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. */
8 #include "src/portable.h"
9 #ifdef HAVE_SYS_PTRACE_H
10 # include <sys/types.h>
11 # include <sys/ptrace.h>
14 #include "src/surf/surf_interface.hpp"
15 #include "src/surf/storage_interface.hpp"
16 #include "src/surf/xml/platf.hpp"
17 #include "smx_private.h"
18 #include "smx_private.hpp"
20 #include "xbt/ex.h" /* ex_backtrace_display */
22 #include "src/mc/mc_replay.h"
23 #include "simgrid/sg_config.h"
26 #include "src/mc/mc_private.h"
27 #include "src/mc/mc_protocol.h"
28 #include "src/mc/mc_client.h"
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 static void* SIMIX_synchro_mallocator_new_f(void);
58 static void SIMIX_synchro_mallocator_free_f(void* synchro);
59 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
61 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
64 int _sg_do_verbose_exit = 1;
65 static void inthandler(int ignored)
67 if ( _sg_do_verbose_exit ) {
68 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
69 SIMIX_display_process_status();
72 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
78 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
80 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
82 "Access violation detected.\n"
83 "This can result from a programming error in your code or, although less likely,\n"
84 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
85 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
87 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
88 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
89 smx_context_stack_size / 1024);
90 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
92 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
93 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
95 } else if (siginfo->si_signo == SIGSEGV) {
96 fprintf(stderr, "Segmentation fault.\n");
98 if (smpi_enabled() && !smpi_privatize_global_variables) {
99 #ifdef HAVE_PRIVATIZATION
101 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
104 "Sadly, your system does not support --cfg=smpi/privatize_global_variables:yes (yet).\n");
105 #endif /* HAVE_PRIVATIZATION */
107 #endif /* HAVE_SMPI */
112 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
115 * Install signal handler for SIGSEGV. Check that nobody has already installed
116 * its own handler. For example, the Java VM does this.
118 static void install_segvhandler(void)
120 stack_t stack, old_stack;
121 stack.ss_sp = sigsegv_stack;
122 stack.ss_size = sizeof sigsegv_stack;
125 if (sigaltstack(&stack, &old_stack) == -1) {
126 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
129 if (!(old_stack.ss_flags & SS_DISABLE)) {
130 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
131 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
132 sigaltstack(&old_stack, NULL);
135 struct sigaction action, old_action;
136 action.sa_sigaction = segvhandler;
137 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
138 sigemptyset(&action.sa_mask);
140 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
141 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
144 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
145 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
146 (old_action.sa_flags & SA_SIGINFO) ?
147 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
148 sigaction(SIGSEGV, &old_action, NULL);
153 /********************************* SIMIX **************************************/
155 double SIMIX_timer_next(void)
157 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
160 static void kill_process(smx_process_t process)
162 SIMIX_process_kill(process, NULL);
165 static void SIMIX_storage_create_(smx_storage_t storage)
167 const char* key = xbt_dict_get_elm_key(storage);
168 SIMIX_storage_create(key, storage, NULL);
171 static std::function<void()> maestro_code;
176 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
178 maestro_code = std::move(code);
184 void SIMIX_set_maestro(void (*code)(void*), void* data)
186 maestro_code = std::bind(code, data);
191 * \brief Initialize SIMIX internal data.
