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
 471         ret = pthread_mutex_init(&reconn_list.mutex, NULL);
 472         if (ret < 0) {
 473                 RTE_LOG(ERR, VHOST_CONFIG, "failed to initialize mutex");
 474                 return ret;
 475         }
 476         TAILQ_INIT(&reconn_list.head);
 477
 478         ret = rte_ctrl_thread_create(&reconn_tid, "vhost_reconn", NULL,
 479                              vhost_user_client_reconnect, NULL);
 480         if (ret != 0) {
 481                 RTE_LOG(ERR, VHOST_CONFIG, "failed to create reconnect thread");
 482                 if (pthread_mutex_destroy(&reconn_list.mutex)) {
 483                         RTE_LOG(ERR, VHOST_CONFIG,
 484                                 "failed to destroy reconnect mutex");
 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_attach_vdpa_device(const char *path, int did)
 550 {
 551         struct vhost_user_socket *vsocket;
 552
 553         if (rte_vdpa_get_device(did) == NULL)
 554                 return -1;
 555
 556         pthread_mutex_lock(&vhost_user.mutex);
 557         vsocket = find_vhost_user_socket(path);
 558         if (vsocket)
 559                 vsocket->vdpa_dev_id = did;
 560         pthread_mutex_unlock(&vhost_user.mutex);
 561
 562         return vsocket ? 0 : -1;
 563 }
 564
 565 int
 566 rte_vhost_driver_detach_vdpa_device(const char *path)
 567 {
 568         struct vhost_user_socket *vsocket;
 569
 570         pthread_mutex_lock(&vhost_user.mutex);
 571         vsocket = find_vhost_user_socket(path);
 572         if (vsocket)
 573                 vsocket->vdpa_dev_id = -1;
 574         pthread_mutex_unlock(&vhost_user.mutex);
 575
 576         return vsocket ? 0 : -1;
 577 }
 578
 579 int
 580 rte_vhost_driver_get_vdpa_device_id(const char *path)
 581 {
 582         struct vhost_user_socket *vsocket;
 583         int did = -1;
 584
 585         pthread_mutex_lock(&vhost_user.mutex);
 586         vsocket = find_vhost_user_socket(path);
 587         if (vsocket)
 588                 did = vsocket->vdpa_dev_id;
 589         pthread_mutex_unlock(&vhost_user.mutex);
 590
 591         return did;
 592 }
 593
 594 int
 595 rte_vhost_driver_disable_features(const char *path, uint64_t features)
 596 {
 597         struct vhost_user_socket *vsocket;
 598
 599         pthread_mutex_lock(&vhost_user.mutex);
 600         vsocket = find_vhost_user_socket(path);
 601
 602         /* Note that use_builtin_virtio_net is not affected by this function
 603          * since callers may want to selectively disable features of the
 604          * built-in vhost net device backend.
 605          */
 606
 607         if (vsocket)
 608                 vsocket->features &= ~features;
 609         pthread_mutex_unlock(&vhost_user.mutex);
 610
 611         return vsocket ? 0 : -1;
 612 }
 613
 614 int
 615 rte_vhost_driver_enable_features(const char *path, uint64_t features)
 616 {
 617         struct vhost_user_socket *vsocket;
 618
 619         pthread_mutex_lock(&vhost_user.mutex);
 620         vsocket = find_vhost_user_socket(path);
 621         if (vsocket) {
 622                 if ((vsocket->supported_features & features) != features) {
 623                         /*
 624                          * trying to enable features the driver doesn't
 625                          * support.
 626                          */
 627                         pthread_mutex_unlock(&vhost_user.mutex);
 628                         return -1;
 629                 }
 630                 vsocket->features |= features;
 631         }
 632         pthread_mutex_unlock(&vhost_user.mutex);
 633
 634         return vsocket ? 0 : -1;
 635 }
 636
 637 int
 638 rte_vhost_driver_set_features(const char *path, uint64_t features)
 639 {
 640         struct vhost_user_socket *vsocket;
 641
 642         pthread_mutex_lock(&vhost_user.mutex);
 643         vsocket = find_vhost_user_socket(path);
 644         if (vsocket) {
 645                 vsocket->supported_features = features;
 646                 vsocket->features = features;
 647
 648                 /* Anyone setting feature bits is implementing their own vhost
 649                  * device backend.
