1 /* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "simgrid/s4u/Engine.hpp"
8 #include "simgrid/s4u/Host.hpp"
9 #include "src/smpi/include/smpi_actor.hpp"
11 #include "simgrid/sg_config.hpp"
12 #include "src/kernel/activity/ExecImpl.hpp"
13 #include "src/kernel/activity/IoImpl.hpp"
14 #include "src/kernel/activity/MailboxImpl.hpp"
15 #include "src/kernel/activity/SleepImpl.hpp"
16 #include "src/kernel/activity/SynchroRaw.hpp"
17 #include "src/mc/mc_record.hpp"
18 #include "src/mc/mc_replay.hpp"
19 #include "src/simix/smx_host_private.hpp"
20 #include "src/simix/smx_private.hpp"
21 #include "src/surf/StorageImpl.hpp"
22 #include "src/surf/xml/platf.hpp"
25 #include "src/mc/remote/Client.hpp"
29 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
30 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
32 std::unique_ptr<simgrid::simix::Global> simix_global;
34 void (*SMPI_switch_data_segment)(simgrid::s4u::ActorPtr) = nullptr;
36 bool _sg_do_verbose_exit = true;
37 static void inthandler(int)
39 if ( _sg_do_verbose_exit ) {
40 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
41 SIMIX_display_process_status();
44 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
50 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
52 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
53 fprintf(stderr, "Access violation detected.\n"
54 "This probably comes from a programming error in your code, or from a stack\n"
55 "overflow. If you are certain of your code, try increasing the stack size\n"
56 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
58 "If it does not help, this may have one of the following causes:\n"
59 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
60 "Failing hardware can sometimes generate such errors too.\n"
62 "If you think you've found a bug in SimGrid, please report it along with a\n"
63 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
64 "of the fault captured with gdb or valgrind.\n",
65 smx_context_stack_size / 1024);
66 } else if (siginfo->si_signo == SIGSEGV) {
67 fprintf(stderr, "Segmentation fault.\n");
69 if (smpi_enabled() && smpi_privatize_global_variables == SmpiPrivStrategies::NONE) {
70 #if HAVE_PRIVATIZATION
71 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
73 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
74 #endif /* HAVE_PRIVATIZATION */
76 #endif /* HAVE_SMPI */
81 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
84 * Install signal handler for SIGSEGV. Check that nobody has already installed
85 * its own handler. For example, the Java VM does this.
87 static void install_segvhandler()
91 stack.ss_sp = sigsegv_stack;
92 stack.ss_size = sizeof sigsegv_stack;
95 if (sigaltstack(&stack, &old_stack) == -1) {
96 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
99 if (not(old_stack.ss_flags & SS_DISABLE)) {
100 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
101 old_stack.ss_size, (unsigned)old_stack.ss_flags);
102 sigaltstack(&old_stack, nullptr);
105 struct sigaction action;
106 struct sigaction old_action;
107 action.sa_sigaction = &segvhandler;
108 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
109 sigemptyset(&action.sa_mask);
111 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
112 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
115 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
116 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
117 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
118 sigaction(SIGSEGV, &old_action, nullptr);
124 /********************************* SIMIX **************************************/
128 Timer* Timer::set(double date, void (*callback)(void*), void* arg)
130 Timer* timer = new Timer(date, simgrid::xbt::make_task([callback, arg]() { callback(arg); }));
131 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
135 Timer* Timer::set(double date, simgrid::xbt::Task<void()> callback)
137 Timer* timer = new Timer(date, std::move(callback));
138 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
142 /** @brief cancels a timer that was added earlier */
145 simgrid::simix::simix_timers.erase(handle_);
149 void Global::empty_trash()
151 while (not actors_to_destroy.empty()) {
152 smx_actor_t actor = &actors_to_destroy.front();
153 actors_to_destroy.pop_front();
154 XBT_DEBUG("Getting rid of %p", actor);
155 intrusive_ptr_release(actor);
158 xbt_dynar_reset(simix_global->dead_actors_vector);
162 * @brief Executes the actors in simix_global->actors_to_run.
