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 void *
 445 async_reply_handle(void *arg __rte_unused)
 446 {
 447         struct pending_request *sr;
 448         struct timeval now;
 449         struct timespec timeout, ts_now;
 450         while (1) {
 451                 struct pending_request *trigger = NULL;
 452                 int ret;
 453                 bool nowait = false;
 454                 bool timedwait = false;
 455
 456                 pthread_mutex_lock(&pending_requests.lock);
 457
 458                 /* scan through the list and see if there are any timeouts that
 459                  * are earlier than our current timeout.
 460                  */
 461                 TAILQ_FOREACH(sr, &pending_requests.requests, next) {
 462                         if (sr->type != REQUEST_TYPE_ASYNC)
 463                                 continue;
 464                         if (!timedwait || timespec_cmp(&sr->async.param->end,
 465                                         &timeout) < 0) {
 466                                 memcpy(&timeout, &sr->async.param->end,
 467                                         sizeof(timeout));
 468                                 timedwait = true;
 469                         }
 470
 471                         /* sometimes, we don't even wait */
 472                         if (sr->reply_received) {
 473                                 nowait = true;
 474                                 break;
 475                         }
 476                 }
 477
 478                 if (nowait)
 479                         ret = 0;
 480                 else if (timedwait)
 481                         ret = pthread_cond_timedwait(
 482                                         &pending_requests.async_cond,
 483                                         &pending_requests.lock, &timeout);
 484                 else
 485                         ret = pthread_cond_wait(&pending_requests.async_cond,
 486                                         &pending_requests.lock);
 487
 488                 if (gettimeofday(&now, NULL) < 0) {
 489                         RTE_LOG(ERR, EAL, "Cannot get current time\n");
 490                         break;
 491                 }
 492                 ts_now.tv_nsec = now.tv_usec * 1000;
 493                 ts_now.tv_sec = now.tv_sec;
 494
 495                 if (ret == 0 || ret == ETIMEDOUT) {
 496                         struct pending_request *next;
 497                         /* we've either been woken up, or we timed out */
 498
 499                         /* we have still the lock, check if anything needs
 500                          * processing.
 501                          */
 502                         TAILQ_FOREACH_SAFE(sr, &pending_requests.requests, next,
 503                                         next) {
 504                                 enum async_action action;
 505                                 if (sr->type != REQUEST_TYPE_ASYNC)
 506                                         continue;
 507
 508                                 action = process_async_request(sr, &ts_now);
 509                                 if (action == ACTION_FREE) {
 510                                         TAILQ_REMOVE(&pending_requests.requests,
 511                                                         sr, next);
 512                                         free(sr);
 513                                 } else if (action == ACTION_TRIGGER &&
 514                                                 trigger == NULL) {
 515                                         TAILQ_REMOVE(&pending_requests.requests,
 516                                                         sr, next);
 517                                         trigger = sr;
 518                                 }
 519                         }
 520                 }
 521                 pthread_mutex_unlock(&pending_requests.lock);
 522                 if (trigger) {
 523                         trigger_async_action(trigger);
 524                         free(trigger);
 525                 }
 526         };
 527
 528         RTE_LOG(ERR, EAL, "ERROR: asynchronous requests disabled\n");
 529
 530         return NULL;
 531 }
 532
 533 static int
 534 open_socket_fd(void)
 535 {
 536         char peer_name[PATH_MAX] = {0};
 537         struct sockaddr_un un;
 538
 539         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 540                 snprintf(peer_name, sizeof(peer_name),
 541                                 "%d_%"PRIx64, getpid(), rte_rdtsc());
 542
