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. */
7 #include <signal.h> /* Signal handling */
9 #include "src/internal_config.h"
11 #include "src/surf/surf_interface.hpp"
12 #include "src/surf/storage_interface.hpp"
13 #include "src/surf/xml/platf.hpp"
14 #include "smx_private.h"
16 #include "xbt/ex.h" /* ex_backtrace_display */
18 #include "src/mc/mc_replay.h"
19 #include "simgrid/sg_config.h"
21 #include "src/simix/SynchroExec.hpp"
22 #include "src/simix/SynchroComm.hpp"
23 #include "src/simix/SynchroSleep.hpp"
24 #include "src/simix/SynchroIo.hpp"
25 #include "src/simix/SynchroRaw.hpp"
28 #include "src/mc/mc_private.h"
29 #include "src/mc/mc_protocol.h"
30 #include "src/mc/Client.hpp"
33 #include "src/mc/mc_protocol.h"
36 #include "src/mc/mc_record.h"
39 #include "src/smpi/private.h"
42 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
43 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
45 smx_global_t simix_global = NULL;
46 static xbt_heap_t simix_timers = NULL;
48 /** @brief Timer datatype */
49 typedef struct s_smx_timer {
55 void (*SMPI_switch_data_segment)(int) = NULL;
57 int _sg_do_verbose_exit = 1;
58 static void inthandler(int ignored)
60 if ( _sg_do_verbose_exit ) {
61 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
62 SIMIX_display_process_status();
65 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
71 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
73 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
75 "Access violation detected.\n"
76 "This can result from a programming error in your code or, although less likely,\n"
77 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
78 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
80 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
81 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
82 smx_context_stack_size / 1024);
83 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
85 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
86 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
88 } else if (siginfo->si_signo == SIGSEGV) {
89 fprintf(stderr, "Segmentation fault.\n");
91 if (smpi_enabled() && !smpi_privatize_global_variables) {
92 #if HAVE_PRIVATIZATION
94 "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
97 "Sadly, your system does not support --cfg=smpi/privatize-global-variables:yes (yet).\n");
98 #endif /* HAVE_PRIVATIZATION */
100 #endif /* HAVE_SMPI */
105 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
108 * Install signal handler for SIGSEGV. Check that nobody has already installed
109 * its own handler. For example, the Java VM does this.
111 static void install_segvhandler(void)
113 stack_t stack, old_stack;
114 stack.ss_sp = sigsegv_stack;
115 stack.ss_size = sizeof sigsegv_stack;
118 if (sigaltstack(&stack, &old_stack) == -1) {
119 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
122 if (!(old_stack.ss_flags & SS_DISABLE)) {
123 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
124 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
125 sigaltstack(&old_stack, NULL);
128 struct sigaction action, old_action;
129 action.sa_sigaction = segvhandler;
130 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
131 sigemptyset(&action.sa_mask);
133 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
134 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
137 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
138 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
139 (old_action.sa_flags & SA_SIGINFO) ?
140 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
141 sigaction(SIGSEGV, &old_action, NULL);
147 /********************************* SIMIX **************************************/
148 double SIMIX_timer_next(void)
150 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
153 static void kill_process(smx_process_t process)
155 SIMIX_process_kill(process, NULL);
158 static std::function<void()> maestro_code;
163 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
165 maestro_code = std::move(code);
171 void SIMIX_set_maestro(void (*code)(void*), void* data)
173 maestro_code = std::bind(code, data);
178 * \brief Initialize SIMIX internal data.
183 void SIMIX_global_init(int *argc, char **argv)
186 // The communication initialization is done ASAP.
187 // We need to communicate initialization of the different layers to the model-checker.
