1 /* Copyright (c) 2007-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
10 #include <signal.h> /* Signal handling */
12 #include "src/internal_config.h"
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/storage_interface.hpp"
21 #include "src/surf/xml/platf.hpp"
22 #include "smx_private.h"
24 #include "xbt/ex.h" /* ex_backtrace_display */
26 #include "src/mc/mc_replay.h"
27 #include "simgrid/sg_config.h"
29 #include "src/kernel/activity/SynchroExec.hpp"
30 #include "src/kernel/activity/SynchroComm.hpp"
31 #include "src/kernel/activity/SynchroSleep.hpp"
32 #include "src/kernel/activity/SynchroIo.hpp"
33 #include "src/kernel/activity/SynchroRaw.hpp"
36 #include "src/mc/mc_private.h"
37 #include "src/mc/remote/Client.hpp"
38 #include "src/mc/remote/mc_protocol.h"
42 #include "src/mc/mc_record.h"
45 #include "src/smpi/private.h"
48 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
49 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
51 std::unique_ptr<simgrid::simix::Global> simix_global;
52 static xbt_heap_t simix_timers = nullptr;
54 /** @brief Timer datatype */
55 typedef struct s_smx_timer {
57 simgrid::xbt::Task<void()> callback;
59 s_smx_timer()=default;
60 s_smx_timer(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
63 void (*SMPI_switch_data_segment)(int) = nullptr;
65 int _sg_do_verbose_exit = 1;
66 static void inthandler(int ignored)
68 if ( _sg_do_verbose_exit ) {
69 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
70 SIMIX_display_process_status();
73 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
79 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
81 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
82 fprintf(stderr, "Access violation detected.\n"
83 "This probably comes from a programming error in your code, or from a stack\n"
84 "overflow. If you are certain of your code, try increasing the stack size\n"
85 " --cfg=contexts/stack-size=XXX (current size is %d KiB).\n"
87 "If it does not help, this may have one of the following causes:\n"
88 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
89 "Failing hardware can sometimes generate such errors too.\n"
91 "If you think you've found a bug in SimGrid, please report it along with a\n"
92 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
93 "of the fault captured with gdb or valgrind.\n",
94 smx_context_stack_size / 1024);
95 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
96 fprintf(stderr, "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
97 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
99 } else if (siginfo->si_signo == SIGSEGV) {
100 fprintf(stderr, "Segmentation fault.\n");
102 if (smpi_enabled() && !smpi_privatize_global_variables) {
103 #if HAVE_PRIVATIZATION
104 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
106 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatize-global-variables: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 (!(old_stack.ss_flags & SS_DISABLE)) {
133 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
134 old_stack.ss_sp, old_stack.ss_size, 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();
207 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
209 simgrid::simix::ActorImpl proc;
210 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_actor_t), nullptr);
211 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_actor_t), nullptr);
212 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
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 host.extension_set<simgrid::simix::Host>(new simgrid::simix::Host());
239 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
240 const char* name = storage->cname();
241 // TODO, create sg_storage_by_name
242 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, name);
243 xbt_assert(s != nullptr, "Storage not found for name %s", name);
245 SIMIX_storage_create(name, s, nullptr);
248 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
251 simix_timers = xbt_heap_new(8, [](void* p) {
252 delete static_cast<smx_timer_t>(p);
255 if (xbt_cfg_get_boolean("clean-atexit"))
258 if (_sg_cfg_exit_asap)
265 * \brief Clean the SIMIX simulation
267 * This functions remove the memory used by SIMIX
272 return; // to avoid double cleaning by java and C
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");
286 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
287 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() <= 0.0) {
289 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
290 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
291 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
293 /* Kill all processes (but maestro) */
294 SIMIX_process_killall(simix_global->maestro_process, 1);
296 /* Exit the SIMIX network module */
297 SIMIX_mailbox_exit();
299 xbt_heap_free(simix_timers);
300 simix_timers = nullptr;
301 /* Free the remaining data structures */
302 xbt_dynar_free(&simix_global->process_to_run);
303 xbt_dynar_free(&simix_global->process_that_ran);
304 xbt_swag_free(simix_global->process_to_destroy);
305 simix_global->process_list.clear();
306 simix_global->process_to_destroy = nullptr;
308 xbt_os_mutex_destroy(simix_global->mutex);
309 simix_global->mutex = nullptr;
311 /* Let's free maestro now */
312 delete simix_global->maestro_process->context;
313 simix_global->maestro_process->context = nullptr;
314 delete simix_global->maestro_process;
315 simix_global->maestro_process = nullptr;
317 /* Finish context module and SURF */
318 SIMIX_context_mod_exit();
322 simix_global = nullptr;
328 * \brief A clock (in second).