 543         mp_fd = socket(AF_UNIX, SOCK_DGRAM, 0);
 544         if (mp_fd < 0) {
 545                 RTE_LOG(ERR, EAL, "failed to create unix socket\n");
 546                 return -1;
 547         }
 548
 549         memset(&un, 0, sizeof(un));
 550         un.sun_family = AF_UNIX;
 551
 552         create_socket_path(peer_name, un.sun_path, sizeof(un.sun_path));
 553
 554         unlink(un.sun_path); /* May still exist since last run */
 555
 556         if (bind(mp_fd, (struct sockaddr *)&un, sizeof(un)) < 0) {
 557                 RTE_LOG(ERR, EAL, "failed to bind %s: %s\n",
 558                         un.sun_path, strerror(errno));
 559                 close(mp_fd);
 560                 return -1;
 561         }
 562
 563         RTE_LOG(INFO, EAL, "Multi-process socket %s\n", un.sun_path);
 564         return mp_fd;
 565 }
 566
 567 static int
 568 unlink_sockets(const char *filter)
 569 {
 570         int dir_fd;
 571         DIR *mp_dir;
 572         struct dirent *ent;
 573
 574         mp_dir = opendir(mp_dir_path);
 575         if (!mp_dir) {
 576                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
 577                 return -1;
 578         }
 579         dir_fd = dirfd(mp_dir);
 580
 581         while ((ent = readdir(mp_dir))) {
 582                 if (fnmatch(filter, ent->d_name, 0) == 0)
 583                         unlinkat(dir_fd, ent->d_name, 0);
 584         }
 585
 586         closedir(mp_dir);
 587         return 0;
 588 }
 589
 590 int
 591 rte_mp_channel_init(void)
 592 {
 593         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 594         char path[PATH_MAX];
 595         int dir_fd;
 596         pthread_t mp_handle_tid, async_reply_handle_tid;
 597
 598         /* create filter path */
 599         create_socket_path("*", path, sizeof(path));
 600         snprintf(mp_filter, sizeof(mp_filter), "%s", basename(path));
 601
 602         /* path may have been modified, so recreate it */
 603         create_socket_path("*", path, sizeof(path));
 604         snprintf(mp_dir_path, sizeof(mp_dir_path), "%s", dirname(path));
 605
 606         /* lock the directory */
 607         dir_fd = open(mp_dir_path, O_RDONLY);
 608         if (dir_fd < 0) {
 609                 RTE_LOG(ERR, EAL, "failed to open %s: %s\n",
 610                         mp_dir_path, strerror(errno));
 611                 return -1;
 612         }
 613
 614         if (flock(dir_fd, LOCK_EX)) {
 615                 RTE_LOG(ERR, EAL, "failed to lock %s: %s\n",
 616                         mp_dir_path, strerror(errno));
 617                 close(dir_fd);
 618                 return -1;
 619         }
 620
 621         if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
 622                         unlink_sockets(mp_filter)) {
 623                 RTE_LOG(ERR, EAL, "failed to unlink mp sockets\n");
 624                 close(dir_fd);
 625                 return -1;
 626         }
 627
 628         if (open_socket_fd() < 0) {
 629                 close(dir_fd);
 630                 return -1;
 631         }
 632
 633         if (pthread_create(&mp_handle_tid, NULL, mp_handle, NULL) < 0) {
 634                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 635                         strerror(errno));
 636                 close(mp_fd);
 637                 mp_fd = -1;
 638                 return -1;
 639         }
 640
 641         if (pthread_create(&async_reply_handle_tid, NULL,
 642                         async_reply_handle, NULL) < 0) {
 643                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 644                         strerror(errno));
 645                 close(mp_fd);
 646                 close(dir_fd);
 647                 mp_fd = -1;
 648                 return -1;
 649         }
 650
 651         /* try best to set thread name */
 652         snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN, "rte_mp_handle");
 653         rte_thread_setname(mp_handle_tid, thread_name);
 654
 655         /* try best to set thread name */
 656         snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN, "rte_mp_async_handle");
 657         rte_thread_setname(async_reply_handle_tid, thread_name);
 658
 659         /* unlock the directory */
 660         flock(dir_fd, LOCK_UN);
 661         close(dir_fd);
 662
 663         return 0;
 664 }
 665
 666 /**
 667  * Return -1, as fail to send message and it's caused by the local side.
 668  * Return 0, as fail to send message and it's caused by the remote side.
 669  * Return 1, as succeed to send message.
