1 /* Copyright (c) 2007-2017. 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. */
6 #include <boost/heap/fibonacci_heap.hpp>
10 #include "src/internal_config.h"
11 #include <csignal> /* Signal handling */
14 #include <xbt/algorithm.hpp>
15 #include <xbt/functional.hpp>
16 #include <xbt/utility.hpp>
18 #include "simgrid/s4u/Engine.hpp"
19 #include "simgrid/s4u/Host.hpp"
21 #include "smx_private.hpp"
22 #include "src/surf/surf_interface.hpp"
23 #include "src/surf/xml/platf.hpp"
24 #include "xbt/ex.h" /* ex_backtrace_display */
27 #include "simgrid/sg_config.h"
28 #include "src/mc/mc_replay.hpp"
29 #include "src/surf/StorageImpl.hpp"
31 #include "src/smpi/include/smpi_process.hpp"
33 #include "src/kernel/activity/CommImpl.hpp"
34 #include "src/kernel/activity/ExecImpl.hpp"
35 #include "src/kernel/activity/SleepImpl.hpp"
36 #include "src/kernel/activity/SynchroIo.hpp"
37 #include "src/kernel/activity/SynchroRaw.hpp"
40 #include "src/mc/mc_private.hpp"
41 #include "src/mc/remote/Client.hpp"
42 #include "src/mc/remote/mc_protocol.h"
45 #include "src/mc/mc_record.hpp"
48 #include "src/smpi/include/private.hpp"
51 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
52 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
54 std::unique_ptr<simgrid::simix::Global> simix_global;
57 typedef std::pair<double, smx_timer_t> TimerQelt;
58 boost::heap::fibonacci_heap<TimerQelt, boost::heap::compare<simgrid::xbt::HeapComparator<TimerQelt>>> simix_timers;
61 /** @brief Timer datatype */
66 decltype(simix_timers)::handle_type handle_;
67 simgrid::xbt::Task<void()> callback;
68 double getDate() { return date; }
69 s_smx_timer_t(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
72 void (*SMPI_switch_data_segment)(int) = nullptr;
74 int _sg_do_verbose_exit = 1;
75 static void inthandler(int)
77 if ( _sg_do_verbose_exit ) {
78 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
79 SIMIX_display_process_status();
82 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
88 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
90 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
91 fprintf(stderr, "Access violation detected.\n"
92 "This probably comes from a programming error in your code, or from a stack\n"
93 "overflow. If you are certain of your code, try increasing the stack size\n"
94 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
96 "If it does not help, this may have one of the following causes:\n"
97 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
98 "Failing hardware can sometimes generate such errors too.\n"
100 "If you think you've found a bug in SimGrid, please report it along with a\n"
101 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
102 "of the fault captured with gdb or valgrind.\n",
103 smx_context_stack_size / 1024);
104 } else if (siginfo->si_signo == SIGSEGV) {
105 fprintf(stderr, "Segmentation fault.\n");
107 if (smpi_enabled() && smpi_privatize_global_variables == SMPI_PRIVATIZE_NONE) {
108 #if HAVE_PRIVATIZATION
109 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
111 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
112 #endif /* HAVE_PRIVATIZATION */
114 #endif /* HAVE_SMPI */
119 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
122 * Install signal handler for SIGSEGV. Check that nobody has already installed
123 * its own handler. For example, the Java VM does this.
