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 /* Pimple to get an histogram of message sizes in the simulation */
13 xbt_dict_t msg_sizes = NULL;
14 #ifdef HAVE_LATENCY_BOUND_TRACKING
15 xbt_dict_t latency_limited_dict = NULL;
18 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_network, simix,
19 "Logging specific to SIMIX (network)");
21 /******************************************************************************/
22 /* Rendez-Vous Points */
23 /******************************************************************************/
26 * \brief Creates a new rendez-vous point
27 * \param name The name of the rendez-vous point
28 * \return The created rendez-vous point
30 smx_rdv_t SIMIX_rdv_create(const char *name)
32 smx_rdv_t rdv = xbt_new0(s_smx_rvpoint_t, 1);
33 rdv->name = name ? xbt_strdup(name) : NULL;
34 rdv->read = SIMIX_mutex_init();
35 rdv->write = SIMIX_mutex_init();
36 rdv->comm_fifo = xbt_fifo_new();
42 * \brief Destroy a rendez-vous point
43 * \param name The rendez-vous point to destroy
45 void SIMIX_rdv_destroy(smx_rdv_t rdv)
49 SIMIX_mutex_destroy(rdv->read);
50 SIMIX_mutex_destroy(rdv->write);
51 xbt_fifo_free(rdv->comm_fifo);
56 * \brief Push a communication request into a rendez-vous point
57 * \param rdv The rendez-vous point
58 * \param comm The communication request
60 static XBT_INLINE void SIMIX_rdv_push(smx_rdv_t rdv, smx_comm_t comm)
62 xbt_fifo_push(rdv->comm_fifo, comm);
67 * \brief Remove a communication request from a rendez-vous point
68 * \param rdv The rendez-vous point
69 * \param comm The communication request
71 static XBT_INLINE void SIMIX_rdv_remove(smx_rdv_t rdv, smx_comm_t comm)
73 xbt_fifo_remove(rdv->comm_fifo, comm);
78 * \brief Checks if there is a communication request queued in a rendez-vous matching our needs
79 * \param type The type of communication we are looking for (comm_send, comm_recv)
80 * \return The communication request if found, NULL otherwise.
82 smx_comm_t SIMIX_rdv_get_request(smx_rdv_t rdv, smx_comm_type_t type) {
83 smx_comm_t comm = (smx_comm_t)xbt_fifo_get_item_content(
84 xbt_fifo_get_first_item(rdv->comm_fifo));
86 if(comm && comm->type == type){
87 DEBUG0("Communication request found!");
88 xbt_fifo_shift(rdv->comm_fifo);
89 SIMIX_communication_use(comm);
94 DEBUG0("Communication request not found");
99 * \brief counts the number of communication requests of a given host pending
100 * on a rendez-vous point
101 * \param rdv The rendez-vous point
102 * \param host The host to be counted
103 * \return The number of comm request pending in the rdv
106 SIMIX_rdv_get_count_waiting_comm(smx_rdv_t rdv, smx_host_t host)
108 smx_comm_t comm = NULL;
109 xbt_fifo_item_t item = NULL;
112 xbt_fifo_foreach(rdv->comm_fifo, item, comm, smx_comm_t) {
113 if (comm->src_proc->smx_host == host)
121 * \brief returns the communication at the head of the rendez-vous
122 * \param rdv The rendez-vous point
123 * \return The communication or NULL if empty
125 XBT_INLINE smx_comm_t SIMIX_rdv_get_head(smx_rdv_t rdv)
127 return (smx_comm_t)xbt_fifo_get_item_content(xbt_fifo_get_first_item(rdv->comm_fifo));
130 /** @brief adds some API-related data to the rendez-vous point */
131 XBT_INLINE void SIMIX_rdv_set_data(smx_rdv_t rdv,void *data) {
134 /** @brief gets API-related data from the rendez-vous point */
135 XBT_INLINE void *SIMIX_rdv_get_data(smx_rdv_t rdv) {
139 /******************************************************************************/
140 /* Communication Requests */
141 /******************************************************************************/
144 * \brief Creates a new communication request
145 * \param type The type of communication (comm_send, comm_recv)
146 * \return The new communication request
148 smx_comm_t SIMIX_communication_new(smx_comm_type_t type)
150 /* alloc structures */
151 smx_comm_t comm = xbt_new0(s_smx_comm_t, 1);
153 comm->sem = SIMIX_sem_init(0);
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,comm->refcount);
167 #ifdef HAVE_LATENCY_BOUND_TRACKING
168 //save is latency limited flag to use afterwards
169 if (latency_limited_dict == NULL) {
170 latency_limited_dict = xbt_dict_new();
173 DEBUG2("adding key %p with latency limited value %d to the dict", comm, SIMIX_action_is_latency_bounded(comm->act));
174 xbt_dicti_set(latency_limited_dict, (uintptr_t)comm, SIMIX_action_is_latency_bounded(comm->act));
179 if(comm->refcount > 0)
183 SIMIX_sem_destroy(comm->sem);
188 SIMIX_action_destroy(comm->act);
192 if(comm->src_timeout){
193 SIMIX_action_destroy(comm->src_timeout);
194 comm->src_timeout = NULL;
197 if(comm->dst_timeout){
198 SIMIX_action_destroy(comm->dst_timeout);
199 comm->dst_timeout = NULL;
208 * \brief Increase the number of users of the communication.
