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 "src/surf/storage_interface.hpp"
16 #include "src/surf/xml/platf.hpp"
17 #include "smx_private.h"
18 #include "smx_private.hpp"
20 #include "xbt/sysdep.h"
23 #include "xbt/ex.h" /* ex_backtrace_display */
25 #include "src/mc/mc_replay.h"
26 #include "simgrid/sg_config.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"
42 #include "src/smpi/private.h"
45 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
46 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
47 "Logging specific to SIMIX (kernel)");
49 smx_global_t simix_global = NULL;
50 static xbt_heap_t simix_timers = NULL;
52 /** @brief Timer datatype */
53 typedef struct s_smx_timer {
59 void (*SMPI_switch_data_segment)(int) = NULL;
61 static void* SIMIX_synchro_mallocator_new_f(void);
62 static void SIMIX_synchro_mallocator_free_f(void* synchro);
63 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
65 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
68 int _sg_do_verbose_exit = 1;
69 static void _XBT_CALL inthandler(int ignored)
71 if ( _sg_do_verbose_exit ) {
72 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
73 SIMIX_display_process_status();
76 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
82 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
84 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
86 "Access violation detected.\n"
87 "This can result from a programming error in your code or, although less likely,\n"
88 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
89 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
91 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
92 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
93 smx_context_stack_size / 1024);
94 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
96 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
97 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
99 } else if (siginfo->si_signo == SIGSEGV) {
100 fprintf(stderr, "Segmentation fault.\n");
102 if (smpi_enabled() && !smpi_privatize_global_variables) {
103 #ifdef HAVE_PRIVATIZATION
105 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
108 "Sadly, your system does not support --cfg=smpi/privatize_global_variables:yes (yet).\n");
116 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
119 * Install signal handler for SIGSEGV. Check that nobody has already installed
120 * its own handler. For example, the Java VM does this.
122 static void install_segvhandler(void)
124 stack_t stack, old_stack;
125 stack.ss_sp = sigsegv_stack;
126 stack.ss_size = sizeof sigsegv_stack;
129 if (sigaltstack(&stack, &old_stack) == -1) {
130 XBT_WARN("Failed to register alternate signal stack: %s",
134 if (!(old_stack.ss_flags & SS_DISABLE)) {
135 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
136 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
137 sigaltstack(&old_stack, NULL);
140 struct sigaction action, old_action;
141 action.sa_sigaction = segvhandler;
142 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
143 sigemptyset(&action.sa_mask);
145 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
146 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
150 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
151 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
152 (old_action.sa_flags & SA_SIGINFO) ?
153 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
154 sigaction(SIGSEGV, &old_action, NULL);
159 /********************************* SIMIX **************************************/
161 double SIMIX_timer_next(void)
163 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
166 static void kill_process(smx_process_t process)
168 SIMIX_process_kill(process, NULL);
171 static void SIMIX_storage_create_(smx_storage_t storage)
173 const char* key = xbt_dict_get_elm_key(storage);
174 SIMIX_storage_create(key, storage, NULL);
177 static std::function<void()> maestro_code;
182 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
184 maestro_code = std::move(code);
190 void SIMIX_set_maestro(void (*code)(void*), void* data)
192 maestro_code = std::bind(code, data);
197 * \brief Initialize SIMIX internal data.
