xref: /dpdk/lib/eal/common/eal_common_proc.c (revision bbbe38a6d59ccdda25917712701e629d0b10af6f)
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_memcfg.h"
33 #include "eal_private.h"
34 #include "eal_filesystem.h"
35 #include "eal_internal_cfg.h"
36 
37 static int mp_fd = -1;
38 static char mp_filter[PATH_MAX];   /* Filter for secondary process sockets */
39 static char mp_dir_path[PATH_MAX]; /* The directory path for all mp sockets */
40 static pthread_mutex_t mp_mutex_action = PTHREAD_MUTEX_INITIALIZER;
41 static char peer_name[PATH_MAX];
42 
43 struct action_entry {
44 	TAILQ_ENTRY(action_entry) next;
45 	char action_name[RTE_MP_MAX_NAME_LEN];
46 	rte_mp_t action;
47 };
48 
49 /** Double linked list of actions. */
50 TAILQ_HEAD(action_entry_list, action_entry);
51 
52 static struct action_entry_list action_entry_list =
53 	TAILQ_HEAD_INITIALIZER(action_entry_list);
54 
55 enum mp_type {
56 	MP_MSG, /* Share message with peers, will not block */
57 	MP_REQ, /* Request for information, Will block for a reply */
58 	MP_REP, /* Response to previously-received request */
59 	MP_IGN, /* Response telling requester to ignore this response */
60 };
61 
62 struct mp_msg_internal {
63 	int type;
64 	struct rte_mp_msg msg;
65 };
66 
67 struct async_request_param {
68 	rte_mp_async_reply_t clb;
69 	struct rte_mp_reply user_reply;
70 	struct timespec end;
71 	int n_responses_processed;
72 };
73 
74 struct pending_request {
75 	TAILQ_ENTRY(pending_request) next;
76 	enum {
77 		REQUEST_TYPE_SYNC,
78 		REQUEST_TYPE_ASYNC
79 	} type;
80 	char dst[PATH_MAX];
81 	struct rte_mp_msg *request;
82 	struct rte_mp_msg *reply;
83 	int reply_received;
84 	RTE_STD_C11
85 	union {
86 		struct {
87 			struct async_request_param *param;
88 		} async;
89 		struct {
90 			pthread_cond_t cond;
91 		} sync;
92 	};
93 };
94 
95 TAILQ_HEAD(pending_request_list, pending_request);
96 
97 static struct {
98 	struct pending_request_list requests;
99 	pthread_mutex_t lock;
100 } pending_requests = {
101 	.requests = TAILQ_HEAD_INITIALIZER(pending_requests.requests),
102 	.lock = PTHREAD_MUTEX_INITIALIZER,
103 	/**< used in async requests only */
104 };
105 
106 /* forward declarations */
107 static int
108 mp_send(struct rte_mp_msg *msg, const char *peer, int type);
109 
110 /* for use with alarm callback */
111 static void
112 async_reply_handle(void *arg);
113 
114 /* for use with process_msg */
115 static struct pending_request *
116 async_reply_handle_thread_unsafe(void *arg);
117 
118 static void
119 trigger_async_action(struct pending_request *req);
120 
121 static struct pending_request *
122 find_pending_request(const char *dst, const char *act_name)
123 {
124 	struct pending_request *r;
125 
126 	TAILQ_FOREACH(r, &pending_requests.requests, next) {
127 		if (!strcmp(r->dst, dst) &&
128 		    !strcmp(r->request->name, act_name))
129 			break;
130 	}
131 
132 	return r;
133 }
134 
135 static void
136 create_socket_path(const char *name, char *buf, int len)
137 {
138 	const char *prefix = eal_mp_socket_path();
139 
140 	if (strlen(name) > 0)
141 		snprintf(buf, len, "%s_%s", prefix, name);
142 	else
143 		strlcpy(buf, prefix, len);
144 }
145 
146 int
147 rte_eal_primary_proc_alive(const char *config_file_path)
148 {
149 	int config_fd;
150 
151 	if (config_file_path)
152 		config_fd = open(config_file_path, O_RDONLY);
153 	else {
154 		const char *path;
155 
156 		path = eal_runtime_config_path();
157 		config_fd = open(path, O_RDONLY);
158 	}
159 	if (config_fd < 0)
160 		return 0;
161 
162 	int ret = lockf(config_fd, F_TEST, 0);
163 	close(config_fd);
164 
165 	return !!ret;
166 }
167 
168 static struct action_entry *
169 find_action_entry_by_name(const char *name)
170 {
171 	struct action_entry *entry;
172 
173 	TAILQ_FOREACH(entry, &action_entry_list, next) {
174 		if (strncmp(entry->action_name, name, RTE_MP_MAX_NAME_LEN) == 0)
175 			break;
176 	}
177 
178 	return entry;
179 }
180 
181 static int
182 validate_action_name(const char *name)
183 {
184 	if (name == NULL) {
185 		RTE_LOG(ERR, EAL, "Action name cannot be NULL\n");
186 		rte_errno = EINVAL;
187 		return -1;
188 	}
189 	if (strnlen(name, RTE_MP_MAX_NAME_LEN) == 0) {
190 		RTE_LOG(ERR, EAL, "Length of action name is zero\n");
191 		rte_errno = EINVAL;
192 		return -1;
193 	}
194 	if (strnlen(name, RTE_MP_MAX_NAME_LEN) == RTE_MP_MAX_NAME_LEN) {
195 		rte_errno = E2BIG;
196 		return -1;
197 	}
198 	return 0;
199 }
200 
201 int
202 rte_mp_action_register(const char *name, rte_mp_t action)
203 {
204 	struct action_entry *entry;
205 	const struct internal_config *internal_conf =
206 		eal_get_internal_configuration();
207 
208 	if (validate_action_name(name) != 0)
209 		return -1;
210 
211 	if (internal_conf->no_shconf) {
