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. */
8 #include "src/portable.h"
9 #ifdef HAVE_SYS_PTRACE_H
10 # include <sys/types.h>
11 # include <sys/ptrace.h>
14 #include "src/surf/surf_interface.hpp"
15 #include "smx_private.h"
16 #include "smx_private.hpp"
18 #include "xbt/sysdep.h"
21 #include "xbt/ex.h" /* ex_backtrace_display */
23 #include "src/mc/mc_replay.h"
24 #include "simgrid/sg_config.h"
26 #include "src/surf/callbacks.h"
29 #include "src/mc/mc_private.h"
30 #include "src/mc/mc_protocol.h"
31 #include "src/mc/mc_client.h"
36 #include "src/mc/mc_protocol.h"
39 #include "src/mc/mc_record.h"
40 #include "src/surf/platform.hpp"
43 #include "src/smpi/private.h"
46 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
47 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
48 "Logging specific to SIMIX (kernel)");
50 smx_global_t simix_global = NULL;
51 static xbt_heap_t simix_timers = NULL;
53 /** @brief Timer datatype */
54 typedef struct s_smx_timer {
60 void (*SMPI_switch_data_segment)(int) = NULL;
62 static void* SIMIX_synchro_mallocator_new_f(void);
63 static void SIMIX_synchro_mallocator_free_f(void* synchro);
64 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
66 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
69 int _sg_do_verbose_exit = 1;
70 static void _XBT_CALL inthandler(int ignored)
72 if ( _sg_do_verbose_exit ) {
73 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
74 SIMIX_display_process_status();
77 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
83 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
85 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
87 "Access violation detected.\n"
88 "This can result from a programming error in your code or, although less likely,\n"
89 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
90 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
92 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
93 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
94 smx_context_stack_size / 1024);
95 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
97 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
98 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
100 } else if (siginfo->si_signo == SIGSEGV) {
101 fprintf(stderr, "Segmentation fault.\n");
103 if (smpi_enabled() && !smpi_privatize_global_variables) {
105 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
112 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
115 * Install signal handler for SIGSEGV. Check that nobody has already installed
116 * its own handler. For example, the Java VM does this.
118 static void install_segvhandler(void)
120 stack_t stack, old_stack;
121 stack.ss_sp = sigsegv_stack;
122 stack.ss_size = sizeof sigsegv_stack;
125 if (sigaltstack(&stack, &old_stack) == -1) {
126 XBT_WARN("Failed to register alternate signal stack: %s",
130 if (!(old_stack.ss_flags & SS_DISABLE)) {
131 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
132 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
133 sigaltstack(&old_stack, NULL);
136 struct sigaction action, old_action;
137 action.sa_sigaction = segvhandler;
138 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
139 sigemptyset(&action.sa_mask);
141 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
142 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
146 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
147 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
148 (old_action.sa_flags & SA_SIGINFO) ?
149 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
150 sigaction(SIGSEGV, &old_action, NULL);
155 /********************************* SIMIX **************************************/
157 double SIMIX_timer_next(void)
159 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
162 static void kill_process(smx_process_t process)
164 SIMIX_process_kill(process, NULL);
167 static void SIMIX_storage_create_(smx_storage_t storage)
169 const char* key = xbt_dict_get_elm_key(storage);
170 SIMIX_storage_create(key, storage, NULL);
175 * \brief Initialize SIMIX internal data.
