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
 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) != 0)
 206                 return -1;
 207
 208         if (internal_config.no_shconf) {
 209                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
 210                 rte_errno = ENOTSUP;
 211                 return -1;
 212         }
 213
 214         entry = malloc(sizeof(struct action_entry));
 215         if (entry == NULL) {
 216                 rte_errno = ENOMEM;
 217                 return -1;
 218         }
 219         strlcpy(entry->action_name, name, sizeof(entry->action_name));
 220         entry->action = action;
 221
 222         pthread_mutex_lock(&mp_mutex_action);
 223         if (find_action_entry_by_name(name) != NULL) {
 224                 pthread_mutex_unlock(&mp_mutex_action);
 225                 rte_errno = EEXIST;
 226                 free(entry);
 227                 return -1;
 228         }
 229         TAILQ_INSERT_TAIL(&action_entry_list, entry, next);
 230         pthread_mutex_unlock(&mp_mutex_action);
 231         return 0;
 232 }
 233
 234 void
 235 rte_mp_action_unregister(const char *name)
 236 {
 237         struct action_entry *entry;
 238
 239         if (validate_action_name(name) != 0)
 240                 return;
 241
 242         if (internal_config.no_shconf) {
 243                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
 244                 return;
 245         }
 246
 247         pthread_mutex_lock(&mp_mutex_action);
 248         entry = find_action_entry_by_name(name);
 249         if (entry == NULL) {
 250                 pthread_mutex_unlock(&mp_mutex_action);
 251                 return;
 252         }
 253         TAILQ_REMOVE(&action_entry_list, entry, next);
 254         pthread_mutex_unlock(&mp_mutex_action);
 255         free(entry);
 256 }
 257
 258 static int
 259 read_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
 260 {
 261         int msglen;
 262         struct iovec iov;
 263         struct msghdr msgh;
 264         char control[CMSG_SPACE(sizeof(m->msg.fds))];
 265         struct cmsghdr *cmsg;
 266         int buflen = sizeof(*m) - sizeof(m->msg.fds);
 267
 268         memset(&msgh, 0, sizeof(msgh));
 269         iov.iov_base = m;
 270         iov.iov_len  = buflen;
 271
 272         msgh.msg_name = s;
 273         msgh.msg_namelen = sizeof(*s);
 274         msgh.msg_iov = &iov;
 275         msgh.msg_iovlen = 1;
 276         msgh.msg_control = control;
 277         msgh.msg_controllen = sizeof(control);
 278
 279         msglen = recvmsg(mp_fd, &msgh, 0);
 280         if (msglen < 0) {
 281                 RTE_LOG(ERR, EAL, "recvmsg failed, %s\n", strerror(errno));
 282                 return -1;
 283         }
 284
 285         if (msglen != buflen || (msgh.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
 286                 RTE_LOG(ERR, EAL, "truncted msg\n");
 287                 return -1;
 288         }
 289
 290         /* read auxiliary FDs if any */
 291         for (cmsg = CMSG_FIRSTHDR(&msgh); cmsg != NULL;
 292                 cmsg = CMSG_NXTHDR(&msgh, cmsg)) {
 293                 if ((cmsg->cmsg_level == SOL_SOCKET) &&
 294                         (cmsg->cmsg_type == SCM_RIGHTS)) {
 295                         memcpy(m->msg.fds, CMSG_DATA(cmsg), sizeof(m->msg.fds));
 296                         break;
 297                 }
 298         }
 299         /* sanity-check the response */
 300         if (m->msg.num_fds < 0 || m->msg.num_fds > RTE_MP_MAX_FD_NUM) {
 301                 RTE_LOG(ERR, EAL, "invalid number of fd's received\n");
 302                 return -1;
 303         }
 304         if (m->msg.len_param < 0 || m->msg.len_param > RTE_MP_MAX_PARAM_LEN) {
 305                 RTE_LOG(ERR, EAL, "invalid received data length\n");
 306                 return -1;
 307         }
 308         return 0;
 309 }
 310
 311 static void
 312 process_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
 313 {
 314         struct pending_request *pending_req;
 315         struct action_entry *entry;
 316         struct rte_mp_msg *msg = &m->msg;
 317         rte_mp_t action = NULL;
 318
 319         RTE_LOG(DEBUG, EAL, "msg: %s\n", msg->name);
 320
 321         if (m->type == MP_REP || m->type == MP_IGN) {
 322                 struct pending_request *req = NULL;
 323
 324                 pthread_mutex_lock(&pending_requests.lock);
 325                 pending_req = find_pending_request(s->sun_path, msg->name);
