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 static void* SIMIX_synchro_mallocator_new_f(void);
59 static void SIMIX_synchro_mallocator_free_f(void* synchro);
60 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
62 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
65 int _sg_do_verbose_exit = 1;
66 static void _XBT_CALL inthandler(int ignored)
68 if ( _sg_do_verbose_exit ) {
69 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
70 SIMIX_display_process_status();
73 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
79 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
81 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
83 "Access violation detected.\n"
84 "This can result from a programming error in your code or, although less likely,\n"
85 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
86 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
88 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
89 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
90 smx_context_stack_size / 1024);
91 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
93 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
94 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
96 } else if (siginfo->si_signo == SIGSEGV) {
97 fprintf(stderr, "Segmentation fault.\n");
99 if (smpi_enabled() && !smpi_privatize_global_variables) {
101 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
108 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
111 * Install signal handler for SIGSEGV. Check that nobody has already installed
112 * its own handler. For example, the Java VM does this.
114 static void install_segvhandler(void)
116 stack_t stack, old_stack;
117 stack.ss_sp = sigsegv_stack;
118 stack.ss_size = sizeof sigsegv_stack;
121 if (sigaltstack(&stack, &old_stack) == -1) {
122 XBT_WARN("Failed to register alternate signal stack: %s",
126 if (!(old_stack.ss_flags & SS_DISABLE)) {
127 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
128 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
129 sigaltstack(&old_stack, NULL);
132 struct sigaction action, old_action;
133 action.sa_sigaction = segvhandler;
134 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
135 sigemptyset(&action.sa_mask);
137 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
138 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
142 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
143 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
144 (old_action.sa_flags & SA_SIGINFO) ?
145 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
146 sigaction(SIGSEGV, &old_action, NULL);
151 /********************************* SIMIX **************************************/
153 double SIMIX_timer_next(void)
155 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
158 static void kill_process(smx_process_t process)
160 SIMIX_process_kill(process, NULL);
163 static void SIMIX_storage_create_(smx_storage_t storage)
165 const char* key = xbt_dict_get_elm_key(storage);
166 SIMIX_storage_create(key, storage, NULL);
171 * \brief Initialize SIMIX internal data.
176 void SIMIX_global_init(int *argc, char **argv)
179 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
182 s_smx_process_t proc;
185 simix_global = xbt_new0(s_smx_global_t, 1);
187 #ifdef TIME_BENCH_AMDAHL
188 simix_global->timer_seq = xbt_os_timer_new();
189 simix_global->timer_par = xbt_os_timer_new();
190 xbt_os_cputimer_start(simix_global->timer_seq);
192 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
193 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
194 simix_global->process_list =
195 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
196 simix_global->process_to_destroy =
197 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
199 simix_global->maestro_process = NULL;
200 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
202 simix_global->create_process_function = SIMIX_process_create;
203 simix_global->kill_process_function = kill_process;
204 simix_global->cleanup_process_function = SIMIX_process_cleanup;
205 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
206 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
207 SIMIX_synchro_mallocator_reset_f);
208 simix_global->autorestart = SIMIX_host_restart_processes;
209 simix_global->mutex = xbt_os_mutex_init();
211 surf_init(argc, argv); /* Initialize SURF structures */
212 SIMIX_context_mod_init();
213 SIMIX_create_maestro_process();
215 /* context exception handlers */
216 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
217 __xbt_ex_terminate = SIMIX_process_exception_terminate;
219 SIMIX_network_init();
221 /* Prepare to display some more info when dying on Ctrl-C pressing */
222 signal(SIGINT, inthandler);
225 /* Install SEGV handler */
226 install_segvhandler();
228 /* register a function to be called by SURF after the environment creation */
230 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
231 surf_on_host_created(SIMIX_host_create);
232 surf_on_storage_created(SIMIX_storage_create_);
236 simix_timers = xbt_heap_new(8, &free);
239 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
241 if (sg_cfg_get_boolean("clean_atexit"))
245 // The communication initialization is done ASAP.
246 // We need to communicate initialization of the different layers to the model-checker.
