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"
16 #include "xbt/sysdep.h"
19 #include "xbt/ex.h" /* ex_backtrace_display */
21 #include "src/mc/mc_replay.h"
22 #include "simgrid/sg_config.h"
24 #include "src/surf/callbacks.h"
27 #include "src/mc/mc_private.h"
28 #include "src/mc/mc_protocol.h"
29 #include "src/mc/mc_client.h"
34 #include "src/mc/mc_protocol.h"
37 #include "src/mc/mc_record.h"
40 #include "src/smpi/private.h"
43 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
44 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
45 "Logging specific to SIMIX (kernel)");
47 smx_global_t simix_global = NULL;
48 static xbt_heap_t simix_timers = NULL;
50 /** @brief Timer datatype */
51 typedef struct s_smx_timer {
57 static void* SIMIX_synchro_mallocator_new_f(void);
58 static void SIMIX_synchro_mallocator_free_f(void* synchro);
59 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
61 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
64 int _sg_do_verbose_exit = 1;
65 static void _XBT_CALL inthandler(int ignored)
67 if ( _sg_do_verbose_exit ) {
68 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
69 SIMIX_display_process_status();
72 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
78 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
80 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
82 "Access violation detected.\n"
83 "This can result from a programming error in your code or, although less likely,\n"
84 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
85 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
87 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
88 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
89 smx_context_stack_size / 1024);
90 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
92 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
93 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
95 } else if (siginfo->si_signo == SIGSEGV) {
96 fprintf(stderr, "Segmentation fault.\n");
98 if (smpi_enabled() && !smpi_privatize_global_variables) {
100 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
107 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
110 * Install signal handler for SIGSEGV. Check that nobody has already installed
111 * its own handler. For example, the Java VM does this.
113 static void install_segvhandler(void)
115 stack_t stack, old_stack;
116 stack.ss_sp = sigsegv_stack;
117 stack.ss_size = sizeof sigsegv_stack;
120 if (sigaltstack(&stack, &old_stack) == -1) {
121 XBT_WARN("Failed to register alternate signal stack: %s",
125 if (!(old_stack.ss_flags & SS_DISABLE)) {
126 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
127 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
128 sigaltstack(&old_stack, NULL);
131 struct sigaction action, old_action;
132 action.sa_sigaction = segvhandler;
133 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
134 sigemptyset(&action.sa_mask);
136 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
137 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
141 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
142 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
143 (old_action.sa_flags & SA_SIGINFO) ?
144 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
145 sigaction(SIGSEGV, &old_action, NULL);
150 /********************************* SIMIX **************************************/
152 XBT_INLINE double SIMIX_timer_next(void)
154 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
157 static void kill_process(smx_process_t process)
159 SIMIX_process_kill(process, NULL);
162 static void SIMIX_storage_create_(smx_storage_t storage)
164 const char* key = xbt_dict_get_elm_key(storage);
165 SIMIX_storage_create(key, storage, NULL);
170 * \brief Initialize SIMIX internal data.
175 void SIMIX_global_init(int *argc, char **argv)
178 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
181 s_smx_process_t proc;
184 simix_global = xbt_new0(s_smx_global_t, 1);
186 #ifdef TIME_BENCH_AMDAHL
187 simix_global->timer_seq = xbt_os_timer_new();
188 simix_global->timer_par = xbt_os_timer_new();
189 xbt_os_cputimer_start(simix_global->timer_seq);
191 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
192 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
193 simix_global->process_list =
194 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
195 simix_global->process_to_destroy =
196 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
198 simix_global->maestro_process = NULL;
199 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
201 simix_global->create_process_function = SIMIX_process_create;
202 simix_global->kill_process_function = kill_process;
203 simix_global->cleanup_process_function = SIMIX_process_cleanup;
204 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
205 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
206 SIMIX_synchro_mallocator_reset_f);
207 simix_global->autorestart = SIMIX_host_restart_processes;
208 simix_global->mutex = xbt_os_mutex_init();
210 surf_init(argc, argv); /* Initialize SURF structures */
211 SIMIX_context_mod_init();
212 SIMIX_create_maestro_process();
214 /* context exception handlers */
215 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
216 __xbt_ex_terminate = SIMIX_process_exception_terminate;
218 SIMIX_network_init();
220 /* Prepare to display some more info when dying on Ctrl-C pressing */
221 signal(SIGINT, inthandler);
224 /* Install SEGV handler */
225 install_segvhandler();
227 /* register a function to be called by SURF after the environment creation */
229 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
230 surf_host_created_callback(SIMIX_host_create);
231 surf_storage_created_callback(SIMIX_storage_create_);
235 simix_timers = xbt_heap_new(8, &free);
238 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
240 if (sg_cfg_get_boolean("clean_atexit"))
244 // The communication initialization is done ASAP.
