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/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/mc_protocol.h"
37 #include "src/mc/Client.hpp"
40 #include "src/mc/mc_protocol.h"
43 #include "src/mc/mc_record.h"
46 #include "src/smpi/private.h"
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 */
56 typedef struct s_smx_timer {
58 simgrid::xbt::Task<void()> callback;
60 s_smx_timer()=default;
61 s_smx_timer(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
64 void (*SMPI_switch_data_segment)(int) = nullptr;
66 int _sg_do_verbose_exit = 1;
67 static void inthandler(int ignored)
69 if ( _sg_do_verbose_exit ) {
70 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
71 SIMIX_display_process_status();
74 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
80 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
82 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
84 "Access violation detected.\n"
85 "This can result from a programming error in your code or, although less likely,\n"
86 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
87 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
89 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
90 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
91 smx_context_stack_size / 1024);
92 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
93 fprintf(stderr, "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
94 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
96 } else if (siginfo->si_signo == SIGSEGV) {
97 fprintf(stderr, "Segmentation fault.\n");
99 if (smpi_enabled() && !smpi_privatize_global_variables) {
100 #if HAVE_PRIVATIZATION
101 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
103 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatize-global-variables:yes (yet).\n");
104 #endif /* HAVE_PRIVATIZATION */
106 #endif /* HAVE_SMPI */
111 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
114 * Install signal handler for SIGSEGV. Check that nobody has already installed
115 * its own handler. For example, the Java VM does this.
117 static void install_segvhandler()
119 stack_t stack, old_stack;
120 stack.ss_sp = sigsegv_stack;
121 stack.ss_size = sizeof sigsegv_stack;
124 if (sigaltstack(&stack, &old_stack) == -1) {
125 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
128 if (!(old_stack.ss_flags & SS_DISABLE)) {
129 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
130 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
131 sigaltstack(&old_stack, nullptr);
134 struct sigaction action, 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_process_t process)
160 SIMIX_process_kill(process, nullptr);
163 static std::function<void()> maestro_code;
168 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
170 maestro_code = std::move(code);
176 void SIMIX_set_maestro(void (*code)(void*), void* data)
178 maestro_code = std::bind(code, data);
183 * \brief Initialize SIMIX internal data.
188 void SIMIX_global_init(int *argc, char **argv)
191 // The communication initialization is done ASAP.
192 // We need to communicate initialization of the different layers to the model-checker.
193 simgrid::mc::Client::initialize();
197 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
199 simgrid::simix::ActorImpl proc;
200 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), nullptr);
201 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), nullptr);
202 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
203 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
204 simix_global->maestro_process = nullptr;
205 simix_global->create_process_function = &SIMIX_process_create;
206 simix_global->kill_process_function = &kill_process;
207 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
208 simix_global->mutex = xbt_os_mutex_init();
210 surf_init(argc, argv); /* Initialize SURF structures */
211 SIMIX_context_mod_init();
213 // Either create a new context with maestro or create
214 // a context object with the current context mestro):
215 simgrid::simix::create_maestro(maestro_code);
217 /* Prepare to display some more info when dying on Ctrl-C pressing */
218 signal(SIGINT, inthandler);
221 install_segvhandler();
223 /* register a function to be called by SURF after the environment creation */
225 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
226 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
227 SIMIX_host_create(&host);
229 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
231 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
232 const char* name = storage->getName();
233 // TODO, create sg_storage_by_name
234 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, name);
235 xbt_assert(s != nullptr, "Storage not found for name %s", name);
237 SIMIX_storage_create(name, s, nullptr);
240 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
243 simix_timers = xbt_heap_new(8, [](void* p) {
244 delete static_cast<smx_timer_t>(p);
247 if (xbt_cfg_get_boolean("clean-atexit"))
250 if (_sg_cfg_exit_asap)
257 * \brief Clean the SIMIX simulation
259 * This functions remove the memory used by SIMIX
264 return; // to avoid double cleaning by java and C
267 if (SIMIX_process_count()>0){
268 if(smpi_process_initialized()){
269 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
271 XBT_WARN("Process called exit when leaving - Skipping cleanups");
278 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
279 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() <= 0.0) {
281 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
282 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
283 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
285 /* Kill all processes (but maestro) */
286 SIMIX_process_killall(simix_global->maestro_process, 1);
288 /* Exit the SIMIX network module */
289 SIMIX_mailbox_exit();
291 xbt_heap_free(simix_timers);
292 simix_timers = nullptr;
293 /* Free the remaining data structures */
294 xbt_dynar_free(&simix_global->process_to_run);
295 xbt_dynar_free(&simix_global->process_that_ran);
296 xbt_swag_free(simix_global->process_to_destroy);
297 xbt_swag_free(simix_global->process_list);
298 simix_global->process_list = nullptr;
299 simix_global->process_to_destroy = nullptr;
301 xbt_os_mutex_destroy(simix_global->mutex);
302 simix_global->mutex = nullptr;
304 /* Let's free maestro now */
305 delete simix_global->maestro_process->context;
306 simix_global->maestro_process->context = nullptr;
307 delete simix_global->maestro_process;
308 simix_global->maestro_process = nullptr;
310 /* Finish context module and SURF */
311 SIMIX_context_mod_exit();
315 simix_global = nullptr;
322 * \brief A clock (in second).