180 void SIMIX_global_init(int *argc, char **argv)
183 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
186 s_smx_process_t proc;
189 simix_global = xbt_new0(s_smx_global_t, 1);
191 #ifdef TIME_BENCH_AMDAHL
192 simix_global->timer_seq = xbt_os_timer_new();
193 simix_global->timer_par = xbt_os_timer_new();
194 xbt_os_cputimer_start(simix_global->timer_seq);
196 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
197 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
198 simix_global->process_list =
199 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
200 simix_global->process_to_destroy =
201 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
203 simix_global->maestro_process = NULL;
204 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
206 simix_global->create_process_function = SIMIX_process_create;
207 simix_global->kill_process_function = kill_process;
208 simix_global->cleanup_process_function = SIMIX_process_cleanup;
209 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
210 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
211 SIMIX_synchro_mallocator_reset_f);
212 simix_global->autorestart = SIMIX_host_restart_processes;
213 simix_global->mutex = xbt_os_mutex_init();
215 surf_init(argc, argv); /* Initialize SURF structures */
216 SIMIX_context_mod_init();
217 SIMIX_create_maestro_process();
219 /* context exception handlers */
220 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
221 __xbt_ex_terminate = SIMIX_process_exception_terminate;
223 SIMIX_network_init();
225 /* Prepare to display some more info when dying on Ctrl-C pressing */
226 signal(SIGINT, inthandler);
229 /* Install SEGV handler */
230 install_segvhandler();
232 /* register a function to be called by SURF after the environment creation */
234 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
235 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
236 SIMIX_host_create(&host);
238 surf_on_storage_created(SIMIX_storage_create_);
242 simix_timers = xbt_heap_new(8, &free);
245 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
247 if (sg_cfg_get_boolean("clean_atexit"))
251 // The communication initialization is done ASAP.
252 // We need to communicate initialization of the different layers to the model-checker.
256 if (_sg_cfg_exit_asap)
263 * \brief Clean the SIMIX simulation
265 * This functions remove the memory used by SIMIX
267 void SIMIX_clean(void)
269 #ifdef TIME_BENCH_PER_SR
270 smx_ctx_raw_new_sr();
272 if (smx_cleaned) return; // to avoid double cleaning by java and C
274 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
275 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
277 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
278 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
279 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
281 /* Kill all processes (but maestro) */
282 SIMIX_process_killall(simix_global->maestro_process, 1);
284 /* Exit the SIMIX network module */
285 SIMIX_network_exit();
287 xbt_heap_free(simix_timers);
289 /* Free the remaining data structures */
290 xbt_dynar_free(&simix_global->process_to_run);
291 xbt_dynar_free(&simix_global->process_that_ran);
292 xbt_swag_free(simix_global->process_to_destroy);
293 xbt_swag_free(simix_global->process_list);
294 simix_global->process_list = NULL;
295 simix_global->process_to_destroy = NULL;
296 xbt_dict_free(&(simix_global->registered_functions));
298 xbt_os_mutex_destroy(simix_global->mutex);
299 simix_global->mutex = NULL;
301 /* Let's free maestro now */
302 SIMIX_context_free(simix_global->maestro_process->context);
303 xbt_free(simix_global->maestro_process->running_ctx);
304 xbt_free(simix_global->maestro_process);
305 simix_global->maestro_process = NULL;
307 /* Restore the default exception setup */
308 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
309 __xbt_ex_terminate = &__xbt_ex_terminate_default;
311 /* Finish context module and SURF */
312 SIMIX_context_mod_exit();
316 #ifdef TIME_BENCH_AMDAHL
317 xbt_os_cputimer_stop(simix_global->timer_seq);
318 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
319 xbt_os_timer_elapsed(simix_global->timer_seq),
320 xbt_os_timer_elapsed(simix_global->timer_par));
321 xbt_os_timer_free(simix_global->timer_seq);
322 xbt_os_timer_free(simix_global->timer_par);
325 xbt_mallocator_free(simix_global->synchro_mallocator);
326 xbt_free(simix_global);
335 * \brief A clock (in second).
337 * \return Return the clock.
339 double SIMIX_get_clock(void)
341 if(MC_is_active() || MC_record_replay_is_active()){
342 return MC_process_clock_get(SIMIX_process_self());
344 return surf_get_clock();
348 static int process_syscall_color(void *p)
350 switch ((*(smx_process_t *)p)->simcall.call) {
352 case SIMCALL_PROCESS_KILL:
354 case SIMCALL_PROCESS_RESUME:
363 * \brief Run the main simulation loop.
