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         if (vsocket->dequeue_zero_copy)
 219                 vhost_enable_dequeue_zero_copy(vid);
 220
 221         RTE_LOG(INFO, VHOST_CONFIG, "new device, handle is %d\n", vid);
 222
 223         if (vsocket->notify_ops->new_connection) {
 224                 ret = vsocket->notify_ops->new_connection(vid);
 225                 if (ret < 0) {
 226                         RTE_LOG(ERR, VHOST_CONFIG,
 227                                 "failed to add vhost user connection with fd %d\n",
 228                                 fd);
 229                         goto err;
 230                 }
 231         }
 232
 233         conn->connfd = fd;
 234         conn->vsocket = vsocket;
 235         conn->vid = vid;
 236         ret = fdset_add(&vhost_user.fdset, fd, vhost_user_read_cb,
 237                         NULL, conn);
 238         if (ret < 0) {
 239                 RTE_LOG(ERR, VHOST_CONFIG,
 240                         "failed to add fd %d into vhost server fdset\n",
 241                         fd);
 242
 243                 if (vsocket->notify_ops->destroy_connection)
 244                         vsocket->notify_ops->destroy_connection(conn->vid);
 245
 246                 goto err;
 247         }
 248
 249         pthread_mutex_lock(&vsocket->conn_mutex);
 250         TAILQ_INSERT_TAIL(&vsocket->conn_list, conn, next);
 251         pthread_mutex_unlock(&vsocket->conn_mutex);
 252
 253         fdset_pipe_notify(&vhost_user.fdset);
 254         return;
 255
 256 err:
 257         free(conn);
 258         close(fd);
 259 }
 260
 261 /* call back when there is new vhost-user connection from client  */
 262 static void
 263 vhost_user_server_new_connection(int fd, void *dat, int *remove __rte_unused)
 264 {
 265         struct vhost_user_socket *vsocket = dat;
 266
 267         fd = accept(fd, NULL, NULL);
 268         if (fd < 0)
 269                 return;
 270
 271         RTE_LOG(INFO, VHOST_CONFIG, "new vhost user connection is %d\n", fd);
 272         vhost_user_add_connection(fd, vsocket);
 273 }
 274
 275 static void
 276 vhost_user_read_cb(int connfd, void *dat, int *remove)
 277 {
 278         struct vhost_user_connection *conn = dat;
 279         struct vhost_user_socket *vsocket = conn->vsocket;
 280         int ret;
 281
 282         ret = vhost_user_msg_handler(conn->vid, connfd);
 283         if (ret < 0) {
 284                 close(connfd);
 285                 *remove = 1;
 286                 vhost_destroy_device(conn->vid);
 287
 288                 if (vsocket->notify_ops->destroy_connection)
 289                         vsocket->notify_ops->destroy_connection(conn->vid);
 290
 291                 pthread_mutex_lock(&vsocket->conn_mutex);
 292                 TAILQ_REMOVE(&vsocket->conn_list, conn, next);
 293                 pthread_mutex_unlock(&vsocket->conn_mutex);
 294
 295                 free(conn);
 296
 297                 if (vsocket->reconnect) {
 298                         create_unix_socket(vsocket);
 299                         vhost_user_start_client(vsocket);
 300                 }
 301         }
 302 }
 303
 304 static int
 305 create_unix_socket(struct vhost_user_socket *vsocket)
 306 {
 307         int fd;
 308         struct sockaddr_un *un = &vsocket->un;
 309
 310         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 311         if (fd < 0)
 312                 return -1;
 313         RTE_LOG(INFO, VHOST_CONFIG, "vhost-user %s: socket created, fd: %d\n",
 314                 vsocket->is_server ? "server" : "client", fd);
 315
 316         if (!vsocket->is_server && fcntl(fd, F_SETFL, O_NONBLOCK)) {
 317                 RTE_LOG(ERR, VHOST_CONFIG,
 318                         "vhost-user: can't set nonblocking mode for socket, fd: "
 319                         "%d (%s)\n", fd, strerror(errno));
 320                 close(fd);
 321                 return -1;
 322         }
 323
 324         memset(un, 0, sizeof(*un));
 325         un->sun_family = AF_UNIX;
 326         strncpy(un->sun_path, vsocket->path, sizeof(un->sun_path));
 327         un->sun_path[sizeof(un->sun_path) - 1] = '\0';
 328
 329         vsocket->socket_fd = fd;
 330         return 0;
 331 }
 332
 333 static int
 334 vhost_user_start_server(struct vhost_user_socket *vsocket)
 335 {
 336         int ret;
 337         int fd = vsocket->socket_fd;
 338         const char *path = vsocket->path;
 339
 340         /*
 341          * bind () may fail if the socket file with the same name already
 342          * exists. But the library obviously should not delete the file
 343          * provided by the user, since we can not be sure that it is not
 344          * being used by other applications. Moreover, many applications form
 345          * socket names based on user input, which is prone to errors.
