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                         snprintf(dummy.name, sizeof(dummy.name),
 327                                         "%s", msg->name);
 328                         mp_send(&dummy, s->sun_path, MP_IGN);
 329                 } else {
 330                         RTE_LOG(ERR, EAL, "Cannot find action: %s\n",
 331                                 msg->name);
 332                 }
 333         } else if (action(msg, s->sun_path) < 0) {
 334                 RTE_LOG(ERR, EAL, "Fail to handle message: %s\n", msg->name);
 335         }
 336 }
 337
 338 static void *
 339 mp_handle(void *arg __rte_unused)
 340 {
 341         struct mp_msg_internal msg;
 342         struct sockaddr_un sa;
 343
 344         while (1) {
 345                 if (read_msg(&msg, &sa) == 0)
 346                         process_msg(&msg, &sa);
 347         }
 348
 349         return NULL;
 350 }
 351
 352 static int
 353 timespec_cmp(const struct timespec *a, const struct timespec *b)
 354 {
 355         if (a->tv_sec < b->tv_sec)
 356                 return -1;
 357         if (a->tv_sec > b->tv_sec)
 358                 return 1;
 359         if (a->tv_nsec < b->tv_nsec)
 360                 return -1;
 361         if (a->tv_nsec > b->tv_nsec)
 362                 return 1;
 363         return 0;
 364 }
 365
 366 enum async_action {
 367         ACTION_NONE, /**< don't do anything */
 368         ACTION_FREE, /**< free the action entry, but don't trigger callback */
 369         ACTION_TRIGGER /**< trigger callback, then free action entry */
 370 };
 371
 372 static enum async_action
 373 process_async_request(struct pending_request *sr, const struct timespec *now)
 374 {
 375         struct async_request_param *param;
 376         struct rte_mp_reply *reply;
 377         bool timeout, received, last_msg;
 378
 379         param = sr->async.param;
 380         reply = &param->user_reply;
 381
 382         /* did we timeout? */
 383         timeout = timespec_cmp(&param->end, now) <= 0;
 384
 385         /* did we receive a response? */
 386         received = sr->reply_received != 0;
 387
 388         /* if we didn't time out, and we didn't receive a response, ignore */
 389         if (!timeout && !received)
 390                 return ACTION_NONE;
 391
 392         /* if we received a response, adjust relevant data and copy mesasge. */
 393         if (sr->reply_received == 1 && sr->reply) {
 394                 struct rte_mp_msg *msg, *user_msgs, *tmp;
 395
 396                 msg = sr->reply;
 397                 user_msgs = reply->msgs;
 398
 399                 tmp = realloc(user_msgs, sizeof(*msg) *
 400                                 (reply->nb_received + 1));
 401                 if (!tmp) {
 402                         RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 403                                 sr->dst, sr->request->name);
 404                         /* this entry is going to be removed and its message
 405                          * dropped, but we don't want to leak memory, so
 406                          * continue.
 407                          */
 408                 } else {
 409                         user_msgs = tmp;
 410                         reply->msgs = user_msgs;
 411                         memcpy(&user_msgs[reply->nb_received],
 412                                         msg, sizeof(*msg));
 413                         reply->nb_received++;
 414                 }
 415
 416                 /* mark this request as processed */
 417                 param->n_responses_processed++;
 418         } else if (sr->reply_received == -1) {
 419                 /* we were asked to ignore this process */
 420                 reply->nb_sent--;
 421         }
 422         free(sr->reply);
 423
 424         last_msg = param->n_responses_processed == reply->nb_sent;
 425
 426         return last_msg ? ACTION_TRIGGER : ACTION_FREE;
 427 }
 428
 429 static void
 430 trigger_async_action(struct pending_request *sr)
 431 {
 432         struct async_request_param *param;
 433         struct rte_mp_reply *reply;
 434
 435         param = sr->async.param;
 436         reply = &param->user_reply;
 437
 438         param->clb(sr->request, reply);
 439
 440         /* clean up */
 441         free(sr->async.param->user_reply.msgs);
 442         free(sr->async.param);
 443         free(sr->request);
 444 }
 445
 446 static struct pending_request *
 447 check_trigger(struct timespec *ts)
 448 {
 449         struct pending_request *next, *cur, *trigger = NULL;
 450
 451         TAILQ_FOREACH_SAFE(cur, &pending_requests.requests, next, next) {
 452                 enum async_action action;
 453                 if (cur->type != REQUEST_TYPE_ASYNC)
 454                         continue;
 455
 456                 action = process_async_request(cur, ts);
 457                 if (action == ACTION_FREE) {
 458                         TAILQ_REMOVE(&pending_requests.requests, cur, next);
 459                         free(cur);
 460                 } else if (action == ACTION_TRIGGER) {
 461                         TAILQ_REMOVE(&pending_requests.requests, cur, next);
 462                         trigger = cur;
 463                         break;
 464                 }
 465         }
 466         return trigger;
 467 }
 468
 469 static void
 470 wait_for_async_messages(void)
 471 {
 472         struct pending_request *sr;
 473         struct timespec timeout;
 474         bool timedwait = false;
 475         bool nowait = false;
 476         int ret;
 477
 478         /* scan through the list and see if there are any timeouts that
 479          * are earlier than our current timeout.
