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 "src/surf/surf_interface.hpp"
15 #include "src/surf/storage_interface.hpp"
16 #include "src/surf/xml/platf.hpp"
17 #include "smx_private.h"
18 #include "smx_private.hpp"
20 #include "xbt/ex.h" /* ex_backtrace_display */
22 #include "src/mc/mc_replay.h"
23 #include "simgrid/sg_config.h"
26 #include "src/mc/mc_private.h"
27 #include "src/mc/mc_protocol.h"
28 #include "src/mc/mc_client.h"
33 #include "src/mc/mc_protocol.h"
36 #include "src/mc/mc_record.h"
39 #include "src/smpi/private.h"
42 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
43 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
44 "Logging specific to SIMIX (kernel)");
46 smx_global_t simix_global = NULL;
47 static xbt_heap_t simix_timers = NULL;
49 /** @brief Timer datatype */
50 typedef struct s_smx_timer {
56 void (*SMPI_switch_data_segment)(int) = NULL;
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) {
100 #ifdef HAVE_PRIVATIZATION
102 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
105 "Sadly, your system does not support --cfg=smpi/privatize_global_variables:yes (yet).\n");
113 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
116 * Install signal handler for SIGSEGV. Check that nobody has already installed
117 * its own handler. For example, the Java VM does this.
119 static void install_segvhandler(void)
121 stack_t stack, old_stack;
122 stack.ss_sp = sigsegv_stack;
123 stack.ss_size = sizeof sigsegv_stack;
126 if (sigaltstack(&stack, &old_stack) == -1) {
127 XBT_WARN("Failed to register alternate signal stack: %s",
131 if (!(old_stack.ss_flags & SS_DISABLE)) {
132 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
133 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
134 sigaltstack(&old_stack, NULL);
137 struct sigaction action, old_action;
138 action.sa_sigaction = segvhandler;
139 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
140 sigemptyset(&action.sa_mask);
142 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
143 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
147 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
148 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
149 (old_action.sa_flags & SA_SIGINFO) ?
150 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
151 sigaction(SIGSEGV, &old_action, NULL);
156 /********************************* SIMIX **************************************/
158 double SIMIX_timer_next(void)
160 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
163 static void kill_process(smx_process_t process)
165 SIMIX_process_kill(process, NULL);
168 static void SIMIX_storage_create_(smx_storage_t storage)
170 const char* key = xbt_dict_get_elm_key(storage);
171 SIMIX_storage_create(key, storage, NULL);
174 static std::function<void()> maestro_code;
179 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
181 maestro_code = std::move(code);
187 void SIMIX_set_maestro(void (*code)(void*), void* data)
189 maestro_code = std::bind(code, data);
194 * \brief Initialize SIMIX internal data.
199 void SIMIX_global_init(int *argc, char **argv)
202 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
205 s_smx_process_t proc;
208 simix_global = xbt_new0(s_smx_global_t, 1);
210 #ifdef TIME_BENCH_AMDAHL
211 simix_global->timer_seq = xbt_os_timer_new();
212 simix_global->timer_par = xbt_os_timer_new();
213 xbt_os_cputimer_start(simix_global->timer_seq);
215 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
216 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
217 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
218 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
220 simix_global->maestro_process = NULL;
221 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
223 simix_global->create_process_function = SIMIX_process_create;
224 simix_global->kill_process_function = kill_process;
225 simix_global->cleanup_process_function = SIMIX_process_cleanup;
226 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
227 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
228 SIMIX_synchro_mallocator_reset_f);
229 simix_global->mutex = xbt_os_mutex_init();
231 surf_init(argc, argv); /* Initialize SURF structures */
232 SIMIX_context_mod_init();
234 // Either create a new context with maestro or create
235 // a context object with the current context mestro):
236 simgrid::simix::create_maestro(maestro_code);
238 /* context exception handlers */
239 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
240 __xbt_ex_terminate = SIMIX_process_exception_terminate;
242 SIMIX_network_init();
244 /* Prepare to display some more info when dying on Ctrl-C pressing */
245 signal(SIGINT, inthandler);
248 install_segvhandler();
250 /* register a function to be called by SURF after the environment creation */
252 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
253 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
254 SIMIX_host_create(&host);
256 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
258 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
259 const char* id = storage->getName();
260 // TODO, create sg_storage_by_name
261 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, id);
262 xbt_assert(s != NULL, "Storage not found for name %s", id);
263 SIMIX_storage_create_(s);
269 simix_timers = xbt_heap_new(8, &free);
272 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
274 if (sg_cfg_get_boolean("clean_atexit"))
278 // The communication initialization is done ASAP.