188 simgrid::mc::Client::initialize();
191 s_smx_process_t proc;
194 simix_global = xbt_new0(s_smx_global_t, 1);
196 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
197 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
198 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
199 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
201 simix_global->maestro_process = NULL;
202 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
204 simix_global->create_process_function = SIMIX_process_create;
205 simix_global->kill_process_function = kill_process;
206 simix_global->cleanup_process_function = SIMIX_process_cleanup;
207 simix_global->mutex = xbt_os_mutex_init();
209 surf_init(argc, argv); /* Initialize SURF structures */
210 SIMIX_context_mod_init();
212 // Either create a new context with maestro or create
213 // a context object with the current context mestro):
214 simgrid::simix::create_maestro(maestro_code);
216 /* context exception handlers */
217 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
218 __xbt_ex_terminate = SIMIX_process_exception_terminate;
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 != NULL, "Storage not found for name %s", name);
240 SIMIX_storage_create(name, s, NULL);
243 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
246 simix_timers = xbt_heap_new(8, &free);
249 if (xbt_cfg_get_boolean("clean-atexit"))
252 if (_sg_cfg_exit_asap)
259 * \brief Clean the SIMIX simulation
261 * This functions remove the memory used by SIMIX
263 void SIMIX_clean(void)
265 if (smx_cleaned) return; // to avoid double cleaning by java and C
267 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
268 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
270 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
271 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
272 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
274 /* Kill all processes (but maestro) */
275 SIMIX_process_killall(simix_global->maestro_process, 1);
277 /* Exit the SIMIX network module */
278 SIMIX_mailbox_exit();
280 xbt_heap_free(simix_timers);
282 /* Free the remaining data structures */
283 xbt_dynar_free(&simix_global->process_to_run);
284 xbt_dynar_free(&simix_global->process_that_ran);
285 xbt_swag_free(simix_global->process_to_destroy);
286 xbt_swag_free(simix_global->process_list);
287 simix_global->process_list = NULL;
288 simix_global->process_to_destroy = NULL;
289 xbt_dict_free(&(simix_global->registered_functions));
291 xbt_os_mutex_destroy(simix_global->mutex);
292 simix_global->mutex = NULL;
294 /* Let's free maestro now */
295 delete simix_global->maestro_process->context;
296 xbt_free(simix_global->maestro_process->running_ctx);
297 xbt_free(simix_global->maestro_process);
298 simix_global->maestro_process = NULL;
300 /* Restore the default exception setup */
301 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
302 __xbt_ex_terminate = &__xbt_ex_terminate_default;
304 /* Finish context module and SURF */
305 SIMIX_context_mod_exit();
309 xbt_free(simix_global);
318 * \brief A clock (in second).
320 * \return Return the clock.
322 double SIMIX_get_clock(void)
324 if(MC_is_active() || MC_record_replay_is_active()){
325 return MC_process_clock_get(SIMIX_process_self());
327 return surf_get_clock();
331 static int process_syscall_color(void *p)
333 switch ((*(smx_process_t *)p)->simcall.call) {
335 case SIMCALL_PROCESS_KILL:
337 case SIMCALL_PROCESS_RESUME:
346 * \brief Run the main simulation loop.
350 if (MC_record_path) {
351 simgrid::mc::replay(MC_record_path);
356 smx_process_t process;
357 surf_action_t action;
363 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
364 xbt_dynar_length(simix_global->process_to_run));
365 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
366 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
367 xbt_dynar_length(simix_global->process_to_run));
369 /* Run all processes that are ready to run, possibly in parallel */
370 SIMIX_process_runall();
372 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
373 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
375 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
377 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
379 /* Here, the order is ok because:
381 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
383 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
385 * - if there is no kill during the simulation, processes remain sorted according by their PID.
386 * rational: This can be proved inductively.
387 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
388 * Let's show that it is still so at the end of this round.
389 * - if a process is added when being created, that's from maestro. It can be either at startup
390 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
391 * in arbitrary order (inductive hypothesis), we are fine.
392 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
393 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
394 * of a simcall terminates, we're still good. Proof:
395 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
396 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
397 * - For communications, this function is called from SIMIX_comm_finish().
398 * This function itself don't mess with the order since simcalls are handled in FIFO order.
399 * The function is called:
400 * - before the comm starts (invalid parameters, or resource already dead or whatever).
401 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
402 * - because the communication failed or were canceled after startup. In this case, it's called from the function
403 * we are in, by the chunk:
404 * set = model->states.failed_action_set;
405 * while ((synchro = xbt_swag_extract(set)))
406 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
407 * This order is also fixed because it depends of the order in which the surf actions were
408 * added to the system, and only maestro can add stuff this way, through simcalls.
409 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
410 * poped out of the swag does not depend on the user code's execution order.
411 * - because the communication terminated. In this case, synchros are served in the order given by
412 * set = model->states.done_action_set;
413 * while ((synchro = xbt_swag_extract(set)))
414 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
415 * and the argument is very similar to the previous one.
416 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
417 * 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.
418 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
419 * - If there is some process killings, the order is changed by this decision that comes from user-land
420 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
421 * So, even the order change induced by the process killing is perfectly reproducible.
423 * So science works, bitches [http://xkcd.com/54/].
425 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
426 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
427 * That would thus be a pure waste of time.
