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 "src/surf/surf_interface.hpp"
20 #include "src/surf/xml/platf.hpp"
21 #include "smx_private.h"
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.h"
39 #include "src/mc/remote/Client.hpp"
40 #include "src/mc/remote/mc_protocol.h"
43 #include "src/mc/mc_record.h"
46 #include "src/smpi/include/private.h"
47 #include "src/smpi/include/private.hpp"
50 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
51 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
53 std::unique_ptr<simgrid::simix::Global> simix_global;
54 static xbt_heap_t simix_timers = nullptr;
56 /** @brief Timer datatype */
57 typedef class s_smx_timer {
59 s_smx_timer() = default;
62 simgrid::xbt::Task<void()> callback;
63 double getDate() { return date; }
64 s_smx_timer(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
67 void (*SMPI_switch_data_segment)(int) = nullptr;
69 int _sg_do_verbose_exit = 1;
70 static void inthandler(int ignored)
72 if ( _sg_do_verbose_exit ) {
73 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
74 SIMIX_display_process_status();
77 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
83 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
85 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
86 fprintf(stderr, "Access violation detected.\n"
87 "This probably comes from a programming error in your code, or from a stack\n"
88 "overflow. If you are certain of your code, try increasing the stack size\n"
89 " --cfg=contexts/stack-size=XXX (current size is %d KiB).\n"
91 "If it does not help, this may have one of the following causes:\n"
92 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
93 "Failing hardware can sometimes generate such errors too.\n"
95 "If you think you've found a bug in SimGrid, please report it along with a\n"
96 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
97 "of the fault captured with gdb or valgrind.\n",
98 smx_context_stack_size / 1024);
99 } else if (siginfo->si_signo == SIGSEGV) {
100 fprintf(stderr, "Segmentation fault.\n");
102 if (smpi_enabled() && smpi_privatize_global_variables == SMPI_PRIVATIZE_NONE) {
103 #if HAVE_PRIVATIZATION
104 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
106 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
107 #endif /* HAVE_PRIVATIZATION */
109 #endif /* HAVE_SMPI */
114 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
117 * Install signal handler for SIGSEGV. Check that nobody has already installed
118 * its own handler. For example, the Java VM does this.
120 static void install_segvhandler()
124 stack.ss_sp = sigsegv_stack;
125 stack.ss_size = sizeof sigsegv_stack;
128 if (sigaltstack(&stack, &old_stack) == -1) {
129 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
132 if (not(old_stack.ss_flags & SS_DISABLE)) {
133 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
134 old_stack.ss_size, (unsigned)old_stack.ss_flags);
135 sigaltstack(&old_stack, nullptr);
138 struct sigaction action;
139 struct sigaction old_action;
140 action.sa_sigaction = &segvhandler;
141 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
142 sigemptyset(&action.sa_mask);
144 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
145 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
148 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
149 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
150 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
151 sigaction(SIGSEGV, &old_action, nullptr);
157 /********************************* SIMIX **************************************/
158 double SIMIX_timer_next()
160 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
163 static void kill_process(smx_actor_t process)
165 SIMIX_process_kill(process, nullptr);
168 static std::function<void()> maestro_code;
173 simgrid::xbt::signal<void()> onDeadlock;
175 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
177 maestro_code = std::move(code);
183 void SIMIX_set_maestro(void (*code)(void*), void* data)
186 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
188 maestro_code = std::bind(code, data);
193 * \brief Initialize SIMIX internal data.
198 void SIMIX_global_init(int *argc, char **argv)
201 // The communication initialization is done ASAP.
202 // We need to communicate initialization of the different layers to the model-checker.
