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;
61 s_smx_timer(double date, simgrid::xbt::Task<void()> callback)
62 : 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) {
85 "Access violation detected.\n"
86 "This can result from a programming error in your code or, although less likely,\n"
87 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
88 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
90 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
91 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
92 smx_context_stack_size / 1024);
93 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
95 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
96 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
98 } else if (siginfo->si_signo == SIGSEGV) {
99 fprintf(stderr, "Segmentation fault.\n");
101 if (smpi_enabled() && !smpi_privatize_global_variables) {
102 #if HAVE_PRIVATIZATION
104 "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
107 "Sadly, your system does not support --cfg=smpi/privatize-global-variables:yes (yet).\n");
108 #endif /* HAVE_PRIVATIZATION */
110 #endif /* HAVE_SMPI */
115 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
118 * Install signal handler for SIGSEGV. Check that nobody has already installed
119 * its own handler. For example, the Java VM does this.
121 static void install_segvhandler(void)
123 stack_t stack, old_stack;
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, old_action;
139 action.sa_sigaction = &segvhandler;
140 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
141 sigemptyset(&action.sa_mask);
143 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
144 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
147 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
148 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
149 (old_action.sa_flags & SA_SIGINFO) ?
150 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
151 sigaction(SIGSEGV, &old_action, nullptr);
157 /********************************* SIMIX **************************************/
158 double SIMIX_timer_next(void)
160 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
163 static void kill_process(smx_process_t process)
165 SIMIX_process_kill(process, nullptr);
168 static std::function<void()> maestro_code;
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)
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::Process proc;
205 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), nullptr);
206 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), nullptr);
207 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
208 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
209 simix_global->maestro_process = nullptr;
210 simix_global->create_process_function = &SIMIX_process_create;
211 simix_global->kill_process_function = &kill_process;
212 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
213 simix_global->mutex = xbt_os_mutex_init();
215 surf_init(argc, argv); /* Initialize SURF structures */
216 SIMIX_context_mod_init();
218 // Either create a new context with maestro or create
219 // a context object with the current context mestro):
220 simgrid::simix::create_maestro(maestro_code);
222 /* Prepare to display some more info when dying on Ctrl-C pressing */
223 signal(SIGINT, inthandler);
226 install_segvhandler();
228 /* register a function to be called by SURF after the environment creation */
230 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
231 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
232 SIMIX_host_create(&host);
234 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
236 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
237 const char* name = storage->getName();
238 // TODO, create sg_storage_by_name
239 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, name);
240 xbt_assert(s != nullptr, "Storage not found for name %s", name);
242 SIMIX_storage_create(name, s, nullptr);
245 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
248 simix_timers = xbt_heap_new(8, [](void* p) {
249 delete static_cast<smx_timer_t>(p);
252 if (xbt_cfg_get_boolean("clean-atexit"))
255 if (_sg_cfg_exit_asap)
262 * \brief Clean the SIMIX simulation
264 * This functions remove the memory used by SIMIX
266 void SIMIX_clean(void)
268 if (smx_cleaned) return; // to avoid double cleaning by java and C
271 if (SIMIX_process_count()>0){
272 if(smpi_process_initialized()){
273 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
275 XBT_WARN("Process called exit when leaving - Skipping cleanups");
282 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
283 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
285 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
286 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
287 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
289 /* Kill all processes (but maestro) */
290 SIMIX_process_killall(simix_global->maestro_process, 1);
292 /* Exit the SIMIX network module */
293 SIMIX_mailbox_exit();
295 xbt_heap_free(simix_timers);
296 simix_timers = nullptr;
297 /* Free the remaining data structures */
298 xbt_dynar_free(&simix_global->process_to_run);
299 xbt_dynar_free(&simix_global->process_that_ran);
300 xbt_swag_free(simix_global->process_to_destroy);
301 xbt_swag_free(simix_global->process_list);
302 simix_global->process_list = nullptr;
303 simix_global->process_to_destroy = nullptr;
305 xbt_os_mutex_destroy(simix_global->mutex);
306 simix_global->mutex = nullptr;
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;
326 * \brief A clock (in second).
328 * \return Return the clock.
330 double SIMIX_get_clock(void)
332 if(MC_is_active() || MC_record_replay_is_active()){
333 return MC_process_clock_get(SIMIX_process_self());
335 return surf_get_clock();
339 static int process_syscall_color(void *p)
341 switch ((*(smx_process_t *)p)->simcall.call) {
343 case SIMCALL_PROCESS_KILL:
345 case SIMCALL_PROCESS_RESUME:
352 /** Wake up all processes waiting for a Surf action to finish */
353 static void SIMIX_wake_processes()
357 surf_action_t action;
359 xbt_dynar_foreach(all_existing_models, iter, model) {
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_synchro_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_synchro_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);
434 smx_process_t process;
437 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
438 xbt_dynar_length(simix_global->process_to_run));
440 SIMIX_execute_tasks();
442 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
443 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
444 xbt_dynar_length(simix_global->process_to_run));
446 /* Run all processes that are ready to run, possibly in parallel */
447 SIMIX_process_runall();
449 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
450 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
452 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
454 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
456 /* Here, the order is ok because:
458 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
460 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
462 * - if there is no kill during the simulation, processes remain sorted according by their PID.
463 * rational: This can be proved inductively.
464 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
465 * Let's show that it is still so at the end of this round.
466 * - if a process is added when being created, that's from maestro. It can be either at startup
467 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
468 * in arbitrary order (inductive hypothesis), we are fine.
469 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
470 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
471 * of a simcall terminates, we're still good. Proof:
472 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
473 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
474 * - For communications, this function is called from SIMIX_comm_finish().
