1 /* Copyright (c) 2007-2014. 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_model_checker.h"
22 #include "mc/mc_protocol.h"
23 #include "mc/mc_client.h"
25 #include "mc/mc_record.h"
28 #include "smpi/private.h"
31 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
32 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
33 "Logging specific to SIMIX (kernel)");
35 smx_global_t simix_global = NULL;
36 static xbt_heap_t simix_timers = NULL;
38 static void* SIMIX_synchro_mallocator_new_f(void);
39 static void SIMIX_synchro_mallocator_free_f(void* synchro);
40 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
42 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
45 int _sg_do_verbose_exit = 1;
46 static void _XBT_CALL inthandler(int ignored)
48 if ( _sg_do_verbose_exit ) {
49 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
50 SIMIX_display_process_status();
53 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
59 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
61 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
63 "Access violation detected.\n"
64 "This can result from a programming error in your code or, although less likely,\n"
65 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
66 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
68 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
69 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
70 smx_context_stack_size / 1024);
71 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
73 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
74 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
76 } else if (siginfo->si_signo == SIGSEGV) {
77 fprintf(stderr, "Segmentation fault.\n");
79 if (smpi_enabled() && !smpi_privatize_global_variables) {
81 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
88 MC_dump_stack_safety(mc_stack);
90 MC_print_statistics(mc_stats);
96 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
99 * Install signal handler for SIGSEGV. Check that nobody has already installed
100 * its own handler. For example, the Java VM does this.
102 static void install_segvhandler(void)
104 stack_t stack, old_stack;
105 stack.ss_sp = sigsegv_stack;
106 stack.ss_size = sizeof sigsegv_stack;
109 if (sigaltstack(&stack, &old_stack) == -1) {
110 XBT_WARN("Failed to register alternate signal stack: %s",
114 if (!(old_stack.ss_flags & SS_DISABLE)) {
115 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
116 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
117 sigaltstack(&old_stack, NULL);
120 struct sigaction action, old_action;
121 action.sa_sigaction = segvhandler;
122 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
123 sigemptyset(&action.sa_mask);
125 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
126 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
130 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
131 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
132 (old_action.sa_flags & SA_SIGINFO) ?
133 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
134 sigaction(SIGSEGV, &old_action, NULL);
139 /********************************* SIMIX **************************************/
141 XBT_INLINE double SIMIX_timer_next(void)
143 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
148 * \brief Initialize SIMIX internal data.
153 void SIMIX_global_init(int *argc, char **argv)
155 s_smx_process_t proc;
158 simix_global = xbt_new0(s_smx_global_t, 1);
160 #ifdef TIME_BENCH_AMDAHL
161 simix_global->timer_seq = xbt_os_timer_new();
162 simix_global->timer_par = xbt_os_timer_new();
163 xbt_os_cputimer_start(simix_global->timer_seq);
165 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
166 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
167 simix_global->process_list =
168 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
169 simix_global->process_to_destroy =
170 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
172 simix_global->maestro_process = NULL;
173 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
175 simix_global->create_process_function = SIMIX_process_create;
176 simix_global->kill_process_function = SIMIX_process_kill;
177 simix_global->cleanup_process_function = SIMIX_process_cleanup;
178 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
179 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
180 SIMIX_synchro_mallocator_reset_f);
181 simix_global->autorestart = SIMIX_host_restart_processes;
183 surf_init(argc, argv); /* Initialize SURF structures */
184 SIMIX_context_mod_init();
185 SIMIX_create_maestro_process();
187 /* context exception handlers */
188 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
189 __xbt_ex_terminate = SIMIX_process_exception_terminate;
191 SIMIX_network_init();
193 /* Prepare to display some more info when dying on Ctrl-C pressing */
194 signal(SIGINT, inthandler);
197 /* Install SEGV handler */
198 install_segvhandler();
200 /* register a function to be called by SURF after the environment creation */
202 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
206 simix_timers = xbt_heap_new(8, &free);
209 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
210 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
212 if (sg_cfg_get_boolean("clean_atexit"))
216 // The communication initialisation is done ASAP.
217 // We need to commuicate initialisation of the different layers to the model-checker.
