lib/librte_eal/common/eal_common_proc.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016-2018 Intel Corporation
   3  */
   4
   5 #include <dirent.h>
   6 #include <errno.h>
   7 #include <fcntl.h>
   8 #include <fnmatch.h>
   9 #include <inttypes.h>
  10 #include <libgen.h>
  11 #include <limits.h>
  12 #include <pthread.h>
  13 #include <stdio.h>
  14 #include <stdlib.h>
  15 #include <string.h>
  16 #include <sys/file.h>
  17 #include <sys/time.h>
  18 #include <sys/types.h>
  19 #include <sys/socket.h>
  20 #include <sys/un.h>
  21 #include <unistd.h>
  22
  23 #include <rte_alarm.h>
  24 #include <rte_common.h>
  25 #include <rte_cycles.h>
  26 #include <rte_eal.h>
  27 #include <rte_errno.h>
  28 #include <rte_lcore.h>
  29 #include <rte_log.h>
  30 #include <rte_tailq.h>
  31
  32 #include "eal_private.h"
  33 #include "eal_filesystem.h"
  34 #include "eal_internal_cfg.h"
  35
  36 static int mp_fd = -1;
  37 static char mp_filter[PATH_MAX];   /* Filter for secondary process sockets */
  38 static char mp_dir_path[PATH_MAX]; /* The directory path for all mp sockets */
  39 static pthread_mutex_t mp_mutex_action = PTHREAD_MUTEX_INITIALIZER;
  40 static char peer_name[PATH_MAX];
  41
  42 struct action_entry {
  43         TAILQ_ENTRY(action_entry) next;
  44         char action_name[RTE_MP_MAX_NAME_LEN];
  45         rte_mp_t action;
  46 };
  47
  48 /** Double linked list of actions. */
  49 TAILQ_HEAD(action_entry_list, action_entry);
  50
  51 static struct action_entry_list action_entry_list =
  52         TAILQ_HEAD_INITIALIZER(action_entry_list);
  53
  54 enum mp_type {
  55         MP_MSG, /* Share message with peers, will not block */
  56         MP_REQ, /* Request for information, Will block for a reply */
  57         MP_REP, /* Response to previously-received request */
  58         MP_IGN, /* Response telling requester to ignore this response */
  59 };
  60
  61 struct mp_msg_internal {
  62         int type;
  63         struct rte_mp_msg msg;
  64 };
  65
  66 struct async_request_param {
  67         rte_mp_async_reply_t clb;
  68         struct rte_mp_reply user_reply;
  69         struct timespec end;
  70         int n_responses_processed;
  71 };
  72
  73 struct pending_request {
  74         TAILQ_ENTRY(pending_request) next;
  75         enum {
  76                 REQUEST_TYPE_SYNC,
  77                 REQUEST_TYPE_ASYNC
  78         } type;
  79         char dst[PATH_MAX];
  80         struct rte_mp_msg *request;
  81         struct rte_mp_msg *reply;
  82         int reply_received;
  83         RTE_STD_C11
  84         union {
  85                 struct {
  86                         struct async_request_param *param;
  87                 } async;
  88                 struct {
  89                         pthread_cond_t cond;
  90                 } sync;
  91         };
  92 };
  93
  94 TAILQ_HEAD(pending_request_list, pending_request);
  95
  96 static struct {
  97         struct pending_request_list requests;
  98         pthread_mutex_t lock;
  99 } pending_requests = {
 100         .requests = TAILQ_HEAD_INITIALIZER(pending_requests.requests),
 101         .lock = PTHREAD_MUTEX_INITIALIZER,
 102         /**< used in async requests only */
 103 };
 104
 105 /* forward declarations */
 106 static int
 107 mp_send(struct rte_mp_msg *msg, const char *peer, int type);
 108
 109 /* for use with alarm callback */
 110 static void
 111 async_reply_handle(void *arg);
 112
 113 /* for use with process_msg */
 114 static struct pending_request *
 115 async_reply_handle_thread_unsafe(void *arg);
 116
 117 static void
 118 trigger_async_action(struct pending_request *req);
 119
 120 static struct pending_request *
 121 find_pending_request(const char *dst, const char *act_name)
 122 {
 123         struct pending_request *r;
 124
 125         TAILQ_FOREACH(r, &pending_requests.requests, next) {
 126                 if (!strcmp(r->dst, dst) &&
 127                     !strcmp(r->request->name, act_name))
 128                         break;
 129         }
 130
 131         return r;
 132 }
 133
 134 static void
 135 create_socket_path(const char *name, char *buf, int len)
 136 {
 137         const char *prefix = eal_mp_socket_path();
 138
 139         if (strlen(name) > 0)
 140                 snprintf(buf, len, "%s_%s", prefix, name);
 141         else
 142                 strlcpy(buf, prefix, len);
 143 }
 144
 145 int
 146 rte_eal_primary_proc_alive(const char *config_file_path)
 147 {
 148         int config_fd;
 149
 150         if (config_file_path)
 151                 config_fd = open(config_file_path, O_RDONLY);
 152         else {
 153                 const char *path;
 154
 155                 path = eal_runtime_config_path();
 156                 config_fd = open(path, O_RDONLY);
 157         }
 158         if (config_fd < 0)
 159                 return 0;
 160
 161         int ret = lockf(config_fd, F_TEST, 0);
 162         close(config_fd);
 163
 164         return !!ret;
 165 }
 166
 167 static struct action_entry *
 168 find_action_entry_by_name(const char *name)
 169 {
 170         struct action_entry *entry;
 171
