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 <signal.h> /* Signal handling */
11 #include "src/internal_config.h"
13 #include <xbt/functional.hpp>
15 #include "simgrid/s4u/Engine.hpp"
16 #include "simgrid/s4u/Host.hpp"
18 #include "src/surf/surf_interface.hpp"
19 #include "src/surf/xml/platf.hpp"
20 #include "smx_private.h"
21 #include "xbt/ex.h" /* ex_backtrace_display */
24 #include "simgrid/sg_config.h"
25 #include "src/mc/mc_replay.h"
26 #include "src/surf/StorageImpl.hpp"
28 #include "src/smpi/include/smpi_process.hpp"
30 #include "src/kernel/activity/CommImpl.hpp"
31 #include "src/kernel/activity/ExecImpl.hpp"
32 #include "src/kernel/activity/SleepImpl.hpp"
33 #include "src/kernel/activity/SynchroIo.hpp"
34 #include "src/kernel/activity/SynchroRaw.hpp"
37 #include "src/mc/mc_private.h"
38 #include "src/mc/remote/Client.hpp"
39 #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 struct s_smx_timer {
59 simgrid::xbt::Task<void()> callback;
61 s_smx_timer()=default;
62 s_smx_timer(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
65 void (*SMPI_switch_data_segment)(int) = nullptr;
67 int _sg_do_verbose_exit = 1;
68 static void inthandler(int ignored)
70 if ( _sg_do_verbose_exit ) {
71 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
72 SIMIX_display_process_status();
75 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
81 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
83 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
84 fprintf(stderr, "Access violation detected.\n"
85 "This probably comes from a programming error in your code, or from a stack\n"
86 "overflow. If you are certain of your code, try increasing the stack size\n"
87 " --cfg=contexts/stack-size=XXX (current size is %d KiB).\n"
89 "If it does not help, this may have one of the following causes:\n"
90 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
91 "Failing hardware can sometimes generate such errors too.\n"
93 "If you think you've found a bug in SimGrid, please report it along with a\n"
94 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
95 "of the fault captured with gdb or valgrind.\n",
96 smx_context_stack_size / 1024);
97 } else if (siginfo->si_signo == SIGSEGV) {
98 fprintf(stderr, "Segmentation fault.\n");
100 if (smpi_enabled() && smpi_privatize_global_variables == SMPI_PRIVATIZE_NONE) {
101 #if HAVE_PRIVATIZATION
102 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
104 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
105 #endif /* HAVE_PRIVATIZATION */
107 #endif /* HAVE_SMPI */
112 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
115 * Install signal handler for SIGSEGV. Check that nobody has already installed
116 * its own handler. For example, the Java VM does this.
118 static void install_segvhandler()
122 stack.ss_sp = sigsegv_stack;
123 stack.ss_size = sizeof sigsegv_stack;
126 if (sigaltstack(&stack, &old_stack) == -1) {
127 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
130 if (not(old_stack.ss_flags & SS_DISABLE)) {
131 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
132 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
133 sigaltstack(&old_stack, nullptr);
136 struct sigaction action;
137 struct sigaction old_action;
138 action.sa_sigaction = &segvhandler;
139 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
140 sigemptyset(&action.sa_mask);
142 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
143 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
146 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
147 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
148 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
149 sigaction(SIGSEGV, &old_action, nullptr);
155 /********************************* SIMIX **************************************/
156 double SIMIX_timer_next()
158 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
161 static void kill_process(smx_actor_t process)
163 SIMIX_process_kill(process, nullptr);
166 static std::function<void()> maestro_code;
171 simgrid::xbt::signal<void()> onDeadlock;
173 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
175 maestro_code = std::move(code);
181 void SIMIX_set_maestro(void (*code)(void*), void* data)
184 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
186 maestro_code = std::bind(code, data);
191 * \brief Initialize SIMIX internal data.
196 void SIMIX_global_init(int *argc, char **argv)
199 // The communication initialization is done ASAP.
200 // We need to communicate initialization of the different layers to the model-checker.
