lib/librte_eal/common/eal_common_proc.c

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