 650                  */
 651                 vsocket->use_builtin_virtio_net = false;
 652         }
 653         pthread_mutex_unlock(&vhost_user.mutex);
 654
 655         return vsocket ? 0 : -1;
 656 }
 657
 658 int
 659 rte_vhost_driver_get_features(const char *path, uint64_t *features)
 660 {
 661         struct vhost_user_socket *vsocket;
 662         uint64_t vdpa_features;
 663         struct rte_vdpa_device *vdpa_dev;
 664         int did = -1;
 665         int ret = 0;
 666
 667         pthread_mutex_lock(&vhost_user.mutex);
 668         vsocket = find_vhost_user_socket(path);
 669         if (!vsocket) {
 670                 RTE_LOG(ERR, VHOST_CONFIG,
 671                         "socket file %s is not registered yet.\n", path);
 672                 ret = -1;
 673                 goto unlock_exit;
 674         }
 675
 676         did = vsocket->vdpa_dev_id;
 677         vdpa_dev = rte_vdpa_get_device(did);
 678         if (!vdpa_dev || !vdpa_dev->ops->get_features) {
 679                 *features = vsocket->features;
 680                 goto unlock_exit;
 681         }
 682
 683         if (vdpa_dev->ops->get_features(did, &vdpa_features) < 0) {
 684                 RTE_LOG(ERR, VHOST_CONFIG,
 685                                 "failed to get vdpa features "
 686                                 "for socket file %s.\n", path);
 687                 ret = -1;
 688                 goto unlock_exit;
 689         }
 690
 691         *features = vsocket->features & vdpa_features;
 692
 693 unlock_exit:
 694         pthread_mutex_unlock(&vhost_user.mutex);
 695         return ret;
 696 }
 697
 698 int
 699 rte_vhost_driver_get_protocol_features(const char *path,
 700                 uint64_t *protocol_features)
 701 {
 702         struct vhost_user_socket *vsocket;
 703         uint64_t vdpa_protocol_features;
 704         struct rte_vdpa_device *vdpa_dev;
 705         int did = -1;
 706         int ret = 0;
 707
 708         pthread_mutex_lock(&vhost_user.mutex);
 709         vsocket = find_vhost_user_socket(path);
 710         if (!vsocket) {
 711                 RTE_LOG(ERR, VHOST_CONFIG,
 712                         "socket file %s is not registered yet.\n", path);
 713                 ret = -1;
 714                 goto unlock_exit;
 715         }
 716
 717         did = vsocket->vdpa_dev_id;
 718         vdpa_dev = rte_vdpa_get_device(did);
 719         if (!vdpa_dev || !vdpa_dev->ops->get_protocol_features) {
 720                 *protocol_features = VHOST_USER_PROTOCOL_FEATURES;
 721                 goto unlock_exit;
 722         }
 723
 724         if (vdpa_dev->ops->get_protocol_features(did,
 725                                 &vdpa_protocol_features) < 0) {
 726                 RTE_LOG(ERR, VHOST_CONFIG,
 727                                 "failed to get vdpa protocol features "
 728                                 "for socket file %s.\n", path);
 729                 ret = -1;
 730                 goto unlock_exit;
 731         }
 732
 733         *protocol_features = VHOST_USER_PROTOCOL_FEATURES
 734                 & vdpa_protocol_features;
 735
 736 unlock_exit:
 737         pthread_mutex_unlock(&vhost_user.mutex);
 738         return ret;
 739 }
 740
 741 int
 742 rte_vhost_driver_get_queue_num(const char *path, uint32_t *queue_num)
 743 {
 744         struct vhost_user_socket *vsocket;
 745         uint32_t vdpa_queue_num;
 746         struct rte_vdpa_device *vdpa_dev;
 747         int did = -1;
 748         int ret = 0;
 749
 750         pthread_mutex_lock(&vhost_user.mutex);
 751         vsocket = find_vhost_user_socket(path);
 752         if (!vsocket) {
 753                 RTE_LOG(ERR, VHOST_CONFIG,
 754                         "socket file %s is not registered yet.\n", path);
 755                 ret = -1;
 756                 goto unlock_exit;
 757         }
 758
 759         did = vsocket->vdpa_dev_id;
 760         vdpa_dev = rte_vdpa_get_device(did);
 761         if (!vdpa_dev || !vdpa_dev->ops->get_queue_num) {
 762                 *queue_num = VHOST_MAX_QUEUE_PAIRS;
 763                 goto unlock_exit;
 764         }
 765
 766         if (vdpa_dev->ops->get_queue_num(did, &vdpa_queue_num) < 0) {
 767                 RTE_LOG(ERR, VHOST_CONFIG,
 768                                 "failed to get vdpa queue number "
 769                                 "for socket file %s.\n", path);
 770                 ret = -1;
 771                 goto unlock_exit;
 772         }
 773
 774         *queue_num = RTE_MIN((uint32_t)VHOST_MAX_QUEUE_PAIRS, vdpa_queue_num);
 775
 776 unlock_exit:
 777         pthread_mutex_unlock(&vhost_user.mutex);
 778         return ret;
 779 }
 780
 781 /*
 782  * Register a new vhost-user socket; here we could act as server
 783  * (the default case), or client (when RTE_VHOST_USER_CLIENT) flag
 784  * is set.
