1 /* Copyright (c) 2007-2021. 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/EngineImpl.hpp"
13 #include "src/kernel/activity/ExecImpl.hpp"
14 #include "src/kernel/activity/IoImpl.hpp"
15 #include "src/kernel/activity/MailboxImpl.hpp"
16 #include "src/kernel/activity/SleepImpl.hpp"
17 #include "src/kernel/activity/SynchroRaw.hpp"
18 #include "src/mc/mc_record.hpp"
19 #include "src/mc/mc_replay.hpp"
20 #include "src/simix/smx_private.hpp"
21 #include "src/surf/xml/platf.hpp"
23 #include "simgrid/kernel/resource/Model.hpp"
26 #include "src/mc/remote/AppSide.hpp"
31 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
32 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
34 std::unique_ptr<simgrid::simix::Global> simix_global;
36 void (*SMPI_switch_data_segment)(simgrid::s4u::ActorPtr) = nullptr;
40 config::Flag<bool> cfg_verbose_exit{"debug/verbose-exit", "Display the actor status at exit", true};
42 void* simix_global_get_actors_addr()
45 return simix_global->actors_vector;
47 xbt_die("This function is intended to be used when compiling with MC");
50 void* simix_global_get_dead_actors_addr()
53 return simix_global->dead_actors_vector;
55 xbt_die("This function is intended to be used when compiling with MC");
60 } // namespace simgrid
62 XBT_ATTRIB_NORETURN static void inthandler(int)
64 if (simgrid::simix::cfg_verbose_exit) {
65 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option "
66 "'debug/verbose-exit').");
67 simix_global->display_all_actor_status();
69 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'debug/verbose-exit' is set to false.");
75 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
77 if ((siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) || siginfo->si_signo == SIGBUS) {
79 "Access violation or Bus error detected.\n"
80 "This probably comes from a programming error in your code, or from a stack\n"
81 "overflow. If you are certain of your code, try increasing the stack size\n"
82 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
84 "If it does not help, this may have one of the following causes:\n"
85 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
86 "Failing hardware can sometimes generate such errors too.\n"
88 "If you think you've found a bug in SimGrid, please report it along with a\n"
89 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
90 "of the fault captured with gdb or valgrind.\n",
91 smx_context_stack_size / 1024);
92 } else if (siginfo->si_signo == SIGSEGV) {
93 fprintf(stderr, "Segmentation fault.\n");
95 if (smpi_enabled() && smpi_cfg_privatization() == SmpiPrivStrategies::NONE) {
96 #if HAVE_PRIVATIZATION
97 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
99 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
100 #endif /* HAVE_PRIVATIZATION */
102 #endif /* HAVE_SMPI */
108 * Install signal handler for SIGSEGV. Check that nobody has already installed
109 * its own handler. For example, the Java VM does this.
111 static void install_segvhandler()
115 if (simgrid::kernel::context::Context::install_sigsegv_stack(&old_stack, true) == -1) {
116 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
119 if (not(old_stack.ss_flags & SS_DISABLE)) {
120 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
121 old_stack.ss_size, (unsigned)old_stack.ss_flags);
122 sigaltstack(&old_stack, nullptr);
125 struct sigaction action;
126 struct sigaction old_action;
127 action.sa_sigaction = &segvhandler;
128 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
129 sigemptyset(&action.sa_mask);
131 /* Linux tend to raise only SIGSEGV where other systems also raise SIGBUS on severe error */
132 for (int sig : {SIGSEGV, SIGBUS}) {
133 if (sigaction(sig, &action, &old_action) == -1) {
134 XBT_WARN("Failed to register signal handler for signal %d: %s", sig, strerror(errno));
137 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
138 XBT_DEBUG("A signal handler was already installed for signal %d (%p). Restore it.", sig,
139 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
140 sigaction(sig, &old_action, nullptr);
147 /********************************* SIMIX **************************************/
151 Timer* Timer::set(double date, xbt::Task<void()>&& callback)
153 auto* timer = new Timer(date, std::move(callback));
154 timer->handle_ = simix_timers().emplace(std::make_pair(date, timer));
158 /** @brief cancels a timer that was added earlier */
161 simix_timers().erase(handle_);
165 /** Execute all the tasks that are queued, e.g. `.then()` callbacks of futures. */
166 bool Global::execute_tasks()
168 xbt_assert(tasksTemp.empty());
174 // We don't want the callbacks to modify the vector we are iterating over:
175 tasks.swap(tasksTemp);
177 // Execute all the queued tasks:
178 for (auto& task : tasksTemp)
182 } while (not tasks.empty());
187 void Global::empty_trash()
189 while (not actors_to_destroy.empty()) {
190 kernel::actor::ActorImpl* actor = &actors_to_destroy.front();
191 actors_to_destroy.pop_front();
192 XBT_DEBUG("Getting rid of %s (refcount: %d)", actor->get_cname(), actor->get_refcount());
193 intrusive_ptr_release(actor);
196 xbt_dynar_reset(dead_actors_vector);
200 * @brief Executes the actors in actors_to_run.
