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. */
9 #include <signal.h> /* Signal handling */
11 #include "src/internal_config.h"
13 #include "src/surf/surf_interface.hpp"
14 #include "src/surf/storage_interface.hpp"
15 #include "src/surf/xml/platf.hpp"
16 #include "smx_private.h"
18 #include "xbt/ex.h" /* ex_backtrace_display */
20 #include "src/mc/mc_replay.h"
21 #include "simgrid/sg_config.h"
23 #include "src/simix/SynchroExec.hpp"
24 #include "src/simix/SynchroComm.hpp"
25 #include "src/simix/SynchroSleep.hpp"
26 #include "src/simix/SynchroIo.hpp"
27 #include "src/simix/SynchroRaw.hpp"
30 #include "src/mc/mc_private.h"
31 #include "src/mc/mc_protocol.h"
32 #include "src/mc/Client.hpp"
35 #include "src/mc/mc_protocol.h"
38 #include "src/mc/mc_record.h"
41 #include "src/smpi/private.h"
44 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
45 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
47 smx_global_t simix_global = nullptr;
48 static xbt_heap_t simix_timers = nullptr;
50 /** @brief Timer datatype */
51 typedef struct s_smx_timer {
53 std::function<void()> callback;
56 s_smx_timer(double date, std::function<void()> callback)
57 : date(date), callback(std::move(callback)) {}
60 void (*SMPI_switch_data_segment)(int) = nullptr;
62 int _sg_do_verbose_exit = 1;
63 static void inthandler(int ignored)
65 if ( _sg_do_verbose_exit ) {
66 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
67 SIMIX_display_process_status();
70 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
76 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
78 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
80 "Access violation detected.\n"
81 "This can result from a programming error in your code or, although less likely,\n"
82 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
83 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
85 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
86 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
87 smx_context_stack_size / 1024);
88 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
90 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
91 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
93 } else if (siginfo->si_signo == SIGSEGV) {
94 fprintf(stderr, "Segmentation fault.\n");
96 if (smpi_enabled() && !smpi_privatize_global_variables) {
97 #if HAVE_PRIVATIZATION
99 "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
102 "Sadly, your system does not support --cfg=smpi/privatize-global-variables:yes (yet).\n");
103 #endif /* HAVE_PRIVATIZATION */
105 #endif /* HAVE_SMPI */
110 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
113 * Install signal handler for SIGSEGV. Check that nobody has already installed
114 * its own handler. For example, the Java VM does this.
116 static void install_segvhandler(void)
118 stack_t stack, old_stack;
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, old_action;
134 action.sa_sigaction = &segvhandler;
135 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
136 sigemptyset(&action.sa_mask);
138 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
139 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
142 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
143 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
144 (old_action.sa_flags & SA_SIGINFO) ?
145 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
146 sigaction(SIGSEGV, &old_action, nullptr);
152 /********************************* SIMIX **************************************/
153 double SIMIX_timer_next(void)
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 = xbt_new0(s_smx_global_t, 1);
199 simgrid::simix::Process 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));
205 simix_global->maestro_process = nullptr;
206 simix_global->registered_functions = xbt_dict_new_homogeneous(nullptr);
208 simix_global->create_process_function = &SIMIX_process_create;
209 simix_global->kill_process_function = &kill_process;
210 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
211 simix_global->mutex = xbt_os_mutex_init();
213 surf_init(argc, argv); /* Initialize SURF structures */
214 SIMIX_context_mod_init();
216 // Either create a new context with maestro or create
217 // a context object with the current context mestro):
218 simgrid::simix::create_maestro(maestro_code);
220 /* Prepare to display some more info when dying on Ctrl-C pressing */
221 signal(SIGINT, inthandler);
224 install_segvhandler();
226 /* register a function to be called by SURF after the environment creation */
228 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
229 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
230 SIMIX_host_create(&host);
232 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
234 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
235 const char* name = storage->getName();
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
264 void SIMIX_clean(void)
266 if (smx_cleaned) 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 (!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 xbt_swag_free(simix_global->process_list);
300 simix_global->process_list = nullptr;
301 simix_global->process_to_destroy = nullptr;
302 xbt_dict_free(&(simix_global->registered_functions));
304 xbt_os_mutex_destroy(simix_global->mutex);
305 simix_global->mutex = nullptr;
307 /* Let's free maestro now */
308 delete simix_global->maestro_process->context;
309 simix_global->maestro_process->context = nullptr;
310 delete simix_global->maestro_process;
311 simix_global->maestro_process = nullptr;
313 /* Finish context module and SURF */
314 SIMIX_context_mod_exit();
318 xbt_free(simix_global);
319 simix_global = nullptr;
327 * \brief A clock (in second).
