vsocket);
 445 remove_fd:
 446                         TAILQ_REMOVE(&reconn_list.head, reconn, next);
 447                         free(reconn);
 448                 }
 449
 450                 pthread_mutex_unlock(&reconn_list.mutex);
 451                 sleep(1);
 452         }
 453
 454         return NULL;
 455 }
 456
 457 static int
 458 vhost_user_reconnect_init(void)
 459 {
 460         int ret;
 461         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 462
 463         ret = pthread_mutex_init(&reconn_list.mutex, NULL);
 464         if (ret < 0) {
 465                 RTE_LOG(ERR, VHOST_CONFIG, "failed to initialize mutex");
 466                 return ret;
 467         }
 468         TAILQ_INIT(&reconn_list.head);
 469
 470         ret = pthread_create(&reconn_tid, NULL,
 471                              vhost_user_client_reconnect, NULL);
 472         if (ret != 0) {
 473                 RTE_LOG(ERR, VHOST_CONFIG, "failed to create reconnect thread");
 474                 if (pthread_mutex_destroy(&reconn_list.mutex)) {
 475                         RTE_LOG(ERR, VHOST_CONFIG,
 476                                 "failed to destroy reconnect mutex");
 477                 }
 478         } else {
 479                 snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN,
 480                          "vhost-reconn");
 481
 482                 if (rte_thread_setname(reconn_tid, thread_name)) {
 483                         RTE_LOG(DEBUG, VHOST_CONFIG,
 484                                 "failed to set reconnect thread name");
 485                 }
 486         }
 487
 488         return ret;
 489 }
 490
 491 static int
 492 vhost_user_start_client(struct vhost_user_socket *vsocket)
 493 {
 494         int ret;
 495         int fd = vsocket->socket_fd;
 496         const char *path = vsocket->path;
 497         struct vhost_user_reconnect *reconn;
 498
 499         ret = vhost_user_connect_nonblock(fd, (struct sockaddr *)&vsocket->un,
 500                                           sizeof(vsocket->un));
 501         if (ret == 0) {
 502                 vhost_user_add_connection(fd, vsocket);
 503                 return 0;
 504         }
 505
 506         RTE_LOG(WARNING, VHOST_CONFIG,
 507                 "failed to connect to %s: %s\n",
 508                 path, strerror(errno));
 509
 510         if (ret == -2 || !vsocket->reconnect) {
 511                 close(fd);
 512                 return -1;
 513         }
 514
 515         RTE_LOG(INFO, VHOST_CONFIG, "%s: reconnecting...\n", path);
 516         reconn = malloc(sizeof(*reconn));
 517         if (reconn == NULL) {
 518                 RTE_LOG(ERR, VHOST_CONFIG,
 519                         "failed to allocate memory for reconnect\n");
 520                 close(fd);
 521                 return -1;
 522         }
 523         reconn->un = vsocket->un;
 524         reconn->fd = fd;
 525         reconn->vsocket = vsocket;
 526         pthread_mutex_lock(&reconn_list.mutex);
 527         TAILQ_INSERT_TAIL(&reconn_list.head, reconn, next);
 528         pthread_mutex_unlock(&reconn_list.mutex);
 529
 530         return 0;
 531 }
 532
 533 static struct vhost_user_socket *
 534 find_vhost_user_socket(const char *path)
 535 {
 536         int i;
 537
 538         for (i = 0; i < vhost_user.vsocket_cnt; i++) {
 539                 struct vhost_user_socket *vsocket = vhost_user.vsockets[i];
 540
 541                 if (!strcmp(vsocket->path, path))
 542                         return vsocket;
 543         }
 544
 545         return NULL;
 546 }
 547
 548 int
 549 rte_vhost_driver_disable_features(const char *path, uint64_t features)
 550 {
 551         struct vhost_user_socket *vsocket;
 552
 553         pthread_mutex_lock(&vhost_user.mutex);
 554         vsocket = find_vhost_user_socket(path);
 555
 556         /* Note that use_builtin_virtio_net is not affected by this function
 557          * since callers may want to selectively disable features of the
 558          * built-in vhost net device backend.
