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"
10 #include "../kernel/activity/IoImpl.hpp"
11 #include "simgrid/sg_config.hpp"
12 #include "smx_private.hpp"
13 #include "src/kernel/activity/SleepImpl.hpp"
14 #include "src/kernel/activity/SynchroRaw.hpp"
15 #include "src/mc/mc_record.hpp"
16 #include "src/mc/mc_replay.hpp"
17 #include "src/simix/smx_host_private.hpp"
18 #include "src/smpi/include/smpi_process.hpp"
19 #include "src/surf/StorageImpl.hpp"
20 #include "src/surf/xml/platf.hpp"
23 #include "src/mc/mc_private.hpp"
24 #include "src/mc/remote/Client.hpp"
25 #include "src/mc/remote/mc_protocol.h"
29 #include "src/smpi/include/private.hpp"
32 #include <boost/heap/fibonacci_heap.hpp>
35 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
36 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
38 std::unique_ptr<simgrid::simix::Global> simix_global;
41 typedef std::pair<double, smx_timer_t> TimerQelt;
42 boost::heap::fibonacci_heap<TimerQelt, boost::heap::compare<simgrid::xbt::HeapComparator<TimerQelt>>> simix_timers;
45 /** @brief Timer datatype */
50 decltype(simix_timers)::handle_type handle_;
51 simgrid::xbt::Task<void()> callback;
52 double getDate() { return date; }
53 s_smx_timer_t(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
56 void (*SMPI_switch_data_segment)(simgrid::s4u::ActorPtr) = nullptr;
58 bool _sg_do_verbose_exit = true;
59 static void inthandler(int)
61 if ( _sg_do_verbose_exit ) {
62 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
63 SIMIX_display_process_status();
66 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
72 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
74 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
75 fprintf(stderr, "Access violation detected.\n"
76 "This probably comes from a programming error in your code, or from a stack\n"
77 "overflow. If you are certain of your code, try increasing the stack size\n"
78 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
80 "If it does not help, this may have one of the following causes:\n"
81 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
82 "Failing hardware can sometimes generate such errors too.\n"
84 "If you think you've found a bug in SimGrid, please report it along with a\n"
85 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
86 "of the fault captured with gdb or valgrind.\n",
87 smx_context_stack_size / 1024);
88 } else if (siginfo->si_signo == SIGSEGV) {
89 fprintf(stderr, "Segmentation fault.\n");
91 if (smpi_enabled() && smpi_privatize_global_variables == SmpiPrivStrategies::NONE) {
92 #if HAVE_PRIVATIZATION
93 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
95 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
96 #endif /* HAVE_PRIVATIZATION */
98 #endif /* HAVE_SMPI */
103 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
106 * Install signal handler for SIGSEGV. Check that nobody has already installed
107 * its own handler. For example, the Java VM does this.
109 static void install_segvhandler()
113 stack.ss_sp = sigsegv_stack;
114 stack.ss_size = sizeof sigsegv_stack;
117 if (sigaltstack(&stack, &old_stack) == -1) {
118 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
121 if (not(old_stack.ss_flags & SS_DISABLE)) {
122 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
123 old_stack.ss_size, (unsigned)old_stack.ss_flags);
124 sigaltstack(&old_stack, nullptr);
127 struct sigaction action;
128 struct sigaction old_action;
129 action.sa_sigaction = &segvhandler;
130 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
131 sigemptyset(&action.sa_mask);
133 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
134 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", 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 SIGSEGV (%p). Restore it.",
139 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
140 sigaction(SIGSEGV, &old_action, nullptr);
146 /********************************* SIMIX **************************************/
147 double SIMIX_timer_next()
149 return simix_timers.empty() ? -1.0 : simix_timers.top().first;
152 static void kill_process(smx_actor_t process)
154 SIMIX_process_kill(process, nullptr);
161 simgrid::xbt::signal<void()> onDeadlock;
163 simgrid::config::Flag<double> breakpoint{"simix/breakpoint",
164 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
168 static std::function<void()> maestro_code;
169 void SIMIX_set_maestro(void (*code)(void*), void* data)
172 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
174 maestro_code = std::bind(code, data);
179 * \brief Initialize SIMIX internal data.