203 simgrid::mc::Client::initialize();
206 if (not simix_global) {
207 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
209 simgrid::simix::ActorImpl proc;
210 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
211 simix_global->maestro_process = nullptr;
212 simix_global->create_process_function = &SIMIX_process_create;
213 simix_global->kill_process_function = &kill_process;
214 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
215 simix_global->mutex = xbt_os_mutex_init();
217 surf_init(argc, argv); /* Initialize SURF structures */
218 SIMIX_context_mod_init();
220 // Either create a new context with maestro or create
221 // a context object with the current context mestro):
222 simgrid::simix::create_maestro(maestro_code);
224 /* Prepare to display some more info when dying on Ctrl-C pressing */
225 signal(SIGINT, inthandler);
228 install_segvhandler();
230 /* register a function to be called by SURF after the environment creation */
232 simgrid::s4u::onPlatformCreated.connect(SIMIX_post_create_environment);
233 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
234 if (host.extension<simgrid::simix::Host>() == nullptr) // another callback to the same signal may have created it
235 host.extension_set<simgrid::simix::Host>(new simgrid::simix::Host());
238 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::StorageImpl* storage) {
239 sg_storage_t s = simgrid::s4u::Storage::byName(storage->cname());
240 xbt_assert(s != nullptr, "Storage not found for name %s", storage->cname());
244 if (not simix_timers)
245 simix_timers = xbt_heap_new(8, [](void* p) {
246 delete static_cast<smx_timer_t>(p);
249 if (xbt_cfg_get_boolean("clean-atexit"))
252 if (_sg_cfg_exit_asap)
259 * \brief Clean the SIMIX simulation
261 * This functions remove the memory used by SIMIX
266 return; // to avoid double cleaning by java and C
269 if (SIMIX_process_count()>0){
270 if(smpi_process()->initialized()){
271 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
273 XBT_WARN("Process called exit when leaving - Skipping cleanups");
280 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
281 if (not simix_global->process_to_run.empty() && SIMIX_get_clock() <= 0.0) {
283 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
284 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
285 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
287 /* Kill all processes (but maestro) */
288 SIMIX_process_killall(simix_global->maestro_process, 1);
290 /* Exit the SIMIX network module */
291 SIMIX_mailbox_exit();
293 xbt_heap_free(simix_timers);
294 simix_timers = nullptr;
295 /* Free the remaining data structures */
296 simix_global->process_to_run.clear();
297 simix_global->process_that_ran.clear();
298 xbt_swag_free(simix_global->process_to_destroy);
299 simix_global->process_list.clear();
300 simix_global->process_to_destroy = nullptr;
302 xbt_os_mutex_destroy(simix_global->mutex);
303 simix_global->mutex = nullptr;
305 xbt_dynar_free(&simix_global->actors_vector);
308 /* Let's free maestro now */
309 delete simix_global->maestro_process->context;
310 simix_global->maestro_process->context = nullptr;
311 delete simix_global->maestro_process;
312 simix_global->maestro_process = nullptr;
314 /* Finish context module and SURF */
315 SIMIX_context_mod_exit();
319 simix_global = nullptr;
325 * \brief A clock (in second).
327 * \return Return the clock.
329 double SIMIX_get_clock()
331 if(MC_is_active() || MC_record_replay_is_active()){
332 return MC_process_clock_get(SIMIX_process_self());
334 return surf_get_clock();
338 /** Wake up all processes waiting for a Surf action to finish */
339 static void SIMIX_wake_processes()
341 surf_action_t action;
343 for (auto const& model : *all_existing_models) {
344 XBT_DEBUG("Handling the processes whose action failed (if any)");
345 while ((action = surf_model_extract_failed_action_set(model))) {
346 XBT_DEBUG(" Handling Action %p",action);
347 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->getData()));
349 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
350 while ((action = surf_model_extract_done_action_set(model))) {
351 XBT_DEBUG(" Handling Action %p",action);
352 if (action->getData() == nullptr)
353 XBT_DEBUG("probably vcpu's action %p, skip", action);
355 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->getData()));
360 /** Handle any pending timer */
361 static bool SIMIX_execute_timers()
364 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
366 //FIXME: make the timers being real callbacks
367 // (i.e. provide dispatchers that read and expand the args)
368 smx_timer_t timer = (smx_timer_t) xbt_heap_pop(simix_timers);
373 xbt_die("Exception throwed ouf of timer callback");
380 /** Execute all the tasks that are queued
382 * e.g. `.then()` callbacks of futures.
384 static bool SIMIX_execute_tasks()
386 xbt_assert(simix_global->tasksTemp.empty());
388 if (simix_global->tasks.empty())
393 // We don't want the callbacks to modify the vector we are iterating over:
394 swap(simix_global->tasks, simix_global->tasksTemp);
396 // Execute all the queued tasks:
397 for (auto& task : simix_global->tasksTemp)
400 simix_global->tasksTemp.clear();
401 } while (not simix_global->tasks.empty());
408 * \brief Run the main simulation loop.
412 if (MC_record_path) {
413 simgrid::mc::replay(MC_record_path);
420 XBT_DEBUG("New Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
422 SIMIX_execute_tasks();
424 while (not simix_global->process_to_run.empty()) {
425 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", simix_global->process_to_run.size());
427 /* Run all processes that are ready to run, possibly in parallel */
428 SIMIX_process_runall();
430 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
432 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
434 /* Here, the order is ok because:
436 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
438 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
440 * - if there is no kill during the simulation, processes remain sorted according by their PID.
441 * rational: This can be proved inductively.
442 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
443 * Let's show that it is still so at the end of this round.
444 * - if a process is added when being created, that's from maestro. It can be either at startup
445 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
446 * in arbitrary order (inductive hypothesis), we are fine.
447 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
448 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
449 * of a simcall terminates, we're still good. Proof:
450 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
451 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
452 * - For communications, this function is called from SIMIX_comm_finish().
453 * This function itself don't mess with the order since simcalls are handled in FIFO order.
454 * The function is called:
455 * - before the comm starts (invalid parameters, or resource already dead or whatever).
456 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
457 * - because the communication failed or were canceled after startup. In this case, it's called from the function
458 * we are in, by the chunk:
459 * set = model->states.failed_action_set;
460 * while ((synchro = xbt_swag_extract(set)))
461 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
462 * This order is also fixed because it depends of the order in which the surf actions were
463 * added to the system, and only maestro can add stuff this way, through simcalls.
