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"
15 #include "smx_private.hpp"
17 #include "xbt/ex.h" /* ex_backtrace_display */
19 #include "src/mc/mc_replay.h"
20 #include "simgrid/sg_config.h"
23 #include "src/mc/mc_private.h"
24 #include "src/mc/mc_protocol.h"
25 #include "src/mc/Client.hpp"
28 #include "src/mc/mc_protocol.h"
31 #include "src/mc/mc_record.h"
34 #include "src/smpi/private.h"
37 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
38 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
40 smx_global_t simix_global = NULL;
41 static xbt_heap_t simix_timers = NULL;
43 /** @brief Timer datatype */
44 typedef struct s_smx_timer {
50 void (*SMPI_switch_data_segment)(int) = NULL;
52 static void* SIMIX_synchro_mallocator_new_f(void);
53 static void SIMIX_synchro_mallocator_free_f(void* synchro);
54 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
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->synchro_mallocator = xbt_mallocator_new(65536,
208 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
209 SIMIX_synchro_mallocator_reset_f);
210 simix_global->mutex = xbt_os_mutex_init();
212 surf_init(argc, argv); /* Initialize SURF structures */
213 SIMIX_context_mod_init();
215 // Either create a new context with maestro or create
216 // a context object with the current context mestro):
217 simgrid::simix::create_maestro(maestro_code);
219 /* context exception handlers */
220 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
221 __xbt_ex_terminate = SIMIX_process_exception_terminate;
223 /* Prepare to display some more info when dying on Ctrl-C pressing */
224 signal(SIGINT, inthandler);
227 install_segvhandler();
229 /* register a function to be called by SURF after the environment creation */
231 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
232 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
233 SIMIX_host_create(&host);
235 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
237 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
238 const char* name = storage->getName();
239 // TODO, create sg_storage_by_name
240 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, name);
241 xbt_assert(s != NULL, "Storage not found for name %s", name);
243 SIMIX_storage_create(name, s, NULL);
246 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
249 simix_timers = xbt_heap_new(8, &free);
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
270 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
271 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
273 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
274 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
275 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
277 /* Kill all processes (but maestro) */
278 SIMIX_process_killall(simix_global->maestro_process, 1);
280 /* Exit the SIMIX network module */
281 SIMIX_mailbox_exit();
283 xbt_heap_free(simix_timers);
285 /* Free the remaining data structures */
286 xbt_dynar_free(&simix_global->process_to_run);
287 xbt_dynar_free(&simix_global->process_that_ran);
288 xbt_swag_free(simix_global->process_to_destroy);
289 xbt_swag_free(simix_global->process_list);
290 simix_global->process_list = NULL;
291 simix_global->process_to_destroy = NULL;
292 xbt_dict_free(&(simix_global->registered_functions));
294 xbt_os_mutex_destroy(simix_global->mutex);
295 simix_global->mutex = NULL;
297 /* Let's free maestro now */
298 SIMIX_context_free(simix_global->maestro_process->context);
299 xbt_free(simix_global->maestro_process->running_ctx);
300 xbt_free(simix_global->maestro_process);
301 simix_global->maestro_process = NULL;
303 /* Restore the default exception setup */
304 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
305 __xbt_ex_terminate = &__xbt_ex_terminate_default;
307 /* Finish context module and SURF */
308 SIMIX_context_mod_exit();
312 xbt_mallocator_free(simix_global->synchro_mallocator);
313 xbt_free(simix_global);
322 * \brief A clock (in second).
324 * \return Return the clock.
326 double SIMIX_get_clock(void)
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:
350 * \brief Run the main simulation loop.
354 if (MC_record_path) {
355 simgrid::mc::replay(MC_record_path);
360 smx_process_t process;
361 surf_action_t action;
367 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
368 xbt_dynar_length(simix_global->process_to_run));
369 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
370 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
371 xbt_dynar_length(simix_global->process_to_run));
373 /* Run all processes that are ready to run, possibly in parallel */
374 SIMIX_process_runall();
376 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
377 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
379 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
381 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
383 /* Here, the order is ok because:
385 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
387 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
389 * - if there is no kill during the simulation, processes remain sorted according by their PID.