201 simgrid::mc::Client::initialize();
204 if (not simix_global) {
205 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
207 simgrid::simix::ActorImpl proc;
208 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_actor_t), nullptr);
209 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_actor_t), nullptr);
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 xbt_dynar_is_empty(simix_global->process_to_run) && 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 xbt_dynar_free(&simix_global->process_to_run);
297 xbt_dynar_free(&simix_global->process_that_ran);
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 static int process_syscall_color(void *p)
340 switch ((*(smx_actor_t *)p)->simcall.call) {
342 case SIMCALL_PROCESS_KILL:
344 // case SIMCALL_PROCESS_RESUME:
351 /** Wake up all processes waiting for a Surf action to finish */
352 static void SIMIX_wake_processes()
354 surf_action_t action;
356 for(auto model : *all_existing_models) {
357 XBT_DEBUG("Handling the processes whose action failed (if any)");
358 while ((action = surf_model_extract_failed_action_set(model))) {
359 XBT_DEBUG(" Handling Action %p",action);
360 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->getData()));
362 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
363 while ((action = surf_model_extract_done_action_set(model))) {
364 XBT_DEBUG(" Handling Action %p",action);
365 if (action->getData() == nullptr)
366 XBT_DEBUG("probably vcpu's action %p, skip", action);
368 SIMIX_simcall_exit(static_cast<simgrid::kernel::activity::ActivityImpl*>(action->getData()));
373 /** Handle any pending timer */
374 static bool SIMIX_execute_timers()
377 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
379 //FIXME: make the timers being real callbacks
380 // (i.e. provide dispatchers that read and expand the args)
381 smx_timer_t timer = (smx_timer_t) xbt_heap_pop(simix_timers);
386 xbt_die("Exception throwed ouf of timer callback");
393 /** Execute all the tasks that are queued
395 * e.g. `.then()` callbacks of futures.
397 static bool SIMIX_execute_tasks()
399 xbt_assert(simix_global->tasksTemp.empty());
401 if (simix_global->tasks.empty())
406 // We don't want the callbacks to modify the vector we are iterating over:
407 swap(simix_global->tasks, simix_global->tasksTemp);
409 // Execute all the queued tasks:
410 for (auto& task : simix_global->tasksTemp)
413 simix_global->tasksTemp.clear();
414 } while (not simix_global->tasks.empty());
421 * \brief Run the main simulation loop.
425 if (MC_record_path) {
426 simgrid::mc::replay(MC_record_path);
433 XBT_DEBUG("New Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
435 SIMIX_execute_tasks();
437 while (not xbt_dynar_is_empty(simix_global->process_to_run)) {
438 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
440 /* Run all processes that are ready to run, possibly in parallel */
441 SIMIX_process_runall();
443 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
444 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
446 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
448 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
450 /* Here, the order is ok because:
452 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
454 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
456 * - if there is no kill during the simulation, processes remain sorted according by their PID.
457 * rational: This can be proved inductively.
458 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
459 * Let's show that it is still so at the end of this round.
460 * - if a process is added when being created, that's from maestro. It can be either at startup
461 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
462 * in arbitrary order (inductive hypothesis), we are fine.
463 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
464 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
465 * of a simcall terminates, we're still good. Proof:
466 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
467 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
468 * - For communications, this function is called from SIMIX_comm_finish().
469 * This function itself don't mess with the order since simcalls are handled in FIFO order.
470 * The function is called:
471 * - before the comm starts (invalid parameters, or resource already dead or whatever).
472 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
473 * - because the communication failed or were canceled after startup. In this case, it's called from the function
474 * we are in, by the chunk:
475 * set = model->states.failed_action_set;
476 * while ((synchro = xbt_swag_extract(set)))
477 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
478 * This order is also fixed because it depends of the order in which the surf actions were
479 * added to the system, and only maestro can add stuff this way, through simcalls.
480 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
481 * poped out of the swag does not depend on the user code's execution order.
482 * - because the communication terminated. In this case, synchros are served in the order given by
483 * set = model->states.done_action_set;
484 * while ((synchro = xbt_swag_extract(set)))
485 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
486 * and the argument is very similar to the previous one.
487 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
488 * 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.
489 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
490 * - If there is some process killings, the order is changed by this decision that comes from user-land
491 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
492 * So, even the order change induced by the process killing is perfectly reproducible.
494 * So science works, bitches [http://xkcd.com/54/].
496 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
497 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
498 * That would thus be a pure waste of time.
