1 /* Copyright (c) 2009, 2010. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
12 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_network, simix,
13 "Logging specific to SIMIX (network)");
15 /******************************************************************************/
16 /* Rendez-Vous Points */
17 /******************************************************************************/
20 * \brief Creates a new rendez-vous point
21 * \param name The name of the rendez-vous point
22 * \return The created rendez-vous point
24 smx_rdv_t SIMIX_rdv_create(const char *name)
26 smx_rdv_t rdv = xbt_new0(s_smx_rvpoint_t, 1);
27 rdv->name = name ? xbt_strdup(name) : NULL;
28 rdv->read = SIMIX_mutex_init();
29 rdv->write = SIMIX_mutex_init();
30 rdv->comm_fifo = xbt_fifo_new();
36 * \brief Destroy a rendez-vous point
37 * \param name The rendez-vous point to destroy
39 void SIMIX_rdv_destroy(smx_rdv_t rdv)
43 SIMIX_mutex_destroy(rdv->read);
44 SIMIX_mutex_destroy(rdv->write);
45 xbt_fifo_free(rdv->comm_fifo);
50 * \brief Push a communication request into a rendez-vous point
51 * \param rdv The rendez-vous point
52 * \param comm The communication request
54 static XBT_INLINE void SIMIX_rdv_push(smx_rdv_t rdv, smx_comm_t comm)
56 xbt_fifo_push(rdv->comm_fifo, comm);
61 * \brief Remove a communication request from a rendez-vous point
62 * \param rdv The rendez-vous point
63 * \param comm The communication request
65 static XBT_INLINE void SIMIX_rdv_remove(smx_rdv_t rdv, smx_comm_t comm)
67 xbt_fifo_remove(rdv->comm_fifo, comm);
72 * \brief Checks if there is a communication request queued in a rendez-vous matching our needs
73 * \param type The type of communication we are looking for (comm_send, comm_recv)
74 * \return The communication request if found, NULL otherwise.
76 smx_comm_t SIMIX_rdv_get_request(smx_rdv_t rdv, smx_comm_type_t type)
78 smx_comm_t comm = (smx_comm_t)
79 xbt_fifo_get_item_content(xbt_fifo_get_first_item(rdv->comm_fifo));
81 if (comm && comm->type == type) {
82 DEBUG0("Communication request found!");
83 xbt_fifo_shift(rdv->comm_fifo);
84 SIMIX_communication_use(comm);
89 DEBUG0("Communication request not found");
94 * \brief counts the number of communication requests of a given host pending
95 * on a rendez-vous point
96 * \param rdv The rendez-vous point
97 * \param host The host to be counted
98 * \return The number of comm request pending in the rdv
100 int SIMIX_rdv_get_count_waiting_comm(smx_rdv_t rdv, smx_host_t host)
102 smx_comm_t comm = NULL;
103 xbt_fifo_item_t item = NULL;
106 xbt_fifo_foreach(rdv->comm_fifo, item, comm, smx_comm_t) {
107 if (comm->src_proc->smx_host == host)
115 * \brief returns the communication at the head of the rendez-vous
116 * \param rdv The rendez-vous point
117 * \return The communication or NULL if empty
119 XBT_INLINE smx_comm_t SIMIX_rdv_get_head(smx_rdv_t rdv)
122 xbt_fifo_get_item_content(xbt_fifo_get_first_item(rdv->comm_fifo));
125 /** @brief adds some API-related data to the rendez-vous point */
126 XBT_INLINE void SIMIX_rdv_set_data(smx_rdv_t rdv, void *data)
131 /** @brief gets API-related data from the rendez-vous point */
132 XBT_INLINE void *SIMIX_rdv_get_data(smx_rdv_t rdv)
137 /******************************************************************************/
138 /* Communication Requests */
139 /******************************************************************************/
142 * \brief Creates a new communication request
143 * \param type The type of communication (comm_send, comm_recv)
144 * \return The new communication request
146 smx_comm_t SIMIX_communication_new(smx_comm_type_t type)
148 /* alloc structures */
149 smx_comm_t comm = xbt_new0(s_smx_comm_t, 1);
151 comm->sem = SIMIX_sem_init(0);
153 VERB2("Create communication %p; refcount initially %d", comm,
160 * \brief Destroy a communication request
161 * \param comm The request to be destroyed
163 void SIMIX_communication_destroy(smx_comm_t comm)
165 VERB2("Destroy communication %p; refcount initially %d", comm,
168 if(!(comm->refcount>0))
169 xbt_die(bprintf("the refcount of comm %p is already 0 before decreasing it. That's a bug!",comm));
171 #ifdef HAVE_LATENCY_BOUND_TRACKING
172 //save is latency limited flag to use afterwards
174 DEBUG2("adding key %p with latency limited value %d to the dict", comm,
175 SIMIX_action_is_latency_bounded(comm->act));
176 xbt_dicti_set(simix_global->latency_limited_dict, (uintptr_t) comm,
177 SIMIX_action_is_latency_bounded(comm->act));
182 if (comm->refcount > 0)
184 VERB2("Really free communication %p; refcount is now %d", comm,
188 SIMIX_sem_destroy(comm->sem);
193 SIMIX_action_destroy(comm->act);
197 if (comm->src_timeout) {
198 SIMIX_action_destroy(comm->src_timeout);
199 comm->src_timeout = NULL;
202 if (comm->dst_timeout) {
203 SIMIX_action_destroy(comm->dst_timeout);
204 comm->dst_timeout = NULL;
213 * \brief Increase the number of users of the communication.
