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. */
7 #include "smx_private.h"
9 #include "xbt/sysdep.h"
12 #include "xbt/ex.h" /* ex_backtrace_display */
14 #include "simgrid/sg_config.h"
17 #include "mc/mc_private.h"
21 #include "smpi/private.h"
24 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
25 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
26 "Logging specific to SIMIX (kernel)");
28 smx_global_t simix_global = NULL;
29 static xbt_heap_t simix_timers = NULL;
31 static void* SIMIX_action_mallocator_new_f(void);
32 static void SIMIX_action_mallocator_free_f(void* action);
33 static void SIMIX_action_mallocator_reset_f(void* action);
35 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
38 int _sg_do_verbose_exit = 1;
39 static void _XBT_CALL inthandler(int ignored)
41 if ( _sg_do_verbose_exit ) {
42 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
43 SIMIX_display_process_status();
46 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
52 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
54 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
56 "Access violation detected.\n"
57 "This can result from a programming error in your code or, although less likely,\n"
58 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
59 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
61 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
62 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
63 smx_context_stack_size / 1024);
64 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
66 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
67 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
69 } else if (siginfo->si_signo == SIGSEGV) {
70 fprintf(stderr, "Segmentation fault.\n");
72 if (smpi_enabled() && !smpi_privatize_global_variables) {
74 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
81 MC_dump_stack_safety(mc_stack);
83 MC_print_statistics(mc_stats);
89 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
92 * Install signal handler for SIGSEGV. Check that nobody has already installed
93 * its own handler. For example, the Java VM does this.
95 static void install_segvhandler(void)
97 stack_t stack, old_stack;
98 stack.ss_sp = sigsegv_stack;
99 stack.ss_size = sizeof sigsegv_stack;
102 if (sigaltstack(&stack, &old_stack) == -1) {
103 XBT_WARN("Failed to register alternate signal stack: %s",
107 if (!(old_stack.ss_flags & SS_DISABLE)) {
108 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
109 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
110 sigaltstack(&old_stack, NULL);
113 struct sigaction action, old_action;
114 action.sa_sigaction = segvhandler;
115 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
116 sigemptyset(&action.sa_mask);
118 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
119 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
123 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
124 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
125 (old_action.sa_flags & SA_SIGINFO) ?
126 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
127 sigaction(SIGSEGV, &old_action, NULL);
132 /********************************* SIMIX **************************************/
134 XBT_INLINE double SIMIX_timer_next(void)
136 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
141 * \brief Initialize SIMIX internal data.
146 void SIMIX_global_init(int *argc, char **argv)
148 s_smx_process_t proc;
151 simix_global = xbt_new0(s_smx_global_t, 1);
153 #ifdef TIME_BENCH_AMDAHL
154 simix_global->timer_seq = xbt_os_timer_new();
155 simix_global->timer_par = xbt_os_timer_new();
156 xbt_os_cputimer_start(simix_global->timer_seq);
158 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
159 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
160 simix_global->process_list =
161 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
162 simix_global->process_to_destroy =
163 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
165 simix_global->maestro_process = NULL;
166 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
168 simix_global->create_process_function = SIMIX_process_create;
169 simix_global->kill_process_function = SIMIX_process_kill;
170 simix_global->cleanup_process_function = SIMIX_process_cleanup;
171 simix_global->action_mallocator = xbt_mallocator_new(65536,
172 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
173 SIMIX_action_mallocator_reset_f);
174 simix_global->autorestart = SIMIX_host_restart_processes;
176 surf_init(argc, argv); /* Initialize SURF structures */
177 SIMIX_context_mod_init();
178 SIMIX_create_maestro_process();
180 /* context exception handlers */
181 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
182 __xbt_ex_terminate = SIMIX_process_exception_terminate;
184 SIMIX_network_init();
186 /* Prepare to display some more info when dying on Ctrl-C pressing */
187 signal(SIGINT, inthandler);
190 /* Install SEGV handler */
191 install_segvhandler();
193 /* register a function to be called by SURF after the environment creation */
195 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
199 simix_timers = xbt_heap_new(8, &free);
202 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
203 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
205 if (sg_cfg_get_boolean("clean_atexit"))
208 if (_sg_cfg_exit_asap)
214 * \brief Clean the SIMIX simulation
216 * This functions remove the memory used by SIMIX
218 void SIMIX_clean(void)
220 static int cleaned = 0;
221 #ifdef TIME_BENCH_PER_SR
222 smx_ctx_raw_new_sr();
224 if (cleaned) return; // to avoid double cleaning by java and C
226 /* Kill everyone (except maestro) */
227 SIMIX_process_killall(simix_global->maestro_process, 1);
229 /* Exit the SIMIX network module */
230 SIMIX_network_exit();
232 xbt_heap_free(simix_timers);
234 /* Free the remaining data structures */
235 xbt_dynar_free(&simix_global->process_to_run);
236 xbt_dynar_free(&simix_global->process_that_ran);
237 xbt_swag_free(simix_global->process_to_destroy);
238 xbt_swag_free(simix_global->process_list);
239 simix_global->process_list = NULL;
240 simix_global->process_to_destroy = NULL;
241 xbt_dict_free(&(simix_global->registered_functions));
243 /* Let's free maestro now */
244 SIMIX_context_free(simix_global->maestro_process->context);
245 xbt_free(simix_global->maestro_process->running_ctx);
246 xbt_free(simix_global->maestro_process);
247 simix_global->maestro_process = NULL;
249 /* Restore the default exception setup */
250 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
251 __xbt_ex_terminate = &__xbt_ex_terminate_default;
253 /* Finish context module and SURF */
254 SIMIX_context_mod_exit();
258 #ifdef TIME_BENCH_AMDAHL
259 xbt_os_cputimer_stop(simix_global->timer_seq);
260 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
261 xbt_os_timer_elapsed(simix_global->timer_seq),
262 xbt_os_timer_elapsed(simix_global->timer_par));
263 xbt_os_timer_free(simix_global->timer_seq);
264 xbt_os_timer_free(simix_global->timer_par);
267 xbt_mallocator_free(simix_global->action_mallocator);
268 xbt_free(simix_global);
277 * \brief A clock (in second).
