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         /*
  55          * Device id to identify a specific backend device.
  56          * It's set to -1 for the default software implementation.
  57          * If valid, one socket can have 1 connection only.
  58          */
  59         int vdpa_dev_id;
  60
  61         struct vhost_device_ops const *notify_ops;
  62 };
  63
  64 struct vhost_user_connection {
  65         struct vhost_user_socket *vsocket;
  66         int connfd;
  67         int vid;
  68
  69         TAILQ_ENTRY(vhost_user_connection) next;
  70 };
  71
  72 #define MAX_VHOST_SOCKET 1024
  73 struct vhost_user {
  74         struct vhost_user_socket *vsockets[MAX_VHOST_SOCKET];
  75         struct fdset fdset;
  76         int vsocket_cnt;
  77         pthread_mutex_t mutex;
  78 };
  79
  80 #define MAX_VIRTIO_BACKLOG 128
  81
  82 static void vhost_user_server_new_connection(int fd, void *data, int *remove);
  83 static void vhost_user_read_cb(int fd, void *dat, int *remove);
  84 static int create_unix_socket(struct vhost_user_socket *vsocket);
  85 static int vhost_user_start_client(struct vhost_user_socket *vsocket);
  86
  87 static struct vhost_user vhost_user = {
  88         .fdset = {
  89                 .fd = { [0 ... MAX_FDS - 1] = {-1, NULL, NULL, NULL, 0} },
  90                 .fd_mutex = PTHREAD_MUTEX_INITIALIZER,
  91                 .num = 0
  92         },
  93         .vsocket_cnt = 0,
  94         .mutex = PTHREAD_MUTEX_INITIALIZER,
  95 };
  96
  97 /* return bytes# of read on success or negative val on failure. */
  98 int
  99 read_fd_message(int sockfd, char *buf, int buflen, int *fds, int fd_num)
 100 {
 101         struct iovec iov;
 102         struct msghdr msgh;
 103         size_t fdsize = fd_num * sizeof(int);
 104         char control[CMSG_SPACE(fdsize)];
 105         struct cmsghdr *cmsg;
 106         int got_fds = 0;
 107         int ret;
 108
 109         memset(&msgh, 0, sizeof(msgh));
 110         iov.iov_base = buf;
 111         iov.iov_len  = buflen;
 112
 113         msgh.msg_iov = &iov;
 114         msgh.msg_iovlen = 1;
 115         msgh.msg_control = control;
 116         msgh.msg_controllen = sizeof(control);
 117
 118         ret = recvmsg(sockfd, &msgh, 0);
 119         if (ret <= 0) {
 120                 RTE_LOG(ERR, VHOST_CONFIG, "recvmsg failed\n");
 121                 return ret;
 122         }
 123
 124         if (msgh.msg_flags & (MSG_TRUNC | MSG_CTRUNC)) {
 125                 RTE_LOG(ERR, VHOST_CONFIG, "truncted msg\n");
 126                 return -1;
 127         }
 128
 129         for (cmsg = CMSG_FIRSTHDR(&msgh); cmsg != NULL;
 130                 cmsg = CMSG_NXTHDR(&msgh, cmsg)) {
 131                 if ((cmsg->cmsg_level == SOL_SOCKET) &&
 132                         (cmsg->cmsg_type == SCM_RIGHTS)) {
 133                         got_fds = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int);
 134                         memcpy(fds, CMSG_DATA(cmsg), got_fds * sizeof(int));
 135                         break;
 136                 }
 137         }
 138
 139         /* Clear out unused file descriptors */
 140         while (got_fds < fd_num)
 141                 fds[got_fds++] = -1;
 142
 143         return ret;
 144 }
 145
 146 int
 147 send_fd_message(int sockfd, char *buf, int buflen, int *fds, int fd_num)
 148 {
 149
 150         struct iovec iov;
 151         struct msghdr msgh;
 152         size_t fdsize = fd_num * sizeof(int);
 153         char control[CMSG_SPACE(fdsize)];
 154         struct cmsghdr *cmsg;
 155         int ret;
 156
 157         memset(&msgh, 0, sizeof(msgh));
 158         iov.iov_base = buf;
 159         iov.iov_len = buflen;
 160
 161         msgh.msg_iov = &iov;
 162         msgh.msg_iovlen = 1;
 163
 164         if (fds && fd_num > 0) {
 165                 msgh.msg_control = control;
 166                 msgh.msg_controllen = sizeof(control);
 167                 cmsg = CMSG_FIRSTHDR(&msgh);
 168                 if (cmsg == NULL) {
 169                         RTE_LOG(ERR, VHOST_CONFIG, "cmsg == NULL\n");
 170                         errno = EINVAL;
 171                         return -1;
 172                 }
 173                 cmsg->cmsg_len = CMSG_LEN(fdsize);
 174                 cmsg->cmsg_level = SOL_SOCKET;
 175                 cmsg->cmsg_type = SCM_RIGHTS;
 176                 memcpy(CMSG_DATA(cmsg), fds, fdsize);
 177         } else {