 480          */
 481         TAILQ_FOREACH(sr, &pending_requests.requests, next) {
 482                 if (sr->type != REQUEST_TYPE_ASYNC)
 483                         continue;
 484                 if (!timedwait || timespec_cmp(&sr->async.param->end,
 485                                 &timeout) < 0) {
 486                         memcpy(&timeout, &sr->async.param->end,
 487                                 sizeof(timeout));
 488                         timedwait = true;
 489                 }
 490
 491                 /* sometimes, we don't even wait */
 492                 if (sr->reply_received) {
 493                         nowait = true;
 494                         break;
 495                 }
 496         }
 497
 498         if (nowait)
 499                 return;
 500
 501         do {
 502                 ret = timedwait ?
 503                         pthread_cond_timedwait(
 504                                 &pending_requests.async_cond,
 505                                 &pending_requests.lock,
 506                                 &timeout) :
 507                         pthread_cond_wait(
 508                                 &pending_requests.async_cond,
 509                                 &pending_requests.lock);
 510         } while (ret != 0 && ret != ETIMEDOUT);
 511
 512         /* we've been woken up or timed out */
 513 }
 514
 515 static void *
 516 async_reply_handle(void *arg __rte_unused)
 517 {
 518         struct timeval now;
 519         struct timespec ts_now;
 520         while (1) {
 521                 struct pending_request *trigger = NULL;
 522
 523                 pthread_mutex_lock(&pending_requests.lock);
 524
 525                 /* we exit this function holding the lock */
 526                 wait_for_async_messages();
 527
 528                 if (gettimeofday(&now, NULL) < 0) {
 529                         RTE_LOG(ERR, EAL, "Cannot get current time\n");
 530                         break;
 531                 }
 532                 ts_now.tv_nsec = now.tv_usec * 1000;
 533                 ts_now.tv_sec = now.tv_sec;
 534
 535                 do {
 536                         trigger = check_trigger(&ts_now);
 537                         /* unlock request list */
 538                         pthread_mutex_unlock(&pending_requests.lock);
 539
 540                         if (trigger) {
 541                                 trigger_async_action(trigger);
 542                                 free(trigger);
 543
 544                                 /* we've triggered a callback, but there may be
 545                                  * more, so lock the list and check again.