279 // We need to communicate initialization of the different layers to the model-checker.
283 if (_sg_cfg_exit_asap)
290 * \brief Clean the SIMIX simulation
292 * This functions remove the memory used by SIMIX
294 void SIMIX_clean(void)
296 #ifdef TIME_BENCH_PER_SR
297 smx_ctx_raw_new_sr();
299 if (smx_cleaned) return; // to avoid double cleaning by java and C
301 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
302 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
304 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
305 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
306 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
308 /* Kill all processes (but maestro) */
309 SIMIX_process_killall(simix_global->maestro_process, 1);
311 /* Exit the SIMIX network module */
312 SIMIX_network_exit();
314 xbt_heap_free(simix_timers);
316 /* Free the remaining data structures */
317 xbt_dynar_free(&simix_global->process_to_run);
318 xbt_dynar_free(&simix_global->process_that_ran);
319 xbt_swag_free(simix_global->process_to_destroy);
320 xbt_swag_free(simix_global->process_list);
321 simix_global->process_list = NULL;
322 simix_global->process_to_destroy = NULL;
323 xbt_dict_free(&(simix_global->registered_functions));
325 xbt_os_mutex_destroy(simix_global->mutex);
326 simix_global->mutex = NULL;
328 /* Let's free maestro now */
329 SIMIX_context_free(simix_global->maestro_process->context);
330 xbt_free(simix_global->maestro_process->running_ctx);
331 xbt_free(simix_global->maestro_process);
332 simix_global->maestro_process = NULL;
334 /* Restore the default exception setup */
335 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
336 __xbt_ex_terminate = &__xbt_ex_terminate_default;
338 /* Finish context module and SURF */
339 SIMIX_context_mod_exit();
343 #ifdef TIME_BENCH_AMDAHL
344 xbt_os_cputimer_stop(simix_global->timer_seq);
345 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
346 xbt_os_timer_elapsed(simix_global->timer_seq),
347 xbt_os_timer_elapsed(simix_global->timer_par));
348 xbt_os_timer_free(simix_global->timer_seq);
349 xbt_os_timer_free(simix_global->timer_par);
352 xbt_mallocator_free(simix_global->synchro_mallocator);
353 xbt_free(simix_global);
362 * \brief A clock (in second).
364 * \return Return the clock.
366 double SIMIX_get_clock(void)
368 if(MC_is_active() || MC_record_replay_is_active()){
369 return MC_process_clock_get(SIMIX_process_self());
371 return surf_get_clock();
375 static int process_syscall_color(void *p)
377 switch ((*(smx_process_t *)p)->simcall.call) {
379 case SIMCALL_PROCESS_KILL:
381 case SIMCALL_PROCESS_RESUME:
390 * \brief Run the main simulation loop.
395 MC_record_replay_init();
396 MC_record_replay_from_string(MC_record_path);
401 smx_process_t process;
402 surf_action_t action;
408 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
409 xbt_dynar_length(simix_global->process_to_run));
410 #ifdef TIME_BENCH_PER_SR
411 smx_ctx_raw_new_sr();
413 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
414 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
415 xbt_dynar_length(simix_global->process_to_run));
417 /* Run all processes that are ready to run, possibly in parallel */
418 #ifdef TIME_BENCH_AMDAHL
419 xbt_os_cputimer_stop(simix_global->timer_seq);
420 xbt_os_cputimer_resume(simix_global->timer_par);
422 SIMIX_process_runall();
423 #ifdef TIME_BENCH_AMDAHL
424 xbt_os_cputimer_stop(simix_global->timer_par);
425 xbt_os_cputimer_resume(simix_global->timer_seq);
428 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
429 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
431 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
433 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
435 /* Here, the order is ok because:
437 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
439 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
441 * - if there is no kill during the simulation, processes remain sorted according by their PID.
