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. */
9 #include <signal.h> /* Signal handling */
11 #include "src/internal_config.h"
13 #include "src/surf/surf_interface.hpp"
14 #include "src/surf/storage_interface.hpp"
15 #include "src/surf/xml/platf.hpp"
16 #include "smx_private.h"
18 #include "xbt/ex.h" /* ex_backtrace_display */
20 #include "src/mc/mc_replay.h"
21 #include "simgrid/sg_config.h"
23 #include "src/simix/SynchroExec.hpp"
24 #include "src/simix/SynchroComm.hpp"
25 #include "src/simix/SynchroSleep.hpp"
26 #include "src/simix/SynchroIo.hpp"
27 #include "src/simix/SynchroRaw.hpp"
30 #include "src/mc/mc_private.h"
31 #include "src/mc/mc_protocol.h"
32 #include "src/mc/Client.hpp"
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, "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 {
53 std::function<void()> callback;
56 s_smx_timer(double date, std::function<void()> callback)
57 : date(date), callback(std::move(callback)) {}
60 void (*SMPI_switch_data_segment)(int) = NULL;
62 int _sg_do_verbose_exit = 1;
63 static void inthandler(int ignored)
65 if ( _sg_do_verbose_exit ) {
66 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
67 SIMIX_display_process_status();
70 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
76 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
78 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
80 "Access violation detected.\n"
81 "This can result from a programming error in your code or, although less likely,\n"
82 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
83 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
85 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
86 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
87 smx_context_stack_size / 1024);
88 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
90 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
91 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
93 } else if (siginfo->si_signo == SIGSEGV) {
94 fprintf(stderr, "Segmentation fault.\n");
96 if (smpi_enabled() && !smpi_privatize_global_variables) {
97 #if HAVE_PRIVATIZATION
99 "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
102 "Sadly, your system does not support --cfg=smpi/privatize-global-variables:yes (yet).\n");
103 #endif /* HAVE_PRIVATIZATION */
105 #endif /* HAVE_SMPI */
110 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
113 * Install signal handler for SIGSEGV. Check that nobody has already installed
114 * its own handler. For example, the Java VM does this.
116 static void install_segvhandler(void)
118 stack_t stack, old_stack;
119 stack.ss_sp = sigsegv_stack;
120 stack.ss_size = sizeof sigsegv_stack;
123 if (sigaltstack(&stack, &old_stack) == -1) {
124 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
127 if (!(old_stack.ss_flags & SS_DISABLE)) {
128 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
129 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
130 sigaltstack(&old_stack, NULL);
133 struct sigaction action, old_action;
134 action.sa_sigaction = &segvhandler;
135 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
136 sigemptyset(&action.sa_mask);
138 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
139 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
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);
152 /********************************* 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 std::function<void()> maestro_code;
168 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
170 maestro_code = std::move(code);
176 void SIMIX_set_maestro(void (*code)(void*), void* data)
178 maestro_code = std::bind(code, data);
183 * \brief Initialize SIMIX internal data.
188 void SIMIX_global_init(int *argc, char **argv)
191 // The communication initialization is done ASAP.
192 // We need to communicate initialization of the different layers to the model-checker.
193 simgrid::mc::Client::initialize();
197 simix_global = xbt_new0(s_smx_global_t, 1);
199 simgrid::simix::Process proc;
200 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
201 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
202 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
203 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
205 simix_global->maestro_process = NULL;
206 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
208 simix_global->create_process_function = &SIMIX_process_create;
209 simix_global->kill_process_function = &kill_process;
210 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
211 simix_global->mutex = xbt_os_mutex_init();
213 surf_init(argc, argv); /* Initialize SURF structures */
214 SIMIX_context_mod_init();
216 // Either create a new context with maestro or create
217 // a context object with the current context mestro):
218 simgrid::simix::create_maestro(maestro_code);
220 /* context exception handlers */
221 __xbt_running_ctx_fetch = &SIMIX_process_get_running_context;
222 __xbt_ex_terminate = &SIMIX_process_exception_terminate;
224 /* Prepare to display some more info when dying on Ctrl-C pressing */
225 signal(SIGINT, inthandler);
228 install_segvhandler();
230 /* register a function to be called by SURF after the environment creation */
232 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
233 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
234 SIMIX_host_create(&host);
236 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
238 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
239 const char* name = storage->getName();
240 // TODO, create sg_storage_by_name
241 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, name);
242 xbt_assert(s != NULL, "Storage not found for name %s", name);
244 SIMIX_storage_create(name, s, NULL);
247 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
250 simix_timers = xbt_heap_new(8, [](void* p) {
251 delete static_cast<smx_timer_t>(p);
254 if (xbt_cfg_get_boolean("clean-atexit"))
257 if (_sg_cfg_exit_asap)
264 * \brief Clean the SIMIX simulation
266 * This functions remove the memory used by SIMIX
268 void SIMIX_clean(void)
270 if (smx_cleaned) return; // to avoid double cleaning by java and C
272 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
273 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
275 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
276 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
277 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
279 /* Kill all processes (but maestro) */
280 SIMIX_process_killall(simix_global->maestro_process, 1);
282 /* Exit the SIMIX network module */
283 SIMIX_mailbox_exit();
285 xbt_heap_free(simix_timers);
287 /* Free the remaining data structures */
288 xbt_dynar_free(&simix_global->process_to_run);
289 xbt_dynar_free(&simix_global->process_that_ran);
290 xbt_swag_free(simix_global->process_to_destroy);
291 xbt_swag_free(simix_global->process_list);
292 simix_global->process_list = NULL;
293 simix_global->process_to_destroy = NULL;
294 xbt_dict_free(&(simix_global->registered_functions));
296 xbt_os_mutex_destroy(simix_global->mutex);
297 simix_global->mutex = NULL;
299 /* Let's free maestro now */
300 delete simix_global->maestro_process->context;
301 simix_global->maestro_process->context = nullptr;
302 xbt_free(simix_global->maestro_process->running_ctx);
303 simix_global->maestro_process->running_ctx = nullptr;
304 delete simix_global->maestro_process;
305 simix_global->maestro_process = NULL;
307 /* Restore the default exception setup */
308 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
309 __xbt_ex_terminate = &__xbt_ex_terminate_default;
311 /* Finish context module and SURF */
312 SIMIX_context_mod_exit();
316 xbt_free(simix_global);
325 * \brief A clock (in second).