 546                                  */
 547                                 pthread_mutex_lock(&pending_requests.lock);
 548                         }
 549                 } while (trigger);
 550         }
 551
 552         RTE_LOG(ERR, EAL, "ERROR: asynchronous requests disabled\n");
 553
 554         return NULL;
 555 }
 556
 557 static int
 558 open_socket_fd(void)
 559 {
 560         char peer_name[PATH_MAX] = {0};
 561         struct sockaddr_un un;
 562
 563         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 564                 snprintf(peer_name, sizeof(peer_name),
 565                                 "%d_%"PRIx64, getpid(), rte_rdtsc());
 566
 567         mp_fd = socket(AF_UNIX, SOCK_DGRAM, 0);
 568         if (mp_fd < 0) {
 569                 RTE_LOG(ERR, EAL, "failed to create unix socket\n");
 570                 return -1;
 571         }
 572
 573         memset(&un, 0, sizeof(un));
 574         un.sun_family = AF_UNIX;
 575
 576         create_socket_path(peer_name, un.sun_path, sizeof(un.sun_path));
 577
 578         unlink(un.sun_path); /* May still exist since last run */
 579
 580         if (bind(mp_fd, (struct sockaddr *)&un, sizeof(un)) < 0) {
 581                 RTE_LOG(ERR, EAL, "failed to bind %s: %s\n",
 582                         un.sun_path, strerror(errno));
 583                 close(mp_fd);
 584                 return -1;
 585         }
 586
 587         RTE_LOG(INFO, EAL, "Multi-process socket %s\n", un.sun_path);
 588         return mp_fd;
 589 }
 590
 591 static int
 592 unlink_sockets(const char *filter)
 593 {
 594         int dir_fd;
 595         DIR *mp_dir;
 596         struct dirent *ent;
 597
 598         mp_dir = opendir(mp_dir_path);
 599         if (!mp_dir) {
 600                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
 601                 return -1;
 602         }
 603         dir_fd = dirfd(mp_dir);
 604
 605         while ((ent = readdir(mp_dir))) {
 606                 if (fnmatch(filter, ent->d_name, 0) == 0)
 607                         unlinkat(dir_fd, ent->d_name, 0);
 608         }
 609
 610         closedir(mp_dir);
 611         return 0;
 612 }
 613
 614 int
 615 rte_mp_channel_init(void)
 616 {
 617         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 618         char path[PATH_MAX];
 619         int dir_fd;
 620         pthread_t mp_handle_tid, async_reply_handle_tid;
 621
 622         /* create filter path */
 623         create_socket_path("*", path, sizeof(path));
 624         snprintf(mp_filter, sizeof(mp_filter), "%s", basename(path));
 625
 626         /* path may have been modified, so recreate it */
 627         create_socket_path("*", path, sizeof(path));
 628         snprintf(mp_dir_path, sizeof(mp_dir_path), "%s", dirname(path));
 629
 630         /* lock the directory */
 631         dir_fd = open(mp_dir_path, O_RDONLY);
 632         if (dir_fd < 0) {
 633                 RTE_LOG(ERR, EAL, "failed to open %s: %s\n",
 634                         mp_dir_path, strerror(errno));
 635                 return -1;
 636         }
 637
 638         if (flock(dir_fd, LOCK_EX)) {
 639                 RTE_LOG(ERR, EAL, "failed to lock %s: %s\n",
 640                         mp_dir_path, strerror(errno));
 641                 close(dir_fd);
 642                 return -1;
 643         }
 644
 645         if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
 646                         unlink_sockets(mp_filter)) {
 647                 RTE_LOG(ERR, EAL, "failed to unlink mp sockets\n");
 648                 close(dir_fd);
 649                 return -1;
 650         }
 651
 652         if (open_socket_fd() < 0) {
 653                 close(dir_fd);
 654                 return -1;
 655         }
 656
 657         if (pthread_create(&mp_handle_tid, NULL, mp_handle, NULL) < 0) {
 658                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 659                         strerror(errno));
 660                 close(mp_fd);
 661                 mp_fd = -1;
 662                 return -1;
 663         }
 664
 665         if (pthread_create(&async_reply_handle_tid, NULL,
 666                         async_reply_handle, NULL) < 0) {
 667                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 668                         strerror(errno));
 669                 close(mp_fd);
 670                 close(dir_fd);
 671                 mp_fd = -1;
 672                 return -1;
 673         }
 674
 675         /* try best to set thread name */
 676         snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN, "rte_mp_handle");
 677         rte_thread_setname(mp_handle_tid, thread_name);
 678
 679         /* try best to set thread name */
 680         snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN, "rte_mp_async_handle");
 681         rte_thread_setname(async_reply_handle_tid, thread_name);
 682
 683         /* unlock the directory */
 684         flock(dir_fd, LOCK_UN);
 685         close(dir_fd);
 686
 687         return 0;
 688 }
 689
 690 /**
 691  * Return -1, as fail to send message and it's caused by the local side.
 692  * Return 0, as fail to send message and it's caused by the remote side.
 693  * Return 1, as succeed to send message.
