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 "simgrid/sg_config.hpp"
10 #include "smx_private.hpp"
11 #include "src/kernel/activity/SleepImpl.hpp"
12 #include "src/kernel/activity/SynchroIo.hpp"
13 #include "src/kernel/activity/SynchroRaw.hpp"
14 #include "src/mc/mc_record.hpp"
15 #include "src/mc/mc_replay.hpp"
16 #include "src/simix/smx_host_private.hpp"
17 #include "src/smpi/include/smpi_process.hpp"
18 #include "src/surf/StorageImpl.hpp"
19 #include "src/surf/xml/platf.hpp"
22 #include "src/mc/mc_private.hpp"
23 #include "src/mc/remote/Client.hpp"
24 #include "src/mc/remote/mc_protocol.h"
28 #include "src/smpi/include/private.hpp"
31 #include <boost/heap/fibonacci_heap.hpp>
34 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
35 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
37 std::unique_ptr<simgrid::simix::Global> simix_global;
40 typedef std::pair<double, smx_timer_t> TimerQelt;
41 boost::heap::fibonacci_heap<TimerQelt, boost::heap::compare<simgrid::xbt::HeapComparator<TimerQelt>>> simix_timers;
44 /** @brief Timer datatype */
49 decltype(simix_timers)::handle_type handle_;
50 simgrid::xbt::Task<void()> callback;
51 double getDate() { return date; }
52 s_smx_timer_t(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
55 void (*SMPI_switch_data_segment)(simgrid::s4u::ActorPtr) = nullptr;
57 bool _sg_do_verbose_exit = true;
58 static void inthandler(int)
60 if ( _sg_do_verbose_exit ) {
61 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
62 SIMIX_display_process_status();
65 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
71 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
73 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
74 fprintf(stderr, "Access violation detected.\n"
75 "This probably comes from a programming error in your code, or from a stack\n"
76 "overflow. If you are certain of your code, try increasing the stack size\n"
77 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
79 "If it does not help, this may have one of the following causes:\n"
80 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
81 "Failing hardware can sometimes generate such errors too.\n"
83 "If you think you've found a bug in SimGrid, please report it along with a\n"
84 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
85 "of the fault captured with gdb or valgrind.\n",
86 smx_context_stack_size / 1024);
87 } else if (siginfo->si_signo == SIGSEGV) {
88 fprintf(stderr, "Segmentation fault.\n");
90 if (smpi_enabled() && smpi_privatize_global_variables == SmpiPrivStrategies::None) {
91 #if HAVE_PRIVATIZATION
92 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
94 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
95 #endif /* HAVE_PRIVATIZATION */
97 #endif /* HAVE_SMPI */
102 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
105 * Install signal handler for SIGSEGV. Check that nobody has already installed
106 * its own handler. For example, the Java VM does this.
108 static void install_segvhandler()
112 stack.ss_sp = sigsegv_stack;
113 stack.ss_size = sizeof sigsegv_stack;
116 if (sigaltstack(&stack, &old_stack) == -1) {
117 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
120 if (not(old_stack.ss_flags & SS_DISABLE)) {
121 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
122 old_stack.ss_size, (unsigned)old_stack.ss_flags);
123 sigaltstack(&old_stack, nullptr);
126 struct sigaction action;
127 struct sigaction old_action;
128 action.sa_sigaction = &segvhandler;
129 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
130 sigemptyset(&action.sa_mask);
132 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
133 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
136 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
137 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
138 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
139 sigaction(SIGSEGV, &old_action, nullptr);
145 /********************************* SIMIX **************************************/
146 double SIMIX_timer_next()
148 return simix_timers.empty() ? -1.0 : simix_timers.top().first;
151 static void kill_process(smx_actor_t process)
153 SIMIX_process_kill(process, nullptr);
160 simgrid::xbt::signal<void()> onDeadlock;
162 simgrid::config::Flag<double> breakpoint{"simix/breakpoint",
163 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
167 static std::function<void()> maestro_code;
168 void SIMIX_set_maestro(void (*code)(void*), void* data)
171 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
173 maestro_code = std::bind(code, data);
178 * \brief Initialize SIMIX internal data.
