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"
27 #include "mc/mc_protocol.h"
30 #include "mc/mc_record.h"
33 #include "smpi/private.h"
36 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
37 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
38 "Logging specific to SIMIX (kernel)");
40 smx_global_t simix_global = NULL;
41 static xbt_heap_t simix_timers = NULL;
43 static void* SIMIX_synchro_mallocator_new_f(void);
44 static void SIMIX_synchro_mallocator_free_f(void* synchro);
45 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
47 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
50 int _sg_do_verbose_exit = 1;
51 static void _XBT_CALL inthandler(int ignored)
53 if ( _sg_do_verbose_exit ) {
54 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
55 SIMIX_display_process_status();
58 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
64 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
66 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
68 "Access violation detected.\n"
69 "This can result from a programming error in your code or, although less likely,\n"
70 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
71 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
73 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
74 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
75 smx_context_stack_size / 1024);
76 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
78 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
79 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
81 } else if (siginfo->si_signo == SIGSEGV) {
82 fprintf(stderr, "Segmentation fault.\n");
84 if (smpi_enabled() && !smpi_privatize_global_variables) {
86 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
93 MC_dump_stack_safety(mc_stack);
95 MC_print_statistics(mc_stats);
101 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
104 * Install signal handler for SIGSEGV. Check that nobody has already installed
105 * its own handler. For example, the Java VM does this.
107 static void install_segvhandler(void)
109 stack_t stack, old_stack;
110 stack.ss_sp = sigsegv_stack;
111 stack.ss_size = sizeof sigsegv_stack;
114 if (sigaltstack(&stack, &old_stack) == -1) {
115 XBT_WARN("Failed to register alternate signal stack: %s",
119 if (!(old_stack.ss_flags & SS_DISABLE)) {
120 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
121 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
122 sigaltstack(&old_stack, NULL);
125 struct sigaction action, old_action;
126 action.sa_sigaction = segvhandler;
127 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
128 sigemptyset(&action.sa_mask);
130 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
131 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
135 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
136 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
137 (old_action.sa_flags & SA_SIGINFO) ?
138 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
139 sigaction(SIGSEGV, &old_action, NULL);
144 /********************************* SIMIX **************************************/
146 XBT_INLINE double SIMIX_timer_next(void)
148 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
153 * \brief Initialize SIMIX internal data.
158 void SIMIX_global_init(int *argc, char **argv)
161 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
164 s_smx_process_t proc;
167 simix_global = xbt_new0(s_smx_global_t, 1);
169 #ifdef TIME_BENCH_AMDAHL
170 simix_global->timer_seq = xbt_os_timer_new();
171 simix_global->timer_par = xbt_os_timer_new();
172 xbt_os_cputimer_start(simix_global->timer_seq);
174 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
175 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
176 simix_global->process_list =
177 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
178 simix_global->process_to_destroy =
179 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
181 simix_global->maestro_process = NULL;
182 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
184 simix_global->create_process_function = SIMIX_process_create;
185 simix_global->kill_process_function = SIMIX_process_kill;
186 simix_global->cleanup_process_function = SIMIX_process_cleanup;
187 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
188 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
189 SIMIX_synchro_mallocator_reset_f);
190 simix_global->autorestart = SIMIX_host_restart_processes;
191 simix_global->mutex = xbt_os_mutex_init();
193 surf_init(argc, argv); /* Initialize SURF structures */
194 SIMIX_context_mod_init();
195 SIMIX_create_maestro_process();
197 /* context exception handlers */
198 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
199 __xbt_ex_terminate = SIMIX_process_exception_terminate;
201 SIMIX_network_init();
203 /* Prepare to display some more info when dying on Ctrl-C pressing */
204 signal(SIGINT, inthandler);
207 /* Install SEGV handler */
208 install_segvhandler();
210 /* register a function to be called by SURF after the environment creation */
212 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
216 simix_timers = xbt_heap_new(8, &free);
219 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
221 if (sg_cfg_get_boolean("clean_atexit"))
225 // The communication initialization is done ASAP.
