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