503 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
504 if (process->simcall.call != SIMCALL_NONE) {
505 SIMIX_simcall_handle(&process->simcall, 0);
509 SIMIX_execute_tasks();
511 SIMIX_wake_processes();
512 } while (SIMIX_execute_tasks());
514 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
515 if (simix_global->process_list.size() == simix_global->daemons.size())
516 for (const auto& dmon : simix_global->daemons) {
517 XBT_DEBUG("Kill %s", dmon->cname());
518 SIMIX_process_kill(dmon, simix_global->maestro_process);
522 time = SIMIX_timer_next();
523 if (time > -1.0 || simix_global->process_list.empty() == false) {
524 XBT_DEBUG("Calling surf_solve");
525 time = surf_solve(time);
526 XBT_DEBUG("Moving time ahead : %g", time);
529 /* Notify all the hosts that have failed */
530 /* FIXME: iterate through the list of failed host and mark each of them */
531 /* as failed. On each host, signal all the running processes with host_fail */
533 // Execute timers and tasks until there isn't anything to be done:
536 again = SIMIX_execute_timers();
537 if (SIMIX_execute_tasks())
539 SIMIX_wake_processes();
542 /* Autorestart all process */
543 for (auto host: host_that_restart) {
544 XBT_INFO("Restart processes on host %s", host->getCname());
545 SIMIX_host_autorestart(host);
547 host_that_restart.clear();
549 /* Clean processes to destroy */
550 SIMIX_process_empty_trash();
552 XBT_DEBUG("### time %f, #processes %zu, #to_run %lu", time, simix_global->process_list.size(),
553 xbt_dynar_length(simix_global->process_to_run));
555 if (xbt_dynar_is_empty(simix_global->process_to_run) && not simix_global->process_list.empty())
556 simgrid::simix::onDeadlock();
558 } while (time > -1.0 || not xbt_dynar_is_empty(simix_global->process_to_run));
560 if (simix_global->process_list.size() != 0) {
564 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
565 SIMIX_display_process_status();
568 simgrid::s4u::onSimulationEnd();
572 * \brief Set the date to execute a function
574 * Set the date to execute the function on the surf.
575 * \param date Date to execute function
576 * \param callback Function to be executed
577 * \param arg Parameters of the function
580 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
582 smx_timer_t timer = new s_smx_timer_t(date, [callback, arg]() { callback(arg); });
583 xbt_heap_push(simix_timers, timer, date);
587 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
589 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
590 xbt_heap_push(simix_timers, timer, date);
594 /** @brief cancels a timer that was added earlier */
595 void SIMIX_timer_remove(smx_timer_t timer) {
596 xbt_heap_rm_elm(simix_timers, timer, timer->date);
599 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
600 double SIMIX_timer_get_date(smx_timer_t timer) {
601 return timer?timer->date:0;
605 * \brief Registers a function to create a process.
607 * This function registers a function to be called
608 * when a new process is created. The function has
609 * to call SIMIX_process_create().
610 * \param function create process function
612 void SIMIX_function_register_process_create(smx_creation_func_t function)
614 simix_global->create_process_function = function;
618 * \brief Registers a function to kill a process.
620 * This function registers a function to be called when a process is killed. The function has to call the
621 * SIMIX_process_kill().
623 * \param function Kill process function
625 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
627 simix_global->kill_process_function = function;
631 * \brief Registers a function to cleanup a process.
633 * This function registers a user function to be called when a process ends properly.
635 * \param function cleanup process function
637 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
639 simix_global->cleanup_process_function = function;
643 void SIMIX_display_process_status()
645 int nbprocess = simix_global->process_list.size();
647 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
648 /* List the process and their state */
649 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
650 for (auto kv : simix_global->process_list) {
651 smx_actor_t process = kv.second;
653 if (process->waiting_synchro) {
655 const char* synchro_description = "unknown";
657 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::ExecImpl>(process->waiting_synchro) != nullptr)
658 synchro_description = "execution";
660 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(process->waiting_synchro) != nullptr)
661 synchro_description = "communication";
663 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::SleepImpl>(process->waiting_synchro) != nullptr)
664 synchro_description = "sleeping";
666 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::RawImpl>(process->waiting_synchro) != nullptr)
667 synchro_description = "synchronization";
669 if (boost::dynamic_pointer_cast<simgrid::kernel::activity::IoImpl>(process->waiting_synchro) != nullptr)
670 synchro_description = "I/O";
672 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid,
673 process->cname(), process->host->getCname(), synchro_description, process->waiting_synchro.get(),
674 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
677 XBT_INFO("Process %lu (%s@%s)", process->pid, process->cname(), process->host->getCname());
682 int SIMIX_is_maestro()
684 smx_actor_t self = SIMIX_process_self();
685 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;