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/internal_config.h"
10 #include "src/surf/surf_interface.hpp"
11 #include "src/surf/storage_interface.hpp"
12 #include "src/surf/xml/platf.hpp"
13 #include "smx_private.h"
14 #include "smx_private.hpp"
16 #include "xbt/ex.h" /* ex_backtrace_display */
18 #include "src/mc/mc_replay.h"
19 #include "simgrid/sg_config.h"
22 #include "src/mc/mc_private.h"
23 #include "src/mc/mc_protocol.h"
24 #include "src/mc/Client.hpp"
27 #include "src/mc/mc_protocol.h"
30 #include "src/mc/mc_record.h"
33 #include "src/smpi/private.h"
36 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
37 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
39 smx_global_t simix_global = NULL;
40 static xbt_heap_t simix_timers = NULL;
42 /** @brief Timer datatype */
43 typedef struct s_smx_timer {
49 void (*SMPI_switch_data_segment)(int) = NULL;
51 static void* SIMIX_synchro_mallocator_new_f(void);
52 static void SIMIX_synchro_mallocator_free_f(void* synchro);
53 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
55 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
58 int _sg_do_verbose_exit = 1;
59 static void inthandler(int ignored)
61 if ( _sg_do_verbose_exit ) {
62 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
63 SIMIX_display_process_status();
66 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
72 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
74 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
76 "Access violation detected.\n"
77 "This can result from a programming error in your code or, although less likely,\n"
78 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
79 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
81 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
82 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
83 smx_context_stack_size / 1024);
84 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
86 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
87 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
89 } else if (siginfo->si_signo == SIGSEGV) {
90 fprintf(stderr, "Segmentation fault.\n");
92 if (smpi_enabled() && !smpi_privatize_global_variables) {
93 #if HAVE_PRIVATIZATION
95 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
98 "Sadly, your system does not support --cfg=smpi/privatize_global_variables:yes (yet).\n");
99 #endif /* HAVE_PRIVATIZATION */
101 #endif /* HAVE_SMPI */
106 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
109 * Install signal handler for SIGSEGV. Check that nobody has already installed
110 * its own handler. For example, the Java VM does this.
112 static void install_segvhandler(void)
114 stack_t stack, old_stack;
115 stack.ss_sp = sigsegv_stack;
116 stack.ss_size = sizeof sigsegv_stack;
119 if (sigaltstack(&stack, &old_stack) == -1) {
120 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
123 if (!(old_stack.ss_flags & SS_DISABLE)) {
124 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
125 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
126 sigaltstack(&old_stack, NULL);
129 struct sigaction action, old_action;
130 action.sa_sigaction = segvhandler;
131 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
132 sigemptyset(&action.sa_mask);
134 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
135 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
138 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
139 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
140 (old_action.sa_flags & SA_SIGINFO) ?
141 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
142 sigaction(SIGSEGV, &old_action, NULL);
147 /********************************* SIMIX **************************************/
149 double SIMIX_timer_next(void)
151 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
154 static void kill_process(smx_process_t process)
156 SIMIX_process_kill(process, NULL);
159 static void SIMIX_storage_create_(smx_storage_t storage)
161 const char* key = xbt_dict_get_elm_key(storage);
162 SIMIX_storage_create(key, storage, NULL);
165 static std::function<void()> maestro_code;
170 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
172 maestro_code = std::move(code);
178 void SIMIX_set_maestro(void (*code)(void*), void* data)
180 maestro_code = std::bind(code, data);
185 * \brief Initialize SIMIX internal data.