245 // We need to communicate initialization of the different layers to the model-checker.
249 if (_sg_cfg_exit_asap)
256 * \brief Clean the SIMIX simulation
258 * This functions remove the memory used by SIMIX
260 void SIMIX_clean(void)
262 #ifdef TIME_BENCH_PER_SR
263 smx_ctx_raw_new_sr();
265 if (smx_cleaned) return; // to avoid double cleaning by java and C
267 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
268 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
270 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
271 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
272 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
274 /* Kill all processes (but maestro) */
275 SIMIX_process_killall(simix_global->maestro_process, 1);
277 /* Exit the SIMIX network module */
278 SIMIX_network_exit();
280 xbt_heap_free(simix_timers);
282 /* Free the remaining data structures */
283 xbt_dynar_free(&simix_global->process_to_run);
284 xbt_dynar_free(&simix_global->process_that_ran);
285 xbt_swag_free(simix_global->process_to_destroy);
286 xbt_swag_free(simix_global->process_list);
287 simix_global->process_list = NULL;
288 simix_global->process_to_destroy = NULL;
289 xbt_dict_free(&(simix_global->registered_functions));
291 xbt_os_mutex_destroy(simix_global->mutex);
292 simix_global->mutex = NULL;
294 /* Let's free maestro now */
295 SIMIX_context_free(simix_global->maestro_process->context);
296 xbt_free(simix_global->maestro_process->running_ctx);
297 xbt_free(simix_global->maestro_process);
298 simix_global->maestro_process = NULL;
300 /* Restore the default exception setup */
301 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
302 __xbt_ex_terminate = &__xbt_ex_terminate_default;
304 /* Finish context module and SURF */
305 SIMIX_context_mod_exit();
309 #ifdef TIME_BENCH_AMDAHL
310 xbt_os_cputimer_stop(simix_global->timer_seq);
311 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
312 xbt_os_timer_elapsed(simix_global->timer_seq),
313 xbt_os_timer_elapsed(simix_global->timer_par));
314 xbt_os_timer_free(simix_global->timer_seq);
315 xbt_os_timer_free(simix_global->timer_par);
318 xbt_mallocator_free(simix_global->synchro_mallocator);
319 xbt_free(simix_global);
328 * \brief A clock (in second).
330 * \return Return the clock.
332 XBT_INLINE double SIMIX_get_clock(void)
334 if(MC_is_active() || MC_record_replay_is_active()){
335 return MC_process_clock_get(SIMIX_process_self());
337 return surf_get_clock();
341 static int process_syscall_color(void *p)
343 switch ((*(smx_process_t *)p)->simcall.call) {
345 case SIMCALL_PROCESS_KILL:
347 case SIMCALL_PROCESS_RESUME:
356 * \brief Run the main simulation loop.