 178                 msgh.msg_control = NULL;
 179                 msgh.msg_controllen = 0;
 180         }
 181
 182         do {
 183                 ret = sendmsg(sockfd, &msgh, MSG_NOSIGNAL);
 184         } while (ret < 0 && errno == EINTR);
 185
 186         if (ret < 0) {
 187                 RTE_LOG(ERR, VHOST_CONFIG,  "sendmsg error\n");
 188                 return ret;
 189         }
 190
 191         return ret;
 192 }
 193
 194 static void
 195 vhost_user_add_connection(int fd, struct vhost_user_socket *vsocket)
 196 {
 197         int vid;
 198         size_t size;
 199         struct vhost_user_connection *conn;
 200         int ret;
 201
 202         conn = malloc(sizeof(*conn));
 203         if (conn == NULL) {
 204                 close(fd);
 205                 return;
 206         }
 207
 208         vid = vhost_new_device();
 209         if (vid == -1) {
 210                 goto err;
 211         }
 212
 213         size = strnlen(vsocket->path, PATH_MAX);
 214         vhost_set_ifname(vid, vsocket->path, size);
 215
 216         vhost_set_builtin_virtio_net(vid, vsocket->use_builtin_virtio_net);
 217
 218         vhost_attach_vdpa_device(vid, vsocket->vdpa_dev_id);
 219
 220         if (vsocket->dequeue_zero_copy)
 221                 vhost_enable_dequeue_zero_copy(vid);
 222
 223         RTE_LOG(INFO, VHOST_CONFIG, "new device, handle is %d\n", vid);
 224
 225         if (vsocket->notify_ops->new_connection) {
 226                 ret = vsocket->notify_ops->new_connection(vid);
 227                 if (ret < 0) {
 228                         RTE_LOG(ERR, VHOST_CONFIG,
 229                                 "failed to add vhost user connection with fd %d\n",
 230                                 fd);
 231                         goto err;
 232                 }
 233         }
 234
 235         conn->connfd = fd;
 236         conn->vsocket = vsocket;
 237         conn->vid = vid;
 238         ret = fdset_add(&vhost_user.fdset, fd, vhost_user_read_cb,
 239                         NULL, conn);
 240         if (ret < 0) {
 241                 RTE_LOG(ERR, VHOST_CONFIG,
 242                         "failed to add fd %d into vhost server fdset\n",
 243                         fd);
 244
 245                 if (vsocket->notify_ops->destroy_connection)
 246                         vsocket->notify_ops->destroy_connection(conn->vid);
 247
 248                 goto err;
 249         }
 250
 251         pthread_mutex_lock(&vsocket->conn_mutex);
 252         TAILQ_INSERT_TAIL(&vsocket->conn_list, conn, next);
 253         pthread_mutex_unlock(&vsocket->conn_mutex);
 254
 255         fdset_pipe_notify(&vhost_user.fdset);
 256         return;
 257
 258 err:
 259         free(conn);
 260         close(fd);
 261 }
 262
 263 /* call back when there is new vhost-user connection from client  */
 264 static void
 265 vhost_user_server_new_connection(int fd, void *dat, int *remove __rte_unused)
 266 {
 267         struct vhost_user_socket *vsocket = dat;
 268
 269         fd = accept(fd, NULL, NULL);
 270         if (fd < 0)
 271                 return;
 272
 273         RTE_LOG(INFO, VHOST_CONFIG, "new vhost user connection is %d\n", fd);
 274         vhost_user_add_connection(fd, vsocket);
 275 }
 276
 277 static void
 278 vhost_user_read_cb(int connfd, void *dat, int *remove)
 279 {
 280         struct vhost_user_connection *conn = dat;
 281         struct vhost_user_socket *vsocket = conn->vsocket;
 282         int ret;
 283
 284         ret = vhost_user_msg_handler(conn->vid, connfd);
 285         if (ret < 0) {
 286                 close(connfd);
 287                 *remove = 1;
 288                 vhost_destroy_device(conn->vid);
 289
 290                 if (vsocket->notify_ops->destroy_connection)
 291                         vsocket->notify_ops->destroy_connection(conn->vid);
 292
 293                 pthread_mutex_lock(&vsocket->conn_mutex);
 294                 TAILQ_REMOVE(&vsocket->conn_list, conn, next);
 295                 pthread_mutex_unlock(&vsocket->conn_mutex);
 296
 297                 free(conn);
 298
 299                 if (vsocket->reconnect) {
 300                         create_unix_socket(vsocket);
 301                         vhost_user_start_client(vsocket);
 302                 }
 303         }
 304 }
 305
 306 static int
 307 create_unix_socket(struct vhost_user_socket *vsocket)
 308 {
 309         int fd;
 310         struct sockaddr_un *un = &vsocket->un;