279 * \return Return the clock.
281 XBT_INLINE double SIMIX_get_clock(void)
284 return MC_process_clock_get(SIMIX_process_self());
286 return surf_get_clock();
290 static int process_syscall_color(void *p)
292 switch ((*(smx_process_t *)p)->simcall.call) {
294 case SIMCALL_PROCESS_KILL:
296 case SIMCALL_PROCESS_RESUME:
305 * \brief Run the main simulation loop.
310 smx_process_t process;
311 surf_action_t action;
317 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
318 xbt_dynar_length(simix_global->process_to_run));
319 #ifdef TIME_BENCH_PER_SR
320 smx_ctx_raw_new_sr();
322 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
323 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
324 xbt_dynar_length(simix_global->process_to_run));
326 /* Run all processes that are ready to run, possibly in parallel */
327 #ifdef TIME_BENCH_AMDAHL
328 xbt_os_cputimer_stop(simix_global->timer_seq);
329 xbt_os_cputimer_resume(simix_global->timer_par);
331 SIMIX_process_runall();
332 #ifdef TIME_BENCH_AMDAHL
333 xbt_os_cputimer_stop(simix_global->timer_par);
334 xbt_os_cputimer_resume(simix_global->timer_seq);
337 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
338 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
340 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
342 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
344 /* Here, the order is ok because:
346 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
348 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
350 * - if there is no kill during the simulation, processes remain sorted according by their PID.
351 * rational: This can be proved inductively.
352 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
353 * Let's show that it is still so at the end of this round.
354 * - if a process is added when being created, that's from maestro. It can be either at startup
355 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
356 * in arbitrary order (inductive hypothesis), we are fine.
357 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
358 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
359 * of a simcall terminates, we're still good. Proof:
360 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
361 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
362 * - For communications, this function is called from SIMIX_comm_finish().
363 * This function itself don't mess with the order since simcalls are handled in FIFO order.
364 * The function is called:
365 * - before the comm starts (invalid parameters, or resource already dead or whatever).
366 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
367 * - because the communication failed or were canceled after startup. In this case, it's called from the function
368 * we are in, by the chunk:
369 * set = model->states.failed_action_set;
370 * while ((action = xbt_swag_extract(set)))
371 * SIMIX_simcall_post((smx_action_t) action->data);
372 * This order is also fixed because it depends of the order in which the surf actions were
373 * added to the system, and only maestro can add stuff this way, through simcalls.
374 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
375 * poped out of the swag does not depend on the user code's execution order.
376 * - because the communication terminated. In this case, actions are served in the order given by
377 * set = model->states.done_action_set;
378 * while ((action = xbt_swag_extract(set)))
379 * SIMIX_simcall_post((smx_action_t) action->data);
380 * and the argument is very similar to the previous one.
381 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
382 * 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.
383 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
384 * - If there is some process killings, the order is changed by this decision that comes from user-land
385 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
386 * So, even the order change induced by the process killing is perfectly reproducible.
388 * So science works, bitches [http://xkcd.com/54/].
390 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
391 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
392 * That would thus be a pure waste of time.