212 		RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
213 		rte_errno = ENOTSUP;
214 		return -1;
215 	}
216 
217 	entry = malloc(sizeof(struct action_entry));
218 	if (entry == NULL) {
219 		rte_errno = ENOMEM;
220 		return -1;
221 	}
222 	strlcpy(entry->action_name, name, sizeof(entry->action_name));
223 	entry->action = action;
224 
225 	pthread_mutex_lock(&mp_mutex_action);
226 	if (find_action_entry_by_name(name) != NULL) {
227 		pthread_mutex_unlock(&mp_mutex_action);
228 		rte_errno = EEXIST;
229 		free(entry);
230 		return -1;
231 	}
232 	TAILQ_INSERT_TAIL(&action_entry_list, entry, next);
233 	pthread_mutex_unlock(&mp_mutex_action);
234 	return 0;
235 }
236 
237 void
238 rte_mp_action_unregister(const char *name)
239 {
240 	struct action_entry *entry;
241 	const struct internal_config *internal_conf =
242 		eal_get_internal_configuration();
243 
244 	if (validate_action_name(name) != 0)
245 		return;
246 
247 	if (internal_conf->no_shconf) {
248 		RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
249 		return;
250 	}
251 
252 	pthread_mutex_lock(&mp_mutex_action);
253 	entry = find_action_entry_by_name(name);
254 	if (entry == NULL) {
255 		pthread_mutex_unlock(&mp_mutex_action);
256 		return;
257 	}
258 	TAILQ_REMOVE(&action_entry_list, entry, next);
259 	pthread_mutex_unlock(&mp_mutex_action);
260 	free(entry);
261 }
262 
263 static int
264 read_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
265 {
266 	int msglen;
267 	struct iovec iov;
268 	struct msghdr msgh;
269 	char control[CMSG_SPACE(sizeof(m->msg.fds))];
270 	struct cmsghdr *cmsg;
271 	int buflen = sizeof(*m) - sizeof(m->msg.fds);
272 
273 	memset(&msgh, 0, sizeof(msgh));
274 	iov.iov_base = m;
275 	iov.iov_len  = buflen;
276 
277 	msgh.msg_name = s;
278 	msgh.msg_namelen = sizeof(*s);
279 	msgh.msg_iov = &iov;
280 	msgh.msg_iovlen = 1;
281 	msgh.msg_control = control;
282 	msgh.msg_controllen = sizeof(control);
283 
284 	msglen = recvmsg(mp_fd, &msgh, 0);
285 	if (msglen < 0) {
286 		RTE_LOG(ERR, EAL, "recvmsg failed, %s\n", strerror(errno));
287 		return -1;
288 	}
289 
290 	if (msglen != buflen || (msgh.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
291 		RTE_LOG(ERR, EAL, "truncated msg\n");
292 		return -1;
293 	}
294 
295 	/* read auxiliary FDs if any */
296 	for (cmsg = CMSG_FIRSTHDR(&msgh); cmsg != NULL;
297 		cmsg = CMSG_NXTHDR(&msgh, cmsg)) {
298 		if ((cmsg->cmsg_level == SOL_SOCKET) &&
299 			(cmsg->cmsg_type == SCM_RIGHTS)) {
300 			memcpy(m->msg.fds, CMSG_DATA(cmsg), sizeof(m->msg.fds));
301 			break;
302 		}
303 	}
304 	/* sanity-check the response */
305 	if (m->msg.num_fds < 0 || m->msg.num_fds > RTE_MP_MAX_FD_NUM) {
306 		RTE_LOG(ERR, EAL, "invalid number of fd's received\n");
307 		return -1;
308 	}
309 	if (m->msg.len_param < 0 || m->msg.len_param > RTE_MP_MAX_PARAM_LEN) {
310 		RTE_LOG(ERR, EAL, "invalid received data length\n");
311 		return -1;
312 	}
313 	return 0;
314 }
315 
316 static void
317 process_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
318 {
319 	struct pending_request *pending_req;
320 	struct action_entry *entry;
321 	struct rte_mp_msg *msg = &m->msg;
322 	rte_mp_t action = NULL;
323 	const struct internal_config *internal_conf =
324 		eal_get_internal_configuration();
325 
326 	RTE_LOG(DEBUG, EAL, "msg: %s\n", msg->name);
327 
328 	if (m->type == MP_REP || m->type == MP_IGN) {
329 		struct pending_request *req = NULL;
330 
331 		pthread_mutex_lock(&pending_requests.lock);
332 		pending_req = find_pending_request(s->sun_path, msg->name);
333 		if (pending_req) {
334 			memcpy(pending_req->reply, msg, sizeof(*msg));
335 			/* -1 indicates that we've been asked to ignore */
336 			pending_req->reply_received =
337 				m->type == MP_REP ? 1 : -1;
338 
339 			if (pending_req->type == REQUEST_TYPE_SYNC)
340 				pthread_cond_signal(&pending_req->sync.cond);
341 			else if (pending_req->type == REQUEST_TYPE_ASYNC)
342 				req = async_reply_handle_thread_unsafe(
343 						pending_req);
344 		} else
345 			RTE_LOG(ERR, EAL, "Drop mp reply: %s\n", msg->name);
346 		pthread_mutex_unlock(&pending_requests.lock);
347 
348 		if (req != NULL)
349 			trigger_async_action(req);
350 		return;
351 	}
352 
353 	pthread_mutex_lock(&mp_mutex_action);
354 	entry = find_action_entry_by_name(msg->name);
355 	if (entry != NULL)
356 		action = entry->action;
357 	pthread_mutex_unlock(&mp_mutex_action);
358 
359 	if (!action) {
360 		if (m->type == MP_REQ && !internal_conf->init_complete) {
361 			/* if this is a request, and init is not yet complete,
362 			 * and callback wasn't registered, we should tell the
363 			 * requester to ignore our existence because we're not
364 			 * yet ready to process this request.