190 void SIMIX_global_init(int *argc, char **argv)
193 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
196 s_smx_process_t proc;
199 simix_global = xbt_new0(s_smx_global_t, 1);
201 #ifdef TIME_BENCH_AMDAHL
202 simix_global->timer_seq = xbt_os_timer_new();
203 simix_global->timer_par = xbt_os_timer_new();
204 xbt_os_cputimer_start(simix_global->timer_seq);
206 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
207 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
208 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
209 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
211 simix_global->maestro_process = NULL;
212 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
214 simix_global->create_process_function = SIMIX_process_create;
215 simix_global->kill_process_function = kill_process;
216 simix_global->cleanup_process_function = SIMIX_process_cleanup;
217 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
218 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
219 SIMIX_synchro_mallocator_reset_f);
220 simix_global->mutex = xbt_os_mutex_init();
222 surf_init(argc, argv); /* Initialize SURF structures */
223 SIMIX_context_mod_init();
225 // Either create a new context with maestro or create
226 // a context object with the current context mestro):
227 simgrid::simix::create_maestro(maestro_code);
229 /* context exception handlers */
230 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
231 __xbt_ex_terminate = SIMIX_process_exception_terminate;
233 /* Prepare to display some more info when dying on Ctrl-C pressing */
234 signal(SIGINT, inthandler);
237 install_segvhandler();
239 /* register a function to be called by SURF after the environment creation */
241 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
242 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
243 SIMIX_host_create(&host);
245 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
247 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
248 const char* id = storage->getName();
249 // TODO, create sg_storage_by_name
250 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, id);
251 xbt_assert(s != NULL, "Storage not found for name %s", id);
252 SIMIX_storage_create_(s);
255 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
258 simix_timers = xbt_heap_new(8, &free);
261 if (xbt_cfg_get_boolean("clean_atexit"))
265 // The communication initialization is done ASAP.
266 // We need to communicate initialization of the different layers to the model-checker.
267 simgrid::mc::Client::initialize();
270 if (_sg_cfg_exit_asap)
277 * \brief Clean the SIMIX simulation
279 * This functions remove the memory used by SIMIX
281 void SIMIX_clean(void)
283 #ifdef TIME_BENCH_PER_SR
284 smx_ctx_raw_new_sr();
286 if (smx_cleaned) return; // to avoid double cleaning by java and C
288 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
289 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
291 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
292 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
293 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
295 /* Kill all processes (but maestro) */
296 SIMIX_process_killall(simix_global->maestro_process, 1);
298 /* Exit the SIMIX network module */
299 SIMIX_mailbox_exit();
301 xbt_heap_free(simix_timers);
303 /* Free the remaining data structures */
304 xbt_dynar_free(&simix_global->process_to_run);
305 xbt_dynar_free(&simix_global->process_that_ran);
306 xbt_swag_free(simix_global->process_to_destroy);
307 xbt_swag_free(simix_global->process_list);
308 simix_global->process_list = NULL;
309 simix_global->process_to_destroy = NULL;
310 xbt_dict_free(&(simix_global->registered_functions));
312 xbt_os_mutex_destroy(simix_global->mutex);
313 simix_global->mutex = NULL;
315 /* Let's free maestro now */
316 SIMIX_context_free(simix_global->maestro_process->context);
317 xbt_free(simix_global->maestro_process->running_ctx);
318 xbt_free(simix_global->maestro_process);
319 simix_global->maestro_process = NULL;
321 /* Restore the default exception setup */
322 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
323 __xbt_ex_terminate = &__xbt_ex_terminate_default;
325 /* Finish context module and SURF */
326 SIMIX_context_mod_exit();
330 #ifdef TIME_BENCH_AMDAHL
331 xbt_os_cputimer_stop(simix_global->timer_seq);
332 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
333 xbt_os_timer_elapsed(simix_global->timer_seq),
334 xbt_os_timer_elapsed(simix_global->timer_par));
335 xbt_os_timer_free(simix_global->timer_seq);
336 xbt_os_timer_free(simix_global->timer_par);
339 xbt_mallocator_free(simix_global->synchro_mallocator);
340 xbt_free(simix_global);
349 * \brief A clock (in second).
351 * \return Return the clock.
353 double SIMIX_get_clock(void)
355 if(MC_is_active() || MC_record_replay_is_active()){
356 return MC_process_clock_get(SIMIX_process_self());
358 return surf_get_clock();
362 static int process_syscall_color(void *p)
364 switch ((*(smx_process_t *)p)->simcall.call) {
366 case SIMCALL_PROCESS_KILL:
368 case SIMCALL_PROCESS_RESUME:
377 * \brief Run the main simulation loop.