 346          *
 347          * The user must ensure that the socket does not exist before
 348          * registering the vhost driver in server mode.
 349          */
 350         ret = bind(fd, (struct sockaddr *)&vsocket->un, sizeof(vsocket->un));
 351         if (ret < 0) {
 352                 RTE_LOG(ERR, VHOST_CONFIG,
 353                         "failed to bind to %s: %s; remove it and try again\n",
 354                         path, strerror(errno));
 355                 goto err;
 356         }
 357         RTE_LOG(INFO, VHOST_CONFIG, "bind to %s\n", path);
 358
 359         ret = listen(fd, MAX_VIRTIO_BACKLOG);
 360         if (ret < 0)
 361                 goto err;
 362
 363         ret = fdset_add(&vhost_user.fdset, fd, vhost_user_server_new_connection,
 364                   NULL, vsocket);
 365         if (ret < 0) {
 366                 RTE_LOG(ERR, VHOST_CONFIG,
 367                         "failed to add listen fd %d to vhost server fdset\n",
 368                         fd);
 369                 goto err;
 370         }
 371
 372         return 0;
 373
 374 err:
 375         close(fd);
 376         return -1;
 377 }
 378
 379 struct vhost_user_reconnect {
 380         struct sockaddr_un un;
 381         int fd;
 382         struct vhost_user_socket *vsocket;
 383
 384         TAILQ_ENTRY(vhost_user_reconnect) next;
 385 };
 386
 387 TAILQ_HEAD(vhost_user_reconnect_tailq_list, vhost_user_reconnect);
 388 struct vhost_user_reconnect_list {
 389         struct vhost_user_reconnect_tailq_list head;
 390         pthread_mutex_t mutex;
 391 };
 392
 393 static struct vhost_user_reconnect_list reconn_list;
 394 static pthread_t reconn_tid;
 395
 396 static int
 397 vhost_user_connect_nonblock(int fd, struct sockaddr *un, size_t sz)
 398 {
 399         int ret, flags;
 400
 401         ret = connect(fd, un, sz);
 402         if (ret < 0 && errno != EISCONN)
 403                 return -1;
 404
 405         flags = fcntl(fd, F_GETFL, 0);
 406         if (flags < 0) {
 407                 RTE_LOG(ERR, VHOST_CONFIG,
 408                         "can't get flags for connfd %d\n", fd);
 409                 return -2;
 410         }
 411         if ((flags & O_NONBLOCK) && fcntl(fd, F_SETFL, flags & ~O_NONBLOCK)) {
 412                 RTE_LOG(ERR, VHOST_CONFIG,
 413                                 "can't disable nonblocking on fd %d\n", fd);
 414                 return -2;
 415         }
 416         return 0;
 417 }
 418
 419 static void *
 420 vhost_user_client_reconnect(void *arg __rte_unused)
 421 {
 422         int ret;
 423         struct vhost_user_reconnect *reconn, *next;
 424
 425         while (1) {
 426                 pthread_mutex_lock(&reconn_list.mutex);
 427
 428                 /*
 429                  * An equal implementation of TAILQ_FOREACH_SAFE,
 430                  * which does not exist on all platforms.