 326                 if (pending_req) {
 327                         memcpy(pending_req->reply, msg, sizeof(*msg));
 328                         /* -1 indicates that we've been asked to ignore */
 329                         pending_req->reply_received =
 330                                 m->type == MP_REP ? 1 : -1;
 331
 332                         if (pending_req->type == REQUEST_TYPE_SYNC)
 333                                 pthread_cond_signal(&pending_req->sync.cond);
 334                         else if (pending_req->type == REQUEST_TYPE_ASYNC)
 335                                 req = async_reply_handle_thread_unsafe(
 336                                                 pending_req);
 337                 } else
 338                         RTE_LOG(ERR, EAL, "Drop mp reply: %s\n", msg->name);
 339                 pthread_mutex_unlock(&pending_requests.lock);
 340
 341                 if (req != NULL)
 342                         trigger_async_action(req);
 343                 return;
 344         }
 345
 346         pthread_mutex_lock(&mp_mutex_action);
 347         entry = find_action_entry_by_name(msg->name);
 348         if (entry != NULL)
 349                 action = entry->action;
 350         pthread_mutex_unlock(&mp_mutex_action);
 351
 352         if (!action) {
 353                 if (m->type == MP_REQ && !internal_config.init_complete) {
 354                         /* if this is a request, and init is not yet complete,
 355                          * and callback wasn't registered, we should tell the
 356                          * requester to ignore our existence because we're not
 357                          * yet ready to process this request.
 358                          */
 359                         struct rte_mp_msg dummy;
 360
 361                         memset(&dummy, 0, sizeof(dummy));
 362                         strlcpy(dummy.name, msg->name, sizeof(dummy.name));
 363                         mp_send(&dummy, s->sun_path, MP_IGN);
 364                 } else {
 365                         RTE_LOG(ERR, EAL, "Cannot find action: %s\n",
 366                                 msg->name);
 367                 }
 368         } else if (action(msg, s->sun_path) < 0) {
 369                 RTE_LOG(ERR, EAL, "Fail to handle message: %s\n", msg->name);
 370         }
 371 }
 372
 373 static void *
 374 mp_handle(void *arg __rte_unused)
 375 {
 376         struct mp_msg_internal msg;
 377         struct sockaddr_un sa;
 378
 379         while (1) {
 380                 if (read_msg(&msg, &sa) == 0)
 381                         process_msg(&msg, &sa);
 382         }
 383
 384         return NULL;
 385 }
 386
 387 static int
 388 timespec_cmp(const struct timespec *a, const struct timespec *b)
 389 {
 390         if (a->tv_sec < b->tv_sec)
 391                 return -1;
 392         if (a->tv_sec > b->tv_sec)
 393                 return 1;
 394         if (a->tv_nsec < b->tv_nsec)
 395                 return -1;
 396         if (a->tv_nsec > b->tv_nsec)
 397                 return 1;
 398         return 0;
 399 }
 400
 401 enum async_action {
 402         ACTION_FREE, /**< free the action entry, but don't trigger callback */
 403         ACTION_TRIGGER /**< trigger callback, then free action entry */
 404 };
 405
 406 static enum async_action
 407 process_async_request(struct pending_request *sr, const struct timespec *now)
 408 {
 409         struct async_request_param *param;
 410         struct rte_mp_reply *reply;
 411         bool timeout, last_msg;
 412
 413         param = sr->async.param;
 414         reply = &param->user_reply;
 415
 416         /* did we timeout? */
 417         timeout = timespec_cmp(&param->end, now) <= 0;
 418
 419         /* if we received a response, adjust relevant data and copy mesasge. */
 420         if (sr->reply_received == 1 && sr->reply) {
 421                 struct rte_mp_msg *msg, *user_msgs, *tmp;
 422
 423                 msg = sr->reply;
 424                 user_msgs = reply->msgs;
 425
 426                 tmp = realloc(user_msgs, sizeof(*msg) *
 427                                 (reply->nb_received + 1));
 428                 if (!tmp) {
 429                         RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 430                                 sr->dst, sr->request->name);
 431                         /* this entry is going to be removed and its message
 432                          * dropped, but we don't want to leak memory, so
 433                          * continue.
