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/storage_interface.hpp"
20 #include "src/surf/xml/platf.hpp"
21 #include "smx_private.h"
23 #include "xbt/ex.h" /* ex_backtrace_display */
25 #include "src/mc/mc_replay.h"
26 #include "simgrid/sg_config.h"
28 #include "src/kernel/activity/SynchroExec.hpp"
29 #include "src/kernel/activity/SynchroComm.hpp"
30 #include "src/kernel/activity/SynchroSleep.hpp"
31 #include "src/kernel/activity/SynchroIo.hpp"
32 #include "src/kernel/activity/SynchroRaw.hpp"
35 #include "src/mc/mc_private.h"
36 #include "src/mc/remote/Client.hpp"
37 #include "src/mc/remote/mc_protocol.h"
41 #include "src/mc/mc_record.h"
44 #include "src/smpi/private.h"
47 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
48 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
50 std::unique_ptr<simgrid::simix::Global> simix_global;
51 static xbt_heap_t simix_timers = nullptr;
53 /** @brief Timer datatype */
54 typedef struct s_smx_timer {
56 simgrid::xbt::Task<void()> callback;
58 s_smx_timer()=default;
59 s_smx_timer(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
62 void (*SMPI_switch_data_segment)(int) = nullptr;
64 int _sg_do_verbose_exit = 1;
65 static void inthandler(int ignored)
67 if ( _sg_do_verbose_exit ) {
68 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
69 SIMIX_display_process_status();
72 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
78 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
80 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
81 fprintf(stderr, "Access violation detected.\n"
82 "This probably comes from a programming error in your code, or from a stack\n"
83 "overflow. If you are certain of your code, try increasing the stack size\n"
84 " --cfg=contexts/stack-size=XXX (current size is %d KiB).\n"
86 "If it does not help, this may have one of the following causes:\n"
87 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
88 "Failing hardware can sometimes generate such errors too.\n"
90 "If you think you've found a bug in SimGrid, please report it along with a\n"
91 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
92 "of the fault captured with gdb or valgrind.\n",
93 smx_context_stack_size / 1024);
94 } else if (siginfo->si_signo == SIGSEGV) {
95 fprintf(stderr, "Segmentation fault.\n");
97 if (smpi_enabled() && !smpi_privatize_global_variables) {
98 #if HAVE_PRIVATIZATION
99 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
101 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatize-global-variables:yes (yet).\n");
102 #endif /* HAVE_PRIVATIZATION */
104 #endif /* HAVE_SMPI */
109 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
112 * Install signal handler for SIGSEGV. Check that nobody has already installed
113 * its own handler. For example, the Java VM does this.
115 static void install_segvhandler()
119 stack.ss_sp = sigsegv_stack;
120 stack.ss_size = sizeof sigsegv_stack;
123 if (sigaltstack(&stack, &old_stack) == -1) {
124 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
127 if (!(old_stack.ss_flags & SS_DISABLE)) {
128 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
129 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
130 sigaltstack(&old_stack, nullptr);
133 struct sigaction action;
134 struct sigaction old_action;
135 action.sa_sigaction = &segvhandler;
136 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
137 sigemptyset(&action.sa_mask);
139 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
140 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
143 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
144 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
145 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
146 sigaction(SIGSEGV, &old_action, nullptr);
152 /********************************* SIMIX **************************************/
153 double SIMIX_timer_next()
155 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
158 static void kill_process(smx_actor_t process)
160 SIMIX_process_kill(process, nullptr);
163 static std::function<void()> maestro_code;
168 simgrid::xbt::signal<void()> onDeadlock;
170 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
172 maestro_code = std::move(code);
178 void SIMIX_set_maestro(void (*code)(void*), void* data)
181 XBT_INFO("WARNING, SIMIX_set_maestro is believed to not work on windows. Please help us investigating this issue if you need that feature");
183 maestro_code = std::bind(code, data);
188 * \brief Initialize SIMIX internal data.
193 void SIMIX_global_init(int *argc, char **argv)
196 // The communication initialization is done ASAP.
197 // We need to communicate initialization of the different layers to the model-checker.