 431                  */
 432                 for (reconn = TAILQ_FIRST(&reconn_list.head);
 433                      reconn != NULL; reconn = next) {
 434                         next = TAILQ_NEXT(reconn, next);
 435
 436                         ret = vhost_user_connect_nonblock(reconn->fd,
 437                                                 (struct sockaddr *)&reconn->un,
 438                                                 sizeof(reconn->un));
 439                         if (ret == -2) {
 440                                 close(reconn->fd);
 441                                 RTE_LOG(ERR, VHOST_CONFIG,
 442                                         "reconnection for fd %d failed\n",
 443                                         reconn->fd);
 444                                 goto remove_fd;
 445                         }
 446                         if (ret == -1)
 447                                 continue;
 448
 449                         RTE_LOG(INFO, VHOST_CONFIG,
 450                                 "%s: connected\n", reconn->vsocket->path);
 451                         vhost_user_add_connection(reconn->fd, reconn->vsocket);
 452 remove_fd:
 453                         TAILQ_REMOVE(&reconn_list.head, reconn, next);
 454                         free(reconn);
 455                 }
 456
 457                 pthread_mutex_unlock(&reconn_list.mutex);
 458                 sleep(1);
 459         }
 460
 461         return NULL;
 462 }
 463
 464 static int
 465 vhost_user_reconnect_init(void)
 466 {
 467         int ret;
 468         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 469
 470         ret = pthread_mutex_init(&reconn_list.mutex, NULL);
 471         if (ret < 0) {
 472                 RTE_LOG(ERR, VHOST_CONFIG, "failed to initialize mutex");
 473                 return ret;
 474         }
 475         TAILQ_INIT(&reconn_list.head);
 476
 477         ret = pthread_create(&reconn_tid, NULL,
 478                              vhost_user_client_reconnect, NULL);
 479         if (ret != 0) {
 480                 RTE_LOG(ERR, VHOST_CONFIG, "failed to create reconnect thread");
 481                 if (pthread_mutex_destroy(&reconn_list.mutex)) {
 482                         RTE_LOG(ERR, VHOST_CONFIG,
 483                                 "failed to destroy reconnect mutex");
 484                 }
 485         } else {
 486                 snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN,
 487                          "vhost-reconn");
 488
 489                 if (rte_thread_setname(reconn_tid, thread_name)) {
 490                         RTE_LOG(DEBUG, VHOST_CONFIG,
 491                                 "failed to set reconnect thread name");
 492                 }
 493         }
 494
 495         return ret;
 496 }
 497
 498 static int
 499 vhost_user_start_client(struct vhost_user_socket *vsocket)
 500 {
 501         int ret;
 502         int fd = vsocket->socket_fd;
 503         const char *path = vsocket->path;
 504         struct vhost_user_reconnect *reconn;
 505
 506         ret = vhost_user_connect_nonblock(fd, (struct sockaddr *)&vsocket->un,
 507                                           sizeof(vsocket->un));
 508         if (ret == 0) {
 509                 vhost_user_add_connection(fd, vsocket);
 510                 return 0;
 511         }
 512
 513         RTE_LOG(WARNING, VHOST_CONFIG,
 514                 "failed to connect to %s: %s\n",
 515                 path, strerror(errno));
 516
 517         if (ret == -2 || !vsocket->reconnect) {
 518                 close(fd);
 519                 return -1;
 520         }
 521
 522         RTE_LOG(INFO, VHOST_CONFIG, "%s: reconnecting...\n", path);
 523         reconn = malloc(sizeof(*reconn));
 524         if (reconn == NULL) {
 525                 RTE_LOG(ERR, VHOST_CONFIG,
 526                         "failed to allocate memory for reconnect\n");
 527                 close(fd);
 528                 return -1;
 529         }
 530         reconn->un = vsocket->un;
 531         reconn->fd = fd;
 532         reconn->vsocket = vsocket;
 533         pthread_mutex_lock(&reconn_list.mutex);
 534         TAILQ_INSERT_TAIL(&reconn_list.head, reconn, next);
 535         pthread_mutex_unlock(&reconn_list.mutex);
 536
 537         return 0;
 538 }
 539
 540 static struct vhost_user_socket *
 541 find_vhost_user_socket(const char *path)
 542 {
 543         int i;
 544
 545         for (i = 0; i < vhost_user.vsocket_cnt; i++) {
 546                 struct vhost_user_socket *vsocket = vhost_user.vsockets[i];
 547
 548                 if (!strcmp(vsocket->path, path))
 549                         return vsocket;
 550         }
 551
 552         return NULL;
 553 }
 554
 555 int
 556 rte_vhost_driver_attach_vdpa_device(const char *path, int did)
 557 {
 558         struct vhost_user_socket *vsocket;
 559
 560         if (rte_vdpa_get_device(did) == NULL)
 561                 return -1;
 562
 563         pthread_mutex_lock(&vhost_user.mutex);
 564         vsocket = find_vhost_user_socket(path);
 565         if (vsocket)
 566                 vsocket->vdpa_dev_id = did;
 567         pthread_mutex_unlock(&vhost_user.mutex);
 568
 569         return vsocket ? 0 : -1;
 570 }
 571
 572 int
 573 rte_vhost_driver_detach_vdpa_device(const char *path)
 574 {
 575         struct vhost_user_socket *vsocket;
 576
 577         pthread_mutex_lock(&vhost_user.mutex);
 578         vsocket = find_vhost_user_socket(path);
 579         if (vsocket)
 580                 vsocket->vdpa_dev_id = -1;
 581         pthread_mutex_unlock(&vhost_user.mutex);
 582
 583         return vsocket ? 0 : -1;
 584 }
 585
 586 int
 587 rte_vhost_driver_get_vdpa_device_id(const char *path)
 588 {
 589         struct vhost_user_socket *vsocket;
 590         int did = -1;
 591
 592         pthread_mutex_lock(&vhost_user.mutex);
 593         vsocket = find_vhost_user_socket(path);
 594         if (vsocket)
 595                 did = vsocket->vdpa_dev_id;
 596         pthread_mutex_unlock(&vhost_user.mutex);
 597
 598         return did;
 599 }
 600
 601 int
 602 rte_vhost_driver_disable_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
 609         /* Note that use_builtin_virtio_net is not affected by this function
 610          * since callers may want to selectively disable features of the
 611          * built-in vhost net device backend.