164 * The actors in simix_global->actors_to_run are run (in parallel if possible). On exit, simix_global->actors_to_run
165 * is empty, and simix_global->actors_that_ran contains the list of actors that just ran.
166 * The two lists are swapped so, be careful when using them before and after a call to this function.
168 void Global::run_all_actors()
170 SIMIX_context_runall();
172 simix_global->actors_to_run.swap(simix_global->actors_that_ran);
173 simix_global->actors_to_run.clear();
176 simgrid::config::Flag<double> breakpoint{"simix/breakpoint",
177 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
181 static simgrid::simix::ActorCode maestro_code;
182 void SIMIX_set_maestro(void (*code)(void*), void* data)
185 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
187 maestro_code = std::bind(code, data);
192 * @brief Initialize SIMIX internal data.
194 void SIMIX_global_init(int *argc, char **argv)
197 // The communication initialization is done ASAP.
198 // We need to communicate initialization of the different layers to the model-checker.
199 simgrid::mc::Client::initialize();
202 if (simix_global == nullptr) {
203 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
204 simix_global->maestro_process = nullptr;
206 surf_init(argc, argv); /* Initialize SURF structures */
207 SIMIX_context_mod_init();
209 // Either create a new context with maestro or create
210 // a context object with the current context mestro):
211 simgrid::kernel::actor::create_maestro(maestro_code);
213 /* Prepare to display some more info when dying on Ctrl-C pressing */
214 std::signal(SIGINT, inthandler);
217 install_segvhandler();
219 /* register a function to be called by SURF after the environment creation */
221 simgrid::s4u::on_platform_created.connect(surf_presolve);
223 simgrid::s4u::Storage::on_creation.connect([](simgrid::s4u::Storage& storage) {
224 sg_storage_t s = simgrid::s4u::Storage::by_name(storage.get_name());
225 xbt_assert(s != nullptr, "Storage not found for name %s", storage.get_cname());
229 if (simgrid::config::get_value<bool>("clean-atexit"))
232 if (_sg_cfg_exit_asap)
239 * @brief Clean the SIMIX simulation
241 * This functions remove the memory used by SIMIX
246 return; // to avoid double cleaning by java and C
249 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
250 if (not simix_global->actors_to_run.empty() && SIMIX_get_clock() <= 0.0) {
252 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
253 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
254 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
258 if (SIMIX_process_count()>0){
259 if(smpi_process()->initialized()){
260 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
262 XBT_WARN("Process called exit when leaving - Skipping cleanups");
268 /* Kill all processes (but maestro) */
269 simix_global->maestro_process->kill_all();
270 SIMIX_context_runall();
271 simix_global->empty_trash();
273 /* Exit the SIMIX network module */
274 SIMIX_mailbox_exit();
276 while (not simgrid::simix::simix_timers.empty()) {
277 delete simgrid::simix::simix_timers.top().second;
278 simgrid::simix::simix_timers.pop();
280 /* Free the remaining data structures */
281 simix_global->actors_to_run.clear();
282 simix_global->actors_that_ran.clear();
283 simix_global->actors_to_destroy.clear();
284 simix_global->process_list.clear();
287 xbt_dynar_free(&simix_global->actors_vector);
288 xbt_dynar_free(&simix_global->dead_actors_vector);
291 /* Let's free maestro now */
292 delete simix_global->maestro_process->context_;
293 simix_global->maestro_process->context_ = nullptr;
294 delete simix_global->maestro_process;
295 simix_global->maestro_process = nullptr;
297 /* Finish context module and SURF */
298 SIMIX_context_mod_exit();
302 simix_global = nullptr;
307 * @brief A clock (in second).
309 * @return Return the clock.
