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                 strlcpy(buf, prefix, len);
 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         strlcpy(entry->action_name, name, sizeof(entry->action_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                         strlcpy(dummy.name, msg->name, sizeof(dummy.name));
 327                         mp_send(&dummy, s->sun_path, MP_IGN);
 328                 } else {
 329                         RTE_LOG(ERR, EAL, "Cannot find action: %s\n",
 330                                 msg->name);
 331                 }
 332         } else if (action(msg, s->sun_path) < 0) {
 333                 RTE_LOG(ERR, EAL, "Fail to handle message: %s\n", msg->name);
 334         }
 335 }
 336
 337 static void *
 338 mp_handle(void *arg __rte_unused)
 339 {
 340         struct mp_msg_internal msg;
 341         struct sockaddr_un sa;
 342
 343         while (1) {
 344                 if (read_msg(&msg, &sa) == 0)
 345                         process_msg(&msg, &sa);
 346         }
 347
 348         return NULL;
 349 }
 350
 351 static int
 352 timespec_cmp(const struct timespec *a, const struct timespec *b)
 353 {
 354         if (a->tv_sec < b->tv_sec)
 355                 return -1;
 356         if (a->tv_sec > b->tv_sec)
 357                 return 1;
 358         if (a->tv_nsec < b->tv_nsec)
 359                 return -1;
 360         if (a->tv_nsec > b->tv_nsec)
 361                 return 1;
 362         return 0;
 363 }
 364
 365 enum async_action {
 366         ACTION_NONE, /**< don't do anything */
 367         ACTION_FREE, /**< free the action entry, but don't trigger callback */
 368         ACTION_TRIGGER /**< trigger callback, then free action entry */
 369 };
 370
 371 static enum async_action
 372 process_async_request(struct pending_request *sr, const struct timespec *now)
 373 {
 374         struct async_request_param *param;
 375         struct rte_mp_reply *reply;
 376         bool timeout, received, last_msg;
 377
 378         param = sr->async.param;
 379         reply = &param->user_reply;
 380
 381         /* did we timeout? */
 382         timeout = timespec_cmp(&param->end, now) <= 0;
 383
 384         /* did we receive a response? */
 385         received = sr->reply_received != 0;
 386
 387         /* if we didn't time out, and we didn't receive a response, ignore */
 388         if (!timeout && !received)
 389                 return ACTION_NONE;
 390
 391         /* if we received a response, adjust relevant data and copy mesasge. */
 392         if (sr->reply_received == 1 && sr->reply) {
 393                 struct rte_mp_msg *msg, *user_msgs, *tmp;
 394
 395                 msg = sr->reply;
 396                 user_msgs = reply->msgs;
 397
 398                 tmp = realloc(user_msgs, sizeof(*msg) *
 399                                 (reply->nb_received + 1));
 400                 if (!tmp) {
 401                         RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 402                                 sr->dst, sr->request->name);
 403                         /* this entry is going to be removed and its message
 404                          * dropped, but we don't want to leak memory, so
 405                          * continue.
 406                          */
 407                 } else {
 408                         user_msgs = tmp;
 409                         reply->msgs = user_msgs;
 410                         memcpy(&user_msgs[reply->nb_received],
 411                                         msg, sizeof(*msg));
 412                         reply->nb_received++;
 413                 }
 414
 415                 /* mark this request as processed */
 416                 param->n_responses_processed++;
 417         } else if (sr->reply_received == -1) {
 418                 /* we were asked to ignore this process */
 419                 reply->nb_sent--;
 420         }
 421         free(sr->reply);
 422
 423         last_msg = param->n_responses_processed == reply->nb_sent;
 424
 425         return last_msg ? ACTION_TRIGGER : ACTION_FREE;
 426 }
 427
 428 static void
 429 trigger_async_action(struct pending_request *sr)
 430 {
 431         struct async_request_param *param;
 432         struct rte_mp_reply *reply;
 433
 434         param = sr->async.param;
 435         reply = &param->user_reply;
 436
 437         param->clb(sr->request, reply);
 438
 439         /* clean up */
 440         free(sr->async.param->user_reply.msgs);
 441         free(sr->async.param);
 442         free(sr->request);
 443 }
 444
 445 static struct pending_request *
 446 check_trigger(struct timespec *ts)
 447 {
 448         struct pending_request *next, *cur, *trigger = NULL;
 449
 450         TAILQ_FOREACH_SAFE(cur, &pending_requests.requests, next, next) {
 451                 enum async_action action;
 452                 if (cur->type != REQUEST_TYPE_ASYNC)
 453                         continue;
 454
 455                 action = process_async_request(cur, ts);
 456                 if (action == ACTION_FREE) {
 457                         TAILQ_REMOVE(&pending_requests.requests, cur, next);
 458                         free(cur);
 459                 } else if (action == ACTION_TRIGGER) {
 460                         TAILQ_REMOVE(&pending_requests.requests, cur, next);
 461                         trigger = cur;
 462                         break;
 463                 }
 464         }
 465         return trigger;
 466 }
 467
 468 static void
 469 wait_for_async_messages(void)
 470 {
 471         struct pending_request *sr;
 472         struct timespec timeout;
 473         bool timedwait = false;
 474         bool nowait = false;
 475         int ret;
 476
 477         /* scan through the list and see if there are any timeouts that
 478          * are earlier than our current timeout.