214 * \param comm The communication request
215 * Each communication request can be used by more than one process, so it is
216 * necessary to know number of them at destroy time, to avoid freeing stuff that
217 * maybe is in use by others.
220 static XBT_INLINE void SIMIX_communication_use(smx_comm_t comm)
224 VERB2("Use communication %p; refcount is now %d", comm,
229 * \brief Start the simulation of a communication request
230 * \param comm The communication request
232 static XBT_INLINE void SIMIX_communication_start(smx_comm_t comm)
234 /* If both the sender and the receiver are already there, start the communication */
235 if (comm->src_proc && comm->dst_proc) {
236 DEBUG1("Starting communication %p", comm);
237 comm->act = SIMIX_action_communicate(comm->src_proc->smx_host,
238 comm->dst_proc->smx_host, NULL,
239 comm->task_size, comm->rate);
241 TRACE_smx_action_communicate(comm->act, comm->src_proc);
242 TRACE_surf_action(comm->act->surf_action, comm->act->category);
245 /* If any of the process is suspend, create the action but stop its execution,
246 it will be restarted when the sender process resume */
247 if (SIMIX_process_is_suspended(comm->src_proc) ||
248 SIMIX_process_is_suspended(comm->dst_proc)) {
249 SIMIX_action_suspend(comm->act);
252 /* Add the communication as user data of the action */
253 comm->act->data = comm;
255 /* The semaphore will only get signaled once, but since the first unlocked guy will
256 * release_forever() the semaphore, that will unlock the second (and any other)
257 * communication partner */
258 SIMIX_register_action_to_semaphore(comm->act, comm->sem);
263 * \brief Performs error checking and cleanup
264 * \param comm The communication
266 static XBT_INLINE void SIMIX_communication_cleanup(smx_comm_t comm)
268 DEBUG1("Checking errors and cleaning communication %p", comm);
270 /* Make sure that everyone sleeping on that semaphore is awake, and that nobody will ever block on it */
271 SIMIX_sem_release_forever(comm->sem);
273 /* Check for errors other than timeouts */
274 if (!SIMIX_host_get_state(SIMIX_host_self())) {
276 SIMIX_rdv_remove(comm->rdv, comm);
277 SIMIX_communication_destroy(comm);
278 THROW0(host_error, 0, "Host failed");
279 } else if (SIMIX_action_get_state(comm->act) == SURF_ACTION_FAILED) {
280 SIMIX_communication_destroy(comm);
281 THROW0(network_error, 0, "Link failure");
282 } else if (!SIMIX_host_get_state(SIMIX_process_get_host(comm->dst_proc))
284 !SIMIX_host_get_state(SIMIX_process_get_host(comm->src_proc)))
286 /* We test both src&dst because we dunno who we are today, and we already tested myself above.
287 * So, at the end, we test the remote peer only
288 * Moreover, we have to test it because if the remote peer fails, the action comm->act is not done nor failed.