198 simgrid::mc::Client::initialize();
202 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
204 simgrid::simix::ActorImpl proc;
205 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_actor_t), nullptr);
206 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_actor_t), nullptr);
207 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
208 simix_global->maestro_process = nullptr;
209 simix_global->create_process_function = &SIMIX_process_create;
210 simix_global->kill_process_function = &kill_process;
211 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
212 simix_global->mutex = xbt_os_mutex_init();
214 surf_init(argc, argv); /* Initialize SURF structures */
215 SIMIX_context_mod_init();
217 // Either create a new context with maestro or create
218 // a context object with the current context mestro):
219 simgrid::simix::create_maestro(maestro_code);
221 /* Prepare to display some more info when dying on Ctrl-C pressing */
222 signal(SIGINT, inthandler);
225 install_segvhandler();
227 /* register a function to be called by SURF after the environment creation */
229 simgrid::s4u::onPlatformCreated.connect(SIMIX_post_create_environment);
230 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
231 host.extension_set<simgrid::simix::Host>(new simgrid::simix::Host());
234 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
235 const char* name = storage->cname();
236 // TODO, create sg_storage_by_name
237 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, name);
238 xbt_assert(s != nullptr, "Storage not found for name %s", name);
240 SIMIX_storage_create(name, s, nullptr);
243 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
246 simix_timers = xbt_heap_new(8, [](void* p) {
247 delete static_cast<smx_timer_t>(p);
250 if (xbt_cfg_get_boolean("clean-atexit"))
253 if (_sg_cfg_exit_asap)
260 * \brief Clean the SIMIX simulation
262 * This functions remove the memory used by SIMIX
267 return; // to avoid double cleaning by java and C
270 if (SIMIX_process_count()>0){
271 if(smpi_process()->initialized()){
272 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
274 XBT_WARN("Process called exit when leaving - Skipping cleanups");
281 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
282 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() <= 0.0) {
284 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
285 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
286 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
288 /* Kill all processes (but maestro) */
289 SIMIX_process_killall(simix_global->maestro_process, 1);
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 xbt_dynar_free(&simix_global->process_to_run);
298 xbt_dynar_free(&simix_global->process_that_ran);
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 /* Let's free maestro now */
307 delete simix_global->maestro_process->context;
308 simix_global->maestro_process->context = nullptr;
309 delete simix_global->maestro_process;
310 simix_global->maestro_process = nullptr;
312 /* Finish context module and SURF */
313 SIMIX_context_mod_exit();
317 simix_global = nullptr;
323 * \brief A clock (in second).
325 * \return Return the clock.
327 double SIMIX_get_clock()
329 if(MC_is_active() || MC_record_replay_is_active()){
330 return MC_process_clock_get(SIMIX_process_self());
332 return surf_get_clock();
336 static int process_syscall_color(void *p)
338 switch ((*(smx_actor_t *)p)->simcall.call) {
340 case SIMCALL_PROCESS_KILL:
342 case SIMCALL_PROCESS_RESUME:
349 /** Wake up all processes waiting for a Surf action to finish */
350 static void SIMIX_wake_processes()
352 surf_action_t action;
354 for(auto model : *all_existing_models) {
355 XBT_DEBUG("Handling the processes whose action failed (if any)");
356 while ((action = surf_model_extract_failed_action_set(model))) {
357 XBT_DEBUG(" Handling Action %p",action);
358 SIMIX_simcall_exit((smx_activity_t) action->getData());
360 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
361 while ((action = surf_model_extract_done_action_set(model))) {
362 XBT_DEBUG(" Handling Action %p",action);
363 if (action->getData() == nullptr)
364 XBT_DEBUG("probably vcpu's action %p, skip", action);
366 SIMIX_simcall_exit((smx_activity_t) action->getData());
371 /** Handle any pending timer */
372 static bool SIMIX_execute_timers()
375 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
377 //FIXME: make the timers being real callbacks
378 // (i.e. provide dispatchers that read and expand the args)
379 smx_timer_t timer = (smx_timer_t) xbt_heap_pop(simix_timers);
384 xbt_die("Exception throwed ouf of timer callback");
391 /** Execute all the tasks that are queued
393 * e.g. `.then()` callbacks of futures.
395 static bool SIMIX_execute_tasks()
397 xbt_assert(simix_global->tasksTemp.empty());
399 if (simix_global->tasks.empty())
404 // We don't want the callbacks to modify the vector we are iterating over:
405 swap(simix_global->tasks, simix_global->tasksTemp);
407 // Execute all the queued tasks:
408 for (auto& task : simix_global->tasksTemp)
411 simix_global->tasksTemp.clear();
412 } while (!simix_global->tasks.empty());
419 * \brief Run the main simulation loop.
423 if (MC_record_path) {
424 simgrid::mc::replay(MC_record_path);
431 XBT_DEBUG("New Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
433 SIMIX_execute_tasks();
435 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
436 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
438 /* Run all processes that are ready to run, possibly in parallel */
439 SIMIX_process_runall();
441 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
442 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
444 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
446 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
448 /* Here, the order is ok because:
450 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
452 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
454 * - if there is no kill during the simulation, processes remain sorted according by their PID.
455 * rational: This can be proved inductively.
456 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
457 * Let's show that it is still so at the end of this round.
458 * - if a process is added when being created, that's from maestro. It can be either at startup
459 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
460 * in arbitrary order (inductive hypothesis), we are fine.