202 void SIMIX_global_init(int *argc, char **argv)
205 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
208 s_smx_process_t proc;
211 simix_global = xbt_new0(s_smx_global_t, 1);
213 #ifdef TIME_BENCH_AMDAHL
214 simix_global->timer_seq = xbt_os_timer_new();
215 simix_global->timer_par = xbt_os_timer_new();
216 xbt_os_cputimer_start(simix_global->timer_seq);
218 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
219 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
220 simix_global->process_list =
221 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
222 simix_global->process_to_destroy =
223 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
225 simix_global->maestro_process = NULL;
226 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
228 simix_global->create_process_function = SIMIX_process_create;
229 simix_global->kill_process_function = kill_process;
230 simix_global->cleanup_process_function = SIMIX_process_cleanup;
231 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
232 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
233 SIMIX_synchro_mallocator_reset_f);
234 simix_global->mutex = xbt_os_mutex_init();
236 surf_init(argc, argv); /* Initialize SURF structures */
237 SIMIX_context_mod_init();
239 // Either create a new context with maestro or create
240 // a context object with the current context mestro):
241 simgrid::simix::create_maestro(maestro_code);
243 /* context exception handlers */
244 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
245 __xbt_ex_terminate = SIMIX_process_exception_terminate;
247 SIMIX_network_init();
249 /* Prepare to display some more info when dying on Ctrl-C pressing */
250 signal(SIGINT, inthandler);
253 install_segvhandler();
255 /* register a function to be called by SURF after the environment creation */
257 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
258 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
259 SIMIX_host_create(&host);
261 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
262 const char* id = storage->getName();
263 // TODO, create sg_storage_by_name
264 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, id);
265 xbt_assert(s != NULL, "Storage not found for name %s", id);
266 SIMIX_storage_create_(s);
272 simix_timers = xbt_heap_new(8, &free);
275 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
277 if (sg_cfg_get_boolean("clean_atexit"))
281 // The communication initialization is done ASAP.
282 // We need to communicate initialization of the different layers to the model-checker.
286 if (_sg_cfg_exit_asap)
293 * \brief Clean the SIMIX simulation
295 * This functions remove the memory used by SIMIX
297 void SIMIX_clean(void)
299 #ifdef TIME_BENCH_PER_SR
300 smx_ctx_raw_new_sr();
302 if (smx_cleaned) return; // to avoid double cleaning by java and C
304 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
305 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
307 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
308 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
309 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
311 /* Kill all processes (but maestro) */
312 SIMIX_process_killall(simix_global->maestro_process, 1);
314 /* Exit the SIMIX network module */
315 SIMIX_network_exit();
317 xbt_heap_free(simix_timers);
319 /* Free the remaining data structures */
320 xbt_dynar_free(&simix_global->process_to_run);
321 xbt_dynar_free(&simix_global->process_that_ran);
322 xbt_swag_free(simix_global->process_to_destroy);
323 xbt_swag_free(simix_global->process_list);
324 simix_global->process_list = NULL;
325 simix_global->process_to_destroy = NULL;
326 xbt_dict_free(&(simix_global->registered_functions));
328 xbt_os_mutex_destroy(simix_global->mutex);
329 simix_global->mutex = NULL;
331 /* Let's free maestro now */
332 SIMIX_context_free(simix_global->maestro_process->context);
333 xbt_free(simix_global->maestro_process->running_ctx);
334 xbt_free(simix_global->maestro_process);
335 simix_global->maestro_process = NULL;
337 /* Restore the default exception setup */
338 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
339 __xbt_ex_terminate = &__xbt_ex_terminate_default;
341 /* Finish context module and SURF */
342 SIMIX_context_mod_exit();
346 #ifdef TIME_BENCH_AMDAHL
347 xbt_os_cputimer_stop(simix_global->timer_seq);
348 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
349 xbt_os_timer_elapsed(simix_global->timer_seq),
350 xbt_os_timer_elapsed(simix_global->timer_par));
351 xbt_os_timer_free(simix_global->timer_seq);
352 xbt_os_timer_free(simix_global->timer_par);
355 xbt_mallocator_free(simix_global->synchro_mallocator);
356 xbt_free(simix_global);
365 * \brief A clock (in second).
367 * \return Return the clock.
369 double SIMIX_get_clock(void)
371 if(MC_is_active() || MC_record_replay_is_active()){
372 return MC_process_clock_get(SIMIX_process_self());
374 return surf_get_clock();
378 static int process_syscall_color(void *p)
380 switch ((*(smx_process_t *)p)->simcall.call) {
382 case SIMCALL_PROCESS_KILL:
384 case SIMCALL_PROCESS_RESUME:
393 * \brief Run the main simulation loop.
