1 /* Copyright (c) 2007-2015. 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. */
9 #include "smx_private.h"
11 #include "xbt/sysdep.h"
14 #include "xbt/ex.h" /* ex_backtrace_display */
16 #include "mc/mc_replay.h"
17 #include "simgrid/sg_config.h"
20 #include "mc/mc_private.h"
21 #include "mc/mc_protocol.h"
22 #include "mc/mc_client.h"
24 #include "mc/mc_record.h"
27 #include "smpi/private.h"
30 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
31 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
32 "Logging specific to SIMIX (kernel)");
34 smx_global_t simix_global = NULL;
35 static xbt_heap_t simix_timers = NULL;
37 static void* SIMIX_synchro_mallocator_new_f(void);
38 static void SIMIX_synchro_mallocator_free_f(void* synchro);
39 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
41 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
44 int _sg_do_verbose_exit = 1;
45 static void _XBT_CALL inthandler(int ignored)
47 if ( _sg_do_verbose_exit ) {
48 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
49 SIMIX_display_process_status();
52 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
58 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
60 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
62 "Access violation detected.\n"
63 "This can result from a programming error in your code or, although less likely,\n"
64 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
65 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
67 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
68 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
69 smx_context_stack_size / 1024);
70 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
72 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
73 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
75 } else if (siginfo->si_signo == SIGSEGV) {
76 fprintf(stderr, "Segmentation fault.\n");
78 if (smpi_enabled() && !smpi_privatize_global_variables) {
80 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
87 MC_dump_stack_safety(mc_stack);
89 MC_print_statistics(mc_stats);
95 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
98 * Install signal handler for SIGSEGV. Check that nobody has already installed
99 * its own handler. For example, the Java VM does this.
101 static void install_segvhandler(void)
103 stack_t stack, old_stack;
104 stack.ss_sp = sigsegv_stack;
105 stack.ss_size = sizeof sigsegv_stack;
108 if (sigaltstack(&stack, &old_stack) == -1) {
109 XBT_WARN("Failed to register alternate signal stack: %s",
113 if (!(old_stack.ss_flags & SS_DISABLE)) {
114 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
115 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
116 sigaltstack(&old_stack, NULL);
119 struct sigaction action, old_action;
120 action.sa_sigaction = segvhandler;
121 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
122 sigemptyset(&action.sa_mask);
124 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
125 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
129 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
130 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
131 (old_action.sa_flags & SA_SIGINFO) ?
132 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
133 sigaction(SIGSEGV, &old_action, NULL);
138 /********************************* SIMIX **************************************/
140 XBT_INLINE double SIMIX_timer_next(void)
142 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
147 * \brief Initialize SIMIX internal data.
152 void SIMIX_global_init(int *argc, char **argv)
154 s_smx_process_t proc;
157 simix_global = xbt_new0(s_smx_global_t, 1);
159 #ifdef TIME_BENCH_AMDAHL
160 simix_global->timer_seq = xbt_os_timer_new();
161 simix_global->timer_par = xbt_os_timer_new();
162 xbt_os_cputimer_start(simix_global->timer_seq);
164 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
165 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
166 simix_global->process_list =
167 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
168 simix_global->process_to_destroy =
169 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
171 simix_global->maestro_process = NULL;
172 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
174 simix_global->create_process_function = SIMIX_process_create;
175 simix_global->kill_process_function = SIMIX_process_kill;
176 simix_global->cleanup_process_function = SIMIX_process_cleanup;
177 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
178 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
179 SIMIX_synchro_mallocator_reset_f);
180 simix_global->autorestart = SIMIX_host_restart_processes;
181 simix_global->mutex = xbt_os_mutex_init();
183 surf_init(argc, argv); /* Initialize SURF structures */
184 SIMIX_context_mod_init();
185 SIMIX_create_maestro_process();
187 /* context exception handlers */
188 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
189 __xbt_ex_terminate = SIMIX_process_exception_terminate;
191 SIMIX_network_init();
193 /* Prepare to display some more info when dying on Ctrl-C pressing */
194 signal(SIGINT, inthandler);
197 /* Install SEGV handler */
198 install_segvhandler();
200 /* register a function to be called by SURF after the environment creation */
202 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
206 simix_timers = xbt_heap_new(8, &free);
209 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
211 if (sg_cfg_get_boolean("clean_atexit"))
215 // The communication initialization is done ASAP.