 311
 312         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 313         if (fd < 0)
 314                 return -1;
 315         RTE_LOG(INFO, VHOST_CONFIG, "vhost-user %s: socket created, fd: %d\n",
 316                 vsocket->is_server ? "server" : "client", fd);
 317
 318         if (!vsocket->is_server && fcntl(fd, F_SETFL, O_NONBLOCK)) {
 319                 RTE_LOG(ERR, VHOST_CONFIG,
 320                         "vhost-user: can't set nonblocking mode for socket, fd: "
 321                         "%d (%s)\n", fd, strerror(errno));
 322                 close(fd);
 323                 return -1;
 324         }
 325
 326         memset(un, 0, sizeof(*un));
 327         un->sun_family = AF_UNIX;
 328         strncpy(un->sun_path, vsocket->path, sizeof(un->sun_path));
 329         un->sun_path[sizeof(un->sun_path) - 1] = '\0';
 330
 331         vsocket->socket_fd = fd;
 332         return 0;
 333 }
 334
 335 static int
 336 vhost_user_start_server(struct vhost_user_socket *vsocket)
 337 {
 338         int ret;
 339         int fd = vsocket->socket_fd;
 340         const char *path = vsocket->path;
 341
 342         /*
 343          * bind () may fail if the socket file with the same name already
 344          * exists. But the library obviously should not delete the file
 345          * provided by the user, since we can not be sure that it is not
 346          * being used by other applications. Moreover, many applications form
 347          * socket names based on user input, which is prone to errors.
 348          *
 349          * The user must ensure that the socket does not exist before
 350          * registering the vhost driver in server mode.
 351          */
 352         ret = bind(fd, (struct sockaddr *)&vsocket->un, sizeof(vsocket->un));
 353         if (ret < 0) {
 354                 RTE_LOG(ERR, VHOST_CONFIG,
 355                         "failed to bind to %s: %s; remove it and try again\n",
 356                         path, strerror(errno));
 357                 goto err;
 358         }
 359         RTE_LOG(INFO, VHOST_CONFIG, "bind to %s\n", path);
 360
 361         ret = listen(fd, MAX_VIRTIO_BACKLOG);
 362         if (ret < 0)
 363                 goto err;
 364
 365         ret = fdset_add(&vhost_user.fdset, fd, vhost_user_server_new_connection,
 366                   NULL, vsocket);
 367         if (ret < 0) {
 368                 RTE_LOG(ERR, VHOST_CONFIG,
 369                         "failed to add listen fd %d to vhost server fdset\n",
 370                         fd);
 371                 goto err;
 372         }
 373
 374         return 0;
 375
 376 err:
 377         close(fd);
 378         return -1;
 379 }
 380
 381 struct vhost_user_reconnect {
 382         struct sockaddr_un un;
 383         int fd;
 384         struct vhost_user_socket *vsocket;
 385
 386         TAILQ_ENTRY(vhost_user_reconnect) next;
 387 };
 388
 389 TAILQ_HEAD(vhost_user_reconnect_tailq_list, vhost_user_reconnect);
 390 struct vhost_user_reconnect_list {
 391         struct vhost_user_reconnect_tailq_list head;
 392         pthread_mutex_t mutex;
 393 };
 394
 395 static struct vhost_user_reconnect_list reconn_list;
 396 static pthread_t reconn_tid;
 397
 398 static int
 399 vhost_user_connect_nonblock(int fd, struct sockaddr *un, size_t sz)
 400 {
 401         int ret, flags;
 402
 403         ret = connect(fd, un, sz);
 404         if (ret < 0 && errno != EISCONN)
 405                 return -1;
 406
 407         flags = fcntl(fd, F_GETFL, 0);
 408         if (flags < 0) {
 409                 RTE_LOG(ERR, VHOST_CONFIG,
 410                         "can't get flags for connfd %d\n", fd);
 411                 return -2;
 412         }
 413         if ((flags & O_NONBLOCK) && fcntl(fd, F_SETFL, flags & ~O_NONBLOCK)) {
 414                 RTE_LOG(ERR, VHOST_CONFIG,
 415                                 "can't disable nonblocking on fd %d\n", fd);
 416                 return -2;
 417         }
 418         return 0;
 419 }
 420
 421 static void *
 422 vhost_user_client_reconnect(void *arg __rte_unused)
 423 {
 424         int ret;
 425         struct vhost_user_reconnect *reconn, *next;
 426
 427         while (1) {
 428                 pthread_mutex_lock(&reconn_list.mutex);
 429
 430                 /*
 431                  * An equal implementation of TAILQ_FOREACH_SAFE,
 432                  * which does not exist on all platforms.