368 MC_record_replay_init();
369 MC_record_replay_from_string(MC_record_path);
374 smx_process_t process;
375 surf_action_t action;
381 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
382 xbt_dynar_length(simix_global->process_to_run));
383 #ifdef TIME_BENCH_PER_SR
384 smx_ctx_raw_new_sr();
386 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
387 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
388 xbt_dynar_length(simix_global->process_to_run));
390 /* Run all processes that are ready to run, possibly in parallel */
391 #ifdef TIME_BENCH_AMDAHL
392 xbt_os_cputimer_stop(simix_global->timer_seq);
393 xbt_os_cputimer_resume(simix_global->timer_par);
395 SIMIX_process_runall();
396 #ifdef TIME_BENCH_AMDAHL
397 xbt_os_cputimer_stop(simix_global->timer_par);
398 xbt_os_cputimer_resume(simix_global->timer_seq);
401 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
402 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
404 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
406 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
408 /* Here, the order is ok because:
410 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
412 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
414 * - if there is no kill during the simulation, processes remain sorted according by their PID.
415 * rational: This can be proved inductively.
416 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
417 * Let's show that it is still so at the end of this round.
418 * - if a process is added when being created, that's from maestro. It can be either at startup
419 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
420 * in arbitrary order (inductive hypothesis), we are fine.
421 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
422 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
423 * of a simcall terminates, we're still good. Proof:
424 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
425 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
426 * - For communications, this function is called from SIMIX_comm_finish().
427 * This function itself don't mess with the order since simcalls are handled in FIFO order.
428 * The function is called:
429 * - before the comm starts (invalid parameters, or resource already dead or whatever).
430 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
431 * - because the communication failed or were canceled after startup. In this case, it's called from the function
432 * we are in, by the chunk:
433 * set = model->states.failed_action_set;
434 * while ((synchro = xbt_swag_extract(set)))
435 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
436 * This order is also fixed because it depends of the order in which the surf actions were
437 * added to the system, and only maestro can add stuff this way, through simcalls.
438 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
439 * poped out of the swag does not depend on the user code's execution order.
440 * - because the communication terminated. In this case, synchros are served in the order given by
441 * set = model->states.done_action_set;
442 * while ((synchro = xbt_swag_extract(set)))
443 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
444 * and the argument is very similar to the previous one.
445 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
446 * 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.
447 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
448 * - If there is some process killings, the order is changed by this decision that comes from user-land
449 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
450 * So, even the order change induced by the process killing is perfectly reproducible.
452 * So science works, bitches [http://xkcd.com/54/].
454 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
455 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
456 * That would thus be a pure waste of time.
459 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
460 if (process->simcall.call != SIMCALL_NONE) {
461 SIMIX_simcall_handle(&process->simcall, 0);
464 /* Wake up all processes waiting for a Surf action to finish */
465 xbt_dynar_foreach(all_existing_models, iter, model) {
466 XBT_DEBUG("Handling process whose action failed");
467 while ((action = surf_model_extract_failed_action_set(model))) {
468 XBT_DEBUG(" Handling Action %p",action);
469 SIMIX_simcall_exit((smx_synchro_t) action->getData());
471 XBT_DEBUG("Handling process whose action terminated normally");
472 while ((action = surf_model_extract_done_action_set(model))) {
473 XBT_DEBUG(" Handling Action %p",action);
474 if (action->getData() == NULL)
475 XBT_DEBUG("probably vcpu's action %p, skip", action);
477 SIMIX_simcall_exit((smx_synchro_t) action->getData());
482 time = SIMIX_timer_next();
483 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
484 XBT_DEBUG("Calling surf_solve");
485 time = surf_solve(time);
486 XBT_DEBUG("Moving time ahead : %g", time);
488 /* Notify all the hosts that have failed */
489 /* FIXME: iterate through the list of failed host and mark each of them */
490 /* as failed. On each host, signal all the running processes with host_fail */
492 /* Handle any pending timer */
493 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
494 //FIXME: make the timers being real callbacks
495 // (i.e. provide dispatchers that read and expand the args)
496 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
498 timer->func(timer->args);
502 /* Wake up all processes waiting for a Surf action to finish */
503 xbt_dynar_foreach(all_existing_models, iter, model) {
504 XBT_DEBUG("Handling process whose action failed");
505 while ((action = surf_model_extract_failed_action_set(model))) {
506 XBT_DEBUG(" Handling Action %p",action);
507 SIMIX_simcall_exit((smx_synchro_t) action->getData());
509 XBT_DEBUG("Handling process whose action terminated normally");
510 while ((action = surf_model_extract_done_action_set(model))) {
511 XBT_DEBUG(" Handling Action %p",action);
512 if (action->getData() == NULL)
513 XBT_DEBUG("probably vcpu's action %p, skip", action);
515 SIMIX_simcall_exit((smx_synchro_t) action->getData());
519 /* Autorestart all process */
520 if(host_that_restart) {
521 char *hostname = NULL;
522 xbt_dynar_foreach(host_that_restart,iter,hostname) {
523 XBT_INFO("Restart processes on host: %s",hostname);
524 SIMIX_host_autorestart(sg_host_by_name(hostname));
526 xbt_dynar_reset(host_that_restart);
529 /* Clean processes to destroy */
530 SIMIX_process_empty_trash();
533 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
535 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
537 if (xbt_swag_size(simix_global->process_list) != 0) {
541 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
542 SIMIX_display_process_status();
548 * \brief Set the date to execute a function
550 * Set the date to execute the function on the surf.