 612          */
 613
 614         if (vsocket)
 615                 vsocket->features &= ~features;
 616         pthread_mutex_unlock(&vhost_user.mutex);
 617
 618         return vsocket ? 0 : -1;
 619 }
 620
 621 int
 622 rte_vhost_driver_enable_features(const char *path, uint64_t features)
 623 {
 624         struct vhost_user_socket *vsocket;
 625
 626         pthread_mutex_lock(&vhost_user.mutex);
 627         vsocket = find_vhost_user_socket(path);
 628         if (vsocket) {
 629                 if ((vsocket->supported_features & features) != features) {
 630                         /*
 631                          * trying to enable features the driver doesn't
 632                          * support.
 633                          */
 634                         pthread_mutex_unlock(&vhost_user.mutex);
 635                         return -1;
 636                 }
 637                 vsocket->features |= features;
 638         }
 639         pthread_mutex_unlock(&vhost_user.mutex);
 640
 641         return vsocket ? 0 : -1;
 642 }
 643
 644 int
 645 rte_vhost_driver_set_features(const char *path, uint64_t features)
 646 {
 647         struct vhost_user_socket *vsocket;
 648
 649         pthread_mutex_lock(&vhost_user.mutex);
 650         vsocket = find_vhost_user_socket(path);
 651         if (vsocket) {
 652                 vsocket->supported_features = features;
 653                 vsocket->features = features;
 654
 655                 /* Anyone setting feature bits is implementing their own vhost
 656                  * device backend.
 657                  */
 658                 vsocket->use_builtin_virtio_net = false;
 659         }
 660         pthread_mutex_unlock(&vhost_user.mutex);
 661
 662         return vsocket ? 0 : -1;
 663 }
 664
 665 int
 666 rte_vhost_driver_get_features(const char *path, uint64_t *features)
 667 {
 668         struct vhost_user_socket *vsocket;
 669
 670         pthread_mutex_lock(&vhost_user.mutex);
 671         vsocket = find_vhost_user_socket(path);
 672         if (vsocket)
 673                 *features = vsocket->features;
 674         pthread_mutex_unlock(&vhost_user.mutex);
 675
 676         if (!vsocket) {
 677                 RTE_LOG(ERR, VHOST_CONFIG,
 678                         "socket file %s is not registered yet.\n", path);
 679                 return -1;
 680         } else {
 681                 return 0;
 682         }
 683 }
 684
 685 /*
 686  * Register a new vhost-user socket; here we could act as server
 687  * (the default case), or client (when RTE_VHOST_USER_CLIENT) flag
 688  * is set.