125 static void install_segvhandler()
129 stack.ss_sp = sigsegv_stack;
130 stack.ss_size = sizeof sigsegv_stack;
133 if (sigaltstack(&stack, &old_stack) == -1) {
134 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
137 if (not(old_stack.ss_flags & SS_DISABLE)) {
138 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
139 old_stack.ss_size, (unsigned)old_stack.ss_flags);
140 sigaltstack(&old_stack, nullptr);
143 struct sigaction action;
144 struct sigaction old_action;
145 action.sa_sigaction = &segvhandler;
146 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
147 sigemptyset(&action.sa_mask);
149 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
150 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
153 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
154 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
155 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
156 sigaction(SIGSEGV, &old_action, nullptr);
162 /********************************* SIMIX **************************************/
163 double SIMIX_timer_next()
165 return simix_timers.empty() ? -1.0 : simix_timers.top().first;
168 static void kill_process(smx_actor_t process)
170 SIMIX_process_kill(process, nullptr);
173 static std::function<void()> maestro_code;
178 simgrid::xbt::signal<void()> onDeadlock;
180 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
182 maestro_code = std::move(code);
188 void SIMIX_set_maestro(void (*code)(void*), void* data)
191 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
193 maestro_code = std::bind(code, data);
198 * \brief Initialize SIMIX internal data.
203 void SIMIX_global_init(int *argc, char **argv)
206 // The communication initialization is done ASAP.
207 // We need to communicate initialization of the different layers to the model-checker.
208 simgrid::mc::Client::initialize();
211 if (not simix_global) {
212 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
213 simix_global->maestro_process = nullptr;
214 simix_global->create_process_function = &SIMIX_process_create;
215 simix_global->kill_process_function = &kill_process;
216 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
217 simix_global->mutex = xbt_os_mutex_init();
219 surf_init(argc, argv); /* Initialize SURF structures */
220 SIMIX_context_mod_init();
222 // Either create a new context with maestro or create
223 // a context object with the current context mestro):
224 simgrid::simix::create_maestro(maestro_code);
226 /* Prepare to display some more info when dying on Ctrl-C pressing */
227 signal(SIGINT, inthandler);
230 install_segvhandler();
232 /* register a function to be called by SURF after the environment creation */
234 simgrid::s4u::onPlatformCreated.connect(SIMIX_post_create_environment);
235 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
236 if (host.extension<simgrid::simix::Host>() == nullptr) // another callback to the same signal may have created it
237 host.extension_set<simgrid::simix::Host>(new simgrid::simix::Host());
240 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::StorageImpl* storage) {
241 sg_storage_t s = simgrid::s4u::Storage::byName(storage->getCname());
242 xbt_assert(s != nullptr, "Storage not found for name %s", storage->getCname());
246 if (xbt_cfg_get_boolean("clean-atexit"))
249 if (_sg_cfg_exit_asap)
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->process_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_process_killall(simix_global->maestro_process, 1);
287 SIMIX_context_runall();
288 SIMIX_process_empty_trash();
290 /* Exit the SIMIX network module */
291 SIMIX_mailbox_exit();
293 while (not simix_timers.empty()) {
294 delete simix_timers.top().second;
297 /* Free the remaining data structures */
298 simix_global->process_to_run.clear();
299 simix_global->process_that_ran.clear();
300 simix_global->process_to_destroy.clear();
301 simix_global->process_list.clear();
303 xbt_os_mutex_destroy(simix_global->mutex);
304 simix_global->mutex = nullptr;
306 xbt_dynar_free(&simix_global->actors_vector);
307 xbt_dynar_free(&simix_global->dead_actors_vector);
310 /* Let's free maestro now */
311 delete simix_global->maestro_process->context;
312 simix_global->maestro_process->context = nullptr;
313 delete simix_global->maestro_process;
314 simix_global->maestro_process = nullptr;
316 /* Finish context module and SURF */
317 SIMIX_context_mod_exit();
321 simix_global = nullptr;
327 * \brief A clock (in second).
329 * \return Return the clock.
