1 /* Copyright (c) 2007-2020. 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_private.hpp"
20 #include "src/surf/StorageImpl.hpp"
21 #include "src/surf/xml/platf.hpp"
24 #include "src/mc/remote/AppSide.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;
38 config::Flag<bool> cfg_verbose_exit{"debug/verbose-exit",
39 "Display the actor status at exit",
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();
52 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'debug/verbose-exit' is set to false.");
58 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
60 if ((siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) || siginfo->si_signo == SIGBUS) {
62 "Access violation or Bus error detected.\n"
63 "This probably comes from a programming error in your code, or from a stack\n"
64 "overflow. If you are certain of your code, try increasing the stack size\n"
65 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
67 "If it does not help, this may have one of the following causes:\n"
68 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
69 "Failing hardware can sometimes generate such errors too.\n"
71 "If you think you've found a bug in SimGrid, please report it along with a\n"
72 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
73 "of the fault captured with gdb or valgrind.\n",
74 smx_context_stack_size / 1024);
75 } else if (siginfo->si_signo == SIGSEGV) {
76 fprintf(stderr, "Segmentation fault.\n");
78 if (smpi_enabled() && smpi_cfg_privatization() == SmpiPrivStrategies::NONE) {
79 #if HAVE_PRIVATIZATION
80 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
82 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
83 #endif /* HAVE_PRIVATIZATION */
85 #endif /* HAVE_SMPI */
90 std::array<unsigned char, SIGSTKSZ> sigsegv_stack; /* alternate stack for SIGSEGV handler */
93 * Install signal handler for SIGSEGV. Check that nobody has already installed
94 * its own handler. For example, the Java VM does this.
96 static void install_segvhandler()
100 if (simgrid::kernel::context::Context::install_sigsegv_stack(&old_stack, true) == -1) {
101 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
104 if (not(old_stack.ss_flags & SS_DISABLE)) {
105 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
106 old_stack.ss_size, (unsigned)old_stack.ss_flags);
107 sigaltstack(&old_stack, nullptr);
110 struct sigaction action;
111 struct sigaction old_action;
112 action.sa_sigaction = &segvhandler;
113 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
114 sigemptyset(&action.sa_mask);
116 /* Linux tend to raise only SIGSEGV where other systems also raise SIGBUS on severe error */
117 for (int sig : {SIGSEGV, SIGBUS}) {
118 if (sigaction(sig, &action, &old_action) == -1) {
119 XBT_WARN("Failed to register signal handler for signal %d: %s", sig, strerror(errno));
122 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
123 XBT_DEBUG("A signal handler was already installed for signal %d (%p). Restore it.", sig,
124 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
125 sigaction(sig, &old_action, nullptr);
132 /********************************* SIMIX **************************************/
136 Timer* Timer::set(double date, xbt::Task<void()>&& callback)
138 auto* timer = new Timer(date, std::move(callback));
139 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
143 /** @brief cancels a timer that was added earlier */
146 simix_timers.erase(handle_);
150 /** Execute all the tasks that are queued, e.g. `.then()` callbacks of futures. */
151 bool Global::execute_tasks()
153 xbt_assert(tasksTemp.empty());
159 // We don't want the callbacks to modify the vector we are iterating over:
160 tasks.swap(tasksTemp);
162 // Execute all the queued tasks:
163 for (auto& task : tasksTemp)
167 } while (not tasks.empty());
172 void Global::empty_trash()
174 while (not actors_to_destroy.empty()) {
175 kernel::actor::ActorImpl* actor = &actors_to_destroy.front();
176 actors_to_destroy.pop_front();
177 XBT_DEBUG("Getting rid of %s (refcount: %d)", actor->get_cname(), actor->get_refcount());
178 intrusive_ptr_release(actor);
181 xbt_dynar_reset(dead_actors_vector);
185 * @brief Executes the actors in actors_to_run.
187 * The actors in actors_to_run are run (in parallel if possible). On exit, actors_to_run is empty, and actors_that_ran
188 * contains the list of actors that just ran. The two lists are swapped so, be careful when using them before and after
189 * a call to this function.