226 // We need to communicate initialization of the different layers to the model-checker.
227 if (mc_mode == MC_MODE_NONE) {
228 if (getenv(MC_ENV_SOCKET_FD)) {
229 mc_mode = MC_MODE_CLIENT;
232 MC_client_handle_messages();
237 if (_sg_cfg_exit_asap)
244 * \brief Clean the SIMIX simulation
246 * This functions remove the memory used by SIMIX
248 void SIMIX_clean(void)
250 #ifdef TIME_BENCH_PER_SR
251 smx_ctx_raw_new_sr();
253 if (smx_cleaned) return; // to avoid double cleaning by java and C
255 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
256 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
258 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
259 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
260 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
262 /* Kill all processes (but maestro) */
263 SIMIX_process_killall(simix_global->maestro_process, 1);
265 /* Exit the SIMIX network module */
266 SIMIX_network_exit();
268 xbt_heap_free(simix_timers);
270 /* Free the remaining data structures */
271 xbt_dynar_free(&simix_global->process_to_run);
272 xbt_dynar_free(&simix_global->process_that_ran);
273 xbt_swag_free(simix_global->process_to_destroy);
274 xbt_swag_free(simix_global->process_list);
275 simix_global->process_list = NULL;
276 simix_global->process_to_destroy = NULL;
277 xbt_dict_free(&(simix_global->registered_functions));
279 xbt_os_mutex_destroy(simix_global->mutex);
280 simix_global->mutex = NULL;
282 /* Let's free maestro now */
283 SIMIX_context_free(simix_global->maestro_process->context);
284 xbt_free(simix_global->maestro_process->running_ctx);
285 xbt_free(simix_global->maestro_process);
286 simix_global->maestro_process = NULL;
288 /* Restore the default exception setup */
289 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
290 __xbt_ex_terminate = &__xbt_ex_terminate_default;
292 /* Finish context module and SURF */
293 SIMIX_context_mod_exit();
297 #ifdef TIME_BENCH_AMDAHL
298 xbt_os_cputimer_stop(simix_global->timer_seq);
299 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
300 xbt_os_timer_elapsed(simix_global->timer_seq),
301 xbt_os_timer_elapsed(simix_global->timer_par));
302 xbt_os_timer_free(simix_global->timer_seq);
303 xbt_os_timer_free(simix_global->timer_par);
306 xbt_mallocator_free(simix_global->synchro_mallocator);
307 xbt_free(simix_global);
316 * \brief A clock (in second).
318 * \return Return the clock.
320 XBT_INLINE double SIMIX_get_clock(void)
322 if(MC_is_active() || MC_record_replay_is_active()){
323 return MC_process_clock_get(SIMIX_process_self());
325 return surf_get_clock();
329 static int process_syscall_color(void *p)
331 switch ((*(smx_process_t *)p)->simcall.call) {
333 case SIMCALL_PROCESS_KILL:
335 case SIMCALL_PROCESS_RESUME:
344 * \brief Run the main simulation loop.
349 MC_record_replay_init();
350 MC_record_replay_from_string(MC_record_path);
355 smx_process_t process;
356 surf_action_t action;
362 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
363 xbt_dynar_length(simix_global->process_to_run));
364 #ifdef TIME_BENCH_PER_SR
365 smx_ctx_raw_new_sr();
367 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
368 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
369 xbt_dynar_length(simix_global->process_to_run));
371 /* Run all processes that are ready to run, possibly in parallel */
372 #ifdef TIME_BENCH_AMDAHL
373 xbt_os_cputimer_stop(simix_global->timer_seq);
374 xbt_os_cputimer_resume(simix_global->timer_par);
376 SIMIX_process_runall();
377 #ifdef TIME_BENCH_AMDAHL
378 xbt_os_cputimer_stop(simix_global->timer_par);
379 xbt_os_cputimer_resume(simix_global->timer_seq);
382 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
383 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
385 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
387 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
389 /* Here, the order is ok because:
391 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
393 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
395 * - if there is no kill during the simulation, processes remain sorted according by their PID.