 434                          */
 435                 } else {
 436                         user_msgs = tmp;
 437                         reply->msgs = user_msgs;
 438                         memcpy(&user_msgs[reply->nb_received],
 439                                         msg, sizeof(*msg));
 440                         reply->nb_received++;
 441                 }
 442
 443                 /* mark this request as processed */
 444                 param->n_responses_processed++;
 445         } else if (sr->reply_received == -1) {
 446                 /* we were asked to ignore this process */
 447                 reply->nb_sent--;
 448         } else if (timeout) {
 449                 /* count it as processed response, but don't increment
 450                  * nb_received.
 451                  */
 452                 param->n_responses_processed++;
 453         }
 454
 455         free(sr->reply);
 456
 457         last_msg = param->n_responses_processed == reply->nb_sent;
 458
 459         return last_msg ? ACTION_TRIGGER : ACTION_FREE;
 460 }
 461
 462 static void
 463 trigger_async_action(struct pending_request *sr)
 464 {
 465         struct async_request_param *param;
 466         struct rte_mp_reply *reply;
 467
 468         param = sr->async.param;
 469         reply = &param->user_reply;
 470
 471         param->clb(sr->request, reply);
 472
 473         /* clean up */
 474         free(sr->async.param->user_reply.msgs);
 475         free(sr->async.param);
 476         free(sr->request);
 477         free(sr);
 478 }
 479
 480 static struct pending_request *
 481 async_reply_handle_thread_unsafe(void *arg)
 482 {
 483         struct pending_request *req = (struct pending_request *)arg;
 484         enum async_action action;
 485         struct timespec ts_now;
 486         struct timeval now;
 487
 488         if (gettimeofday(&now, NULL) < 0) {
 489                 RTE_LOG(ERR, EAL, "Cannot get current time\n");
 490                 goto no_trigger;
 491         }
 492         ts_now.tv_nsec = now.tv_usec * 1000;
 493         ts_now.tv_sec = now.tv_sec;
 494
 495         action = process_async_request(req, &ts_now);
 496
 497         TAILQ_REMOVE(&pending_requests.requests, req, next);
 498
 499         if (rte_eal_alarm_cancel(async_reply_handle, req) < 0) {
 500                 /* if we failed to cancel the alarm because it's already in
 501                  * progress, don't proceed because otherwise we will end up
 502                  * handling the same message twice.
 503                  */
 504                 if (rte_errno == EINPROGRESS) {
 505                         RTE_LOG(DEBUG, EAL, "Request handling is already in progress\n");
 506                         goto no_trigger;
 507                 }
 508                 RTE_LOG(ERR, EAL, "Failed to cancel alarm\n");
 509         }
 510
 511         if (action == ACTION_TRIGGER)
 512                 return req;
 513 no_trigger:
 514         free(req);
 515         return NULL;
 516 }
 517
 518 static void
 519 async_reply_handle(void *arg)
 520 {
 521         struct pending_request *req;
 522
 523         pthread_mutex_lock(&pending_requests.lock);
 524         req = async_reply_handle_thread_unsafe(arg);
 525         pthread_mutex_unlock(&pending_requests.lock);
 526
 527         if (req != NULL)
 528                 trigger_async_action(req);
 529 }
 530
 531 static int
 532 open_socket_fd(void)
 533 {
 534         struct sockaddr_un un;
 535
 536         peer_name[0] = '\0';
 537         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 538                 snprintf(peer_name, sizeof(peer_name),
 539                                 "%d_%"PRIx64, getpid(), rte_rdtsc());
 540
 541         mp_fd = socket(AF_UNIX, SOCK_DGRAM, 0);
 542         if (mp_fd < 0) {
 543                 RTE_LOG(ERR, EAL, "failed to create unix socket\n");
 544                 return -1;
 545         }
 546
 547         memset(&un, 0, sizeof(un));
 548         un.sun_family = AF_UNIX;
 549
 550         create_socket_path(peer_name, un.sun_path, sizeof(un.sun_path));
 551
 552         unlink(un.sun_path); /* May still exist since last run */
 553
 554         if (bind(mp_fd, (struct sockaddr *)&un, sizeof(un)) < 0) {
 555                 RTE_LOG(ERR, EAL, "failed to bind %s: %s\n",
 556                         un.sun_path, strerror(errno));
 557                 close(mp_fd);
 558                 return -1;
 559         }
 560
 561         RTE_LOG(INFO, EAL, "Multi-process socket %s\n", un.sun_path);
 562         return mp_fd;
 563 }
 564
 565 static void
 566 close_socket_fd(void)
 567 {
 568         char path[PATH_MAX];
 569
 570         if (mp_fd < 0)
 571                 return;
 572
 573         close(mp_fd);
 574         create_socket_path(peer_name, path, sizeof(path));
 575         unlink(path);
 576 }
 577
 578 int
 579 rte_mp_channel_init(void)
 580 {
 581         char path[PATH_MAX];
 582         int dir_fd;
 583         pthread_t mp_handle_tid;
 584
 585         /* in no shared files mode, we do not have secondary processes support,
 586          * so no need to initialize IPC.
