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) {
104 #ifdef HAVE_PRIVATIZATION
106 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
109 "Sadly, your system does not support --cfg=smpi/privatize_global_variables:yes (yet).\n");
117 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
120 * Install signal handler for SIGSEGV. Check that nobody has already installed
121 * its own handler. For example, the Java VM does this.
123 static void install_segvhandler(void)
125 stack_t stack, old_stack;
126 stack.ss_sp = sigsegv_stack;
127 stack.ss_size = sizeof sigsegv_stack;
130 if (sigaltstack(&stack, &old_stack) == -1) {
131 XBT_WARN("Failed to register alternate signal stack: %s",
135 if (!(old_stack.ss_flags & SS_DISABLE)) {
136 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
137 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
138 sigaltstack(&old_stack, NULL);
141 struct sigaction action, old_action;
142 action.sa_sigaction = segvhandler;
143 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
144 sigemptyset(&action.sa_mask);
146 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
147 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
151 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
152 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
153 (old_action.sa_flags & SA_SIGINFO) ?
154 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
155 sigaction(SIGSEGV, &old_action, NULL);
160 /********************************* SIMIX **************************************/
162 double SIMIX_timer_next(void)
164 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
167 static void kill_process(smx_process_t process)
169 SIMIX_process_kill(process, NULL);
172 static void SIMIX_storage_create_(smx_storage_t storage)
174 const char* key = xbt_dict_get_elm_key(storage);
175 SIMIX_storage_create(key, storage, NULL);
180 * \brief Initialize SIMIX internal data.
185 void SIMIX_global_init(int *argc, char **argv)
188 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
191 s_smx_process_t proc;
194 simix_global = xbt_new0(s_smx_global_t, 1);
196 #ifdef TIME_BENCH_AMDAHL
197 simix_global->timer_seq = xbt_os_timer_new();
198 simix_global->timer_par = xbt_os_timer_new();
199 xbt_os_cputimer_start(simix_global->timer_seq);
201 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
202 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
203 simix_global->process_list =
204 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
205 simix_global->process_to_destroy =
206 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
208 simix_global->maestro_process = NULL;
209 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
211 simix_global->create_process_function = SIMIX_process_create;
212 simix_global->kill_process_function = kill_process;
213 simix_global->cleanup_process_function = SIMIX_process_cleanup;
214 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
215 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
216 SIMIX_synchro_mallocator_reset_f);
217 simix_global->autorestart = SIMIX_host_restart_processes;
218 simix_global->mutex = xbt_os_mutex_init();
220 surf_init(argc, argv); /* Initialize SURF structures */
221 SIMIX_context_mod_init();
222 SIMIX_create_maestro_process();
224 /* context exception handlers */
225 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
226 __xbt_ex_terminate = SIMIX_process_exception_terminate;
228 SIMIX_network_init();
230 /* Prepare to display some more info when dying on Ctrl-C pressing */
231 signal(SIGINT, inthandler);
234 /* Install SEGV handler */
235 install_segvhandler();
237 /* register a function to be called by SURF after the environment creation */
239 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
240 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
241 SIMIX_host_create(&host);
243 surf_on_storage_created(SIMIX_storage_create_);
247 simix_timers = xbt_heap_new(8, &free);
250 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
252 if (sg_cfg_get_boolean("clean_atexit"))
256 // The communication initialization is done ASAP.
257 // We need to communicate initialization of the different layers to the model-checker.
261 if (_sg_cfg_exit_asap)
268 * \brief Clean the SIMIX simulation
270 * This functions remove the memory used by SIMIX
272 void SIMIX_clean(void)
274 #ifdef TIME_BENCH_PER_SR
275 smx_ctx_raw_new_sr();
277 if (smx_cleaned) return; // to avoid double cleaning by java and C
279 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
280 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
282 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
283 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
284 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
286 /* Kill all processes (but maestro) */
287 SIMIX_process_killall(simix_global->maestro_process, 1);
289 /* Exit the SIMIX network module */
290 SIMIX_network_exit();
292 xbt_heap_free(simix_timers);
294 /* Free the remaining data structures */
295 xbt_dynar_free(&simix_global->process_to_run);
296 xbt_dynar_free(&simix_global->process_that_ran);
297 xbt_swag_free(simix_global->process_to_destroy);
298 xbt_swag_free(simix_global->process_list);
299 simix_global->process_list = NULL;
300 simix_global->process_to_destroy = NULL;
301 xbt_dict_free(&(simix_global->registered_functions));
303 xbt_os_mutex_destroy(simix_global->mutex);
304 simix_global->mutex = NULL;
306 /* Let's free maestro now */
307 SIMIX_context_free(simix_global->maestro_process->context);
308 xbt_free(simix_global->maestro_process->running_ctx);
309 xbt_free(simix_global->maestro_process);
310 simix_global->maestro_process = NULL;
312 /* Restore the default exception setup */
313 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
314 __xbt_ex_terminate = &__xbt_ex_terminate_default;
316 /* Finish context module and SURF */
317 SIMIX_context_mod_exit();
321 #ifdef TIME_BENCH_AMDAHL
322 xbt_os_cputimer_stop(simix_global->timer_seq);
323 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
324 xbt_os_timer_elapsed(simix_global->timer_seq),
325 xbt_os_timer_elapsed(simix_global->timer_par));
326 xbt_os_timer_free(simix_global->timer_seq);
327 xbt_os_timer_free(simix_global->timer_par);
330 xbt_mallocator_free(simix_global->synchro_mallocator);
331 xbt_free(simix_global);
340 * \brief A clock (in second).