209 * \param comm The communication request
210 * Each communication request can be used by more than one process, so it is
211 * necessary to know number of them at destroy time, to avoid freeing stuff that
212 * maybe is in use by others.
215 static XBT_INLINE void SIMIX_communication_use(smx_comm_t comm)
221 * \brief Start the simulation of a communication request
222 * \param comm The communication request
224 static XBT_INLINE void SIMIX_communication_start(smx_comm_t comm)
226 /* If both the sender and the receiver are already there, start the communication */
227 if(comm->src_proc && comm->dst_proc){
228 DEBUG1("Starting communication %p", comm);
229 comm->act = SIMIX_action_communicate(comm->src_proc->smx_host,
230 comm->dst_proc->smx_host, NULL,
231 comm->task_size, comm->rate);
233 TRACE_smx_action_communicate (comm->act, comm->src_proc);
236 /* If any of the process is suspend, create the action but stop its execution,
237 it will be restarted when the sender process resume */
238 if(SIMIX_process_is_suspended(comm->src_proc) ||
239 SIMIX_process_is_suspended(comm->dst_proc)) {
240 SIMIX_action_suspend(comm->act);
243 /* Add the communication as user data of the action */
244 comm->act->data = comm;
246 /* The semaphore will only get signaled once, but since the first unlocked guy will
247 * release_forever() the semaphore, that will unlock the second (and any other)
248 * communication partner */
249 SIMIX_register_action_to_semaphore(comm->act, comm->sem);
254 * \brief Performs error checking and cleanup
255 * \param comm The communication
257 static XBT_INLINE void SIMIX_communication_cleanup(smx_comm_t comm)
259 DEBUG1("Checking errors and cleaning communication %p", comm);
261 /* Make sure that everyone sleeping on that semaphore is awake, and that nobody will ever block on it */
262 SIMIX_sem_release_forever(comm->sem);
264 /* Check for errors other than timeouts */
265 if (!SIMIX_host_get_state(SIMIX_host_self())){
267 SIMIX_rdv_remove(comm->rdv, comm);
268 SIMIX_communication_destroy(comm);
269 THROW0(host_error, 0, "Host failed");
270 } else if (SIMIX_action_get_state(comm->act) == SURF_ACTION_FAILED){
271 SIMIX_communication_destroy(comm);
272 THROW0(network_error, 0, "Link failure");
273 } else if (!SIMIX_host_get_state(SIMIX_process_get_host(comm->dst_proc)) ||
274 !SIMIX_host_get_state(SIMIX_process_get_host(comm->src_proc))) {
275 /* We test both src&dst because we dunno who we are today, and we already tested myself above.
276 * So, at the end, we test the remote peer only
277 * Moreover, we have to test it because if the remote peer fails, the action comm->act is not done nor failed.