381 if (MC_record_path) {
382 simgrid::mc::replay(MC_record_path);
387 smx_process_t process;
388 surf_action_t action;
394 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
395 xbt_dynar_length(simix_global->process_to_run));
396 #ifdef TIME_BENCH_PER_SR
397 smx_ctx_raw_new_sr();
399 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
400 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
401 xbt_dynar_length(simix_global->process_to_run));
403 /* Run all processes that are ready to run, possibly in parallel */
404 #ifdef TIME_BENCH_AMDAHL
405 xbt_os_cputimer_stop(simix_global->timer_seq);
406 xbt_os_cputimer_resume(simix_global->timer_par);
408 SIMIX_process_runall();
409 #ifdef TIME_BENCH_AMDAHL
410 xbt_os_cputimer_stop(simix_global->timer_par);
411 xbt_os_cputimer_resume(simix_global->timer_seq);
414 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
415 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
417 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
419 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
421 /* Here, the order is ok because:
423 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
425 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
427 * - if there is no kill during the simulation, processes remain sorted according by their PID.
428 * rational: This can be proved inductively.
429 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
430 * Let's show that it is still so at the end of this round.
431 * - if a process is added when being created, that's from maestro. It can be either at startup
432 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
433 * in arbitrary order (inductive hypothesis), we are fine.
434 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
435 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
436 * of a simcall terminates, we're still good. Proof:
437 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
438 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
439 * - For communications, this function is called from SIMIX_comm_finish().
440 * This function itself don't mess with the order since simcalls are handled in FIFO order.
441 * The function is called:
442 * - before the comm starts (invalid parameters, or resource already dead or whatever).
443 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
444 * - because the communication failed or were canceled after startup. In this case, it's called from the function
445 * we are in, by the chunk:
446 * set = model->states.failed_action_set;
447 * while ((synchro = xbt_swag_extract(set)))
448 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
449 * This order is also fixed because it depends of the order in which the surf actions were
450 * added to the system, and only maestro can add stuff this way, through simcalls.
451 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
452 * poped out of the swag does not depend on the user code's execution order.
453 * - because the communication terminated. In this case, synchros are served in the order given by
454 * set = model->states.done_action_set;
455 * while ((synchro = xbt_swag_extract(set)))
456 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
457 * and the argument is very similar to the previous one.
458 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
459 * 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.
460 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
461 * - If there is some process killings, the order is changed by this decision that comes from user-land
462 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
463 * So, even the order change induced by the process killing is perfectly reproducible.
465 * So science works, bitches [http://xkcd.com/54/].
467 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
468 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
469 * That would thus be a pure waste of time.
472 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
473 if (process->simcall.call != SIMCALL_NONE) {
474 SIMIX_simcall_handle(&process->simcall, 0);
477 /* Wake up all processes waiting for a Surf action to finish */
478 xbt_dynar_foreach(all_existing_models, iter, model) {
479 XBT_DEBUG("Handling process whose action failed");
480 while ((action = surf_model_extract_failed_action_set(model))) {
481 XBT_DEBUG(" Handling Action %p",action);
482 SIMIX_simcall_exit((smx_synchro_t) action->getData());
484 XBT_DEBUG("Handling process whose action terminated normally");
485 while ((action = surf_model_extract_done_action_set(model))) {
486 XBT_DEBUG(" Handling Action %p",action);
487 if (action->getData() == NULL)
488 XBT_DEBUG("probably vcpu's action %p, skip", action);
490 SIMIX_simcall_exit((smx_synchro_t) action->getData());
495 time = SIMIX_timer_next();
496 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
497 XBT_DEBUG("Calling surf_solve");
498 time = surf_solve(time);
499 XBT_DEBUG("Moving time ahead : %g", time);
501 /* Notify all the hosts that have failed */
502 /* FIXME: iterate through the list of failed host and mark each of them */
503 /* as failed. On each host, signal all the running processes with host_fail */
505 /* Handle any pending timer */
506 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
507 //FIXME: make the timers being real callbacks
508 // (i.e. provide dispatchers that read and expand the args)
509 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
511 timer->func(timer->args);
515 /* Wake up all processes waiting for a Surf action to finish */
516 xbt_dynar_foreach(all_existing_models, iter, model) {
517 XBT_DEBUG("Handling process whose action failed");
518 while ((action = surf_model_extract_failed_action_set(model))) {
519 XBT_DEBUG(" Handling Action %p",action);
520 SIMIX_simcall_exit((smx_synchro_t) action->getData());
522 XBT_DEBUG("Handling process whose action terminated normally");
523 while ((action = surf_model_extract_done_action_set(model))) {
524 XBT_DEBUG(" Handling Action %p",action);
525 if (action->getData() == NULL)
526 XBT_DEBUG("probably vcpu's action %p, skip", action);
528 SIMIX_simcall_exit((smx_synchro_t) action->getData());
532 /* Autorestart all process */
533 char *hostname = NULL;
534 xbt_dynar_foreach(host_that_restart,iter,hostname) {
535 XBT_INFO("Restart processes on host: %s",hostname);
536 SIMIX_host_autorestart(sg_host_by_name(hostname));
538 xbt_dynar_reset(host_that_restart);
540 /* Clean processes to destroy */
541 SIMIX_process_empty_trash();
544 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
546 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
548 if (xbt_swag_size(simix_global->process_list) != 0) {
552 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
553 SIMIX_display_process_status();
559 * \brief Set the date to execute a function
561 * Set the date to execute the function on the surf.