390 * rational: This can be proved inductively.
391 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
392 * Let's show that it is still so at the end of this round.
393 * - if a process is added when being created, that's from maestro. It can be either at startup
394 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
395 * in arbitrary order (inductive hypothesis), we are fine.
396 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
397 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
398 * of a simcall terminates, we're still good. Proof:
399 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
400 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
401 * - For communications, this function is called from SIMIX_comm_finish().
402 * This function itself don't mess with the order since simcalls are handled in FIFO order.
403 * The function is called:
404 * - before the comm starts (invalid parameters, or resource already dead or whatever).
405 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
406 * - because the communication failed or were canceled after startup. In this case, it's called from the function
407 * we are in, by the chunk:
408 * set = model->states.failed_action_set;
409 * while ((synchro = xbt_swag_extract(set)))
410 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
411 * This order is also fixed because it depends of the order in which the surf actions were
412 * added to the system, and only maestro can add stuff this way, through simcalls.
413 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
414 * poped out of the swag does not depend on the user code's execution order.
415 * - because the communication terminated. In this case, synchros are served in the order given by
416 * set = model->states.done_action_set;
417 * while ((synchro = xbt_swag_extract(set)))
418 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
419 * and the argument is very similar to the previous one.
420 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
421 * 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.
422 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
423 * - If there is some process killings, the order is changed by this decision that comes from user-land
424 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
425 * So, even the order change induced by the process killing is perfectly reproducible.
427 * So science works, bitches [http://xkcd.com/54/].
429 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
430 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
431 * That would thus be a pure waste of time.
434 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
435 if (process->simcall.call != SIMCALL_NONE) {
436 SIMIX_simcall_handle(&process->simcall, 0);
439 /* Wake up all processes waiting for a Surf action to finish */
440 xbt_dynar_foreach(all_existing_models, iter, model) {
441 XBT_DEBUG("Handling process whose action failed");
442 while ((action = surf_model_extract_failed_action_set(model))) {
443 XBT_DEBUG(" Handling Action %p",action);
444 SIMIX_simcall_exit((smx_synchro_t) action->getData());
446 XBT_DEBUG("Handling process whose action terminated normally");
447 while ((action = surf_model_extract_done_action_set(model))) {
448 XBT_DEBUG(" Handling Action %p",action);
449 if (action->getData() == NULL)
450 XBT_DEBUG("probably vcpu's action %p, skip", action);
452 SIMIX_simcall_exit((smx_synchro_t) action->getData());
457 time = SIMIX_timer_next();
458 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
459 XBT_DEBUG("Calling surf_solve");
460 time = surf_solve(time);
461 XBT_DEBUG("Moving time ahead : %g", time);
463 /* Notify all the hosts that have failed */
464 /* FIXME: iterate through the list of failed host and mark each of them */
465 /* as failed. On each host, signal all the running processes with host_fail */
467 /* Handle any pending timer */
468 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
469 //FIXME: make the timers being real callbacks
470 // (i.e. provide dispatchers that read and expand the args)
471 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
473 timer->func(timer->args);
477 /* Wake up all processes waiting for a Surf action to finish */
478 xbt_dynar_foreach(all_existing_models, iter, model) {
479 XBT_DEBUG("Handling process whose action failed");
480 while ((action = surf_model_extract_failed_action_set(model))) {
481 XBT_DEBUG(" Handling Action %p",action);
482 SIMIX_simcall_exit((smx_synchro_t) action->getData());
484 XBT_DEBUG("Handling process whose action terminated normally");
485 while ((action = surf_model_extract_done_action_set(model))) {
486 XBT_DEBUG(" Handling Action %p",action);
487 if (action->getData() == NULL)
488 XBT_DEBUG("probably vcpu's action %p, skip", action);
490 SIMIX_simcall_exit((smx_synchro_t) action->getData());
494 /* Autorestart all process */
495 char *hostname = NULL;
496 xbt_dynar_foreach(host_that_restart,iter,hostname) {
497 XBT_INFO("Restart processes on host: %s",hostname);
498 SIMIX_host_autorestart(sg_host_by_name(hostname));
500 xbt_dynar_reset(host_that_restart);
502 /* Clean processes to destroy */
503 SIMIX_process_empty_trash();
506 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
508 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
510 if (xbt_swag_size(simix_global->process_list) != 0) {
514 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
515 SIMIX_display_process_status();
521 * \brief Set the date to execute a function
523 * Set the date to execute the function on the surf.