329 * \return Return the clock.
331 double SIMIX_get_clock(void)
333 if(MC_is_active() || MC_record_replay_is_active()){
334 return MC_process_clock_get(SIMIX_process_self());
336 return surf_get_clock();
340 static int process_syscall_color(void *p)
342 switch ((*(smx_process_t *)p)->simcall.call) {
344 case SIMCALL_PROCESS_KILL:
346 case SIMCALL_PROCESS_RESUME:
355 * \brief Run the main simulation loop.
359 if (MC_record_path) {
360 simgrid::mc::replay(MC_record_path);
365 smx_process_t process;
366 surf_action_t action;
372 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
373 xbt_dynar_length(simix_global->process_to_run));
374 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
375 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
376 xbt_dynar_length(simix_global->process_to_run));
378 /* Run all processes that are ready to run, possibly in parallel */
379 SIMIX_process_runall();
381 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
382 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
384 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
386 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
388 /* Here, the order is ok because:
390 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
392 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
394 * - if there is no kill during the simulation, processes remain sorted according by their PID.
395 * rational: This can be proved inductively.
396 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
397 * Let's show that it is still so at the end of this round.
398 * - if a process is added when being created, that's from maestro. It can be either at startup
399 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
400 * in arbitrary order (inductive hypothesis), we are fine.
401 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
402 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
403 * of a simcall terminates, we're still good. Proof:
404 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
405 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
406 * - For communications, this function is called from SIMIX_comm_finish().
407 * This function itself don't mess with the order since simcalls are handled in FIFO order.
408 * The function is called:
409 * - before the comm starts (invalid parameters, or resource already dead or whatever).
410 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
411 * - because the communication failed or were canceled after startup. In this case, it's called from the function
412 * we are in, by the chunk:
413 * set = model->states.failed_action_set;
414 * while ((synchro = xbt_swag_extract(set)))
415 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
416 * This order is also fixed because it depends of the order in which the surf actions were
417 * added to the system, and only maestro can add stuff this way, through simcalls.
418 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
419 * poped out of the swag does not depend on the user code's execution order.
420 * - because the communication terminated. In this case, synchros are served in the order given by
421 * set = model->states.done_action_set;
422 * while ((synchro = xbt_swag_extract(set)))
423 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
424 * and the argument is very similar to the previous one.
425 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
426 * 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.
427 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
428 * - If there is some process killings, the order is changed by this decision that comes from user-land
429 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
430 * So, even the order change induced by the process killing is perfectly reproducible.
432 * So science works, bitches [http://xkcd.com/54/].
434 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
435 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
436 * That would thus be a pure waste of time.
439 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
440 if (process->simcall.call != SIMCALL_NONE) {
441 SIMIX_simcall_handle(&process->simcall, 0);
444 /* Wake up all processes waiting for a Surf action to finish */
445 xbt_dynar_foreach(all_existing_models, iter, model) {
446 XBT_DEBUG("Handling process whose action failed");
447 while ((action = surf_model_extract_failed_action_set(model))) {
448 XBT_DEBUG(" Handling Action %p",action);
449 SIMIX_simcall_exit((smx_synchro_t) action->getData());
451 XBT_DEBUG("Handling process whose action terminated normally");
452 while ((action = surf_model_extract_done_action_set(model))) {
453 XBT_DEBUG(" Handling Action %p",action);
454 if (action->getData() == nullptr)
455 XBT_DEBUG("probably vcpu's action %p, skip", action);
457 SIMIX_simcall_exit((smx_synchro_t) action->getData());
462 time = SIMIX_timer_next();
463 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
464 XBT_DEBUG("Calling surf_solve");
465 time = surf_solve(time);
466 XBT_DEBUG("Moving time ahead : %g", time);
468 /* Notify all the hosts that have failed */
469 /* FIXME: iterate through the list of failed host and mark each of them */
470 /* as failed. On each host, signal all the running processes with host_fail */
472 /* Handle any pending timer */
473 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
474 //FIXME: make the timers being real callbacks
475 // (i.e. provide dispatchers that read and expand the args)
476 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
477 if (timer->callback) {
482 xbt_die("Exception throwed ouf of timer callback");
488 /* Wake up all processes waiting for a Surf action to finish */
489 xbt_dynar_foreach(all_existing_models, iter, model) {
490 XBT_DEBUG("Handling process whose action failed");
491 while ((action = surf_model_extract_failed_action_set(model))) {
492 XBT_DEBUG(" Handling Action %p",action);
493 SIMIX_simcall_exit((smx_synchro_t) action->getData());
495 XBT_DEBUG("Handling process whose action terminated normally");
496 while ((action = surf_model_extract_done_action_set(model))) {
497 XBT_DEBUG(" Handling Action %p",action);
498 if (action->getData() == nullptr)
499 XBT_DEBUG("probably vcpu's action %p, skip", action);
501 SIMIX_simcall_exit((smx_synchro_t) action->getData());
505 /* Autorestart all process */
506 char *hostname = nullptr;
507 xbt_dynar_foreach(host_that_restart,iter,hostname) {
508 XBT_INFO("Restart processes on host: %s",hostname);
509 SIMIX_host_autorestart(sg_host_by_name(hostname));
511 xbt_dynar_reset(host_that_restart);
513 /* Clean processes to destroy */
514 SIMIX_process_empty_trash();
517 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
519 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
521 if (xbt_swag_size(simix_global->process_list) != 0) {
525 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
526 SIMIX_display_process_status();
532 * \brief Set the date to execute a function
534 * Set the date to execute the function on the surf.