361 MC_record_replay_init();
362 MC_record_replay_from_string(MC_record_path);
367 smx_process_t process;
368 surf_action_t action;
374 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
375 xbt_dynar_length(simix_global->process_to_run));
376 #ifdef TIME_BENCH_PER_SR
377 smx_ctx_raw_new_sr();
379 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
380 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
381 xbt_dynar_length(simix_global->process_to_run));
383 /* Run all processes that are ready to run, possibly in parallel */
384 #ifdef TIME_BENCH_AMDAHL
385 xbt_os_cputimer_stop(simix_global->timer_seq);
386 xbt_os_cputimer_resume(simix_global->timer_par);
388 SIMIX_process_runall();
389 #ifdef TIME_BENCH_AMDAHL
390 xbt_os_cputimer_stop(simix_global->timer_par);
391 xbt_os_cputimer_resume(simix_global->timer_seq);
394 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
395 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
397 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
399 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
401 /* Here, the order is ok because:
403 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
405 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
407 * - if there is no kill during the simulation, processes remain sorted according by their PID.
408 * rational: This can be proved inductively.
409 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
410 * Let's show that it is still so at the end of this round.
411 * - if a process is added when being created, that's from maestro. It can be either at startup
412 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
413 * in arbitrary order (inductive hypothesis), we are fine.
414 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
415 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
416 * of a simcall terminates, we're still good. Proof:
417 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
418 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
419 * - For communications, this function is called from SIMIX_comm_finish().
420 * This function itself don't mess with the order since simcalls are handled in FIFO order.
421 * The function is called:
422 * - before the comm starts (invalid parameters, or resource already dead or whatever).
423 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
424 * - because the communication failed or were canceled after startup. In this case, it's called from the function
425 * we are in, by the chunk:
426 * set = model->states.failed_action_set;
427 * while ((synchro = xbt_swag_extract(set)))
428 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
429 * This order is also fixed because it depends of the order in which the surf actions were
430 * added to the system, and only maestro can add stuff this way, through simcalls.
431 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
432 * poped out of the swag does not depend on the user code's execution order.
433 * - because the communication terminated. In this case, synchros are served in the order given by
434 * set = model->states.done_action_set;
435 * while ((synchro = xbt_swag_extract(set)))
436 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
437 * and the argument is very similar to the previous one.
438 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
439 * 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.
440 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
441 * - If there is some process killings, the order is changed by this decision that comes from user-land
442 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
443 * So, even the order change induced by the process killing is perfectly reproducible.
445 * So science works, bitches [http://xkcd.com/54/].
447 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
448 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
449 * That would thus be a pure waste of time.
452 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
453 if (process->simcall.call != SIMCALL_NONE) {
454 SIMIX_simcall_handle(&process->simcall, 0);
457 /* Wake up all processes waiting for a Surf action to finish */
458 xbt_dynar_foreach(all_existing_models, iter, model) {
459 XBT_DEBUG("Handling process whose action failed");
460 while ((action = surf_model_extract_failed_action_set(model))) {
461 XBT_DEBUG(" Handling Action %p",action);
462 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
464 XBT_DEBUG("Handling process whose action terminated normally");
465 while ((action = surf_model_extract_done_action_set(model))) {
466 XBT_DEBUG(" Handling Action %p",action);
467 if (surf_action_get_data(action) == NULL)
468 XBT_DEBUG("probably vcpu's action %p, skip", action);
470 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
475 time = SIMIX_timer_next();
476 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
477 XBT_DEBUG("Calling surf_solve");
478 time = surf_solve(time);
479 XBT_DEBUG("Moving time ahead : %g", time);
481 /* Notify all the hosts that have failed */
482 /* FIXME: iterate through the list of failed host and mark each of them */
483 /* as failed. On each host, signal all the running processes with host_fail */
485 /* Handle any pending timer */
486 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
487 //FIXME: make the timers being real callbacks
488 // (i.e. provide dispatchers that read and expand the args)
489 timer = xbt_heap_pop(simix_timers);
491 timer->func(timer->args);
495 /* Wake up all processes waiting for a Surf action to finish */
496 xbt_dynar_foreach(all_existing_models, iter, model) {
497 XBT_DEBUG("Handling process whose action failed");
498 while ((action = surf_model_extract_failed_action_set(model))) {
499 XBT_DEBUG(" Handling Action %p",action);
500 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
502 XBT_DEBUG("Handling process whose action terminated normally");
503 while ((action = surf_model_extract_done_action_set(model))) {
504 XBT_DEBUG(" Handling Action %p",action);
505 if (surf_action_get_data(action) == NULL)
506 XBT_DEBUG("probably vcpu's action %p, skip", action);
508 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
512 /* Autorestart all process */
513 if(host_that_restart) {
514 char *hostname = NULL;
515 xbt_dynar_foreach(host_that_restart,iter,hostname) {
516 XBT_INFO("Restart processes on host: %s",hostname);
517 SIMIX_host_autorestart(sg_host_by_name(hostname));
519 xbt_dynar_reset(host_that_restart);
522 /* Clean processes to destroy */
523 SIMIX_process_empty_trash();
526 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
528 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
530 if (xbt_swag_size(simix_global->process_list) != 0) {
534 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
535 SIMIX_display_process_status();
541 * \brief Set the date to execute a function
543 * Set the date to execute the function on the surf.