342 * \return Return the clock.
344 double SIMIX_get_clock(void)
346 if(MC_is_active() || MC_record_replay_is_active()){
347 return MC_process_clock_get(SIMIX_process_self());
349 return surf_get_clock();
353 static int process_syscall_color(void *p)
355 switch ((*(smx_process_t *)p)->simcall.call) {
357 case SIMCALL_PROCESS_KILL:
359 case SIMCALL_PROCESS_RESUME:
368 * \brief Run the main simulation loop.
373 MC_record_replay_init();
374 MC_record_replay_from_string(MC_record_path);
379 smx_process_t process;
380 surf_action_t action;
386 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
387 xbt_dynar_length(simix_global->process_to_run));
388 #ifdef TIME_BENCH_PER_SR
389 smx_ctx_raw_new_sr();
391 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
392 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
393 xbt_dynar_length(simix_global->process_to_run));
395 /* Run all processes that are ready to run, possibly in parallel */
396 #ifdef TIME_BENCH_AMDAHL
397 xbt_os_cputimer_stop(simix_global->timer_seq);
398 xbt_os_cputimer_resume(simix_global->timer_par);
400 SIMIX_process_runall();
401 #ifdef TIME_BENCH_AMDAHL
402 xbt_os_cputimer_stop(simix_global->timer_par);
403 xbt_os_cputimer_resume(simix_global->timer_seq);
406 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
407 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
409 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
411 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
413 /* Here, the order is ok because:
415 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
417 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
419 * - if there is no kill during the simulation, processes remain sorted according by their PID.
420 * rational: This can be proved inductively.
421 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
422 * Let's show that it is still so at the end of this round.
423 * - if a process is added when being created, that's from maestro. It can be either at startup
424 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
425 * in arbitrary order (inductive hypothesis), we are fine.
426 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
427 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
428 * of a simcall terminates, we're still good. Proof:
429 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
430 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
431 * - For communications, this function is called from SIMIX_comm_finish().
432 * This function itself don't mess with the order since simcalls are handled in FIFO order.
433 * The function is called:
434 * - before the comm starts (invalid parameters, or resource already dead or whatever).
435 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
436 * - because the communication failed or were canceled after startup. In this case, it's called from the function
437 * we are in, by the chunk:
438 * set = model->states.failed_action_set;
439 * while ((synchro = xbt_swag_extract(set)))
440 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
441 * This order is also fixed because it depends of the order in which the surf actions were
442 * added to the system, and only maestro can add stuff this way, through simcalls.
443 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
444 * poped out of the swag does not depend on the user code's execution order.
445 * - because the communication terminated. In this case, synchros are served in the order given by
446 * set = model->states.done_action_set;
447 * while ((synchro = xbt_swag_extract(set)))
448 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
449 * and the argument is very similar to the previous one.
450 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
451 * 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.
452 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
453 * - If there is some process killings, the order is changed by this decision that comes from user-land
454 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
455 * So, even the order change induced by the process killing is perfectly reproducible.
457 * So science works, bitches [http://xkcd.com/54/].
459 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
460 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
461 * That would thus be a pure waste of time.