 479          */
 480         TAILQ_FOREACH(sr, &pending_requests.requests, next) {
 481                 if (sr->type != REQUEST_TYPE_ASYNC)
 482                         continue;
 483                 if (!timedwait || timespec_cmp(&sr->async.param->end,
 484                                 &timeout) < 0) {
 485                         memcpy(&timeout, &sr->async.param->end,
 486                                 sizeof(timeout));
 487                         timedwait = true;
 488                 }
 489
 490                 /* sometimes, we don't even wait */
 491                 if (sr->reply_received) {
 492                         nowait = true;
 493                         break;
 494                 }
 495         }
 496
 497         if (nowait)
 498                 return;
 499
 500         do {
 501                 ret = timedwait ?
 502                         pthread_cond_timedwait(
 503                                 &pending_requests.async_cond,
 504                                 &pending_requests.lock,
 505                                 &timeout) :
 506                         pthread_cond_wait(
 507                                 &pending_requests.async_cond,
 508                                 &pending_requests.lock);
 509         } while (ret != 0 && ret != ETIMEDOUT);
 510
 511         /* we've been woken up or timed out */
 512 }
 513
 514 static void *
 515 async_reply_handle(void *arg __rte_unused)
 516 {
 517         struct timeval now;
 518         struct timespec ts_now;
 519         while (1) {
 520                 struct pending_request *trigger = NULL;
 521
 522                 pthread_mutex_lock(&pending_requests.lock);
 523
 524                 /* we exit this function holding the lock */
 525                 wait_for_async_messages();
 526
 527                 if (gettimeofday(&now, NULL) < 0) {
 528                         RTE_LOG(ERR, EAL, "Cannot get current time\n");
 529                         break;
 530                 }
 531                 ts_now.tv_nsec = now.tv_usec * 1000;
 532                 ts_now.tv_sec = now.tv_sec;
 533
 534                 do {
 535                         trigger = check_trigger(&ts_now);
 536                         /* unlock request list */
 537                         pthread_mutex_unlock(&pending_requests.lock);
 538
 539                         if (trigger) {
 540                                 trigger_async_action(trigger);
 541                                 free(trigger);
 542
 543                                 /* we've triggered a callback, but there may be
 544                                  * more, so lock the list and check again.