196 void SIMIX_global_init(int *argc, char **argv)
199 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
202 s_smx_process_t proc;
205 simix_global = xbt_new0(s_smx_global_t, 1);
207 #ifdef TIME_BENCH_AMDAHL
208 simix_global->timer_seq = xbt_os_timer_new();
209 simix_global->timer_par = xbt_os_timer_new();
210 xbt_os_cputimer_start(simix_global->timer_seq);
212 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
213 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
214 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
215 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
217 simix_global->maestro_process = NULL;
218 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
220 simix_global->create_process_function = SIMIX_process_create;
221 simix_global->kill_process_function = kill_process;
222 simix_global->cleanup_process_function = SIMIX_process_cleanup;
223 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
224 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
225 SIMIX_synchro_mallocator_reset_f);
226 simix_global->mutex = xbt_os_mutex_init();
228 surf_init(argc, argv); /* Initialize SURF structures */
229 SIMIX_context_mod_init();
231 // Either create a new context with maestro or create
232 // a context object with the current context mestro):
233 simgrid::simix::create_maestro(maestro_code);
235 /* context exception handlers */
236 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
237 __xbt_ex_terminate = SIMIX_process_exception_terminate;
239 SIMIX_network_init();
241 /* Prepare to display some more info when dying on Ctrl-C pressing */
242 signal(SIGINT, inthandler);
245 install_segvhandler();
247 /* register a function to be called by SURF after the environment creation */
249 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
250 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
251 SIMIX_host_create(&host);
253 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
255 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
256 const char* id = storage->getName();
257 // TODO, create sg_storage_by_name
258 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, id);
259 xbt_assert(s != NULL, "Storage not found for name %s", id);
260 SIMIX_storage_create_(s);
263 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
266 simix_timers = xbt_heap_new(8, &free);
269 if (sg_cfg_get_boolean("clean_atexit"))
273 // The communication initialization is done ASAP.
274 // We need to communicate initialization of the different layers to the model-checker.
278 if (_sg_cfg_exit_asap)
285 * \brief Clean the SIMIX simulation
287 * This functions remove the memory used by SIMIX
289 void SIMIX_clean(void)
291 #ifdef TIME_BENCH_PER_SR
292 smx_ctx_raw_new_sr();
294 if (smx_cleaned) return; // to avoid double cleaning by java and C
296 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
297 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
299 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
300 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
301 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
303 /* Kill all processes (but maestro) */
304 SIMIX_process_killall(simix_global->maestro_process, 1);
306 /* Exit the SIMIX network module */
307 SIMIX_network_exit();
309 xbt_heap_free(simix_timers);
311 /* Free the remaining data structures */
312 xbt_dynar_free(&simix_global->process_to_run);
313 xbt_dynar_free(&simix_global->process_that_ran);
314 xbt_swag_free(simix_global->process_to_destroy);
315 xbt_swag_free(simix_global->process_list);
316 simix_global->process_list = NULL;
317 simix_global->process_to_destroy = NULL;
318 xbt_dict_free(&(simix_global->registered_functions));
320 xbt_os_mutex_destroy(simix_global->mutex);
321 simix_global->mutex = NULL;
323 /* Let's free maestro now */
324 SIMIX_context_free(simix_global->maestro_process->context);
325 xbt_free(simix_global->maestro_process->running_ctx);
326 xbt_free(simix_global->maestro_process);
327 simix_global->maestro_process = NULL;
329 /* Restore the default exception setup */
330 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
331 __xbt_ex_terminate = &__xbt_ex_terminate_default;
333 /* Finish context module and SURF */
334 SIMIX_context_mod_exit();
338 #ifdef TIME_BENCH_AMDAHL
339 xbt_os_cputimer_stop(simix_global->timer_seq);
340 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
341 xbt_os_timer_elapsed(simix_global->timer_seq),
342 xbt_os_timer_elapsed(simix_global->timer_par));
343 xbt_os_timer_free(simix_global->timer_seq);
344 xbt_os_timer_free(simix_global->timer_par);
347 xbt_mallocator_free(simix_global->synchro_mallocator);
348 xbt_free(simix_global);
357 * \brief A clock (in second).
359 * \return Return the clock.
361 double SIMIX_get_clock(void)
363 if(MC_is_active() || MC_record_replay_is_active()){
364 return MC_process_clock_get(SIMIX_process_self());
366 return surf_get_clock();
370 static int process_syscall_color(void *p)
372 switch ((*(smx_process_t *)p)->simcall.call) {
374 case SIMCALL_PROCESS_KILL:
376 case SIMCALL_PROCESS_RESUME:
385 * \brief Run the main simulation loop.