 559          */
 560
 561         if (vsocket)
 562                 vsocket->features &= ~features;
 563         pthread_mutex_unlock(&vhost_user.mutex);
 564
 565         return vsocket ? 0 : -1;
 566 }
 567
 568 int
 569 rte_vhost_driver_enable_features(const char *path, uint64_t features)
 570 {
 571         struct vhost_user_socket *vsocket;
 572
 573         pthread_mutex_lock(&vhost_user.mutex);
 574         vsocket = find_vhost_user_socket(path);
 575         if (vsocket) {
 576                 if ((vsocket->supported_features & features) != features) {
 577                         /*
 578                          * trying to enable features the driver doesn't
 579                          * support.
 580                          */
 581                         pthread_mutex_unlock(&vhost_user.mutex);
 582                         return -1;
 583                 }
 584                 vsocket->features |= features;
 585         }
 586         pthread_mutex_unlock(&vhost_user.mutex);
 587
 588         return vsocket ? 0 : -1;
 589 }
 590
 591 int
 592 rte_vhost_driver_set_features(const char *path, uint64_t features)
 593 {
 594         struct vhost_user_socket *vsocket;
 595
 596         pthread_mutex_lock(&vhost_user.mutex);
 597         vsocket = find_vhost_user_socket(path);
 598         if (vsocket) {
 599                 vsocket->supported_features = features;
 600                 vsocket->features = features;
 601
 602                 /* Anyone setting feature bits is implementing their own vhost
 603                  * device backend.
 604                  */
 605                 vsocket->use_builtin_virtio_net = false;
 606         }
 607         pthread_mutex_unlock(&vhost_user.mutex);
 608
 609         return vsocket ? 0 : -1;
 610 }
 611
 612 int
 613 rte_vhost_driver_get_features(const char *path, uint64_t *features)
 614 {
 615         struct vhost_user_socket *vsocket;
 616
 617         pthread_mutex_lock(&vhost_user.mutex);
 618         vsocket = find_vhost_user_socket(path);
 619         if (vsocket)
 620                 *features = vsocket->features;
 621         pthread_mutex_unlock(&vhost_user.mutex);
 622
 623         if (!vsocket) {
 624                 RTE_LOG(ERR, VHOST_CONFIG,
 625                         "socket file %s is not registered yet.\n", path);
 626                 return -1;
 627         } else {
 628                 return 0;
 629         }
 630 }
 631
 632 /*
 633  * Register a new vhost-user socket; here we could act as server
 634  * (the default case), or client (when RTE_VHOST_USER_CLIENT) flag
 635  * is set.
 636  */
 637 int
 638 rte_vhost_driver_register(const char *path, uint64_t flags)
 639 {
 640         int ret = -1;
 641         struct vhost_user_socket *vsocket;
 642
 643         if (!path)
 644                 return -1;
 645
 646         pthread_mutex_lock(&vhost_user.mutex);
 647
 648         if (vhost_user.vsocket_cnt == MAX_VHOST_SOCKET) {
 649                 RTE_LOG(ERR, VHOST_CONFIG,
 650                         "error: the number of vhost sockets reaches maximum\n");
 651                 goto out;
 652         }
 653
 654         vsocket = malloc(sizeof(struct vhost_user_socket));
 655         if (!vsocket)
 656                 goto out;
 657         memset(vsocket, 0, sizeof(struct vhost_user_socket));
 658         vsocket->path = strdup(path);
 659         if (vsocket->path == NULL) {
 660                 RTE_LOG(ERR, VHOST_CONFIG,
 661                         "error: failed to copy socket path string\n");
 662                 free(vsocket);
 663                 goto out;
 664         }
 665         TAILQ_INIT(&vsocket->conn_list);
 666         ret = pthread_mutex_init(&vsocket->conn_mutex, NULL);
 667         if (ret) {
 668                 RTE_LOG(ERR, VHOST_CONFIG,
 669                         "error: failed to init connection mutex\n");
 670                 goto out_free;
 671         }
 672         vsocket->dequeue_zero_copy = flags & RTE_VHOST_USER_DEQUEUE_ZERO_COPY;
 673
 674         /*
 675          * Set the supported features correctly for the builtin vhost-user
 676          * net driver.