 670  *
 671  */
 672 static int
 673 send_msg(const char *dst_path, struct rte_mp_msg *msg, int type)
 674 {
 675         int snd;
 676         struct iovec iov;
 677         struct msghdr msgh;
 678         struct cmsghdr *cmsg;
 679         struct sockaddr_un dst;
 680         struct mp_msg_internal m;
 681         int fd_size = msg->num_fds * sizeof(int);
 682         char control[CMSG_SPACE(fd_size)];
 683
 684         m.type = type;
 685         memcpy(&m.msg, msg, sizeof(*msg));
 686
 687         memset(&dst, 0, sizeof(dst));
 688         dst.sun_family = AF_UNIX;
 689         snprintf(dst.sun_path, sizeof(dst.sun_path), "%s", dst_path);
 690
 691         memset(&msgh, 0, sizeof(msgh));
 692         memset(control, 0, sizeof(control));
 693
 694         iov.iov_base = &m;
 695         iov.iov_len = sizeof(m) - sizeof(msg->fds);
 696
 697         msgh.msg_name = &dst;
 698         msgh.msg_namelen = sizeof(dst);
 699         msgh.msg_iov = &iov;
 700         msgh.msg_iovlen = 1;
 701         msgh.msg_control = control;
 702         msgh.msg_controllen = sizeof(control);
 703
 704         cmsg = CMSG_FIRSTHDR(&msgh);
 705         cmsg->cmsg_len = CMSG_LEN(fd_size);
 706         cmsg->cmsg_level = SOL_SOCKET;
 707         cmsg->cmsg_type = SCM_RIGHTS;
 708         memcpy(CMSG_DATA(cmsg), msg->fds, fd_size);
 709
 710         do {
 711                 snd = sendmsg(mp_fd, &msgh, 0);
 712         } while (snd < 0 && errno == EINTR);
 713
 714         if (snd < 0) {
 715                 rte_errno = errno;
 716                 /* Check if it caused by peer process exits */
 717                 if (errno == ECONNREFUSED &&
 718                                 rte_eal_process_type() == RTE_PROC_PRIMARY) {
 719                         unlink(dst_path);
 720                         return 0;
 721                 }
 722                 if (errno == ENOBUFS) {
 723                         RTE_LOG(ERR, EAL, "Peer cannot receive message %s\n",
 724                                 dst_path);
 725                         return 0;
 726                 }
 727                 RTE_LOG(ERR, EAL, "failed to send to (%s) due to %s\n",
 728                         dst_path, strerror(errno));
 729                 return -1;
 730         }
 731
 732         return 1;
 733 }
 734
 735 static int
 736 mp_send(struct rte_mp_msg *msg, const char *peer, int type)
 737 {
 738         int dir_fd, ret = 0;
 739         DIR *mp_dir;
 740         struct dirent *ent;
 741
 742         if (!peer && (rte_eal_process_type() == RTE_PROC_SECONDARY))
 743                 peer = eal_mp_socket_path();
 744
 745         if (peer) {
 746                 if (send_msg(peer, msg, type) < 0)
 747                         return -1;
 748                 else
 749                         return 0;
 750         }
 751
 752         /* broadcast to all secondary processes */
 753         mp_dir = opendir(mp_dir_path);
 754         if (!mp_dir) {
 755                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n",
 756                                 mp_dir_path);
 757                 rte_errno = errno;
 758                 return -1;
 759         }
 760
 761         dir_fd = dirfd(mp_dir);
 762         /* lock the directory to prevent processes spinning up while we send */
 763         if (flock(dir_fd, LOCK_EX)) {
 764                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
 765                         mp_dir_path);
 766                 rte_errno = errno;
 767                 closedir(mp_dir);
 768                 return -1;
 769         }
 770
 771         while ((ent = readdir(mp_dir))) {
 772                 char path[PATH_MAX];
 773
 774                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
 775                         continue;
 776
 777                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
 778                          ent->d_name);
 779                 if (send_msg(path, msg, type) < 0)
 780                         ret = -1;
 781         }
 782         /* unlock the dir */
 783         flock(dir_fd, LOCK_UN);
 784
 785         /* dir_fd automatically closed on closedir */
 786         closedir(mp_dir);
 787         return ret;
 788 }
 789
 790 static bool
 791 check_input(const struct rte_mp_msg *msg)
 792 {
 793         if (msg == NULL) {
 794                 RTE_LOG(ERR, EAL, "Msg cannot be NULL\n");
 795                 rte_errno = EINVAL;
 796                 return false;
 797         }
 798
 799         if (validate_action_name(msg->name))
 800                 return false;
 801
 802         if (msg->len_param > RTE_MP_MAX_PARAM_LEN) {
 803                 RTE_LOG(ERR, EAL, "Message data is too long\n");
 804                 rte_errno = E2BIG;
 805                 return false;
 806         }
 807
 808         if (msg->num_fds > RTE_MP_MAX_FD_NUM) {
 809                 RTE_LOG(ERR, EAL, "Cannot send more than %d FDs\n",
 810                         RTE_MP_MAX_FD_NUM);
 811                 rte_errno = E2BIG;