 694  *
 695  */
 696 static int
 697 send_msg(const char *dst_path, struct rte_mp_msg *msg, int type)
 698 {
 699         int snd;
 700         struct iovec iov;
 701         struct msghdr msgh;
 702         struct cmsghdr *cmsg;
 703         struct sockaddr_un dst;
 704         struct mp_msg_internal m;
 705         int fd_size = msg->num_fds * sizeof(int);
 706         char control[CMSG_SPACE(fd_size)];
 707
 708         m.type = type;
 709         memcpy(&m.msg, msg, sizeof(*msg));
 710
 711         memset(&dst, 0, sizeof(dst));
 712         dst.sun_family = AF_UNIX;
 713         snprintf(dst.sun_path, sizeof(dst.sun_path), "%s", dst_path);
 714
 715         memset(&msgh, 0, sizeof(msgh));
 716         memset(control, 0, sizeof(control));
 717
 718         iov.iov_base = &m;
 719         iov.iov_len = sizeof(m) - sizeof(msg->fds);
 720
 721         msgh.msg_name = &dst;
 722         msgh.msg_namelen = sizeof(dst);
 723         msgh.msg_iov = &iov;
 724         msgh.msg_iovlen = 1;
 725         msgh.msg_control = control;
 726         msgh.msg_controllen = sizeof(control);
 727
 728         cmsg = CMSG_FIRSTHDR(&msgh);
 729         cmsg->cmsg_len = CMSG_LEN(fd_size);
 730         cmsg->cmsg_level = SOL_SOCKET;
 731         cmsg->cmsg_type = SCM_RIGHTS;
 732         memcpy(CMSG_DATA(cmsg), msg->fds, fd_size);
 733
 734         do {
 735                 snd = sendmsg(mp_fd, &msgh, 0);
 736         } while (snd < 0 && errno == EINTR);
 737
 738         if (snd < 0) {
 739                 rte_errno = errno;
 740                 /* Check if it caused by peer process exits */
 741                 if (errno == ECONNREFUSED &&
 742                                 rte_eal_process_type() == RTE_PROC_PRIMARY) {
 743                         unlink(dst_path);
 744                         return 0;
 745                 }
 746                 if (errno == ENOBUFS) {
 747                         RTE_LOG(ERR, EAL, "Peer cannot receive message %s\n",
 748                                 dst_path);
 749                         return 0;
 750                 }
 751                 RTE_LOG(ERR, EAL, "failed to send to (%s) due to %s\n",
 752                         dst_path, strerror(errno));
 753                 return -1;
 754         }
 755
 756         return 1;
 757 }
 758
 759 static int
 760 mp_send(struct rte_mp_msg *msg, const char *peer, int type)
 761 {
 762         int dir_fd, ret = 0;
 763         DIR *mp_dir;
 764         struct dirent *ent;
 765
 766         if (!peer && (rte_eal_process_type() == RTE_PROC_SECONDARY))
 767                 peer = eal_mp_socket_path();
 768
 769         if (peer) {
 770                 if (send_msg(peer, msg, type) < 0)
 771                         return -1;
 772                 else
 773                         return 0;
 774         }
 775
 776         /* broadcast to all secondary processes */
 777         mp_dir = opendir(mp_dir_path);
 778         if (!mp_dir) {
 779                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n",
 780                                 mp_dir_path);
 781                 rte_errno = errno;
 782                 return -1;
 783         }
 784
 785         dir_fd = dirfd(mp_dir);
 786         /* lock the directory to prevent processes spinning up while we send */
 787         if (flock(dir_fd, LOCK_EX)) {
 788                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
 789                         mp_dir_path);
 790                 rte_errno = errno;
 791                 closedir(mp_dir);
 792                 return -1;
 793         }
 794
 795         while ((ent = readdir(mp_dir))) {
 796                 char path[PATH_MAX];
 797
 798                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
 799                         continue;
 800
 801                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
 802                          ent->d_name);
 803                 if (send_msg(path, msg, type) < 0)
 804                         ret = -1;
 805         }
 806         /* unlock the dir */
 807         flock(dir_fd, LOCK_UN);
 808
 809         /* dir_fd automatically closed on closedir */
 810         closedir(mp_dir);
 811         return ret;
 812 }
 813
 814 static bool
 815 check_input(const struct rte_mp_msg *msg)
 816 {
 817         if (msg == NULL) {
 818                 RTE_LOG(ERR, EAL, "Msg cannot be NULL\n");
 819                 rte_errno = EINVAL;
 820                 return false;
 821         }
 822
 823         if (validate_action_name(msg->name))
 824                 return false;
 825
 826         if (msg->len_param > RTE_MP_MAX_PARAM_LEN) {
 827                 RTE_LOG(ERR, EAL, "Message data is too long\n");
 828                 rte_errno = E2BIG;
 829                 return false;
 830         }
 831
 832         if (msg->num_fds > RTE_MP_MAX_FD_NUM) {