396 * rational: This can be proved inductively.
397 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
398 * Let's show that it is still so at the end of this round.
399 * - if a process is added when being created, that's from maestro. It can be either at startup
400 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
401 * in arbitrary order (inductive hypothesis), we are fine.
402 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
403 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
404 * of a simcall terminates, we're still good. Proof:
405 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
406 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
407 * - For communications, this function is called from SIMIX_comm_finish().
408 * This function itself don't mess with the order since simcalls are handled in FIFO order.
409 * The function is called:
410 * - before the comm starts (invalid parameters, or resource already dead or whatever).
411 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
412 * - because the communication failed or were canceled after startup. In this case, it's called from the function
413 * we are in, by the chunk:
414 * set = model->states.failed_action_set;
415 * while ((synchro = xbt_swag_extract(set)))
416 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
417 * This order is also fixed because it depends of the order in which the surf actions were
418 * added to the system, and only maestro can add stuff this way, through simcalls.
419 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
420 * poped out of the swag does not depend on the user code's execution order.
421 * - because the communication terminated. In this case, synchros are served in the order given by
422 * set = model->states.done_action_set;
423 * while ((synchro = xbt_swag_extract(set)))
424 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
425 * and the argument is very similar to the previous one.
426 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
427 * 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.
428 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
429 * - If there is some process killings, the order is changed by this decision that comes from user-land
430 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
431 * So, even the order change induced by the process killing is perfectly reproducible.
433 * So science works, bitches [http://xkcd.com/54/].
435 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
436 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
437 * That would thus be a pure waste of time.
440 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
441 if (process->simcall.call != SIMCALL_NONE) {
442 SIMIX_simcall_handle(&process->simcall, 0);
445 /* Wake up all processes waiting for a Surf action to finish */
446 xbt_dynar_foreach(model_list, iter, model) {
447 XBT_DEBUG("Handling process whose action failed");
448 while ((action = surf_model_extract_failed_action_set(model))) {
449 XBT_DEBUG(" Handling Action %p",action);
450 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
452 XBT_DEBUG("Handling process whose action terminated normally");
453 while ((action = surf_model_extract_done_action_set(model))) {
454 XBT_DEBUG(" Handling Action %p",action);
455 if (surf_action_get_data(action) == NULL)
456 XBT_DEBUG("probably vcpu's action %p, skip", action);
458 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
463 time = SIMIX_timer_next();
464 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
465 XBT_DEBUG("Calling surf_solve");
466 time = surf_solve(time);
467 XBT_DEBUG("Moving time ahead : %g", time);
469 /* Notify all the hosts that have failed */
470 /* FIXME: iterate through the list of failed host and mark each of them */
471 /* as failed. On each host, signal all the running processes with host_fail */
473 /* Handle any pending timer */
474 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
475 //FIXME: make the timers being real callbacks
476 // (i.e. provide dispatchers that read and expand the args)
477 timer = xbt_heap_pop(simix_timers);
479 ((void (*)(void*))timer->func)(timer->args);
483 /* Wake up all processes waiting for a Surf action to finish */
484 xbt_dynar_foreach(model_list, iter, model) {
485 XBT_DEBUG("Handling process whose action failed");
486 while ((action = surf_model_extract_failed_action_set(model))) {
487 XBT_DEBUG(" Handling Action %p",action);
488 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
490 XBT_DEBUG("Handling process whose action terminated normally");
491 while ((action = surf_model_extract_done_action_set(model))) {
492 XBT_DEBUG(" Handling Action %p",action);
493 if (surf_action_get_data(action) == NULL)
494 XBT_DEBUG("probably vcpu's action %p, skip", action);
496 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
500 /* Autorestart all process */
501 if(host_that_restart) {
502 char *hostname = NULL;
503 xbt_dynar_foreach(host_that_restart,iter,hostname) {
504 XBT_INFO("Restart processes on host: %s",hostname);
505 SIMIX_host_autorestart(sg_host_by_name(hostname));
507 xbt_dynar_reset(host_that_restart);
510 /* Clean processes to destroy */
511 SIMIX_process_empty_trash();
514 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
516 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
518 if (xbt_swag_size(simix_global->process_list) != 0) {
522 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
523 SIMIX_display_process_status();
529 * \brief Set the date to execute a function
531 * Set the date to execute the function on the surf.