398 MC_record_replay_init();
399 MC_record_replay_from_string(MC_record_path);
404 smx_process_t process;
405 surf_action_t action;
411 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
412 xbt_dynar_length(simix_global->process_to_run));
413 #ifdef TIME_BENCH_PER_SR
414 smx_ctx_raw_new_sr();
416 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
417 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
418 xbt_dynar_length(simix_global->process_to_run));
420 /* Run all processes that are ready to run, possibly in parallel */
421 #ifdef TIME_BENCH_AMDAHL
422 xbt_os_cputimer_stop(simix_global->timer_seq);
423 xbt_os_cputimer_resume(simix_global->timer_par);
425 SIMIX_process_runall();
426 #ifdef TIME_BENCH_AMDAHL
427 xbt_os_cputimer_stop(simix_global->timer_par);
428 xbt_os_cputimer_resume(simix_global->timer_seq);
431 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
432 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
434 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
436 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
438 /* Here, the order is ok because:
440 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
442 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
444 * - if there is no kill during the simulation, processes remain sorted according by their PID.
445 * rational: This can be proved inductively.
446 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
447 * Let's show that it is still so at the end of this round.
448 * - if a process is added when being created, that's from maestro. It can be either at startup
449 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
450 * in arbitrary order (inductive hypothesis), we are fine.
451 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
452 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
453 * of a simcall terminates, we're still good. Proof:
454 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
455 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
456 * - For communications, this function is called from SIMIX_comm_finish().
457 * This function itself don't mess with the order since simcalls are handled in FIFO order.
458 * The function is called:
459 * - before the comm starts (invalid parameters, or resource already dead or whatever).
460 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
461 * - because the communication failed or were canceled after startup. In this case, it's called from the function
462 * we are in, by the chunk:
463 * set = model->states.failed_action_set;
464 * while ((synchro = xbt_swag_extract(set)))
465 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
466 * This order is also fixed because it depends of the order in which the surf actions were
467 * added to the system, and only maestro can add stuff this way, through simcalls.
468 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
469 * poped out of the swag does not depend on the user code's execution order.
470 * - because the communication terminated. In this case, synchros are served in the order given by
471 * set = model->states.done_action_set;
472 * while ((synchro = xbt_swag_extract(set)))
473 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
474 * and the argument is very similar to the previous one.
475 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
476 * 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.
477 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
478 * - If there is some process killings, the order is changed by this decision that comes from user-land
479 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
480 * So, even the order change induced by the process killing is perfectly reproducible.
482 * So science works, bitches [http://xkcd.com/54/].
484 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
485 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
486 * That would thus be a pure waste of time.
489 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
490 if (process->simcall.call != SIMCALL_NONE) {
491 SIMIX_simcall_handle(&process->simcall, 0);
494 /* Wake up all processes waiting for a Surf action to finish */
495 xbt_dynar_foreach(all_existing_models, iter, model) {
496 XBT_DEBUG("Handling process whose action failed");
497 while ((action = surf_model_extract_failed_action_set(model))) {
498 XBT_DEBUG(" Handling Action %p",action);
499 SIMIX_simcall_exit((smx_synchro_t) action->getData());
501 XBT_DEBUG("Handling process whose action terminated normally");
502 while ((action = surf_model_extract_done_action_set(model))) {
503 XBT_DEBUG(" Handling Action %p",action);
504 if (action->getData() == NULL)
505 XBT_DEBUG("probably vcpu's action %p, skip", action);
507 SIMIX_simcall_exit((smx_synchro_t) action->getData());
512 time = SIMIX_timer_next();
513 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
514 XBT_DEBUG("Calling surf_solve");
515 time = surf_solve(time);
516 XBT_DEBUG("Moving time ahead : %g", time);
518 /* Notify all the hosts that have failed */
519 /* FIXME: iterate through the list of failed host and mark each of them */
520 /* as failed. On each host, signal all the running processes with host_fail */
522 /* Handle any pending timer */
523 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
524 //FIXME: make the timers being real callbacks
525 // (i.e. provide dispatchers that read and expand the args)
526 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
528 timer->func(timer->args);
532 /* Wake up all processes waiting for a Surf action to finish */
533 xbt_dynar_foreach(all_existing_models, iter, model) {
534 XBT_DEBUG("Handling process whose action failed");
535 while ((action = surf_model_extract_failed_action_set(model))) {
536 XBT_DEBUG(" Handling Action %p",action);
537 SIMIX_simcall_exit((smx_synchro_t) action->getData());
539 XBT_DEBUG("Handling process whose action terminated normally");
540 while ((action = surf_model_extract_done_action_set(model))) {
541 XBT_DEBUG(" Handling Action %p",action);
542 if (action->getData() == NULL)
543 XBT_DEBUG("probably vcpu's action %p, skip", action);
545 SIMIX_simcall_exit((smx_synchro_t) action->getData());
549 /* Autorestart all process */
550 char *hostname = NULL;
551 xbt_dynar_foreach(host_that_restart,iter,hostname) {
552 XBT_INFO("Restart processes on host: %s",hostname);
553 SIMIX_host_autorestart(sg_host_by_name(hostname));
555 xbt_dynar_reset(host_that_restart);
557 /* Clean processes to destroy */
558 SIMIX_process_empty_trash();
561 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
563 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
565 if (xbt_swag_size(simix_global->process_list) != 0) {
569 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
570 SIMIX_display_process_status();
576 * \brief Set the date to execute a function
578 * Set the date to execute the function on the surf.