191 void Global::run_all_actors()
193 simix_global->context_factory->run_all();
195 actors_to_run.swap(actors_that_ran);
196 actors_to_run.clear();
199 /** Wake up all actors waiting for a Surf action to finish */
200 void Global::wake_all_waiting_actors() const
202 for (auto const& model : all_existing_models) {
203 kernel::resource::Action* action;
205 XBT_DEBUG("Handling the failed actions (if any)");
206 while ((action = model->extract_failed_action())) {
207 XBT_DEBUG(" Handling Action %p", action);
208 if (action->get_activity() != nullptr)
209 kernel::activity::ActivityImplPtr(action->get_activity())->post();
211 XBT_DEBUG("Handling the terminated actions (if any)");
212 while ((action = model->extract_done_action())) {
213 XBT_DEBUG(" Handling Action %p", action);
214 if (action->get_activity() == nullptr)
215 XBT_DEBUG("probably vcpu's action %p, skip", action);
217 kernel::activity::ActivityImplPtr(action->get_activity())->post();
222 void Global::display_all_actor_status() const
224 XBT_INFO("%zu actors are still running, waiting for something.", process_list.size());
225 /* List the actors and their state */
226 XBT_INFO("Legend of the following listing: \"Actor <pid> (<name>@<host>): <status>\"");
227 for (auto const& kv : process_list) {
228 kernel::actor::ActorImpl* actor = kv.second;
230 if (actor->waiting_synchro_) {
231 const char* synchro_description = "unknown";
233 if (boost::dynamic_pointer_cast<kernel::activity::ExecImpl>(actor->waiting_synchro_) != nullptr)
234 synchro_description = "execution";
236 if (boost::dynamic_pointer_cast<kernel::activity::CommImpl>(actor->waiting_synchro_) != nullptr)
237 synchro_description = "communication";
239 if (boost::dynamic_pointer_cast<kernel::activity::SleepImpl>(actor->waiting_synchro_) != nullptr)
240 synchro_description = "sleeping";
242 if (boost::dynamic_pointer_cast<kernel::activity::RawImpl>(actor->waiting_synchro_) != nullptr)
243 synchro_description = "synchronization";
245 if (boost::dynamic_pointer_cast<kernel::activity::IoImpl>(actor->waiting_synchro_) != nullptr)
246 synchro_description = "I/O";
248 XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %d to finish", actor->get_pid(),
249 actor->get_cname(), actor->get_host()->get_cname(), synchro_description,
250 (xbt_log_no_loc ? (size_t)0xDEADBEEF : (size_t)actor->waiting_synchro_.get()),
251 actor->waiting_synchro_->get_cname(), (int)actor->waiting_synchro_->state_);
253 XBT_INFO("Actor %ld (%s@%s)", actor->get_pid(), actor->get_cname(), actor->get_host()->get_cname());
258 config::Flag<double> cfg_breakpoint{"debug/breakpoint",
259 "When non-negative, raise a SIGTRAP after given (simulated) time",
262 } // namespace simgrid
264 static simgrid::kernel::actor::ActorCode maestro_code;
265 void SIMIX_set_maestro(void (*code)(void*), void* data)
268 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
270 maestro_code = std::bind(code, data);
275 * @brief Initialize SIMIX internal data.
277 void SIMIX_global_init(int *argc, char **argv)
280 // The communication initialization is done ASAP.
281 // We need to communicate initialization of the different layers to the model-checker.
