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. */
9 #include "src/internal_config.h"
10 #include <csignal> /* Signal handling */
13 #include <xbt/algorithm.hpp>
14 #include <xbt/functional.hpp>
16 #include "simgrid/s4u/Engine.hpp"
17 #include "simgrid/s4u/Host.hpp"
19 #include "smx_private.hpp"
20 #include "src/surf/surf_interface.hpp"
21 #include "src/surf/xml/platf.hpp"
22 #include "xbt/ex.h" /* ex_backtrace_display */
25 #include "simgrid/sg_config.h"
26 #include "src/mc/mc_replay.h"
27 #include "src/surf/StorageImpl.hpp"
29 #include "src/smpi/include/smpi_process.hpp"
31 #include "src/kernel/activity/CommImpl.hpp"
32 #include "src/kernel/activity/ExecImpl.hpp"
33 #include "src/kernel/activity/SleepImpl.hpp"
34 #include "src/kernel/activity/SynchroIo.hpp"
35 #include "src/kernel/activity/SynchroRaw.hpp"
38 #include "src/mc/mc_private.hpp"
39 #include "src/mc/remote/Client.hpp"
40 #include "src/mc/remote/mc_protocol.h"
43 #include "src/mc/mc_record.hpp"
46 #include "src/smpi/include/private.hpp"
49 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
50 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
52 std::unique_ptr<simgrid::simix::Global> simix_global;
53 static xbt_heap_t simix_timers = nullptr;
55 /** @brief Timer datatype */
58 s_smx_timer_t() = default;
61 simgrid::xbt::Task<void()> callback;
62 double getDate() { return date; }
63 s_smx_timer_t(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
66 void (*SMPI_switch_data_segment)(int) = nullptr;
68 int _sg_do_verbose_exit = 1;
69 static void inthandler(int ignored)
71 if ( _sg_do_verbose_exit ) {
72 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
73 SIMIX_display_process_status();
76 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
82 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
84 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
85 fprintf(stderr, "Access violation detected.\n"
86 "This probably comes from a programming error in your code, or from a stack\n"
87 "overflow. If you are certain of your code, try increasing the stack size\n"
88 " --cfg=contexts/stack-size=XXX (current size is %u KiB).\n"
90 "If it does not help, this may have one of the following causes:\n"
91 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
92 "Failing hardware can sometimes generate such errors too.\n"
94 "If you think you've found a bug in SimGrid, please report it along with a\n"
95 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
96 "of the fault captured with gdb or valgrind.\n",
97 smx_context_stack_size / 1024);
98 } else if (siginfo->si_signo == SIGSEGV) {
99 fprintf(stderr, "Segmentation fault.\n");
101 if (smpi_enabled() && smpi_privatize_global_variables == SMPI_PRIVATIZE_NONE) {
102 #if HAVE_PRIVATIZATION
103 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
105 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
106 #endif /* HAVE_PRIVATIZATION */
108 #endif /* HAVE_SMPI */
113 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
116 * Install signal handler for SIGSEGV. Check that nobody has already installed
117 * its own handler. For example, the Java VM does this.
119 static void install_segvhandler()
123 stack.ss_sp = sigsegv_stack;
124 stack.ss_size = sizeof sigsegv_stack;
127 if (sigaltstack(&stack, &old_stack) == -1) {
128 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
131 if (not(old_stack.ss_flags & SS_DISABLE)) {
132 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
133 old_stack.ss_size, (unsigned)old_stack.ss_flags);
134 sigaltstack(&old_stack, nullptr);
137 struct sigaction action;
138 struct sigaction old_action;
139 action.sa_sigaction = &segvhandler;
140 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
141 sigemptyset(&action.sa_mask);
143 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
144 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
147 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
148 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
149 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
150 sigaction(SIGSEGV, &old_action, nullptr);
156 /********************************* SIMIX **************************************/
157 double SIMIX_timer_next()
159 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
162 static void kill_process(smx_actor_t process)
164 SIMIX_process_kill(process, nullptr);
167 static std::function<void()> maestro_code;
172 simgrid::xbt::signal<void()> onDeadlock;
174 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
176 maestro_code = std::move(code);
182 void SIMIX_set_maestro(void (*code)(void*), void* data)
185 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
187 maestro_code = std::bind(code, data);
192 * \brief Initialize SIMIX internal data.