180 void SIMIX_global_init(int *argc, char **argv)
183 // The communication initialization is done ASAP.
184 // We need to communicate initialization of the different layers to the model-checker.
185 simgrid::mc::Client::initialize();
188 if (not simix_global) {
189 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
190 simix_global->maestro_process = nullptr;
191 simix_global->create_process_function = &SIMIX_process_create;
192 simix_global->kill_process_function = &kill_process;
193 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
194 simix_global->mutex = xbt_os_mutex_init();
196 surf_init(argc, argv); /* Initialize SURF structures */
197 SIMIX_context_mod_init();
199 // Either create a new context with maestro or create
200 // a context object with the current context mestro):
201 simgrid::kernel::actor::create_maestro(maestro_code);
203 /* Prepare to display some more info when dying on Ctrl-C pressing */
204 std::signal(SIGINT, inthandler);
207 install_segvhandler();
209 /* register a function to be called by SURF after the environment creation */
211 simgrid::s4u::on_platform_created.connect(SIMIX_post_create_environment);
212 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
213 if (host.extension<simgrid::simix::Host>() == nullptr) // another callback to the same signal may have created it
214 host.extension_set<simgrid::simix::Host>(new simgrid::simix::Host());
217 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::StorageImpl* storage) {
218 sg_storage_t s = simgrid::s4u::Storage::byName(storage->get_cname());
219 xbt_assert(s != nullptr, "Storage not found for name %s", storage->get_cname());
223 if (simgrid::config::get_value<bool>("clean-atexit"))
226 if (_sg_cfg_exit_asap)
233 * \brief Clean the SIMIX simulation
235 * This functions remove the memory used by SIMIX
240 return; // to avoid double cleaning by java and C
243 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
244 if (not simix_global->process_to_run.empty() && SIMIX_get_clock() <= 0.0) {
246 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
247 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
248 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
252 if (SIMIX_process_count()>0){
253 if(smpi_process()->initialized()){
254 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
256 XBT_WARN("Process called exit when leaving - Skipping cleanups");
262 /* Kill all processes (but maestro) */
263 SIMIX_process_killall(simix_global->maestro_process);
264 SIMIX_context_runall();
265 SIMIX_process_empty_trash();
267 /* Exit the SIMIX network module */
268 SIMIX_mailbox_exit();
270 while (not simix_timers.empty()) {
271 delete simix_timers.top().second;
274 /* Free the remaining data structures */
275 simix_global->process_to_run.clear();
276 simix_global->process_that_ran.clear();
277 simix_global->process_to_destroy.clear();
278 simix_global->process_list.clear();
280 xbt_os_mutex_destroy(simix_global->mutex);
281 simix_global->mutex = nullptr;
283 xbt_dynar_free(&simix_global->actors_vector);
284 xbt_dynar_free(&simix_global->dead_actors_vector);
287 /* Let's free maestro now */
288 delete simix_global->maestro_process->context;
289 simix_global->maestro_process->context = nullptr;
290 delete simix_global->maestro_process;
291 simix_global->maestro_process = nullptr;
293 /* Finish context module and SURF */
294 SIMIX_context_mod_exit();
298 simix_global = nullptr;
304 * \brief A clock (in second).
306 * \return Return the clock.
308 double SIMIX_get_clock()
310 if(MC_is_active() || MC_record_replay_is_active()){
311 return MC_process_clock_get(SIMIX_process_self());
313 return surf_get_clock();
317 /** Wake up all processes waiting for a Surf action to finish */
318 static void SIMIX_wake_processes()
320 for (auto const& model : *all_existing_models) {
321 simgrid::kernel::resource::Action* action;
323 XBT_DEBUG("Handling the processes whose action failed (if any)");
324 while ((action = surf_model_extract_failed_action_set(model))) {
325 XBT_DEBUG(" Handling Action %p",action);
326 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
328 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
329 while ((action = surf_model_extract_done_action_set(model))) {
330 XBT_DEBUG(" Handling Action %p",action);
331 if (action->get_data() == nullptr)
332 XBT_DEBUG("probably vcpu's action %p, skip", action);
334 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->get_data()));
339 /** Handle any pending timer */
340 static bool SIMIX_execute_timers()
343 while (not simix_timers.empty() && SIMIX_get_clock() >= simix_timers.top().first) {
345 // FIXME: make the timers being real callbacks
346 // (i.e. provide dispatchers that read and expand the args)
347 smx_timer_t timer = simix_timers.top().second;
352 xbt_die("Exception thrown ouf of timer callback");
359 /** Execute all the tasks that are queued
361 * e.g. `.then()` callbacks of futures.