 833                 RTE_LOG(ERR, EAL, "Cannot send more than %d FDs\n",
 834                         RTE_MP_MAX_FD_NUM);
 835                 rte_errno = E2BIG;
 836                 return false;
 837         }
 838
 839         return true;
 840 }
 841
 842 int __rte_experimental
 843 rte_mp_sendmsg(struct rte_mp_msg *msg)
 844 {
 845         if (!check_input(msg))
 846                 return -1;
 847
 848         RTE_LOG(DEBUG, EAL, "sendmsg: %s\n", msg->name);
 849         return mp_send(msg, NULL, MP_MSG);
 850 }
 851
 852 static int
 853 mp_request_async(const char *dst, struct rte_mp_msg *req,
 854                 struct async_request_param *param)
 855 {
 856         struct rte_mp_msg *reply_msg;
 857         struct pending_request *sync_req, *exist;
 858         int ret;
 859
 860         sync_req = malloc(sizeof(*sync_req));
 861         reply_msg = malloc(sizeof(*reply_msg));
 862         if (sync_req == NULL || reply_msg == NULL) {
 863                 RTE_LOG(ERR, EAL, "Could not allocate space for sync request\n");
 864                 rte_errno = ENOMEM;
 865                 ret = -1;
 866                 goto fail;
 867         }
 868
 869         memset(sync_req, 0, sizeof(*sync_req));
 870         memset(reply_msg, 0, sizeof(*reply_msg));
 871
 872         sync_req->type = REQUEST_TYPE_ASYNC;
 873         strcpy(sync_req->dst, dst);
 874         sync_req->request = req;
 875         sync_req->reply = reply_msg;
 876         sync_req->async.param = param;
 877
 878         /* queue already locked by caller */
 879
 880         exist = find_sync_request(dst, req->name);
 881         if (exist) {
 882                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 883                 rte_errno = EEXIST;
 884                 ret = -1;
 885                 goto fail;
 886         }
 887
 888         ret = send_msg(dst, req, MP_REQ);
 889         if (ret < 0) {
 890                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 891                         dst, req->name);
 892                 ret = -1;
 893                 goto fail;
 894         } else if (ret == 0) {
 895                 ret = 0;
 896                 goto fail;
 897         }
 898         TAILQ_INSERT_TAIL(&pending_requests.requests, sync_req, next);
 899
 900         param->user_reply.nb_sent++;
 901
 902         return 0;
 903 fail:
 904         free(sync_req);
 905         free(reply_msg);
 906         return ret;
 907 }
 908
 909 static int
 910 mp_request_sync(const char *dst, struct rte_mp_msg *req,
 911                struct rte_mp_reply *reply, const struct timespec *ts)
 912 {
 913         int ret;
 914         struct rte_mp_msg msg, *tmp;
 915         struct pending_request sync_req, *exist;
 916
 917         sync_req.type = REQUEST_TYPE_SYNC;
 918         sync_req.reply_received = 0;
 919         strcpy(sync_req.dst, dst);
 920         sync_req.request = req;
 921         sync_req.reply = &msg;
 922         pthread_cond_init(&sync_req.sync.cond, NULL);
 923
 924         exist = find_sync_request(dst, req->name);
 925         if (exist) {
 926                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 927                 rte_errno = EEXIST;
 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         TAILQ_REMOVE(&pending_requests.requests, &sync_req, next);
 949
 950         if (sync_req.reply_received == 0) {
 951                 RTE_LOG(ERR, EAL, "Fail to recv reply for request %s:%s\n",
 952                         dst, req->name);
 953                 rte_errno = ETIMEDOUT;
 954                 return -1;
 955         }
 956         if (sync_req.reply_received == -1) {
 957                 RTE_LOG(DEBUG, EAL, "Asked to ignore response\n");
 958                 /* not receiving this message is not an error, so decrement
 959                  * number of sent messages
 960                  */
 961                 reply->nb_sent--;
 962                 return 0;
 963         }
 964
 965         tmp = realloc(reply->msgs, sizeof(msg) * (reply->nb_received + 1));
 966         if (!tmp) {
 967                 RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 968                         dst, req->name);
 969                 rte_errno = ENOMEM;
 970                 return -1;
 971         }
 972         memcpy(&tmp[reply->nb_received], &msg, sizeof(msg));
 973         reply->msgs = tmp;
 974         reply->nb_received++;
 975         return 0;
 976 }
 977
 978 int __rte_experimental
 979 rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
 980                 const struct timespec *ts)
 981 {
 982         int dir_fd, ret = 0;
 983         DIR *mp_dir;
 984         struct dirent *ent;
 985         struct timeval now;
 986         struct timespec end;
 987
 988         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
 989