464 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
465 * poped out of the swag does not depend on the user code's execution order.
466 * - because the communication terminated. In this case, synchros are served in the order given by
467 * set = model->states.done_action_set;
468 * while ((synchro = xbt_swag_extract(set)))
469 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
470 * and the argument is very similar to the previous one.
471 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
472 * 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.
473 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
474 * - If there is some process killings, the order is changed by this decision that comes from user-land
475 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
476 * So, even the order change induced by the process killing is perfectly reproducible.
478 * So science works, bitches [http://xkcd.com/54/].
480 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
481 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
482 * That would thus be a pure waste of time.
485 for (smx_actor_t const& process : simix_global->process_that_ran) {
486 if (process->simcall.call != SIMCALL_NONE) {
487 SIMIX_simcall_handle(&process->simcall, 0);
491 SIMIX_execute_tasks();
493 SIMIX_wake_processes();
494 } while (SIMIX_execute_tasks());
496 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
497 if (simix_global->process_list.size() == simix_global->daemons.size())
498 for (auto const& dmon : simix_global->daemons) {
499 XBT_DEBUG("Kill %s", dmon->cname());
500 SIMIX_process_kill(dmon, simix_global->maestro_process);
504 time = SIMIX_timer_next();
505 if (time > -1.0 || simix_global->process_list.empty() == false) {
506 XBT_DEBUG("Calling surf_solve");
507 time = surf_solve(time);
508 XBT_DEBUG("Moving time ahead : %g", time);
511 /* Notify all the hosts that have failed */
512 /* FIXME: iterate through the list of failed host and mark each of them */
513 /* as failed. On each host, signal all the running processes with host_fail */
515 // Execute timers and tasks until there isn't anything to be done:
518 again = SIMIX_execute_timers();
519 if (SIMIX_execute_tasks())
521 SIMIX_wake_processes();
524 /* Autorestart all process */
525 for (auto const& host : host_that_restart) {
526 XBT_INFO("Restart processes on host %s", host->getCname());
527 SIMIX_host_autorestart(host);
529 host_that_restart.clear();
531 /* Clean processes to destroy */
532 SIMIX_process_empty_trash();
534 XBT_DEBUG("### time %f, #processes %zu, #to_run %zu", time, simix_global->process_list.size(),
535 simix_global->process_to_run.size());
537 if (simix_global->process_to_run.empty() && not simix_global->process_list.empty())
538 simgrid::simix::onDeadlock();
540 } while (time > -1.0 || not simix_global->process_to_run.empty());
542 if (simix_global->process_list.size() != 0) {
546 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
547 SIMIX_display_process_status();
550 simgrid::s4u::onSimulationEnd();
554 * \brief Set the date to execute a function
556 * Set the date to execute the function on the surf.
557 * \param date Date to execute function
558 * \param callback Function to be executed
559 * \param arg Parameters of the function
562 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
564 smx_timer_t timer = new s_smx_timer_t(date, [callback, arg]() { callback(arg); });
565 xbt_heap_push(simix_timers, timer, date);
569 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
571 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
572 xbt_heap_push(simix_timers, timer, date);
576 /** @brief cancels a timer that was added earlier */
577 void SIMIX_timer_remove(smx_timer_t timer) {
578 xbt_heap_rm_elm(simix_timers, timer, timer->getDate());
581 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
582 double SIMIX_timer_get_date(smx_timer_t timer) {
583 return timer ? timer->getDate() : 0;
587 * \brief Registers a function to create a process.
589 * This function registers a function to be called
590 * when a new process is created. The function has
591 * to call SIMIX_process_create().
592 * \param function create process function
594 void SIMIX_function_register_process_create(smx_creation_func_t function)
596 simix_global->create_process_function = function;
600 * \brief Registers a function to kill a process.
602 * This function registers a function to be called when a process is killed. The function has to call the
603 * SIMIX_process_kill().
605 * \param function Kill process function
607 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
609 simix_global->kill_process_function = function;
613 * \brief Registers a function to cleanup a process.
615 * This function registers a user function to be called when a process ends properly.
617 * \param function cleanup process function
619 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
621 simix_global->cleanup_process_function = function;
625 void SIMIX_display_process_status()
627 int nbprocess = simix_global->process_list.size();
629 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
630 /* List the process and their state */
631 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
632 for (auto const& kv : simix_global->process_list) {
633 smx_actor_t process = kv.second;
635 if (process->waiting_synchro) {
637 const char* synchro_description = "unknown";
639 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
640 synchro_description = "execution";
642 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
643 synchro_description = "communication";
645 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
646 synchro_description = "sleeping";
648 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
649 synchro_description = "synchronization";
651 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
652 synchro_description = "I/O";
654 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid,
655 process->cname(), process->host->getCname(), synchro_description, process->waiting_synchro.get(),
656 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
659 XBT_INFO("Process %lu (%s@%s)", process->pid, process->cname(), process->host->getCname());
664 int SIMIX_is_maestro()
666 smx_actor_t self = SIMIX_process_self();
667 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;