535 * \param date Date to execute function
536 * \param function Function to be executed
537 * \param arg Parameters of the function
540 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
542 smx_timer_t timer = new s_smx_timer_t(date, std::bind(function, arg));
543 xbt_heap_push(simix_timers, timer, date);
547 smx_timer_t SIMIX_timer_set(double date, std::function<void()> callback)
549 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
550 xbt_heap_push(simix_timers, timer, date);
554 /** @brief cancels a timer that was added earlier */
555 void SIMIX_timer_remove(smx_timer_t timer) {
556 xbt_heap_rm_elm(simix_timers, timer, timer->date);
559 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
560 double SIMIX_timer_get_date(smx_timer_t timer) {
561 return timer?timer->date:0;
565 * \brief Registers a function to create a process.
567 * This function registers a function to be called
568 * when a new process is created. The function has
569 * to call SIMIX_process_create().
570 * \param function create process function
572 void SIMIX_function_register_process_create(smx_creation_func_t function)
574 simix_global->create_process_function = function;
578 * \brief Registers a function to kill a process.
580 * This function registers a function to be called when a
581 * process is killed. The function has to call the SIMIX_process_kill().
583 * \param function Kill process function
585 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
588 simix_global->kill_process_function = function;
592 * \brief Registers a function to cleanup a process.
594 * This function registers a user function to be called when
595 * a process ends properly.
597 * \param function cleanup process function
599 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
602 simix_global->cleanup_process_function = function;
606 void SIMIX_display_process_status(void)
608 if (simix_global->process_list == nullptr) {
612 smx_process_t process = nullptr;
613 int nbprocess = xbt_swag_size(simix_global->process_list);
615 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
616 /* List the process and their state */
618 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
619 xbt_swag_foreach(process, simix_global->process_list) {
621 if (process->waiting_synchro) {
623 const char* synchro_description = "unknown";
625 if (dynamic_cast<simgrid::simix::Exec*>(process->waiting_synchro) != nullptr)
626 synchro_description = "execution";
628 if (dynamic_cast<simgrid::simix::Comm*>(process->waiting_synchro) != nullptr)
629 synchro_description = "communication";
631 if (dynamic_cast<simgrid::simix::Sleep*>(process->waiting_synchro) != nullptr)
632 synchro_description = "sleeping";
634 if (dynamic_cast<simgrid::simix::Raw*>(process->waiting_synchro) != nullptr)
635 synchro_description = "synchronization";
637 if (dynamic_cast<simgrid::simix::Io*>(process->waiting_synchro) != nullptr)
638 synchro_description = "I/O";
642 switch (process->waiting_synchro->type) {
643 case SIMIX_SYNC_PARALLEL_EXECUTE:
644 synchro_description = "parallel execution";
647 case SIMIX_SYNC_JOIN:
648 synchro_description = "joining";
652 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
653 process->pid, process->name.c_str(), sg_host_get_name(process->host),
654 synchro_description, process->waiting_synchro,
655 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
658 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name.c_str(), sg_host_get_name(process->host));
663 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
664 return SIMIX_asr_get_properties(name);
666 xbt_dict_t SIMIX_asr_get_properties(const char *name)
668 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
671 int SIMIX_is_maestro()
673 return simix_global==nullptr /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;