181 void SIMIX_global_init(int *argc, char **argv)
184 // The communication initialization is done ASAP.
185 // We need to communicate initialization of the different layers to the model-checker.
186 simgrid::mc::Client::initialize();
189 if (not simix_global) {
190 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
191 simix_global->maestro_process = nullptr;
192 simix_global->create_process_function = &SIMIX_process_create;
193 simix_global->kill_process_function = &kill_process;
194 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
195 simix_global->mutex = xbt_os_mutex_init();
197 surf_init(argc, argv); /* Initialize SURF structures */
198 SIMIX_context_mod_init();
200 // Either create a new context with maestro or create
201 // a context object with the current context mestro):
202 simgrid::kernel::actor::create_maestro(maestro_code);
204 /* Prepare to display some more info when dying on Ctrl-C pressing */
205 std::signal(SIGINT, inthandler);
208 install_segvhandler();
210 /* register a function to be called by SURF after the environment creation */
212 simgrid::s4u::on_platform_created.connect(SIMIX_post_create_environment);
213 simgrid::s4u::Host::on_creation.connect([](simgrid::s4u::Host& host) {
214 if (host.extension<simgrid::simix::Host>() == nullptr) // another callback to the same signal may have created it
215 host.extension_set<simgrid::simix::Host>(new simgrid::simix::Host());
218 simgrid::s4u::Storage::on_creation.connect([](simgrid::s4u::Storage& storage) {
219 sg_storage_t s = simgrid::s4u::Storage::by_name(storage.get_cname());
220 xbt_assert(s != nullptr, "Storage not found for name %s", storage.get_cname());
224 if (simgrid::config::get_value<bool>("clean-atexit"))
227 if (_sg_cfg_exit_asap)
234 * \brief Clean the SIMIX simulation
236 * This functions remove the memory used by SIMIX
241 return; // to avoid double cleaning by java and C
244 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
245 if (not simix_global->process_to_run.empty() && SIMIX_get_clock() <= 0.0) {
247 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
248 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
249 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
253 if (SIMIX_process_count()>0){
254 if(smpi_process()->initialized()){
255 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
257 XBT_WARN("Process called exit when leaving - Skipping cleanups");
263 /* Kill all processes (but maestro) */
264 SIMIX_process_killall(simix_global->maestro_process);
265 SIMIX_context_runall();
266 SIMIX_process_empty_trash();
268 /* Exit the SIMIX network module */
269 SIMIX_mailbox_exit();
271 while (not simix_timers.empty()) {
272 delete simix_timers.top().second;
275 /* Free the remaining data structures */
276 simix_global->process_to_run.clear();
277 simix_global->process_that_ran.clear();
278 simix_global->process_to_destroy.clear();
279 simix_global->process_list.clear();
281 xbt_os_mutex_destroy(simix_global->mutex);
282 simix_global->mutex = nullptr;
284 xbt_dynar_free(&simix_global->actors_vector);
285 xbt_dynar_free(&simix_global->dead_actors_vector);
288 /* Let's free maestro now */
289 delete simix_global->maestro_process->context;
290 simix_global->maestro_process->context = nullptr;
291 delete simix_global->maestro_process;
292 simix_global->maestro_process = nullptr;
294 /* Finish context module and SURF */
295 SIMIX_context_mod_exit();
299 simix_global = nullptr;
305 * \brief A clock (in second).
307 * \return Return the clock.