 172         TAILQ_FOREACH(entry, &action_entry_list, next) {
 173                 if (strncmp(entry->action_name, name, RTE_MP_MAX_NAME_LEN) == 0)
 174                         break;
 175         }
 176
 177         return entry;
 178 }
 179
 180 static int
 181 validate_action_name(const char *name)
 182 {
 183         if (name == NULL) {
 184                 RTE_LOG(ERR, EAL, "Action name cannot be NULL\n");
 185                 rte_errno = EINVAL;
 186                 return -1;
 187         }
 188         if (strnlen(name, RTE_MP_MAX_NAME_LEN) == 0) {
 189                 RTE_LOG(ERR, EAL, "Length of action name is zero\n");
 190                 rte_errno = EINVAL;
 191                 return -1;
 192         }
 193         if (strnlen(name, RTE_MP_MAX_NAME_LEN) == RTE_MP_MAX_NAME_LEN) {
 194                 rte_errno = E2BIG;
 195                 return -1;
 196         }
 197         return 0;
 198 }
 199
 200 int
 201 rte_mp_action_register(const char *name, rte_mp_t action)
 202 {
 203         struct action_entry *entry;
 204         const struct internal_config *internal_conf =
 205                 eal_get_internal_configuration();
 206
 207         if (validate_action_name(name) != 0)
 208                 return -1;
 209
 210         if (internal_conf->no_shconf) {
 211                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
 212                 rte_errno = ENOTSUP;
 213                 return -1;
 214         }
 215
 216         entry = malloc(sizeof(struct action_entry));
 217         if (entry == NULL) {
 218                 rte_errno = ENOMEM;
 219                 return -1;
 220         }
 221         strlcpy(entry->action_name, name, sizeof(entry->action_name));
 222         entry->action = action;
 223
 224         pthread_mutex_lock(&mp_mutex_action);
 225         if (find_action_entry_by_name(name) != NULL) {
 226                 pthread_mutex_unlock(&mp_mutex_action);
 227                 rte_errno = EEXIST;
 228                 free(entry);
 229                 return -1;
 230         }
 231         TAILQ_INSERT_TAIL(&action_entry_list, entry, next);
 232         pthread_mutex_unlock(&mp_mutex_action);
 233         return 0;
 234 }
 235
 236 void
 237 rte_mp_action_unregister(const char *name)
 238 {
 239         struct action_entry *entry;
 240         const struct internal_config *internal_conf =
 241                 eal_get_internal_configuration();
 242
 243         if (validate_action_name(name) != 0)
 244                 return;
 245
 246         if (internal_conf->no_shconf) {
 247                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
 248                 return;
 249         }
 250
 251         pthread_mutex_lock(&mp_mutex_action);
 252         entry = find_action_entry_by_name(name);
 253         if (entry == NULL) {
 254                 pthread_mutex_unlock(&mp_mutex_action);
 255                 return;
 256         }
 257         TAILQ_REMOVE(&action_entry_list, entry, next);
 258         pthread_mutex_unlock(&mp_mutex_action);
 259         free(entry);
 260 }
 261
 262 static int
 263 read_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
 264 {
 265         int msglen;
 266         struct iovec iov;
 267         struct msghdr msgh;
 268         char control[CMSG_SPACE(sizeof(m->msg.fds))];
 269         struct cmsghdr *cmsg;
 270         int buflen = sizeof(*m) - sizeof(m->msg.fds);
 271
 272         memset(&msgh, 0, sizeof(msgh));
 273         iov.iov_base = m;
 274         iov.iov_len  = buflen;
 275
 276         msgh.msg_name = s;
 277         msgh.msg_namelen = sizeof(*s);
 278         msgh.msg_iov = &iov;
 279         msgh.msg_iovlen = 1;
 280         msgh.msg_control = control;
 281         msgh.msg_controllen = sizeof(control);
 282
 283         msglen = recvmsg(mp_fd, &msgh, 0);
 284         if (msglen < 0) {
 285                 RTE_LOG(ERR, EAL, "recvmsg failed, %s\n", strerror(errno));
 286                 return -1;
 287         }
 288
 289         if (msglen != buflen || (msgh.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
 290                 RTE_LOG(ERR, EAL, "truncated msg\n");
 291                 return -1;
 292         }
 293
 294         /* read auxiliary FDs if any */
 295         for (cmsg = CMSG_FIRSTHDR(&msgh); cmsg != NULL;
 296                 cmsg = CMSG_NXTHDR(&msgh, cmsg)) {
 297                 if ((cmsg->cmsg_level == SOL_SOCKET) &&
 298                         (cmsg->cmsg_type == SCM_RIGHTS)) {
 299                         memcpy(m->msg.fds, CMSG_DATA(cmsg), sizeof(m->msg.fds));
 300                         break;
 301                 }
 302         }
 303         /* sanity-check the response */
 304         if (m->msg.num_fds < 0 || m->msg.num_fds > RTE_MP_MAX_FD_NUM) {
 305                 RTE_LOG(ERR, EAL, "invalid number of fd's received\n");
 306                 return -1;
 307         }
 308         if (m->msg.len_param < 0 || m->msg.len_param > RTE_MP_MAX_PARAM_LEN) {
 309                 RTE_LOG(ERR, EAL, "invalid received data length\n");
 310                 return -1;
 311         }
 312         return 0;
 313 }
 314
 315 static void
 316 process_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