289 * In that case, we got awaken by the little endless actions created in the SIMIX_sem_acquire(comm->sem)
290 * at the beginning of this function. */
291 SIMIX_communication_destroy(comm);
292 THROW0(network_error, 0, "Remote peer failed");
295 /* Copy network data */
296 SIMIX_network_copy_data(comm);
298 SIMIX_communication_destroy(comm);
302 * \brief Waits for communication completion
303 * \param comm The communication
304 * \param timeout The max amount of time to wait for the communication to finish
307 * - host_error if local peer failed
308 * - timeout_error if communication reached the timeout specified (either because of local peer or remote peer)
309 * - network_error if network failed or remote peer failed
311 static XBT_INLINE void SIMIX_communication_wait_for_completion(smx_comm_t
316 smx_action_t act_sleep = NULL;
320 DEBUG1("Waiting for the completion of communication %p", comm);
323 act_sleep = SIMIX_action_sleep(SIMIX_host_self(), timeout);
324 if (SIMIX_process_self() == comm->src_proc)
325 comm->src_timeout = act_sleep;
327 comm->dst_timeout = act_sleep;
328 SIMIX_action_set_name(act_sleep,
330 ("Timeout for comm %p and wait on semaphore %p (max_duration:%f)",
331 comm, comm->sem, timeout));
332 SIMIX_register_action_to_semaphore(act_sleep, comm->sem);
333 SIMIX_process_self()->waiting_action = act_sleep;
334 SIMIX_sem_block_onto(comm->sem);
335 SIMIX_process_self()->waiting_action = NULL;
336 SIMIX_unregister_action_to_semaphore(act_sleep, comm->sem);
338 SIMIX_sem_acquire(comm->sem);
341 /* Check for timeouts */
342 if ((src_timeout = ((comm->src_timeout)
343 && (SIMIX_action_get_state(comm->src_timeout) ==
345 || (dst_timeout = ((comm->dst_timeout)
346 && (SIMIX_action_get_state(comm->dst_timeout) ==
347 SURF_ACTION_DONE)))) {
348 /* Somebody did a timeout! */
350 DEBUG1("Communication timeout from the src! %p", comm);
352 DEBUG1("Communication timeout from the dst! %p", comm);
355 && SIMIX_action_get_state(comm->act) == SURF_ACTION_RUNNING)
356 SIMIX_communication_cancel(comm);
358 SIMIX_rdv_remove(comm->rdv, comm);
360 /* Make sure that everyone sleeping on that semaphore is awake, and that nobody will ever block on it */
361 SIMIX_sem_release_forever(comm->sem);
362 SIMIX_communication_destroy(comm);
364 THROW1(timeout_error, 0, "Communication timeouted because of %s",
365 src_timeout ? "the source" : "the destination");
368 DEBUG1("Communication %p complete!", comm);
369 SIMIX_communication_cleanup(comm);
373 * \brief Cancels a communication
374 * \brief comm The communication to cancel
376 XBT_INLINE void SIMIX_communication_cancel(smx_comm_t comm)
379 SIMIX_action_cancel(comm->act);
383 * \brief get the amount remaining from the communication
384 * \param comm The communication
386 XBT_INLINE double SIMIX_communication_get_remains(smx_comm_t comm)
388 DEBUG1("calling SIMIX_action_get_remains(%p)", comm->act);
389 return SIMIX_action_get_remains(comm->act);
392 #ifdef HAVE_LATENCY_BOUND_TRACKING
394 * \brief verify if communication is latency bounded
395 * \param comm The communication
397 XBT_INLINE int SIMIX_communication_is_latency_bounded(smx_comm_t comm)
399 //try to find comm on the list of finished flows
402 xbt_dict_cursor_t cursor;
403 xbt_dict_foreach(simix_global->latency_limited_dict, cursor, key, data) {
404 DEBUG2("comparing key=%p with comm=%p", (void *) key, (void *) comm);
405 if ((void *) comm == (void *) key) {
406 DEBUG2("key %p found, return value latency limited value %d",
407 (void *) key, (int) data);
412 DEBUG1("calling SIMIX_action_is_latency_bounded(%p)", comm->act);
413 return SIMIX_action_is_latency_bounded(comm->act);
417 /******************************************************************************/
418 /* SIMIX_network_copy_data callbacks */
419 /******************************************************************************/
420 static void (*SIMIX_network_copy_data_callback) (smx_comm_t, size_t) =
421 &SIMIX_network_copy_pointer_callback;
424 SIMIX_network_set_copy_data_callback(void (*callback) (smx_comm_t, size_t))
426 SIMIX_network_copy_data_callback = callback;
429 void SIMIX_network_copy_pointer_callback(smx_comm_t comm, size_t buff_size)
431 xbt_assert1((buff_size == sizeof(void *)),
432 "Cannot copy %zu bytes: must be sizeof(void*)", buff_size);
433 *(void **) (comm->dst_buff) = comm->src_buff;
436 void SIMIX_network_copy_buffer_callback(smx_comm_t comm, size_t buff_size)
438 memcpy(comm->dst_buff, comm->src_buff, buff_size);
442 * \brief Copy the communication data from the sender's buffer to the receiver's one
443 * \param comm The communication
445 void SIMIX_network_copy_data(smx_comm_t comm)
447 size_t buff_size = comm->src_buff_size;
448 uintptr_t casted_size = 0;
449 uintptr_t amount = 0;
450 /* If there is no data to be copy then return */
451 if (!comm->src_buff || !comm->dst_buff || comm->copied == 1)