461 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
462 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
463 * of a simcall terminates, we're still good. Proof:
464 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
465 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
466 * - For communications, this function is called from SIMIX_comm_finish().
467 * This function itself don't mess with the order since simcalls are handled in FIFO order.
468 * The function is called:
469 * - before the comm starts (invalid parameters, or resource already dead or whatever).
470 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
471 * - because the communication failed or were canceled after startup. In this case, it's called from the function
472 * we are in, by the chunk:
473 * set = model->states.failed_action_set;
474 * while ((synchro = xbt_swag_extract(set)))
475 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
476 * This order is also fixed because it depends of the order in which the surf actions were
477 * added to the system, and only maestro can add stuff this way, through simcalls.
478 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
479 * poped out of the swag does not depend on the user code's execution order.
480 * - because the communication terminated. In this case, synchros are served in the order given by
481 * set = model->states.done_action_set;
482 * while ((synchro = xbt_swag_extract(set)))
483 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
484 * and the argument is very similar to the previous one.
485 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
486 * 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.
487 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
488 * - If there is some process killings, the order is changed by this decision that comes from user-land
489 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
490 * So, even the order change induced by the process killing is perfectly reproducible.
492 * So science works, bitches [http://xkcd.com/54/].
494 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
495 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
496 * That would thus be a pure waste of time.
501 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
502 if (process->simcall.call != SIMCALL_NONE) {
503 SIMIX_simcall_handle(&process->simcall, 0);
507 SIMIX_execute_tasks();
509 SIMIX_wake_processes();
510 } while (SIMIX_execute_tasks());
514 time = SIMIX_timer_next();
515 if (time > -1.0 || simix_global->process_list.empty() == false) {
516 XBT_DEBUG("Calling surf_solve");
517 time = surf_solve(time);
518 XBT_DEBUG("Moving time ahead : %g", time);
521 /* Notify all the hosts that have failed */
522 /* FIXME: iterate through the list of failed host and mark each of them */
523 /* as failed. On each host, signal all the running processes with host_fail */
525 // Execute timers and tasks until there isn't anything to be done:
528 again = SIMIX_execute_timers();
529 if (SIMIX_execute_tasks())
531 SIMIX_wake_processes();
534 /* Autorestart all process */
535 for (auto host: host_that_restart) {
536 XBT_INFO("Restart processes on host %s", host->cname());
537 SIMIX_host_autorestart(host);
539 host_that_restart.clear();
541 /* Clean processes to destroy */
542 SIMIX_process_empty_trash();
544 XBT_DEBUG("### time %f, #processes %zu, #to_run %lu", time, simix_global->process_list.size(),
545 xbt_dynar_length(simix_global->process_to_run));
547 /* If only daemon processes remain, cancel their actions, mark them to die and reschedule them */
548 if (simix_global->process_list.size() == simix_global->daemons.size())
549 for (const auto& dmon : simix_global->daemons) {
550 XBT_DEBUG("Kill %s", dmon->cname());
551 SIMIX_process_kill(dmon, simix_global->maestro_process);
554 if (xbt_dynar_is_empty(simix_global->process_to_run) &&
555 !simix_global->process_list.empty())
556 simgrid::simix::onDeadlock();
558 } while (time > -1.0 || !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); });
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 (dynamic_cast<simgrid::kernel::activity::Exec*>(process->waiting_synchro) != nullptr)
658 synchro_description = "execution";
660 if (dynamic_cast<simgrid::kernel::activity::Comm*>(process->waiting_synchro) != nullptr)
661 synchro_description = "communication";
663 if (dynamic_cast<simgrid::kernel::activity::Sleep*>(process->waiting_synchro) != nullptr)
664 synchro_description = "sleeping";
666 if (dynamic_cast<simgrid::kernel::activity::Raw*>(process->waiting_synchro) != nullptr)
667 synchro_description = "synchronization";
669 if (dynamic_cast<simgrid::kernel::activity::Io*>(process->waiting_synchro) != nullptr)
670 synchro_description = "I/O";
674 switch (process->waiting_synchro->type) {
675 case SIMIX_SYNC_PARALLEL_EXECUTE:
676 synchro_description = "parallel execution";
679 case SIMIX_SYNC_JOIN:
680 synchro_description = "joining";
684 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid,
685 process->cname(), process->host->cname(), synchro_description, process->waiting_synchro,
686 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
689 XBT_INFO("Process %lu (%s@%s)", process->pid, process->cname(), process->host->cname());
694 int SIMIX_is_maestro()
696 smx_actor_t self = SIMIX_process_self();
697 return simix_global == nullptr /*SimDag*/ || self == nullptr || self == simix_global->maestro_process;