202 * The actors in actors_to_run are run (in parallel if possible). On exit, actors_to_run is empty, and actors_that_ran
203 * contains the list of actors that just ran. The two lists are swapped so, be careful when using them before and after
204 * a call to this function.
206 void Global::run_all_actors()
208 simix_global->context_factory->run_all();
210 actors_to_run.swap(actors_that_ran);
211 actors_to_run.clear();
214 /** Wake up all actors waiting for a Surf action to finish */
215 void Global::wake_all_waiting_actors() const
217 for (auto const& model : simgrid::kernel::EngineImpl::get_instance()->get_all_models()) {
218 XBT_DEBUG("Handling the failed actions (if any)");
219 while (auto* action = model->extract_failed_action()) {
220 XBT_DEBUG(" Handling Action %p", action);
221 if (action->get_activity() != nullptr)
222 kernel::activity::ActivityImplPtr(action->get_activity())->post();
224 XBT_DEBUG("Handling the terminated actions (if any)");
225 while (auto* action = model->extract_done_action()) {
226 XBT_DEBUG(" Handling Action %p", action);
227 if (action->get_activity() == nullptr)
228 XBT_DEBUG("probably vcpu's action %p, skip", action);
230 kernel::activity::ActivityImplPtr(action->get_activity())->post();
235 void Global::display_all_actor_status() const
237 XBT_INFO("%zu actors are still running, waiting for something.", process_list.size());
238 /* List the actors and their state */
239 XBT_INFO("Legend of the following listing: \"Actor <pid> (<name>@<host>): <status>\"");
240 for (auto const& kv : process_list) {
241 kernel::actor::ActorImpl* actor = kv.second;
243 if (actor->waiting_synchro_) {
244 const char* synchro_description = "unknown";
246 if (boost::dynamic_pointer_cast<kernel::activity::ExecImpl>(actor->waiting_synchro_) != nullptr)
247 synchro_description = "execution";
249 if (boost::dynamic_pointer_cast<kernel::activity::CommImpl>(actor->waiting_synchro_) != nullptr)
250 synchro_description = "communication";
252 if (boost::dynamic_pointer_cast<kernel::activity::SleepImpl>(actor->waiting_synchro_) != nullptr)
253 synchro_description = "sleeping";
255 if (boost::dynamic_pointer_cast<kernel::activity::RawImpl>(actor->waiting_synchro_) != nullptr)
256 synchro_description = "synchronization";
258 if (boost::dynamic_pointer_cast<kernel::activity::IoImpl>(actor->waiting_synchro_) != nullptr)
259 synchro_description = "I/O";
261 XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %d to finish", actor->get_pid(),
262 actor->get_cname(), actor->get_host()->get_cname(), synchro_description,
263 (xbt_log_no_loc ? (size_t)0xDEADBEEF : (size_t)actor->waiting_synchro_.get()),
264 actor->waiting_synchro_->get_cname(), (int)actor->waiting_synchro_->state_);
266 XBT_INFO("Actor %ld (%s@%s) simcall %s", actor->get_pid(), actor->get_cname(), actor->get_host()->get_cname(), SIMIX_simcall_name(actor->simcall_.call_));
271 config::Flag<double> cfg_breakpoint{"debug/breakpoint",
272 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
274 } // namespace simgrid
276 static simgrid::kernel::actor::ActorCode maestro_code;
277 void SIMIX_set_maestro(void (*code)(void*), void* data)
280 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if "
281 "you need that feature");
283 maestro_code = std::bind(code, data);
286 void SIMIX_global_init(int* argc, char** argv)
288 if (simix_global == nullptr) {
289 simix_global = std::make_unique<simgrid::simix::Global>();
292 // The communication initialization is done ASAP, as we need to get some init parameters from the MC for different layers.