430 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
431 if (process->simcall.call != SIMCALL_NONE) {
432 SIMIX_simcall_handle(&process->simcall, 0);
435 /* Wake up all processes waiting for a Surf action to finish */
436 xbt_dynar_foreach(all_existing_models, iter, model) {
437 XBT_DEBUG("Handling process whose action failed");
438 while ((action = surf_model_extract_failed_action_set(model))) {
439 XBT_DEBUG(" Handling Action %p",action);
440 SIMIX_simcall_exit((smx_synchro_t) action->getData());
442 XBT_DEBUG("Handling process whose action terminated normally");
443 while ((action = surf_model_extract_done_action_set(model))) {
444 XBT_DEBUG(" Handling Action %p",action);
445 if (action->getData() == NULL)
446 XBT_DEBUG("probably vcpu's action %p, skip", action);
448 SIMIX_simcall_exit((smx_synchro_t) action->getData());
453 time = SIMIX_timer_next();
454 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
455 XBT_DEBUG("Calling surf_solve");
456 time = surf_solve(time);
457 XBT_DEBUG("Moving time ahead : %g", time);
459 /* Notify all the hosts that have failed */
460 /* FIXME: iterate through the list of failed host and mark each of them */
461 /* as failed. On each host, signal all the running processes with host_fail */
463 /* Handle any pending timer */
464 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
465 //FIXME: make the timers being real callbacks
466 // (i.e. provide dispatchers that read and expand the args)
467 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
469 timer->func(timer->args);
473 /* Wake up all processes waiting for a Surf action to finish */
474 xbt_dynar_foreach(all_existing_models, iter, model) {
475 XBT_DEBUG("Handling process whose action failed");
476 while ((action = surf_model_extract_failed_action_set(model))) {
477 XBT_DEBUG(" Handling Action %p",action);
478 SIMIX_simcall_exit((smx_synchro_t) action->getData());
480 XBT_DEBUG("Handling process whose action terminated normally");
481 while ((action = surf_model_extract_done_action_set(model))) {
482 XBT_DEBUG(" Handling Action %p",action);
483 if (action->getData() == NULL)
484 XBT_DEBUG("probably vcpu's action %p, skip", action);
486 SIMIX_simcall_exit((smx_synchro_t) action->getData());
490 /* Autorestart all process */
491 char *hostname = NULL;
492 xbt_dynar_foreach(host_that_restart,iter,hostname) {
493 XBT_INFO("Restart processes on host: %s",hostname);
494 SIMIX_host_autorestart(sg_host_by_name(hostname));
496 xbt_dynar_reset(host_that_restart);
498 /* Clean processes to destroy */
499 SIMIX_process_empty_trash();
502 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
504 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
506 if (xbt_swag_size(simix_global->process_list) != 0) {
510 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
511 SIMIX_display_process_status();
517 * \brief Set the date to execute a function
519 * Set the date to execute the function on the surf.
520 * \param date Date to execute function
521 * \param function Function to be executed
522 * \param arg Parameters of the function
525 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
527 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
530 timer->func = function;
532 xbt_heap_push(simix_timers, timer, date);
535 /** @brief cancels a timer that was added earlier */
536 void SIMIX_timer_remove(smx_timer_t timer) {
537 xbt_heap_rm_elm(simix_timers, timer, timer->date);
540 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
541 double SIMIX_timer_get_date(smx_timer_t timer) {
542 return timer?timer->date:0;
546 * \brief Registers a function to create a process.
548 * This function registers a function to be called
549 * when a new process is created. The function has
550 * to call SIMIX_process_create().
551 * \param function create process function
553 void SIMIX_function_register_process_create(smx_creation_func_t
556 simix_global->create_process_function = function;
560 * \brief Registers a function to kill a process.
562 * This function registers a function to be called when a
563 * process is killed. The function has to call the SIMIX_process_kill().
565 * \param function Kill process function
567 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
570 simix_global->kill_process_function = function;
574 * \brief Registers a function to cleanup a process.
576 * This function registers a user function to be called when
577 * a process ends properly.
579 * \param function cleanup process function
581 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
584 simix_global->cleanup_process_function = function;
588 void SIMIX_display_process_status(void)
590 if (simix_global->process_list == NULL) {
594 smx_process_t process = NULL;
595 int nbprocess = xbt_swag_size(simix_global->process_list);
597 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
598 /* List the process and their state */
600 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
601 xbt_swag_foreach(process, simix_global->process_list) {
603 if (process->waiting_synchro) {
605 const char* synchro_description = "unknown";
607 if (dynamic_cast<simgrid::simix::Exec*>(process->waiting_synchro) != nullptr)
608 synchro_description = "execution";
610 if (dynamic_cast<simgrid::simix::Comm*>(process->waiting_synchro) != nullptr)
611 synchro_description = "communication";
613 if (dynamic_cast<simgrid::simix::Sleep*>(process->waiting_synchro) != nullptr)
614 synchro_description = "sleeping";
616 if (dynamic_cast<simgrid::simix::Raw*>(process->waiting_synchro) != nullptr)
617 synchro_description = "synchronization";
619 if (dynamic_cast<simgrid::simix::Io*>(process->waiting_synchro) != nullptr)
620 synchro_description = "I/O";
624 switch (process->waiting_synchro->type) {
625 case SIMIX_SYNC_PARALLEL_EXECUTE:
626 synchro_description = "parallel execution";
629 case SIMIX_SYNC_JOIN:
630 synchro_description = "joining";
634 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
635 process->pid, process->name, sg_host_get_name(process->host),
636 synchro_description, process->waiting_synchro,
637 process->waiting_synchro->name, (int)process->waiting_synchro->state);
640 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
645 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
646 return SIMIX_asr_get_properties(name);
648 xbt_dict_t SIMIX_asr_get_properties(const char *name)
650 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
653 int SIMIX_is_maestro()
655 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;