 587          */
 588         if (internal_config.no_shconf) {
 589                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC will be disabled\n");
 590                 rte_errno = ENOTSUP;
 591                 return -1;
 592         }
 593
 594         /* create filter path */
 595         create_socket_path("*", path, sizeof(path));
 596         strlcpy(mp_filter, basename(path), sizeof(mp_filter));
 597
 598         /* path may have been modified, so recreate it */
 599         create_socket_path("*", path, sizeof(path));
 600         strlcpy(mp_dir_path, dirname(path), sizeof(mp_dir_path));
 601
 602         /* lock the directory */
 603         dir_fd = open(mp_dir_path, O_RDONLY);
 604         if (dir_fd < 0) {
 605                 RTE_LOG(ERR, EAL, "failed to open %s: %s\n",
 606                         mp_dir_path, strerror(errno));
 607                 return -1;
 608         }
 609
 610         if (flock(dir_fd, LOCK_EX)) {
 611                 RTE_LOG(ERR, EAL, "failed to lock %s: %s\n",
 612                         mp_dir_path, strerror(errno));
 613                 close(dir_fd);
 614                 return -1;
 615         }
 616
 617         if (open_socket_fd() < 0) {
 618                 close(dir_fd);
 619                 return -1;
 620         }
 621
 622         if (rte_ctrl_thread_create(&mp_handle_tid, "rte_mp_handle",
 623                         NULL, mp_handle, NULL) < 0) {
 624                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 625                         strerror(errno));
 626                 close(mp_fd);
 627                 close(dir_fd);
 628                 mp_fd = -1;
 629                 return -1;
 630         }
 631
 632         /* unlock the directory */
 633         flock(dir_fd, LOCK_UN);
 634         close(dir_fd);
 635
 636         return 0;
 637 }
 638
 639 void
 640 rte_mp_channel_cleanup(void)
 641 {
 642         close_socket_fd();
 643 }
 644
 645 /**
 646  * Return -1, as fail to send message and it's caused by the local side.
 647  * Return 0, as fail to send message and it's caused by the remote side.
 648  * Return 1, as succeed to send message.
 649  *
 650  */
 651 static int
 652 send_msg(const char *dst_path, struct rte_mp_msg *msg, int type)
 653 {
 654         int snd;
 655         struct iovec iov;
 656         struct msghdr msgh;
 657         struct cmsghdr *cmsg;
 658         struct sockaddr_un dst;
 659         struct mp_msg_internal m;
 660         int fd_size = msg->num_fds * sizeof(int);
 661         char control[CMSG_SPACE(fd_size)];
 662
 663         m.type = type;
 664         memcpy(&m.msg, msg, sizeof(*msg));
 665
 666         memset(&dst, 0, sizeof(dst));
 667         dst.sun_family = AF_UNIX;
 668         strlcpy(dst.sun_path, dst_path, sizeof(dst.sun_path));
 669
 670         memset(&msgh, 0, sizeof(msgh));
 671         memset(control, 0, sizeof(control));
 672
 673         iov.iov_base = &m;
 674         iov.iov_len = sizeof(m) - sizeof(msg->fds);
 675
 676         msgh.msg_name = &dst;
 677         msgh.msg_namelen = sizeof(dst);
 678         msgh.msg_iov = &iov;
 679         msgh.msg_iovlen = 1;
 680         msgh.msg_control = control;
 681         msgh.msg_controllen = sizeof(control);
 682
 683         cmsg = CMSG_FIRSTHDR(&msgh);
 684         cmsg->cmsg_len = CMSG_LEN(fd_size);
 685         cmsg->cmsg_level = SOL_SOCKET;
 686         cmsg->cmsg_type = SCM_RIGHTS;
 687         memcpy(CMSG_DATA(cmsg), msg->fds, fd_size);
 688
 689         do {
 690                 snd = sendmsg(mp_fd, &msgh, 0);
 691         } while (snd < 0 && errno == EINTR);
 692
 693         if (snd < 0) {
 694                 rte_errno = errno;
 695                 /* Check if it caused by peer process exits */
 696                 if (errno == ECONNREFUSED &&
 697                                 rte_eal_process_type() == RTE_PROC_PRIMARY) {
 698                         unlink(dst_path);
 699                         return 0;
 700                 }
 701                 RTE_LOG(ERR, EAL, "failed to send to (%s) due to %s\n",
 702                         dst_path, strerror(errno));
 703                 return -1;
 704         }
 705
 706         return 1;
 707 }
 708
 709 static int
 710 mp_send(struct rte_mp_msg *msg, const char *peer, int type)
 711 {
 712         int dir_fd, ret = 0;
 713         DIR *mp_dir;
 714         struct dirent *ent;
 715
 716         if (!peer && (rte_eal_process_type() == RTE_PROC_SECONDARY))
 717                 peer = eal_mp_socket_path();
 718
 719         if (peer) {
 720                 if (send_msg(peer, msg, type) < 0)
 721                         return -1;