309 double SIMIX_get_clock()
311 if(MC_is_active() || MC_record_replay_is_active()){
312 return MC_process_clock_get(SIMIX_process_self());
314 return surf_get_clock();
318 /** Wake up all processes waiting for a Surf action to finish */
319 static void SIMIX_wake_processes()
321 for (auto const& model : *all_existing_models) {
322 simgrid::kernel::resource::Action* action;
324 XBT_DEBUG("Handling the processes whose action failed (if any)");
325 while ((action = surf_model_extract_failed_action_set(model))) {
326 XBT_DEBUG(" Handling Action %p",action);
327 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
329 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
330 while ((action = surf_model_extract_done_action_set(model))) {
331 XBT_DEBUG(" Handling Action %p",action);
332 if (action->get_data() == nullptr)
333 XBT_DEBUG("probably vcpu's action %p, skip", action);
335 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
340 /** Handle any pending timer */
341 static bool SIMIX_execute_timers()
344 while (not simix_timers.empty() && SIMIX_get_clock() >= simix_timers.top().first) {
346 // FIXME: make the timers being real callbacks
347 // (i.e. provide dispatchers that read and expand the args)
348 smx_timer_t timer = simix_timers.top().second;
353 xbt_die("Exception thrown ouf of timer callback");
360 /** Execute all the tasks that are queued
362 * e.g. `.then()` callbacks of futures.
364 static bool SIMIX_execute_tasks()
366 xbt_assert(simix_global->tasksTemp.empty());
368 if (simix_global->tasks.empty())
373 // We don't want the callbacks to modify the vector we are iterating over:
374 swap(simix_global->tasks, simix_global->tasksTemp);
376 // Execute all the queued tasks:
377 for (auto& task : simix_global->tasksTemp)
380 simix_global->tasksTemp.clear();
381 } while (not simix_global->tasks.empty());
388 * \brief Run the main simulation loop.
392 if (not MC_record_path.empty()) {
393 simgrid::mc::replay(MC_record_path);
400 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
402 if (simgrid::simix::breakpoint >= 0.0 && time >= simgrid::simix::breakpoint) {
403 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::breakpoint.get());
404 simgrid::simix::breakpoint = -1.0;
412 SIMIX_execute_tasks();
414 while (not simix_global->process_to_run.empty()) {
415 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
417 /* Run all processes that are ready to run, possibly in parallel */
418 SIMIX_process_runall();
420 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
422 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
424 /* Here, the order is ok because:
426 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do
427 * not impact its order.
429 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
431 * - if there is no kill during the simulation, processes remain sorted according by their PID.
432 * Rationale: This can be proved inductively.
433 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
434 * is parsed linearly).
435 * Let's show that it is still so at the end of this round.
436 * - if a process is added when being created, that's from maestro. It can be either at startup
437 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
438 * in arbitrary order (inductive hypothesis), we are fine.
439 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
440 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
441 * of a simcall terminates, we're still good. Proof:
442 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
443 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an
445 * - For communications, this function is called from SIMIX_comm_finish().
446 * This function itself don't mess with the order since simcalls are handled in FIFO order.
447 * The function is called:
448 * - before the comm starts (invalid parameters, or resource already dead or whatever).
449 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
450 * - because the communication failed or were canceled after startup. In this case, it's called from
451 * the function we are in, by the chunk:
452 * set = model->states.failed_action_set;
453 * while ((synchro = extract(set)))
454 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
455 * This order is also fixed because it depends of the order in which the surf actions were
456 * added to the system, and only maestro can add stuff this way, through simcalls.
457 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
458 * poped out of the set does not depend on the user code's execution order.
459 * - because the communication terminated. In this case, synchros are served in the order given by
460 * set = model->states.done_action_set;
461 * while ((synchro = extract(set)))
462 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
463 * and the argument is very similar to the previous one.
464 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user
465 * processes are executed.
466 * So, in any cases, the orders of processes within process_to_run do not depend on the order in which
467 * user processes were executed previously.
468 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
469 * - If there is some process killings, the order is changed by this decision that comes from user-land
470 * But this decision may not have been motivated by a situation that were different because the simulation is
472 * So, even the order change induced by the process killing is perfectly reproducible.
474 * So science works, bitches [http://xkcd.com/54/].
476 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are
477 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
479 * That would thus be a pure waste of time.
