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_private.hpp"
20 #include "src/surf/StorageImpl.hpp"
21 #include "src/surf/xml/platf.hpp"
24 #include "src/mc/remote/Client.hpp"
28 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
29 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
31 std::unique_ptr<simgrid::simix::Global> simix_global;
33 void (*SMPI_switch_data_segment)(simgrid::s4u::ActorPtr) = nullptr;
37 simgrid::config::Flag<double> cfg_verbose_exit{
38 "debug/verbose-exit", {"verbose-exit"}, "Display the actor status at exit", true};
40 } // namespace simgrid
41 XBT_ATTRIB_NORETURN static void inthandler(int)
43 if (simgrid::simix::cfg_verbose_exit) {
44 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option "
45 "'debug/verbose-exit').");
46 SIMIX_display_process_status();
49 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'debug/verbose-exit' is set to false.");
55 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
57 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
58 fprintf(stderr, "Access violation detected.\n"
59 "This probably comes from a programming error in your code, or from a stack\n"
60 "overflow. If you are certain of your code, try increasing the stack size\n"
61 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
63 "If it does not help, this may have one of the following causes:\n"
64 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
65 "Failing hardware can sometimes generate such errors too.\n"
67 "If you think you've found a bug in SimGrid, please report it along with a\n"
68 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
69 "of the fault captured with gdb or valgrind.\n",
70 smx_context_stack_size / 1024);
71 } else if (siginfo->si_signo == SIGSEGV) {
72 fprintf(stderr, "Segmentation fault.\n");
74 if (smpi_enabled() && smpi_privatize_global_variables == SmpiPrivStrategies::NONE) {
75 #if HAVE_PRIVATIZATION
76 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
78 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
79 #endif /* HAVE_PRIVATIZATION */
81 #endif /* HAVE_SMPI */
86 unsigned char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
89 * Install signal handler for SIGSEGV. Check that nobody has already installed
90 * its own handler. For example, the Java VM does this.
92 static void install_segvhandler()
96 stack.ss_sp = sigsegv_stack;
97 stack.ss_size = sizeof sigsegv_stack;
100 if (sigaltstack(&stack, &old_stack) == -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 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
117 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
120 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
121 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
122 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
123 sigaction(SIGSEGV, &old_action, nullptr);
129 /********************************* SIMIX **************************************/
133 Timer* Timer::set(double date, simgrid::xbt::Task<void()>&& callback)
135 Timer* 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 simgrid::simix::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 smx_actor_t 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 simgrid::config::Flag<double> cfg_breakpoint{
197 "debug/breakpoint", {"simix/breakpoint"}, "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
201 static simgrid::simix::ActorCode maestro_code;
202 void SIMIX_set_maestro(void (*code)(void*), void* data)
205 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
207 maestro_code = std::bind(code, data);
212 * @brief Initialize SIMIX internal data.
214 void SIMIX_global_init(int *argc, char **argv)
217 // The communication initialization is done ASAP.
218 // We need to communicate initialization of the different layers to the model-checker.
219 simgrid::mc::Client::initialize();
222 if (simix_global == nullptr) {
223 surf_init(argc, argv); /* Initialize SURF structures */
225 simix_global.reset(new simgrid::simix::Global());
226 simix_global->maestro_process = nullptr;
227 SIMIX_context_mod_init();
229 // Either create a new context with maestro or create
230 // a context object with the current context mestro):
231 simgrid::kernel::actor::create_maestro(maestro_code);
233 /* Prepare to display some more info when dying on Ctrl-C pressing */
234 std::signal(SIGINT, inthandler);
237 install_segvhandler();
239 /* register a function to be called by SURF after the environment creation */
241 simgrid::s4u::on_platform_created.connect(surf_presolve);
243 simgrid::s4u::Storage::on_creation.connect([](simgrid::s4u::Storage const& storage) {
244 sg_storage_t s = simgrid::s4u::Storage::by_name(storage.get_name());
245 xbt_assert(s != nullptr, "Storage not found for name %s", storage.get_cname());
249 if (simgrid::config::get_value<bool>("debug/clean-atexit"))
256 * @brief Clean the SIMIX simulation
258 * This functions remove the memory used by SIMIX
263 return; // to avoid double cleaning by java and C
266 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
267 if (not simix_global->actors_to_run.empty() && SIMIX_get_clock() <= 0.0) {
269 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
270 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
271 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
275 if (SIMIX_process_count()>0){
276 if(smpi_process()->initialized()){
277 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
279 XBT_WARN("Process called exit when leaving - Skipping cleanups");
285 /* Kill all processes (but maestro) */
286 simix_global->maestro_process->kill_all();
287 simix_global->context_factory->run_all();
288 simix_global->empty_trash();
290 /* Exit the SIMIX network module */
291 SIMIX_mailbox_exit();
293 while (not simgrid::simix::simix_timers.empty()) {
294 delete simgrid::simix::simix_timers.top().second;
295 simgrid::simix::simix_timers.pop();
297 /* Free the remaining data structures */
298 simix_global->actors_to_run.clear();
299 simix_global->actors_that_ran.clear();
300 simix_global->actors_to_destroy.clear();
301 simix_global->process_list.clear();
304 xbt_dynar_free(&simix_global->actors_vector);
305 xbt_dynar_free(&simix_global->dead_actors_vector);
308 /* Let's free maestro now */
309 delete simix_global->maestro_process;
310 simix_global->maestro_process = nullptr;
312 /* Finish context module and SURF */
313 SIMIX_context_mod_exit();
317 simix_global = nullptr;
322 * @brief A clock (in second).