 722                 else
 723                         return 0;
 724         }
 725
 726         /* broadcast to all secondary processes */
 727         mp_dir = opendir(mp_dir_path);
 728         if (!mp_dir) {
 729                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n",
 730                                 mp_dir_path);
 731                 rte_errno = errno;
 732                 return -1;
 733         }
 734
 735         dir_fd = dirfd(mp_dir);
 736         /* lock the directory to prevent processes spinning up while we send */
 737         if (flock(dir_fd, LOCK_SH)) {
 738                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
 739                         mp_dir_path);
 740                 rte_errno = errno;
 741                 closedir(mp_dir);
 742                 return -1;
 743         }
 744
 745         while ((ent = readdir(mp_dir))) {
 746                 char path[PATH_MAX];
 747
 748                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
 749                         continue;
 750
 751                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
 752                          ent->d_name);
 753                 if (send_msg(path, msg, type) < 0)
 754                         ret = -1;
 755         }
 756         /* unlock the dir */
 757         flock(dir_fd, LOCK_UN);
 758
 759         /* dir_fd automatically closed on closedir */
 760         closedir(mp_dir);
 761         return ret;
 762 }
 763
 764 static int
 765 check_input(const struct rte_mp_msg *msg)
 766 {
 767         if (msg == NULL) {
 768                 RTE_LOG(ERR, EAL, "Msg cannot be NULL\n");
 769                 rte_errno = EINVAL;
 770                 return -1;
 771         }
 772
 773         if (validate_action_name(msg->name) != 0)
 774                 return -1;
 775
 776         if (msg->len_param < 0) {
 777                 RTE_LOG(ERR, EAL, "Message data length is negative\n");
 778                 rte_errno = EINVAL;
 779                 return -1;
 780         }
 781
 782         if (msg->num_fds < 0) {
 783                 RTE_LOG(ERR, EAL, "Number of fd's is negative\n");
 784                 rte_errno = EINVAL;
 785                 return -1;
 786         }
 787
 788         if (msg->len_param > RTE_MP_MAX_PARAM_LEN) {
 789                 RTE_LOG(ERR, EAL, "Message data is too long\n");
 790                 rte_errno = E2BIG;
 791                 return -1;
 792         }
 793
 794         if (msg->num_fds > RTE_MP_MAX_FD_NUM) {
 795                 RTE_LOG(ERR, EAL, "Cannot send more than %d FDs\n",
 796                         RTE_MP_MAX_FD_NUM);
 797                 rte_errno = E2BIG;
 798                 return -1;
 799         }
 800
 801         return 0;
 802 }
 803
 804 int
 805 rte_mp_sendmsg(struct rte_mp_msg *msg)
 806 {
 807         if (check_input(msg) != 0)
 808                 return -1;
 809
 810         if (internal_config.no_shconf) {
 811                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
 812                 rte_errno = ENOTSUP;
 813                 return -1;
 814         }
 815
 816         RTE_LOG(DEBUG, EAL, "sendmsg: %s\n", msg->name);
 817         return mp_send(msg, NULL, MP_MSG);
 818 }
 819
 820 static int
 821 mp_request_async(const char *dst, struct rte_mp_msg *req,
 822                 struct async_request_param *param, const struct timespec *ts)
 823 {
 824         struct rte_mp_msg *reply_msg;
 825         struct pending_request *pending_req, *exist;
 826         int ret = -1;
 827
 828         pending_req = calloc(1, sizeof(*pending_req));
 829         reply_msg = calloc(1, sizeof(*reply_msg));
 830         if (pending_req == NULL || reply_msg == NULL) {
 831                 RTE_LOG(ERR, EAL, "Could not allocate space for sync request\n");
 832                 rte_errno = ENOMEM;
 833                 ret = -1;
 834                 goto fail;
 835         }
 836
 837         pending_req->type = REQUEST_TYPE_ASYNC;
 838         strlcpy(pending_req->dst, dst, sizeof(pending_req->dst));
 839         pending_req->request = req;
 840         pending_req->reply = reply_msg;
 841         pending_req->async.param = param;
 842
 843         /* queue already locked by caller */
 844
 845         exist = find_pending_request(dst, req->name);
 846         if (exist) {
 847                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 848                 rte_errno = EEXIST;
 849                 ret = -1;
 850                 goto fail;
 851         }
 852
 853         ret = send_msg(dst, req, MP_REQ);
 854         if (ret < 0) {