324 * \return Return the clock.
326 double SIMIX_get_clock()
328 if(MC_is_active() || MC_record_replay_is_active()){
329 return MC_process_clock_get(SIMIX_process_self());
331 return surf_get_clock();
335 static int process_syscall_color(void *p)
337 switch ((*(smx_process_t *)p)->simcall.call) {
339 case SIMCALL_PROCESS_KILL:
341 case SIMCALL_PROCESS_RESUME:
348 /** Wake up all processes waiting for a Surf action to finish */
349 static void SIMIX_wake_processes()
353 surf_action_t action;
355 xbt_dynar_foreach(all_existing_models, iter, model) {
356 XBT_DEBUG("Handling the processes whose action failed (if any)");
357 while ((action = surf_model_extract_failed_action_set(model))) {
358 XBT_DEBUG(" Handling Action %p",action);
359 SIMIX_simcall_exit((smx_activity_t) action->getData());
361 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
362 while ((action = surf_model_extract_done_action_set(model))) {
363 XBT_DEBUG(" Handling Action %p",action);
364 if (action->getData() == nullptr)
365 XBT_DEBUG("probably vcpu's action %p, skip", action);
367 SIMIX_simcall_exit((smx_activity_t) action->getData());
372 /** Handle any pending timer */
373 static bool SIMIX_execute_timers()
376 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
378 //FIXME: make the timers being real callbacks
379 // (i.e. provide dispatchers that read and expand the args)
380 smx_timer_t timer = (smx_timer_t) xbt_heap_pop(simix_timers);
385 xbt_die("Exception throwed ouf of timer callback");
392 /** Execute all the tasks that are queued
394 * e.g. `.then()` callbacks of futures.
396 static bool SIMIX_execute_tasks()
398 xbt_assert(simix_global->tasksTemp.empty());
400 if (simix_global->tasks.empty())
405 // We don't want the callbacks to modify the vector we are iterating over:
406 swap(simix_global->tasks, simix_global->tasksTemp);
408 // Execute all the queued tasks:
409 for (auto& task : simix_global->tasksTemp)
412 simix_global->tasksTemp.clear();
413 } while (!simix_global->tasks.empty());
420 * \brief Run the main simulation loop.
424 if (MC_record_path) {
425 simgrid::mc::replay(MC_record_path);
430 smx_process_t process;
433 XBT_DEBUG("New Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
435 SIMIX_execute_tasks();
437 while (!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.