 545                                  */
 546                                 pthread_mutex_lock(&pending_requests.lock);
 547                         }
 548                 } while (trigger);
 549         }
 550
 551         RTE_LOG(ERR, EAL, "ERROR: asynchronous requests disabled\n");
 552
 553         return NULL;
 554 }
 555
 556 static int
 557 open_socket_fd(void)
 558 {
 559         char peer_name[PATH_MAX] = {0};
 560         struct sockaddr_un un;
 561
 562         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 563                 snprintf(peer_name, sizeof(peer_name),
 564                                 "%d_%"PRIx64, getpid(), rte_rdtsc());
 565
 566         mp_fd = socket(AF_UNIX, SOCK_DGRAM, 0);
 567         if (mp_fd < 0) {
 568                 RTE_LOG(ERR, EAL, "failed to create unix socket\n");
 569                 return -1;
 570         }
 571
 572         memset(&un, 0, sizeof(un));
 573         un.sun_family = AF_UNIX;
 574
 575         create_socket_path(peer_name, un.sun_path, sizeof(un.sun_path));
 576
 577         unlink(un.sun_path); /* May still exist since last run */
 578
 579         if (bind(mp_fd, (struct sockaddr *)&un, sizeof(un)) < 0) {
 580                 RTE_LOG(ERR, EAL, "failed to bind %s: %s\n",
 581                         un.sun_path, strerror(errno));
 582                 close(mp_fd);
 583                 return -1;
 584         }
 585
 586         RTE_LOG(INFO, EAL, "Multi-process socket %s\n", un.sun_path);
 587         return mp_fd;
 588 }
 589
 590 static int
 591 unlink_sockets(const char *filter)
 592 {
 593         int dir_fd;
 594         DIR *mp_dir;
 595         struct dirent *ent;
 596
 597         mp_dir = opendir(mp_dir_path);
 598         if (!mp_dir) {
 599                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
 600                 return -1;
 601         }
 602         dir_fd = dirfd(mp_dir);
 603
 604         while ((ent = readdir(mp_dir))) {
 605                 if (fnmatch(filter, ent->d_name, 0) == 0)
 606                         unlinkat(dir_fd, ent->d_name, 0);
 607         }
 608
 609         closedir(mp_dir);
 610         return 0;
 611 }
 612
 613 int
 614 rte_mp_channel_init(void)
 615 {
 616         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 617         char path[PATH_MAX];
 618         int dir_fd;
 619         pthread_t mp_handle_tid, async_reply_handle_tid;
 620
 621         /* create filter path */
 622         create_socket_path("*", path, sizeof(path));
 623         strlcpy(mp_filter, basename(path), sizeof(mp_filter));
 624
 625         /* path may have been modified, so recreate it */
 626         create_socket_path("*", path, sizeof(path));
 627         strlcpy(mp_dir_path, dirname(path), sizeof(mp_dir_path));
 628
 629         /* lock the directory */
 630         dir_fd = open(mp_dir_path, O_RDONLY);
 631         if (dir_fd < 0) {
 632                 RTE_LOG(ERR, EAL, "failed to open %s: %s\n",
 633                         mp_dir_path, strerror(errno));
 634                 return -1;
 635         }
 636
 637         if (flock(dir_fd, LOCK_EX)) {
 638                 RTE_LOG(ERR, EAL, "failed to lock %s: %s\n",
 639                         mp_dir_path, strerror(errno));
 640                 close(dir_fd);
 641                 return -1;
 642         }
 643
 644         if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
 645                         unlink_sockets(mp_filter)) {
 646                 RTE_LOG(ERR, EAL, "failed to unlink mp sockets\n");
 647                 close(dir_fd);
 648                 return -1;
 649         }
 650
 651         if (open_socket_fd() < 0) {
 652                 close(dir_fd);
 653                 return -1;
 654         }
 655
 656         if (pthread_create(&mp_handle_tid, NULL, mp_handle, NULL) < 0) {
 657                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 658                         strerror(errno));
 659                 close(mp_fd);
 660                 mp_fd = -1;
 661                 return -1;
 662         }
 663
 664         if (pthread_create(&async_reply_handle_tid, NULL,
 665                         async_reply_handle, NULL) < 0) {
 666                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 667                         strerror(errno));
 668                 close(mp_fd);
 669                 close(dir_fd);
 670                 mp_fd = -1;
 671                 return -1;
 672         }
 673
 674         /* try best to set thread name */
 675         strlcpy(thread_name, "rte_mp_handle", RTE_MAX_THREAD_NAME_LEN);
 676         rte_thread_setname(mp_handle_tid, thread_name);
 677
 678         /* try best to set thread name */
 679         strlcpy(thread_name, "rte_mp_async_handle", RTE_MAX_THREAD_NAME_LEN);
 680         rte_thread_setname(async_reply_handle_tid, thread_name);
 681
 682         /* unlock the directory */
 683         flock(dir_fd, LOCK_UN);
 684         close(dir_fd);
 685
 686         return 0;
 687 }
 688
 689 /**
 690  * Return -1, as fail to send message and it's caused by the local side.
 691  * Return 0, as fail to send message and it's caused by the remote side.
 692  * Return 1, as succeed to send message.