250 if (_sg_cfg_exit_asap)
257 * \brief Clean the SIMIX simulation
259 * This functions remove the memory used by SIMIX
261 void SIMIX_clean(void)
263 #ifdef TIME_BENCH_PER_SR
264 smx_ctx_raw_new_sr();
266 if (smx_cleaned) return; // to avoid double cleaning by java and C
268 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
269 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
271 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
272 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
273 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
275 /* Kill all processes (but maestro) */
276 SIMIX_process_killall(simix_global->maestro_process, 1);
278 /* Exit the SIMIX network module */
279 SIMIX_network_exit();
281 xbt_heap_free(simix_timers);
283 /* Free the remaining data structures */
284 xbt_dynar_free(&simix_global->process_to_run);
285 xbt_dynar_free(&simix_global->process_that_ran);
286 xbt_swag_free(simix_global->process_to_destroy);
287 xbt_swag_free(simix_global->process_list);
288 simix_global->process_list = NULL;
289 simix_global->process_to_destroy = NULL;
290 xbt_dict_free(&(simix_global->registered_functions));
292 xbt_os_mutex_destroy(simix_global->mutex);
293 simix_global->mutex = NULL;
295 /* Let's free maestro now */
296 SIMIX_context_free(simix_global->maestro_process->context);
297 xbt_free(simix_global->maestro_process->running_ctx);
298 xbt_free(simix_global->maestro_process);
299 simix_global->maestro_process = NULL;
301 /* Restore the default exception setup */
302 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
303 __xbt_ex_terminate = &__xbt_ex_terminate_default;
305 /* Finish context module and SURF */
306 SIMIX_context_mod_exit();
310 #ifdef TIME_BENCH_AMDAHL
311 xbt_os_cputimer_stop(simix_global->timer_seq);
312 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
313 xbt_os_timer_elapsed(simix_global->timer_seq),
314 xbt_os_timer_elapsed(simix_global->timer_par));
315 xbt_os_timer_free(simix_global->timer_seq);
316 xbt_os_timer_free(simix_global->timer_par);
319 xbt_mallocator_free(simix_global->synchro_mallocator);
320 xbt_free(simix_global);
329 * \brief A clock (in second).
331 * \return Return the clock.
333 double SIMIX_get_clock(void)
335 if(MC_is_active() || MC_record_replay_is_active()){
336 return MC_process_clock_get(SIMIX_process_self());
338 return surf_get_clock();
342 static int process_syscall_color(void *p)
344 switch ((*(smx_process_t *)p)->simcall.call) {
346 case SIMCALL_PROCESS_KILL:
348 case SIMCALL_PROCESS_RESUME:
357 * \brief Run the main simulation loop.
362 MC_record_replay_init();
363 MC_record_replay_from_string(MC_record_path);
368 smx_process_t process;
369 surf_action_t action;
375 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
376 xbt_dynar_length(simix_global->process_to_run));
377 #ifdef TIME_BENCH_PER_SR
378 smx_ctx_raw_new_sr();
380 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
381 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
382 xbt_dynar_length(simix_global->process_to_run));
384 /* Run all processes that are ready to run, possibly in parallel */
385 #ifdef TIME_BENCH_AMDAHL
386 xbt_os_cputimer_stop(simix_global->timer_seq);
387 xbt_os_cputimer_resume(simix_global->timer_par);
389 SIMIX_process_runall();
390 #ifdef TIME_BENCH_AMDAHL
391 xbt_os_cputimer_stop(simix_global->timer_par);
392 xbt_os_cputimer_resume(simix_global->timer_seq);
395 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
396 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
398 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
400 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
402 /* Here, the order is ok because:
404 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
406 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
408 * - if there is no kill during the simulation, processes remain sorted according by their PID.
409 * rational: This can be proved inductively.
410 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
411 * Let's show that it is still so at the end of this round.
412 * - if a process is added when being created, that's from maestro. It can be either at startup
413 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
414 * in arbitrary order (inductive hypothesis), we are fine.
415 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
416 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
417 * of a simcall terminates, we're still good. Proof:
418 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
419 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
420 * - For communications, this function is called from SIMIX_comm_finish().
421 * This function itself don't mess with the order since simcalls are handled in FIFO order.
422 * The function is called:
423 * - before the comm starts (invalid parameters, or resource already dead or whatever).
424 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
425 * - because the communication failed or were canceled after startup. In this case, it's called from the function
426 * we are in, by the chunk:
427 * set = model->states.failed_action_set;
428 * while ((synchro = xbt_swag_extract(set)))
429 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
430 * This order is also fixed because it depends of the order in which the surf actions were
431 * added to the system, and only maestro can add stuff this way, through simcalls.
432 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
433 * poped out of the swag does not depend on the user code's execution order.
434 * - because the communication terminated. In this case, synchros are served in the order given by
435 * set = model->states.done_action_set;
436 * while ((synchro = xbt_swag_extract(set)))
437 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
438 * and the argument is very similar to the previous one.
439 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
440 * 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.
441 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
442 * - If there is some process killings, the order is changed by this decision that comes from user-land
443 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
444 * So, even the order change induced by the process killing is perfectly reproducible.
446 * So science works, bitches [http://xkcd.com/54/].
448 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
449 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
450 * That would thus be a pure waste of time.