502 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
503 if (process->simcall.call != SIMCALL_NONE) {
504 SIMIX_simcall_handle(&process->simcall, 0);
508 SIMIX_execute_tasks();
510 SIMIX_wake_processes();
511 } while (SIMIX_execute_tasks());
515 time = SIMIX_timer_next();
516 if (time > -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
517 XBT_DEBUG("Calling surf_solve");
518 time = surf_solve(time);
519 XBT_DEBUG("Moving time ahead : %g", time);
522 /* Notify all the hosts that have failed */
523 /* FIXME: iterate through the list of failed host and mark each of them */
524 /* as failed. On each host, signal all the running processes with host_fail */
526 // Execute timers and tasks until there isn't anything to be done:
529 again = SIMIX_execute_timers();
530 if (SIMIX_execute_tasks())
532 SIMIX_wake_processes();
535 /* Autorestart all process */
536 char *hostname = nullptr;
538 xbt_dynar_foreach(host_that_restart,iter,hostname) {
539 XBT_INFO("Restart processes on host: %s",hostname);
540 SIMIX_host_autorestart(sg_host_by_name(hostname));
542 xbt_dynar_reset(host_that_restart);
544 /* Clean processes to destroy */
545 SIMIX_process_empty_trash();
547 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
549 } while (time > -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
551 if (xbt_swag_size(simix_global->process_list) != 0) {
555 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
556 SIMIX_display_process_status();
562 * \brief Set the date to execute a function
564 * Set the date to execute the function on the surf.
565 * \param date Date to execute function
566 * \param function Function to be executed
567 * \param arg Parameters of the function
570 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
572 smx_timer_t timer = new s_smx_timer_t(date, [=](){ callback(arg); });
573 xbt_heap_push(simix_timers, timer, date);
577 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
579 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
580 xbt_heap_push(simix_timers, timer, date);
584 /** @brief cancels a timer that was added earlier */
585 void SIMIX_timer_remove(smx_timer_t timer) {
586 xbt_heap_rm_elm(simix_timers, timer, timer->date);
589 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
590 double SIMIX_timer_get_date(smx_timer_t timer) {
591 return timer?timer->date:0;
595 * \brief Registers a function to create a process.
597 * This function registers a function to be called
598 * when a new process is created. The function has
599 * to call SIMIX_process_create().
600 * \param function create process function
602 void SIMIX_function_register_process_create(smx_creation_func_t function)
604 simix_global->create_process_function = function;
608 * \brief Registers a function to kill a process.
610 * This function registers a function to be called when a process is killed. The function has to call the
611 * SIMIX_process_kill().
613 * \param function Kill process function
615 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
617 simix_global->kill_process_function = function;
621 * \brief Registers a function to cleanup a process.
623 * This function registers a user function to be called when a process ends properly.
625 * \param function cleanup process function
627 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
629 simix_global->cleanup_process_function = function;
633 void SIMIX_display_process_status()
635 if (simix_global->process_list == nullptr) {
639 smx_process_t process = nullptr;
640 int nbprocess = xbt_swag_size(simix_global->process_list);
642 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
643 /* List the process and their state */
644 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
645 xbt_swag_foreach(process, simix_global->process_list) {
647 if (process->waiting_synchro) {
649 const char* synchro_description = "unknown";
651 if (dynamic_cast<simgrid::kernel::activity::Exec*>(process->waiting_synchro) != nullptr)
652 synchro_description = "execution";
654 if (dynamic_cast<simgrid::kernel::activity::Comm*>(process->waiting_synchro) != nullptr)
655 synchro_description = "communication";
657 if (dynamic_cast<simgrid::kernel::activity::Sleep*>(process->waiting_synchro) != nullptr)
658 synchro_description = "sleeping";
660 if (dynamic_cast<simgrid::kernel::activity::Raw*>(process->waiting_synchro) != nullptr)
661 synchro_description = "synchronization";
663 if (dynamic_cast<simgrid::kernel::activity::Io*>(process->waiting_synchro) != nullptr)
664 synchro_description = "I/O";
668 switch (process->waiting_synchro->type) {
669 case SIMIX_SYNC_PARALLEL_EXECUTE:
670 synchro_description = "parallel execution";
673 case SIMIX_SYNC_JOIN:
674 synchro_description = "joining";
678 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
679 process->pid, process->name.c_str(), sg_host_get_name(process->host),
680 synchro_description, process->waiting_synchro,
681 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
684 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name.c_str(), sg_host_get_name(process->host));
689 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
690 return SIMIX_asr_get_properties(name);
692 xbt_dict_t SIMIX_asr_get_properties(const char *name)
694 return static_cast<xbt_dict_t>(xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL));
697 int SIMIX_is_maestro()
699 return simix_global==nullptr /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;