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 } // namespace simgrid
44 XBT_ATTRIB_NORETURN static void inthandler(int)
46 if (simgrid::simix::cfg_verbose_exit) {
47 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option "
48 "'debug/verbose-exit').");
49 simix_global->display_all_actor_status();
51 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'debug/verbose-exit' is set to false.");
57 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
59 if ((siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) || siginfo->si_signo == SIGBUS) {
61 "Access violation or Bus error detected.\n"
62 "This probably comes from a programming error in your code, or from a stack\n"
63 "overflow. If you are certain of your code, try increasing the stack size\n"
64 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
66 "If it does not help, this may have one of the following causes:\n"
67 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
68 "Failing hardware can sometimes generate such errors too.\n"
70 "If you think you've found a bug in SimGrid, please report it along with a\n"
71 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
72 "of the fault captured with gdb or valgrind.\n",
73 smx_context_stack_size / 1024);
74 } else if (siginfo->si_signo == SIGSEGV) {
75 fprintf(stderr, "Segmentation fault.\n");
77 if (smpi_enabled() && smpi_cfg_privatization() == SmpiPrivStrategies::NONE) {
78 #if HAVE_PRIVATIZATION
79 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
81 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
82 #endif /* HAVE_PRIVATIZATION */
84 #endif /* HAVE_SMPI */
90 * Install signal handler for SIGSEGV. Check that nobody has already installed
91 * its own handler. For example, the Java VM does this.
93 static void install_segvhandler()
97 if (simgrid::kernel::context::Context::install_sigsegv_stack(&old_stack, true) == -1) {
98 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
101 if (not(old_stack.ss_flags & SS_DISABLE)) {
102 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
103 old_stack.ss_size, (unsigned)old_stack.ss_flags);
104 sigaltstack(&old_stack, nullptr);
107 struct sigaction action;
108 struct sigaction old_action;
109 action.sa_sigaction = &segvhandler;
110 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
111 sigemptyset(&action.sa_mask);
113 /* Linux tend to raise only SIGSEGV where other systems also raise SIGBUS on severe error */
114 for (int sig : {SIGSEGV, SIGBUS}) {
115 if (sigaction(sig, &action, &old_action) == -1) {
116 XBT_WARN("Failed to register signal handler for signal %d: %s", sig, strerror(errno));
119 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
120 XBT_DEBUG("A signal handler was already installed for signal %d (%p). Restore it.", sig,
121 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
122 sigaction(sig, &old_action, nullptr);
129 /********************************* SIMIX **************************************/
133 Timer* Timer::set(double date, xbt::Task<void()>&& callback)
135 auto* timer = new Timer(date, std::move(callback));
136 timer->handle_ = simix_timers().emplace(std::make_pair(date, timer));
140 /** @brief cancels a timer that was added earlier */
143 simix_timers().erase(handle_);
147 /** Execute all the tasks that are queued, e.g. `.then()` callbacks of futures. */
148 bool Global::execute_tasks()
150 xbt_assert(tasksTemp.empty());
156 // We don't want the callbacks to modify the vector we are iterating over:
157 tasks.swap(tasksTemp);
159 // Execute all the queued tasks:
160 for (auto& task : tasksTemp)
164 } while (not tasks.empty());
169 void Global::empty_trash()
171 while (not actors_to_destroy.empty()) {
172 kernel::actor::ActorImpl* actor = &actors_to_destroy.front();
173 actors_to_destroy.pop_front();
174 XBT_DEBUG("Getting rid of %s (refcount: %d)", actor->get_cname(), actor->get_refcount());
175 intrusive_ptr_release(actor);
178 xbt_dynar_reset(dead_actors_vector);
182 * @brief Executes the actors in actors_to_run.
184 * The actors in actors_to_run are run (in parallel if possible). On exit, actors_to_run is empty, and actors_that_ran
185 * contains the list of actors that just ran. The two lists are swapped so, be careful when using them before and after
186 * a call to this function.