330 * \return Return the clock.
332 double SIMIX_get_clock()
334 if(MC_is_active() || MC_record_replay_is_active()){
335 return MC_process_clock_get(SIMIX_process_self());
337 return surf_get_clock();
341 static int process_syscall_color(void *p)
343 switch ((*(smx_actor_t *)p)->simcall.call) {
345 case SIMCALL_PROCESS_KILL:
347 case SIMCALL_PROCESS_RESUME:
354 /** Wake up all processes waiting for a Surf action to finish */
355 static void SIMIX_wake_processes()
357 surf_action_t action;
359 for(auto model : *all_existing_models) {
360 XBT_DEBUG("Handling the processes whose action failed (if any)");
361 while ((action = surf_model_extract_failed_action_set(model))) {
362 XBT_DEBUG(" Handling Action %p",action);
363 SIMIX_simcall_exit((smx_activity_t) action->getData());
365 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
366 while ((action = surf_model_extract_done_action_set(model))) {
367 XBT_DEBUG(" Handling Action %p",action);
368 if (action->getData() == nullptr)
369 XBT_DEBUG("probably vcpu's action %p, skip", action);
371 SIMIX_simcall_exit((smx_activity_t) action->getData());
376 /** Handle any pending timer */
377 static bool SIMIX_execute_timers()
380 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
382 //FIXME: make the timers being real callbacks
383 // (i.e. provide dispatchers that read and expand the args)
384 smx_timer_t timer = (smx_timer_t) xbt_heap_pop(simix_timers);
389 xbt_die("Exception throwed ouf of timer callback");
396 /** Execute all the tasks that are queued
398 * e.g. `.then()` callbacks of futures.
400 static bool SIMIX_execute_tasks()
402 xbt_assert(simix_global->tasksTemp.empty());
404 if (simix_global->tasks.empty())
409 // We don't want the callbacks to modify the vector we are iterating over:
410 swap(simix_global->tasks, simix_global->tasksTemp);
412 // Execute all the queued tasks:
413 for (auto& task : simix_global->tasksTemp)
416 simix_global->tasksTemp.clear();
417 } while (!simix_global->tasks.empty());
424 * \brief Run the main simulation loop.
428 if (MC_record_path) {
429 simgrid::mc::replay(MC_record_path);
436 XBT_DEBUG("New Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
438 SIMIX_execute_tasks();
440 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
441 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
443 /* Run all processes that are ready to run, possibly in parallel */
444 SIMIX_process_runall();
446 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
447 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
449 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
451 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
453 /* Here, the order is ok because:
455 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
457 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
459 * - if there is no kill during the simulation, processes remain sorted according by their PID.
460 * rational: This can be proved inductively.
461 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
462 * Let's show that it is still so at the end of this round.
463 * - if a process is added when being created, that's from maestro. It can be either at startup
464 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
465 * in arbitrary order (inductive hypothesis), we are fine.
466 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
467 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
468 * of a simcall terminates, we're still good. Proof:
469 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
470 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
471 * - For communications, this function is called from SIMIX_comm_finish().
472 * This function itself don't mess with the order since simcalls are handled in FIFO order.
473 * The function is called:
474 * - before the comm starts (invalid parameters, or resource already dead or whatever).
475 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
476 * - because the communication failed or were canceled after startup. In this case, it's called from the function
477 * we are in, by the chunk:
478 * set = model->states.failed_action_set;
479 * while ((synchro = xbt_swag_extract(set)))
480 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
481 * This order is also fixed because it depends of the order in which the surf actions were
482 * added to the system, and only maestro can add stuff this way, through simcalls.
483 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
484 * poped out of the swag does not depend on the user code's execution order.
485 * - because the communication terminated. In this case, synchros are served in the order given by
486 * set = model->states.done_action_set;
487 * while ((synchro = xbt_swag_extract(set)))
488 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
489 * and the argument is very similar to the previous one.
490 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
491 * 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.
492 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
493 * - If there is some process killings, the order is changed by this decision that comes from user-land
494 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
495 * So, even the order change induced by the process killing is perfectly reproducible.
497 * So science works, bitches [http://xkcd.com/54/].
499 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
500 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
501 * That would thus be a pure waste of time.