282 simgrid::mc::AppSide::initialize();
285 if (simix_global == nullptr) {
286 surf_init(argc, argv); /* Initialize SURF structures */
288 simix_global = std::make_unique<simgrid::simix::Global>();
289 simix_global->maestro_ = nullptr;
290 SIMIX_context_mod_init();
292 // Either create a new context with maestro or create
293 // a context object with the current context maestro):
294 simgrid::kernel::actor::create_maestro(maestro_code);
296 /* Prepare to display some more info when dying on Ctrl-C pressing */
297 std::signal(SIGINT, inthandler);
300 install_segvhandler();
302 /* register a function to be called by SURF after the environment creation */
304 simgrid::s4u::Engine::on_platform_created.connect(surf_presolve);
307 if (simgrid::config::get_value<bool>("debug/clean-atexit"))
314 * @brief Clean the SIMIX simulation
316 * This functions remove the memory used by SIMIX
321 return; // to avoid double cleaning by java and C
324 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
325 if (not simix_global->actors_to_run.empty() && SIMIX_get_clock() <= 0.0) {
327 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
328 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
329 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
333 if (not simix_global->process_list.empty()) {
334 if(smpi_process()->initialized()){
335 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
337 XBT_WARN("Process called exit when leaving - Skipping cleanups");
343 /* Kill all processes (but maestro) */
344 simix_global->maestro_->kill_all();
345 simix_global->run_all_actors();
346 simix_global->empty_trash();
348 /* Exit the SIMIX network module */
349 SIMIX_mailbox_exit();
351 while (not simgrid::simix::simix_timers.empty()) {
352 delete simgrid::simix::simix_timers.top().second;
353 simgrid::simix::simix_timers.pop();
355 /* Free the remaining data structures */
356 simix_global->actors_to_run.clear();
357 simix_global->actors_that_ran.clear();
358 simix_global->actors_to_destroy.clear();
359 simix_global->process_list.clear();
362 xbt_dynar_free(&simix_global->actors_vector);
363 xbt_dynar_free(&simix_global->dead_actors_vector);
366 /* Let's free maestro now */
367 delete simix_global->maestro_;
368 simix_global->maestro_ = nullptr;
370 /* Finish context module and SURF */
371 SIMIX_context_mod_exit();
375 simix_global = nullptr;
380 * @brief A clock (in second).
382 * @return Return the clock.
384 double SIMIX_get_clock()
386 if(MC_is_active() || MC_record_replay_is_active()){
387 return MC_process_clock_get(SIMIX_process_self());
389 return surf_get_clock();
393 /** Handle any pending timer. Returns if something was actually run. */
394 static bool SIMIX_execute_timers()
397 while (not simgrid::simix::simix_timers.empty() && SIMIX_get_clock() >= simgrid::simix::simix_timers.top().first) {
399 // FIXME: make the timers being real callbacks (i.e. provide dispatchers that read and expand the args)
400 smx_timer_t timer = simgrid::simix::simix_timers.top().second;
401 simgrid::simix::simix_timers.pop();
410 * @brief Run the main simulation loop.
414 if (MC_record_replay_is_active()) {
415 simgrid::mc::replay(MC_record_path);
422 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
424 if (simgrid::simix::cfg_breakpoint >= 0.0 && surf_get_clock() >= simgrid::simix::cfg_breakpoint) {
425 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::cfg_breakpoint.get());
426 simgrid::simix::cfg_breakpoint = -1.0;
434 simix_global->execute_tasks();
436 while (not simix_global->actors_to_run.empty()) {
437 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
439 /* Run all processes that are ready to run, possibly in parallel */
440 simix_global->run_all_actors();
442 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
444 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
446 /* Here, the order is ok because:
448 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
449 * not impact its order.
451 * Long proof: actors remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
453 * - if there is no kill during the simulation, actors remain sorted according by their PID.
454 * Rationale: This can be proved inductively.
455 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
456 * is parsed linearly).
457 * Let's show that it is still so at the end of this round.
458 * - if an actor is added when being created, that's from maestro. It can be either at startup
459 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
460 * in arbitrary order (inductive hypothesis), we are fine.
461 * - If an actor is added because it's getting killed, its subsequent actions shouldn't matter
462 * - If an actor gets added to actors_to_run because one of their blocking action constituting the meat
463 * of a simcall terminates, we're still good. Proof:
464 * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the
465 * resource kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications
467 * - For communications, this function is called from SIMIX_comm_finish().
468 * This function itself don't mess with the order since simcalls are handled in FIFO order.
469 * The function is called:
470 * - before the comm starts (invalid parameters, or resource already dead or whatever).