311 double SIMIX_get_clock()
313 if(MC_is_active() || MC_record_replay_is_active()){
314 return MC_process_clock_get(SIMIX_process_self());
316 return surf_get_clock();
320 /** Wake up all processes waiting for a Surf action to finish */
321 static void SIMIX_wake_processes()
323 for (auto const& model : all_existing_models) {
324 simgrid::kernel::resource::Action* action;
326 XBT_DEBUG("Handling the processes whose action failed (if any)");
327 while ((action = model->extract_failed_action())) {
328 XBT_DEBUG(" Handling Action %p",action);
329 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
331 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
332 while ((action = model->extract_done_action())) {
333 XBT_DEBUG(" Handling Action %p",action);
334 if (action->get_data() == nullptr)
335 XBT_DEBUG("probably vcpu's action %p, skip", action);
337 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
342 /** Handle any pending timer */
343 static bool SIMIX_execute_timers()
346 while (not simgrid::simix::simix_timers.empty() && SIMIX_get_clock() >= simgrid::simix::simix_timers.top().first) {
348 // FIXME: make the timers being real callbacks (i.e. provide dispatchers that read and expand the args)
349 smx_timer_t timer = simgrid::simix::simix_timers.top().second;
350 simgrid::simix::simix_timers.pop();
354 xbt_die("Exception thrown ouf of timer callback");
361 /** Execute all the tasks that are queued
363 * e.g. `.then()` callbacks of futures.
365 static bool SIMIX_execute_tasks()
367 xbt_assert(simix_global->tasksTemp.empty());
369 if (simix_global->tasks.empty())
374 // We don't want the callbacks to modify the vector we are iterating over:
375 swap(simix_global->tasks, simix_global->tasksTemp);
377 // Execute all the queued tasks:
378 for (auto& task : simix_global->tasksTemp)
381 simix_global->tasksTemp.clear();
382 } while (not simix_global->tasks.empty());
389 * @brief Run the main simulation loop.
393 if (not MC_record_path.empty()) {
394 simgrid::mc::replay(MC_record_path);
401 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
403 if (simgrid::simix::breakpoint >= 0.0 && surf_get_clock() >= simgrid::simix::breakpoint) {
404 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::breakpoint.get());
405 simgrid::simix::breakpoint = -1.0;
413 SIMIX_execute_tasks();
415 while (not simix_global->actors_to_run.empty()) {
416 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
418 /* Run all processes that are ready to run, possibly in parallel */
419 simix_global->run_all_actors();
421 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
423 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
425 /* Here, the order is ok because:
427 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
428 * not impact its order.
430 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
432 * - if there is no kill during the simulation, processes remain sorted according by their PID.
433 * Rationale: This can be proved inductively.
434 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
435 * is parsed linearly).
436 * Let's show that it is still so at the end of this round.
437 * - if a process is added when being created, that's from maestro. It can be either at startup
438 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
439 * in arbitrary order (inductive hypothesis), we are fine.
440 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
441 * - If a process gets added to actors_to_run because one of their blocking action constituting the meat
442 * of a simcall terminates, we're still good. Proof:
443 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
444 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an
446 * - For communications, this function is called from SIMIX_comm_finish().
447 * This function itself don't mess with the order since simcalls are handled in FIFO order.
448 * The function is called:
449 * - before the comm starts (invalid parameters, or resource already dead or whatever).
450 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
451 * - because the communication failed or were canceled after startup. In this case, it's called from
452 * the function we are in, by the chunk:
453 * set = model->states.failed_action_set;
454 * while ((synchro = extract(set)))
455 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
456 * This order is also fixed because it depends of the order in which the surf actions were
457 * added to the system, and only maestro can add stuff this way, through simcalls.
458 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
459 * poped out of the set does not depend on the user code's execution order.
460 * - because the communication terminated. In this case, synchros are served in the order given by
461 * set = model->states.done_action_set;
462 * while ((synchro = extract(set)))
463 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
464 * and the argument is very similar to the previous one.