278 * In that case, we got awaken by the little endless actions created in the SIMIX_sem_acquire(comm->sem)
279 * at the beginning of this function. */
280 SIMIX_communication_destroy(comm);
281 THROW0(network_error, 0, "Remote peer failed");
284 /* Copy network data */
285 SIMIX_network_copy_data(comm);
287 SIMIX_communication_destroy(comm);
291 * \brief Waits for communication completion
292 * \param comm The communication
293 * \param timeout The max amount of time to wait for the communication to finish
296 * - host_error if local peer failed
297 * - timeout_error if communication reached the timeout specified (either because of local peer or remote peer)
298 * - network_error if network failed or remote peer failed
300 static XBT_INLINE void SIMIX_communication_wait_for_completion(smx_comm_t comm, double timeout)
302 smx_action_t act_sleep = NULL;
306 DEBUG1("Waiting for the completion of communication %p", comm);
309 act_sleep = SIMIX_action_sleep(SIMIX_host_self(), timeout);
310 if(SIMIX_process_self()==comm->src_proc)
311 comm->src_timeout = act_sleep;
313 comm->dst_timeout = act_sleep;
314 SIMIX_action_set_name(act_sleep,bprintf("Timeout for comm %p and wait on semaphore %p (max_duration:%f)", comm, comm->sem,timeout));
315 SIMIX_register_action_to_semaphore(act_sleep, comm->sem);
316 SIMIX_process_self()->waiting_action = act_sleep;
317 SIMIX_sem_block_onto(comm->sem);
318 SIMIX_process_self()->waiting_action = NULL;
319 SIMIX_unregister_action_to_semaphore(act_sleep, comm->sem);
321 SIMIX_sem_acquire(comm->sem);
324 /* Check for timeouts */
325 if ((src_timeout = ((comm->src_timeout) && (SIMIX_action_get_state(comm->src_timeout) == SURF_ACTION_DONE))) ||
326 (dst_timeout = ((comm->dst_timeout) && (SIMIX_action_get_state(comm->dst_timeout) == SURF_ACTION_DONE))) ) {
327 /* Somebody did a timeout! */
328 if (src_timeout) DEBUG1("Communication timeout from the src! %p", comm);
329 if (dst_timeout) DEBUG1("Communication timeout from the dst! %p", comm);
331 if(comm->act && SIMIX_action_get_state(comm->act) == SURF_ACTION_RUNNING)
332 SIMIX_communication_cancel(comm);
334 SIMIX_rdv_remove(comm->rdv, comm);
336 /* Make sure that everyone sleeping on that semaphore is awake, and that nobody will ever block on it */
337 SIMIX_sem_release_forever(comm->sem);
338 SIMIX_communication_destroy(comm);
340 THROW1(timeout_error, 0, "Communication timeouted because of %s",src_timeout?"the source":"the destination");
343 DEBUG1("Communication %p complete!", comm);
344 SIMIX_communication_cleanup(comm);
348 * \brief Cancels a communication
349 * \brief comm The communication to cancel
351 XBT_INLINE void SIMIX_communication_cancel(smx_comm_t comm)
354 SIMIX_action_cancel(comm->act);
358 * \brief get the amount remaining from the communication
359 * \param comm The communication
361 XBT_INLINE double SIMIX_communication_get_remains(smx_comm_t comm)
363 DEBUG1("calling SIMIX_action_get_remains(%p)", comm->act);
364 return SIMIX_action_get_remains(comm->act);
367 #ifdef HAVE_LATENCY_BOUND_TRACKING
369 * \brief verify if communication is latency bounded
370 * \param comm The communication
372 XBT_INLINE int SIMIX_communication_is_latency_bounded(smx_comm_t comm)
374 //try to find comm on the list of finished flows
377 xbt_dict_cursor_t cursor;
378 xbt_dict_foreach(latency_limited_dict,cursor,key,data) {
379 DEBUG2("comparing key=%p with comm=%p", (void*)key, (void*)comm);
380 if((void*)comm == (void*)key){
381 DEBUG2("key %p found, return value latency limited value %d", (void*)key, (int)data);
386 DEBUG1("calling SIMIX_action_is_latency_bounded(%p)", comm->act);
387 return SIMIX_action_is_latency_bounded(comm->act);
391 /******************************************************************************/
392 /* SIMIX_network_copy_data callbacks */
393 /******************************************************************************/
394 static void (*SIMIX_network_copy_data_callback)(smx_comm_t, size_t) = &SIMIX_network_copy_pointer_callback;
396 void SIMIX_network_set_copy_data_callback(void (*callback)(smx_comm_t, size_t)) {
397 SIMIX_network_copy_data_callback = callback;
400 void SIMIX_network_copy_pointer_callback(smx_comm_t comm, size_t buff_size) {
401 xbt_assert1((buff_size == sizeof(void*)), "Cannot copy %zu bytes: must be sizeof(void*)",buff_size);
402 *(void**)(comm->dst_buff) = comm->src_buff;
405 void SIMIX_network_copy_buffer_callback(smx_comm_t comm, size_t buff_size) {
406 memcpy(comm->dst_buff, comm->src_buff, buff_size);
410 * \brief Copy the communication data from the sender's buffer to the receiver's one
411 * \param comm The communication
413 void SIMIX_network_copy_data(smx_comm_t comm)
415 size_t buff_size = comm->src_buff_size;
416 uintptr_t casted_size = 0;
417 uintptr_t amount = 0;