216 // We need to communicate initialization of the different layers to the model-checker.
217 if (mc_mode == MC_MODE_NONE) {
218 if (getenv(MC_ENV_SOCKET_FD)) {
219 mc_mode = MC_MODE_CLIENT;
222 MC_client_handle_messages();
227 if (_sg_cfg_exit_asap)
233 * \brief Clean the SIMIX simulation
235 * This functions remove the memory used by SIMIX
237 void SIMIX_clean(void)
239 static int cleaned = 0;
240 #ifdef TIME_BENCH_PER_SR
241 smx_ctx_raw_new_sr();
243 if (cleaned) return; // to avoid double cleaning by java and C
245 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
246 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
248 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
249 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
250 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
252 /* Kill all processes (but maestro) */
253 SIMIX_process_killall(simix_global->maestro_process, 1);
255 /* Exit the SIMIX network module */
256 SIMIX_network_exit();
258 xbt_heap_free(simix_timers);
260 /* Free the remaining data structures */
261 xbt_dynar_free(&simix_global->process_to_run);
262 xbt_dynar_free(&simix_global->process_that_ran);
263 xbt_swag_free(simix_global->process_to_destroy);
264 xbt_swag_free(simix_global->process_list);
265 simix_global->process_list = NULL;
266 simix_global->process_to_destroy = NULL;
267 xbt_dict_free(&(simix_global->registered_functions));
269 xbt_os_mutex_destroy(simix_global->mutex);
270 simix_global->mutex = NULL;
272 /* Let's free maestro now */
273 SIMIX_context_free(simix_global->maestro_process->context);
274 xbt_free(simix_global->maestro_process->running_ctx);
275 xbt_free(simix_global->maestro_process);
276 simix_global->maestro_process = NULL;
278 /* Restore the default exception setup */
279 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
280 __xbt_ex_terminate = &__xbt_ex_terminate_default;
282 /* Finish context module and SURF */
283 SIMIX_context_mod_exit();
287 #ifdef TIME_BENCH_AMDAHL
288 xbt_os_cputimer_stop(simix_global->timer_seq);
289 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
290 xbt_os_timer_elapsed(simix_global->timer_seq),
291 xbt_os_timer_elapsed(simix_global->timer_par));
292 xbt_os_timer_free(simix_global->timer_seq);
293 xbt_os_timer_free(simix_global->timer_par);
296 xbt_mallocator_free(simix_global->synchro_mallocator);
297 xbt_free(simix_global);
306 * \brief A clock (in second).
308 * \return Return the clock.
310 XBT_INLINE double SIMIX_get_clock(void)
312 if(MC_is_active() || MC_record_replay_is_active()){
313 return MC_process_clock_get(SIMIX_process_self());
315 return surf_get_clock();
319 static int process_syscall_color(void *p)
321 switch ((*(smx_process_t *)p)->simcall.call) {
323 case SIMCALL_PROCESS_KILL:
325 case SIMCALL_PROCESS_RESUME:
334 * \brief Run the main simulation loop.
339 MC_record_replay_init();
340 MC_record_replay_from_string(MC_record_path);
345 smx_process_t process;
346 surf_action_t action;
352 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
353 xbt_dynar_length(simix_global->process_to_run));
354 #ifdef TIME_BENCH_PER_SR
355 smx_ctx_raw_new_sr();
357 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
358 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
359 xbt_dynar_length(simix_global->process_to_run));
361 /* Run all processes that are ready to run, possibly in parallel */
362 #ifdef TIME_BENCH_AMDAHL
363 xbt_os_cputimer_stop(simix_global->timer_seq);
364 xbt_os_cputimer_resume(simix_global->timer_par);
366 SIMIX_process_runall();
367 #ifdef TIME_BENCH_AMDAHL
368 xbt_os_cputimer_stop(simix_global->timer_par);
369 xbt_os_cputimer_resume(simix_global->timer_seq);
372 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
373 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
375 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
377 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
379 /* Here, the order is ok because:
381 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
383 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
385 * - if there is no kill during the simulation, processes remain sorted according by their PID.