 677          *
 678          * Applications know nothing about features the builtin virtio net
 679          * driver (virtio_net.c) supports, thus it's not possible for them
 680          * to invoke rte_vhost_driver_set_features(). To workaround it, here
 681          * we set it unconditionally. If the application want to implement
 682          * another vhost-user driver (say SCSI), it should call the
 683          * rte_vhost_driver_set_features(), which will overwrite following
 684          * two values.
 685          */
 686         vsocket->use_builtin_virtio_net = true;
 687         vsocket->supported_features = VIRTIO_NET_SUPPORTED_FEATURES;
 688         vsocket->features           = VIRTIO_NET_SUPPORTED_FEATURES;
 689
 690         if (!(flags & RTE_VHOST_USER_IOMMU_SUPPORT)) {
 691                 vsocket->supported_features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 692                 vsocket->features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 693         }
 694
 695         if ((flags & RTE_VHOST_USER_CLIENT) != 0) {
 696                 vsocket->reconnect = !(flags & RTE_VHOST_USER_NO_RECONNECT);
 697                 if (vsocket->reconnect && reconn_tid == 0) {
 698                         if (vhost_user_reconnect_init() != 0)
 699                                 goto out_mutex;
 700                 }
 701         } else {
 702                 vsocket->is_server = true;
 703         }
 704         ret = create_unix_socket(vsocket);
 705         if (ret < 0) {
 706                 goto out_mutex;
 707         }
 708
 709         vhost_user.vsockets[vhost_user.vsocket_cnt++] = vsocket;
 710
 711         pthread_mutex_unlock(&vhost_user.mutex);
 712         return ret;
 713
 714 out_mutex:
 715         if (pthread_mutex_destroy(&vsocket->conn_mutex)) {
 716                 RTE_LOG(ERR, VHOST_CONFIG,
 717                         "error: failed to destroy connection mutex\n");
 718         }
 719 out_free:
 720         free(vsocket->path);
 721         free(vsocket);
 722 out:
 723         pthread_mutex_unlock(&vhost_user.mutex);
 724
 725         return ret;
 726 }
 727
 728 static bool
 729 vhost_user_remove_reconnect(struct vhost_user_socket *vsocket)
 730 {
 731         int found = false;
 732         struct vhost_user_reconnect *reconn, *next;
 733
 734         pthread_mutex_lock(&reconn_list.mutex);
 735
 736         for (reconn = TAILQ_FIRST(&reconn_list.head);
 737              reconn != NULL; reconn = next) {
 738                 next = TAILQ_NEXT(reconn, next);
 739
 740                 if (reconn->vsocket == vsocket) {
 741                         TAILQ_REMOVE(&reconn_list.head, reconn, next);
 742                         close(reconn->fd);
 743                         free(reconn);
 744                         found = true;
 745                         break;
 746                 }
 747         }
 748         pthread_mutex_unlock(&reconn_list.mutex);
 749         return found;
 750 }
 751
 752 /**
 753  * Unregister the specified vhost socket
 754  */
 755 int
 756 rte_vhost_driver_unregister(const char *path)
 757 {
 758         int i;
 759         int count;
 760         struct vhost_user_connection *conn, *next;
 761
 762         pthread_mutex_lock(&vhost_user.mutex);
 763
 764         for (i = 0; i < vhost_user.vsocket_cnt; i++) {
 765                 struct vhost_user_socket *vsocket = vhost_user.vsockets[i];
 766
 767                 if (!strcmp(vsocket->path, path)) {
 768                         if (vsocket->is_server) {
 769                                 fdset_del(&vhost_user.fdset, vsocket->socket_fd);
 770                                 close(vsocket->socket_fd);
 771                                 unlink(path);
 772                         } else if (vsocket->reconnect) {
 773                                 vhost_user_remove_reconnect(vsocket);
 774                         }
 775
 776                         pthread_mutex_lock(&vsocket->conn_mutex);
 777                         for (conn = TAILQ_FIRST(&vsocket->conn_list);
 778                              conn != NULL;
 779                              conn = next) {
 780                                 next = TAILQ_NEXT(conn, next);
 781