 785  */
 786 int
 787 rte_vhost_driver_register(const char *path, uint64_t flags)
 788 {
 789         int ret = -1;
 790         struct vhost_user_socket *vsocket;
 791
 792         if (!path)
 793                 return -1;
 794
 795         pthread_mutex_lock(&vhost_user.mutex);
 796
 797         if (vhost_user.vsocket_cnt == MAX_VHOST_SOCKET) {
 798                 RTE_LOG(ERR, VHOST_CONFIG,
 799                         "error: the number of vhost sockets reaches maximum\n");
 800                 goto out;
 801         }
 802
 803         vsocket = malloc(sizeof(struct vhost_user_socket));
 804         if (!vsocket)
 805                 goto out;
 806         memset(vsocket, 0, sizeof(struct vhost_user_socket));
 807         vsocket->path = strdup(path);
 808         if (vsocket->path == NULL) {
 809                 RTE_LOG(ERR, VHOST_CONFIG,
 810                         "error: failed to copy socket path string\n");
 811                 free(vsocket);
 812                 goto out;
 813         }
 814         TAILQ_INIT(&vsocket->conn_list);
 815         ret = pthread_mutex_init(&vsocket->conn_mutex, NULL);
 816         if (ret) {
 817                 RTE_LOG(ERR, VHOST_CONFIG,
 818                         "error: failed to init connection mutex\n");
 819                 goto out_free;
 820         }
 821         vsocket->dequeue_zero_copy = flags & RTE_VHOST_USER_DEQUEUE_ZERO_COPY;
 822
 823         /*
 824          * Set the supported features correctly for the builtin vhost-user
 825          * net driver.
 826          *
 827          * Applications know nothing about features the builtin virtio net
 828          * driver (virtio_net.c) supports, thus it's not possible for them
 829          * to invoke rte_vhost_driver_set_features(). To workaround it, here
 830          * we set it unconditionally. If the application want to implement
 831          * another vhost-user driver (say SCSI), it should call the
 832          * rte_vhost_driver_set_features(), which will overwrite following
 833          * two values.
 834          */
 835         vsocket->use_builtin_virtio_net = true;
 836         vsocket->supported_features = VIRTIO_NET_SUPPORTED_FEATURES;
 837         vsocket->features           = VIRTIO_NET_SUPPORTED_FEATURES;
 838
 839         if (!(flags & RTE_VHOST_USER_IOMMU_SUPPORT)) {
 840                 vsocket->supported_features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 841                 vsocket->features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 842         }
 843
 844         if ((flags & RTE_VHOST_USER_CLIENT) != 0) {
 845                 vsocket->reconnect = !(flags & RTE_VHOST_USER_NO_RECONNECT);
 846                 if (vsocket->reconnect && reconn_tid == 0) {
 847                         if (vhost_user_reconnect_init() != 0)
 848                                 goto out_mutex;
 849                 }
 850         } else {
 851                 vsocket->is_server = true;
 852         }
 853         ret = create_unix_socket(vsocket);
 854         if (ret < 0) {
 855                 goto out_mutex;
 856         }
 857
 858         vhost_user.vsockets[vhost_user.vsocket_cnt++] = vsocket;
 859
 860         pthread_mutex_unlock(&vhost_user.mutex);
 861         return ret;
 862
 863 out_mutex:
 864         if (pthread_mutex_destroy(&vsocket->conn_mutex)) {
 865                 RTE_LOG(ERR, VHOST_CONFIG,
 866                         "error: failed to destroy connection mutex\n");
 867         }
 868 out_free:
 869         free(vsocket->path);
 870         free(vsocket);
 871 out:
 872         pthread_mutex_unlock(&vhost_user.mutex);
 873
 874         return ret;
 875 }
 876
 877 static bool
 878 vhost_user_remove_reconnect(struct vhost_user_socket *vsocket)
 879 {
 880         int found = false;
 881         struct vhost_user_reconnect *reconn, *next;
 882
 883         pthread_mutex_lock(&reconn_list.mutex);
 884
 885         for (reconn = TAILQ_FIRST(&reconn_list.head);
 886              reconn != NULL; reconn = next) {
 887                 next = TAILQ_NEXT(reconn, next);
 888
 889                 if (reconn->vsocket == vsocket) {
 890                         TAILQ_REMOVE(&reconn_list.head, reconn, next);
 891                         close(reconn->fd);
 892                         free(reconn);
 893                         found = true;
 894                         break;
 895                 }
 896         }
 897         pthread_mutex_unlock(&reconn_list.mutex);
 898         return found;
 899 }
 900
 901 /**
 902  * Unregister the specified vhost socket
 903  */
 904 int
 905 rte_vhost_driver_unregister(const char *path)
 906 {
 907         int i;
 908         int count;
 909         struct vhost_user_connection *conn, *next;
 910