386 * rational: This can be proved inductively.
387 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
388 * Let's show that it is still so at the end of this round.
389 * - if a process is added when being created, that's from maestro. It can be either at startup
390 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
391 * in arbitrary order (inductive hypothesis), we are fine.
392 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
393 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
394 * of a simcall terminates, we're still good. Proof:
395 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
396 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
397 * - For communications, this function is called from SIMIX_comm_finish().
398 * This function itself don't mess with the order since simcalls are handled in FIFO order.
399 * The function is called:
400 * - before the comm starts (invalid parameters, or resource already dead or whatever).
401 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
402 * - because the communication failed or were canceled after startup. In this case, it's called from the function
403 * we are in, by the chunk:
404 * set = model->states.failed_action_set;
405 * while ((synchro = xbt_swag_extract(set)))
406 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
407 * This order is also fixed because it depends of the order in which the surf actions were
408 * added to the system, and only maestro can add stuff this way, through simcalls.
409 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
410 * poped out of the swag does not depend on the user code's execution order.
411 * - because the communication terminated. In this case, synchros are served in the order given by
412 * set = model->states.done_action_set;
413 * while ((synchro = xbt_swag_extract(set)))
414 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
415 * and the argument is very similar to the previous one.
416 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
417 * So, in any cases, the orders of processes within process_to_run do not depend on the order in which user processes were executed previously.
418 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
419 * - If there is some process killings, the order is changed by this decision that comes from user-land
420 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
421 * So, even the order change induced by the process killing is perfectly reproducible.
423 * So science works, bitches [http://xkcd.com/54/].
425 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
426 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
427 * That would thus be a pure waste of time.
430 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
431 if (process->simcall.call != SIMCALL_NONE) {
432 SIMIX_simcall_handle(&process->simcall, 0);
435 /* Wake up all processes waiting for a Surf action to finish */
436 xbt_dynar_foreach(model_list, iter, model) {
437 XBT_DEBUG("Handling process whose action failed");
438 while ((action = surf_model_extract_failed_action_set(model))) {
439 XBT_DEBUG(" Handling Action %p",action);
440 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
442 XBT_DEBUG("Handling process whose action terminated normally");
443 while ((action = surf_model_extract_done_action_set(model))) {
444 XBT_DEBUG(" Handling Action %p",action);
445 if (surf_action_get_data(action) == NULL)
446 XBT_DEBUG("probably vcpu's action %p, skip", action);
448 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
453 time = SIMIX_timer_next();
454 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
455 XBT_DEBUG("Calling surf_solve");
456 time = surf_solve(time);
457 XBT_DEBUG("Moving time ahead : %g", time);
459 /* Notify all the hosts that have failed */
460 /* FIXME: iterate through the list of failed host and mark each of them */
461 /* as failed. On each host, signal all the running processes with host_fail */
463 /* Handle any pending timer */
464 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
465 //FIXME: make the timers being real callbacks
466 // (i.e. provide dispatchers that read and expand the args)
467 timer = xbt_heap_pop(simix_timers);
469 ((void (*)(void*))timer->func)(timer->args);
473 /* Wake up all processes waiting for a Surf action to finish */
474 xbt_dynar_foreach(model_list, iter, model) {
475 XBT_DEBUG("Handling process whose action failed");
476 while ((action = surf_model_extract_failed_action_set(model))) {
477 XBT_DEBUG(" Handling Action %p",action);
478 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
480 XBT_DEBUG("Handling process whose action terminated normally");
481 while ((action = surf_model_extract_done_action_set(model))) {
482 XBT_DEBUG(" Handling Action %p",action);
483 if (surf_action_get_data(action) == NULL)
484 XBT_DEBUG("probably vcpu's action %p, skip", action);
486 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
490 /* Autorestart all process */
491 if(host_that_restart) {
492 char *hostname = NULL;
493 xbt_dynar_foreach(host_that_restart,iter,hostname) {
494 XBT_INFO("Restart processes on host: %s",hostname);
495 SIMIX_host_autorestart(sg_host_by_name(hostname));
497 xbt_dynar_reset(host_that_restart);
500 /* Clean processes to destroy */
501 SIMIX_process_empty_trash();
504 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
506 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
508 if (xbt_swag_size(simix_global->process_list) != 0) {
512 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
513 SIMIX_display_process_status();
519 * \brief Set the date to execute a function
521 * Set the date to execute the function on the surf.