 433                  */
 434                 for (reconn = TAILQ_FIRST(&reconn_list.head);
 435                      reconn != NULL; reconn = next) {
 436                         next = TAILQ_NEXT(reconn, next);
 437
 438                         ret = vhost_user_connect_nonblock(reconn->fd,
 439                                                 (struct sockaddr *)&reconn->un,
 440                                                 sizeof(reconn->un));
 441                         if (ret == -2) {
 442                                 close(reconn->fd);
 443                                 RTE_LOG(ERR, VHOST_CONFIG,
 444                                         "reconnection for fd %d failed\n",
 445                                         reconn->fd);
 446                                 goto remove_fd;
 447                         }
 448                         if (ret == -1)
 449                                 continue;
 450
 451                         RTE_LOG(INFO, VHOST_CONFIG,
 452                                 "%s: connected\n", reconn->vsocket->path);
 453                         vhost_user_add_connection(reconn->fd, reconn->vsocket);
 454 remove_fd:
 455                         TAILQ_REMOVE(&reconn_list.head, reconn, next);
 456                         free(reconn);
 457                 }
 458
 459                 pthread_mutex_unlock(&reconn_list.mutex);
 460                 sleep(1);
 461         }
 462
 463         return NULL;
 464 }
 465
 466 static int
 467 vhost_user_reconnect_init(void)
 468 {
 469         int ret;
 470         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 471
 472         ret = pthread_mutex_init(&reconn_list.mutex, NULL);
 473         if (ret < 0) {
 474                 RTE_LOG(ERR, VHOST_CONFIG, "failed to initialize mutex");
 475                 return ret;
 476         }
 477         TAILQ_INIT(&reconn_list.head);
 478
 479         ret = rte_ctrl_thread_create(&reconn_tid, NULL,
 480                              vhost_user_client_reconnect, NULL);
 481         if (ret != 0) {
 482                 RTE_LOG(ERR, VHOST_CONFIG, "failed to create reconnect thread");
 483                 if (pthread_mutex_destroy(&reconn_list.mutex)) {
 484                         RTE_LOG(ERR, VHOST_CONFIG,
 485                                 "failed to destroy reconnect mutex");
 486                 }
 487         } else {
 488                 snprintf(thread_name, sizeof(thread_name),
 489                          "vhost-reconn");
 490
 491                 if (rte_thread_setname(reconn_tid, thread_name)) {
 492                         RTE_LOG(DEBUG, VHOST_CONFIG,
 493                                 "failed to set reconnect thread name");
 494                 }
 495         }
 496
 497         return ret;
 498 }
 499
 500 static int
 501 vhost_user_start_client(struct vhost_user_socket *vsocket)
 502 {
 503         int ret;
 504         int fd = vsocket->socket_fd;
 505         const char *path = vsocket->path;
 506         struct vhost_user_reconnect *reconn;
 507
 508         ret = vhost_user_connect_nonblock(fd, (struct sockaddr *)&vsocket->un,
 509                                           sizeof(vsocket->un));
 510         if (ret == 0) {
 511                 vhost_user_add_connection(fd, vsocket);
 512                 return 0;
 513         }
 514
 515         RTE_LOG(WARNING, VHOST_CONFIG,
 516                 "failed to connect to %s: %s\n",
 517                 path, strerror(errno));
 518
 519         if (ret == -2 || !vsocket->reconnect) {
 520                 close(fd);
 521                 return -1;
 522         }
 523
 524         RTE_LOG(INFO, VHOST_CONFIG, "%s: reconnecting...\n", path);
 525         reconn = malloc(sizeof(*reconn));
 526         if (reconn == NULL) {
 527                 RTE_LOG(ERR, VHOST_CONFIG,
 528                         "failed to allocate memory for reconnect\n");
 529                 close(fd);
 530                 return -1;
 531         }
 532         reconn->un = vsocket->un;
 533         reconn->fd = fd;