442 * rational: This can be proved inductively.
443 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
444 * Let's show that it is still so at the end of this round.
445 * - if a process is added when being created, that's from maestro. It can be either at startup
446 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
447 * in arbitrary order (inductive hypothesis), we are fine.
448 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
449 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
450 * of a simcall terminates, we're still good. Proof:
451 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
452 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
453 * - For communications, this function is called from SIMIX_comm_finish().
454 * This function itself don't mess with the order since simcalls are handled in FIFO order.
455 * The function is called:
456 * - before the comm starts (invalid parameters, or resource already dead or whatever).
457 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
458 * - because the communication failed or were canceled after startup. In this case, it's called from the function
459 * we are in, by the chunk:
460 * set = model->states.failed_action_set;
461 * while ((synchro = xbt_swag_extract(set)))
462 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
463 * This order is also fixed because it depends of the order in which the surf actions were
464 * added to the system, and only maestro can add stuff this way, through simcalls.
465 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
466 * poped out of the swag does not depend on the user code's execution order.
467 * - because the communication terminated. In this case, synchros are served in the order given by
468 * set = model->states.done_action_set;
469 * while ((synchro = xbt_swag_extract(set)))
470 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
471 * and the argument is very similar to the previous one.
472 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
473 * 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.
474 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
475 * - If there is some process killings, the order is changed by this decision that comes from user-land
476 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
477 * So, even the order change induced by the process killing is perfectly reproducible.
479 * So science works, bitches [http://xkcd.com/54/].
481 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
482 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
483 * That would thus be a pure waste of time.
486 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
487 if (process->simcall.call != SIMCALL_NONE) {
488 SIMIX_simcall_handle(&process->simcall, 0);
491 /* Wake up all processes waiting for a Surf action to finish */
492 xbt_dynar_foreach(all_existing_models, iter, model) {
493 XBT_DEBUG("Handling process whose action failed");
494 while ((action = surf_model_extract_failed_action_set(model))) {
495 XBT_DEBUG(" Handling Action %p",action);
496 SIMIX_simcall_exit((smx_synchro_t) action->getData());
498 XBT_DEBUG("Handling process whose action terminated normally");
499 while ((action = surf_model_extract_done_action_set(model))) {
500 XBT_DEBUG(" Handling Action %p",action);
501 if (action->getData() == NULL)
502 XBT_DEBUG("probably vcpu's action %p, skip", action);
504 SIMIX_simcall_exit((smx_synchro_t) action->getData());
509 time = SIMIX_timer_next();
510 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
511 XBT_DEBUG("Calling surf_solve");
512 time = surf_solve(time);
513 XBT_DEBUG("Moving time ahead : %g", time);
515 /* Notify all the hosts that have failed */
516 /* FIXME: iterate through the list of failed host and mark each of them */
517 /* as failed. On each host, signal all the running processes with host_fail */
519 /* Handle any pending timer */
520 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
521 //FIXME: make the timers being real callbacks
522 // (i.e. provide dispatchers that read and expand the args)
523 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
525 timer->func(timer->args);
529 /* Wake up all processes waiting for a Surf action to finish */
530 xbt_dynar_foreach(all_existing_models, iter, model) {
531 XBT_DEBUG("Handling process whose action failed");
532 while ((action = surf_model_extract_failed_action_set(model))) {
533 XBT_DEBUG(" Handling Action %p",action);
534 SIMIX_simcall_exit((smx_synchro_t) action->getData());
536 XBT_DEBUG("Handling process whose action terminated normally");
537 while ((action = surf_model_extract_done_action_set(model))) {
538 XBT_DEBUG(" Handling Action %p",action);
539 if (action->getData() == NULL)
540 XBT_DEBUG("probably vcpu's action %p, skip", action);
542 SIMIX_simcall_exit((smx_synchro_t) action->getData());
546 /* Autorestart all process */
547 char *hostname = NULL;
548 xbt_dynar_foreach(host_that_restart,iter,hostname) {
549 XBT_INFO("Restart processes on host: %s",hostname);
550 SIMIX_host_autorestart(sg_host_by_name(hostname));
552 xbt_dynar_reset(host_that_restart);
554 /* Clean processes to destroy */
555 SIMIX_process_empty_trash();
558 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
560 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
562 if (xbt_swag_size(simix_global->process_list) != 0) {
566 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
567 SIMIX_display_process_status();
573 * \brief Set the date to execute a function
575 * Set the date to execute the function on the surf.