532 * \param date Date to execute function
533 * \param function Function to be executed
534 * \param arg Parameters of the function
537 XBT_INLINE smx_timer_t SIMIX_timer_set(double date, void *function, void *arg)
539 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
542 timer->func = function;
544 xbt_heap_push(simix_timers, timer, date);
547 /** @brief cancels a timer that was added earlier */
548 XBT_INLINE void SIMIX_timer_remove(smx_timer_t timer) {
549 xbt_heap_rm_elm(simix_timers, timer, timer->date);
552 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
553 XBT_INLINE double SIMIX_timer_get_date(smx_timer_t timer) {
554 return timer?timer->date:0;
558 * \brief Registers a function to create a process.
560 * This function registers a function to be called
561 * when a new process is created. The function has
562 * to call SIMIX_process_create().
563 * \param function create process function
565 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
568 simix_global->create_process_function = function;
572 * \brief Registers a function to kill a process.
574 * This function registers a function to be called when a
575 * process is killed. The function has to call the SIMIX_process_kill().
577 * \param function Kill process function
579 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
582 simix_global->kill_process_function = function;
586 * \brief Registers a function to cleanup a process.
588 * This function registers a user function to be called when
589 * a process ends properly.
591 * \param function cleanup process function
593 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
596 simix_global->cleanup_process_function = function;
600 void SIMIX_display_process_status(void)
602 if (simix_global->process_list == NULL) {
606 smx_process_t process = NULL;
607 int nbprocess = xbt_swag_size(simix_global->process_list);
609 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
610 /* List the process and their state */
612 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
613 xbt_swag_foreach(process, simix_global->process_list) {
615 if (process->waiting_synchro) {
617 const char* synchro_description = "unknown";
618 switch (process->waiting_synchro->type) {
620 case SIMIX_SYNC_EXECUTE:
621 synchro_description = "execution";
624 case SIMIX_SYNC_PARALLEL_EXECUTE:
625 synchro_description = "parallel execution";
628 case SIMIX_SYNC_COMMUNICATE:
629 synchro_description = "communication";
632 case SIMIX_SYNC_SLEEP:
633 synchro_description = "sleeping";
636 case SIMIX_SYNC_JOIN:
637 synchro_description = "joining";
640 case SIMIX_SYNC_SYNCHRO:
641 synchro_description = "synchronization";
645 synchro_description = "I/O";
648 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
649 process->pid, process->name, sg_host_name(process->host),
650 synchro_description, process->waiting_synchro,
651 process->waiting_synchro->name, (int)process->waiting_synchro->state);
654 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->host));
659 static void* SIMIX_synchro_mallocator_new_f(void) {
660 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
661 synchro->simcalls = xbt_fifo_new();
665 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
666 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
670 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
672 // we also recycle the simcall list
673 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
674 xbt_fifo_reset(fifo);
675 memset(synchro, 0, sizeof(s_smx_synchro_t));
676 ((smx_synchro_t) synchro)->simcalls = fifo;
679 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
680 return SIMIX_asr_get_properties(name);
682 xbt_dict_t SIMIX_asr_get_properties(const char *name)
684 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);