475 * This function itself don't mess with the order since simcalls are handled in FIFO order.
476 * The function is called:
477 * - before the comm starts (invalid parameters, or resource already dead or whatever).
478 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
479 * - because the communication failed or were canceled after startup. In this case, it's called from the function
480 * we are in, by the chunk:
481 * set = model->states.failed_action_set;
482 * while ((synchro = xbt_swag_extract(set)))
483 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
484 * This order is also fixed because it depends of the order in which the surf actions were
485 * added to the system, and only maestro can add stuff this way, through simcalls.
486 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
487 * poped out of the swag does not depend on the user code's execution order.
488 * - because the communication terminated. In this case, synchros are served in the order given by
489 * set = model->states.done_action_set;
490 * while ((synchro = xbt_swag_extract(set)))
491 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
492 * and the argument is very similar to the previous one.
493 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
494 * 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.
495 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
496 * - If there is some process killings, the order is changed by this decision that comes from user-land
497 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
498 * So, even the order change induced by the process killing is perfectly reproducible.
500 * So science works, bitches [http://xkcd.com/54/].
502 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
503 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
504 * That would thus be a pure waste of time.
508 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
509 if (process->simcall.call != SIMCALL_NONE) {
510 SIMIX_simcall_handle(&process->simcall, 0);
514 SIMIX_execute_tasks();
516 SIMIX_wake_processes();
517 } while (SIMIX_execute_tasks());
521 time = SIMIX_timer_next();
522 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
523 XBT_DEBUG("Calling surf_solve");
524 time = surf_solve(time);
525 XBT_DEBUG("Moving time ahead : %g", time);
528 /* Notify all the hosts that have failed */
529 /* FIXME: iterate through the list of failed host and mark each of them */
530 /* as failed. On each host, signal all the running processes with host_fail */
532 // Execute timers and tasks until there isn't anything to be done:
535 again = SIMIX_execute_timers();
536 if (SIMIX_execute_tasks())
538 SIMIX_wake_processes();
541 /* Autorestart all process */
542 char *hostname = nullptr;
544 xbt_dynar_foreach(host_that_restart,iter,hostname) {
545 XBT_INFO("Restart processes on host: %s",hostname);
546 SIMIX_host_autorestart(sg_host_by_name(hostname));
548 xbt_dynar_reset(host_that_restart);
550 /* Clean processes to destroy */
551 SIMIX_process_empty_trash();
553 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
555 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
557 if (xbt_swag_size(simix_global->process_list) != 0) {
561 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
562 SIMIX_display_process_status();
568 * \brief Set the date to execute a function
570 * Set the date to execute the function on the surf.
571 * \param date Date to execute function
572 * \param function Function to be executed
573 * \param arg Parameters of the function
576 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
578 smx_timer_t timer = new s_smx_timer_t(date, [=](){ callback(arg); });
579 xbt_heap_push(simix_timers, timer, date);
583 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
585 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
586 xbt_heap_push(simix_timers, timer, date);
590 /** @brief cancels a timer that was added earlier */
591 void SIMIX_timer_remove(smx_timer_t timer) {
592 xbt_heap_rm_elm(simix_timers, timer, timer->date);
595 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
596 double SIMIX_timer_get_date(smx_timer_t timer) {
597 return timer?timer->date:0;
601 * \brief Registers a function to create a process.
603 * This function registers a function to be called
604 * when a new process is created. The function has
605 * to call SIMIX_process_create().
606 * \param function create process function
608 void SIMIX_function_register_process_create(smx_creation_func_t function)
610 simix_global->create_process_function = function;
614 * \brief Registers a function to kill a process.
616 * This function registers a function to be called when a
617 * process is killed. The function has to call the SIMIX_process_kill().
619 * \param function Kill process function
621 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
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
631 * a process ends properly.
633 * \param function cleanup process function
635 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
638 simix_global->cleanup_process_function = function;
642 void SIMIX_display_process_status(void)
644 if (simix_global->process_list == nullptr) {
648 smx_process_t process = nullptr;
649 int nbprocess = xbt_swag_size(simix_global->process_list);
651 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
652 /* List the process and their state */
654 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
655 xbt_swag_foreach(process, simix_global->process_list) {
657 if (process->waiting_synchro) {
659 const char* synchro_description = "unknown";
661 if (dynamic_cast<simgrid::kernel::activity::Exec*>(process->waiting_synchro) != nullptr)
662 synchro_description = "execution";
664 if (dynamic_cast<simgrid::kernel::activity::Comm*>(process->waiting_synchro) != nullptr)
665 synchro_description = "communication";
667 if (dynamic_cast<simgrid::kernel::activity::Sleep*>(process->waiting_synchro) != nullptr)
668 synchro_description = "sleeping";
670 if (dynamic_cast<simgrid::kernel::activity::Raw*>(process->waiting_synchro) != nullptr)
671 synchro_description = "synchronization";
673 if (dynamic_cast<simgrid::kernel::activity::Io*>(process->waiting_synchro) != nullptr)
674 synchro_description = "I/O";
678 switch (process->waiting_synchro->type) {
679 case SIMIX_SYNC_PARALLEL_EXECUTE:
680 synchro_description = "parallel execution";
683 case SIMIX_SYNC_JOIN:
684 synchro_description = "joining";
688 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
689 process->pid, process->name.c_str(), sg_host_get_name(process->host),
690 synchro_description, process->waiting_synchro,
691 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
694 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name.c_str(), sg_host_get_name(process->host));
699 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
700 return SIMIX_asr_get_properties(name);
702 xbt_dict_t SIMIX_asr_get_properties(const char *name)
704 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
707 int SIMIX_is_maestro()
709 return simix_global==nullptr /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;