188 void Global::run_all_actors()
190 simix_global->context_factory->run_all();
192 actors_to_run.swap(actors_that_ran);
193 actors_to_run.clear();
196 /** Wake up all actors waiting for a Surf action to finish */
197 void Global::wake_all_waiting_actors() const
199 for (auto const& model : simgrid::kernel::EngineImpl::get_instance()->get_all_models()) {
200 kernel::resource::Action* action;
202 XBT_DEBUG("Handling the failed actions (if any)");
203 while ((action = model->extract_failed_action())) {
204 XBT_DEBUG(" Handling Action %p", action);
205 if (action->get_activity() != nullptr)
206 kernel::activity::ActivityImplPtr(action->get_activity())->post();
208 XBT_DEBUG("Handling the terminated actions (if any)");
209 while ((action = model->extract_done_action())) {
210 XBT_DEBUG(" Handling Action %p", action);
211 if (action->get_activity() == nullptr)
212 XBT_DEBUG("probably vcpu's action %p, skip", action);
214 kernel::activity::ActivityImplPtr(action->get_activity())->post();
219 void Global::display_all_actor_status() const
221 XBT_INFO("%zu actors are still running, waiting for something.", process_list.size());
222 /* List the actors and their state */
223 XBT_INFO("Legend of the following listing: \"Actor <pid> (<name>@<host>): <status>\"");
224 for (auto const& kv : process_list) {
225 kernel::actor::ActorImpl* actor = kv.second;
227 if (actor->waiting_synchro_) {
228 const char* synchro_description = "unknown";
230 if (boost::dynamic_pointer_cast<kernel::activity::ExecImpl>(actor->waiting_synchro_) != nullptr)
231 synchro_description = "execution";
233 if (boost::dynamic_pointer_cast<kernel::activity::CommImpl>(actor->waiting_synchro_) != nullptr)
234 synchro_description = "communication";
236 if (boost::dynamic_pointer_cast<kernel::activity::SleepImpl>(actor->waiting_synchro_) != nullptr)
237 synchro_description = "sleeping";
239 if (boost::dynamic_pointer_cast<kernel::activity::RawImpl>(actor->waiting_synchro_) != nullptr)
240 synchro_description = "synchronization";
242 if (boost::dynamic_pointer_cast<kernel::activity::IoImpl>(actor->waiting_synchro_) != nullptr)
243 synchro_description = "I/O";
245 XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %d to finish", actor->get_pid(),
246 actor->get_cname(), actor->get_host()->get_cname(), synchro_description,
247 (xbt_log_no_loc ? (size_t)0xDEADBEEF : (size_t)actor->waiting_synchro_.get()),
248 actor->waiting_synchro_->get_cname(), (int)actor->waiting_synchro_->state_);
250 XBT_INFO("Actor %ld (%s@%s)", actor->get_pid(), actor->get_cname(), actor->get_host()->get_cname());
255 config::Flag<double> cfg_breakpoint{"debug/breakpoint",
256 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
258 } // namespace simgrid
260 static simgrid::kernel::actor::ActorCode maestro_code;
261 void SIMIX_set_maestro(void (*code)(void*), void* data)
264 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if "
265 "you need that feature");
267 maestro_code = std::bind(code, data);
272 * @brief Initialize SIMIX internal data.
274 void SIMIX_global_init(int* argc, char** argv)
277 // The communication initialization is done ASAP.
278 // We need to communicate initialization of the different layers to the model-checker.