 855                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 856                         dst, req->name);
 857                 ret = -1;
 858                 goto fail;
 859         } else if (ret == 0) {
 860                 ret = 0;
 861                 goto fail;
 862         }
 863         param->user_reply.nb_sent++;
 864
 865         /* if alarm set fails, we simply ignore the reply */
 866         if (rte_eal_alarm_set(ts->tv_sec * 1000000 + ts->tv_nsec / 1000,
 867                               async_reply_handle, pending_req) < 0) {
 868                 RTE_LOG(ERR, EAL, "Fail to set alarm for request %s:%s\n",
 869                         dst, req->name);
 870                 ret = -1;
 871                 goto fail;
 872         }
 873         TAILQ_INSERT_TAIL(&pending_requests.requests, pending_req, next);
 874
 875         return 0;
 876 fail:
 877         free(pending_req);
 878         free(reply_msg);
 879         return ret;
 880 }
 881
 882 static int
 883 mp_request_sync(const char *dst, struct rte_mp_msg *req,
 884                struct rte_mp_reply *reply, const struct timespec *ts)
 885 {
 886         int ret;
 887         struct rte_mp_msg msg, *tmp;
 888         struct pending_request pending_req, *exist;
 889
 890         pending_req.type = REQUEST_TYPE_SYNC;
 891         pending_req.reply_received = 0;
 892         strlcpy(pending_req.dst, dst, sizeof(pending_req.dst));
 893         pending_req.request = req;
 894         pending_req.reply = &msg;
 895         pthread_cond_init(&pending_req.sync.cond, NULL);
 896
 897         exist = find_pending_request(dst, req->name);
 898         if (exist) {
 899                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 900                 rte_errno = EEXIST;
 901                 return -1;
 902         }
 903
 904         ret = send_msg(dst, req, MP_REQ);
 905         if (ret < 0) {
 906                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 907                         dst, req->name);
 908                 return -1;
 909         } else if (ret == 0)
 910                 return 0;
 911
 912         TAILQ_INSERT_TAIL(&pending_requests.requests, &pending_req, next);
 913
 914         reply->nb_sent++;
 915
 916         do {
 917                 ret = pthread_cond_timedwait(&pending_req.sync.cond,
 918                                 &pending_requests.lock, ts);
 919         } while (ret != 0 && ret != ETIMEDOUT);
 920
 921         TAILQ_REMOVE(&pending_requests.requests, &pending_req, next);
 922
 923         if (pending_req.reply_received == 0) {
 924                 RTE_LOG(ERR, EAL, "Fail to recv reply for request %s:%s\n",
 925                         dst, req->name);
 926                 rte_errno = ETIMEDOUT;
 927                 return -1;
 928         }
 929         if (pending_req.reply_received == -1) {
 930                 RTE_LOG(DEBUG, EAL, "Asked to ignore response\n");
 931                 /* not receiving this message is not an error, so decrement
 932                  * number of sent messages
 933                  */
 934                 reply->nb_sent--;
 935                 return 0;
 936         }
 937
 938         tmp = realloc(reply->msgs, sizeof(msg) * (reply->nb_received + 1));
 939         if (!tmp) {
 940                 RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 941                         dst, req->name);
 942                 rte_errno = ENOMEM;
 943                 return -1;
 944         }
 945         memcpy(&tmp[reply->nb_received], &msg, sizeof(msg));
 946         reply->msgs = tmp;
 947         reply->nb_received++;
 948         return 0;
 949 }
 950
 951 int
 952 rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
 953                 const struct timespec *ts)
 954 {
 955         int dir_fd, ret = -1;
 956         DIR *mp_dir;
 957         struct dirent *ent;
 958         struct timeval now;
 959         struct timespec end;
 960
 961         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
 962
 963         reply->nb_sent = 0;
 964         reply->nb_received = 0;
 965         reply->msgs = NULL;
 966
 967         if (check_input(req) != 0)
 968                 goto end;
 969
 970         if (internal_config.no_shconf) {
 971                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
 972                 rte_errno = ENOTSUP;
 973                 return -1;
 974         }
 975
 976         if (gettimeofday(&now, NULL) < 0) {
 977                 RTE_LOG(ERR, EAL, "Failed to get current time\n");
 978                 rte_errno = errno;
 979                 goto end;
 980         }
 981
 982         end.tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