 812                 return false;
 813         }
 814
 815         return true;
 816 }
 817
 818 int __rte_experimental
 819 rte_mp_sendmsg(struct rte_mp_msg *msg)
 820 {
 821         if (!check_input(msg))
 822                 return -1;
 823
 824         RTE_LOG(DEBUG, EAL, "sendmsg: %s\n", msg->name);
 825         return mp_send(msg, NULL, MP_MSG);
 826 }
 827
 828 static int
 829 mp_request_async(const char *dst, struct rte_mp_msg *req,
 830                 struct async_request_param *param)
 831 {
 832         struct rte_mp_msg *reply_msg;
 833         struct pending_request *sync_req, *exist;
 834         int ret;
 835
 836         sync_req = malloc(sizeof(*sync_req));
 837         reply_msg = malloc(sizeof(*reply_msg));
 838         if (sync_req == NULL || reply_msg == NULL) {
 839                 RTE_LOG(ERR, EAL, "Could not allocate space for sync request\n");
 840                 rte_errno = ENOMEM;
 841                 ret = -1;
 842                 goto fail;
 843         }
 844
 845         memset(sync_req, 0, sizeof(*sync_req));
 846         memset(reply_msg, 0, sizeof(*reply_msg));
 847
 848         sync_req->type = REQUEST_TYPE_ASYNC;
 849         strcpy(sync_req->dst, dst);
 850         sync_req->request = req;
 851         sync_req->reply = reply_msg;
 852         sync_req->async.param = param;
 853
 854         /* queue already locked by caller */
 855
 856         exist = find_sync_request(dst, req->name);
 857         if (!exist) {
 858                 TAILQ_INSERT_TAIL(&pending_requests.requests, sync_req, next);
 859         } else {
 860                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 861                 rte_errno = EEXIST;
 862                 ret = -1;
 863                 goto fail;
 864         }
 865
 866         ret = send_msg(dst, req, MP_REQ);
 867         if (ret < 0) {
 868                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 869                         dst, req->name);
 870                 ret = -1;
 871                 goto fail;
 872         } else if (ret == 0) {
 873                 ret = 0;
 874                 goto fail;
 875         }
 876
 877         param->user_reply.nb_sent++;
 878
 879         return 0;
 880 fail:
 881         free(sync_req);
 882         free(reply_msg);
 883         return ret;
 884 }
 885
 886 static int
 887 mp_request_sync(const char *dst, struct rte_mp_msg *req,
 888                struct rte_mp_reply *reply, const struct timespec *ts)
 889 {
 890         int ret;
 891         struct rte_mp_msg msg, *tmp;
 892         struct pending_request sync_req, *exist;
 893
 894         sync_req.type = REQUEST_TYPE_SYNC;
 895         sync_req.reply_received = 0;
 896         strcpy(sync_req.dst, dst);
 897         sync_req.request = req;
 898         sync_req.reply = &msg;
 899         pthread_cond_init(&sync_req.sync.cond, NULL);
 900
 901         pthread_mutex_lock(&pending_requests.lock);
 902         exist = find_sync_request(dst, req->name);
 903         if (!exist)
 904                 TAILQ_INSERT_TAIL(&pending_requests.requests, &sync_req, next);
 905         if (exist) {
 906                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 907                 rte_errno = EEXIST;
 908                 pthread_mutex_unlock(&pending_requests.lock);
 909                 return -1;
 910         }
 911
 912         ret = send_msg(dst, req, MP_REQ);
 913         if (ret < 0) {
 914                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 915                         dst, req->name);
 916                 return -1;
 917         } else if (ret == 0)
 918                 return 0;
 919
 920         reply->nb_sent++;
 921
 922         do {
 923                 ret = pthread_cond_timedwait(&sync_req.sync.cond,
 924                                 &pending_requests.lock, ts);
 925         } while (ret != 0 && ret != ETIMEDOUT);
 926
 927         /* We got the lock now */
 928         TAILQ_REMOVE(&pending_requests.requests, &sync_req, next);
 929         pthread_mutex_unlock(&pending_requests.lock);
 930
 931         if (sync_req.reply_received == 0) {
 932                 RTE_LOG(ERR, EAL, "Fail to recv reply for request %s:%s\n",
 933                         dst, req->name);
 934                 rte_errno = ETIMEDOUT;
 935                 return -1;
 936         }
 937         if (sync_req.reply_received == -1) {
 938                 RTE_LOG(DEBUG, EAL, "Asked to ignore response\n");
 939                 /* not receiving this message is not an error, so decrement
 940                  * number of sent messages
 941                  */
 942                 reply->nb_sent--;
 943                 return 0;
 944         }
 945
 946         tmp = realloc(reply->msgs, sizeof(msg) * (reply->nb_received + 1));
 947         if (!tmp) {
 948                 RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 949                         dst, req->name);
 950                 rte_errno = ENOMEM;