218 if (mc_mode == MC_MODE_NONE) {
219 if (getenv(MC_ENV_SOCKET_FD)) {
220 mc_mode = MC_MODE_CLIENT;
223 MC_client_handle_messages();
228 if (_sg_cfg_exit_asap)
234 * \brief Clean the SIMIX simulation
236 * This functions remove the memory used by SIMIX
238 void SIMIX_clean(void)
240 static int cleaned = 0;
241 #ifdef TIME_BENCH_PER_SR
242 smx_ctx_raw_new_sr();
244 if (cleaned) return; // to avoid double cleaning by java and C
246 /* Kill everyone (except maestro) */
247 SIMIX_process_killall(simix_global->maestro_process, 1);
249 /* Exit the SIMIX network module */
250 SIMIX_network_exit();
252 xbt_heap_free(simix_timers);
254 /* Free the remaining data structures */
255 xbt_dynar_free(&simix_global->process_to_run);
256 xbt_dynar_free(&simix_global->process_that_ran);
257 xbt_swag_free(simix_global->process_to_destroy);
258 xbt_swag_free(simix_global->process_list);
259 simix_global->process_list = NULL;
260 simix_global->process_to_destroy = NULL;
261 xbt_dict_free(&(simix_global->registered_functions));
263 /* Let's free maestro now */
264 SIMIX_context_free(simix_global->maestro_process->context);
265 xbt_free(simix_global->maestro_process->running_ctx);
266 xbt_free(simix_global->maestro_process);
267 simix_global->maestro_process = NULL;
269 /* Restore the default exception setup */
270 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
271 __xbt_ex_terminate = &__xbt_ex_terminate_default;
273 /* Finish context module and SURF */
274 SIMIX_context_mod_exit();
278 #ifdef TIME_BENCH_AMDAHL
279 xbt_os_cputimer_stop(simix_global->timer_seq);
280 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
281 xbt_os_timer_elapsed(simix_global->timer_seq),
282 xbt_os_timer_elapsed(simix_global->timer_par));
283 xbt_os_timer_free(simix_global->timer_seq);
284 xbt_os_timer_free(simix_global->timer_par);
287 xbt_mallocator_free(simix_global->synchro_mallocator);
288 xbt_free(simix_global);
297 * \brief A clock (in second).
299 * \return Return the clock.
301 XBT_INLINE double SIMIX_get_clock(void)
303 if(MC_is_active() || MC_record_replay_is_active()){
304 return MC_process_clock_get(SIMIX_process_self());
306 return surf_get_clock();
310 static int process_syscall_color(void *p)
312 switch ((*(smx_process_t *)p)->simcall.call) {
314 case SIMCALL_PROCESS_KILL:
316 case SIMCALL_PROCESS_RESUME:
325 * \brief Run the main simulation loop.
330 MC_record_replay_init();
331 MC_record_replay_from_string(MC_record_path);
336 smx_process_t process;
337 surf_action_t action;
343 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
344 xbt_dynar_length(simix_global->process_to_run));
345 #ifdef TIME_BENCH_PER_SR
346 smx_ctx_raw_new_sr();
348 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
349 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
350 xbt_dynar_length(simix_global->process_to_run));
352 /* Run all processes that are ready to run, possibly in parallel */
353 #ifdef TIME_BENCH_AMDAHL
354 xbt_os_cputimer_stop(simix_global->timer_seq);
355 xbt_os_cputimer_resume(simix_global->timer_par);
357 SIMIX_process_runall();
358 #ifdef TIME_BENCH_AMDAHL
359 xbt_os_cputimer_stop(simix_global->timer_par);
360 xbt_os_cputimer_resume(simix_global->timer_seq);
363 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
364 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
366 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
368 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
370 /* Here, the order is ok because:
372 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
374 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
376 * - if there is no kill during the simulation, processes remain sorted according by their PID.
377 * rational: This can be proved inductively.
378 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
379 * Let's show that it is still so at the end of this round.
380 * - if a process is added when being created, that's from maestro. It can be either at startup
381 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
382 * in arbitrary order (inductive hypothesis), we are fine.
383 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
384 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
385 * of a simcall terminates, we're still good. Proof:
386 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
387 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
388 * - For communications, this function is called from SIMIX_comm_finish().
389 * This function itself don't mess with the order since simcalls are handled in FIFO order.
390 * The function is called:
391 * - before the comm starts (invalid parameters, or resource already dead or whatever).
392 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
393 * - because the communication failed or were canceled after startup. In this case, it's called from the function
394 * we are in, by the chunk:
395 * set = model->states.failed_action_set;
396 * while ((synchro = xbt_swag_extract(set)))
397 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
398 * This order is also fixed because it depends of the order in which the surf actions were
399 * added to the system, and only maestro can add stuff this way, through simcalls.
400 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
401 * poped out of the swag does not depend on the user code's execution order.
402 * - because the communication terminated. In this case, synchros are served in the order given by
403 * set = model->states.done_action_set;
404 * while ((synchro = xbt_swag_extract(set)))
405 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
406 * and the argument is very similar to the previous one.
407 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
408 * 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.
409 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
410 * - If there is some process killings, the order is changed by this decision that comes from user-land
411 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
412 * So, even the order change induced by the process killing is perfectly reproducible.
414 * So science works, bitches [http://xkcd.com/54/].
416 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
417 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
418 * That would thus be a pure waste of time.
