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 "smx_private.h"
15 #include "smx_private.hpp"
17 #include "xbt/sysdep.h"
20 #include "xbt/ex.h" /* ex_backtrace_display */
22 #include "src/mc/mc_replay.h"
23 #include "simgrid/sg_config.h"
25 #include "src/surf/callbacks.h"
28 #include "src/mc/mc_private.h"
29 #include "src/mc/mc_protocol.h"
30 #include "src/mc/mc_client.h"
35 #include "src/mc/mc_protocol.h"
38 #include "src/mc/mc_record.h"
41 #include "src/smpi/private.h"
44 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
45 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
46 "Logging specific to SIMIX (kernel)");
48 smx_global_t simix_global = NULL;
49 static xbt_heap_t simix_timers = NULL;
51 /** @brief Timer datatype */
52 typedef struct s_smx_timer {
58 void (*SMPI_switch_data_segment)(int) = NULL;
60 static void* SIMIX_synchro_mallocator_new_f(void);
61 static void SIMIX_synchro_mallocator_free_f(void* synchro);
62 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
64 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
67 int _sg_do_verbose_exit = 1;
68 static void _XBT_CALL inthandler(int ignored)
70 if ( _sg_do_verbose_exit ) {
71 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
72 SIMIX_display_process_status();
75 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
81 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
83 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
85 "Access violation detected.\n"
86 "This can result from a programming error in your code or, although less likely,\n"
87 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
88 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
90 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
91 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
92 smx_context_stack_size / 1024);
93 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
95 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
96 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
98 } else if (siginfo->si_signo == SIGSEGV) {
99 fprintf(stderr, "Segmentation fault.\n");
101 if (smpi_enabled() && !smpi_privatize_global_variables) {
103 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
110 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
113 * Install signal handler for SIGSEGV. Check that nobody has already installed
114 * its own handler. For example, the Java VM does this.
116 static void install_segvhandler(void)
118 stack_t stack, old_stack;
119 stack.ss_sp = sigsegv_stack;
120 stack.ss_size = sizeof sigsegv_stack;
123 if (sigaltstack(&stack, &old_stack) == -1) {
124 XBT_WARN("Failed to register alternate signal stack: %s",
128 if (!(old_stack.ss_flags & SS_DISABLE)) {
129 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
130 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
131 sigaltstack(&old_stack, NULL);
134 struct sigaction action, old_action;
135 action.sa_sigaction = segvhandler;
136 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
137 sigemptyset(&action.sa_mask);
139 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
140 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
144 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
145 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
146 (old_action.sa_flags & SA_SIGINFO) ?
147 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
148 sigaction(SIGSEGV, &old_action, NULL);
153 /********************************* SIMIX **************************************/
155 double SIMIX_timer_next(void)
157 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
160 static void kill_process(smx_process_t process)
162 SIMIX_process_kill(process, NULL);
165 static void SIMIX_storage_create_(smx_storage_t storage)
167 const char* key = xbt_dict_get_elm_key(storage);
168 SIMIX_storage_create(key, storage, NULL);
173 * \brief Initialize SIMIX internal data.
178 void SIMIX_global_init(int *argc, char **argv)
181 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
184 s_smx_process_t proc;
187 simix_global = xbt_new0(s_smx_global_t, 1);
189 #ifdef TIME_BENCH_AMDAHL
190 simix_global->timer_seq = xbt_os_timer_new();
191 simix_global->timer_par = xbt_os_timer_new();
192 xbt_os_cputimer_start(simix_global->timer_seq);
194 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
195 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
196 simix_global->process_list =
197 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
198 simix_global->process_to_destroy =
199 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
201 simix_global->maestro_process = NULL;
202 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
204 simix_global->create_process_function = SIMIX_process_create;
205 simix_global->kill_process_function = kill_process;
206 simix_global->cleanup_process_function = SIMIX_process_cleanup;
207 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
208 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
209 SIMIX_synchro_mallocator_reset_f);
210 simix_global->autorestart = SIMIX_host_restart_processes;
211 simix_global->mutex = xbt_os_mutex_init();
213 surf_init(argc, argv); /* Initialize SURF structures */
214 SIMIX_context_mod_init();
215 SIMIX_create_maestro_process();
217 /* context exception handlers */
218 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
219 __xbt_ex_terminate = SIMIX_process_exception_terminate;
221 SIMIX_network_init();
223 /* Prepare to display some more info when dying on Ctrl-C pressing */
224 signal(SIGINT, inthandler);
227 /* Install SEGV handler */
228 install_segvhandler();
230 /* register a function to be called by SURF after the environment creation */
232 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
233 surf_on_host_created(SIMIX_host_create);
234 surf_on_storage_created(SIMIX_storage_create_);
238 simix_timers = xbt_heap_new(8, &free);
241 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
243 if (sg_cfg_get_boolean("clean_atexit"))
247 // The communication initialization is done ASAP.
