1 /* Copyright (c) 2007-2018. 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"
28 #include <boost/heap/fibonacci_heap.hpp>
30 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
31 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
33 std::unique_ptr<simgrid::simix::Global> simix_global;
36 typedef std::pair<double, smx_timer_t> TimerQelt;
37 boost::heap::fibonacci_heap<TimerQelt, boost::heap::compare<simgrid::xbt::HeapComparator<TimerQelt>>> simix_timers;
40 /** @brief Timer datatype */
45 decltype(simix_timers)::handle_type handle_;
46 simgrid::xbt::Task<void()> callback;
47 double getDate() { return date; }
48 s_smx_timer_t(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
51 void (*SMPI_switch_data_segment)(simgrid::s4u::ActorPtr) = nullptr;
53 bool _sg_do_verbose_exit = true;
54 static void inthandler(int)
56 if ( _sg_do_verbose_exit ) {
57 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
58 SIMIX_display_process_status();
61 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
67 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
69 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
70 fprintf(stderr, "Access violation detected.\n"
71 "This probably comes from a programming error in your code, or from a stack\n"
72 "overflow. If you are certain of your code, try increasing the stack size\n"
73 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
75 "If it does not help, this may have one of the following causes:\n"
76 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
77 "Failing hardware can sometimes generate such errors too.\n"
79 "If you think you've found a bug in SimGrid, please report it along with a\n"
80 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
81 "of the fault captured with gdb or valgrind.\n",
82 smx_context_stack_size / 1024);
83 } else if (siginfo->si_signo == SIGSEGV) {
84 fprintf(stderr, "Segmentation fault.\n");
86 if (smpi_enabled() && smpi_privatize_global_variables == SmpiPrivStrategies::NONE) {
87 #if HAVE_PRIVATIZATION
88 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
90 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
91 #endif /* HAVE_PRIVATIZATION */
93 #endif /* HAVE_SMPI */
98 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
101 * Install signal handler for SIGSEGV. Check that nobody has already installed
102 * its own handler. For example, the Java VM does this.
104 static void install_segvhandler()
108 stack.ss_sp = sigsegv_stack;
109 stack.ss_size = sizeof sigsegv_stack;
112 if (sigaltstack(&stack, &old_stack) == -1) {
113 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
116 if (not(old_stack.ss_flags & SS_DISABLE)) {
117 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
118 old_stack.ss_size, (unsigned)old_stack.ss_flags);
119 sigaltstack(&old_stack, nullptr);
122 struct sigaction action;
123 struct sigaction old_action;
124 action.sa_sigaction = &segvhandler;
125 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
126 sigemptyset(&action.sa_mask);
128 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
129 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
132 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
133 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
134 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
135 sigaction(SIGSEGV, &old_action, nullptr);
141 /********************************* SIMIX **************************************/
142 double SIMIX_timer_next()
144 return simix_timers.empty() ? -1.0 : simix_timers.top().first;
147 static void kill_process(smx_actor_t process)
149 SIMIX_process_kill(process, nullptr);
156 simgrid::config::Flag<double> breakpoint{"simix/breakpoint",
157 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
161 static simgrid::simix::ActorCode maestro_code;
162 void SIMIX_set_maestro(void (*code)(void*), void* data)
165 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
167 maestro_code = std::bind(code, data);
172 * @brief Initialize SIMIX internal data.
174 void SIMIX_global_init(int *argc, char **argv)
177 // The communication initialization is done ASAP.
178 // We need to communicate initialization of the different layers to the model-checker.