421 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
422 if (process->simcall.call != SIMCALL_NONE) {
423 SIMIX_simcall_handle(&process->simcall, 0);
426 /* Wake up all processes waiting for a Surf action to finish */
427 xbt_dynar_foreach(model_list, iter, model) {
428 XBT_DEBUG("Handling process whose action failed");
429 while ((action = surf_model_extract_failed_action_set(model))) {
430 XBT_DEBUG(" Handling Action %p",action);
431 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
433 XBT_DEBUG("Handling process whose action terminated normally");
434 while ((action = surf_model_extract_done_action_set(model))) {
435 XBT_DEBUG(" Handling Action %p",action);
436 if (surf_action_get_data(action) == NULL)
437 XBT_DEBUG("probably vcpu's action %p, skip", action);
439 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
444 time = SIMIX_timer_next();
445 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
446 XBT_DEBUG("Calling surf_solve");
447 time = surf_solve(time);
448 XBT_DEBUG("Moving time ahead : %g", time);
450 /* Notify all the hosts that have failed */
451 /* FIXME: iterate through the list of failed host and mark each of them */
452 /* as failed. On each host, signal all the running processes with host_fail */
454 /* Handle any pending timer */
455 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
456 //FIXME: make the timers being real callbacks
457 // (i.e. provide dispatchers that read and expand the args)
458 timer = xbt_heap_pop(simix_timers);
460 ((void (*)(void*))timer->func)(timer->args);
464 /* Wake up all processes waiting for a Surf action to finish */
465 xbt_dynar_foreach(model_list, 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) surf_action_get_data(action));
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 (surf_action_get_data(action) == NULL)
475 XBT_DEBUG("probably vcpu's action %p, skip", action);
477 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
481 /* Autorestart all process */
482 if(host_that_restart) {
483 char *hostname = NULL;
484 xbt_dynar_foreach(host_that_restart,iter,hostname) {
485 XBT_INFO("Restart processes on host: %s",hostname);
486 SIMIX_host_autorestart(SIMIX_host_get_by_name(hostname));
488 xbt_dynar_reset(host_that_restart);
491 /* Clean processes to destroy */
492 SIMIX_process_empty_trash();
495 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
497 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
499 if (xbt_swag_size(simix_global->process_list) != 0) {
503 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
504 SIMIX_display_process_status();
510 * \brief Set the date to execute a function
512 * Set the date to execute the function on the surf.
513 * \param date Date to execute function
514 * \param function Function to be executed
515 * \param arg Parameters of the function
518 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
520 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
523 timer->func = function;
525 xbt_heap_push(simix_timers, timer, date);
529 * \brief Registers a function to create a process.
531 * This function registers a function to be called
532 * when a new process is created. The function has
533 * to call SIMIX_process_create().
534 * \param function create process function
536 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
539 simix_global->create_process_function = function;
543 * \brief Registers a function to kill a process.
545 * This function registers a function to be called when a
546 * process is killed. The function has to call the SIMIX_process_kill().
548 * \param function Kill process function
550 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
553 simix_global->kill_process_function = function;
557 * \brief Registers a function to cleanup a process.
559 * This function registers a user function to be called when
560 * a process ends properly.
562 * \param function cleanup process function
564 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
567 simix_global->cleanup_process_function = function;
571 void SIMIX_display_process_status(void)
573 if (simix_global->process_list == NULL) {
577 smx_process_t process = NULL;
578 int nbprocess = xbt_swag_size(simix_global->process_list);
580 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
581 /* List the process and their state */
583 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
584 xbt_swag_foreach(process, simix_global->process_list) {
586 if (process->waiting_synchro) {
588 const char* synchro_description = "unknown";
589 switch (process->waiting_synchro->type) {
591 case SIMIX_SYNC_EXECUTE:
592 synchro_description = "execution";
595 case SIMIX_SYNC_PARALLEL_EXECUTE:
596 synchro_description = "parallel execution";
599 case SIMIX_SYNC_COMMUNICATE:
600 synchro_description = "communication";
603 case SIMIX_SYNC_SLEEP:
604 synchro_description = "sleeping";
607 case SIMIX_SYNC_JOIN:
608 synchro_description = "joining";
611 case SIMIX_SYNC_SYNCHRO:
612 synchro_description = "synchronization";
616 synchro_description = "I/O";
619 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
620 process->pid, process->name, sg_host_name(process->smx_host),
621 synchro_description, process->waiting_synchro,
622 process->waiting_synchro->name, (int)process->waiting_synchro->state);
625 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host));
630 static void* SIMIX_synchro_mallocator_new_f(void) {
631 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
632 synchro->simcalls = xbt_fifo_new();
636 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
637 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
641 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
643 // we also recycle the simcall list
644 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
645 xbt_fifo_reset(fifo);
646 memset(synchro, 0, sizeof(s_smx_synchro_t));
647 ((smx_synchro_t) synchro)->simcalls = fifo;
650 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
651 return SIMIX_asr_get_properties(name);
653 xbt_dict_t SIMIX_asr_get_properties(const char *name)
655 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);