 990         if (check_input(req) == false)
 991                 return -1;
 992         if (gettimeofday(&now, NULL) < 0) {
 993                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
 994                 rte_errno = errno;
 995                 return -1;
 996         }
 997
 998         end.tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
 999         end.tv_sec = now.tv_sec + ts->tv_sec +
1000                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1001
1002         reply->nb_sent = 0;
1003         reply->nb_received = 0;
1004         reply->msgs = NULL;
1005
1006         /* for secondary process, send request to the primary process only */
1007         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1008                 pthread_mutex_lock(&pending_requests.lock);
1009                 ret = mp_request_sync(eal_mp_socket_path(), req, reply, &end);
1010                 pthread_mutex_unlock(&pending_requests.lock);
1011                 return ret;
1012         }
1013
1014         /* for primary process, broadcast request, and collect reply 1 by 1 */
1015         mp_dir = opendir(mp_dir_path);
1016         if (!mp_dir) {
1017                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1018                 rte_errno = errno;
1019                 return -1;
1020         }
1021
1022         dir_fd = dirfd(mp_dir);
1023         /* lock the directory to prevent processes spinning up while we send */
1024         if (flock(dir_fd, LOCK_EX)) {
1025                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1026                         mp_dir_path);
1027                 closedir(mp_dir);
1028                 rte_errno = errno;
1029                 return -1;
1030         }
1031
1032         pthread_mutex_lock(&pending_requests.lock);
1033         while ((ent = readdir(mp_dir))) {
1034                 char path[PATH_MAX];
1035
1036                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1037                         continue;
1038
1039                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1040                          ent->d_name);
1041
1042                 /* unlocks the mutex while waiting for response,
1043                  * locks on receive
1044                  */
1045                 if (mp_request_sync(path, req, reply, &end))
1046                         ret = -1;
1047         }
1048         pthread_mutex_unlock(&pending_requests.lock);
1049         /* unlock the directory */
1050         flock(dir_fd, LOCK_UN);
1051
1052         /* dir_fd automatically closed on closedir */
1053         closedir(mp_dir);
1054         return ret;
1055 }
1056
1057 int __rte_experimental
1058 rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
1059                 rte_mp_async_reply_t clb)
1060 {
1061         struct rte_mp_msg *copy;
1062         struct pending_request *dummy;
1063         struct async_request_param *param;
1064         struct rte_mp_reply *reply;
1065         int dir_fd, ret = 0;
1066         DIR *mp_dir;
1067         struct dirent *ent;
1068         struct timeval now;
1069         struct timespec *end;
1070         bool dummy_used = false;
1071
1072         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
1073
1074         if (check_input(req) == false)
1075                 return -1;
1076         if (gettimeofday(&now, NULL) < 0) {
1077                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
1078                 rte_errno = errno;
1079                 return -1;
1080         }
1081         copy = malloc(sizeof(*copy));
1082         dummy = malloc(sizeof(*dummy));
1083         param = malloc(sizeof(*param));
1084         if (copy == NULL || dummy == NULL || param == NULL) {
1085                 RTE_LOG(ERR, EAL, "Failed to allocate memory for async reply\n");
1086                 rte_errno = ENOMEM;
1087                 goto fail;
1088         }
1089
1090         memset(copy, 0, sizeof(*copy));
1091         memset(dummy, 0, sizeof(*dummy));
1092         memset(param, 0, sizeof(*param));
1093
1094         /* copy message */
1095         memcpy(copy, req, sizeof(*copy));
1096
1097         param->n_responses_processed = 0;
1098         param->clb = clb;
1099         end = &param->end;
1100         reply = &param->user_reply;
1101
1102         end->tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
1103         end->tv_sec = now.tv_sec + ts->tv_sec +
1104                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1105         reply->nb_sent = 0;
1106         reply->nb_received = 0;
1107         reply->msgs = NULL;
1108
1109         /* we have to lock the request queue here, as we will be adding a bunch
1110          * of requests to the queue at once, and some of the replies may arrive
1111          * before we add all of the requests to the queue.