562 * \param date Date to execute function
563 * \param function Function to be executed
564 * \param arg Parameters of the function
567 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
569 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
572 timer->func = function;
574 xbt_heap_push(simix_timers, timer, date);
577 /** @brief cancels a timer that was added earlier */
578 void SIMIX_timer_remove(smx_timer_t timer) {
579 xbt_heap_rm_elm(simix_timers, timer, timer->date);
582 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
583 double SIMIX_timer_get_date(smx_timer_t timer) {
584 return timer?timer->date:0;
588 * \brief Registers a function to create a process.
590 * This function registers a function to be called
591 * when a new process is created. The function has
592 * to call SIMIX_process_create().
593 * \param function create process function
595 void SIMIX_function_register_process_create(smx_creation_func_t
598 simix_global->create_process_function = function;
602 * \brief Registers a function to kill a process.
604 * This function registers a function to be called when a
605 * process is killed. The function has to call the SIMIX_process_kill().
607 * \param function Kill process function
609 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
612 simix_global->kill_process_function = function;
616 * \brief Registers a function to cleanup a process.
618 * This function registers a user function to be called when
619 * a process ends properly.
621 * \param function cleanup process function
623 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
626 simix_global->cleanup_process_function = function;
630 void SIMIX_display_process_status(void)
632 if (simix_global->process_list == NULL) {
636 smx_process_t process = NULL;
637 int nbprocess = xbt_swag_size(simix_global->process_list);
639 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
640 /* List the process and their state */
642 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
643 xbt_swag_foreach(process, simix_global->process_list) {
645 if (process->waiting_synchro) {
647 const char* synchro_description = "unknown";
648 switch (process->waiting_synchro->type) {
650 case SIMIX_SYNC_EXECUTE:
651 synchro_description = "execution";
654 case SIMIX_SYNC_PARALLEL_EXECUTE:
655 synchro_description = "parallel execution";
658 case SIMIX_SYNC_COMMUNICATE:
659 synchro_description = "communication";
662 case SIMIX_SYNC_SLEEP:
663 synchro_description = "sleeping";
666 case SIMIX_SYNC_JOIN:
667 synchro_description = "joining";
670 case SIMIX_SYNC_SYNCHRO:
671 synchro_description = "synchronization";
675 synchro_description = "I/O";
678 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
679 process->pid, process->name, sg_host_get_name(process->host),
680 synchro_description, process->waiting_synchro,
681 process->waiting_synchro->name, (int)process->waiting_synchro->state);
684 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
689 static void* SIMIX_synchro_mallocator_new_f(void) {
690 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
691 synchro->simcalls = xbt_fifo_new();
695 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
696 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
700 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
702 // we also recycle the simcall list
703 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
704 xbt_fifo_reset(fifo);
705 memset(synchro, 0, sizeof(s_smx_synchro_t));
706 ((smx_synchro_t) synchro)->simcalls = fifo;
709 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
710 return SIMIX_asr_get_properties(name);
712 xbt_dict_t SIMIX_asr_get_properties(const char *name)
714 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
717 int SIMIX_is_maestro()
719 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;