 534         reconn->vsocket = vsocket;
 535         pthread_mutex_lock(&reconn_list.mutex);
 536         TAILQ_INSERT_TAIL(&reconn_list.head, reconn, next);
 537         pthread_mutex_unlock(&reconn_list.mutex);
 538
 539         return 0;
 540 }
 541
 542 static struct vhost_user_socket *
 543 find_vhost_user_socket(const char *path)
 544 {
 545         int i;
 546
 547         for (i = 0; i < vhost_user.vsocket_cnt; i++) {
 548                 struct vhost_user_socket *vsocket = vhost_user.vsockets[i];
 549
 550                 if (!strcmp(vsocket->path, path))
 551                         return vsocket;
 552         }
 553
 554         return NULL;
 555 }
 556
 557 int
 558 rte_vhost_driver_attach_vdpa_device(const char *path, int did)
 559 {
 560         struct vhost_user_socket *vsocket;
 561
 562         if (rte_vdpa_get_device(did) == NULL)
 563                 return -1;
 564
 565         pthread_mutex_lock(&vhost_user.mutex);
 566         vsocket = find_vhost_user_socket(path);
 567         if (vsocket)
 568                 vsocket->vdpa_dev_id = did;
 569         pthread_mutex_unlock(&vhost_user.mutex);
 570
 571         return vsocket ? 0 : -1;
 572 }
 573
 574 int
 575 rte_vhost_driver_detach_vdpa_device(const char *path)
 576 {
 577         struct vhost_user_socket *vsocket;
 578
 579         pthread_mutex_lock(&vhost_user.mutex);
 580         vsocket = find_vhost_user_socket(path);
 581         if (vsocket)
 582                 vsocket->vdpa_dev_id = -1;
 583         pthread_mutex_unlock(&vhost_user.mutex);
 584
 585         return vsocket ? 0 : -1;
 586 }
 587
 588 int
 589 rte_vhost_driver_get_vdpa_device_id(const char *path)
 590 {
 591         struct vhost_user_socket *vsocket;
 592         int did = -1;
 593
 594         pthread_mutex_lock(&vhost_user.mutex);
 595         vsocket = find_vhost_user_socket(path);
 596         if (vsocket)
 597                 did = vsocket->vdpa_dev_id;
 598         pthread_mutex_unlock(&vhost_user.mutex);
 599
 600         return did;
 601 }
 602
 603 int
 604 rte_vhost_driver_disable_features(const char *path, uint64_t features)
 605 {
 606         struct vhost_user_socket *vsocket;
 607
 608         pthread_mutex_lock(&vhost_user.mutex);
 609         vsocket = find_vhost_user_socket(path);
 610
 611         /* Note that use_builtin_virtio_net is not affected by this function
 612          * since callers may want to selectively disable features of the
 613          * built-in vhost net device backend.
 614          */
 615
 616         if (vsocket)
 617                 vsocket->features &= ~features;
 618         pthread_mutex_unlock(&vhost_user.mutex);
 619
 620         return vsocket ? 0 : -1;
 621 }
 622
 623 int
 624 rte_vhost_driver_enable_features(const char *path, uint64_t features)
 625 {
 626         struct vhost_user_socket *vsocket;
 627
 628         pthread_mutex_lock(&vhost_user.mutex);
 629         vsocket = find_vhost_user_socket(path);
 630         if (vsocket) {
 631                 if ((vsocket->supported_features & features) != features) {
 632                         /*
 633                          * trying to enable features the driver doesn't
 634                          * support.
 635                          */
 636                         pthread_mutex_unlock(&vhost_user.mutex);
 637                         return -1;
 638                 }
 639                 vsocket->features |= features;
 640         }
 641         pthread_mutex_unlock(&vhost_user.mutex);
 642
 643         return vsocket ? 0 : -1;
 644 }
 645
 646 int
 647 rte_vhost_driver_set_features(const char *path, uint64_t features)
 648 {
 649         struct vhost_user_socket *vsocket;
 650
 651         pthread_mutex_lock(&vhost_user.mutex);
 652         vsocket = find_vhost_user_socket(path);
 653         if (vsocket) {
 654                 vsocket->supported_features = features;
 655                 vsocket->features = features;
 656
 657                 /* Anyone setting feature bits is implementing their own vhost
 658                  * device backend.