 983         end.tv_sec = now.tv_sec + ts->tv_sec +
 984                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
 985
 986         /* for secondary process, send request to the primary process only */
 987         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 988                 pthread_mutex_lock(&pending_requests.lock);
 989                 ret = mp_request_sync(eal_mp_socket_path(), req, reply, &end);
 990                 pthread_mutex_unlock(&pending_requests.lock);
 991                 goto end;
 992         }
 993
 994         /* for primary process, broadcast request, and collect reply 1 by 1 */
 995         mp_dir = opendir(mp_dir_path);
 996         if (!mp_dir) {
 997                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
 998                 rte_errno = errno;
 999                 goto end;
1000         }
1001
1002         dir_fd = dirfd(mp_dir);
1003         /* lock the directory to prevent processes spinning up while we send */
1004         if (flock(dir_fd, LOCK_SH)) {
1005                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1006                         mp_dir_path);
1007                 rte_errno = errno;
1008                 goto close_end;
1009         }
1010
1011         pthread_mutex_lock(&pending_requests.lock);
1012         while ((ent = readdir(mp_dir))) {
1013                 char path[PATH_MAX];
1014
1015                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1016                         continue;
1017
1018                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1019                          ent->d_name);
1020
1021                 /* unlocks the mutex while waiting for response,
1022                  * locks on receive
1023                  */
1024                 if (mp_request_sync(path, req, reply, &end))
1025                         goto unlock_end;
1026         }
1027         ret = 0;
1028
1029 unlock_end:
1030         pthread_mutex_unlock(&pending_requests.lock);
1031         /* unlock the directory */
1032         flock(dir_fd, LOCK_UN);
1033
1034 close_end:
1035         /* dir_fd automatically closed on closedir */
1036         closedir(mp_dir);
1037
1038 end:
1039         if (ret) {
1040                 free(reply->msgs);
1041                 reply->nb_received = 0;
1042                 reply->msgs = NULL;
1043         }
1044         return ret;
1045 }
1046
1047 int
1048 rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
1049                 rte_mp_async_reply_t clb)
1050 {
1051         struct rte_mp_msg *copy;
1052         struct pending_request *dummy;
1053         struct async_request_param *param;
1054         struct rte_mp_reply *reply;
1055         int dir_fd, ret = 0;
1056         DIR *mp_dir;
1057         struct dirent *ent;
1058         struct timeval now;
1059         struct timespec *end;
1060         bool dummy_used = false;
1061
1062         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
1063
1064         if (check_input(req) != 0)
1065                 return -1;
1066
1067         if (internal_config.no_shconf) {
1068                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
1069                 rte_errno = ENOTSUP;
1070                 return -1;
1071         }
1072
1073         if (gettimeofday(&now, NULL) < 0) {
1074                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
1075                 rte_errno = errno;
1076                 return -1;
1077         }
1078         copy = calloc(1, sizeof(*copy));
1079         dummy = calloc(1, sizeof(*dummy));
1080         param = calloc(1, sizeof(*param));
1081         if (copy == NULL || dummy == NULL || param == NULL) {
1082                 RTE_LOG(ERR, EAL, "Failed to allocate memory for async reply\n");
1083                 rte_errno = ENOMEM;
1084                 goto fail;
1085         }
1086
1087         /* copy message */
1088         memcpy(copy, req, sizeof(*copy));
1089
1090         param->n_responses_processed = 0;
1091         param->clb = clb;
1092         end = &param->end;
1093         reply = &param->user_reply;
1094
1095         end->tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
1096         end->tv_sec = now.tv_sec + ts->tv_sec +
1097                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1098         reply->nb_sent = 0;
1099         reply->nb_received = 0;
1100         reply->msgs = NULL;
1101
1102         /* we have to lock the request queue here, as we will be adding a bunch
1103          * of requests to the queue at once, and some of the replies may arrive
1104          * before we add all of the requests to the queue.