248 // We need to communicate initialization of the different layers to the model-checker.
252 if (_sg_cfg_exit_asap)
259 * \brief Clean the SIMIX simulation
261 * This functions remove the memory used by SIMIX
263 void SIMIX_clean(void)
265 #ifdef TIME_BENCH_PER_SR
266 smx_ctx_raw_new_sr();
268 if (smx_cleaned) return; // to avoid double cleaning by java and C
270 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
271 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
273 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
274 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
275 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
277 /* Kill all processes (but maestro) */
278 SIMIX_process_killall(simix_global->maestro_process, 1);
280 /* Exit the SIMIX network module */
281 SIMIX_network_exit();
283 xbt_heap_free(simix_timers);
285 /* Free the remaining data structures */
286 xbt_dynar_free(&simix_global->process_to_run);
287 xbt_dynar_free(&simix_global->process_that_ran);
288 xbt_swag_free(simix_global->process_to_destroy);
289 xbt_swag_free(simix_global->process_list);
290 simix_global->process_list = NULL;
291 simix_global->process_to_destroy = NULL;
292 xbt_dict_free(&(simix_global->registered_functions));
294 xbt_os_mutex_destroy(simix_global->mutex);
295 simix_global->mutex = NULL;
297 /* Let's free maestro now */
298 SIMIX_context_free(simix_global->maestro_process->context);
299 xbt_free(simix_global->maestro_process->running_ctx);
300 xbt_free(simix_global->maestro_process);
301 simix_global->maestro_process = NULL;
303 /* Restore the default exception setup */
304 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
305 __xbt_ex_terminate = &__xbt_ex_terminate_default;
307 /* Finish context module and SURF */
308 SIMIX_context_mod_exit();
312 #ifdef TIME_BENCH_AMDAHL
313 xbt_os_cputimer_stop(simix_global->timer_seq);
314 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
315 xbt_os_timer_elapsed(simix_global->timer_seq),
316 xbt_os_timer_elapsed(simix_global->timer_par));
317 xbt_os_timer_free(simix_global->timer_seq);
318 xbt_os_timer_free(simix_global->timer_par);
321 xbt_mallocator_free(simix_global->synchro_mallocator);
322 xbt_free(simix_global);
331 * \brief A clock (in second).
333 * \return Return the clock.
335 double SIMIX_get_clock(void)
337 if(MC_is_active() || MC_record_replay_is_active()){
338 return MC_process_clock_get(SIMIX_process_self());
340 return surf_get_clock();
344 static int process_syscall_color(void *p)
346 switch ((*(smx_process_t *)p)->simcall.call) {
348 case SIMCALL_PROCESS_KILL:
350 case SIMCALL_PROCESS_RESUME:
359 * \brief Run the main simulation loop.