1112          */
1113         pthread_mutex_lock(&pending_requests.lock);
1114
1115         /* we have to ensure that callback gets triggered even if we don't send
1116          * anything, therefore earlier we have allocated a dummy request. fill
1117          * it, and put it on the queue if we don't send any requests.
1118          */
1119         dummy->type = REQUEST_TYPE_ASYNC;
1120         dummy->request = copy;
1121         dummy->reply = NULL;
1122         dummy->async.param = param;
1123         dummy->reply_received = 1; /* short-circuit the timeout */
1124
1125         /* for secondary process, send request to the primary process only */
1126         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1127                 ret = mp_request_async(eal_mp_socket_path(), copy, param);
1128
1129                 /* if we didn't send anything, put dummy request on the queue */
1130                 if (ret == 0 && reply->nb_sent == 0) {
1131                         TAILQ_INSERT_TAIL(&pending_requests.requests, dummy,
1132                                         next);
1133                         dummy_used = true;
1134                 }
1135
1136                 pthread_mutex_unlock(&pending_requests.lock);
1137
1138                 /* if we couldn't send anything, clean up */
1139                 if (ret != 0)
1140                         goto fail;
1141                 return 0;
1142         }
1143
1144         /* for primary process, broadcast request */
1145         mp_dir = opendir(mp_dir_path);
1146         if (!mp_dir) {
1147                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1148                 rte_errno = errno;
1149                 goto unlock_fail;
1150         }
1151         dir_fd = dirfd(mp_dir);
1152
1153         /* lock the directory to prevent processes spinning up while we send */
1154         if (flock(dir_fd, LOCK_EX)) {
1155                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1156                         mp_dir_path);
1157                 rte_errno = errno;
1158                 goto closedir_fail;
1159         }
1160
1161         while ((ent = readdir(mp_dir))) {
1162                 char path[PATH_MAX];
1163
1164                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1165                         continue;
1166
1167                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1168                          ent->d_name);
1169
1170                 if (mp_request_async(path, copy, param))
1171                         ret = -1;
1172         }
1173         /* if we didn't send anything, put dummy request on the queue */
1174         if (ret == 0 && reply->nb_sent == 0) {
1175                 TAILQ_INSERT_HEAD(&pending_requests.requests, dummy, next);
1176                 dummy_used = true;
1177         }
1178
1179         /* trigger async request thread wake up */
1180         pthread_cond_signal(&pending_requests.async_cond);
1181
1182         /* finally, unlock the queue */
1183         pthread_mutex_unlock(&pending_requests.lock);
1184
1185         /* unlock the directory */
1186         flock(dir_fd, LOCK_UN);
1187
1188         /* dir_fd automatically closed on closedir */
1189         closedir(mp_dir);
1190
1191         /* if dummy was unused, free it */
1192         if (!dummy_used)
1193                 free(dummy);
1194
1195         return ret;
1196 closedir_fail:
1197         closedir(mp_dir);
1198 unlock_fail:
1199         pthread_mutex_unlock(&pending_requests.lock);
1200 fail:
1201         free(dummy);
1202         free(param);
1203         free(copy);
1204         return -1;
1205 }
1206
1207 int __rte_experimental
1208 rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
1209 {
1210         RTE_LOG(DEBUG, EAL, "reply: %s\n", msg->name);
1211
1212         if (check_input(msg) == false)
1213                 return -1;
1214
1215         if (peer == NULL) {
1216                 RTE_LOG(ERR, EAL, "peer is not specified\n");
1217                 rte_errno = EINVAL;
1218                 return -1;
1219         }
1220
1221         return mp_send(msg, peer, MP_REP);
1222 }