551 * \param date Date to execute function
552 * \param function Function to be executed
553 * \param arg Parameters of the function
556 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
558 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
561 timer->func = function;
563 xbt_heap_push(simix_timers, timer, date);
566 /** @brief cancels a timer that was added earlier */
567 void SIMIX_timer_remove(smx_timer_t timer) {
568 xbt_heap_rm_elm(simix_timers, timer, timer->date);
571 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
572 double SIMIX_timer_get_date(smx_timer_t timer) {
573 return timer?timer->date:0;
577 * \brief Registers a function to create a process.
579 * This function registers a function to be called
580 * when a new process is created. The function has
581 * to call SIMIX_process_create().
582 * \param function create process function
584 void SIMIX_function_register_process_create(smx_creation_func_t
587 simix_global->create_process_function = function;
591 * \brief Registers a function to kill a process.
593 * This function registers a function to be called when a
594 * process is killed. The function has to call the SIMIX_process_kill().
596 * \param function Kill process function
598 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
601 simix_global->kill_process_function = function;
605 * \brief Registers a function to cleanup a process.
607 * This function registers a user function to be called when
608 * a process ends properly.
610 * \param function cleanup process function
612 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
615 simix_global->cleanup_process_function = function;
619 void SIMIX_display_process_status(void)
621 if (simix_global->process_list == NULL) {
625 smx_process_t process = NULL;
626 int nbprocess = xbt_swag_size(simix_global->process_list);
628 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
629 /* List the process and their state */
631 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
632 xbt_swag_foreach(process, simix_global->process_list) {
634 if (process->waiting_synchro) {
636 const char* synchro_description = "unknown";
637 switch (process->waiting_synchro->type) {
639 case SIMIX_SYNC_EXECUTE:
640 synchro_description = "execution";
643 case SIMIX_SYNC_PARALLEL_EXECUTE:
644 synchro_description = "parallel execution";
647 case SIMIX_SYNC_COMMUNICATE:
648 synchro_description = "communication";
651 case SIMIX_SYNC_SLEEP:
652 synchro_description = "sleeping";
655 case SIMIX_SYNC_JOIN:
656 synchro_description = "joining";
659 case SIMIX_SYNC_SYNCHRO:
660 synchro_description = "synchronization";
664 synchro_description = "I/O";
667 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
668 process->pid, process->name, sg_host_get_name(process->host),
669 synchro_description, process->waiting_synchro,
670 process->waiting_synchro->name, (int)process->waiting_synchro->state);
673 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
678 static void* SIMIX_synchro_mallocator_new_f(void) {
679 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
680 synchro->simcalls = xbt_fifo_new();
684 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
685 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
689 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
691 // we also recycle the simcall list
692 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
693 xbt_fifo_reset(fifo);
694 memset(synchro, 0, sizeof(s_smx_synchro_t));
695 ((smx_synchro_t) synchro)->simcalls = fifo;
698 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
699 return SIMIX_asr_get_properties(name);
701 xbt_dict_t SIMIX_asr_get_properties(const char *name)
703 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
706 int SIMIX_is_maestro()
708 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;