365 			 */
366 			struct rte_mp_msg dummy;
367 
368 			memset(&dummy, 0, sizeof(dummy));
369 			strlcpy(dummy.name, msg->name, sizeof(dummy.name));
370 			mp_send(&dummy, s->sun_path, MP_IGN);
371 		} else {
372 			RTE_LOG(ERR, EAL, "Cannot find action: %s\n",
373 				msg->name);
374 		}
375 	} else if (action(msg, s->sun_path) < 0) {
376 		RTE_LOG(ERR, EAL, "Fail to handle message: %s\n", msg->name);
377 	}
378 }
379 
380 static void *
381 mp_handle(void *arg __rte_unused)
382 {
383 	struct mp_msg_internal msg;
384 	struct sockaddr_un sa;
385 
386 	while (1) {
387 		if (read_msg(&msg, &sa) == 0)
388 			process_msg(&msg, &sa);
389 	}
390 
391 	return NULL;
392 }
393 
394 static int
395 timespec_cmp(const struct timespec *a, const struct timespec *b)
396 {
397 	if (a->tv_sec < b->tv_sec)
398 		return -1;
399 	if (a->tv_sec > b->tv_sec)
400 		return 1;
401 	if (a->tv_nsec < b->tv_nsec)
402 		return -1;
403 	if (a->tv_nsec > b->tv_nsec)
404 		return 1;
405 	return 0;
406 }
407 
408 enum async_action {
409 	ACTION_FREE, /**< free the action entry, but don't trigger callback */
410 	ACTION_TRIGGER /**< trigger callback, then free action entry */
411 };
412 
413 static enum async_action
414 process_async_request(struct pending_request *sr, const struct timespec *now)
415 {
416 	struct async_request_param *param;
417 	struct rte_mp_reply *reply;
418 	bool timeout, last_msg;
419 
420 	param = sr->async.param;
421 	reply = &param->user_reply;
422 
423 	/* did we timeout? */
424 	timeout = timespec_cmp(&param->end, now) <= 0;
425 
426 	/* if we received a response, adjust relevant data and copy message. */
427 	if (sr->reply_received == 1 && sr->reply) {
428 		struct rte_mp_msg *msg, *user_msgs, *tmp;
429 
430 		msg = sr->reply;
431 		user_msgs = reply->msgs;
432 
433 		tmp = realloc(user_msgs, sizeof(*msg) *
434 				(reply->nb_received + 1));
435 		if (!tmp) {
436 			RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
437 				sr->dst, sr->request->name);
438 			/* this entry is going to be removed and its message
439 			 * dropped, but we don't want to leak memory, so
440 			 * continue.
441 			 */
442 		} else {
443 			user_msgs = tmp;
444 			reply->msgs = user_msgs;
445 			memcpy(&user_msgs[reply->nb_received],
446 					msg, sizeof(*msg));
447 			reply->nb_received++;
448 		}
449 
450 		/* mark this request as processed */
451 		param->n_responses_processed++;
452 	} else if (sr->reply_received == -1) {
453 		/* we were asked to ignore this process */
454 		reply->nb_sent--;
455 	} else if (timeout) {
456 		/* count it as processed response, but don't increment
457 		 * nb_received.
458 		 */
459 		param->n_responses_processed++;
460 	}
461 
462 	free(sr->reply);
463 
464 	last_msg = param->n_responses_processed == reply->nb_sent;
465 
466 	return last_msg ? ACTION_TRIGGER : ACTION_FREE;
467 }
468 
469 static void
470 trigger_async_action(struct pending_request *sr)
471 {
472 	struct async_request_param *param;
473 	struct rte_mp_reply *reply;
474 
475 	param = sr->async.param;
476 	reply = &param->user_reply;
477 
478 	param->clb(sr->request, reply);
479 
480 	/* clean up */
481 	free(sr->async.param->user_reply.msgs);
482 	free(sr->async.param);
483 	free(sr->request);
484 	free(sr);
485 }
486 
487 static struct pending_request *
488 async_reply_handle_thread_unsafe(void *arg)
489 {
490 	struct pending_request *req = (struct pending_request *)arg;
491 	enum async_action action;
492 	struct timespec ts_now;
493 
494 	if (clock_gettime(CLOCK_MONOTONIC, &ts_now) < 0) {
495 		RTE_LOG(ERR, EAL, "Cannot get current time\n");
496 		goto no_trigger;
497 	}
498 
499 	action = process_async_request(req, &ts_now);
500 
501 	TAILQ_REMOVE(&pending_requests.requests, req, next);
502 
503 	if (rte_eal_alarm_cancel(async_reply_handle, req) < 0) {
504 		/* if we failed to cancel the alarm because it's already in
505 		 * progress, don't proceed because otherwise we will end up
506 		 * handling the same message twice.