418 /* If there is no data to be copy then return */
419 if(!comm->src_buff || !comm->dst_buff || comm->copied == 1)
422 DEBUG6("Copying comm %p data from %s (%p) -> %s (%p) (%zu bytes)",
424 comm->src_proc->smx_host->name, comm->src_buff,
425 comm->dst_proc->smx_host->name, comm->dst_buff,
428 /* Copy at most dst_buff_size bytes of the message to receiver's buffer */
429 if (comm->dst_buff_size)
430 buff_size = MIN(buff_size,*(comm->dst_buff_size));
432 /* Update the receiver's buffer size to the copied amount */
433 if (comm->dst_buff_size)
434 *comm->dst_buff_size = buff_size;
438 (*SIMIX_network_copy_data_callback)(comm, buff_size);
440 /* Set the copied flag so we copy data only once */
441 /* (this function might be called from both communication ends)*/
444 /* pimple to display the message sizes */
446 if (msg_sizes == NULL)
447 msg_sizes = xbt_dict_new();
448 casted_size = comm->task_size;
449 amount = xbt_dicti_get(msg_sizes, casted_size);
452 xbt_dicti_set(msg_sizes,casted_size, amount);
456 /* pimple to display the message sizes */
457 void SIMIX_message_sizes_output(const char *filename) {
460 xbt_dict_cursor_t cursor;
462 out = fopen(filename,"w");
463 xbt_assert1(out,"Cannot open file %s",filename);
465 xbt_dict_foreach(msg_sizes,cursor,key,data) {
466 fprintf(out,"%zu %zu\n",key,data);
472 * \brief Return the user data associated to the communication
473 * \param comm The communication
474 * \return the user data
476 XBT_INLINE void *SIMIX_communication_get_data(smx_comm_t comm)
481 XBT_PUBLIC(void *) SIMIX_communication_get_src_buf(smx_comm_t comm)
483 return comm->src_buff;
485 XBT_PUBLIC(void *) SIMIX_communication_get_dst_buf(smx_comm_t comm)
487 return comm->dst_buff;
489 XBT_PUBLIC(size_t) SIMIX_communication_get_src_buf_size(smx_comm_t comm)
491 return comm->src_buff_size;
493 XBT_PUBLIC(size_t) SIMIX_communication_get_dst_buf_size(smx_comm_t comm)
495 return *(comm->dst_buff_size);
498 /******************************************************************************/
499 /* Synchronous Communication */
500 /******************************************************************************/
502 * \brief Put a send communication request in a rendez-vous point and waits for
503 * its completion (blocking)
504 * \param rdv The rendez-vous point
505 * \param task_size The size of the communication action (for surf simulation)
506 * \param rate The rate of the communication action (for surf)
507 * \param timeout The timeout used for the waiting the completion
508 * \param src_buff The source buffer containing the message to be sent
509 * \param src_buff_size The size of the source buffer
510 * \param comm_ref The communication object used for the send (useful if someone else wants to cancel this communication afterward)
511 * \param data User data associated to the communication object
513 * - host_error if peer failed
514 * - timeout_error if communication reached the timeout specified
515 * - network_error if network failed or peer issued a timeout
517 XBT_INLINE void SIMIX_network_send(smx_rdv_t rdv, double task_size, double rate,
518 double timeout, void *src_buff, size_t src_buff_size,
519 smx_comm_t *comm_ref, void *data)
521 *comm_ref = SIMIX_network_isend(rdv,task_size,rate,src_buff,src_buff_size,data);
522 SIMIX_network_wait(*comm_ref,timeout);
526 * \brief Put a receive communication request in a rendez-vous point and waits
527 * for its completion (blocking)
528 * \param rdv The rendez-vous point
529 * \param timeout The timeout used for the waiting the completion
530 * \param dst_buff The destination buffer to copy the received message
531 * \param src_buff_size The size of the destination buffer
532 * \param comm_ref The communication object used for the send (useful if someone else wants to cancel this communication afterward)
534 * - host_error if peer failed
535 * - timeout_error if communication reached the timeout specified
536 * - network_error if network failed or peer issued a timeout
538 XBT_INLINE void SIMIX_network_recv(smx_rdv_t rdv, double timeout, void *dst_buff,
539 size_t *dst_buff_size, smx_comm_t *comm_ref)
541 *comm_ref = (smx_comm_t) SIMIX_network_irecv(rdv,dst_buff,dst_buff_size);
542 SIMIX_network_wait(*comm_ref,timeout);
545 /******************************************************************************/
546 /* Asynchronous Communication */
547 /******************************************************************************/
548 smx_comm_t SIMIX_network_isend(smx_rdv_t rdv, double task_size, double rate,
549 void *src_buff, size_t src_buff_size, void *data)
553 /*If running in model-checking mode then intercept the communication action*/
555 if (_surf_do_model_check)
556 MC_trans_intercept_isend(rdv);
558 /* Look for communication request matching our needs.