 693  *
 694  */
 695 static int
 696 send_msg(const char *dst_path, struct rte_mp_msg *msg, int type)
 697 {
 698         int snd;
 699         struct iovec iov;
 700         struct msghdr msgh;
 701         struct cmsghdr *cmsg;
 702         struct sockaddr_un dst;
 703         struct mp_msg_internal m;
 704         int fd_size = msg->num_fds * sizeof(int);
 705         char control[CMSG_SPACE(fd_size)];
 706
 707         m.type = type;
 708         memcpy(&m.msg, msg, sizeof(*msg));
 709
 710         memset(&dst, 0, sizeof(dst));
 711         dst.sun_family = AF_UNIX;
 712         strlcpy(dst.sun_path, dst_path, sizeof(dst.sun_path));
 713
 714         memset(&msgh, 0, sizeof(msgh));
 715         memset(control, 0, sizeof(control));
 716
 717         iov.iov_base = &m;
 718         iov.iov_len = sizeof(m) - sizeof(msg->fds);
 719
 720         msgh.msg_name = &dst;
 721         msgh.msg_namelen = sizeof(dst);
 722         msgh.msg_iov = &iov;
 723         msgh.msg_iovlen = 1;
 724         msgh.msg_control = control;
 725         msgh.msg_controllen = sizeof(control);
 726
 727         cmsg = CMSG_FIRSTHDR(&msgh);
 728         cmsg->cmsg_len = CMSG_LEN(fd_size);
 729         cmsg->cmsg_level = SOL_SOCKET;
 730         cmsg->cmsg_type = SCM_RIGHTS;
 731         memcpy(CMSG_DATA(cmsg), msg->fds, fd_size);
 732
 733         do {
 734                 snd = sendmsg(mp_fd, &msgh, 0);
 735         } while (snd < 0 && errno == EINTR);
 736
 737         if (snd < 0) {
 738                 rte_errno = errno;
 739                 /* Check if it caused by peer process exits */
 740                 if (errno == ECONNREFUSED &&
 741                                 rte_eal_process_type() == RTE_PROC_PRIMARY) {
 742                         unlink(dst_path);
 743                         return 0;
 744                 }
 745                 if (errno == ENOBUFS) {
 746                         RTE_LOG(ERR, EAL, "Peer cannot receive message %s\n",
 747                                 dst_path);
 748                         return 0;
 749                 }
 750                 RTE_LOG(ERR, EAL, "failed to send to (%s) due to %s\n",
 751                         dst_path, strerror(errno));
 752                 return -1;
 753         }
 754
 755         return 1;
 756 }
 757
 758 static int
 759 mp_send(struct rte_mp_msg *msg, const char *peer, int type)
 760 {
 761         int dir_fd, ret = 0;
 762         DIR *mp_dir;
 763         struct dirent *ent;
 764
 765         if (!peer && (rte_eal_process_type() == RTE_PROC_SECONDARY))
 766                 peer = eal_mp_socket_path();
 767
 768         if (peer) {
 769                 if (send_msg(peer, msg, type) < 0)
 770                         return -1;
 771                 else
 772                         return 0;
 773         }
 774
 775         /* broadcast to all secondary processes */
 776         mp_dir = opendir(mp_dir_path);
 777         if (!mp_dir) {
 778                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n",
 779                                 mp_dir_path);
 780                 rte_errno = errno;
 781                 return -1;
 782         }
 783
 784         dir_fd = dirfd(mp_dir);
 785         /* lock the directory to prevent processes spinning up while we send */
 786         if (flock(dir_fd, LOCK_EX)) {
 787                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
 788                         mp_dir_path);
 789                 rte_errno = errno;
 790                 closedir(mp_dir);
 791                 return -1;
 792         }
 793
 794         while ((ent = readdir(mp_dir))) {
 795                 char path[PATH_MAX];
 796
 797                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
 798                         continue;
 799
 800                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
 801                          ent->d_name);
 802                 if (send_msg(path, msg, type) < 0)
 803                         ret = -1;
 804         }
 805         /* unlock the dir */
 806         flock(dir_fd, LOCK_UN);
 807
 808         /* dir_fd automatically closed on closedir */
 809         closedir(mp_dir);
 810         return ret;
 811 }
 812
 813 static bool
 814 check_input(const struct rte_mp_msg *msg)
 815 {
 816         if (msg == NULL) {
 817                 RTE_LOG(ERR, EAL, "Msg cannot be NULL\n");
 818                 rte_errno = EINVAL;
 819                 return false;
 820         }
 821
 822         if (validate_action_name(msg->name))
 823                 return false;
 824
 825         if (msg->len_param > RTE_MP_MAX_PARAM_LEN) {
 826                 RTE_LOG(ERR, EAL, "Message data is too long\n");
 827                 rte_errno = E2BIG;
 828                 return false;
 829         }
 830
 831         if (msg->num_fds > RTE_MP_MAX_FD_NUM) {
 832                 RTE_LOG(ERR, EAL, "Cannot send more than %d FDs\n",
 833                         RTE_MP_MAX_FD_NUM);
 834                 rte_errno = E2BIG;