327 * \return Return the clock.
329 double SIMIX_get_clock(void)
331 if(MC_is_active() || MC_record_replay_is_active()){
332 return MC_process_clock_get(SIMIX_process_self());
334 return surf_get_clock();
338 static int process_syscall_color(void *p)
340 switch ((*(smx_process_t *)p)->simcall.call) {
342 case SIMCALL_PROCESS_KILL:
344 case SIMCALL_PROCESS_RESUME:
353 * \brief Run the main simulation loop.
357 if (MC_record_path) {
358 simgrid::mc::replay(MC_record_path);
363 smx_process_t process;
364 surf_action_t action;
370 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
371 xbt_dynar_length(simix_global->process_to_run));
372 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
373 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
374 xbt_dynar_length(simix_global->process_to_run));
376 /* Run all processes that are ready to run, possibly in parallel */
377 SIMIX_process_runall();
379 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
380 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
382 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
384 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
386 /* Here, the order is ok because:
388 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
390 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
392 * - if there is no kill during the simulation, processes remain sorted according by their PID.
393 * rational: This can be proved inductively.
394 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
395 * Let's show that it is still so at the end of this round.
396 * - if a process is added when being created, that's from maestro. It can be either at startup
397 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
398 * in arbitrary order (inductive hypothesis), we are fine.
399 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
400 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
401 * of a simcall terminates, we're still good. Proof:
402 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
403 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
404 * - For communications, this function is called from SIMIX_comm_finish().
405 * This function itself don't mess with the order since simcalls are handled in FIFO order.
406 * The function is called:
407 * - before the comm starts (invalid parameters, or resource already dead or whatever).
408 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
409 * - because the communication failed or were canceled after startup. In this case, it's called from the function
410 * we are in, by the chunk:
411 * set = model->states.failed_action_set;
412 * while ((synchro = xbt_swag_extract(set)))
413 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
414 * This order is also fixed because it depends of the order in which the surf actions were
415 * added to the system, and only maestro can add stuff this way, through simcalls.
416 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
417 * poped out of the swag does not depend on the user code's execution order.
418 * - because the communication terminated. In this case, synchros are served in the order given by
419 * set = model->states.done_action_set;
420 * while ((synchro = xbt_swag_extract(set)))
421 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
422 * and the argument is very similar to the previous one.
423 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
424 * 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.
425 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
426 * - If there is some process killings, the order is changed by this decision that comes from user-land
427 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
428 * So, even the order change induced by the process killing is perfectly reproducible.
430 * So science works, bitches [http://xkcd.com/54/].
432 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
433 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
434 * That would thus be a pure waste of time.