179 simgrid::mc::Client::initialize();
182 if (not simix_global) {
183 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
184 simix_global->maestro_process = nullptr;
186 surf_init(argc, argv); /* Initialize SURF structures */
187 SIMIX_context_mod_init();
189 // Either create a new context with maestro or create
190 // a context object with the current context mestro):
191 simgrid::kernel::actor::create_maestro(maestro_code);
193 /* Prepare to display some more info when dying on Ctrl-C pressing */
194 std::signal(SIGINT, inthandler);
197 install_segvhandler();
199 /* register a function to be called by SURF after the environment creation */
201 simgrid::s4u::on_platform_created.connect(SIMIX_post_create_environment);
203 simgrid::s4u::Storage::on_creation.connect([](simgrid::s4u::Storage& storage) {
204 sg_storage_t s = simgrid::s4u::Storage::by_name(storage.get_cname());
205 xbt_assert(s != nullptr, "Storage not found for name %s", storage.get_cname());
209 if (simgrid::config::get_value<bool>("clean-atexit"))
212 if (_sg_cfg_exit_asap)
219 * @brief Clean the SIMIX simulation
221 * This functions remove the memory used by SIMIX
226 return; // to avoid double cleaning by java and C
229 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
230 if (not simix_global->process_to_run.empty() && SIMIX_get_clock() <= 0.0) {
232 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
233 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
234 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
238 if (SIMIX_process_count()>0){
239 if(smpi_process()->initialized()){
240 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
242 XBT_WARN("Process called exit when leaving - Skipping cleanups");
248 /* Kill all processes (but maestro) */
249 SIMIX_process_killall(simix_global->maestro_process);
250 SIMIX_context_runall();
251 SIMIX_process_empty_trash();
253 /* Exit the SIMIX network module */
254 SIMIX_mailbox_exit();
256 while (not simix_timers.empty()) {
257 delete simix_timers.top().second;
260 /* Free the remaining data structures */
261 simix_global->process_to_run.clear();
262 simix_global->process_that_ran.clear();
263 simix_global->process_to_destroy.clear();
264 simix_global->process_list.clear();
267 xbt_dynar_free(&simix_global->actors_vector);
268 xbt_dynar_free(&simix_global->dead_actors_vector);
271 /* Let's free maestro now */
272 delete simix_global->maestro_process->context_;
273 simix_global->maestro_process->context_ = nullptr;
274 delete simix_global->maestro_process;
275 simix_global->maestro_process = nullptr;
277 /* Finish context module and SURF */
278 SIMIX_context_mod_exit();
282 simix_global = nullptr;
287 * @brief A clock (in second).
289 * @return Return the clock.
291 double SIMIX_get_clock()
293 if(MC_is_active() || MC_record_replay_is_active()){
294 return MC_process_clock_get(SIMIX_process_self());
296 return surf_get_clock();
300 /** Wake up all processes waiting for a Surf action to finish */
301 static void SIMIX_wake_processes()
303 for (auto const& model : all_existing_models) {
304 simgrid::kernel::resource::Action* action;
306 XBT_DEBUG("Handling the processes whose action failed (if any)");
307 while ((action = model->extract_failed_action())) {
308 XBT_DEBUG(" Handling Action %p",action);
309 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
311 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
312 while ((action = model->extract_done_action())) {
313 XBT_DEBUG(" Handling Action %p",action);
314 if (action->get_data() == nullptr)
315 XBT_DEBUG("probably vcpu's action %p, skip", action);
317 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
322 /** Handle any pending timer */
323 static bool SIMIX_execute_timers()
326 while (not simix_timers.empty() && SIMIX_get_clock() >= simix_timers.top().first) {
328 // FIXME: make the timers being real callbacks
329 // (i.e. provide dispatchers that read and expand the args)
330 smx_timer_t timer = simix_timers.top().second;
335 xbt_die("Exception thrown ouf of timer callback");
342 /** Execute all the tasks that are queued
344 * e.g. `.then()` callbacks of futures.
346 static bool SIMIX_execute_tasks()
348 xbt_assert(simix_global->tasksTemp.empty());
350 if (simix_global->tasks.empty())
355 // We don't want the callbacks to modify the vector we are iterating over:
356 swap(simix_global->tasks, simix_global->tasksTemp);
358 // Execute all the queued tasks:
359 for (auto& task : simix_global->tasksTemp)
362 simix_global->tasksTemp.clear();
363 } while (not simix_global->tasks.empty());
370 * @brief Run the main simulation loop.
374 if (not MC_record_path.empty()) {
375 simgrid::mc::replay(MC_record_path);
382 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
384 if (simgrid::simix::breakpoint >= 0.0 && surf_get_clock() >= simgrid::simix::breakpoint) {
385 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::breakpoint.get());
386 simgrid::simix::breakpoint = -1.0;
394 SIMIX_execute_tasks();
396 while (not simix_global->process_to_run.empty()) {
397 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
399 /* Run all processes that are ready to run, possibly in parallel */
400 SIMIX_process_runall();
402 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
404 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
406 /* Here, the order is ok because:
408 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do
409 * not impact its order.
411 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
413 * - if there is no kill during the simulation, processes remain sorted according by their PID.
414 * Rationale: This can be proved inductively.
415 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
416 * is parsed linearly).
417 * Let's show that it is still so at the end of this round.
418 * - if a process is added when being created, that's from maestro. It can be either at startup
419 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
420 * in arbitrary order (inductive hypothesis), we are fine.
421 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
422 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
423 * of a simcall terminates, we're still good. Proof:
424 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
425 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an
427 * - For communications, this function is called from SIMIX_comm_finish().
428 * This function itself don't mess with the order since simcalls are handled in FIFO order.