197 void SIMIX_global_init(int *argc, char **argv)
200 // The communication initialization is done ASAP.
201 // We need to communicate initialization of the different layers to the model-checker.
202 simgrid::mc::Client::initialize();
205 if (not simix_global) {
206 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
208 simgrid::simix::ActorImpl proc;
209 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
210 simix_global->maestro_process = nullptr;
211 simix_global->create_process_function = &SIMIX_process_create;
212 simix_global->kill_process_function = &kill_process;
213 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
214 simix_global->mutex = xbt_os_mutex_init();
216 surf_init(argc, argv); /* Initialize SURF structures */
217 SIMIX_context_mod_init();
219 // Either create a new context with maestro or create
220 // a context object with the current context mestro):
221 simgrid::simix::create_maestro(maestro_code);
223 /* Prepare to display some more info when dying on Ctrl-C pressing */
224 signal(SIGINT, inthandler);
227 install_segvhandler();
229 /* register a function to be called by SURF after the environment creation */
231 simgrid::s4u::onPlatformCreated.connect(SIMIX_post_create_environment);
232 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
233 if (host.extension<simgrid::simix::Host>() == nullptr) // another callback to the same signal may have created it
234 host.extension_set<simgrid::simix::Host>(new simgrid::simix::Host());
237 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::StorageImpl* storage) {
238 sg_storage_t s = simgrid::s4u::Storage::byName(storage->getCname());
239 xbt_assert(s != nullptr, "Storage not found for name %s", storage->getCname());
243 if (not simix_timers)
244 simix_timers = xbt_heap_new(8, [](void* p) {
245 delete static_cast<smx_timer_t>(p);
248 if (xbt_cfg_get_boolean("clean-atexit"))
251 if (_sg_cfg_exit_asap)
258 * \brief Clean the SIMIX simulation
260 * This functions remove the memory used by SIMIX
265 return; // to avoid double cleaning by java and C
268 if (SIMIX_process_count()>0){
269 if(smpi_process()->initialized()){
270 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
272 XBT_WARN("Process called exit when leaving - Skipping cleanups");
279 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
280 if (not simix_global->process_to_run.empty() && SIMIX_get_clock() <= 0.0) {
282 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
283 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
284 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
286 /* Kill all processes (but maestro) */
287 SIMIX_process_killall(simix_global->maestro_process, 1);
288 SIMIX_context_runall();
289 SIMIX_process_empty_trash();
291 /* Exit the SIMIX network module */
292 SIMIX_mailbox_exit();
294 xbt_heap_free(simix_timers);
295 simix_timers = nullptr;
296 /* Free the remaining data structures */
297 simix_global->process_to_run.clear();
298 simix_global->process_that_ran.clear();
299 xbt_swag_free(simix_global->process_to_destroy);
300 simix_global->process_list.clear();
301 simix_global->process_to_destroy = nullptr;
303 xbt_os_mutex_destroy(simix_global->mutex);
304 simix_global->mutex = nullptr;
306 xbt_dynar_free(&simix_global->actors_vector);
309 /* Let's free maestro now */
310 delete simix_global->maestro_process->context;
311 simix_global->maestro_process->context = nullptr;
312 delete simix_global->maestro_process;
313 simix_global->maestro_process = nullptr;
315 /* Finish context module and SURF */
316 SIMIX_context_mod_exit();
320 simix_global = nullptr;
326 * \brief A clock (in second).
328 * \return Return the clock.
