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;
185 simix_global->create_process_function = &SIMIX_process_create;
186 simix_global->kill_process_function = &kill_process;
187 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
188 simix_global->mutex = xbt_os_mutex_init();
190 surf_init(argc, argv); /* Initialize SURF structures */
191 SIMIX_context_mod_init();
193 // Either create a new context with maestro or create
194 // a context object with the current context mestro):
195 simgrid::kernel::actor::create_maestro(maestro_code);
197 /* Prepare to display some more info when dying on Ctrl-C pressing */
198 std::signal(SIGINT, inthandler);
201 install_segvhandler();
203 /* register a function to be called by SURF after the environment creation */
205 simgrid::s4u::on_platform_created.connect(SIMIX_post_create_environment);
207 simgrid::s4u::Storage::on_creation.connect([](simgrid::s4u::Storage& storage) {
208 sg_storage_t s = simgrid::s4u::Storage::by_name(storage.get_cname());
209 xbt_assert(s != nullptr, "Storage not found for name %s", storage.get_cname());
213 if (simgrid::config::get_value<bool>("clean-atexit"))
216 if (_sg_cfg_exit_asap)
223 * @brief Clean the SIMIX simulation
225 * This functions remove the memory used by SIMIX
230 return; // to avoid double cleaning by java and C
233 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
234 if (not simix_global->process_to_run.empty() && SIMIX_get_clock() <= 0.0) {
236 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
237 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
238 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
242 if (SIMIX_process_count()>0){
243 if(smpi_process()->initialized()){
244 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
246 XBT_WARN("Process called exit when leaving - Skipping cleanups");
252 /* Kill all processes (but maestro) */
253 SIMIX_process_killall(simix_global->maestro_process);
254 SIMIX_context_runall();
255 SIMIX_process_empty_trash();
257 /* Exit the SIMIX network module */
258 SIMIX_mailbox_exit();
260 while (not simix_timers.empty()) {
261 delete simix_timers.top().second;
264 /* Free the remaining data structures */
265 simix_global->process_to_run.clear();
266 simix_global->process_that_ran.clear();
267 simix_global->process_to_destroy.clear();
268 simix_global->process_list.clear();
270 xbt_os_mutex_destroy(simix_global->mutex);
271 simix_global->mutex = nullptr;
273 xbt_dynar_free(&simix_global->actors_vector);
274 xbt_dynar_free(&simix_global->dead_actors_vector);
277 /* Let's free maestro now */
278 delete simix_global->maestro_process->context_;
279 simix_global->maestro_process->context_ = nullptr;
280 delete simix_global->maestro_process;
281 simix_global->maestro_process = nullptr;
283 /* Finish context module and SURF */
284 SIMIX_context_mod_exit();
288 simix_global = nullptr;
293 * @brief A clock (in second).
295 * @return Return the clock.
297 double SIMIX_get_clock()
299 if(MC_is_active() || MC_record_replay_is_active()){
300 return MC_process_clock_get(SIMIX_process_self());
302 return surf_get_clock();
306 /** Wake up all processes waiting for a Surf action to finish */
307 static void SIMIX_wake_processes()
309 for (auto const& model : all_existing_models) {
310 simgrid::kernel::resource::Action* action;
312 XBT_DEBUG("Handling the processes whose action failed (if any)");
313 while ((action = model->extract_failed_action())) {
314 XBT_DEBUG(" Handling Action %p",action);
315 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
317 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
318 while ((action = model->extract_done_action())) {
319 XBT_DEBUG(" Handling Action %p",action);
320 if (action->get_data() == nullptr)
321 XBT_DEBUG("probably vcpu's action %p, skip", action);
323 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
328 /** Handle any pending timer */
329 static bool SIMIX_execute_timers()
332 while (not simix_timers.empty() && SIMIX_get_clock() >= simix_timers.top().first) {
334 // FIXME: make the timers being real callbacks
335 // (i.e. provide dispatchers that read and expand the args)
336 smx_timer_t timer = simix_timers.top().second;
341 xbt_die("Exception thrown ouf of timer callback");
348 /** Execute all the tasks that are queued
350 * e.g. `.then()` callbacks of futures.
352 static bool SIMIX_execute_tasks()
354 xbt_assert(simix_global->tasksTemp.empty());
356 if (simix_global->tasks.empty())
361 // We don't want the callbacks to modify the vector we are iterating over:
362 swap(simix_global->tasks, simix_global->tasksTemp);
364 // Execute all the queued tasks:
365 for (auto& task : simix_global->tasksTemp)
368 simix_global->tasksTemp.clear();
369 } while (not simix_global->tasks.empty());
376 * @brief Run the main simulation loop.
380 if (not MC_record_path.empty()) {
381 simgrid::mc::replay(MC_record_path);
388 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
390 if (simgrid::simix::breakpoint >= 0.0 && surf_get_clock() >= simgrid::simix::breakpoint) {
391 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::breakpoint.get());
392 simgrid::simix::breakpoint = -1.0;
400 SIMIX_execute_tasks();
402 while (not simix_global->process_to_run.empty()) {
403 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
405 /* Run all processes that are ready to run, possibly in parallel */
406 SIMIX_process_runall();
408 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
410 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
412 /* Here, the order is ok because:
414 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do
415 * not impact its order.
417 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
419 * - if there is no kill during the simulation, processes remain sorted according by their PID.
420 * Rationale: This can be proved inductively.
421 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
422 * is parsed linearly).
423 * Let's show that it is still so at the end of this round.
424 * - if a process is added when being created, that's from maestro. It can be either at startup
425 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
426 * in arbitrary order (inductive hypothesis), we are fine.
427 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
428 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
429 * of a simcall terminates, we're still good. Proof:
430 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
431 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an
433 * - For communications, this function is called from SIMIX_comm_finish().