559 If it is not found then create it and push it into the rendez-vous point */
560 comm = SIMIX_rdv_get_request(rdv, comm_recv);
563 comm = SIMIX_communication_new(comm_send);
564 SIMIX_rdv_push(rdv, comm);
567 /* Setup the communication request */
568 comm->src_proc = SIMIX_process_self();
569 comm->task_size = task_size;
571 comm->src_buff = src_buff;
572 comm->src_buff_size = src_buff_size;
575 SIMIX_communication_start(comm);
579 smx_comm_t SIMIX_network_irecv(smx_rdv_t rdv, void *dst_buff, size_t *dst_buff_size)
583 /*If running in model-checking mode then intercept the communication action*/
585 if (_surf_do_model_check)
586 MC_trans_intercept_irecv(rdv);
588 /* Look for communication request matching our needs.
589 * If it is not found then create it and push it into the rendez-vous point
591 comm = SIMIX_rdv_get_request(rdv, comm_send);
594 comm = SIMIX_communication_new(comm_recv);
595 SIMIX_rdv_push(rdv, comm);
598 /* Setup communication request */
599 comm->dst_proc = SIMIX_process_self();
600 comm->dst_buff = dst_buff;
601 comm->dst_buff_size = dst_buff_size;
603 SIMIX_communication_start(comm);
607 /** @brief blocks until the communication terminates or the timeout occurs */
608 XBT_INLINE void SIMIX_network_wait(smx_comm_t comm, double timeout)
610 /*If running in model-checking mode then intercept the communication action*/
612 if (_surf_do_model_check)
613 MC_trans_intercept_wait(comm);
615 /* Wait for communication completion */
616 SIMIX_communication_wait_for_completion(comm, timeout);
619 /** @Returns whether the (asynchronous) communication is done yet or not */
620 XBT_INLINE int SIMIX_network_test(smx_comm_t comm)
622 /*If running in model-checking mode then intercept the communication action*/
624 if (_surf_do_model_check)
625 MC_trans_intercept_test(comm);
628 /* Copy data if the communication is done */
629 if(comm->sem && !SIMIX_sem_would_block(comm->sem)){
630 /* Copy network data */
631 SIMIX_network_copy_data(comm);
637 /** @brief wait for the completion of any communication of a set
639 * @Returns the rank in the dynar of communication which finished; destroy it after identifying which one it is
641 unsigned int SIMIX_network_waitany(xbt_dynar_t comms)
643 xbt_dynar_t sems = xbt_dynar_new(sizeof(smx_sem_t),NULL);
644 unsigned int cursor, found_comm=-1;
645 smx_comm_t comm,comm_finished=NULL;
647 /*If running in model-checking mode then intercept the communication action*/
649 if (_surf_do_model_check)
650 MC_trans_intercept_waitany(comms);
652 xbt_dynar_foreach(comms,cursor,comm)
653 xbt_dynar_push(sems,&(comm->sem));
655 DEBUG1("Waiting for the completion of communication set %p", comms);
657 found_comm = SIMIX_sem_acquire_any(sems);
658 xbt_dynar_free_container(&sems);
659 xbt_assert0(found_comm!=-1,"Cannot find which communication finished");
660 xbt_dynar_get_cpy(comms,found_comm,&comm_finished);
662 DEBUG2("Communication %p of communication set %p finished", comm_finished, comms);
664 /* let the regular code deal with the communication end (errors checking and cleanup).
665 * A bit of useless work will be done, but that's good for source factorization */
666 SIMIX_sem_release_forever(comm_finished->sem);
667 SIMIX_communication_wait_for_completion(comm_finished, -1);