 317 {
 318         struct pending_request *pending_req;
 319         struct action_entry *entry;
 320         struct rte_mp_msg *msg = &m->msg;
 321         rte_mp_t action = NULL;
 322         const struct internal_config *internal_conf =
 323                 eal_get_internal_configuration();
 324
 325         RTE_LOG(DEBUG, EAL, "msg: %s\n", msg->name);
 326
 327         if (m->type == MP_REP || m->type == MP_IGN) {
 328                 struct pending_request *req = NULL;
 329
 330                 pthread_mutex_lock(&pending_requests.lock);
 331                 pending_req = find_pending_request(s->sun_path, msg->name);
 332                 if (pending_req) {
 333                         memcpy(pending_req->reply, msg, sizeof(*msg));
 334                         /* -1 indicates that we've been asked to ignore */
 335                         pending_req->reply_received =
 336                                 m->type == MP_REP ? 1 : -1;
 337
 338                         if (pending_req->type == REQUEST_TYPE_SYNC)
 339                                 pthread_cond_signal(&pending_req->sync.cond);
 340                         else if (pending_req->type == REQUEST_TYPE_ASYNC)
 341                                 req = async_reply_handle_thread_unsafe(
 342                                                 pending_req);
 343                 } else
 344                         RTE_LOG(ERR, EAL, "Drop mp reply: %s\n", msg->name);
 345                 pthread_mutex_unlock(&pending_requests.lock);
 346
 347                 if (req != NULL)
 348                         trigger_async_action(req);
 349                 return;
 350         }
 351
 352         pthread_mutex_lock(&mp_mutex_action);
 353         entry = find_action_entry_by_name(msg->name);
 354         if (entry != NULL)
 355                 action = entry->action;
 356         pthread_mutex_unlock(&mp_mutex_action);
 357
 358         if (!action) {
 359                 if (m->type == MP_REQ && !internal_conf->init_complete) {
 360                         /* if this is a request, and init is not yet complete,
 361                          * and callback wasn't registered, we should tell the
 362                          * requester to ignore our existence because we're not
 363                          * yet ready to process this request.
 364                          */
 365                         struct rte_mp_msg dummy;
 366
 367                         memset(&dummy, 0, sizeof(dummy));
 368                         strlcpy(dummy.name, msg->name, sizeof(dummy.name));
 369                         mp_send(&dummy, s->sun_path, MP_IGN);
 370                 } else {
 371                         RTE_LOG(ERR, EAL, "Cannot find action: %s\n",
 372                                 msg->name);
 373                 }
 374         } else if (action(msg, s->sun_path) < 0) {
 375                 RTE_LOG(ERR, EAL, "Fail to handle message: %s\n", msg->name);
 376         }
 377 }
 378
 379 static void *
 380 mp_handle(void *arg __rte_unused)
 381 {
 382         struct mp_msg_internal msg;
 383         struct sockaddr_un sa;
 384
 385         while (1) {
 386                 if (read_msg(&msg, &sa) == 0)
 387                         process_msg(&msg, &sa);
 388         }
 389
 390         return NULL;
 391 }
 392
 393 static int
 394 timespec_cmp(const struct timespec *a, const struct timespec *b)
 395 {
 396         if (a->tv_sec < b->tv_sec)
 397                 return -1;
 398         if (a->tv_sec > b->tv_sec)
 399                 return 1;
 400         if (a->tv_nsec < b->tv_nsec)
 401                 return -1;
 402         if (a->tv_nsec > b->tv_nsec)
 403                 return 1;
 404         return 0;
 405 }
 406
 407 enum async_action {
 408         ACTION_FREE, /**< free the action entry, but don't trigger callback */
 409         ACTION_TRIGGER /**< trigger callback, then free action entry */
 410 };
 411
 412 static enum async_action
 413 process_async_request(struct pending_request *sr, const struct timespec *now)
 414 {
 415         struct async_request_param *param;
 416         struct rte_mp_reply *reply;
 417         bool timeout, last_msg;
 418
 419         param = sr->async.param;
 420         reply = &param->user_reply;
 421
 422         /* did we timeout? */
 423         timeout = timespec_cmp(&param->end, now) <= 0;
 424
 425         /* if we received a response, adjust relevant data and copy mesasge. */
 426         if (sr->reply_received == 1 && sr->reply) {
 427                 struct rte_mp_msg *msg, *user_msgs, *tmp;
 428
 429                 msg = sr->reply;
 430                 user_msgs = reply->msgs;
 431
 432                 tmp = realloc(user_msgs, sizeof(*msg) *
 433                                 (reply->nb_received + 1));
 434                 if (!tmp) {
 435                         RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 436                                 sr->dst, sr->request->name);
 437                         /* this entry is going to be removed and its message
 438                          * dropped, but we don't want to leak memory, so
 439                          * continue.