454 DEBUG6("Copying comm %p data from %s (%p) -> %s (%p) (%zu bytes)",
456 comm->src_proc->smx_host->name, comm->src_buff,
457 comm->dst_proc->smx_host->name, comm->dst_buff, buff_size);
459 /* Copy at most dst_buff_size bytes of the message to receiver's buffer */
460 if (comm->dst_buff_size)
461 buff_size = MIN(buff_size, *(comm->dst_buff_size));
463 /* Update the receiver's buffer size to the copied amount */
464 if (comm->dst_buff_size)
465 *comm->dst_buff_size = buff_size;
469 (*SIMIX_network_copy_data_callback) (comm, buff_size);
471 /* Set the copied flag so we copy data only once */
472 /* (this function might be called from both communication ends) */
475 /* pimple to display the message sizes */
477 casted_size = comm->task_size;
478 amount = xbt_dicti_get(simix_global->msg_sizes, casted_size);
481 xbt_dicti_set(simix_global->msg_sizes, casted_size, amount);
486 /* pimple to display the message sizes */
487 void SIMIX_message_sizes_output(const char *filename)
491 xbt_dict_cursor_t cursor;
493 out = fopen(filename, "w");
494 xbt_assert1(out, "Cannot open file %s", filename);
496 xbt_dict_foreach(simix_global->msg_sizes, cursor, key, data) {
497 fprintf(out, "%zu %zu\n", key, data);
503 * \brief Return the user data associated to the communication
504 * \param comm The communication
505 * \return the user data
507 XBT_INLINE void *SIMIX_communication_get_data(smx_comm_t comm)
512 XBT_PUBLIC(void *) SIMIX_communication_get_src_buf(smx_comm_t comm)
514 return comm->src_buff;
517 XBT_PUBLIC(void *) SIMIX_communication_get_dst_buf(smx_comm_t comm)
519 return comm->dst_buff;
522 XBT_PUBLIC(size_t) SIMIX_communication_get_src_buf_size(smx_comm_t comm)
524 return comm->src_buff_size;
527 XBT_PUBLIC(size_t) SIMIX_communication_get_dst_buf_size(smx_comm_t comm)
529 return *(comm->dst_buff_size);
532 /******************************************************************************/
533 /* Synchronous Communication */
534 /******************************************************************************/
536 * \brief Put a send communication request in a rendez-vous point and waits for
537 * its completion (blocking)
538 * \param rdv The rendez-vous point
539 * \param task_size The size of the communication action (for surf simulation)
540 * \param rate The rate of the communication action (for surf)
541 * \param timeout The timeout used for the waiting the completion
542 * \param src_buff The source buffer containing the message to be sent
543 * \param src_buff_size The size of the source buffer
544 * \param comm_ref The communication object used for the send (useful if someone else wants to cancel this communication afterward)
545 * \param data User data associated to the communication object
547 * - host_error if peer failed
548 * - timeout_error if communication reached the timeout specified
549 * - network_error if network failed or peer issued a timeout
551 XBT_INLINE void SIMIX_network_send(smx_rdv_t rdv, double task_size,
552 double rate, double timeout,
553 void *src_buff, size_t src_buff_size,
554 smx_comm_t * comm_ref, void *data)
557 smx_comm_t comm = *comm_ref =
558 SIMIX_network_isend(rdv, task_size, rate, src_buff, src_buff_size,
561 SIMIX_network_wait(comm, timeout);
564 SIMIX_communication_destroy(comm);
572 * \brief Put a receive communication request in a rendez-vous point and waits
573 * for its completion (blocking)
574 * \param rdv The rendez-vous point
575 * \param timeout The timeout used for the waiting the completion
576 * \param dst_buff The destination buffer to copy the received message
577 * \param src_buff_size The size of the destination buffer
578 * \param comm_ref The communication object used for the send (useful if someone else wants to cancel this communication afterward)
580 * - host_error if peer failed
581 * - timeout_error if communication reached the timeout specified
582 * - network_error if network failed or peer issued a timeout
584 XBT_INLINE void SIMIX_network_recv(smx_rdv_t rdv, double timeout,
585 void *dst_buff, size_t * dst_buff_size,
586 smx_comm_t * comm_ref)
589 smx_comm_t comm = *comm_ref =
590 (smx_comm_t) SIMIX_network_irecv(rdv, dst_buff, dst_buff_size);
592 SIMIX_network_wait(comm, timeout);
595 SIMIX_communication_destroy(comm);
602 /******************************************************************************/
603 /* Asynchronous Communication */
604 /******************************************************************************/
605 smx_comm_t SIMIX_network_isend(smx_rdv_t rdv, double task_size,
606 double rate, void *src_buff,
607 size_t src_buff_size, void *data)
611 /*If running in model-checking mode then intercept the communication action */
613 if (_surf_do_model_check)
614 MC_trans_intercept_isend(rdv);
616 /* Look for communication request matching our needs.