506 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
507 if (process->simcall.call != SIMCALL_NONE) {
508 SIMIX_simcall_handle(&process->simcall, 0);
512 SIMIX_execute_tasks();
514 SIMIX_wake_processes();
515 } while (SIMIX_execute_tasks());
519 time = SIMIX_timer_next();
520 if (time > -1.0 || simix_global->process_list.empty() == false) {
521 XBT_DEBUG("Calling surf_solve");
522 time = surf_solve(time);
523 XBT_DEBUG("Moving time ahead : %g", time);
526 /* Notify all the hosts that have failed */
527 /* FIXME: iterate through the list of failed host and mark each of them */
528 /* as failed. On each host, signal all the running processes with host_fail */
530 // Execute timers and tasks until there isn't anything to be done:
533 again = SIMIX_execute_timers();
534 if (SIMIX_execute_tasks())
536 SIMIX_wake_processes();
539 /* Autorestart all process */
540 for (auto host: host_that_restart) {
541 XBT_INFO("Restart processes on host %s", host->cname());
542 SIMIX_host_autorestart(host);
544 host_that_restart.clear();
546 /* Clean processes to destroy */
547 SIMIX_process_empty_trash();
549 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
551 if (xbt_dynar_is_empty(simix_global->process_to_run) &&
552 !simix_global->process_list.empty())
553 simgrid::simix::onDeadlock();
555 } while (time > -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
557 if (simix_global->process_list.size() != 0) {
561 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
562 SIMIX_display_process_status();
565 simgrid::s4u::onSimulationEnd();
569 * \brief Set the date to execute a function
571 * Set the date to execute the function on the surf.
572 * \param date Date to execute function
573 * \param callback Function to be executed
574 * \param arg Parameters of the function
577 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
579 smx_timer_t timer = new s_smx_timer_t(date, [=](){ callback(arg); });
580 xbt_heap_push(simix_timers, timer, date);
584 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
586 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
587 xbt_heap_push(simix_timers, timer, date);
591 /** @brief cancels a timer that was added earlier */
592 void SIMIX_timer_remove(smx_timer_t timer) {
593 xbt_heap_rm_elm(simix_timers, timer, timer->date);
596 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
597 double SIMIX_timer_get_date(smx_timer_t timer) {
598 return timer?timer->date:0;
602 * \brief Registers a function to create a process.
604 * This function registers a function to be called
605 * when a new process is created. The function has
606 * to call SIMIX_process_create().
607 * \param function create process function
609 void SIMIX_function_register_process_create(smx_creation_func_t function)
611 simix_global->create_process_function = function;
615 * \brief Registers a function to kill a process.
617 * This function registers a function to be called when a process is killed. The function has to call the
618 * SIMIX_process_kill().
620 * \param function Kill process function
622 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
624 simix_global->kill_process_function = function;
628 * \brief Registers a function to cleanup a process.
630 * This function registers a user function to be called when a process ends properly.
632 * \param function cleanup process function
634 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
636 simix_global->cleanup_process_function = function;
640 void SIMIX_display_process_status()
642 int nbprocess = simix_global->process_list.size();
644 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
645 /* List the process and their state */
646 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
647 for (auto kv : simix_global->process_list) {
648 smx_actor_t process = kv.second;
650 if (process->waiting_synchro) {
652 const char* synchro_description = "unknown";
654 if (dynamic_cast<simgrid::kernel::activity::Exec*>(process->waiting_synchro) != nullptr)
655 synchro_description = "execution";
657 if (dynamic_cast<simgrid::kernel::activity::Comm*>(process->waiting_synchro) != nullptr)
658 synchro_description = "communication";
660 if (dynamic_cast<simgrid::kernel::activity::Sleep*>(process->waiting_synchro) != nullptr)
661 synchro_description = "sleeping";
663 if (dynamic_cast<simgrid::kernel::activity::Raw*>(process->waiting_synchro) != nullptr)
664 synchro_description = "synchronization";
666 if (dynamic_cast<simgrid::kernel::activity::Io*>(process->waiting_synchro) != nullptr)
667 synchro_description = "I/O";
671 switch (process->waiting_synchro->type) {
672 case SIMIX_SYNC_PARALLEL_EXECUTE:
673 synchro_description = "parallel execution";
676 case SIMIX_SYNC_JOIN:
677 synchro_description = "joining";
681 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish", process->pid,
682 process->cname(), process->host->cname(), synchro_description, process->waiting_synchro,
683 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
686 XBT_INFO("Process %lu (%s@%s)", process->pid, process->cname(), process->host->cname());
691 int SIMIX_is_maestro()
693 return simix_global==nullptr /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;