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