 440                          */
 441                 } else {
 442                         user_msgs = tmp;
 443                         reply->msgs = user_msgs;
 444                         memcpy(&user_msgs[reply->nb_received],
 445                                         msg, sizeof(*msg));
 446                         reply->nb_received++;
 447                 }
 448
 449                 /* mark this request as processed */
 450                 param->n_responses_processed++;
 451         } else if (sr->reply_received == -1) {
 452                 /* we were asked to ignore this process */
 453                 reply->nb_sent--;
 454         } else if (timeout) {
 455                 /* count it as processed response, but don't increment
 456                  * nb_received.
 457                  */
 458                 param->n_responses_processed++;
 459         }
 460
 461         free(sr->reply);
 462
 463         last_msg = param->n_responses_processed == reply->nb_sent;
 464
 465         return last_msg ? ACTION_TRIGGER : ACTION_FREE;
 466 }
 467
 468 static void
 469 trigger_async_action(struct pending_request *sr)
 470 {
 471         struct async_request_param *param;
 472         struct rte_mp_reply *reply;
 473
 474         param = sr->async.param;
 475         reply = &param->user_reply;
 476
 477         param->clb(sr->request, reply);
 478
 479         /* clean up */
 480         free(sr->async.param->user_reply.msgs);
 481         free(sr->async.param);
 482         free(sr->request);
 483         free(sr);
 484 }
 485
 486 static struct pending_request *
 487 async_reply_handle_thread_unsafe(void *arg)
 488 {
 489         struct pending_request *req = (struct pending_request *)arg;
 490         enum async_action action;
 491         struct timespec ts_now;
 492         struct timeval now;
 493
 494         if (gettimeofday(&now, NULL) < 0) {
 495                 RTE_LOG(ERR, EAL, "Cannot get current time\n");
 496                 goto no_trigger;
 497         }
 498         ts_now.tv_nsec = now.tv_usec * 1000;
 499         ts_now.tv_sec = now.tv_sec;
 500
 501         action = process_async_request(req, &ts_now);
 502
 503         TAILQ_REMOVE(&pending_requests.requests, req, next);
 504
 505         if (rte_eal_alarm_cancel(async_reply_handle, req) < 0) {
 506                 /* if we failed to cancel the alarm because it's already in
 507                  * progress, don't proceed because otherwise we will end up
 508                  * handling the same message twice.
 509                  */
 510                 if (rte_errno == EINPROGRESS) {
 511                         RTE_LOG(DEBUG, EAL, "Request handling is already in progress\n");
 512                         goto no_trigger;
 513                 }
 514                 RTE_LOG(ERR, EAL, "Failed to cancel alarm\n");
 515         }
 516
 517         if (action == ACTION_TRIGGER)
 518                 return req;
 519 no_trigger:
 520         free(req);
 521         return NULL;
 522 }
 523
 524 static void
 525 async_reply_handle(void *arg)
 526 {
 527         struct pending_request *req;
 528
 529         pthread_mutex_lock(&pending_requests.lock);
 530         req = async_reply_handle_thread_unsafe(arg);
 531         pthread_mutex_unlock(&pending_requests.lock);
 532
 533         if (req != NULL)
 534                 trigger_async_action(req);
 535 }
 536
 537 static int
 538 open_socket_fd(void)
 539 {
 540         struct sockaddr_un un;
 541
 542         peer_name[0] = '\0';
 543         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 544                 snprintf(peer_name, sizeof(peer_name),
 545                                 "%d_%"PRIx64, getpid(), rte_rdtsc());
 546
 547         mp_fd = socket(AF_UNIX, SOCK_DGRAM, 0);
 548         if (mp_fd < 0) {
 549                 RTE_LOG(ERR, EAL, "failed to create unix socket\n");
 550                 return -1;
 551         }
 552
 553         memset(&un, 0, sizeof(un));
 554         un.sun_family = AF_UNIX;
 555
 556         create_socket_path(peer_name, un.sun_path, sizeof(un.sun_path));
 557
 558         unlink(un.sun_path); /* May still exist since last run */
 559
 560         if (bind(mp_fd, (struct sockaddr *)&un, sizeof(un)) < 0) {
 561                 RTE_LOG(ERR, EAL, "failed to bind %s: %s\n",
 562                         un.sun_path, strerror(errno));
 563                 close(mp_fd);
 564                 return -1;
 565         }
 566
 567         RTE_LOG(INFO, EAL, "Multi-process socket %s\n", un.sun_path);
 568         return mp_fd;
 569 }
 570
 571 static void
 572 close_socket_fd(void)
 573 {
 574         char path[PATH_MAX];
 575
 576         if (mp_fd < 0)
 577                 return;
 578
 579         close(mp_fd);
 580         create_socket_path(peer_name, path, sizeof(path));
 581         unlink(path);
 582 }
 583
 584 int
 585 rte_mp_channel_init(void)
 586 {
 587         char path[PATH_MAX];
 588         int dir_fd;
 589         pthread_t mp_handle_tid;
 590         const struct internal_config *internal_conf =
 591                 eal_get_internal_configuration();
 592
 593         /* in no shared files mode, we do not have secondary processes support,
 594          * so no need to initialize IPC.
