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