331 double SIMIX_get_clock()
333 if(MC_is_active() || MC_record_replay_is_active()){
334 return MC_process_clock_get(SIMIX_process_self());
336 return surf_get_clock();
340 /** Wake up all processes waiting for a Surf action to finish */
341 static void SIMIX_wake_processes()
343 surf_action_t action;
345 for (auto const& model : *all_existing_models) {
346 XBT_DEBUG("Handling the processes whose action failed (if any)");
347 while ((action = surf_model_extract_failed_action_set(model))) {
348 XBT_DEBUG(" Handling Action %p",action);
349 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->getData()));
351 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
352 while ((action = surf_model_extract_done_action_set(model))) {
353 XBT_DEBUG(" Handling Action %p",action);
354 if (action->getData() == nullptr)
355 XBT_DEBUG("probably vcpu's action %p, skip", action);
357 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->getData()));
362 /** Handle any pending timer */
363 static bool SIMIX_execute_timers()
366 while (not simix_timers.empty() && SIMIX_get_clock() >= simix_timers.top().first) {
368 // FIXME: make the timers being real callbacks
369 // (i.e. provide dispatchers that read and expand the args)
370 smx_timer_t timer = simix_timers.top().second;
375 xbt_die("Exception thrown ouf of timer callback");
382 /** Execute all the tasks that are queued
384 * e.g. `.then()` callbacks of futures.
386 static bool SIMIX_execute_tasks()
388 xbt_assert(simix_global->tasksTemp.empty());
390 if (simix_global->tasks.empty())
395 // We don't want the callbacks to modify the vector we are iterating over:
396 swap(simix_global->tasks, simix_global->tasksTemp);
398 // Execute all the queued tasks:
399 for (auto& task : simix_global->tasksTemp)
402 simix_global->tasksTemp.clear();
403 } while (not simix_global->tasks.empty());
410 * \brief Run the main simulation loop.
414 if (not MC_record_path.empty()) {
415 simgrid::mc::replay(MC_record_path);
422 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
424 SIMIX_execute_tasks();
426 while (not simix_global->process_to_run.empty()) {
427 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
429 /* Run all processes that are ready to run, possibly in parallel */
430 SIMIX_process_runall();
432 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
434 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
436 /* Here, the order is ok because:
438 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do
439 * not impact its order.
441 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
443 * - if there is no kill during the simulation, processes remain sorted according by their PID.
444 * rational: This can be proved inductively.
445 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
446 * is parsed linearly).
447 * Let's show that it is still so at the end of this round.
448 * - if a process is added when being created, that's from maestro. It can be either at startup
449 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
450 * in arbitrary order (inductive hypothesis), we are fine.
451 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
452 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
453 * of a simcall terminates, we're still good. Proof:
454 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
455 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an
457 * - For communications, this function is called from SIMIX_comm_finish().
458 * This function itself don't mess with the order since simcalls are handled in FIFO order.
459 * The function is called:
460 * - before the comm starts (invalid parameters, or resource already dead or whatever).
461 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
462 * - because the communication failed or were canceled after startup. In this case, it's called from
463 * the function we are in, by the chunk:
464 * set = model->states.failed_action_set;
465 * while ((synchro = extract(set)))
466 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
467 * This order is also fixed because it depends of the order in which the surf actions were
468 * added to the system, and only maestro can add stuff this way, through simcalls.
469 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
470 * poped out of the set does not depend on the user code's execution order.
471 * - because the communication terminated. In this case, synchros are served in the order given by
472 * set = model->states.done_action_set;
473 * while ((synchro = extract(set)))
474 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
475 * and the argument is very similar to the previous one.
476 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user
477 * processes are executed.
478 * So, in any cases, the orders of processes within process_to_run do not depend on the order in which
479 * user processes were executed previously.
480 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
481 * - If there is some process killings, the order is changed by this decision that comes from user-land
482 * But this decision may not have been motivated by a situation that were different because the simulation is
484 * So, even the order change induced by the process killing is perfectly reproducible.
486 * So science works, bitches [http://xkcd.com/54/].
488 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are
489 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
491 * That would thus be a pure waste of time.