 835                 return false;
 836         }
 837
 838         return true;
 839 }
 840
 841 int __rte_experimental
 842 rte_mp_sendmsg(struct rte_mp_msg *msg)
 843 {
 844         if (!check_input(msg))
 845                 return -1;
 846
 847         RTE_LOG(DEBUG, EAL, "sendmsg: %s\n", msg->name);
 848         return mp_send(msg, NULL, MP_MSG);
 849 }
 850
 851 static int
 852 mp_request_async(const char *dst, struct rte_mp_msg *req,
 853                 struct async_request_param *param)
 854 {
 855         struct rte_mp_msg *reply_msg;
 856         struct pending_request *sync_req, *exist;
 857         int ret;
 858
 859         sync_req = malloc(sizeof(*sync_req));
 860         reply_msg = malloc(sizeof(*reply_msg));
 861         if (sync_req == NULL || reply_msg == NULL) {
 862                 RTE_LOG(ERR, EAL, "Could not allocate space for sync request\n");
 863                 rte_errno = ENOMEM;
 864                 ret = -1;
 865                 goto fail;
 866         }
 867
 868         memset(sync_req, 0, sizeof(*sync_req));
 869         memset(reply_msg, 0, sizeof(*reply_msg));
 870
 871         sync_req->type = REQUEST_TYPE_ASYNC;
 872         strlcpy(sync_req->dst, dst, sizeof(sync_req->dst));
 873         sync_req->request = req;
 874         sync_req->reply = reply_msg;
 875         sync_req->async.param = param;
 876
 877         /* queue already locked by caller */
 878
 879         exist = find_sync_request(dst, req->name);
 880         if (exist) {
 881                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 882                 rte_errno = EEXIST;
 883                 ret = -1;
 884                 goto fail;
 885         }
 886
 887         ret = send_msg(dst, req, MP_REQ);
 888         if (ret < 0) {
 889                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 890                         dst, req->name);
 891                 ret = -1;
 892                 goto fail;
 893         } else if (ret == 0) {
 894                 ret = 0;
 895                 goto fail;
 896         }
 897         TAILQ_INSERT_TAIL(&pending_requests.requests, sync_req, next);
 898
 899         param->user_reply.nb_sent++;
 900
 901         return 0;
 902 fail:
 903         free(sync_req);
 904         free(reply_msg);
 905         return ret;
 906 }
 907
 908 static int
 909 mp_request_sync(const char *dst, struct rte_mp_msg *req,
 910                struct rte_mp_reply *reply, const struct timespec *ts)
 911 {
 912         int ret;
 913         struct rte_mp_msg msg, *tmp;
 914         struct pending_request sync_req, *exist;
 915
 916         sync_req.type = REQUEST_TYPE_SYNC;
 917         sync_req.reply_received = 0;
 918         strlcpy(sync_req.dst, dst, sizeof(sync_req.dst));
 919         sync_req.request = req;
 920         sync_req.reply = &msg;
 921         pthread_cond_init(&sync_req.sync.cond, NULL);
 922
 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                 return -1;
 928         }
 929
 930         ret = send_msg(dst, req, MP_REQ);
 931         if (ret < 0) {
 932                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 933                         dst, req->name);
 934                 return -1;
 935         } else if (ret == 0)
 936                 return 0;
 937
 938         TAILQ_INSERT_TAIL(&pending_requests.requests, &sync_req, next);
 939
 940         reply->nb_sent++;
 941
 942         do {
 943                 ret = pthread_cond_timedwait(&sync_req.sync.cond,
 944                                 &pending_requests.lock, ts);
 945         } while (ret != 0 && ret != ETIMEDOUT);
 946
 947         TAILQ_REMOVE(&pending_requests.requests, &sync_req, next);
 948
 949         if (sync_req.reply_received == 0) {
 950                 RTE_LOG(ERR, EAL, "Fail to recv reply for request %s:%s\n",
 951                         dst, req->name);
 952                 rte_errno = ETIMEDOUT;
 953                 return -1;
 954         }
 955         if (sync_req.reply_received == -1) {
 956                 RTE_LOG(DEBUG, EAL, "Asked to ignore response\n");
 957                 /* not receiving this message is not an error, so decrement
 958                  * number of sent messages
 959                  */
 960                 reply->nb_sent--;
 961                 return 0;
 962         }
 963
 964         tmp = realloc(reply->msgs, sizeof(msg) * (reply->nb_received + 1));
 965         if (!tmp) {
 966                 RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 967                         dst, req->name);
 968                 rte_errno = ENOMEM;
 969                 return -1;
 970         }
 971         memcpy(&tmp[reply->nb_received], &msg, sizeof(msg));
 972         reply->msgs = tmp;
 973         reply->nb_received++;
 974         return 0;
 975 }
 976
 977 int __rte_experimental
 978 rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
 979                 const struct timespec *ts)
 980 {
 981         int dir_fd, ret = 0;
 982         DIR *mp_dir;