 951                 return -1;
 952         }
 953         memcpy(&tmp[reply->nb_received], &msg, sizeof(msg));
 954         reply->msgs = tmp;
 955         reply->nb_received++;
 956         return 0;
 957 }
 958
 959 int __rte_experimental
 960 rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
 961                 const struct timespec *ts)
 962 {
 963         int dir_fd, ret = 0;
 964         DIR *mp_dir;
 965         struct dirent *ent;
 966         struct timeval now;
 967         struct timespec end;
 968
 969         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
 970
 971         if (check_input(req) == false)
 972                 return -1;
 973         if (gettimeofday(&now, NULL) < 0) {
 974                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
 975                 rte_errno = errno;
 976                 return -1;
 977         }
 978
 979         end.tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
 980         end.tv_sec = now.tv_sec + ts->tv_sec +
 981                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
 982
 983         reply->nb_sent = 0;
 984         reply->nb_received = 0;
 985         reply->msgs = NULL;
 986
 987         /* for secondary process, send request to the primary process only */
 988         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 989                 return mp_request_sync(eal_mp_socket_path(), req, reply, &end);
 990
 991         /* for primary process, broadcast request, and collect reply 1 by 1 */
 992         mp_dir = opendir(mp_dir_path);
 993         if (!mp_dir) {
 994                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
 995                 rte_errno = errno;
 996                 return -1;
 997         }
 998
 999         dir_fd = dirfd(mp_dir);
1000         /* lock the directory to prevent processes spinning up while we send */
1001         if (flock(dir_fd, LOCK_EX)) {
1002                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1003                         mp_dir_path);
1004                 closedir(mp_dir);
1005                 rte_errno = errno;
1006                 return -1;
1007         }
1008
1009         while ((ent = readdir(mp_dir))) {
1010                 char path[PATH_MAX];
1011
1012                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1013                         continue;
1014
1015                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1016                          ent->d_name);
1017
1018                 if (mp_request_sync(path, req, reply, &end))
1019                         ret = -1;
1020         }
1021         /* unlock the directory */
1022         flock(dir_fd, LOCK_UN);
1023
1024         /* dir_fd automatically closed on closedir */
1025         closedir(mp_dir);
1026         return ret;
1027 }
1028
1029 int __rte_experimental
1030 rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
1031                 rte_mp_async_reply_t clb)
1032 {
1033         struct rte_mp_msg *copy;
1034         struct pending_request *dummy;
1035         struct async_request_param *param;
1036         struct rte_mp_reply *reply;
1037         int dir_fd, ret = 0;
1038         DIR *mp_dir;
1039         struct dirent *ent;
1040         struct timeval now;
1041         struct timespec *end;
1042         bool dummy_used = false;
1043
1044         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
1045
1046         if (check_input(req) == false)
1047                 return -1;
1048         if (gettimeofday(&now, NULL) < 0) {
1049                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
1050                 rte_errno = errno;
1051                 return -1;
1052         }
1053         copy = malloc(sizeof(*copy));
1054         dummy = malloc(sizeof(*dummy));
1055         param = malloc(sizeof(*param));
1056         if (copy == NULL || dummy == NULL || param == NULL) {
1057                 RTE_LOG(ERR, EAL, "Failed to allocate memory for async reply\n");
1058                 rte_errno = ENOMEM;
1059                 goto fail;
1060         }
1061
1062         memset(copy, 0, sizeof(*copy));
1063         memset(dummy, 0, sizeof(*dummy));
1064         memset(param, 0, sizeof(*param));
1065
1066         /* copy message */
1067         memcpy(copy, req, sizeof(*copy));
1068
1069         param->n_responses_processed = 0;
1070         param->clb = clb;
1071         end = &param->end;
1072         reply = &param->user_reply;
1073
1074         end->tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
1075         end->tv_sec = now.tv_sec + ts->tv_sec +
1076                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1077         reply->nb_sent = 0;
1078         reply->nb_received = 0;
1079         reply->msgs = NULL;
1080
1081         /* we have to lock the request queue here, as we will be adding a bunch
1082          * of requests to the queue at once, and some of the replies may arrive
1083          * before we add all of the requests to the queue.