324 * @return Return the clock.
326 double SIMIX_get_clock()
328 if(MC_is_active() || MC_record_replay_is_active()){
329 return MC_process_clock_get(SIMIX_process_self());
331 return surf_get_clock();
335 /** Wake up all processes waiting for a Surf action to finish */
336 static void SIMIX_wake_processes()
338 for (auto const& model : all_existing_models) {
339 simgrid::kernel::resource::Action* action;
341 XBT_DEBUG("Handling the failed actions (if any)");
342 while ((action = model->extract_failed_action())) {
343 XBT_DEBUG(" Handling Action %p",action);
344 if (action->get_activity() != nullptr)
345 simgrid::kernel::activity::ActivityImplPtr(action->get_activity())->post();
347 XBT_DEBUG("Handling the terminated actions (if any)");
348 while ((action = model->extract_done_action())) {
349 XBT_DEBUG(" Handling Action %p",action);
350 if (action->get_activity() == nullptr)
351 XBT_DEBUG("probably vcpu's action %p, skip", action);
353 simgrid::kernel::activity::ActivityImplPtr(action->get_activity())->post();
358 /** Handle any pending timer. Returns if something was actually run. */
359 static bool SIMIX_execute_timers()
362 while (not simgrid::simix::simix_timers.empty() && SIMIX_get_clock() >= simgrid::simix::simix_timers.top().first) {
364 // FIXME: make the timers being real callbacks (i.e. provide dispatchers that read and expand the args)
365 smx_timer_t timer = simgrid::simix::simix_timers.top().second;
366 simgrid::simix::simix_timers.pop();
375 * @brief Run the main simulation loop.
379 if (MC_record_replay_is_active()) {
380 simgrid::mc::replay(MC_record_path);
387 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
389 if (simgrid::simix::cfg_breakpoint >= 0.0 && surf_get_clock() >= simgrid::simix::cfg_breakpoint) {
390 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::cfg_breakpoint.get());
391 simgrid::simix::cfg_breakpoint = -1.0;
399 simix_global->execute_tasks();
401 while (not simix_global->actors_to_run.empty()) {
402 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->actors_to_run.size());
404 /* Run all processes that are ready to run, possibly in parallel */
405 simix_global->run_all_actors();
407 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
409 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
411 /* Here, the order is ok because:
413 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
414 * not impact its order.
416 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
418 * - if there is no kill during the simulation, processes remain sorted according by their PID.
419 * Rationale: This can be proved inductively.
420 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
421 * is parsed linearly).
422 * Let's show that it is still so at the end of this round.
423 * - if a process is added when being created, that's from maestro. It can be either at startup
424 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
425 * in arbitrary order (inductive hypothesis), we are fine.
426 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
427 * - If a process gets added to actors_to_run because one of their blocking action constituting the meat
428 * of a simcall terminates, we're still good. Proof:
429 * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the resource
430 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an
432 * - For communications, this function is called from SIMIX_comm_finish().
433 * This function itself don't mess with the order since simcalls are handled in FIFO order.
434 * The function is called:
435 * - before the comm starts (invalid parameters, or resource already dead or whatever).
436 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
437 * - because the communication failed or were canceled after startup. In this case, it's called from
438 * the function we are in, by the chunk:
439 * set = model->states.failed_action_set;
440 * while ((synchro = extract(set)))
441 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
442 * This order is also fixed because it depends of the order in which the surf actions were
443 * added to the system, and only maestro can add stuff this way, through simcalls.