 782                                 fdset_del(&vhost_user.fdset, conn->connfd);
 783                                 RTE_LOG(INFO, VHOST_CONFIG,
 784                                         "free connfd = %d for device '%s'\n",
 785                                         conn->connfd, path);
 786                                 close(conn->connfd);
 787                                 vhost_destroy_device(conn->vid);
 788                                 TAILQ_REMOVE(&vsocket->conn_list, conn, next);
 789                                 free(conn);
 790                         }
 791                         pthread_mutex_unlock(&vsocket->conn_mutex);
 792
 793                         pthread_mutex_destroy(&vsocket->conn_mutex);
 794                         free(vsocket->path);
 795                         free(vsocket);
 796
 797                         count = --vhost_user.vsocket_cnt;
 798                         vhost_user.vsockets[i] = vhost_user.vsockets[count];
 799                         vhost_user.vsockets[count] = NULL;
 800                         pthread_mutex_unlock(&vhost_user.mutex);
 801
 802                         return 0;
 803                 }
 804         }
 805         pthread_mutex_unlock(&vhost_user.mutex);
 806
 807         return -1;
 808 }
 809
 810 /*
 811  * Register ops so that we can add/remove device to data core.
 812  */
 813 int
 814 rte_vhost_driver_callback_register(const char *path,
 815         struct vhost_device_ops const * const ops)
 816 {
 817         struct vhost_user_socket *vsocket;
 818
 819         pthread_mutex_lock(&vhost_user.mutex);
 820         vsocket = find_vhost_user_socket(path);
 821         if (vsocket)
 822                 vsocket->notify_ops = ops;
 823         pthread_mutex_unlock(&vhost_user.mutex);
 824
 825         return vsocket ? 0 : -1;
 826 }
 827
 828 struct vhost_device_ops const *
 829 vhost_driver_callback_get(const char *path)
 830 {
 831         struct vhost_user_socket *vsocket;
 832
 833         pthread_mutex_lock(&vhost_user.mutex);
 834         vsocket = find_vhost_user_socket(path);
 835         pthread_mutex_unlock(&vhost_user.mutex);
 836
 837         return vsocket ? vsocket->notify_ops : NULL;
 838 }
 839
 840 int
 841 rte_vhost_driver_start(const char *path)
 842 {
 843         struct vhost_user_socket *vsocket;
 844         static pthread_t fdset_tid;
 845         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 846
 847         pthread_mutex_lock(&vhost_user.mutex);
 848         vsocket = find_vhost_user_socket(path);
 849         pthread_mutex_unlock(&vhost_user.mutex);
 850
 851         if (!vsocket)
 852                 return -1;
 853
 854         if (fdset_tid == 0) {
 855                 /**
 856                  * create a pipe which will be waited by poll and notified to
 857                  * rebuild the wait list of poll.
 858                  */
 859                 if (fdset_pipe_init(&vhost_user.fdset) < 0) {
 860                         RTE_LOG(ERR, VHOST_CONFIG,
 861                                 "failed to create pipe for vhost fdset\n");
 862                         return -1;
 863                 }
 864
 865                 int ret = pthread_create(&fdset_tid, NULL, fdset_event_dispatch,
 866                                      &vhost_user.fdset);
 867                 if (ret != 0) {
 868                         RTE_LOG(ERR, VHOST_CONFIG,
 869                                 "failed to create fdset handling thread");
 870
 871                         fdset_pipe_uninit(&vhost_user.fdset);
 872                         return -1;
 873                 } else {
 874                         snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN,
 875                                  "vhost-events");
 876
 877                         if (rte_thread_setname(fdset_tid, thread_name)) {
 878                                 RTE_LOG(DEBUG, VHOST_CONFIG,
 879                                         "failed to set vhost-event thread name");
 880                         }
 881                 }
 882         }
 883
 884         if (vsocket->is_server)
 885                 return vhost_user_start_server(vsocket);
 886         else
 887                 return vhost_user_start_client(vsocket);
 888 }