1084          */
1085         pthread_mutex_lock(&pending_requests.lock);
1086
1087         /* we have to ensure that callback gets triggered even if we don't send
1088          * anything, therefore earlier we have allocated a dummy request. fill
1089          * it, and put it on the queue if we don't send any requests.
1090          */
1091         dummy->type = REQUEST_TYPE_ASYNC;
1092         dummy->request = copy;
1093         dummy->reply = NULL;
1094         dummy->async.param = param;
1095         dummy->reply_received = 1; /* short-circuit the timeout */
1096
1097         /* for secondary process, send request to the primary process only */
1098         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1099                 ret = mp_request_async(eal_mp_socket_path(), copy, param);
1100
1101                 /* if we didn't send anything, put dummy request on the queue */
1102                 if (ret == 0 && reply->nb_sent == 0) {
1103                         TAILQ_INSERT_TAIL(&pending_requests.requests, dummy,
1104                                         next);
1105                         dummy_used = true;
1106                 }
1107
1108                 pthread_mutex_unlock(&pending_requests.lock);
1109
1110                 /* if we couldn't send anything, clean up */
1111                 if (ret != 0)
1112                         goto fail;
1113                 return 0;
1114         }
1115
1116         /* for primary process, broadcast request */
1117         mp_dir = opendir(mp_dir_path);
1118         if (!mp_dir) {
1119                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1120                 rte_errno = errno;
1121                 goto unlock_fail;
1122         }
1123         dir_fd = dirfd(mp_dir);
1124
1125         /* lock the directory to prevent processes spinning up while we send */
1126         if (flock(dir_fd, LOCK_EX)) {
1127                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1128                         mp_dir_path);
1129                 rte_errno = errno;
1130                 goto closedir_fail;
1131         }
1132
1133         while ((ent = readdir(mp_dir))) {
1134                 char path[PATH_MAX];
1135
1136                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1137                         continue;
1138
1139                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1140                          ent->d_name);
1141
1142                 if (mp_request_async(path, copy, param))
1143                         ret = -1;
1144         }
1145         /* if we didn't send anything, put dummy request on the queue */
1146         if (ret == 0 && reply->nb_sent == 0) {
1147                 TAILQ_INSERT_HEAD(&pending_requests.requests, dummy, next);
1148                 dummy_used = true;
1149         }
1150
1151         /* trigger async request thread wake up */
1152         pthread_cond_signal(&pending_requests.async_cond);
1153
1154         /* finally, unlock the queue */
1155         pthread_mutex_unlock(&pending_requests.lock);
1156
1157         /* unlock the directory */
1158         flock(dir_fd, LOCK_UN);
1159
1160         /* dir_fd automatically closed on closedir */
1161         closedir(mp_dir);
1162
1163         /* if dummy was unused, free it */
1164         if (!dummy_used)
1165                 free(dummy);
1166
1167         return ret;
1168 closedir_fail:
1169         closedir(mp_dir);
1170 unlock_fail:
1171         pthread_mutex_unlock(&pending_requests.lock);
1172 fail:
1173         free(dummy);
1174         free(param);
1175         free(copy);
1176         return -1;
1177 }
1178
1179 int __rte_experimental
1180 rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
1181 {
1182         RTE_LOG(DEBUG, EAL, "reply: %s\n", msg->name);
1183
1184         if (check_input(msg) == false)
1185                 return -1;
1186
1187         if (peer == NULL) {
1188                 RTE_LOG(ERR, EAL, "peer is not specified\n");
1189                 rte_errno = EINVAL;
1190                 return -1;
1191         }
1192
1193         return mp_send(msg, peer, MP_REP);
1194 }