364 MC_record_replay_init();
365 MC_record_replay_from_string(MC_record_path);
370 smx_process_t process;
371 surf_action_t action;
377 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
378 xbt_dynar_length(simix_global->process_to_run));
379 #ifdef TIME_BENCH_PER_SR
380 smx_ctx_raw_new_sr();
382 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
383 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
384 xbt_dynar_length(simix_global->process_to_run));
386 /* Run all processes that are ready to run, possibly in parallel */
387 #ifdef TIME_BENCH_AMDAHL
388 xbt_os_cputimer_stop(simix_global->timer_seq);
389 xbt_os_cputimer_resume(simix_global->timer_par);
391 SIMIX_process_runall();
392 #ifdef TIME_BENCH_AMDAHL
393 xbt_os_cputimer_stop(simix_global->timer_par);
394 xbt_os_cputimer_resume(simix_global->timer_seq);
397 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
398 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
400 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
402 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
404 /* Here, the order is ok because:
406 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
408 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
410 * - if there is no kill during the simulation, processes remain sorted according by their PID.
411 * rational: This can be proved inductively.
412 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
413 * Let's show that it is still so at the end of this round.
414 * - if a process is added when being created, that's from maestro. It can be either at startup
415 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
416 * in arbitrary order (inductive hypothesis), we are fine.
417 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
418 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
419 * of a simcall terminates, we're still good. Proof:
420 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
421 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
422 * - For communications, this function is called from SIMIX_comm_finish().
423 * This function itself don't mess with the order since simcalls are handled in FIFO order.
424 * The function is called:
425 * - before the comm starts (invalid parameters, or resource already dead or whatever).
426 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
427 * - because the communication failed or were canceled after startup. In this case, it's called from the function
428 * we are in, by the chunk:
429 * set = model->states.failed_action_set;
430 * while ((synchro = xbt_swag_extract(set)))
431 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
432 * This order is also fixed because it depends of the order in which the surf actions were
433 * added to the system, and only maestro can add stuff this way, through simcalls.
434 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
435 * poped out of the swag does not depend on the user code's execution order.
436 * - because the communication terminated. In this case, synchros are served in the order given by
437 * set = model->states.done_action_set;
438 * while ((synchro = xbt_swag_extract(set)))
439 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
440 * and the argument is very similar to the previous one.
441 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
442 * 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.
443 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
444 * - If there is some process killings, the order is changed by this decision that comes from user-land
445 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
446 * So, even the order change induced by the process killing is perfectly reproducible.
448 * So science works, bitches [http://xkcd.com/54/].
450 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
451 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
452 * That would thus be a pure waste of time.
455 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
456 if (process->simcall.call != SIMCALL_NONE) {
457 SIMIX_simcall_handle(&process->simcall, 0);
460 /* Wake up all processes waiting for a Surf action to finish */
461 xbt_dynar_foreach(all_existing_models, iter, model) {
462 XBT_DEBUG("Handling process whose action failed");
463 while ((action = surf_model_extract_failed_action_set(model))) {
464 XBT_DEBUG(" Handling Action %p",action);
465 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
467 XBT_DEBUG("Handling process whose action terminated normally");
468 while ((action = surf_model_extract_done_action_set(model))) {
469 XBT_DEBUG(" Handling Action %p",action);
470 if (surf_action_get_data(action) == NULL)
471 XBT_DEBUG("probably vcpu's action %p, skip", action);
473 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
478 time = SIMIX_timer_next();
479 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
480 XBT_DEBUG("Calling surf_solve");
481 time = surf_solve(time);
482 XBT_DEBUG("Moving time ahead : %g", time);
484 /* Notify all the hosts that have failed */
485 /* FIXME: iterate through the list of failed host and mark each of them */
486 /* as failed. On each host, signal all the running processes with host_fail */
488 /* Handle any pending timer */
489 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
490 //FIXME: make the timers being real callbacks
491 // (i.e. provide dispatchers that read and expand the args)
492 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
494 timer->func(timer->args);
498 /* Wake up all processes waiting for a Surf action to finish */
499 xbt_dynar_foreach(all_existing_models, iter, model) {
500 XBT_DEBUG("Handling process whose action failed");
501 while ((action = surf_model_extract_failed_action_set(model))) {
502 XBT_DEBUG(" Handling Action %p",action);
503 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
505 XBT_DEBUG("Handling process whose action terminated normally");
506 while ((action = surf_model_extract_done_action_set(model))) {
507 XBT_DEBUG(" Handling Action %p",action);
508 if (surf_action_get_data(action) == NULL)
509 XBT_DEBUG("probably vcpu's action %p, skip", action);
511 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
515 /* Autorestart all process */
516 if(host_that_restart) {
517 char *hostname = NULL;
518 xbt_dynar_foreach(host_that_restart,iter,hostname) {
519 XBT_INFO("Restart processes on host: %s",hostname);
520 SIMIX_host_autorestart(sg_host_by_name(hostname));
522 xbt_dynar_reset(host_that_restart);
525 /* Clean processes to destroy */
526 SIMIX_process_empty_trash();
529 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
531 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
533 if (xbt_swag_size(simix_global->process_list) != 0) {
537 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
538 SIMIX_display_process_status();
544 * \brief Set the date to execute a function
546 * Set the date to execute the function on the surf.