482 for (smx_actor_t const& process : simix_global->process_that_ran) {
483 if (process->simcall.call != SIMCALL_NONE) {
484 SIMIX_simcall_handle(&process->simcall, 0);
488 SIMIX_execute_tasks();
490 SIMIX_wake_processes();
491 } while (SIMIX_execute_tasks());
493 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
494 if (simix_global->process_list.size() == simix_global->daemons.size())
495 for (auto const& dmon : simix_global->daemons) {
496 XBT_DEBUG("Kill %s", dmon->get_cname());
497 SIMIX_process_kill(dmon, simix_global->maestro_process);
501 time = SIMIX_timer_next();
502 if (time > -1.0 || not simix_global->process_list.empty()) {
503 XBT_DEBUG("Calling surf_solve");
504 time = surf_solve(time);
505 XBT_DEBUG("Moving time ahead : %g", time);
508 /* Notify all the hosts that have failed */
509 /* FIXME: iterate through the list of failed host and mark each of them */
510 /* as failed. On each host, signal all the running processes with host_fail */
512 // Execute timers and tasks until there isn't anything to be done:
515 again = SIMIX_execute_timers();
516 if (SIMIX_execute_tasks())
518 SIMIX_wake_processes();
521 /* Autorestart all process */
522 for (auto const& host : host_that_restart) {
523 XBT_INFO("Restart processes on host %s", host->get_cname());
524 SIMIX_host_autorestart(host);
526 host_that_restart.clear();
528 /* Clean processes to destroy */
529 SIMIX_process_empty_trash();
531 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
532 simix_global->process_to_run.size());
534 if (simix_global->process_to_run.empty() && not simix_global->process_list.empty())
535 simgrid::simix::onDeadlock();
537 } while (time > -1.0 || not simix_global->process_to_run.empty());
539 if (not simix_global->process_list.empty()) {
541 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
542 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
543 "once the simulation is over. Please fix your on_exit() functions.");
545 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
547 SIMIX_display_process_status();
548 simgrid::s4u::on_deadlock();
551 simgrid::s4u::on_simulation_end();
555 * \brief Set the date to execute a function
557 * Set the date to execute the function on the surf.
558 * \param date Date to execute function
559 * \param callback Function to be executed
560 * \param arg Parameters of the function
563 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
565 smx_timer_t timer = new s_smx_timer_t(date, simgrid::xbt::makeTask([callback, arg]() { callback(arg); }));
566 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
570 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
572 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
573 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
577 /** @brief cancels a timer that was added earlier */
578 void SIMIX_timer_remove(smx_timer_t timer) {
579 simix_timers.erase(timer->handle_);
583 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
584 double SIMIX_timer_get_date(smx_timer_t timer) {
585 return timer ? timer->getDate() : 0;
589 * \brief Registers a function to create a process.
591 * This function registers a function to be called
592 * when a new process is created. The function has
593 * to call SIMIX_process_create().
594 * \param function create process function
596 void SIMIX_function_register_process_create(smx_creation_func_t function)
598 simix_global->create_process_function = function;
602 * \brief Registers a function to kill a process.
604 * This function registers a function to be called when a process is killed. The function has to call the
605 * SIMIX_process_kill().
607 * \param function Kill process function
609 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
611 simix_global->kill_process_function = function;
615 * \brief Registers a function to cleanup a process.
617 * This function registers a user function to be called when a process ends properly.
619 * \param function cleanup process function
621 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
623 simix_global->cleanup_process_function = function;
627 void SIMIX_display_process_status()
629 int nbprocess = simix_global->process_list.size();
631 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
632 /* List the process and their state */
633 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
634 for (auto const& kv : simix_global->process_list) {
635 smx_actor_t process = kv.second;
637 if (process->waiting_synchro) {
639 const char* synchro_description = "unknown";
641 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
642 synchro_description = "execution";
644 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
645 synchro_description = "communication";
647 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
648 synchro_description = "sleeping";
650 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
651 synchro_description = "synchronization";
653 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
654 synchro_description = "I/O";
656 XBT_INFO("Process %ld (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid,
657 process->get_cname(), process->host->get_cname(), synchro_description, process->waiting_synchro.get(),
658 process->waiting_synchro->name_.c_str(), (int)process->waiting_synchro->state_);
661 XBT_INFO("Process %ld (%s@%s)", process->pid, process->get_cname(), process->host->get_cname());
666 int SIMIX_is_maestro()
668 smx_actor_t self = SIMIX_process_self();
669 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;