437 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
438 if (process->simcall.call != SIMCALL_NONE) {
439 SIMIX_simcall_handle(&process->simcall, 0);
442 /* Wake up all processes waiting for a Surf action to finish */
443 xbt_dynar_foreach(all_existing_models, iter, model) {
444 XBT_DEBUG("Handling process whose action failed");
445 while ((action = surf_model_extract_failed_action_set(model))) {
446 XBT_DEBUG(" Handling Action %p",action);
447 SIMIX_simcall_exit((smx_synchro_t) action->getData());
449 XBT_DEBUG("Handling process whose action terminated normally");
450 while ((action = surf_model_extract_done_action_set(model))) {
451 XBT_DEBUG(" Handling Action %p",action);
452 if (action->getData() == NULL)
453 XBT_DEBUG("probably vcpu's action %p, skip", action);
455 SIMIX_simcall_exit((smx_synchro_t) action->getData());
460 time = SIMIX_timer_next();
461 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
462 XBT_DEBUG("Calling surf_solve");
463 time = surf_solve(time);
464 XBT_DEBUG("Moving time ahead : %g", time);
466 /* Notify all the hosts that have failed */
467 /* FIXME: iterate through the list of failed host and mark each of them */
468 /* as failed. On each host, signal all the running processes with host_fail */
470 /* Handle any pending timer */
471 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
472 //FIXME: make the timers being real callbacks
473 // (i.e. provide dispatchers that read and expand the args)
474 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
475 if (timer->callback) {
480 xbt_die("Exception throwed ouf of timer callback");
486 /* Wake up all processes waiting for a Surf action to finish */
487 xbt_dynar_foreach(all_existing_models, iter, model) {
488 XBT_DEBUG("Handling process whose action failed");
489 while ((action = surf_model_extract_failed_action_set(model))) {
490 XBT_DEBUG(" Handling Action %p",action);
491 SIMIX_simcall_exit((smx_synchro_t) action->getData());
493 XBT_DEBUG("Handling process whose action terminated normally");
494 while ((action = surf_model_extract_done_action_set(model))) {
495 XBT_DEBUG(" Handling Action %p",action);
496 if (action->getData() == NULL)
497 XBT_DEBUG("probably vcpu's action %p, skip", action);
499 SIMIX_simcall_exit((smx_synchro_t) action->getData());
503 /* Autorestart all process */
504 char *hostname = NULL;
505 xbt_dynar_foreach(host_that_restart,iter,hostname) {
506 XBT_INFO("Restart processes on host: %s",hostname);
507 SIMIX_host_autorestart(sg_host_by_name(hostname));
509 xbt_dynar_reset(host_that_restart);
511 /* Clean processes to destroy */
512 SIMIX_process_empty_trash();
515 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
517 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
519 if (xbt_swag_size(simix_global->process_list) != 0) {
523 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
524 SIMIX_display_process_status();
530 * \brief Set the date to execute a function
532 * Set the date to execute the function on the surf.
533 * \param date Date to execute function
534 * \param function Function to be executed
535 * \param arg Parameters of the function
538 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
540 smx_timer_t timer = new s_smx_timer_t(date, std::bind(function, arg));
541 xbt_heap_push(simix_timers, timer, date);
545 smx_timer_t SIMIX_timer_set(double date, std::function<void()> callback)
547 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
548 xbt_heap_push(simix_timers, timer, date);
552 /** @brief cancels a timer that was added earlier */
553 void SIMIX_timer_remove(smx_timer_t timer) {
554 xbt_heap_rm_elm(simix_timers, timer, timer->date);
557 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
558 double SIMIX_timer_get_date(smx_timer_t timer) {
559 return timer?timer->date:0;
563 * \brief Registers a function to create a process.
565 * This function registers a function to be called
566 * when a new process is created. The function has
567 * to call SIMIX_process_create().
568 * \param function create process function
570 void SIMIX_function_register_process_create(smx_creation_func_t function)
572 simix_global->create_process_function = function;
576 * \brief Registers a function to kill a process.
578 * This function registers a function to be called when a
579 * process is killed. The function has to call the SIMIX_process_kill().
581 * \param function Kill process function
583 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
586 simix_global->kill_process_function = function;
590 * \brief Registers a function to cleanup a process.
592 * This function registers a user function to be called when
593 * a process ends properly.
595 * \param function cleanup process function
597 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
600 simix_global->cleanup_process_function = function;
604 void SIMIX_display_process_status(void)
606 if (simix_global->process_list == NULL) {
610 smx_process_t process = NULL;
611 int nbprocess = xbt_swag_size(simix_global->process_list);
613 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
614 /* List the process and their state */
616 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
617 xbt_swag_foreach(process, simix_global->process_list) {
619 if (process->waiting_synchro) {
621 const char* synchro_description = "unknown";
623 if (dynamic_cast<simgrid::simix::Exec*>(process->waiting_synchro) != nullptr)
624 synchro_description = "execution";
626 if (dynamic_cast<simgrid::simix::Comm*>(process->waiting_synchro) != nullptr)
627 synchro_description = "communication";
629 if (dynamic_cast<simgrid::simix::Sleep*>(process->waiting_synchro) != nullptr)
630 synchro_description = "sleeping";
632 if (dynamic_cast<simgrid::simix::Raw*>(process->waiting_synchro) != nullptr)
633 synchro_description = "synchronization";
635 if (dynamic_cast<simgrid::simix::Io*>(process->waiting_synchro) != nullptr)
636 synchro_description = "I/O";
640 switch (process->waiting_synchro->type) {
641 case SIMIX_SYNC_PARALLEL_EXECUTE:
642 synchro_description = "parallel execution";
645 case SIMIX_SYNC_JOIN:
646 synchro_description = "joining";
650 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
651 process->pid, process->name.c_str(), sg_host_get_name(process->host),
652 synchro_description, process->waiting_synchro,
653 process->waiting_synchro->name, (int)process->waiting_synchro->state);
656 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name.c_str(), sg_host_get_name(process->host));
661 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
662 return SIMIX_asr_get_properties(name);
664 xbt_dict_t SIMIX_asr_get_properties(const char *name)
666 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
669 int SIMIX_is_maestro()
671 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;