279 simgrid::mc::AppSide::initialize();
282 if (simix_global == nullptr) {
283 surf_init(argc, argv); /* Initialize SURF structures */
285 simix_global = std::make_unique<simgrid::simix::Global>();
286 simix_global->maestro_ = nullptr;
287 SIMIX_context_mod_init();
289 // Either create a new context with maestro or create
290 // a context object with the current context maestro):
291 simgrid::kernel::actor::create_maestro(maestro_code);
293 /* Prepare to display some more info when dying on Ctrl-C pressing */
294 std::signal(SIGINT, inthandler);
297 install_segvhandler();
299 /* register a function to be called by SURF after the environment creation */
301 simgrid::s4u::Engine::on_platform_created.connect(surf_presolve);
304 if (simgrid::config::get_value<bool>("debug/clean-atexit"))
310 * @brief Clean the SIMIX simulation
312 * This functions remove the memory used by SIMIX
316 static bool smx_cleaned = false;
318 return; // to avoid double cleaning by java and C
321 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
322 if (not simix_global->actors_to_run.empty() && SIMIX_get_clock() <= 0.0) {
324 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
325 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
326 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
330 if (not simix_global->process_list.empty()) {
331 if (smpi_process()->initialized()) {
332 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
334 XBT_WARN("Process called exit when leaving - Skipping cleanups");
340 /* Kill all processes (but maestro) */
341 simix_global->maestro_->kill_all();
342 simix_global->run_all_actors();
343 simix_global->empty_trash();
345 /* Exit the SIMIX network module */
346 SIMIX_mailbox_exit();
348 while (not simgrid::simix::simix_timers().empty()) {
349 delete simgrid::simix::simix_timers().top().second;
350 simgrid::simix::simix_timers().pop();
352 /* Free the remaining data structures */
353 simix_global->actors_to_run.clear();
354 simix_global->actors_that_ran.clear();
355 simix_global->actors_to_destroy.clear();
356 simix_global->process_list.clear();
359 xbt_dynar_free(&simix_global->actors_vector);
360 xbt_dynar_free(&simix_global->dead_actors_vector);
363 /* Let's free maestro now */
364 delete simix_global->maestro_;
365 simix_global->maestro_ = nullptr;
367 /* Finish context module and SURF */
368 SIMIX_context_mod_exit();
372 simix_global = nullptr;
377 * @brief A clock (in second).
379 * @return Return the clock.
381 double SIMIX_get_clock()
383 if (MC_is_active() || MC_record_replay_is_active()) {
384 return MC_process_clock_get(SIMIX_process_self());
386 return surf_get_clock();
390 /** Handle any pending timer. Returns if something was actually run. */
391 static bool SIMIX_execute_timers()
394 while (not simgrid::simix::simix_timers().empty() &&
395 SIMIX_get_clock() >= simgrid::simix::simix_timers().top().first) {
397 // FIXME: make the timers being real callbacks (i.e. provide dispatchers that read and expand the args)
398 smx_timer_t timer = simgrid::simix::simix_timers().top().second;
399 simgrid::simix::simix_timers().pop();
408 * @brief Run the main simulation loop.
412 if (MC_record_replay_is_active()) {
413 simgrid::mc::replay(MC_record_path());
420 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
422 if (simgrid::simix::cfg_breakpoint >= 0.0 && surf_get_clock() >= simgrid::simix::cfg_breakpoint) {
423 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::cfg_breakpoint.get());
424 simgrid::simix::cfg_breakpoint = -1.0;
432 simix_global->execute_tasks();
434 while (not simix_global->actors_to_run.empty()) {
435 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
437 /* Run all processes that are ready to run, possibly in parallel */
438 simix_global->run_all_actors();
440 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
442 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
444 /* Here, the order is ok because:
446 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
447 * not impact its order.
449 * Long proof: actors remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
451 * - if there is no kill during the simulation, actors remain sorted according by their PID.
452 * Rationale: This can be proved inductively.
453 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
454 * is parsed linearly).
455 * Let's show that it is still so at the end of this round.
456 * - if an actor is added when being created, that's from maestro. It can be either at startup
457 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
458 * in arbitrary order (inductive hypothesis), we are fine.
459 * - If an actor is added because it's getting killed, its subsequent actions shouldn't matter
460 * - If an actor gets added to actors_to_run because one of their blocking action constituting the meat
461 * of a simcall terminates, we're still good. Proof:
462 * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the
463 * resource kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications
465 * - For communications, this function is called from SIMIX_comm_finish().
466 * This function itself don't mess with the order since simcalls are handled in FIFO order.
467 * The function is called:
468 * - before the comm starts (invalid parameters, or resource already dead or whatever).