 983         struct dirent *ent;
 984         struct timeval now;
 985         struct timespec end;
 986
 987         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
 988
 989         if (check_input(req) == false)
 990                 return -1;
 991         if (gettimeofday(&now, NULL) < 0) {
 992                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
 993                 rte_errno = errno;
 994                 return -1;
 995         }
 996
 997         end.tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
 998         end.tv_sec = now.tv_sec + ts->tv_sec +
 999                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1000
1001         reply->nb_sent = 0;
1002         reply->nb_received = 0;
1003         reply->msgs = NULL;
1004
1005         /* for secondary process, send request to the primary process only */
1006         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1007                 pthread_mutex_lock(&pending_requests.lock);
1008                 ret = mp_request_sync(eal_mp_socket_path(), req, reply, &end);
1009                 pthread_mutex_unlock(&pending_requests.lock);
1010                 return ret;
1011         }
1012
1013         /* for primary process, broadcast request, and collect reply 1 by 1 */
1014         mp_dir = opendir(mp_dir_path);
1015         if (!mp_dir) {
1016                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1017                 rte_errno = errno;
1018                 return -1;
1019         }
1020
1021         dir_fd = dirfd(mp_dir);
1022         /* lock the directory to prevent processes spinning up while we send */
1023         if (flock(dir_fd, LOCK_EX)) {
1024                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1025                         mp_dir_path);
1026                 closedir(mp_dir);
1027                 rte_errno = errno;
1028                 return -1;
1029         }
1030
1031         pthread_mutex_lock(&pending_requests.lock);
1032         while ((ent = readdir(mp_dir))) {
1033                 char path[PATH_MAX];
1034
1035                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1036                         continue;
1037
1038                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1039                          ent->d_name);
1040
1041                 /* unlocks the mutex while waiting for response,
1042                  * locks on receive
1043                  */
1044                 if (mp_request_sync(path, req, reply, &end))
1045                         ret = -1;
1046         }
1047         pthread_mutex_unlock(&pending_requests.lock);
1048         /* unlock the directory */
1049         flock(dir_fd, LOCK_UN);
1050
1051         /* dir_fd automatically closed on closedir */
1052         closedir(mp_dir);
1053         return ret;
1054 }
1055
1056 int __rte_experimental
1057 rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
1058                 rte_mp_async_reply_t clb)
1059 {
1060         struct rte_mp_msg *copy;
1061         struct pending_request *dummy;
1062         struct async_request_param *param;
1063         struct rte_mp_reply *reply;
1064         int dir_fd, ret = 0;
1065         DIR *mp_dir;
1066         struct dirent *ent;
1067         struct timeval now;
1068         struct timespec *end;
1069         bool dummy_used = false;
1070
1071         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
1072
1073         if (check_input(req) == false)
1074                 return -1;
1075         if (gettimeofday(&now, NULL) < 0) {
1076                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
1077                 rte_errno = errno;
1078                 return -1;
1079         }
1080         copy = malloc(sizeof(*copy));
1081         dummy = malloc(sizeof(*dummy));
1082         param = malloc(sizeof(*param));
1083         if (copy == NULL || dummy == NULL || param == NULL) {
1084                 RTE_LOG(ERR, EAL, "Failed to allocate memory for async reply\n");
1085                 rte_errno = ENOMEM;
1086                 goto fail;
1087         }
1088
1089         memset(copy, 0, sizeof(*copy));
1090         memset(dummy, 0, sizeof(*dummy));
1091         memset(param, 0, sizeof(*param));
1092
1093         /* copy message */
1094         memcpy(copy, req, sizeof(*copy));
1095
1096         param->n_responses_processed = 0;
1097         param->clb = clb;
1098         end = &param->end;
1099         reply = &param->user_reply;
1100
1101         end->tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
1102         end->tv_sec = now.tv_sec + ts->tv_sec +
1103                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1104         reply->nb_sent = 0;
1105         reply->nb_received = 0;
1106         reply->msgs = NULL;
1107
1108         /* we have to lock the request queue here, as we will be adding a bunch
1109          * of requests to the queue at once, and some of the replies may arrive
1110          * before we add all of the requests to the queue.