1105          */
1106         pthread_mutex_lock(&pending_requests.lock);
1107
1108         /* we have to ensure that callback gets triggered even if we don't send
1109          * anything, therefore earlier we have allocated a dummy request. fill
1110          * it, and put it on the queue if we don't send any requests.
1111          */
1112         dummy->type = REQUEST_TYPE_ASYNC;
1113         dummy->request = copy;
1114         dummy->reply = NULL;
1115         dummy->async.param = param;
1116         dummy->reply_received = 1; /* short-circuit the timeout */
1117
1118         /* for secondary process, send request to the primary process only */
1119         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1120                 ret = mp_request_async(eal_mp_socket_path(), copy, param, ts);
1121
1122                 /* if we didn't send anything, put dummy request on the queue */
1123                 if (ret == 0 && reply->nb_sent == 0) {
1124                         TAILQ_INSERT_TAIL(&pending_requests.requests, dummy,
1125                                         next);
1126                         dummy_used = true;
1127                 }
1128
1129                 pthread_mutex_unlock(&pending_requests.lock);
1130
1131                 /* if we couldn't send anything, clean up */
1132                 if (ret != 0)
1133                         goto fail;
1134                 return 0;
1135         }
1136
1137         /* for primary process, broadcast request */
1138         mp_dir = opendir(mp_dir_path);
1139         if (!mp_dir) {
1140                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1141                 rte_errno = errno;
1142                 goto unlock_fail;
1143         }
1144         dir_fd = dirfd(mp_dir);
1145
1146         /* lock the directory to prevent processes spinning up while we send */
1147         if (flock(dir_fd, LOCK_SH)) {
1148                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1149                         mp_dir_path);
1150                 rte_errno = errno;
1151                 goto closedir_fail;
1152         }
1153
1154         while ((ent = readdir(mp_dir))) {
1155                 char path[PATH_MAX];
1156
1157                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1158                         continue;
1159
1160                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1161                          ent->d_name);
1162
1163                 if (mp_request_async(path, copy, param, ts))
1164                         ret = -1;
1165         }
1166         /* if we didn't send anything, put dummy request on the queue */
1167         if (ret == 0 && reply->nb_sent == 0) {
1168                 TAILQ_INSERT_HEAD(&pending_requests.requests, dummy, next);
1169                 dummy_used = true;
1170         }
1171
1172         /* finally, unlock the queue */
1173         pthread_mutex_unlock(&pending_requests.lock);
1174
1175         /* unlock the directory */
1176         flock(dir_fd, LOCK_UN);
1177
1178         /* dir_fd automatically closed on closedir */
1179         closedir(mp_dir);
1180
1181         /* if dummy was unused, free it */
1182         if (!dummy_used)
1183                 free(dummy);
1184
1185         return ret;
1186 closedir_fail:
1187         closedir(mp_dir);
1188 unlock_fail:
1189         pthread_mutex_unlock(&pending_requests.lock);
1190 fail:
1191         free(dummy);
1192         free(param);
1193         free(copy);
1194         return -1;
1195 }
1196
1197 int
1198 rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
1199 {
1200         RTE_LOG(DEBUG, EAL, "reply: %s\n", msg->name);
1201
1202         if (check_input(msg) != 0)
1203                 return -1;
1204
1205         if (peer == NULL) {
1206                 RTE_LOG(ERR, EAL, "peer is not specified\n");
1207                 rte_errno = EINVAL;
1208                 return -1;
1209         }
1210
1211         if (internal_config.no_shconf) {
1212                 RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
1213                 return 0;
1214         }
1215
1216         return mp_send(msg, peer, MP_REP);
1217 }