363 static bool SIMIX_execute_tasks()
365 xbt_assert(simix_global->tasksTemp.empty());
367 if (simix_global->tasks.empty())
372 // We don't want the callbacks to modify the vector we are iterating over:
373 swap(simix_global->tasks, simix_global->tasksTemp);
375 // Execute all the queued tasks:
376 for (auto& task : simix_global->tasksTemp)
379 simix_global->tasksTemp.clear();
380 } while (not simix_global->tasks.empty());
387 * \brief Run the main simulation loop.
391 if (not MC_record_path.empty()) {
392 simgrid::mc::replay(MC_record_path);
399 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
401 if (simgrid::simix::breakpoint >= 0.0 && time >= simgrid::simix::breakpoint) {
402 XBT_DEBUG("Breakpoint reached (%g)", simgrid::simix::breakpoint.get());
403 simgrid::simix::breakpoint = -1.0;
411 SIMIX_execute_tasks();
413 while (not simix_global->process_to_run.empty()) {
414 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
416 /* Run all processes that are ready to run, possibly in parallel */
417 SIMIX_process_runall();
419 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
421 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
423 /* Here, the order is ok because:
425 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do
426 * not impact its order.
428 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
430 * - if there is no kill during the simulation, processes remain sorted according by their PID.
431 * Rationale: This can be proved inductively.
432 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
433 * is parsed linearly).
434 * Let's show that it is still so at the end of this round.
435 * - if a process is added when being created, that's from maestro. It can be either at startup
436 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
437 * in arbitrary order (inductive hypothesis), we are fine.
438 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
439 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
440 * of a simcall terminates, we're still good. Proof:
441 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
442 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an
444 * - For communications, this function is called from SIMIX_comm_finish().
445 * This function itself don't mess with the order since simcalls are handled in FIFO order.
446 * The function is called:
447 * - before the comm starts (invalid parameters, or resource already dead or whatever).
448 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
449 * - because the communication failed or were canceled after startup. In this case, it's called from
450 * the function we are in, by the chunk:
451 * set = model->states.failed_action_set;
452 * while ((synchro = extract(set)))
453 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
454 * This order is also fixed because it depends of the order in which the surf actions were
455 * added to the system, and only maestro can add stuff this way, through simcalls.
456 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
457 * poped out of the set does not depend on the user code's execution order.
458 * - because the communication terminated. In this case, synchros are served in the order given by
459 * set = model->states.done_action_set;
460 * while ((synchro = extract(set)))
461 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
462 * and the argument is very similar to the previous one.
463 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user
464 * processes are executed.
465 * So, in any cases, the orders of processes within process_to_run do not depend on the order in which
466 * user processes were executed previously.
467 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
468 * - If there is some process killings, the order is changed by this decision that comes from user-land
469 * But this decision may not have been motivated by a situation that were different because the simulation is
471 * So, even the order change induced by the process killing is perfectly reproducible.
473 * So science works, bitches [http://xkcd.com/54/].
475 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are
476 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
478 * That would thus be a pure waste of time.