464 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
465 if (process->simcall.call != SIMCALL_NONE) {
466 SIMIX_simcall_handle(&process->simcall, 0);
469 /* Wake up all processes waiting for a Surf action to finish */
470 xbt_dynar_foreach(all_existing_models, iter, model) {
471 XBT_DEBUG("Handling process whose action failed");
472 while ((action = surf_model_extract_failed_action_set(model))) {
473 XBT_DEBUG(" Handling Action %p",action);
474 SIMIX_simcall_exit((smx_synchro_t) action->getData());
476 XBT_DEBUG("Handling process whose action terminated normally");
477 while ((action = surf_model_extract_done_action_set(model))) {
478 XBT_DEBUG(" Handling Action %p",action);
479 if (action->getData() == NULL)
480 XBT_DEBUG("probably vcpu's action %p, skip", action);
482 SIMIX_simcall_exit((smx_synchro_t) action->getData());
487 time = SIMIX_timer_next();
488 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
489 XBT_DEBUG("Calling surf_solve");
490 time = surf_solve(time);
491 XBT_DEBUG("Moving time ahead : %g", time);
493 /* Notify all the hosts that have failed */
494 /* FIXME: iterate through the list of failed host and mark each of them */
495 /* as failed. On each host, signal all the running processes with host_fail */
497 /* Handle any pending timer */
498 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
499 //FIXME: make the timers being real callbacks
500 // (i.e. provide dispatchers that read and expand the args)
501 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
503 timer->func(timer->args);
507 /* Wake up all processes waiting for a Surf action to finish */
508 xbt_dynar_foreach(all_existing_models, iter, model) {
509 XBT_DEBUG("Handling process whose action failed");
510 while ((action = surf_model_extract_failed_action_set(model))) {
511 XBT_DEBUG(" Handling Action %p",action);
512 SIMIX_simcall_exit((smx_synchro_t) action->getData());
514 XBT_DEBUG("Handling process whose action terminated normally");
515 while ((action = surf_model_extract_done_action_set(model))) {
516 XBT_DEBUG(" Handling Action %p",action);
517 if (action->getData() == NULL)
518 XBT_DEBUG("probably vcpu's action %p, skip", action);
520 SIMIX_simcall_exit((smx_synchro_t) action->getData());
524 /* Autorestart all process */
525 if(host_that_restart) {
526 char *hostname = NULL;
527 xbt_dynar_foreach(host_that_restart,iter,hostname) {
528 XBT_INFO("Restart processes on host: %s",hostname);
529 SIMIX_host_autorestart(sg_host_by_name(hostname));
531 xbt_dynar_reset(host_that_restart);
534 /* Clean processes to destroy */
535 SIMIX_process_empty_trash();
538 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
540 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
542 if (xbt_swag_size(simix_global->process_list) != 0) {
546 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
547 SIMIX_display_process_status();
553 * \brief Set the date to execute a function
555 * Set the date to execute the function on the surf.
556 * \param date Date to execute function
557 * \param function Function to be executed
558 * \param arg Parameters of the function
561 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
563 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
566 timer->func = function;
568 xbt_heap_push(simix_timers, timer, date);
571 /** @brief cancels a timer that was added earlier */
572 void SIMIX_timer_remove(smx_timer_t timer) {
573 xbt_heap_rm_elm(simix_timers, timer, timer->date);
576 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
577 double SIMIX_timer_get_date(smx_timer_t timer) {
578 return timer?timer->date:0;
582 * \brief Registers a function to create a process.
584 * This function registers a function to be called
585 * when a new process is created. The function has
586 * to call SIMIX_process_create().
587 * \param function create process function
589 void SIMIX_function_register_process_create(smx_creation_func_t
592 simix_global->create_process_function = function;
596 * \brief Registers a function to kill a process.
598 * This function registers a function to be called when a
599 * process is killed. The function has to call the SIMIX_process_kill().
601 * \param function Kill process function
603 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
606 simix_global->kill_process_function = function;
610 * \brief Registers a function to cleanup a process.
612 * This function registers a user function to be called when
613 * a process ends properly.
615 * \param function cleanup process function
617 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
620 simix_global->cleanup_process_function = function;
624 void SIMIX_display_process_status(void)
626 if (simix_global->process_list == NULL) {
630 smx_process_t process = NULL;
631 int nbprocess = xbt_swag_size(simix_global->process_list);
633 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
634 /* List the process and their state */
636 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
637 xbt_swag_foreach(process, simix_global->process_list) {
639 if (process->waiting_synchro) {
641 const char* synchro_description = "unknown";
642 switch (process->waiting_synchro->type) {
644 case SIMIX_SYNC_EXECUTE:
645 synchro_description = "execution";
648 case SIMIX_SYNC_PARALLEL_EXECUTE:
649 synchro_description = "parallel execution";
652 case SIMIX_SYNC_COMMUNICATE:
653 synchro_description = "communication";
656 case SIMIX_SYNC_SLEEP:
657 synchro_description = "sleeping";
660 case SIMIX_SYNC_JOIN:
661 synchro_description = "joining";
664 case SIMIX_SYNC_SYNCHRO:
665 synchro_description = "synchronization";
669 synchro_description = "I/O";
672 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
673 process->pid, process->name, sg_host_get_name(process->host),
674 synchro_description, process->waiting_synchro,
675 process->waiting_synchro->name, (int)process->waiting_synchro->state);
678 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
683 static void* SIMIX_synchro_mallocator_new_f(void) {
684 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
685 synchro->simcalls = xbt_fifo_new();
689 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
690 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
694 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
696 // we also recycle the simcall list
697 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
698 xbt_fifo_reset(fifo);
699 memset(synchro, 0, sizeof(s_smx_synchro_t));
700 ((smx_synchro_t) synchro)->simcalls = fifo;
703 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
704 return SIMIX_asr_get_properties(name);
706 xbt_dict_t SIMIX_asr_get_properties(const char *name)
708 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
711 int SIMIX_is_maestro()
713 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;