 689  */
 690 int
 691 rte_vhost_driver_register(const char *path, uint64_t flags)
 692 {
 693         int ret = -1;
 694         struct vhost_user_socket *vsocket;
 695
 696         if (!path)
 697                 return -1;
 698
 699         pthread_mutex_lock(&vhost_user.mutex);
 700
 701         if (vhost_user.vsocket_cnt == MAX_VHOST_SOCKET) {
 702                 RTE_LOG(ERR, VHOST_CONFIG,
 703                         "error: the number of vhost sockets reaches maximum\n");
 704                 goto out;
 705         }
 706
 707         vsocket = malloc(sizeof(struct vhost_user_socket));
 708         if (!vsocket)
 709                 goto out;
 710         memset(vsocket, 0, sizeof(struct vhost_user_socket));
 711         vsocket->path = strdup(path);
 712         if (vsocket->path == NULL) {
 713                 RTE_LOG(ERR, VHOST_CONFIG,
 714                         "error: failed to copy socket path string\n");
 715                 free(vsocket);
 716                 goto out;
 717         }
 718         TAILQ_INIT(&vsocket->conn_list);
 719         ret = pthread_mutex_init(&vsocket->conn_mutex, NULL);
 720         if (ret) {
 721                 RTE_LOG(ERR, VHOST_CONFIG,
 722                         "error: failed to init connection mutex\n");
 723                 goto out_free;
 724         }
 725         vsocket->dequeue_zero_copy = flags & RTE_VHOST_USER_DEQUEUE_ZERO_COPY;
 726
 727         /*
 728          * Set the supported features correctly for the builtin vhost-user
 729          * net driver.
 730          *
 731          * Applications know nothing about features the builtin virtio net
 732          * driver (virtio_net.c) supports, thus it's not possible for them
 733          * to invoke rte_vhost_driver_set_features(). To workaround it, here
 734          * we set it unconditionally. If the application want to implement
 735          * another vhost-user driver (say SCSI), it should call the
 736          * rte_vhost_driver_set_features(), which will overwrite following
 737          * two values.
 738          */
 739         vsocket->use_builtin_virtio_net = true;
 740         vsocket->supported_features = VIRTIO_NET_SUPPORTED_FEATURES;
 741         vsocket->features           = VIRTIO_NET_SUPPORTED_FEATURES;
 742
 743         if (!(flags & RTE_VHOST_USER_IOMMU_SUPPORT)) {
 744                 vsocket->supported_features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 745                 vsocket->features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 746         }
 747
 748         if ((flags & RTE_VHOST_USER_CLIENT) != 0) {
 749                 vsocket->reconnect = !(flags & RTE_VHOST_USER_NO_RECONNECT);
 750                 if (vsocket->reconnect && reconn_tid == 0) {
 751                         if (vhost_user_reconnect_init() != 0)
 752                                 goto out_mutex;
 753                 }
 754         } else {
 755                 vsocket->is_server = true;
 756         }
 757         ret = create_unix_socket(vsocket);
 758         if (ret < 0) {
 759                 goto out_mutex;
 760         }
 761
 762         vhost_user.vsockets[vhost_user.vsocket_cnt++] = vsocket;
 763
 764         pthread_mutex_unlock(&vhost_user.mutex);
 765         return ret;
 766
 767 out_mutex:
 768         if (pthread_mutex_destroy(&vsocket->conn_mutex)) {
 769                 RTE_LOG(ERR, VHOST_CONFIG,
 770                         "error: failed to destroy connection mutex\n");
 771         }
 772 out_free:
 773         free(vsocket->path);
 774         free(vsocket);
 775 out:
 776         pthread_mutex_unlock(&vhost_user.mutex);
 777
 778         return ret;
 779 }
 780
 781 static bool
 782 vhost_user_remove_reconnect(struct vhost_user_socket *vsocket)
 783 {
 784         int found = false;
 785         struct vhost_user_reconnect *reconn, *next;
 786
 787         pthread_mutex_lock(&reconn_list.mutex);
 788
 789         for (reconn = TAILQ_FIRST(&reconn_list.head);
 790              reconn != NULL; reconn = next) {
 791                 next = TAILQ_NEXT(reconn, next);
 792
 793                 if (reconn->vsocket == vsocket) {
 794                         TAILQ_REMOVE(&reconn_list.head, reconn, next);
 795                         close(reconn->fd);
 796                         free(reconn);
 797                         found = true;
 798                         break;
 799                 }
 800         }
 801         pthread_mutex_unlock(&reconn_list.mutex);
 802         return found;
 803 }
 804
 805 /**
 806  * Unregister the specified vhost socket
 807  */
 808 int
 809 rte_vhost_driver_unregister(const char *path)
 810 {
 811         int i;
 812         int count;
 813         struct vhost_user_connection *conn, *next;
 814
 815         pthread_mutex_lock(&vhost_user.mutex);
 816
 817         for (i = 0; i < vhost_user.vsocket_cnt; i++) {
 818                 struct vhost_user_socket *vsocket = vhost_user.vsockets[i];
 819
 820                 if (!strcmp(vsocket->path, path)) {