293 // But simix_global needs to be created, as we send the address of some of its fields to the MC that wants to read them directly.
294 simgrid::mc::AppSide::initialize();
297 surf_init(argc, argv); /* Initialize SURF structures */
299 simix_global->maestro_ = nullptr;
300 SIMIX_context_mod_init();
302 // Either create a new context with maestro or create
303 // a context object with the current context maestro):
304 simgrid::kernel::actor::create_maestro(maestro_code);
306 /* Prepare to display some more info when dying on Ctrl-C pressing */
307 std::signal(SIGINT, inthandler);
310 install_segvhandler();
312 /* register a function to be called by SURF after the environment creation */
314 simgrid::s4u::Engine::on_platform_created.connect(surf_presolve);
317 if (simgrid::config::get_value<bool>("debug/clean-atexit"))
323 * @brief Clean the SIMIX simulation
325 * This functions remove the memory used by SIMIX
329 static bool smx_cleaned = false;
331 return; // to avoid double cleaning by java and C
334 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
335 if (not simix_global->actors_to_run.empty() && SIMIX_get_clock() <= 0.0) {
337 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
338 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
339 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
343 if (not simix_global->process_list.empty()) {
344 if (smpi_process()->initialized()) {
345 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
347 XBT_WARN("Process called exit when leaving - Skipping cleanups");
353 /* Kill all processes (but maestro) */
354 simix_global->maestro_->kill_all();
355 simix_global->run_all_actors();
356 simix_global->empty_trash();
358 /* Exit the SIMIX network module */
359 SIMIX_mailbox_exit();
361 while (not simgrid::simix::simix_timers().empty()) {
362 delete simgrid::simix::simix_timers().top().second;
363 simgrid::simix::simix_timers().pop();
365 /* Free the remaining data structures */
366 simix_global->actors_to_run.clear();
367 simix_global->actors_that_ran.clear();
368 simix_global->actors_to_destroy.clear();
369 simix_global->process_list.clear();
372 xbt_dynar_free(&simix_global->actors_vector);
373 xbt_dynar_free(&simix_global->dead_actors_vector);
376 /* Let's free maestro now */
377 delete simix_global->maestro_;
378 simix_global->maestro_ = nullptr;
380 /* Finish context module and SURF */
381 SIMIX_context_mod_exit();
385 simix_global = nullptr;
390 * @brief A clock (in second).
392 * @return Return the clock.
394 double SIMIX_get_clock()
396 if (MC_is_active() || MC_record_replay_is_active()) {
397 return MC_process_clock_get(SIMIX_process_self());
399 return surf_get_clock();
403 /** Handle any pending timer. Returns if something was actually run. */
404 static bool SIMIX_execute_timers()
407 while (not simgrid::simix::simix_timers().empty() &&
408 SIMIX_get_clock() >= simgrid::simix::simix_timers().top().first) {
410 // FIXME: make the timers being real callbacks (i.e. provide dispatchers that read and expand the args)
411 smx_timer_t timer = simgrid::simix::simix_timers().top().second;
412 simgrid::simix::simix_timers().pop();
421 * @brief Run the main simulation loop.
425 if (MC_record_replay_is_active()) {
426 simgrid::mc::replay(MC_record_path());
433 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
435 if (simgrid::simix::cfg_breakpoint >= 0.0 && surf_get_clock() >= simgrid::simix::cfg_breakpoint) {
436 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::cfg_breakpoint.get());
437 simgrid::simix::cfg_breakpoint = -1.0;
445 simix_global->execute_tasks();
447 while (not simix_global->actors_to_run.empty()) {
448 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
450 /* Run all processes that are ready to run, possibly in parallel */
451 simix_global->run_all_actors();
453 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
455 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
457 /* Here, the order is ok because:
459 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
460 * not impact its order.
462 * Long proof: actors remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
464 * - if there is no kill during the simulation, actors remain sorted according by their PID.
465 * Rationale: This can be proved inductively.
466 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
467 * is parsed linearly).
468 * Let's show that it is still so at the end of this round.
469 * - if an actor is added when being created, that's from maestro. It can be either at startup
470 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
471 * in arbitrary order (inductive hypothesis), we are fine.
472 * - If an actor is added because it's getting killed, its subsequent actions shouldn't matter
473 * - If an actor gets added to actors_to_run because one of their blocking action constituting the meat
474 * of a simcall terminates, we're still good. Proof:
475 * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the
476 * resource kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications
478 * - For communications, this function is called from SIMIX_comm_finish().