465 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user
466 * processes are executed.
467 * So, in any cases, the orders of processes within actors_to_run do not depend on the order in which
468 * user processes were executed previously.
469 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
470 * - If there is some process killings, the order is changed by this decision that comes from user-land
471 * But this decision may not have been motivated by a situation that were different because the simulation is
473 * So, even the order change induced by the process killing is perfectly reproducible.
475 * So science works, bitches [http://xkcd.com/54/].
477 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
478 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
480 * That would thus be a pure waste of time.
483 for (smx_actor_t const& process : simix_global->actors_that_ran) {
484 if (process->simcall.call != SIMCALL_NONE) {
485 SIMIX_simcall_handle(&process->simcall, 0);
489 SIMIX_execute_tasks();
491 SIMIX_wake_processes();
492 } while (SIMIX_execute_tasks());
494 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
495 if (simix_global->process_list.size() == simix_global->daemons.size())
496 for (auto const& dmon : simix_global->daemons) {
497 XBT_DEBUG("Kill %s", dmon->get_cname());
498 simix_global->maestro_process->kill(dmon);
502 time = simgrid::simix::Timer::next();
503 if (time > -1.0 || not simix_global->process_list.empty()) {
504 XBT_DEBUG("Calling surf_solve");
505 time = surf_solve(time);
506 XBT_DEBUG("Moving time ahead : %g", time);
509 /* Notify all the hosts that have failed */
510 /* FIXME: iterate through the list of failed host and mark each of them */
511 /* as failed. On each host, signal all the running processes with host_fail */
513 // Execute timers and tasks until there isn't anything to be done:
516 again = SIMIX_execute_timers();
517 if (SIMIX_execute_tasks())
519 SIMIX_wake_processes();
522 /* Clean actors to destroy */
523 simix_global->empty_trash();
525 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
526 simix_global->actors_to_run.size());
528 } while (time > -1.0 || not simix_global->actors_to_run.empty());
530 if (not simix_global->process_list.empty()) {
532 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
533 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
534 "once the simulation is over. Please fix your on_exit() functions.");
536 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
538 SIMIX_display_process_status();
539 simgrid::s4u::on_deadlock();
542 simgrid::s4u::on_simulation_end();
545 double SIMIX_timer_next()
547 return simgrid::simix::Timer::next();
550 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
552 return simgrid::simix::Timer::set(date, callback, arg);
555 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
557 smx_timer_t timer = new simgrid::simix::Timer(date, std::move(callback));
558 timer->handle_ = simgrid::simix::simix_timers.emplace(std::make_pair(date, timer));
562 /** @brief cancels a timer that was added earlier */
563 void SIMIX_timer_remove(smx_timer_t timer) {
567 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
568 double SIMIX_timer_get_date(smx_timer_t timer) {
569 return timer ? timer->get_date() : 0;
572 void SIMIX_display_process_status()
574 int nbprocess = simix_global->process_list.size();
576 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
577 /* List the process and their state */
578 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
579 for (auto const& kv : simix_global->process_list) {
580 smx_actor_t process = kv.second;
582 if (process->waiting_synchro) {
584 const char* synchro_description = "unknown";
586 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
587 synchro_description = "execution";
589 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
590 synchro_description = "communication";
592 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
593 synchro_description = "sleeping";
595 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
596 synchro_description = "synchronization";
598 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
599 synchro_description = "I/O";
601 XBT_INFO("Process %ld (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->get_pid(),
602 process->get_cname(), process->get_host()->get_cname(), synchro_description,
603 process->waiting_synchro.get(), process->waiting_synchro->get_cname(),
604 (int)process->waiting_synchro->state_);
607 XBT_INFO("Process %ld (%s@%s)", process->get_pid(), process->get_cname(), process->get_host()->get_cname());
612 int SIMIX_is_maestro()
614 smx_actor_t self = SIMIX_process_self();
615 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;