576 * \param date Date to execute function
577 * \param function Function to be executed
578 * \param arg Parameters of the function
581 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
583 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
586 timer->func = function;
588 xbt_heap_push(simix_timers, timer, date);
591 /** @brief cancels a timer that was added earlier */
592 void SIMIX_timer_remove(smx_timer_t timer) {
593 xbt_heap_rm_elm(simix_timers, timer, timer->date);
596 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
597 double SIMIX_timer_get_date(smx_timer_t timer) {
598 return timer?timer->date:0;
602 * \brief Registers a function to create a process.
604 * This function registers a function to be called
605 * when a new process is created. The function has
606 * to call SIMIX_process_create().
607 * \param function create process function
609 void SIMIX_function_register_process_create(smx_creation_func_t
612 simix_global->create_process_function = function;
616 * \brief Registers a function to kill a process.
618 * This function registers a function to be called when a
619 * process is killed. The function has to call the SIMIX_process_kill().
621 * \param function Kill process function
623 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
626 simix_global->kill_process_function = function;
630 * \brief Registers a function to cleanup a process.
632 * This function registers a user function to be called when
633 * a process ends properly.
635 * \param function cleanup process function
637 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
640 simix_global->cleanup_process_function = function;
644 void SIMIX_display_process_status(void)
646 if (simix_global->process_list == NULL) {
650 smx_process_t process = NULL;
651 int nbprocess = xbt_swag_size(simix_global->process_list);
653 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
654 /* List the process and their state */
656 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
657 xbt_swag_foreach(process, simix_global->process_list) {
659 if (process->waiting_synchro) {
661 const char* synchro_description = "unknown";
662 switch (process->waiting_synchro->type) {
664 case SIMIX_SYNC_EXECUTE:
665 synchro_description = "execution";
668 case SIMIX_SYNC_PARALLEL_EXECUTE:
669 synchro_description = "parallel execution";
672 case SIMIX_SYNC_COMMUNICATE:
673 synchro_description = "communication";
676 case SIMIX_SYNC_SLEEP:
677 synchro_description = "sleeping";
680 case SIMIX_SYNC_JOIN:
681 synchro_description = "joining";
684 case SIMIX_SYNC_SYNCHRO:
685 synchro_description = "synchronization";
689 synchro_description = "I/O";
692 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
693 process->pid, process->name, sg_host_get_name(process->host),
694 synchro_description, process->waiting_synchro,
695 process->waiting_synchro->name, (int)process->waiting_synchro->state);
698 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
703 static void* SIMIX_synchro_mallocator_new_f(void) {
704 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
705 synchro->simcalls = xbt_fifo_new();
709 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
710 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
714 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
716 // we also recycle the simcall list
717 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
718 xbt_fifo_reset(fifo);
719 memset(synchro, 0, sizeof(s_smx_synchro_t));
720 ((smx_synchro_t) synchro)->simcalls = fifo;
723 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
724 return SIMIX_asr_get_properties(name);
726 xbt_dict_t SIMIX_asr_get_properties(const char *name)
728 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
731 int SIMIX_is_maestro()
733 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;