469 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
470 * - because the communication failed or were canceled after startup. In this case, it's called from
471 * the function we are in, by the chunk:
472 * set = model->states.failed_action_set;
473 * while ((synchro = extract(set)))
474 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
475 * This order is also fixed because it depends of the order in which the surf actions were
476 * added to the system, and only maestro can add stuff this way, through simcalls.
477 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
478 * popped out of the set does not depend on the user code's execution order.
479 * - because the communication terminated. In this case, synchros are served in the order given by
480 * set = model->states.done_action_set;
481 * while ((synchro = extract(set)))
482 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
483 * and the argument is very similar to the previous one.
484 * So, in any case, the orders of calls to CommImpl::finish() do not depend on the order in which user
485 * actors are executed.
486 * So, in any cases, the orders of actors within actors_to_run do not depend on the order in which
487 * user actors were executed previously.
488 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
489 * - If there is some actor killings, the order is changed by this decision that comes from user-land
490 * But this decision may not have been motivated by a situation that were different because the simulation is
492 * So, even the order change induced by the actor killing is perfectly reproducible.
494 * So science works, bitches [http://xkcd.com/54/].
496 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
497 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
499 * That would thus be a pure waste of time.
502 for (auto const& actor : simix_global->actors_that_ran) {
503 if (actor->simcall_.call_ != simgrid::simix::Simcall::NONE) {
504 actor->simcall_handle(0);
508 simix_global->execute_tasks();
510 simix_global->wake_all_waiting_actors();
511 } while (simix_global->execute_tasks());
513 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
514 if (simix_global->process_list.size() == simix_global->daemons.size())
515 for (auto const& dmon : simix_global->daemons) {
516 XBT_DEBUG("Kill %s", dmon->get_cname());
517 simix_global->maestro_->kill(dmon);
521 time = simgrid::simix::Timer::next();
522 if (time > -1.0 || not simix_global->process_list.empty()) {
523 XBT_DEBUG("Calling surf_solve");
524 time = surf_solve(time);
525 XBT_DEBUG("Moving time ahead : %g", time);
528 /* Notify all the hosts that have failed */
529 /* FIXME: iterate through the list of failed host and mark each of them */
530 /* as failed. On each host, signal all the running processes with host_fail */
532 // Execute timers and tasks until there isn't anything to be done:
535 again = SIMIX_execute_timers();
536 if (simix_global->execute_tasks())
538 simix_global->wake_all_waiting_actors();
541 /* Clean actors to destroy */
542 simix_global->empty_trash();
544 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
545 simix_global->actors_to_run.size());
547 if (time < 0. && simix_global->actors_to_run.empty() && not simix_global->process_list.empty()) {
548 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
549 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
550 "once the simulation is over. Please fix your on_exit() functions.");
552 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
554 simix_global->display_all_actor_status();
555 simgrid::s4u::Engine::on_deadlock();
556 for (auto const& kv : simix_global->process_list) {
557 XBT_DEBUG("Kill %s", kv.second->get_cname());
558 simix_global->maestro_->kill(kv.second);
561 } while (time > -1.0 || not simix_global->actors_to_run.empty());
563 if (not simix_global->process_list.empty())
566 simgrid::s4u::Engine::on_simulation_end();
569 double SIMIX_timer_next() // XBT_ATTRIB_DEPRECATED_v329
571 return simgrid::simix::Timer::next();
574 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void* arg) // XBT_ATTRIB_DEPRECATED_v329
576 return simgrid::simix::Timer::set(date, std::bind(callback, arg));
579 /** @brief cancels a timer that was added earlier */
580 void SIMIX_timer_remove(smx_timer_t timer) // XBT_ATTRIB_DEPRECATED_v329
585 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
586 double SIMIX_timer_get_date(smx_timer_t timer) // XBT_ATTRIB_DEPRECATED_v329
588 return timer ? timer->get_date() : 0;
591 void SIMIX_display_process_status() // XBT_ATTRIB_DEPRECATED_v329
593 simix_global->display_all_actor_status();
596 int SIMIX_is_maestro()
598 if (simix_global == nullptr) // SimDag
600 const simgrid::kernel::actor::ActorImpl* self = SIMIX_process_self();
601 return self == nullptr || self == simix_global->maestro_;