330 double SIMIX_get_clock()
332 if(MC_is_active() || MC_record_replay_is_active()){
333 return MC_process_clock_get(SIMIX_process_self());
335 return surf_get_clock();
339 /** Wake up all processes waiting for a Surf action to finish */
340 static void SIMIX_wake_processes()
342 surf_action_t action;
344 for (auto const& model : *all_existing_models) {
345 XBT_DEBUG("Handling the processes whose action failed (if any)");
346 while ((action = surf_model_extract_failed_action_set(model))) {
347 XBT_DEBUG(" Handling Action %p",action);
348 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->getData()));
350 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
351 while ((action = surf_model_extract_done_action_set(model))) {
352 XBT_DEBUG(" Handling Action %p",action);
353 if (action->getData() == nullptr)
354 XBT_DEBUG("probably vcpu's action %p, skip", action);
356 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->getData()));
361 /** Handle any pending timer */
362 static bool SIMIX_execute_timers()
365 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
367 //FIXME: make the timers being real callbacks
368 // (i.e. provide dispatchers that read and expand the args)
369 smx_timer_t timer = (smx_timer_t) xbt_heap_pop(simix_timers);
374 xbt_die("Exception throwed ouf of timer callback");
381 /** Execute all the tasks that are queued
383 * e.g. `.then()` callbacks of futures.
385 static bool SIMIX_execute_tasks()
387 xbt_assert(simix_global->tasksTemp.empty());
389 if (simix_global->tasks.empty())
394 // We don't want the callbacks to modify the vector we are iterating over:
395 swap(simix_global->tasks, simix_global->tasksTemp);
397 // Execute all the queued tasks:
398 for (auto& task : simix_global->tasksTemp)
401 simix_global->tasksTemp.clear();
402 } while (not simix_global->tasks.empty());
409 * \brief Run the main simulation loop.
413 if (MC_record_path) {
414 simgrid::mc::replay(MC_record_path);
421 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
423 SIMIX_execute_tasks();
425 while (not simix_global->process_to_run.empty()) {
426 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
428 /* Run all processes that are ready to run, possibly in parallel */
429 SIMIX_process_runall();
431 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
433 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
435 /* Here, the order is ok because:
437 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
439 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
441 * - if there is no kill during the simulation, processes remain sorted according by their PID.
442 * rational: This can be proved inductively.
443 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
444 * Let's show that it is still so at the end of this round.
445 * - if a process is added when being created, that's from maestro. It can be either at startup
446 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
447 * in arbitrary order (inductive hypothesis), we are fine.
448 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
449 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
450 * of a simcall terminates, we're still good. Proof:
451 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
452 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
453 * - For communications, this function is called from SIMIX_comm_finish().
454 * This function itself don't mess with the order since simcalls are handled in FIFO order.
455 * The function is called:
456 * - before the comm starts (invalid parameters, or resource already dead or whatever).
457 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
458 * - because the communication failed or were canceled after startup. In this case, it's called from the function
459 * we are in, by the chunk:
460 * set = model->states.failed_action_set;
461 * while ((synchro = xbt_swag_extract(set)))
462 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
463 * This order is also fixed because it depends of the order in which the surf actions were
464 * added to the system, and only maestro can add stuff this way, through simcalls.
465 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
466 * poped out of the swag does not depend on the user code's execution order.
467 * - because the communication terminated. In this case, synchros are served in the order given by
468 * set = model->states.done_action_set;
469 * while ((synchro = xbt_swag_extract(set)))
470 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
471 * and the argument is very similar to the previous one.
472 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
473 * So, in any cases, the orders of processes within process_to_run do not depend on the order in which user processes were executed previously.
474 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
475 * - If there is some process killings, the order is changed by this decision that comes from user-land
476 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
477 * So, even the order change induced by the process killing is perfectly reproducible.
479 * So science works, bitches [http://xkcd.com/54/].
481 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
482 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
483 * That would thus be a pure waste of time.