 595          */
 596         if (internal_conf->no_shconf) {
 597                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC will be disabled\n");
 598                 rte_errno = ENOTSUP;
 599                 return -1;
 600         }
 601
 602         /* create filter path */
 603         create_socket_path("*", path, sizeof(path));
 604         strlcpy(mp_filter, basename(path), sizeof(mp_filter));
 605
 606         /* path may have been modified, so recreate it */
 607         create_socket_path("*", path, sizeof(path));
 608         strlcpy(mp_dir_path, dirname(path), sizeof(mp_dir_path));
 609
 610         /* lock the directory */
 611         dir_fd = open(mp_dir_path, O_RDONLY);
 612         if (dir_fd < 0) {
 613                 RTE_LOG(ERR, EAL, "failed to open %s: %s\n",
 614                         mp_dir_path, strerror(errno));
 615                 return -1;
 616         }
 617
 618         if (flock(dir_fd, LOCK_EX)) {
 619                 RTE_LOG(ERR, EAL, "failed to lock %s: %s\n",
 620                         mp_dir_path, strerror(errno));
 621                 close(dir_fd);
 622                 return -1;
 623         }
 624
 625         if (open_socket_fd() < 0) {
 626                 close(dir_fd);
 627                 return -1;
 628         }
 629
 630         if (rte_ctrl_thread_create(&mp_handle_tid, "rte_mp_handle",
 631                         NULL, mp_handle, NULL) < 0) {
 632                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 633                         strerror(errno));
 634                 close(mp_fd);
 635                 close(dir_fd);
 636                 mp_fd = -1;
 637                 return -1;
 638         }
 639
 640         /* unlock the directory */
 641         flock(dir_fd, LOCK_UN);
 642         close(dir_fd);
 643
 644         return 0;
 645 }
 646
 647 void
 648 rte_mp_channel_cleanup(void)
 649 {
 650         close_socket_fd();
 651 }
 652
 653 /**
 654  * Return -1, as fail to send message and it's caused by the local side.
 655  * Return 0, as fail to send message and it's caused by the remote side.
 656  * Return 1, as succeed to send message.
 657  *
 658  */
 659 static int
 660 send_msg(const char *dst_path, struct rte_mp_msg *msg, int type)
 661 {
 662         int snd;
 663         struct iovec iov;
 664         struct msghdr msgh;
 665         struct cmsghdr *cmsg;
 666         struct sockaddr_un dst;
 667         struct mp_msg_internal m;
 668         int fd_size = msg->num_fds * sizeof(int);
 669         char control[CMSG_SPACE(fd_size)];
 670
 671         m.type = type;
 672         memcpy(&m.msg, msg, sizeof(*msg));
 673
 674         memset(&dst, 0, sizeof(dst));
 675         dst.sun_family = AF_UNIX;
 676         strlcpy(dst.sun_path, dst_path, sizeof(dst.sun_path));
 677
 678         memset(&msgh, 0, sizeof(msgh));
 679         memset(control, 0, sizeof(control));
 680
 681         iov.iov_base = &m;
 682         iov.iov_len = sizeof(m) - sizeof(msg->fds);
 683
 684         msgh.msg_name = &dst;
 685         msgh.msg_namelen = sizeof(dst);
 686         msgh.msg_iov = &iov;
 687         msgh.msg_iovlen = 1;
 688         msgh.msg_control = control;
 689         msgh.msg_controllen = sizeof(control);
 690
 691         cmsg = CMSG_FIRSTHDR(&msgh);
 692         cmsg->cmsg_len = CMSG_LEN(fd_size);
 693         cmsg->cmsg_level = SOL_SOCKET;
 694         cmsg->cmsg_type = SCM_RIGHTS;
 695         memcpy(CMSG_DATA(cmsg), msg->fds, fd_size);
 696
 697         do {
 698                 snd = sendmsg(mp_fd, &msgh, 0);
 699         } while (snd < 0 && errno == EINTR);
 700
 701         if (snd < 0) {
 702                 rte_errno = errno;
 703                 /* Check if it caused by peer process exits */
 704                 if (errno == ECONNREFUSED &&
 705                                 rte_eal_process_type() == RTE_PROC_PRIMARY) {
 706                         unlink(dst_path);
 707                         return 0;
 708                 }
 709                 RTE_LOG(ERR, EAL, "failed to send to (%s) due to %s\n",
 710                         dst_path, strerror(errno));
 711                 return -1;
 712         }
 713
 714         return 1;
 715 }
 716
 717 static int
 718 mp_send(struct rte_mp_msg *msg, const char *peer, int type)
 719 {
 720         int dir_fd, ret = 0;
 721         DIR *mp_dir;
 722         struct dirent *ent;
 723
 724         if (!peer && (rte_eal_process_type() == RTE_PROC_SECONDARY))
 725                 peer = eal_mp_socket_path();
 726
 727         if (peer) {
 728                 if (send_msg(peer, msg, type) < 0)
 729                         return -1;
 730                 else
 731                         return 0;
 732         }
 733
 734         /* broadcast to all secondary processes */
 735         mp_dir = opendir(mp_dir_path);
 736         if (!mp_dir) {
 737                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n",
 738                                 mp_dir_path);
 739                 rte_errno = errno;
 740                 return -1;
 741         }
 742
 743         dir_fd = dirfd(mp_dir);
 744         /* lock the directory to prevent processes spinning up while we send */
 745         if (flock(dir_fd, LOCK_SH)) {
 746                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
 747                         mp_dir_path);
 748                 rte_errno = errno;
 749                 closedir(mp_dir);
 750                 return -1;
 751         }
 752
 753         while ((ent = readdir(mp_dir))) {
 754                 char path[PATH_MAX];
 755
 756                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
 757                         continue;
 758
 759                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