507 		 */
508 		if (rte_errno == EINPROGRESS) {
509 			RTE_LOG(DEBUG, EAL, "Request handling is already in progress\n");
510 			goto no_trigger;
511 		}
512 		RTE_LOG(ERR, EAL, "Failed to cancel alarm\n");
513 	}
514 
515 	if (action == ACTION_TRIGGER)
516 		return req;
517 no_trigger:
518 	free(req);
519 	return NULL;
520 }
521 
522 static void
523 async_reply_handle(void *arg)
524 {
525 	struct pending_request *req;
526 
527 	pthread_mutex_lock(&pending_requests.lock);
528 	req = async_reply_handle_thread_unsafe(arg);
529 	pthread_mutex_unlock(&pending_requests.lock);
530 
531 	if (req != NULL)
532 		trigger_async_action(req);
533 }
534 
535 static int
536 open_socket_fd(void)
537 {
538 	struct sockaddr_un un;
539 
540 	peer_name[0] = '\0';
541 	if (rte_eal_process_type() == RTE_PROC_SECONDARY)
542 		snprintf(peer_name, sizeof(peer_name),
543 				"%d_%"PRIx64, getpid(), rte_rdtsc());
544 
545 	mp_fd = socket(AF_UNIX, SOCK_DGRAM, 0);
546 	if (mp_fd < 0) {
547 		RTE_LOG(ERR, EAL, "failed to create unix socket\n");
548 		return -1;
549 	}
550 
551 	memset(&un, 0, sizeof(un));
552 	un.sun_family = AF_UNIX;
553 
554 	create_socket_path(peer_name, un.sun_path, sizeof(un.sun_path));
555 
556 	unlink(un.sun_path); /* May still exist since last run */
557 
558 	if (bind(mp_fd, (struct sockaddr *)&un, sizeof(un)) < 0) {
559 		RTE_LOG(ERR, EAL, "failed to bind %s: %s\n",
560 			un.sun_path, strerror(errno));
561 		close(mp_fd);
562 		return -1;
563 	}
564 
565 	RTE_LOG(INFO, EAL, "Multi-process socket %s\n", un.sun_path);
566 	return mp_fd;
567 }
568 
569 static void
570 close_socket_fd(void)
571 {
572 	char path[PATH_MAX];
573 
574 	if (mp_fd < 0)
575 		return;
576 
577 	close(mp_fd);
578 	create_socket_path(peer_name, path, sizeof(path));
579 	unlink(path);
580 }
581 
582 int
583 rte_mp_channel_init(void)
584 {
585 	char path[PATH_MAX];
586 	int dir_fd;
587 	pthread_t mp_handle_tid;
588 	const struct internal_config *internal_conf =
589 		eal_get_internal_configuration();
590 
591 	/* in no shared files mode, we do not have secondary processes support,
592 	 * so no need to initialize IPC.
593 	 */
594 	if (internal_conf->no_shconf) {
595 		RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC will be disabled\n");
596 		rte_errno = ENOTSUP;
597 		return -1;
598 	}
599 
600 	/* create filter path */
601 	create_socket_path("*", path, sizeof(path));
602 	strlcpy(mp_filter, basename(path), sizeof(mp_filter));
603 
604 	/* path may have been modified, so recreate it */
605 	create_socket_path("*", path, sizeof(path));
606 	strlcpy(mp_dir_path, dirname(path), sizeof(mp_dir_path));
607 
608 	/* lock the directory */
609 	dir_fd = open(mp_dir_path, O_RDONLY);
610 	if (dir_fd < 0) {
611 		RTE_LOG(ERR, EAL, "failed to open %s: %s\n",
612 			mp_dir_path, strerror(errno));
613 		return -1;
614 	}
615 
616 	if (flock(dir_fd, LOCK_EX)) {
617 		RTE_LOG(ERR, EAL, "failed to lock %s: %s\n",
618 			mp_dir_path, strerror(errno));
619 		close(dir_fd);
620 		return -1;
621 	}
622 
623 	if (open_socket_fd() < 0) {
624 		close(dir_fd);
625 		return -1;
626 	}
627 
628 	if (rte_ctrl_thread_create(&mp_handle_tid, "rte_mp_handle",
629 			NULL, mp_handle, NULL) < 0) {
630 		RTE_LOG(ERR, EAL, "failed to create mp thread: %s\n",
631 			strerror(errno));
632 		close(mp_fd);
633 		close(dir_fd);
634 		mp_fd = -1;
635 		return -1;
636 	}
637 
638 	/* unlock the directory */
639 	flock(dir_fd, LOCK_UN);
640 	close(dir_fd);
641 
642 	return 0;
643 }
644 
645 void
646 rte_mp_channel_cleanup(void)
647 {
648 	close_socket_fd();
649 }
650 
651 /**
652  * Return -1, as fail to send message and it's caused by the local side.
653  * Return 0, as fail to send message and it's caused by the remote side.
654  * Return 1, as succeed to send message.