434 * This function itself don't mess with the order since simcalls are handled in FIFO order.
435 * The function is called:
436 * - before the comm starts (invalid parameters, or resource already dead or whatever).
437 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
438 * - because the communication failed or were canceled after startup. In this case, it's called from
439 * the function we are in, by the chunk:
440 * set = model->states.failed_action_set;
441 * while ((synchro = extract(set)))
442 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
443 * This order is also fixed because it depends of the order in which the surf actions were
444 * added to the system, and only maestro can add stuff this way, through simcalls.
445 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
446 * poped out of the set does not depend on the user code's execution order.
447 * - because the communication terminated. In this case, synchros are served in the order given by
448 * set = model->states.done_action_set;
449 * while ((synchro = extract(set)))
450 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
451 * and the argument is very similar to the previous one.
452 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user
453 * processes are executed.
454 * So, in any cases, the orders of processes within process_to_run do not depend on the order in which
455 * user processes were executed previously.
456 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
457 * - If there is some process killings, the order is changed by this decision that comes from user-land
458 * But this decision may not have been motivated by a situation that were different because the simulation is
460 * So, even the order change induced by the process killing is perfectly reproducible.
462 * So science works, bitches [http://xkcd.com/54/].
464 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are
465 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
467 * That would thus be a pure waste of time.
470 for (smx_actor_t const& process : simix_global->process_that_ran) {
471 if (process->simcall.call != SIMCALL_NONE) {
472 SIMIX_simcall_handle(&process->simcall, 0);
476 SIMIX_execute_tasks();
478 SIMIX_wake_processes();
479 } while (SIMIX_execute_tasks());
481 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
482 if (simix_global->process_list.size() == simix_global->daemons.size())
483 for (auto const& dmon : simix_global->daemons) {
484 XBT_DEBUG("Kill %s", dmon->get_cname());
485 SIMIX_process_kill(dmon, simix_global->maestro_process);
489 time = SIMIX_timer_next();
490 if (time > -1.0 || not simix_global->process_list.empty()) {
491 XBT_DEBUG("Calling surf_solve");
492 time = surf_solve(time);
493 XBT_DEBUG("Moving time ahead : %g", time);
496 /* Notify all the hosts that have failed */
497 /* FIXME: iterate through the list of failed host and mark each of them */
498 /* as failed. On each host, signal all the running processes with host_fail */
500 // Execute timers and tasks until there isn't anything to be done:
503 again = SIMIX_execute_timers();
504 if (SIMIX_execute_tasks())
506 SIMIX_wake_processes();
509 /* Autorestart all process */
510 for (auto const& host : host_that_restart) {
511 XBT_INFO("Restart processes on host %s", host->get_cname());
512 SIMIX_host_autorestart(host);
514 host_that_restart.clear();
516 /* Clean processes to destroy */
517 SIMIX_process_empty_trash();
519 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
520 simix_global->process_to_run.size());
522 } while (time > -1.0 || not simix_global->process_to_run.empty());
524 if (not simix_global->process_list.empty()) {
526 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
527 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
528 "once the simulation is over. Please fix your on_exit() functions.");
530 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
532 SIMIX_display_process_status();
533 simgrid::s4u::on_deadlock();
536 simgrid::s4u::on_simulation_end();
540 * @brief Set the date to execute a function
542 * Set the date to execute the function on the surf.
543 * @param date Date to execute function
544 * @param callback Function to be executed
545 * @param arg Parameters of the function
548 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
550 smx_timer_t timer = new s_smx_timer_t(date, simgrid::xbt::make_task([callback, arg]() { callback(arg); }));
551 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
555 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
557 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
558 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
562 /** @brief cancels a timer that was added earlier */
563 void SIMIX_timer_remove(smx_timer_t timer) {
564 simix_timers.erase(timer->handle_);
568 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
569 double SIMIX_timer_get_date(smx_timer_t timer) {
570 return timer ? timer->getDate() : 0;
574 * @brief Registers a function to create a process.
576 * This function registers a function to be called
577 * when a new process is created. The function has
578 * to call SIMIX_process_create().
579 * @param function create process function
581 void SIMIX_function_register_process_create(smx_creation_func_t function)
583 simix_global->create_process_function = function;
587 * @brief Registers a function to kill a process.
589 * This function registers a function to be called when a process is killed. The function has to call the
590 * SIMIX_process_kill().
592 * @param function Kill process function
594 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
596 simix_global->kill_process_function = function;
600 * @brief Registers a function to cleanup a process.
602 * This function registers a user function to be called when a process ends properly.
604 * @param function cleanup process function
606 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
608 simix_global->cleanup_process_function = function;
612 void SIMIX_display_process_status()
614 int nbprocess = simix_global->process_list.size();
616 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
617 /* List the process and their state */
618 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
619 for (auto const& kv : simix_global->process_list) {
620 smx_actor_t process = kv.second;
622 if (process->waiting_synchro) {
624 const char* synchro_description = "unknown";
626 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
627 synchro_description = "execution";
629 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
630 synchro_description = "communication";
632 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
633 synchro_description = "sleeping";
635 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
636 synchro_description = "synchronization";
638 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
639 synchro_description = "I/O";
641 XBT_INFO("Process %ld (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid_,
642 process->get_cname(), process->host_->get_cname(), synchro_description, process->waiting_synchro.get(),
643 process->waiting_synchro->name_.c_str(), (int)process->waiting_synchro->state_);
646 XBT_INFO("Process %ld (%s@%s)", process->pid_, process->get_cname(), process->host_->get_cname());
651 int SIMIX_is_maestro()
653 smx_actor_t self = SIMIX_process_self();
654 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;