 760                          ent->d_name);
 761                 if (send_msg(path, msg, type) < 0)
 762                         ret = -1;
 763         }
 764         /* unlock the dir */
 765         flock(dir_fd, LOCK_UN);
 766
 767         /* dir_fd automatically closed on closedir */
 768         closedir(mp_dir);
 769         return ret;
 770 }
 771
 772 static int
 773 check_input(const struct rte_mp_msg *msg)
 774 {
 775         if (msg == NULL) {
 776                 RTE_LOG(ERR, EAL, "Msg cannot be NULL\n");
 777                 rte_errno = EINVAL;
 778                 return -1;
 779         }
 780
 781         if (validate_action_name(msg->name) != 0)
 782                 return -1;
 783
 784         if (msg->len_param < 0) {
 785                 RTE_LOG(ERR, EAL, "Message data length is negative\n");
 786                 rte_errno = EINVAL;
 787                 return -1;
 788         }
 789
 790         if (msg->num_fds < 0) {
 791                 RTE_LOG(ERR, EAL, "Number of fd's is negative\n");
 792                 rte_errno = EINVAL;
 793                 return -1;
 794         }
 795
 796         if (msg->len_param > RTE_MP_MAX_PARAM_LEN) {
 797                 RTE_LOG(ERR, EAL, "Message data is too long\n");
 798                 rte_errno = E2BIG;
 799                 return -1;
 800         }
 801
 802         if (msg->num_fds > RTE_MP_MAX_FD_NUM) {
 803                 RTE_LOG(ERR, EAL, "Cannot send more than %d FDs\n",
 804                         RTE_MP_MAX_FD_NUM);
 805                 rte_errno = E2BIG;
 806                 return -1;
 807         }
 808
 809         return 0;
 810 }
 811
 812 int
 813 rte_mp_sendmsg(struct rte_mp_msg *msg)
 814 {
 815         const struct internal_config *internal_conf =
 816                 eal_get_internal_configuration();
 817
 818         if (check_input(msg) != 0)
 819                 return -1;
 820
 821         if (internal_conf->no_shconf) {
 822                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
 823                 rte_errno = ENOTSUP;
 824                 return -1;
 825         }
 826
 827         RTE_LOG(DEBUG, EAL, "sendmsg: %s\n", msg->name);
 828         return mp_send(msg, NULL, MP_MSG);
 829 }
 830
 831 static int
 832 mp_request_async(const char *dst, struct rte_mp_msg *req,
 833                 struct async_request_param *param, const struct timespec *ts)
 834 {
 835         struct rte_mp_msg *reply_msg;
 836         struct pending_request *pending_req, *exist;
 837         int ret = -1;
 838
 839         pending_req = calloc(1, sizeof(*pending_req));
 840         reply_msg = calloc(1, sizeof(*reply_msg));
 841         if (pending_req == NULL || reply_msg == NULL) {
 842                 RTE_LOG(ERR, EAL, "Could not allocate space for sync request\n");
 843                 rte_errno = ENOMEM;
 844                 ret = -1;
 845                 goto fail;
 846         }
 847
 848         pending_req->type = REQUEST_TYPE_ASYNC;
 849         strlcpy(pending_req->dst, dst, sizeof(pending_req->dst));
 850         pending_req->request = req;
 851         pending_req->reply = reply_msg;
 852         pending_req->async.param = param;
 853
 854         /* queue already locked by caller */
 855
 856         exist = find_pending_request(dst, req->name);
 857         if (exist) {
 858                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 859                 rte_errno = EEXIST;
 860                 ret = -1;
 861                 goto fail;
 862         }
 863
 864         ret = send_msg(dst, req, MP_REQ);
 865         if (ret < 0) {
 866                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 867                         dst, req->name);
 868                 ret = -1;
 869                 goto fail;
 870         } else if (ret == 0) {
 871                 ret = 0;
 872                 goto fail;
 873         }
 874         param->user_reply.nb_sent++;
 875
 876         /* if alarm set fails, we simply ignore the reply */
 877         if (rte_eal_alarm_set(ts->tv_sec * 1000000 + ts->tv_nsec / 1000,
 878                               async_reply_handle, pending_req) < 0) {
 879                 RTE_LOG(ERR, EAL, "Fail to set alarm for request %s:%s\n",
 880                         dst, req->name);
 881                 ret = -1;
 882                 goto fail;
 883         }
 884         TAILQ_INSERT_TAIL(&pending_requests.requests, pending_req, next);
 885
 886         return 0;
 887 fail:
 888         free(pending_req);
 889         free(reply_msg);
 890         return ret;
 891 }
 892
 893 static int
 894 mp_request_sync(const char *dst, struct rte_mp_msg *req,
 895                struct rte_mp_reply *reply, const struct timespec *ts)
 896 {
 897         int ret;
 898         struct rte_mp_msg msg, *tmp;
 899         struct pending_request pending_req, *exist;
 900
 901         pending_req.type = REQUEST_TYPE_SYNC;
 902         pending_req.reply_received = 0;
 903         strlcpy(pending_req.dst, dst, sizeof(pending_req.dst));
 904         pending_req.request = req;
 905         pending_req.reply = &msg;
 906         pthread_cond_init(&pending_req.sync.cond, NULL);
 907
 908         exist = find_pending_request(dst, req->name);
 909         if (exist) {
 910                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 911                 rte_errno = EEXIST;
 912                 return -1;
 913         }
 914
 915         ret = send_msg(dst, req, MP_REQ);
 916         if (ret < 0) {
 917                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 918                         dst, req->name);
 919                 return -1;
 920         } else if (ret == 0)
 921                 return 0;
 922
 923         TAILQ_INSERT_TAIL(&pending_requests.requests, &pending_req, next);
 924
 925         reply->nb_sent++;