579 * \param date Date to execute function
580 * \param function Function to be executed
581 * \param arg Parameters of the function
584 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
586 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
589 timer->func = function;
591 xbt_heap_push(simix_timers, timer, date);
594 /** @brief cancels a timer that was added earlier */
595 void SIMIX_timer_remove(smx_timer_t timer) {
596 xbt_heap_rm_elm(simix_timers, timer, timer->date);
599 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
600 double SIMIX_timer_get_date(smx_timer_t timer) {
601 return timer?timer->date:0;
605 * \brief Registers a function to create a process.
607 * This function registers a function to be called
608 * when a new process is created. The function has
609 * to call SIMIX_process_create().
610 * \param function create process function
612 void SIMIX_function_register_process_create(smx_creation_func_t
615 simix_global->create_process_function = function;
619 * \brief Registers a function to kill a process.
621 * This function registers a function to be called when a
622 * process is killed. The function has to call the SIMIX_process_kill().
624 * \param function Kill process function
626 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
629 simix_global->kill_process_function = function;
633 * \brief Registers a function to cleanup a process.
635 * This function registers a user function to be called when
636 * a process ends properly.
638 * \param function cleanup process function
640 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
643 simix_global->cleanup_process_function = function;
647 void SIMIX_display_process_status(void)
649 if (simix_global->process_list == NULL) {
653 smx_process_t process = NULL;
654 int nbprocess = xbt_swag_size(simix_global->process_list);
656 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
657 /* List the process and their state */
659 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
660 xbt_swag_foreach(process, simix_global->process_list) {
662 if (process->waiting_synchro) {
664 const char* synchro_description = "unknown";
665 switch (process->waiting_synchro->type) {
667 case SIMIX_SYNC_EXECUTE:
668 synchro_description = "execution";
671 case SIMIX_SYNC_PARALLEL_EXECUTE:
672 synchro_description = "parallel execution";
675 case SIMIX_SYNC_COMMUNICATE:
676 synchro_description = "communication";
679 case SIMIX_SYNC_SLEEP:
680 synchro_description = "sleeping";
683 case SIMIX_SYNC_JOIN:
684 synchro_description = "joining";
687 case SIMIX_SYNC_SYNCHRO:
688 synchro_description = "synchronization";
692 synchro_description = "I/O";
695 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
696 process->pid, process->name, sg_host_get_name(process->host),
697 synchro_description, process->waiting_synchro,
698 process->waiting_synchro->name, (int)process->waiting_synchro->state);
701 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
706 static void* SIMIX_synchro_mallocator_new_f(void) {
707 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
708 synchro->simcalls = xbt_fifo_new();
712 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
713 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
717 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
719 // we also recycle the simcall list
720 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
721 xbt_fifo_reset(fifo);
722 memset(synchro, 0, sizeof(s_smx_synchro_t));
723 ((smx_synchro_t) synchro)->simcalls = fifo;
726 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
727 return SIMIX_asr_get_properties(name);
729 xbt_dict_t SIMIX_asr_get_properties(const char *name)
731 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
734 int SIMIX_is_maestro()
736 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;