494 for (smx_actor_t const& process : simix_global->process_that_ran) {
495 if (process->simcall.call != SIMCALL_NONE) {
496 SIMIX_simcall_handle(&process->simcall, 0);
500 SIMIX_execute_tasks();
502 SIMIX_wake_processes();
503 } while (SIMIX_execute_tasks());
505 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
506 if (simix_global->process_list.size() == simix_global->daemons.size())
507 for (auto const& dmon : simix_global->daemons) {
508 XBT_DEBUG("Kill %s", dmon->getCname());
509 SIMIX_process_kill(dmon, simix_global->maestro_process);
513 time = SIMIX_timer_next();
514 if (time > -1.0 || not simix_global->process_list.empty()) {
515 XBT_DEBUG("Calling surf_solve");
516 time = surf_solve(time);
517 XBT_DEBUG("Moving time ahead : %g", time);
520 /* Notify all the hosts that have failed */
521 /* FIXME: iterate through the list of failed host and mark each of them */
522 /* as failed. On each host, signal all the running processes with host_fail */
524 // Execute timers and tasks until there isn't anything to be done:
527 again = SIMIX_execute_timers();
528 if (SIMIX_execute_tasks())
530 SIMIX_wake_processes();
533 /* Autorestart all process */
534 for (auto const& host : host_that_restart) {
535 XBT_INFO("Restart processes on host %s", host->getCname());
536 SIMIX_host_autorestart(host);
538 host_that_restart.clear();
540 /* Clean processes to destroy */
541 SIMIX_process_empty_trash();
543 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
544 simix_global->process_to_run.size());
546 if (simix_global->process_to_run.empty() && not simix_global->process_list.empty())
547 simgrid::simix::onDeadlock();
549 } while (time > -1.0 || not simix_global->process_to_run.empty());
551 if (not simix_global->process_list.empty()) {
555 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
556 SIMIX_display_process_status();
559 simgrid::s4u::onSimulationEnd();
563 * \brief Set the date to execute a function
565 * Set the date to execute the function on the surf.
566 * \param date Date to execute function
567 * \param callback Function to be executed
568 * \param arg Parameters of the function
571 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
573 smx_timer_t timer = new s_smx_timer_t(date, [callback, arg]() { callback(arg); });
574 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
578 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
580 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
581 timer->handle_ = simix_timers.emplace(std::make_pair(date, timer));
585 /** @brief cancels a timer that was added earlier */
586 void SIMIX_timer_remove(smx_timer_t timer) {
587 simix_timers.erase(timer->handle_);
591 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
592 double SIMIX_timer_get_date(smx_timer_t timer) {
593 return timer ? timer->getDate() : 0;
597 * \brief Registers a function to create a process.
599 * This function registers a function to be called
600 * when a new process is created. The function has
601 * to call SIMIX_process_create().
602 * \param function create process function
604 void SIMIX_function_register_process_create(smx_creation_func_t function)
606 simix_global->create_process_function = function;
610 * \brief Registers a function to kill a process.
612 * This function registers a function to be called when a process is killed. The function has to call the
613 * SIMIX_process_kill().
615 * \param function Kill process function
617 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
619 simix_global->kill_process_function = function;
623 * \brief Registers a function to cleanup a process.
625 * This function registers a user function to be called when a process ends properly.
627 * \param function cleanup process function
629 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
631 simix_global->cleanup_process_function = function;
635 void SIMIX_display_process_status()
637 int nbprocess = simix_global->process_list.size();
639 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
640 /* List the process and their state */
641 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
642 for (auto const& kv : simix_global->process_list) {
643 smx_actor_t process = kv.second;
645 if (process->waiting_synchro) {
647 const char* synchro_description = "unknown";
649 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
650 synchro_description = "execution";
652 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
653 synchro_description = "communication";
655 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
656 synchro_description = "sleeping";
658 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
659 synchro_description = "synchronization";
661 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
662 synchro_description = "I/O";
664 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid,
665 process->getCname(), process->host->getCname(), synchro_description, process->waiting_synchro.get(),
666 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
669 XBT_INFO("Process %lu (%s@%s)", process->pid, process->getCname(), process->host->getCname());
674 int SIMIX_is_maestro()
676 smx_actor_t self = SIMIX_process_self();
677 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;