547 * \param date Date to execute function
548 * \param function Function to be executed
549 * \param arg Parameters of the function
552 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
554 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
557 timer->func = function;
559 xbt_heap_push(simix_timers, timer, date);
562 /** @brief cancels a timer that was added earlier */
563 void SIMIX_timer_remove(smx_timer_t timer) {
564 xbt_heap_rm_elm(simix_timers, timer, timer->date);
567 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
568 double SIMIX_timer_get_date(smx_timer_t timer) {
569 return timer?timer->date:0;
573 * \brief Registers a function to create a process.
575 * This function registers a function to be called
576 * when a new process is created. The function has
577 * to call SIMIX_process_create().
578 * \param function create process function
580 void SIMIX_function_register_process_create(smx_creation_func_t
583 simix_global->create_process_function = function;
587 * \brief Registers a function to kill a process.
589 * This function registers a function to be called when a
590 * process is killed. The function has to call the SIMIX_process_kill().
592 * \param function Kill process function
594 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
597 simix_global->kill_process_function = function;
601 * \brief Registers a function to cleanup a process.
603 * This function registers a user function to be called when
604 * a process ends properly.
606 * \param function cleanup process function
608 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
611 simix_global->cleanup_process_function = function;
615 void SIMIX_display_process_status(void)
617 if (simix_global->process_list == NULL) {
621 smx_process_t process = NULL;
622 int nbprocess = xbt_swag_size(simix_global->process_list);
624 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
625 /* List the process and their state */
627 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
628 xbt_swag_foreach(process, simix_global->process_list) {
630 if (process->waiting_synchro) {
632 const char* synchro_description = "unknown";
633 switch (process->waiting_synchro->type) {
635 case SIMIX_SYNC_EXECUTE:
636 synchro_description = "execution";
639 case SIMIX_SYNC_PARALLEL_EXECUTE:
640 synchro_description = "parallel execution";
643 case SIMIX_SYNC_COMMUNICATE:
644 synchro_description = "communication";
647 case SIMIX_SYNC_SLEEP:
648 synchro_description = "sleeping";
651 case SIMIX_SYNC_JOIN:
652 synchro_description = "joining";
655 case SIMIX_SYNC_SYNCHRO:
656 synchro_description = "synchronization";
660 synchro_description = "I/O";
663 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
664 process->pid, process->name, sg_host_get_name(process->host),
665 synchro_description, process->waiting_synchro,
666 process->waiting_synchro->name, (int)process->waiting_synchro->state);
669 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
674 static void* SIMIX_synchro_mallocator_new_f(void) {
675 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
676 synchro->simcalls = xbt_fifo_new();
680 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
681 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
685 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
687 // we also recycle the simcall list
688 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
689 xbt_fifo_reset(fifo);
690 memset(synchro, 0, sizeof(s_smx_synchro_t));
691 ((smx_synchro_t) synchro)->simcalls = fifo;
694 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
695 return SIMIX_asr_get_properties(name);
697 xbt_dict_t SIMIX_asr_get_properties(const char *name)
699 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);