395 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
396 if (process->simcall.call != SIMCALL_NONE) {
397 SIMIX_simcall_enter(&process->simcall, 0);
402 time = SIMIX_timer_next();
403 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
404 time = surf_solve(time);
406 /* Notify all the hosts that have failed */
407 /* FIXME: iterate through the list of failed host and mark each of them */
408 /* as failed. On each host, signal all the running processes with host_fail */
410 /* Handle any pending timer */
411 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
412 //FIXME: make the timers being real callbacks
413 // (i.e. provide dispatchers that read and expand the args)
414 timer = xbt_heap_pop(simix_timers);
416 ((void (*)(void*))timer->func)(timer->args);
420 /* Wake up all processes waiting for a Surf action to finish */
421 xbt_dynar_foreach(model_list, iter, model) {
422 while ((action = surf_model_extract_failed_action_set(model)))
423 SIMIX_simcall_exit((smx_action_t) surf_action_get_data(action));
425 while ((action = surf_model_extract_done_action_set(model)))
426 if (surf_action_get_data(action) == NULL)
427 XBT_DEBUG("probably vcpu's action %p, skip", action);
429 SIMIX_simcall_exit((smx_action_t) surf_action_get_data(action));
432 /* Autorestart all process */
433 if(host_that_restart) {
434 char *hostname = NULL;
435 xbt_dynar_foreach(host_that_restart,iter,hostname) {
436 XBT_INFO("Restart processes on host: %s",hostname);
437 SIMIX_host_autorestart(SIMIX_host_get_by_name(hostname));
439 xbt_dynar_reset(host_that_restart);
442 /* Clean processes to destroy */
443 SIMIX_process_empty_trash();
446 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
447 // !(time == -1.0 && xbt_dynar_is_empty())
450 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
452 if (xbt_swag_size(simix_global->process_list) != 0) {
458 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
459 SIMIX_display_process_status();
465 * \brief Set the date to execute a function
467 * Set the date to execute the function on the surf.
468 * \param date Date to execute function
469 * \param function Function to be executed
470 * \param arg Parameters of the function
473 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
475 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
478 timer->func = function;
480 xbt_heap_push(simix_timers, timer, date);
484 * \brief Registers a function to create a process.
486 * This function registers a function to be called
487 * when a new process is created. The function has
488 * to call SIMIX_process_create().
489 * \param function create process function
491 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
494 simix_global->create_process_function = function;
498 * \brief Registers a function to kill a process.
500 * This function registers a function to be called when a
501 * process is killed. The function has to call the SIMIX_process_kill().
503 * \param function Kill process function
505 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
508 simix_global->kill_process_function = function;
512 * \brief Registers a function to cleanup a process.
514 * This function registers a user function to be called when
515 * a process ends properly.
517 * \param function cleanup process function
519 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
522 simix_global->cleanup_process_function = function;
526 void SIMIX_display_process_status(void)
528 if (simix_global->process_list == NULL) {
532 smx_process_t process = NULL;
533 int nbprocess = xbt_swag_size(simix_global->process_list);
535 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
536 /* List the process and their state */
538 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
539 xbt_swag_foreach(process, simix_global->process_list) {
541 if (process->waiting_action) {
543 const char* action_description = "unknown";
544 switch (process->waiting_action->type) {
546 case SIMIX_ACTION_EXECUTE:
547 action_description = "execution";
550 case SIMIX_ACTION_PARALLEL_EXECUTE:
551 action_description = "parallel execution";
554 case SIMIX_ACTION_COMMUNICATE:
555 action_description = "communication";
558 case SIMIX_ACTION_SLEEP:
559 action_description = "sleeping";
562 case SIMIX_ACTION_JOIN:
563 action_description = "joining";
566 case SIMIX_ACTION_SYNCHRO:
567 action_description = "synchronization";
570 case SIMIX_ACTION_IO:
571 action_description = "I/O";
573 /* **************************************/
574 /* TUTORIAL: New API */
575 case SIMIX_ACTION_NEW_API:
576 action_description = "NEW API";
577 /* **************************************/
581 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
582 process->pid, process->name, sg_host_name(process->smx_host),
583 action_description, process->waiting_action,
584 process->waiting_action->name, (int)process->waiting_action->state);
587 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host));
592 static void* SIMIX_action_mallocator_new_f(void) {
593 smx_action_t action = xbt_new(s_smx_action_t, 1);
594 action->simcalls = xbt_fifo_new();
598 static void SIMIX_action_mallocator_free_f(void* action) {
599 xbt_fifo_free(((smx_action_t) action)->simcalls);
603 static void SIMIX_action_mallocator_reset_f(void* action) {
605 // we also recycle the simcall list
606 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
607 xbt_fifo_reset(fifo);
608 memset(action, 0, sizeof(s_smx_action_t));
609 ((smx_action_t) action)->simcalls = fifo;
612 xbt_dict_t SIMIX_pre_asr_get_properties(smx_simcall_t simcall, const char *name){
613 return SIMIX_asr_get_properties(name);
615 xbt_dict_t SIMIX_asr_get_properties(const char *name)
617 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);