522 * \param date Date to execute function
523 * \param function Function to be executed
524 * \param arg Parameters of the function
527 XBT_INLINE smx_timer_t SIMIX_timer_set(double date, void *function, void *arg)
529 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
532 timer->func = function;
534 xbt_heap_push(simix_timers, timer, date);
537 /** @brief cancels a timer that was added earlier */
538 XBT_INLINE void SIMIX_timer_remove(smx_timer_t timer) {
539 xbt_heap_rm_elm(simix_timers, timer, timer->date);
542 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
543 XBT_INLINE double SIMIX_timer_get_date(smx_timer_t timer) {
544 return timer?timer->date:0;
548 * \brief Registers a function to create a process.
550 * This function registers a function to be called
551 * when a new process is created. The function has
552 * to call SIMIX_process_create().
553 * \param function create process function
555 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
558 simix_global->create_process_function = function;
562 * \brief Registers a function to kill a process.
564 * This function registers a function to be called when a
565 * process is killed. The function has to call the SIMIX_process_kill().
567 * \param function Kill process function
569 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
572 simix_global->kill_process_function = function;
576 * \brief Registers a function to cleanup a process.
578 * This function registers a user function to be called when
579 * a process ends properly.
581 * \param function cleanup process function
583 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
586 simix_global->cleanup_process_function = function;
590 void SIMIX_display_process_status(void)
592 if (simix_global->process_list == NULL) {
596 smx_process_t process = NULL;
597 int nbprocess = xbt_swag_size(simix_global->process_list);
599 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
600 /* List the process and their state */
602 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
603 xbt_swag_foreach(process, simix_global->process_list) {
605 if (process->waiting_synchro) {
607 const char* synchro_description = "unknown";
608 switch (process->waiting_synchro->type) {
610 case SIMIX_SYNC_EXECUTE:
611 synchro_description = "execution";
614 case SIMIX_SYNC_PARALLEL_EXECUTE:
615 synchro_description = "parallel execution";
618 case SIMIX_SYNC_COMMUNICATE:
619 synchro_description = "communication";
622 case SIMIX_SYNC_SLEEP:
623 synchro_description = "sleeping";
626 case SIMIX_SYNC_JOIN:
627 synchro_description = "joining";
630 case SIMIX_SYNC_SYNCHRO:
631 synchro_description = "synchronization";
635 synchro_description = "I/O";
638 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
639 process->pid, process->name, sg_host_name(process->host),
640 synchro_description, process->waiting_synchro,
641 process->waiting_synchro->name, (int)process->waiting_synchro->state);
644 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->host));
649 static void* SIMIX_synchro_mallocator_new_f(void) {
650 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
651 synchro->simcalls = xbt_fifo_new();
655 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
656 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
660 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
662 // we also recycle the simcall list
663 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
664 xbt_fifo_reset(fifo);
665 memset(synchro, 0, sizeof(s_smx_synchro_t));
666 ((smx_synchro_t) synchro)->simcalls = fifo;
669 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
670 return SIMIX_asr_get_properties(name);
672 xbt_dict_t SIMIX_asr_get_properties(const char *name)
674 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);