 911         pthread_mutex_lock(&vhost_user.mutex);
 912
 913         for (i = 0; i < vhost_user.vsocket_cnt; i++) {
 914                 struct vhost_user_socket *vsocket = vhost_user.vsockets[i];
 915
 916                 if (!strcmp(vsocket->path, path)) {
 917                         if (vsocket->is_server) {
 918                                 fdset_del(&vhost_user.fdset, vsocket->socket_fd);
 919                                 close(vsocket->socket_fd);
 920                                 unlink(path);
 921                         } else if (vsocket->reconnect) {
 922                                 vhost_user_remove_reconnect(vsocket);
 923                         }
 924
 925                         pthread_mutex_lock(&vsocket->conn_mutex);
 926                         for (conn = TAILQ_FIRST(&vsocket->conn_list);
 927                              conn != NULL;
 928                              conn = next) {
 929                                 next = TAILQ_NEXT(conn, next);
 930
 931                                 fdset_del(&vhost_user.fdset, conn->connfd);
 932                                 RTE_LOG(INFO, VHOST_CONFIG,
 933                                         "free connfd = %d for device '%s'\n",
 934                                         conn->connfd, path);
 935                                 close(conn->connfd);
 936                                 vhost_destroy_device(conn->vid);
 937                                 TAILQ_REMOVE(&vsocket->conn_list, conn, next);
 938                                 free(conn);
 939                         }
 940                         pthread_mutex_unlock(&vsocket->conn_mutex);
 941
 942                         pthread_mutex_destroy(&vsocket->conn_mutex);
 943                         free(vsocket->path);
 944                         free(vsocket);
 945
 946                         count = --vhost_user.vsocket_cnt;
 947                         vhost_user.vsockets[i] = vhost_user.vsockets[count];
 948                         vhost_user.vsockets[count] = NULL;
 949                         pthread_mutex_unlock(&vhost_user.mutex);
 950
 951                         return 0;
 952                 }
 953         }
 954         pthread_mutex_unlock(&vhost_user.mutex);
 955
 956         return -1;
 957 }
 958
 959 /*
 960  * Register ops so that we can add/remove device to data core.
 961  */
 962 int
 963 rte_vhost_driver_callback_register(const char *path,
 964         struct vhost_device_ops const * const ops)
 965 {
 966         struct vhost_user_socket *vsocket;
 967
 968         pthread_mutex_lock(&vhost_user.mutex);
 969         vsocket = find_vhost_user_socket(path);
 970         if (vsocket)
 971                 vsocket->notify_ops = ops;
 972         pthread_mutex_unlock(&vhost_user.mutex);
 973
 974         return vsocket ? 0 : -1;
 975 }
 976
 977 struct vhost_device_ops const *
 978 vhost_driver_callback_get(const char *path)
 979 {
 980         struct vhost_user_socket *vsocket;
 981
 982         pthread_mutex_lock(&vhost_user.mutex);
 983         vsocket = find_vhost_user_socket(path);
 984         pthread_mutex_unlock(&vhost_user.mutex);
 985
 986         return vsocket ? vsocket->notify_ops : NULL;
 987 }
 988
 989 int
 990 rte_vhost_driver_start(const char *path)
 991 {
 992         struct vhost_user_socket *vsocket;
 993         static pthread_t fdset_tid;
 994
 995         pthread_mutex_lock(&vhost_user.mutex);
 996         vsocket = find_vhost_user_socket(path);
 997         pthread_mutex_unlock(&vhost_user.mutex);
 998
 999         if (!vsocket)
1000                 return -1;
1001
1002         if (fdset_tid == 0) {
1003                 /**
1004                  * create a pipe which will be waited by poll and notified to
1005                  * rebuild the wait list of poll.
1006                  */
1007                 if (fdset_pipe_init(&vhost_user.fdset) < 0) {
1008                         RTE_LOG(ERR, VHOST_CONFIG,
1009                                 "failed to create pipe for vhost fdset\n");
1010                         return -1;
1011                 }
1012
1013                 int ret = rte_ctrl_thread_create(&fdset_tid,
1014                         "vhost-events", NULL, fdset_event_dispatch,
1015                         &vhost_user.fdset);
1016                 if (ret != 0) {
1017                         RTE_LOG(ERR, VHOST_CONFIG,
1018                                 "failed to create fdset handling thread");
1019
1020                         fdset_pipe_uninit(&vhost_user.fdset);
1021                         return -1;
1022                 }
1023         }
1024
1025         if (vsocket->is_server)
1026                 return vhost_user_start_server(vsocket);
1027         else
1028                 return vhost_user_start_client(vsocket);
1029 }