444 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
445 * poped out of the set does not depend on the user code's execution order.
446 * - because the communication terminated. In this case, synchros are served in the order given by
447 * set = model->states.done_action_set;
448 * while ((synchro = extract(set)))
449 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
450 * and the argument is very similar to the previous one.
451 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user
452 * processes are executed.
453 * So, in any cases, the orders of processes within actors_to_run do not depend on the order in which
454 * user processes were executed previously.
455 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
456 * - If there is some process killings, the order is changed by this decision that comes from user-land
457 * But this decision may not have been motivated by a situation that were different because the simulation is
459 * So, even the order change induced by the process killing is perfectly reproducible.
461 * So science works, bitches [http://xkcd.com/54/].
463 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
464 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
466 * That would thus be a pure waste of time.
469 for (smx_actor_t const& process : simix_global->actors_that_ran) {
470 if (process->simcall.call_ != SIMCALL_NONE) {
471 process->simcall_handle(0);
475 simix_global->execute_tasks();
477 SIMIX_wake_processes();
478 } while (simix_global->execute_tasks());
480 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
481 if (simix_global->process_list.size() == simix_global->daemons.size())
482 for (auto const& dmon : simix_global->daemons) {
483 XBT_DEBUG("Kill %s", dmon->get_cname());
484 simix_global->maestro_process->kill(dmon);
488 time = simgrid::simix::Timer::next();
489 if (time > -1.0 || not simix_global->process_list.empty()) {
490 XBT_DEBUG("Calling surf_solve");
491 time = surf_solve(time);
492 XBT_DEBUG("Moving time ahead : %g", time);
495 /* Notify all the hosts that have failed */
496 /* FIXME: iterate through the list of failed host and mark each of them */
497 /* as failed. On each host, signal all the running processes with host_fail */
499 // Execute timers and tasks until there isn't anything to be done:
502 again = SIMIX_execute_timers();
503 if (simix_global->execute_tasks())
505 SIMIX_wake_processes();
508 /* Clean actors to destroy */
509 simix_global->empty_trash();
511 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
512 simix_global->actors_to_run.size());
514 } while (time > -1.0 || not simix_global->actors_to_run.empty());
516 if (not simix_global->process_list.empty()) {
518 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
519 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
520 "once the simulation is over. Please fix your on_exit() functions.");
522 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
524 SIMIX_display_process_status();
525 simgrid::s4u::on_deadlock();
528 simgrid::s4u::on_simulation_end();
531 double SIMIX_timer_next()
533 return simgrid::simix::Timer::next();
536 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
538 return simgrid::simix::Timer::set(date, std::bind(callback, arg));
541 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()>&& callback) // deprecated
543 return simgrid::simix::Timer::set(date, std::move(callback));
546 /** @brief cancels a timer that was added earlier */
547 void SIMIX_timer_remove(smx_timer_t timer) {
551 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
552 double SIMIX_timer_get_date(smx_timer_t timer) {
553 return timer ? timer->get_date() : 0;
556 void SIMIX_display_process_status()
558 int nbprocess = simix_global->process_list.size();
560 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
561 /* List the process and their state */
562 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
563 for (auto const& kv : simix_global->process_list) {
564 smx_actor_t actor = kv.second;
566 if (actor->waiting_synchro) {
568 const char* synchro_description = "unknown";
569 // we don't care about the Activity type to get its name, use RawImpl
571 boost::static_pointer_cast<simgrid::kernel::activity::ActivityImpl_T<simgrid::kernel::activity::RawImpl>>(
572 actor->waiting_synchro)
575 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(actor->waiting_synchro) != nullptr)
576 synchro_description = "execution";
578 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(actor->waiting_synchro) != nullptr)
579 synchro_description = "communication";
581 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(actor->waiting_synchro) != nullptr)
582 synchro_description = "sleeping";
584 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(actor->waiting_synchro) != nullptr)
585 synchro_description = "synchronization";
587 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(actor->waiting_synchro) != nullptr)
588 synchro_description = "I/O";
590 XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %p (%s) in state %d to finish", actor->get_pid(),
591 actor->get_cname(), actor->get_host()->get_cname(), synchro_description, actor->waiting_synchro.get(),
592 name, (int)actor->waiting_synchro->state_);
595 XBT_INFO("Actor %ld (%s@%s)", actor->get_pid(), actor->get_cname(), actor->get_host()->get_cname());
600 int SIMIX_is_maestro()
602 smx_actor_t self = SIMIX_process_self();
603 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;