481 for (smx_actor_t const& process : simix_global->process_that_ran) {
482 if (process->simcall.call != SIMCALL_NONE) {
483 SIMIX_simcall_handle(&process->simcall, 0);
487 SIMIX_execute_tasks();
489 SIMIX_wake_processes();
490 } while (SIMIX_execute_tasks());
492 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
493 if (simix_global->process_list.size() == simix_global->daemons.size())
494 for (auto const& dmon : simix_global->daemons) {
495 XBT_DEBUG("Kill %s", dmon->get_cname());
496 SIMIX_process_kill(dmon, simix_global->maestro_process);
500 time = SIMIX_timer_next();
501 if (time > -1.0 || not simix_global->process_list.empty()) {
502 XBT_DEBUG("Calling surf_solve");
503 time = surf_solve(time);
504 XBT_DEBUG("Moving time ahead : %g", time);
507 /* Notify all the hosts that have failed */
508 /* FIXME: iterate through the list of failed host and mark each of them */
509 /* as failed. On each host, signal all the running processes with host_fail */
511 // Execute timers and tasks until there isn't anything to be done:
514 again = SIMIX_execute_timers();
515 if (SIMIX_execute_tasks())
517 SIMIX_wake_processes();
520 /* Autorestart all process */
521 for (auto const& host : host_that_restart) {
522 XBT_INFO("Restart processes on host %s", host->get_cname());
523 SIMIX_host_autorestart(host);
525 host_that_restart.clear();
527 /* Clean processes to destroy */
528 SIMIX_process_empty_trash();
530 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
531 simix_global->process_to_run.size());
533 if (simix_global->process_to_run.empty() && not simix_global->process_list.empty())
534 simgrid::simix::onDeadlock();
536 } while (time > -1.0 || not simix_global->process_to_run.empty());
538 if (not simix_global->process_list.empty()) {
542 if (simix_global->process_list.size() <= simix_global->daemons.size()) {
543 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
544 "once the simulation is over. Please fix your on_exit() functions.");
546 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
548 SIMIX_display_process_status();
549 simgrid::s4u::onDeadlock();
552 simgrid::s4u::onSimulationEnd();
556 * \brief Set the date to execute a function
558 * Set the date to execute the function on the surf.
559 * \param date Date to execute function
560 * \param callback Function to be executed
561 * \param arg Parameters of the function
564 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
566 smx_timer_t timer = new s_smx_timer_t(date, simgrid::xbt::makeTask([callback, arg]() { callback(arg); }));
567 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
571 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
573 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
574 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
578 /** @brief cancels a timer that was added earlier */
579 void SIMIX_timer_remove(smx_timer_t timer) {
580 simix_timers.erase(timer->handle_);
584 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
585 double SIMIX_timer_get_date(smx_timer_t timer) {
586 return timer ? timer->getDate() : 0;
590 * \brief Registers a function to create a process.
592 * This function registers a function to be called
593 * when a new process is created. The function has
594 * to call SIMIX_process_create().
595 * \param function create process function
597 void SIMIX_function_register_process_create(smx_creation_func_t function)
599 simix_global->create_process_function = function;
603 * \brief Registers a function to kill a process.
605 * This function registers a function to be called when a process is killed. The function has to call the
606 * SIMIX_process_kill().
608 * \param function Kill process function
610 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
612 simix_global->kill_process_function = function;
616 * \brief Registers a function to cleanup a process.
618 * This function registers a user function to be called when a process ends properly.
620 * \param function cleanup process function
622 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
624 simix_global->cleanup_process_function = function;
628 void SIMIX_display_process_status()
630 int nbprocess = simix_global->process_list.size();
632 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
633 /* List the process and their state */
634 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
635 for (auto const& kv : simix_global->process_list) {
636 smx_actor_t process = kv.second;
638 if (process->waiting_synchro) {
640 const char* synchro_description = "unknown";
642 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
643 synchro_description = "execution";
645 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
646 synchro_description = "communication";
648 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
649 synchro_description = "sleeping";
651 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
652 synchro_description = "synchronization";
654 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
655 synchro_description = "I/O";
657 XBT_INFO("Process %ld (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid,
658 process->get_cname(), process->host->get_cname(), synchro_description, process->waiting_synchro.get(),
659 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
662 XBT_INFO("Process %ld (%s@%s)", process->pid, process->get_cname(), process->host->get_cname());
667 int SIMIX_is_maestro()
669 smx_actor_t self = SIMIX_process_self();
670 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;