 659                  */
 660                 vsocket->use_builtin_virtio_net = false;
 661         }
 662         pthread_mutex_unlock(&vhost_user.mutex);
 663
 664         return vsocket ? 0 : -1;
 665 }
 666
 667 int
 668 rte_vhost_driver_get_features(const char *path, uint64_t *features)
 669 {
 670         struct vhost_user_socket *vsocket;
 671         uint64_t vdpa_features;
 672         struct rte_vdpa_device *vdpa_dev;
 673         int did = -1;
 674         int ret = 0;
 675
 676         pthread_mutex_lock(&vhost_user.mutex);
 677         vsocket = find_vhost_user_socket(path);
 678         if (!vsocket) {
 679                 RTE_LOG(ERR, VHOST_CONFIG,
 680                         "socket file %s is not registered yet.\n", path);
 681                 ret = -1;
 682                 goto unlock_exit;
 683         }
 684
 685         did = vsocket->vdpa_dev_id;
 686         vdpa_dev = rte_vdpa_get_device(did);
 687         if (!vdpa_dev || !vdpa_dev->ops->get_features) {
 688                 *features = vsocket->features;
 689                 goto unlock_exit;
 690         }
 691
 692         if (vdpa_dev->ops->get_features(did, &vdpa_features) < 0) {
 693                 RTE_LOG(ERR, VHOST_CONFIG,
 694                                 "failed to get vdpa features "
 695                                 "for socket file %s.\n", path);
 696                 ret = -1;
 697                 goto unlock_exit;
 698         }
 699
 700         *features = vsocket->features & vdpa_features;
 701
 702 unlock_exit:
 703         pthread_mutex_unlock(&vhost_user.mutex);
 704         return ret;
 705 }
 706
 707 int
 708 rte_vhost_driver_get_protocol_features(const char *path,
 709                 uint64_t *protocol_features)
 710 {
 711         struct vhost_user_socket *vsocket;
 712         uint64_t vdpa_protocol_features;
 713         struct rte_vdpa_device *vdpa_dev;
 714         int did = -1;
 715         int ret = 0;
 716
 717         pthread_mutex_lock(&vhost_user.mutex);
 718         vsocket = find_vhost_user_socket(path);
 719         if (!vsocket) {
 720                 RTE_LOG(ERR, VHOST_CONFIG,
 721                         "socket file %s is not registered yet.\n", path);
 722                 ret = -1;
 723                 goto unlock_exit;
 724         }
 725
 726         did = vsocket->vdpa_dev_id;
 727         vdpa_dev = rte_vdpa_get_device(did);
 728         if (!vdpa_dev || !vdpa_dev->ops->get_protocol_features) {
 729                 *protocol_features = VHOST_USER_PROTOCOL_FEATURES;
 730                 goto unlock_exit;
 731         }
 732
 733         if (vdpa_dev->ops->get_protocol_features(did,
 734                                 &vdpa_protocol_features) < 0) {
 735                 RTE_LOG(ERR, VHOST_CONFIG,
 736                                 "failed to get vdpa protocol features "
 737                                 "for socket file %s.\n", path);
 738                 ret = -1;
 739                 goto unlock_exit;
 740         }
 741
 742         *protocol_features = VHOST_USER_PROTOCOL_FEATURES
 743                 & vdpa_protocol_features;
 744
 745 unlock_exit:
 746         pthread_mutex_unlock(&vhost_user.mutex);
 747         return ret;
 748 }
 749
 750 int
 751 rte_vhost_driver_get_queue_num(const char *path, uint32_t *queue_num)
 752 {
 753         struct vhost_user_socket *vsocket;
 754         uint32_t vdpa_queue_num;
 755         struct rte_vdpa_device *vdpa_dev;
 756         int did = -1;
 757         int ret = 0;
 758
 759         pthread_mutex_lock(&vhost_user.mutex);
 760         vsocket = find_vhost_user_socket(path);
 761         if (!vsocket) {
 762                 RTE_LOG(ERR, VHOST_CONFIG,
 763                         "socket file %s is not registered yet.\n", path);
 764                 ret = -1;
 765                 goto unlock_exit;
 766         }
 767
 768         did = vsocket->vdpa_dev_id;
 769         vdpa_dev = rte_vdpa_get_device(did);
 770         if (!vdpa_dev || !vdpa_dev->ops->get_queue_num) {
 771                 *queue_num = VHOST_MAX_QUEUE_PAIRS;
 772                 goto unlock_exit;
 773         }
 774
 775         if (vdpa_dev->ops->get_queue_num(did, &vdpa_queue_num) < 0) {
 776                 RTE_LOG(ERR, VHOST_CONFIG,
 777                                 "failed to get vdpa queue number "
 778                                 "for socket file %s.\n", path);
 779                 ret = -1;
 780                 goto unlock_exit;
 781         }
 782
 783         *queue_num = RTE_MIN((uint32_t)VHOST_MAX_QUEUE_PAIRS, vdpa_queue_num);
 784
 785 unlock_exit:
 786         pthread_mutex_unlock(&vhost_user.mutex);
 787         return ret;
 788 }
 789
 790 /*
 791  * Register a new vhost-user socket; here we could act as server
 792  * (the default case), or client (when RTE_VHOST_USER_CLIENT) flag
 793  * is set.