524 * \param date Date to execute function
525 * \param function Function to be executed
526 * \param arg Parameters of the function
529 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
531 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
534 timer->func = function;
536 xbt_heap_push(simix_timers, timer, date);
539 /** @brief cancels a timer that was added earlier */
540 void SIMIX_timer_remove(smx_timer_t timer) {
541 xbt_heap_rm_elm(simix_timers, timer, timer->date);
544 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
545 double SIMIX_timer_get_date(smx_timer_t timer) {
546 return timer?timer->date:0;
550 * \brief Registers a function to create a process.
552 * This function registers a function to be called
553 * when a new process is created. The function has
554 * to call SIMIX_process_create().
555 * \param function create process function
557 void SIMIX_function_register_process_create(smx_creation_func_t
560 simix_global->create_process_function = function;
564 * \brief Registers a function to kill a process.
566 * This function registers a function to be called when a
567 * process is killed. The function has to call the SIMIX_process_kill().
569 * \param function Kill process function
571 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
574 simix_global->kill_process_function = function;
578 * \brief Registers a function to cleanup a process.
580 * This function registers a user function to be called when
581 * a process ends properly.
583 * \param function cleanup process function
585 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
588 simix_global->cleanup_process_function = function;
592 void SIMIX_display_process_status(void)
594 if (simix_global->process_list == NULL) {
598 smx_process_t process = NULL;
599 int nbprocess = xbt_swag_size(simix_global->process_list);
601 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
602 /* List the process and their state */
604 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
605 xbt_swag_foreach(process, simix_global->process_list) {
607 if (process->waiting_synchro) {
609 const char* synchro_description = "unknown";
610 switch (process->waiting_synchro->type) {
612 case SIMIX_SYNC_EXECUTE:
613 synchro_description = "execution";
616 case SIMIX_SYNC_PARALLEL_EXECUTE:
617 synchro_description = "parallel execution";
620 case SIMIX_SYNC_COMMUNICATE:
621 synchro_description = "communication";
624 case SIMIX_SYNC_SLEEP:
625 synchro_description = "sleeping";
628 case SIMIX_SYNC_JOIN:
629 synchro_description = "joining";
632 case SIMIX_SYNC_SYNCHRO:
633 synchro_description = "synchronization";
637 synchro_description = "I/O";
640 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
641 process->pid, process->name, sg_host_get_name(process->host),
642 synchro_description, process->waiting_synchro,
643 process->waiting_synchro->name, (int)process->waiting_synchro->state);
646 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
651 static void* SIMIX_synchro_mallocator_new_f(void) {
652 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
653 synchro->simcalls = xbt_fifo_new();
657 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
658 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
662 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
664 // we also recycle the simcall list
665 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
666 xbt_fifo_reset(fifo);
667 memset(synchro, 0, sizeof(s_smx_synchro_t));
668 ((smx_synchro_t) synchro)->simcalls = fifo;
671 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
672 return SIMIX_asr_get_properties(name);
674 xbt_dict_t SIMIX_asr_get_properties(const char *name)
676 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
679 int SIMIX_is_maestro()
681 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;