453 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
454 if (process->simcall.call != SIMCALL_NONE) {
455 SIMIX_simcall_handle(&process->simcall, 0);
458 /* Wake up all processes waiting for a Surf action to finish */
459 xbt_dynar_foreach(all_existing_models, iter, model) {
460 XBT_DEBUG("Handling process whose action failed");
461 while ((action = surf_model_extract_failed_action_set(model))) {
462 XBT_DEBUG(" Handling Action %p",action);
463 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
465 XBT_DEBUG("Handling process whose action terminated normally");
466 while ((action = surf_model_extract_done_action_set(model))) {
467 XBT_DEBUG(" Handling Action %p",action);
468 if (surf_action_get_data(action) == NULL)
469 XBT_DEBUG("probably vcpu's action %p, skip", action);
471 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
476 time = SIMIX_timer_next();
477 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
478 XBT_DEBUG("Calling surf_solve");
479 time = surf_solve(time);
480 XBT_DEBUG("Moving time ahead : %g", time);
482 /* Notify all the hosts that have failed */
483 /* FIXME: iterate through the list of failed host and mark each of them */
484 /* as failed. On each host, signal all the running processes with host_fail */
486 /* Handle any pending timer */
487 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
488 //FIXME: make the timers being real callbacks
489 // (i.e. provide dispatchers that read and expand the args)
490 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
492 timer->func(timer->args);
496 /* Wake up all processes waiting for a Surf action to finish */
497 xbt_dynar_foreach(all_existing_models, iter, model) {
498 XBT_DEBUG("Handling process whose action failed");
499 while ((action = surf_model_extract_failed_action_set(model))) {
500 XBT_DEBUG(" Handling Action %p",action);
501 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
503 XBT_DEBUG("Handling process whose action terminated normally");
504 while ((action = surf_model_extract_done_action_set(model))) {
505 XBT_DEBUG(" Handling Action %p",action);
506 if (surf_action_get_data(action) == NULL)
507 XBT_DEBUG("probably vcpu's action %p, skip", action);
509 SIMIX_simcall_exit((smx_synchro_t) surf_action_get_data(action));
513 /* Autorestart all process */
514 if(host_that_restart) {
515 char *hostname = NULL;
516 xbt_dynar_foreach(host_that_restart,iter,hostname) {
517 XBT_INFO("Restart processes on host: %s",hostname);
518 SIMIX_host_autorestart(sg_host_by_name(hostname));
520 xbt_dynar_reset(host_that_restart);
523 /* Clean processes to destroy */
524 SIMIX_process_empty_trash();
527 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
529 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
531 if (xbt_swag_size(simix_global->process_list) != 0) {
535 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
536 SIMIX_display_process_status();
542 * \brief Set the date to execute a function
544 * Set the date to execute the function on the surf.
545 * \param date Date to execute function
546 * \param function Function to be executed
547 * \param arg Parameters of the function
550 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
552 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
555 timer->func = function;
557 xbt_heap_push(simix_timers, timer, date);
560 /** @brief cancels a timer that was added earlier */
561 void SIMIX_timer_remove(smx_timer_t timer) {
562 xbt_heap_rm_elm(simix_timers, timer, timer->date);
565 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
566 double SIMIX_timer_get_date(smx_timer_t timer) {
567 return timer?timer->date:0;
571 * \brief Registers a function to create a process.
573 * This function registers a function to be called
574 * when a new process is created. The function has
575 * to call SIMIX_process_create().
576 * \param function create process function
578 void SIMIX_function_register_process_create(smx_creation_func_t
581 simix_global->create_process_function = function;
585 * \brief Registers a function to kill a process.
587 * This function registers a function to be called when a
588 * process is killed. The function has to call the SIMIX_process_kill().
590 * \param function Kill process function
592 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
595 simix_global->kill_process_function = function;
599 * \brief Registers a function to cleanup a process.
601 * This function registers a user function to be called when
602 * a process ends properly.
604 * \param function cleanup process function
606 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
609 simix_global->cleanup_process_function = function;
613 void SIMIX_display_process_status(void)
615 if (simix_global->process_list == NULL) {
619 smx_process_t process = NULL;
620 int nbprocess = xbt_swag_size(simix_global->process_list);
622 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
623 /* List the process and their state */
625 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
626 xbt_swag_foreach(process, simix_global->process_list) {
628 if (process->waiting_synchro) {
630 const char* synchro_description = "unknown";
631 switch (process->waiting_synchro->type) {
633 case SIMIX_SYNC_EXECUTE:
634 synchro_description = "execution";
637 case SIMIX_SYNC_PARALLEL_EXECUTE:
638 synchro_description = "parallel execution";
641 case SIMIX_SYNC_COMMUNICATE:
642 synchro_description = "communication";
645 case SIMIX_SYNC_SLEEP:
646 synchro_description = "sleeping";
649 case SIMIX_SYNC_JOIN:
650 synchro_description = "joining";
653 case SIMIX_SYNC_SYNCHRO:
654 synchro_description = "synchronization";
658 synchro_description = "I/O";
661 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
662 process->pid, process->name, sg_host_get_name(process->host),
663 synchro_description, process->waiting_synchro,
664 process->waiting_synchro->name, (int)process->waiting_synchro->state);
667 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
672 static void* SIMIX_synchro_mallocator_new_f(void) {
673 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
674 synchro->simcalls = xbt_fifo_new();
678 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
679 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
683 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
685 // we also recycle the simcall list
686 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
687 xbt_fifo_reset(fifo);
688 memset(synchro, 0, sizeof(s_smx_synchro_t));
689 ((smx_synchro_t) synchro)->simcalls = fifo;
692 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
693 return SIMIX_asr_get_properties(name);
695 xbt_dict_t SIMIX_asr_get_properties(const char *name)
697 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);