1111          */
1112         pthread_mutex_lock(&pending_requests.lock);
1113
1114         /* we have to ensure that callback gets triggered even if we don't send
1115          * anything, therefore earlier we have allocated a dummy request. fill
1116          * it, and put it on the queue if we don't send any requests.
1117          */
1118         dummy->type = REQUEST_TYPE_ASYNC;
1119         dummy->request = copy;
1120         dummy->reply = NULL;
1121         dummy->async.param = param;
1122         dummy->reply_received = 1; /* short-circuit the timeout */
1123
1124         /* for secondary process, send request to the primary process only */
1125         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1126                 ret = mp_request_async(eal_mp_socket_path(), copy, param);
1127
1128                 /* if we didn't send anything, put dummy request on the queue */
1129                 if (ret == 0 && reply->nb_sent == 0) {
1130                         TAILQ_INSERT_TAIL(&pending_requests.requests, dummy,
1131                                         next);
1132                         dummy_used = true;
1133                 }
1134
1135                 pthread_mutex_unlock(&pending_requests.lock);
1136
1137                 /* if we couldn't send anything, clean up */
1138                 if (ret != 0)
1139                         goto fail;
1140                 return 0;
1141         }
1142
1143         /* for primary process, broadcast request */
1144         mp_dir = opendir(mp_dir_path);
1145         if (!mp_dir) {
1146                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1147                 rte_errno = errno;
1148                 goto unlock_fail;
1149         }
1150         dir_fd = dirfd(mp_dir);
1151
1152         /* lock the directory to prevent processes spinning up while we send */
1153         if (flock(dir_fd, LOCK_EX)) {
1154                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1155                         mp_dir_path);
1156                 rte_errno = errno;
1157                 goto closedir_fail;
1158         }
1159
1160         while ((ent = readdir(mp_dir))) {
1161                 char path[PATH_MAX];
1162
1163                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1164                         continue;
1165
1166                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1167                          ent->d_name);
1168
1169                 if (mp_request_async(path, copy, param))
1170                         ret = -1;
1171         }
1172         /* if we didn't send anything, put dummy request on the queue */
1173         if (ret == 0 && reply->nb_sent == 0) {
1174                 TAILQ_INSERT_HEAD(&pending_requests.requests, dummy, next);
1175                 dummy_used = true;
1176         }
1177
1178         /* trigger async request thread wake up */
1179         pthread_cond_signal(&pending_requests.async_cond);
1180
1181         /* finally, unlock the queue */
1182         pthread_mutex_unlock(&pending_requests.lock);
1183
1184         /* unlock the directory */
1185         flock(dir_fd, LOCK_UN);
1186
1187         /* dir_fd automatically closed on closedir */
1188         closedir(mp_dir);
1189
1190         /* if dummy was unused, free it */
1191         if (!dummy_used)
1192                 free(dummy);
1193
1194         return ret;
1195 closedir_fail:
1196         closedir(mp_dir);
1197 unlock_fail:
1198         pthread_mutex_unlock(&pending_requests.lock);
1199 fail:
1200         free(dummy);
1201         free(param);
1202         free(copy);
1203         return -1;
1204 }
1205
1206 int __rte_experimental
1207 rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
1208 {
1209         RTE_LOG(DEBUG, EAL, "reply: %s\n", msg->name);
1210
1211         if (check_input(msg) == false)
1212                 return -1;
1213
1214         if (peer == NULL) {
1215                 RTE_LOG(ERR, EAL, "peer is not specified\n");
1216                 rte_errno = EINVAL;
1217                 return -1;
1218         }
1219
1220         return mp_send(msg, peer, MP_REP);
1221 }