 821                         if (vsocket->is_server) {
 822                                 fdset_del(&vhost_user.fdset, vsocket->socket_fd);
 823                                 close(vsocket->socket_fd);
 824                                 unlink(path);
 825                         } else if (vsocket->reconnect) {
 826                                 vhost_user_remove_reconnect(vsocket);
 827                         }
 828
 829                         pthread_mutex_lock(&vsocket->conn_mutex);
 830                         for (conn = TAILQ_FIRST(&vsocket->conn_list);
 831                              conn != NULL;
 832                              conn = next) {
 833                                 next = TAILQ_NEXT(conn, next);
 834
 835                                 fdset_del(&vhost_user.fdset, conn->connfd);
 836                                 RTE_LOG(INFO, VHOST_CONFIG,
 837                                         "free connfd = %d for device '%s'\n",
 838                                         conn->connfd, path);
 839                                 close(conn->connfd);
 840                                 vhost_destroy_device(conn->vid);
 841                                 TAILQ_REMOVE(&vsocket->conn_list, conn, next);
 842                                 free(conn);
 843                         }
 844                         pthread_mutex_unlock(&vsocket->conn_mutex);
 845
 846                         pthread_mutex_destroy(&vsocket->conn_mutex);
 847                         free(vsocket->path);
 848                         free(vsocket);
 849
 850                         count = --vhost_user.vsocket_cnt;
 851                         vhost_user.vsockets[i] = vhost_user.vsockets[count];
 852                         vhost_user.vsockets[count] = NULL;
 853                         pthread_mutex_unlock(&vhost_user.mutex);
 854
 855                         return 0;
 856                 }
 857         }
 858         pthread_mutex_unlock(&vhost_user.mutex);
 859
 860         return -1;
 861 }
 862
 863 /*
 864  * Register ops so that we can add/remove device to data core.
 865  */
 866 int
 867 rte_vhost_driver_callback_register(const char *path,
 868         struct vhost_device_ops const * const ops)
 869 {
 870         struct vhost_user_socket *vsocket;
 871
 872         pthread_mutex_lock(&vhost_user.mutex);
 873         vsocket = find_vhost_user_socket(path);
 874         if (vsocket)
 875                 vsocket->notify_ops = ops;
 876         pthread_mutex_unlock(&vhost_user.mutex);
 877
 878         return vsocket ? 0 : -1;
 879 }
 880
 881 struct vhost_device_ops const *
 882 vhost_driver_callback_get(const char *path)
 883 {
 884         struct vhost_user_socket *vsocket;
 885
 886         pthread_mutex_lock(&vhost_user.mutex);
 887         vsocket = find_vhost_user_socket(path);
 888         pthread_mutex_unlock(&vhost_user.mutex);
 889
 890         return vsocket ? vsocket->notify_ops : NULL;
 891 }
 892
 893 int
 894 rte_vhost_driver_start(const char *path)
 895 {
 896         struct vhost_user_socket *vsocket;
 897         static pthread_t fdset_tid;
 898         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 899
 900         pthread_mutex_lock(&vhost_user.mutex);
 901         vsocket = find_vhost_user_socket(path);
 902         pthread_mutex_unlock(&vhost_user.mutex);
 903
 904         if (!vsocket)
 905                 return -1;
 906
 907         if (fdset_tid == 0) {
 908                 /**
 909                  * create a pipe which will be waited by poll and notified to
 910                  * rebuild the wait list of poll.
 911                  */
 912                 if (fdset_pipe_init(&vhost_user.fdset) < 0) {
 913                         RTE_LOG(ERR, VHOST_CONFIG,
 914                                 "failed to create pipe for vhost fdset\n");
 915                         return -1;
 916                 }
 917
 918                 int ret = pthread_create(&fdset_tid, NULL, fdset_event_dispatch,
 919                                      &vhost_user.fdset);
 920                 if (ret != 0) {
 921                         RTE_LOG(ERR, VHOST_CONFIG,
 922                                 "failed to create fdset handling thread");
 923
 924                         fdset_pipe_uninit(&vhost_user.fdset);
 925                         return -1;
 926                 } else {
 927                         snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN,
 928                                  "vhost-events");
 929
 930                         if (rte_thread_setname(fdset_tid, thread_name)) {
 931                                 RTE_LOG(DEBUG, VHOST_CONFIG,
 932                                         "failed to set vhost-event thread name");
 933                         }
 934                 }
 935         }
 936
 937         if (vsocket->is_server)
 938                 return vhost_user_start_server(vsocket);
 939         else
 940                 return vhost_user_start_client(vsocket);
 941 }