655  *
656  */
657 static int
658 send_msg(const char *dst_path, struct rte_mp_msg *msg, int type)
659 {
660 	int snd;
661 	struct iovec iov;
662 	struct msghdr msgh;
663 	struct cmsghdr *cmsg;
664 	struct sockaddr_un dst;
665 	struct mp_msg_internal m;
666 	int fd_size = msg->num_fds * sizeof(int);
667 	char control[CMSG_SPACE(fd_size)];
668 
669 	m.type = type;
670 	memcpy(&m.msg, msg, sizeof(*msg));
671 
672 	memset(&dst, 0, sizeof(dst));
673 	dst.sun_family = AF_UNIX;
674 	strlcpy(dst.sun_path, dst_path, sizeof(dst.sun_path));
675 
676 	memset(&msgh, 0, sizeof(msgh));
677 	memset(control, 0, sizeof(control));
678 
679 	iov.iov_base = &m;
680 	iov.iov_len = sizeof(m) - sizeof(msg->fds);
681 
682 	msgh.msg_name = &dst;
683 	msgh.msg_namelen = sizeof(dst);
684 	msgh.msg_iov = &iov;
685 	msgh.msg_iovlen = 1;
686 	msgh.msg_control = control;
687 	msgh.msg_controllen = sizeof(control);
688 
689 	cmsg = CMSG_FIRSTHDR(&msgh);
690 	cmsg->cmsg_len = CMSG_LEN(fd_size);
691 	cmsg->cmsg_level = SOL_SOCKET;
692 	cmsg->cmsg_type = SCM_RIGHTS;
693 	memcpy(CMSG_DATA(cmsg), msg->fds, fd_size);
694 
695 	do {
696 		snd = sendmsg(mp_fd, &msgh, 0);
697 	} while (snd < 0 && errno == EINTR);
698 
699 	if (snd < 0) {
700 		rte_errno = errno;
701 		/* Check if it caused by peer process exits */
702 		if (errno == ECONNREFUSED &&
703 				rte_eal_process_type() == RTE_PROC_PRIMARY) {
704 			unlink(dst_path);
705 			return 0;
706 		}
707 		RTE_LOG(ERR, EAL, "failed to send to (%s) due to %s\n",
708 			dst_path, strerror(errno));
709 		return -1;
710 	}
711 
712 	return 1;
713 }
714 
715 static int
716 mp_send(struct rte_mp_msg *msg, const char *peer, int type)
717 {
718 	int dir_fd, ret = 0;
719 	DIR *mp_dir;
720 	struct dirent *ent;
721 
722 	if (!peer && (rte_eal_process_type() == RTE_PROC_SECONDARY))
723 		peer = eal_mp_socket_path();
724 
725 	if (peer) {
726 		if (send_msg(peer, msg, type) < 0)
727 			return -1;
728 		else
729 			return 0;
730 	}
731 
732 	/* broadcast to all secondary processes */
733 	mp_dir = opendir(mp_dir_path);
734 	if (!mp_dir) {
735 		RTE_LOG(ERR, EAL, "Unable to open directory %s\n",
736 				mp_dir_path);
737 		rte_errno = errno;
738 		return -1;
739 	}
740 
741 	dir_fd = dirfd(mp_dir);
742 	/* lock the directory to prevent processes spinning up while we send */
743 	if (flock(dir_fd, LOCK_SH)) {
744 		RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
745 			mp_dir_path);
746 		rte_errno = errno;
747 		closedir(mp_dir);
748 		return -1;
749 	}
750 
751 	while ((ent = readdir(mp_dir))) {
752 		char path[PATH_MAX];
753 
754 		if (fnmatch(mp_filter, ent->d_name, 0) != 0)
755 			continue;
756 
757 		snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
758 			 ent->d_name);
759 		if (send_msg(path, msg, type) < 0)
760 			ret = -1;
761 	}
762 	/* unlock the dir */
763 	flock(dir_fd, LOCK_UN);
764 
765 	/* dir_fd automatically closed on closedir */
766 	closedir(mp_dir);
767 	return ret;
768 }
769 
770 static int
771 check_input(const struct rte_mp_msg *msg)
772 {
773 	if (msg == NULL) {
774 		RTE_LOG(ERR, EAL, "Msg cannot be NULL\n");
775 		rte_errno = EINVAL;
776 		return -1;
777 	}
778 
779 	if (validate_action_name(msg->name) != 0)
780 		return -1;
781 
782 	if (msg->len_param < 0) {
783 		RTE_LOG(ERR, EAL, "Message data length is negative\n");
784 		rte_errno = EINVAL;
785 		return -1;
786 	}
787 
788 	if (msg->num_fds < 0) {
789 		RTE_LOG(ERR, EAL, "Number of fd's is negative\n");
790 		rte_errno = EINVAL;
791 		return -1;
792 	}
793 
794 	if (msg->len_param > RTE_MP_MAX_PARAM_LEN) {
795 		RTE_LOG(ERR, EAL, "Message data is too long\n");
796 		rte_errno = E2BIG;
797 		return -1;
798 	}
799 
800 	if (msg->num_fds > RTE_MP_MAX_FD_NUM) {
801 		RTE_LOG(ERR, EAL, "Cannot send more than %d FDs\n",
802 			RTE_MP_MAX_FD_NUM);
803 		rte_errno = E2BIG;
804 		return -1;
805 	}
806 
807 	return 0;
808 }
809 
810 int
811 rte_mp_sendmsg(struct rte_mp_msg *msg)
812 {
813 	const struct internal_config *internal_conf =
814 		eal_get_internal_configuration();
815 
816 	if (check_input(msg) != 0)
817 		return -1;
818 
819 	if (internal_conf->no_shconf) {
820 		RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
821 		rte_errno = ENOTSUP;
822 		return -1;
823 	}
824 
825 	RTE_LOG(DEBUG, EAL, "sendmsg: %s\n", msg->name);
826 	return mp_send(msg, NULL, MP_MSG);