429 * The function is called:
430 * - before the comm starts (invalid parameters, or resource already dead or whatever).
431 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
432 * - because the communication failed or were canceled after startup. In this case, it's called from
433 * the function we are in, by the chunk:
434 * set = model->states.failed_action_set;
435 * while ((synchro = extract(set)))
436 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
437 * This order is also fixed because it depends of the order in which the surf actions were
438 * added to the system, and only maestro can add stuff this way, through simcalls.
439 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
440 * poped out of the set does not depend on the user code's execution order.
441 * - because the communication terminated. In this case, synchros are served in the order given by
442 * set = model->states.done_action_set;
443 * while ((synchro = extract(set)))
444 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
445 * and the argument is very similar to the previous one.
446 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user
447 * processes are executed.
448 * So, in any cases, the orders of processes within process_to_run do not depend on the order in which
449 * user processes were executed previously.
450 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
451 * - If there is some process killings, the order is changed by this decision that comes from user-land
452 * But this decision may not have been motivated by a situation that were different because the simulation is
454 * So, even the order change induced by the process killing is perfectly reproducible.
456 * So science works, bitches [http://xkcd.com/54/].
458 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are
459 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
461 * That would thus be a pure waste of time.
464 for (smx_actor_t const& process : simix_global->process_that_ran) {
465 if (process->simcall.call != SIMCALL_NONE) {
466 SIMIX_simcall_handle(&process->simcall, 0);
470 SIMIX_execute_tasks();
472 SIMIX_wake_processes();
473 } while (SIMIX_execute_tasks());
475 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
476 if (simix_global->process_list.size() == simix_global->daemons.size())
477 for (auto const& dmon : simix_global->daemons) {
478 XBT_DEBUG("Kill %s", dmon->get_cname());
479 SIMIX_process_kill(dmon, simix_global->maestro_process);
483 time = SIMIX_timer_next();
484 if (time > -1.0 || not simix_global->process_list.empty()) {
485 XBT_DEBUG("Calling surf_solve");
486 time = surf_solve(time);
487 XBT_DEBUG("Moving time ahead : %g", time);
490 /* Notify all the hosts that have failed */
491 /* FIXME: iterate through the list of failed host and mark each of them */
492 /* as failed. On each host, signal all the running processes with host_fail */
494 // Execute timers and tasks until there isn't anything to be done:
497 again = SIMIX_execute_timers();
498 if (SIMIX_execute_tasks())
500 SIMIX_wake_processes();
503 /* Clean processes to destroy */
504 SIMIX_process_empty_trash();
506 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
507 simix_global->process_to_run.size());
509 } while (time > -1.0 || not simix_global->process_to_run.empty());
511 if (not simix_global->process_list.empty()) {
513 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
514 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
515 "once the simulation is over. Please fix your on_exit() functions.");
517 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
519 SIMIX_display_process_status();
520 simgrid::s4u::on_deadlock();
523 simgrid::s4u::on_simulation_end();
527 * @brief Set the date to execute a function
529 * Set the date to execute the function on the surf.
530 * @param date Date to execute function
531 * @param callback Function to be executed
532 * @param arg Parameters of the function
535 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
537 smx_timer_t timer = new s_smx_timer_t(date, simgrid::xbt::make_task([callback, arg]() { callback(arg); }));
538 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
542 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
544 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
545 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
549 /** @brief cancels a timer that was added earlier */
550 void SIMIX_timer_remove(smx_timer_t timer) {
551 simix_timers.erase(timer->handle_);
555 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
556 double SIMIX_timer_get_date(smx_timer_t timer) {
557 return timer ? timer->getDate() : 0;
560 void SIMIX_display_process_status()
562 int nbprocess = simix_global->process_list.size();
564 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
565 /* List the process and their state */
566 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
567 for (auto const& kv : simix_global->process_list) {
568 smx_actor_t process = kv.second;
570 if (process->waiting_synchro) {
572 const char* synchro_description = "unknown";
574 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
575 synchro_description = "execution";
577 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
578 synchro_description = "communication";
580 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
581 synchro_description = "sleeping";
583 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
584 synchro_description = "synchronization";
586 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
587 synchro_description = "I/O";
589 XBT_INFO("Process %ld (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid_,
590 process->get_cname(), process->host_->get_cname(), synchro_description, process->waiting_synchro.get(),
591 process->waiting_synchro->name_.c_str(), (int)process->waiting_synchro->state_);
594 XBT_INFO("Process %ld (%s@%s)", process->pid_, process->get_cname(), process->host_->get_cname());
599 int SIMIX_is_maestro()
601 smx_actor_t self = SIMIX_process_self();
602 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;