479 * This function itself don't mess with the order since simcalls are handled in FIFO order.
480 * The function is called:
481 * - before the comm starts (invalid parameters, or resource already dead or whatever).
482 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
483 * - because the communication failed or were canceled after startup. In this case, it's called from
484 * the function we are in, by the chunk:
485 * set = model->states.failed_action_set;
486 * while ((synchro = extract(set)))
487 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
488 * This order is also fixed because it depends of the order in which the surf actions were
489 * added to the system, and only maestro can add stuff this way, through simcalls.
490 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
491 * popped out of the set does not depend on the user code's execution order.
492 * - because the communication terminated. In this case, synchros are served in the order given by
493 * set = model->states.done_action_set;
494 * while ((synchro = extract(set)))
495 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
496 * and the argument is very similar to the previous one.
497 * So, in any case, the orders of calls to CommImpl::finish() do not depend on the order in which user
498 * actors are executed.
499 * So, in any cases, the orders of actors within actors_to_run do not depend on the order in which
500 * user actors were executed previously.
501 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
502 * - If there is some actor killings, the order is changed by this decision that comes from user-land
503 * But this decision may not have been motivated by a situation that were different because the simulation is
505 * So, even the order change induced by the actor killing is perfectly reproducible.
507 * So science works, bitches [http://xkcd.com/54/].
509 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
510 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
512 * That would thus be a pure waste of time.
515 for (auto const& actor : simix_global->actors_that_ran) {
516 if (actor->simcall_.call_ != simgrid::simix::Simcall::NONE) {
517 actor->simcall_handle(0);
521 simix_global->execute_tasks();
523 simix_global->wake_all_waiting_actors();
524 } while (simix_global->execute_tasks());
526 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
527 if (simix_global->process_list.size() == simix_global->daemons.size())
528 for (auto const& dmon : simix_global->daemons) {
529 XBT_DEBUG("Kill %s", dmon->get_cname());
530 simix_global->maestro_->kill(dmon);
534 time = simgrid::simix::Timer::next();
535 if (time > -1.0 || not simix_global->process_list.empty()) {
536 XBT_DEBUG("Calling surf_solve");
537 time = surf_solve(time);
538 XBT_DEBUG("Moving time ahead : %g", time);
541 /* Notify all the hosts that have failed */
542 /* FIXME: iterate through the list of failed host and mark each of them */
543 /* as failed. On each host, signal all the running processes with host_fail */
545 // Execute timers and tasks until there isn't anything to be done:
548 again = SIMIX_execute_timers();
549 if (simix_global->execute_tasks())
551 simix_global->wake_all_waiting_actors();
554 /* Clean actors to destroy */
555 simix_global->empty_trash();
557 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
558 simix_global->actors_to_run.size());
560 if (time < 0. && simix_global->actors_to_run.empty() && not simix_global->process_list.empty()) {
561 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
562 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
563 "once the simulation is over. Please fix your on_exit() functions.");
565 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
567 simix_global->display_all_actor_status();
568 simgrid::s4u::Engine::on_deadlock();
569 for (auto const& kv : simix_global->process_list) {
570 XBT_DEBUG("Kill %s", kv.second->get_cname());
571 simix_global->maestro_->kill(kv.second);
574 } while (time > -1.0 || not simix_global->actors_to_run.empty());
576 if (not simix_global->process_list.empty())
579 simgrid::s4u::Engine::on_simulation_end();
582 double SIMIX_timer_next() // XBT_ATTRIB_DEPRECATED_v329
584 return simgrid::simix::Timer::next();
587 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void* arg) // XBT_ATTRIB_DEPRECATED_v329
589 return simgrid::simix::Timer::set(date, std::bind(callback, arg));
592 /** @brief cancels a timer that was added earlier */
593 void SIMIX_timer_remove(smx_timer_t timer) // XBT_ATTRIB_DEPRECATED_v329
598 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
599 double SIMIX_timer_get_date(smx_timer_t timer) // XBT_ATTRIB_DEPRECATED_v329
601 return timer ? timer->date : 0.0;
604 void SIMIX_display_process_status() // XBT_ATTRIB_DEPRECATED_v329
606 simix_global->display_all_actor_status();
609 int SIMIX_is_maestro()
611 if (simix_global == nullptr) // SimDag
613 const simgrid::kernel::actor::ActorImpl* self = SIMIX_process_self();
614 return self == nullptr || self == simix_global->maestro_;