827 }
828 
829 static int
830 mp_request_async(const char *dst, struct rte_mp_msg *req,
831 		struct async_request_param *param, const struct timespec *ts)
832 {
833 	struct rte_mp_msg *reply_msg;
834 	struct pending_request *pending_req, *exist;
835 	int ret = -1;
836 
837 	pending_req = calloc(1, sizeof(*pending_req));
838 	reply_msg = calloc(1, sizeof(*reply_msg));
839 	if (pending_req == NULL || reply_msg == NULL) {
840 		RTE_LOG(ERR, EAL, "Could not allocate space for sync request\n");
841 		rte_errno = ENOMEM;
842 		ret = -1;
843 		goto fail;
844 	}
845 
846 	pending_req->type = REQUEST_TYPE_ASYNC;
847 	strlcpy(pending_req->dst, dst, sizeof(pending_req->dst));
848 	pending_req->request = req;
849 	pending_req->reply = reply_msg;
850 	pending_req->async.param = param;
851 
852 	/* queue already locked by caller */
853 
854 	exist = find_pending_request(dst, req->name);
855 	if (exist) {
856 		RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
857 		rte_errno = EEXIST;
858 		ret = -1;
859 		goto fail;
860 	}
861 
862 	ret = send_msg(dst, req, MP_REQ);
863 	if (ret < 0) {
864 		RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
865 			dst, req->name);
866 		ret = -1;
867 		goto fail;
868 	} else if (ret == 0) {
869 		ret = 0;
870 		goto fail;
871 	}
872 	param->user_reply.nb_sent++;
873 
874 	/* if alarm set fails, we simply ignore the reply */
875 	if (rte_eal_alarm_set(ts->tv_sec * 1000000 + ts->tv_nsec / 1000,
876 			      async_reply_handle, pending_req) < 0) {
877 		RTE_LOG(ERR, EAL, "Fail to set alarm for request %s:%s\n",
878 			dst, req->name);
879 		ret = -1;
880 		goto fail;
881 	}
882 	TAILQ_INSERT_TAIL(&pending_requests.requests, pending_req, next);
883 
884 	return 0;
885 fail:
886 	free(pending_req);
887 	free(reply_msg);
888 	return ret;
889 }
890 
891 static int
892 mp_request_sync(const char *dst, struct rte_mp_msg *req,
893 	       struct rte_mp_reply *reply, const struct timespec *ts)
894 {
895 	int ret;
896 	pthread_condattr_t attr;
897 	struct rte_mp_msg msg, *tmp;
898 	struct pending_request pending_req, *exist;
899 
900 	pending_req.type = REQUEST_TYPE_SYNC;
901 	pending_req.reply_received = 0;
902 	strlcpy(pending_req.dst, dst, sizeof(pending_req.dst));
903 	pending_req.request = req;
904 	pending_req.reply = &msg;
905 	pthread_condattr_init(&attr);
906 	pthread_condattr_setclock(&attr, CLOCK_MONOTONIC);
907 	pthread_cond_init(&pending_req.sync.cond, &attr);
908 
909 	exist = find_pending_request(dst, req->name);
910 	if (exist) {
911 		RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
912 		rte_errno = EEXIST;
913 		return -1;
914 	}
915 
916 	ret = send_msg(dst, req, MP_REQ);
917 	if (ret < 0) {
918 		RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
919 			dst, req->name);
920 		return -1;
921 	} else if (ret == 0)
922 		return 0;
923 
924 	TAILQ_INSERT_TAIL(&pending_requests.requests, &pending_req, next);
925 
926 	reply->nb_sent++;
927 
928 	do {
929 		ret = pthread_cond_timedwait(&pending_req.sync.cond,
930 				&pending_requests.lock, ts);
931 	} while (ret != 0 && ret != ETIMEDOUT);
932 
933 	TAILQ_REMOVE(&pending_requests.requests, &pending_req, next);
934 
935 	if (pending_req.reply_received == 0) {
936 		RTE_LOG(ERR, EAL, "Fail to recv reply for request %s:%s\n",
937 			dst, req->name);
938 		rte_errno = ETIMEDOUT;
939 		return -1;
940 	}
941 	if (pending_req.reply_received == -1) {
942 		RTE_LOG(DEBUG, EAL, "Asked to ignore response\n");
943 		/* not receiving this message is not an error, so decrement
944 		 * number of sent messages
945 		 */
946 		reply->nb_sent--;
947 		return 0;
948 	}
949 
950 	tmp = realloc(reply->msgs, sizeof(msg) * (reply->nb_received + 1));
951 	if (!tmp) {
952 		RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
953 			dst, req->name);
954 		rte_errno = ENOMEM;
955 		return -1;
956 	}
957 	memcpy(&tmp[reply->nb_received], &msg, sizeof(msg));
958 	reply->msgs = tmp;
959 	reply->nb_received++;
960 	return 0;
961 }
962 
963 int
964 rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
965 		const struct timespec *ts)
966 {
967 	int dir_fd, ret = -1;
968 	DIR *mp_dir;
969 	struct dirent *ent;
970 	struct timespec now, end;
971 	const struct internal_config *internal_conf =
972 		eal_get_internal_configuration();
973 
974 	RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