 926
 927         do {
 928                 ret = pthread_cond_timedwait(&pending_req.sync.cond,
 929                                 &pending_requests.lock, ts);
 930         } while (ret != 0 && ret != ETIMEDOUT);
 931
 932         TAILQ_REMOVE(&pending_requests.requests, &pending_req, next);
 933
 934         if (pending_req.reply_received == 0) {
 935                 RTE_LOG(ERR, EAL, "Fail to recv reply for request %s:%s\n",
 936                         dst, req->name);
 937                 rte_errno = ETIMEDOUT;
 938                 return -1;
 939         }
 940         if (pending_req.reply_received == -1) {
 941                 RTE_LOG(DEBUG, EAL, "Asked to ignore response\n");
 942                 /* not receiving this message is not an error, so decrement
 943                  * number of sent messages
 944                  */
 945                 reply->nb_sent--;
 946                 return 0;
 947         }
 948
 949         tmp = realloc(reply->msgs, sizeof(msg) * (reply->nb_received + 1));
 950         if (!tmp) {
 951                 RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 952                         dst, req->name);
 953                 rte_errno = ENOMEM;
 954                 return -1;
 955         }
 956         memcpy(&tmp[reply->nb_received], &msg, sizeof(msg));
 957         reply->msgs = tmp;
 958         reply->nb_received++;
 959         return 0;
 960 }
 961
 962 int
 963 rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
 964                 const struct timespec *ts)
 965 {
 966         int dir_fd, ret = -1;
 967         DIR *mp_dir;
 968         struct dirent *ent;
 969         struct timeval now;
 970         struct timespec end;
 971         const struct internal_config *internal_conf =
 972                 eal_get_internal_configuration();
 973
 974         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
 975
 976         reply->nb_sent = 0;
 977         reply->nb_received = 0;
 978         reply->msgs = NULL;
 979
 980         if (check_input(req) != 0)
 981                 goto end;
 982
 983         if (internal_conf->no_shconf) {
 984                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
 985                 rte_errno = ENOTSUP;
 986                 return -1;
 987         }
 988
 989         if (gettimeofday(&now, NULL) < 0) {
 990                 RTE_LOG(ERR, EAL, "Failed to get current time\n");
 991                 rte_errno = errno;
 992                 goto end;
 993         }
 994
 995         end.tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
 996         end.tv_sec = now.tv_sec + ts->tv_sec +
 997                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
 998
 999         /* for secondary process, send request to the primary process only */
1000         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1001                 pthread_mutex_lock(&pending_requests.lock);
1002                 ret = mp_request_sync(eal_mp_socket_path(), req, reply, &end);
1003                 pthread_mutex_unlock(&pending_requests.lock);
1004                 goto end;
1005         }
1006
1007         /* for primary process, broadcast request, and collect reply 1 by 1 */
1008         mp_dir = opendir(mp_dir_path);
1009         if (!mp_dir) {
1010                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1011                 rte_errno = errno;
1012                 goto end;
1013         }
1014
1015         dir_fd = dirfd(mp_dir);
1016         /* lock the directory to prevent processes spinning up while we send */
1017         if (flock(dir_fd, LOCK_SH)) {
1018                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1019                         mp_dir_path);
1020                 rte_errno = errno;
1021                 goto close_end;
1022         }
1023
1024         pthread_mutex_lock(&pending_requests.lock);
1025         while ((ent = readdir(mp_dir))) {
1026                 char path[PATH_MAX];
1027
1028                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1029                         continue;
1030
1031                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1032                          ent->d_name);
1033
1034                 /* unlocks the mutex while waiting for response,
1035                  * locks on receive
1036                  */
1037                 if (mp_request_sync(path, req, reply, &end))
1038                         goto unlock_end;
1039         }
1040         ret = 0;
1041
1042 unlock_end:
1043         pthread_mutex_unlock(&pending_requests.lock);
1044         /* unlock the directory */
1045         flock(dir_fd, LOCK_UN);
1046
1047 close_end:
1048         /* dir_fd automatically closed on closedir */
1049         closedir(mp_dir);
1050
1051 end:
1052         if (ret) {
1053                 free(reply->msgs);
1054                 reply->nb_received = 0;
1055                 reply->msgs = NULL;
1056         }
1057         return ret;
1058 }
1059
1060 int
1061 rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
1062                 rte_mp_async_reply_t clb)
1063 {
1064         struct rte_mp_msg *copy;
1065         struct pending_request *dummy;
1066         struct async_request_param *param;
1067         struct rte_mp_reply *reply;
1068         int dir_fd, ret = 0;
1069         DIR *mp_dir;
1070         struct dirent *ent;
1071         struct timeval now;
1072         struct timespec *end;
1073         bool dummy_used = false;
1074         const struct internal_config *internal_conf =
1075                 eal_get_internal_configuration();
1076
1077         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
1078
1079         if (check_input(req) != 0)
1080                 return -1;
1081
1082         if (internal_conf->no_shconf) {