471 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
472 * - because the communication failed or were canceled after startup. In this case, it's called from
473 * the function we are in, by the chunk:
474 * set = model->states.failed_action_set;
475 * while ((synchro = extract(set)))
476 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
477 * This order is also fixed because it depends of the order in which the surf actions were
478 * added to the system, and only maestro can add stuff this way, through simcalls.
479 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
480 * popped out of the set does not depend on the user code's execution order.
481 * - because the communication terminated. In this case, synchros are served in the order given by
482 * set = model->states.done_action_set;
483 * while ((synchro = extract(set)))
484 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
485 * and the argument is very similar to the previous one.
486 * So, in any case, the orders of calls to CommImpl::finish() do not depend on the order in which user
487 * actors are executed.
488 * So, in any cases, the orders of actors within actors_to_run do not depend on the order in which
489 * user actors were executed previously.
490 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
491 * - If there is some actor killings, the order is changed by this decision that comes from user-land
492 * But this decision may not have been motivated by a situation that were different because the simulation is
494 * So, even the order change induced by the actor killing is perfectly reproducible.
496 * So science works, bitches [http://xkcd.com/54/].
498 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
499 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
501 * That would thus be a pure waste of time.
504 for (auto const& actor : simix_global->actors_that_ran) {
505 if (actor->simcall_.call_ != simgrid::simix::Simcall::NONE) {
506 actor->simcall_handle(0);
510 simix_global->execute_tasks();
512 simix_global->wake_all_waiting_actors();
513 } while (simix_global->execute_tasks());
515 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
516 if (simix_global->process_list.size() == simix_global->daemons.size())
517 for (auto const& dmon : simix_global->daemons) {
518 XBT_DEBUG("Kill %s", dmon->get_cname());
519 simix_global->maestro_->kill(dmon);
523 time = simgrid::simix::Timer::next();
524 if (time > -1.0 || not simix_global->process_list.empty()) {
525 XBT_DEBUG("Calling surf_solve");
526 time = surf_solve(time);
527 XBT_DEBUG("Moving time ahead : %g", time);
530 /* Notify all the hosts that have failed */
531 /* FIXME: iterate through the list of failed host and mark each of them */
532 /* as failed. On each host, signal all the running processes with host_fail */
534 // Execute timers and tasks until there isn't anything to be done:
537 again = SIMIX_execute_timers();
538 if (simix_global->execute_tasks())
540 simix_global->wake_all_waiting_actors();
543 /* Clean actors to destroy */
544 simix_global->empty_trash();
546 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
547 simix_global->actors_to_run.size());
549 if (time < 0. && simix_global->actors_to_run.empty() && not simix_global->process_list.empty()) {
550 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
551 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
552 "once the simulation is over. Please fix your on_exit() functions.");
554 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
556 simix_global->display_all_actor_status();
557 simgrid::s4u::Engine::on_deadlock();
558 for (auto const& kv : simix_global->process_list) {
559 XBT_DEBUG("Kill %s", kv.second->get_cname());
560 simix_global->maestro_->kill(kv.second);
563 } while (time > -1.0 || not simix_global->actors_to_run.empty());
565 if (not simix_global->process_list.empty())
568 simgrid::s4u::Engine::on_simulation_end();
571 double SIMIX_timer_next() // XBT_ATTRIB_DEPRECATED_v329
573 return simgrid::simix::Timer::next();
576 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void* arg) // XBT_ATTRIB_DEPRECATED_v329
578 return simgrid::simix::Timer::set(date, std::bind(callback, arg));
581 /** @brief cancels a timer that was added earlier */
582 void SIMIX_timer_remove(smx_timer_t timer) // XBT_ATTRIB_DEPRECATED_v329
587 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
588 double SIMIX_timer_get_date(smx_timer_t timer) // XBT_ATTRIB_DEPRECATED_v329
590 return timer ? timer->get_date() : 0;
593 void SIMIX_display_process_status() // XBT_ATTRIB_DEPRECATED_v329
595 simix_global->display_all_actor_status();
598 int SIMIX_is_maestro()
600 if (simix_global == nullptr) // SimDag
602 const simgrid::kernel::actor::ActorImpl* self = SIMIX_process_self();
603 return self == nullptr || self == simix_global->maestro_;