486 for (smx_actor_t const& process : simix_global->process_that_ran) {
487 if (process->simcall.call != SIMCALL_NONE) {
488 SIMIX_simcall_handle(&process->simcall, 0);
492 SIMIX_execute_tasks();
494 SIMIX_wake_processes();
495 } while (SIMIX_execute_tasks());
497 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
498 if (simix_global->process_list.size() == simix_global->daemons.size())
499 for (auto const& dmon : simix_global->daemons) {
500 XBT_DEBUG("Kill %s", dmon->cname());
501 SIMIX_process_kill(dmon, simix_global->maestro_process);
505 time = SIMIX_timer_next();
506 if (time > -1.0 || not simix_global->process_list.empty()) {
507 XBT_DEBUG("Calling surf_solve");
508 time = surf_solve(time);
509 XBT_DEBUG("Moving time ahead : %g", time);
512 /* Notify all the hosts that have failed */
513 /* FIXME: iterate through the list of failed host and mark each of them */
514 /* as failed. On each host, signal all the running processes with host_fail */
516 // Execute timers and tasks until there isn't anything to be done:
519 again = SIMIX_execute_timers();
520 if (SIMIX_execute_tasks())
522 SIMIX_wake_processes();
525 /* Autorestart all process */
526 for (auto const& host : host_that_restart) {
527 XBT_INFO("Restart processes on host %s", host->getCname());
528 SIMIX_host_autorestart(host);
530 host_that_restart.clear();
532 /* Clean processes to destroy */
533 SIMIX_process_empty_trash();
535 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
536 simix_global->process_to_run.size());
538 if (simix_global->process_to_run.empty() && not simix_global->process_list.empty())
539 simgrid::simix::onDeadlock();
541 } while (time > -1.0 || not simix_global->process_to_run.empty());
543 if (not simix_global->process_list.empty()) {
547 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
548 SIMIX_display_process_status();
551 simgrid::s4u::onSimulationEnd();
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, [callback, arg]() { callback(arg); });
566 xbt_heap_push(simix_timers, timer, date);
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 xbt_heap_push(simix_timers, timer, date);
577 /** @brief cancels a timer that was added earlier */
578 void SIMIX_timer_remove(smx_timer_t timer) {
579 delete static_cast<smx_timer_t>(xbt_heap_rm_elm(simix_timers, timer, timer->getDate()));
582 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
583 double SIMIX_timer_get_date(smx_timer_t timer) {
584 return timer ? timer->getDate() : 0;
588 * \brief Registers a function to create a process.
590 * This function registers a function to be called
591 * when a new process is created. The function has
592 * to call SIMIX_process_create().
593 * \param function create process function
595 void SIMIX_function_register_process_create(smx_creation_func_t function)
597 simix_global->create_process_function = function;
601 * \brief Registers a function to kill a process.
603 * This function registers a function to be called when a process is killed. The function has to call the
604 * SIMIX_process_kill().
606 * \param function Kill process function
608 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
610 simix_global->kill_process_function = function;
614 * \brief Registers a function to cleanup a process.
616 * This function registers a user function to be called when a process ends properly.
618 * \param function cleanup process function
620 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
622 simix_global->cleanup_process_function = function;
626 void SIMIX_display_process_status()
628 int nbprocess = simix_global->process_list.size();
630 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
631 /* List the process and their state */
632 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
633 for (auto const& kv : simix_global->process_list) {
634 smx_actor_t process = kv.second;
636 if (process->waiting_synchro) {
638 const char* synchro_description = "unknown";
640 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
641 synchro_description = "execution";
643 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
644 synchro_description = "communication";
646 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
647 synchro_description = "sleeping";
649 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
650 synchro_description = "synchronization";
652 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
653 synchro_description = "I/O";
655 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid,
656 process->cname(), process->host->getCname(), synchro_description, process->waiting_synchro.get(),
657 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
660 XBT_INFO("Process %lu (%s@%s)", process->pid, process->cname(), process->host->getCname());
665 int SIMIX_is_maestro()
667 smx_actor_t self = SIMIX_process_self();
668 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;