544 * \param date Date to execute function
545 * \param function Function to be executed
546 * \param arg Parameters of the function
549 XBT_INLINE smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
551 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
554 timer->func = function;
556 xbt_heap_push(simix_timers, timer, date);
559 /** @brief cancels a timer that was added earlier */
560 XBT_INLINE void SIMIX_timer_remove(smx_timer_t timer) {
561 xbt_heap_rm_elm(simix_timers, timer, timer->date);
564 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
565 XBT_INLINE double SIMIX_timer_get_date(smx_timer_t timer) {
566 return timer?timer->date:0;
570 * \brief Registers a function to create a process.
572 * This function registers a function to be called
573 * when a new process is created. The function has
574 * to call SIMIX_process_create().
575 * \param function create process function
577 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
580 simix_global->create_process_function = function;
584 * \brief Registers a function to kill a process.
586 * This function registers a function to be called when a
587 * process is killed. The function has to call the SIMIX_process_kill().
589 * \param function Kill process function
591 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
594 simix_global->kill_process_function = function;
598 * \brief Registers a function to cleanup a process.
600 * This function registers a user function to be called when
601 * a process ends properly.
603 * \param function cleanup process function
605 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
608 simix_global->cleanup_process_function = function;
612 void SIMIX_display_process_status(void)
614 if (simix_global->process_list == NULL) {
618 smx_process_t process = NULL;
619 int nbprocess = xbt_swag_size(simix_global->process_list);
621 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
622 /* List the process and their state */
624 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
625 xbt_swag_foreach(process, simix_global->process_list) {
627 if (process->waiting_synchro) {
629 const char* synchro_description = "unknown";
630 switch (process->waiting_synchro->type) {
632 case SIMIX_SYNC_EXECUTE:
633 synchro_description = "execution";
636 case SIMIX_SYNC_PARALLEL_EXECUTE:
637 synchro_description = "parallel execution";
640 case SIMIX_SYNC_COMMUNICATE:
641 synchro_description = "communication";
644 case SIMIX_SYNC_SLEEP:
645 synchro_description = "sleeping";
648 case SIMIX_SYNC_JOIN:
649 synchro_description = "joining";
652 case SIMIX_SYNC_SYNCHRO:
653 synchro_description = "synchronization";
657 synchro_description = "I/O";
660 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
661 process->pid, process->name, sg_host_get_name(process->host),
662 synchro_description, process->waiting_synchro,
663 process->waiting_synchro->name, (int)process->waiting_synchro->state);
666 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
671 static void* SIMIX_synchro_mallocator_new_f(void) {
672 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
673 synchro->simcalls = xbt_fifo_new();
677 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
678 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
682 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
684 // we also recycle the simcall list
685 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
686 xbt_fifo_reset(fifo);
687 memset(synchro, 0, sizeof(s_smx_synchro_t));
688 ((smx_synchro_t) synchro)->simcalls = fifo;
691 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
692 return SIMIX_asr_get_properties(name);
694 xbt_dict_t SIMIX_asr_get_properties(const char *name)
696 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);