975 
976 	reply->nb_sent = 0;
977 	reply->nb_received = 0;
978 	reply->msgs = NULL;
979 
980 	if (check_input(req) != 0)
981 		goto end;
982 
983 	if (internal_conf->no_shconf) {
984 		RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
985 		rte_errno = ENOTSUP;
986 		return -1;
987 	}
988 
989 	if (clock_gettime(CLOCK_MONOTONIC, &now) < 0) {
990 		RTE_LOG(ERR, EAL, "Failed to get current time\n");
991 		rte_errno = errno;
992 		goto end;
993 	}
994 
995 	end.tv_nsec = (now.tv_nsec + ts->tv_nsec) % 1000000000;
996 	end.tv_sec = now.tv_sec + ts->tv_sec +
997 			(now.tv_nsec + ts->tv_nsec) / 1000000000;
998 
999 	/* for secondary process, send request to the primary process only */
1000 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1001 		pthread_mutex_lock(&pending_requests.lock);
1002 		ret = mp_request_sync(eal_mp_socket_path(), req, reply, &end);
1003 		pthread_mutex_unlock(&pending_requests.lock);
1004 		goto end;
1005 	}
1006 
1007 	/* for primary process, broadcast request, and collect reply 1 by 1 */
1008 	mp_dir = opendir(mp_dir_path);
1009 	if (!mp_dir) {
1010 		RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1011 		rte_errno = errno;
1012 		goto end;
1013 	}
1014 
1015 	dir_fd = dirfd(mp_dir);
1016 	/* lock the directory to prevent processes spinning up while we send */
1017 	if (flock(dir_fd, LOCK_SH)) {
1018 		RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1019 			mp_dir_path);
1020 		rte_errno = errno;
1021 		goto close_end;
1022 	}
1023 
1024 	pthread_mutex_lock(&pending_requests.lock);
1025 	while ((ent = readdir(mp_dir))) {
1026 		char path[PATH_MAX];
1027 
1028 		if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1029 			continue;
1030 
1031 		snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1032 			 ent->d_name);
1033 
1034 		/* unlocks the mutex while waiting for response,
1035 		 * locks on receive
1036 		 */
1037 		if (mp_request_sync(path, req, reply, &end))
1038 			goto unlock_end;
1039 	}
1040 	ret = 0;
1041 
1042 unlock_end:
1043 	pthread_mutex_unlock(&pending_requests.lock);
1044 	/* unlock the directory */
1045 	flock(dir_fd, LOCK_UN);
1046 
1047 close_end:
1048 	/* dir_fd automatically closed on closedir */
1049 	closedir(mp_dir);
1050 
1051 end:
1052 	if (ret) {
1053 		free(reply->msgs);
1054 		reply->nb_received = 0;
1055 		reply->msgs = NULL;
1056 	}
1057 	return ret;
1058 }
1059 
1060 int
1061 rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
1062 		rte_mp_async_reply_t clb)
1063 {
1064 	struct rte_mp_msg *copy;
1065 	struct pending_request *dummy;
1066 	struct async_request_param *param;
1067 	struct rte_mp_reply *reply;
1068 	int dir_fd, ret = 0;
1069 	DIR *mp_dir;
1070 	struct dirent *ent;
1071 	struct timespec now;
1072 	struct timespec *end;
1073 	bool dummy_used = false;
1074 	const struct internal_config *internal_conf =
1075 		eal_get_internal_configuration();
1076 
1077 	RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
1078 
1079 	if (check_input(req) != 0)
1080 		return -1;
1081 
1082 	if (internal_conf->no_shconf) {
1083 		RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
1084 		rte_errno = ENOTSUP;
1085 		return -1;
1086 	}
1087 
1088 	if (clock_gettime(CLOCK_MONOTONIC, &now) < 0) {
1089 		RTE_LOG(ERR, EAL, "Failed to get current time\n");
1090 		rte_errno = errno;
1091 		return -1;
1092 	}
1093 	copy = calloc(1, sizeof(*copy));
1094 	dummy = calloc(1, sizeof(*dummy));
1095 	param = calloc(1, sizeof(*param));
1096 	if (copy == NULL || dummy == NULL || param == NULL) {
1097 		RTE_LOG(ERR, EAL, "Failed to allocate memory for async reply\n");
1098 		rte_errno = ENOMEM;
1099 		goto fail;
1100 	}
1101 
1102 	/* copy message */
1103 	memcpy(copy, req, sizeof(*copy));
1104 
1105 	param->n_responses_processed = 0;
1106 	param->clb = clb;
1107 	end = &param->end;
1108 	reply = &param->user_reply;
1109 
1110 	end->tv_nsec = (now.tv_nsec + ts->tv_nsec) % 1000000000;
1111 	end->tv_sec = now.tv_sec + ts->tv_sec +
1112 			(now.tv_nsec + ts->tv_nsec) / 1000000000;
1113 	reply->nb_sent = 0;
1114 	reply->nb_received = 0;
1115 	reply->msgs = NULL;
1116 
1117 	/* we have to lock the request queue here, as we will be adding a bunch
1118 	 * of requests to the queue at once, and some of the replies may arrive
1119 	 * before we add all of the requests to the queue.