 794  */
 795 int
 796 rte_vhost_driver_register(const char *path, uint64_t flags)
 797 {
 798         int ret = -1;
 799         struct vhost_user_socket *vsocket;
 800
 801         if (!path)
 802                 return -1;
 803
 804         pthread_mutex_lock(&vhost_user.mutex);
 805
 806         if (vhost_user.vsocket_cnt == MAX_VHOST_SOCKET) {
 807                 RTE_LOG(ERR, VHOST_CONFIG,
 808                         "error: the number of vhost sockets reaches maximum\n");
 809                 goto out;
 810         }
 811
 812         vsocket = malloc(sizeof(struct vhost_user_socket));
 813         if (!vsocket)
 814                 goto out;
 815         memset(vsocket, 0, sizeof(struct vhost_user_socket));
 816         vsocket->path = strdup(path);
 817         if (vsocket->path == NULL) {
 818                 RTE_LOG(ERR, VHOST_CONFIG,
 819                         "error: failed to copy socket path string\n");
 820                 free(vsocket);
 821                 goto out;
 822         }
 823         TAILQ_INIT(&vsocket->conn_list);
 824         ret = pthread_mutex_init(&vsocket->conn_mutex, NULL);
 825         if (ret) {
 826                 RTE_LOG(ERR, VHOST_CONFIG,
 827                         "error: failed to init connection mutex\n");
 828                 goto out_free;
 829         }
 830         vsocket->dequeue_zero_copy = flags & RTE_VHOST_USER_DEQUEUE_ZERO_COPY;
 831
 832         /*
 833          * Set the supported features correctly for the builtin vhost-user
 834          * net driver.
 835          *
 836          * Applications know nothing about features the builtin virtio net
 837          * driver (virtio_net.c) supports, thus it's not possible for them
 838          * to invoke rte_vhost_driver_set_features(). To workaround it, here
 839          * we set it unconditionally. If the application want to implement
 840          * another vhost-user driver (say SCSI), it should call the
 841          * rte_vhost_driver_set_features(), which will overwrite following
 842          * two values.
 843          */
 844         vsocket->use_builtin_virtio_net = true;
 845         vsocket->supported_features = VIRTIO_NET_SUPPORTED_FEATURES;
 846         vsocket->features           = VIRTIO_NET_SUPPORTED_FEATURES;
 847
 848         if (!(flags & RTE_VHOST_USER_IOMMU_SUPPORT)) {
 849                 vsocket->supported_features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 850                 vsocket->features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 851         }
 852
 853         if ((flags & RTE_VHOST_USER_CLIENT) != 0) {
 854                 vsocket->reconnect = !(flags & RTE_VHOST_USER_NO_RECONNECT);
 855                 if (vsocket->reconnect && reconn_tid == 0) {
 856                         if (vhost_user_reconnect_init() != 0)
 857                                 goto out_mutex;
 858                 }
 859         } else {
 860                 vsocket->is_server = true;
 861         }
 862         ret = create_unix_socket(vsocket);
 863         if (ret < 0) {
 864                 goto out_mutex;
 865         }
 866
 867         vhost_user.vsockets[vhost_user.vsocket_cnt++] = vsocket;
 868
 869         pthread_mutex_unlock(&vhost_user.mutex);
 870         return ret;
 871
 872 out_mutex:
 873         if (pthread_mutex_destroy(&vsocket->conn_mutex)) {
 874                 RTE_LOG(ERR, VHOST_CONFIG,
 875                         "error: failed to destroy connection mutex\n");
 876         }
 877 out_free:
 878         free(vsocket->path);
 879         free(vsocket);
 880 out:
 881         pthread_mutex_unlock(&vhost_user.mutex);
 882
 883         return ret;
 884 }
 885
 886 static bool
 887 vhost_user_remove_reconnect(struct vhost_user_socket *vsocket)
 888 {
 889         int found = false;
 890         struct vhost_user_reconnect *reconn, *next;
 891
 892         pthread_mutex_lock(&reconn_list.mutex);
 893
 894         for (reconn = TAILQ_FIRST(&reconn_list.head);
 895              reconn != NULL; reconn = next) {
 896                 next = TAILQ_NEXT(reconn, next);
 897
 898                 if (reconn->vsocket == vsocket) {
 899                         TAILQ_REMOVE(&reconn_list.head, reconn, next);
 900                         close(reconn->fd);
 901                         free(reconn);
 902                         found = true;
 903                         break;
 904                 }
 905         }
 906         pthread_mutex_unlock(&reconn_list.mutex);
 907         return found;
 908 }
 909
 910 /**
 911  * Unregister the specified vhost socket
 912  */
 913 int
 914 rte_vhost_driver_unregister(const char *path)
 915 {
 916         int i;
 917         int count;
 918         struct vhost_user_connection *conn, *next;
 919
 920         pthread_mutex_lock(&vhost_user.mutex);
 921
 922         for (i = 0; i < vhost_user.vsocket_cnt; i++) {
 923                 struct vhost_user_socket *vsocket = vhost_user.vsockets[i];
 924
 925                 if (!strcmp(vsocket->path, path)) {