617 If it is not found then create it and push it into the rendez-vous point */
618 comm = SIMIX_rdv_get_request(rdv, comm_recv);
621 comm = SIMIX_communication_new(comm_send);
622 SIMIX_rdv_push(rdv, comm);
625 /* Setup the communication request */
626 comm->src_proc = SIMIX_process_self();
627 comm->task_size = task_size;
629 comm->src_buff = src_buff;
630 comm->src_buff_size = src_buff_size;
633 SIMIX_communication_start(comm);
637 smx_comm_t SIMIX_network_irecv(smx_rdv_t rdv, void *dst_buff,
638 size_t * dst_buff_size)
642 /*If running in model-checking mode then intercept the communication action */
644 if (_surf_do_model_check)
645 MC_trans_intercept_irecv(rdv);
647 /* Look for communication request matching our needs.
648 * If it is not found then create it and push it into the rendez-vous point
650 comm = SIMIX_rdv_get_request(rdv, comm_send);
653 comm = SIMIX_communication_new(comm_recv);
654 SIMIX_rdv_push(rdv, comm);
657 /* Setup communication request */
658 comm->dst_proc = SIMIX_process_self();
659 comm->dst_buff = dst_buff;
660 comm->dst_buff_size = dst_buff_size;
662 SIMIX_communication_start(comm);
666 /** @brief blocks until the communication terminates or the timeout occurs */
667 XBT_INLINE void SIMIX_network_wait(smx_comm_t comm, double timeout)
669 /*If running in model-checking mode then intercept the communication action */
671 if (_surf_do_model_check)
672 MC_trans_intercept_wait(comm);
674 SIMIX_communication_use(comm);
675 /* Wait for communication completion */
676 SIMIX_communication_wait_for_completion(comm, timeout);
679 /** @Returns whether the (asynchronous) communication is done yet or not */
680 XBT_INLINE int SIMIX_network_test(smx_comm_t comm)
682 /*If running in model-checking mode then intercept the communication action */
684 if (_surf_do_model_check)
685 MC_trans_intercept_test(comm);
688 /* Copy data if the communication is done */
689 if (comm->sem && !SIMIX_sem_would_block(comm->sem)) {
690 /* Copy network data */
691 SIMIX_network_copy_data(comm);
697 /** @brief wait for the completion of any communication of a set
699 * @Returns the rank in the dynar of communication which finished; destroy it after identifying which one it is
701 unsigned int SIMIX_network_waitany(xbt_dynar_t comms)
703 xbt_dynar_t sems = xbt_dynar_new(sizeof(smx_sem_t), NULL);
704 unsigned int cursor, found_comm = -1;
705 smx_comm_t comm, comm_finished = NULL;
707 /*If running in model-checking mode then intercept the communication action */
709 if (_surf_do_model_check)
710 MC_trans_intercept_waitany(comms);
712 xbt_dynar_foreach(comms, cursor, comm)
713 xbt_dynar_push(sems, &(comm->sem));
715 DEBUG1("Waiting for the completion of communication set %p", comms);
717 found_comm = SIMIX_sem_acquire_any(sems);
718 xbt_dynar_free_container(&sems);
719 xbt_assert0(found_comm != -1,
720 "Cannot find which communication finished");
721 xbt_dynar_get_cpy(comms, found_comm, &comm_finished);
723 DEBUG2("Communication %p of communication set %p finished",
724 comm_finished, comms);
726 /* let the regular code deal with the communication end (errors checking and cleanup).
727 * A bit of useless work will be done, but that's good for source factorization */
728 SIMIX_sem_release_forever(comm_finished->sem);
729 SIMIX_communication_use(comm_finished);
730 SIMIX_communication_wait_for_completion(comm_finished, -1);