1120 	 */
1121 	pthread_mutex_lock(&pending_requests.lock);
1122 
1123 	/* we have to ensure that callback gets triggered even if we don't send
1124 	 * anything, therefore earlier we have allocated a dummy request. fill
1125 	 * it, and put it on the queue if we don't send any requests.
1126 	 */
1127 	dummy->type = REQUEST_TYPE_ASYNC;
1128 	dummy->request = copy;
1129 	dummy->reply = NULL;
1130 	dummy->async.param = param;
1131 	dummy->reply_received = 1; /* short-circuit the timeout */
1132 
1133 	/* for secondary process, send request to the primary process only */
1134 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1135 		ret = mp_request_async(eal_mp_socket_path(), copy, param, ts);
1136 
1137 		/* if we didn't send anything, put dummy request on the queue */
1138 		if (ret == 0 && reply->nb_sent == 0) {
1139 			TAILQ_INSERT_TAIL(&pending_requests.requests, dummy,
1140 					next);
1141 			dummy_used = true;
1142 		}
1143 
1144 		pthread_mutex_unlock(&pending_requests.lock);
1145 
1146 		/* if we couldn't send anything, clean up */
1147 		if (ret != 0)
1148 			goto fail;
1149 		return 0;
1150 	}
1151 
1152 	/* for primary process, broadcast request */
1153 	mp_dir = opendir(mp_dir_path);
1154 	if (!mp_dir) {
1155 		RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1156 		rte_errno = errno;
1157 		goto unlock_fail;
1158 	}
1159 	dir_fd = dirfd(mp_dir);
1160 
1161 	/* lock the directory to prevent processes spinning up while we send */
1162 	if (flock(dir_fd, LOCK_SH)) {
1163 		RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1164 			mp_dir_path);
1165 		rte_errno = errno;
1166 		goto closedir_fail;
1167 	}
1168 
1169 	while ((ent = readdir(mp_dir))) {
1170 		char path[PATH_MAX];
1171 
1172 		if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1173 			continue;
1174 
1175 		snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1176 			 ent->d_name);
1177 
1178 		if (mp_request_async(path, copy, param, ts))
1179 			ret = -1;
1180 	}
1181 	/* if we didn't send anything, put dummy request on the queue */
1182 	if (ret == 0 && reply->nb_sent == 0) {
1183 		TAILQ_INSERT_HEAD(&pending_requests.requests, dummy, next);
1184 		dummy_used = true;
1185 	}
1186 
1187 	/* finally, unlock the queue */
1188 	pthread_mutex_unlock(&pending_requests.lock);
1189 
1190 	/* unlock the directory */
1191 	flock(dir_fd, LOCK_UN);
1192 
1193 	/* dir_fd automatically closed on closedir */
1194 	closedir(mp_dir);
1195 
1196 	/* if dummy was unused, free it */
1197 	if (!dummy_used)
1198 		free(dummy);
1199 
1200 	return ret;
1201 closedir_fail:
1202 	closedir(mp_dir);
1203 unlock_fail:
1204 	pthread_mutex_unlock(&pending_requests.lock);
1205 fail:
1206 	free(dummy);
1207 	free(param);
1208 	free(copy);
1209 	return -1;
1210 }
1211 
1212 int
1213 rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
1214 {
1215 	RTE_LOG(DEBUG, EAL, "reply: %s\n", msg->name);
1216 	const struct internal_config *internal_conf =
1217 		eal_get_internal_configuration();
1218 
1219 	if (check_input(msg) != 0)
1220 		return -1;
1221 
1222 	if (peer == NULL) {
1223 		RTE_LOG(ERR, EAL, "peer is not specified\n");
1224 		rte_errno = EINVAL;
1225 		return -1;
1226 	}
1227 
1228 	if (internal_conf->no_shconf) {
1229 		RTE_LOG(DEBUG, EAL, "No shared files mode enabled, IPC is disabled\n");
1230 		return 0;
1231 	}
1232 
1233 	return mp_send(msg, peer, MP_REP);
1234 }
1235 
1236 /* Internally, the status of the mp feature is represented as a three-state:
1237  * - "unknown" as long as no secondary process attached to a primary process
1238  *   and there was no call to rte_mp_disable yet,
1239  * - "enabled" as soon as a secondary process attaches to a primary process,
1240  * - "disabled" when a primary process successfully called rte_mp_disable,
1241  */
1242 enum mp_status {
1243 	MP_STATUS_UNKNOWN,
1244 	MP_STATUS_DISABLED,
1245 	MP_STATUS_ENABLED,
1246 };
1247 
1248 static bool
1249 set_mp_status(enum mp_status status)
1250 {
1251 	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1252 	uint8_t expected;
1253 	uint8_t desired;
1254 
1255 	expected = MP_STATUS_UNKNOWN;
1256 	desired = status;
1257 	if (__atomic_compare_exchange_n(&mcfg->mp_status, &expected, desired,
1258 			false, __ATOMIC_RELAXED, __ATOMIC_RELAXED))
1259 		return true;
1260 
1261 	return __atomic_load_n(&mcfg->mp_status, __ATOMIC_RELAXED) == desired;
1262 }
1263 
1264 bool
1265 rte_mp_disable(void)
1266 {
1267 	return set_mp_status(MP_STATUS_DISABLED);
1268 }
1269 
1270 bool
1271 __rte_mp_enable(void)
1272 {
1273 	return set_mp_status(MP_STATUS_ENABLED);
1274 }
1275