 926                         if (vsocket->is_server) {
 927                                 fdset_del(&vhost_user.fdset, vsocket->socket_fd);
 928                                 close(vsocket->socket_fd);
 929                                 unlink(path);
 930                         } else if (vsocket->reconnect) {
 931                                 vhost_user_remove_reconnect(vsocket);
 932                         }
 933
 934                         pthread_mutex_lock(&vsocket->conn_mutex);
 935                         for (conn = TAILQ_FIRST(&vsocket->conn_list);
 936                              conn != NULL;
 937                              conn = next) {
 938                                 next = TAILQ_NEXT(conn, next);
 939
 940                                 fdset_del(&vhost_user.fdset, conn->connfd);
 941                                 RTE_LOG(INFO, VHOST_CONFIG,
 942                                         "free connfd = %d for device '%s'\n",
 943                                         conn->connfd, path);
 944                                 close(conn->connfd);
 945                                 vhost_destroy_device(conn->vid);
 946                                 TAILQ_REMOVE(&vsocket->conn_list, conn, next);
 947                                 free(conn);
 948                         }
 949                         pthread_mutex_unlock(&vsocket->conn_mutex);
 950
 951                         pthread_mutex_destroy(&vsocket->conn_mutex);
 952                         free(vsocket->path);
 953                         free(vsocket);
 954
 955                         count = --vhost_user.vsocket_cnt;
 956                         vhost_user.vsockets[i] = vhost_user.vsockets[count];
 957                         vhost_user.vsockets[count] = NULL;
 958                         pthread_mutex_unlock(&vhost_user.mutex);
 959
 960                         return 0;
 961                 }
 962         }
 963         pthread_mutex_unlock(&vhost_user.mutex);
 964
 965         return -1;
 966 }
 967
 968 /*
 969  * Register ops so that we can add/remove device to data core.
 970  */
 971 int
 972 rte_vhost_driver_callback_register(const char *path,
 973         struct vhost_device_ops const * const ops)
 974 {
 975         struct vhost_user_socket *vsocket;
 976
 977         pthread_mutex_lock(&vhost_user.mutex);
 978         vsocket = find_vhost_user_socket(path);
 979         if (vsocket)
 980                 vsocket->notify_ops = ops;
 981         pthread_mutex_unlock(&vhost_user.mutex);
 982
 983         return vsocket ? 0 : -1;
 984 }
 985
 986 struct vhost_device_ops const *
 987 vhost_driver_callback_get(const char *path)
 988 {
 989         struct vhost_user_socket *vsocket;
 990
 991         pthread_mutex_lock(&vhost_user.mutex);
 992         vsocket = find_vhost_user_socket(path);
 993         pthread_mutex_unlock(&vhost_user.mutex);
 994
 995         return vsocket ? vsocket->notify_ops : NULL;
 996 }
 997
 998 int
 999 rte_vhost_driver_start(const char *path)
1000 {
1001         struct vhost_user_socket *vsocket;
1002         static pthread_t fdset_tid;
1003         char thread_name[RTE_MAX_THREAD_NAME_LEN];
1004
1005         pthread_mutex_lock(&vhost_user.mutex);
1006         vsocket = find_vhost_user_socket(path);
1007         pthread_mutex_unlock(&vhost_user.mutex);
1008
1009         if (!vsocket)
1010                 return -1;
1011
1012         if (fdset_tid == 0) {
1013                 /**
1014                  * create a pipe which will be waited by poll and notified to
1015                  * rebuild the wait list of poll.
1016                  */
1017                 if (fdset_pipe_init(&vhost_user.fdset) < 0) {
1018                         RTE_LOG(ERR, VHOST_CONFIG,
1019                                 "failed to create pipe for vhost fdset\n");
1020                         return -1;
1021                 }
1022
1023                 int ret = rte_ctrl_thread_create(&fdset_tid, NULL,
1024                         fdset_event_dispatch, &vhost_user.fdset);
1025                 if (ret != 0) {
1026                         RTE_LOG(ERR, VHOST_CONFIG,
1027                                 "failed to create fdset handling thread");
1028
1029                         fdset_pipe_uninit(&vhost_user.fdset);
1030                         return -1;
1031                 } else {
1032                         snprintf(thread_name, sizeof(thread_name),
1033                                  "vhost-events");
1034
1035                         if (rte_thread_setname(fdset_tid, thread_name)) {
1036                                 RTE_LOG(DEBUG, VHOST_CONFIG,
1037                                         "failed to set vhost-event thread name");
1038                         }
1039                 }
1040         }
1041
1042         if (vsocket->is_server)
1043                 return vhost_user_start_server(vsocket);
1044         else
1045                 return vhost_user_start_client(vsocket);
1046 }