1083                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
1084                 rte_errno = ENOTSUP;
1085                 return -1;
1086         }
1087
1088         if (gettimeofday(&now, NULL) < 0) {
1089                 RTE_LOG(ERR, EAL, "Failed to get current time\n");
1090                 rte_errno = errno;
1091                 return -1;
1092         }
1093         copy = calloc(1, sizeof(*copy));
1094         dummy = calloc(1, sizeof(*dummy));
1095         param = calloc(1, sizeof(*param));
1096         if (copy == NULL || dummy == NULL || param == NULL) {
1097                 RTE_LOG(ERR, EAL, "Failed to allocate memory for async reply\n");
1098                 rte_errno = ENOMEM;
1099                 goto fail;
1100         }
1101
1102         /* copy message */
1103         memcpy(copy, req, sizeof(*copy));
1104
1105         param->n_responses_processed = 0;
1106         param->clb = clb;
1107         end = &param->end;
1108         reply = &param->user_reply;
1109
1110         end->tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
1111         end->tv_sec = now.tv_sec + ts->tv_sec +
1112                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1113         reply->nb_sent = 0;
1114         reply->nb_received = 0;
1115         reply->msgs = NULL;
1116
1117         /* we have to lock the request queue here, as we will be adding a bunch
1118          * of requests to the queue at once, and some of the replies may arrive
1119          * before we add all of the requests to the queue.
1120          */
1121         pthread_mutex_lock(&pending_requests.lock);
1122
1123         /* we have to ensure that callback gets triggered even if we don't send
1124          * anything, therefore earlier we have allocated a dummy request. fill
1125          * it, and put it on the queue if we don't send any requests.
1126          */
1127         dummy->type = REQUEST_TYPE_ASYNC;
1128         dummy->request = copy;
1129         dummy->reply = NULL;
1130         dummy->async.param = param;
1131         dummy->reply_received = 1; /* short-circuit the timeout */
1132
1133         /* for secondary process, send request to the primary process only */
1134         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1135                 ret = mp_request_async(eal_mp_socket_path(), copy, param, ts);
1136
1137                 /* if we didn't send anything, put dummy request on the queue */
1138                 if (ret == 0 && reply->nb_sent == 0) {
1139                         TAILQ_INSERT_TAIL(&pending_requests.requests, dummy,
1140                                         next);
1141                         dummy_used = true;
1142                 }
1143
1144                 pthread_mutex_unlock(&pending_requests.lock);
1145
1146                 /* if we couldn't send anything, clean up */
1147                 if (ret != 0)
1148                         goto fail;
1149                 return 0;
1150         }
1151
1152         /* for primary process, broadcast request */
1153         mp_dir = opendir(mp_dir_path);
1154         if (!mp_dir) {
1155                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1156                 rte_errno = errno;
1157                 goto unlock_fail;
1158         }
1159         dir_fd = dirfd(mp_dir);
1160
1161         /* lock the directory to prevent processes spinning up while we send */
1162         if (flock(dir_fd, LOCK_SH)) {
1163                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1164                         mp_dir_path);
1165                 rte_errno = errno;
1166                 goto closedir_fail;
1167         }
1168
1169         while ((ent = readdir(mp_dir))) {
1170                 char path[PATH_MAX];
1171
1172                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1173                         continue;
1174
1175                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1176                          ent->d_name);
1177
1178                 if (mp_request_async(path, copy, param, ts))
1179                         ret = -1;
1180         }
1181         /* if we didn't send anything, put dummy request on the queue */
1182         if (ret == 0 && reply->nb_sent == 0) {
1183                 TAILQ_INSERT_HEAD(&pending_requests.requests, dummy, next);
1184                 dummy_used = true;
1185         }
1186
1187         /* finally, unlock the queue */
1188         pthread_mutex_unlock(&pending_requests.lock);
1189
1190         /* unlock the directory */
1191         flock(dir_fd, LOCK_UN);
1192
1193         /* dir_fd automatically closed on closedir */
1194         closedir(mp_dir);
1195
1196         /* if dummy was unused, free it */
1197         if (!dummy_used)
1198                 free(dummy);
1199
1200         return ret;
1201 closedir_fail:
1202         closedir(mp_dir);
1203 unlock_fail:
1204         pthread_mutex_unlock(&pending_requests.lock);
1205 fail:
1206         free(dummy);
1207         free(param);
1208         free(copy);
1209         return -1;
1210 }
1211
1212 int
1213 rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
1214 {
1215         RTE_LOG(DEBUG, EAL, "reply: %s\n", msg->name);
1216         const struct internal_config *internal_conf =
1217                 eal_get_internal_configuration();
1218
1219         if (check_input(msg) != 0)
1220                 return -1;
1221
1222         if (peer == NULL) {
1223                 RTE_LOG(ERR, EAL, "peer is not specified\n");
1224                 rte_errno = EINVAL;
1225                 return -1;
1226         }
1227
1228         if (internal_conf->no_shconf) {
1229                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
1230                 return 0;
1231         }
1232
1233         return mp_send(msg, peer, MP_REP);
1234 }