390 MC_record_replay_init();
391 MC_record_replay_from_string(MC_record_path);
396 smx_process_t process;
397 surf_action_t action;
403 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
404 xbt_dynar_length(simix_global->process_to_run));
405 #ifdef TIME_BENCH_PER_SR
406 smx_ctx_raw_new_sr();
408 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
409 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
410 xbt_dynar_length(simix_global->process_to_run));
412 /* Run all processes that are ready to run, possibly in parallel */
413 #ifdef TIME_BENCH_AMDAHL
414 xbt_os_cputimer_stop(simix_global->timer_seq);
415 xbt_os_cputimer_resume(simix_global->timer_par);
417 SIMIX_process_runall();
418 #ifdef TIME_BENCH_AMDAHL
419 xbt_os_cputimer_stop(simix_global->timer_par);
420 xbt_os_cputimer_resume(simix_global->timer_seq);
423 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
424 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
426 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
428 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
430 /* Here, the order is ok because:
432 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
434 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
436 * - if there is no kill during the simulation, processes remain sorted according by their PID.
437 * rational: This can be proved inductively.
438 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
439 * Let's show that it is still so at the end of this round.
440 * - if a process is added when being created, that's from maestro. It can be either at startup
441 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
442 * in arbitrary order (inductive hypothesis), we are fine.
443 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
444 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
445 * of a simcall terminates, we're still good. Proof:
446 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
447 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
448 * - For communications, this function is called from SIMIX_comm_finish().
449 * This function itself don't mess with the order since simcalls are handled in FIFO order.
450 * The function is called:
451 * - before the comm starts (invalid parameters, or resource already dead or whatever).
452 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
453 * - because the communication failed or were canceled after startup. In this case, it's called from the function
454 * we are in, by the chunk:
455 * set = model->states.failed_action_set;
456 * while ((synchro = xbt_swag_extract(set)))
457 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
458 * This order is also fixed because it depends of the order in which the surf actions were
459 * added to the system, and only maestro can add stuff this way, through simcalls.
460 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
461 * poped out of the swag does not depend on the user code's execution order.
462 * - because the communication terminated. In this case, synchros are served in the order given by
463 * set = model->states.done_action_set;
464 * while ((synchro = xbt_swag_extract(set)))
465 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
466 * and the argument is very similar to the previous one.
467 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
468 * 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.
469 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
470 * - If there is some process killings, the order is changed by this decision that comes from user-land
471 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
472 * So, even the order change induced by the process killing is perfectly reproducible.
474 * So science works, bitches [http://xkcd.com/54/].
476 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
477 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
478 * That would thus be a pure waste of time.
481 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
482 if (process->simcall.call != SIMCALL_NONE) {
483 SIMIX_simcall_handle(&process->simcall, 0);
486 /* Wake up all processes waiting for a Surf action to finish */
487 xbt_dynar_foreach(all_existing_models, iter, model) {
488 XBT_DEBUG("Handling process whose action failed");
489 while ((action = surf_model_extract_failed_action_set(model))) {
490 XBT_DEBUG(" Handling Action %p",action);
491 SIMIX_simcall_exit((smx_synchro_t) action->getData());
493 XBT_DEBUG("Handling process whose action terminated normally");
494 while ((action = surf_model_extract_done_action_set(model))) {
495 XBT_DEBUG(" Handling Action %p",action);
496 if (action->getData() == NULL)
497 XBT_DEBUG("probably vcpu's action %p, skip", action);
499 SIMIX_simcall_exit((smx_synchro_t) action->getData());
504 time = SIMIX_timer_next();
505 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
506 XBT_DEBUG("Calling surf_solve");
507 time = surf_solve(time);
508 XBT_DEBUG("Moving time ahead : %g", time);
510 /* Notify all the hosts that have failed */
511 /* FIXME: iterate through the list of failed host and mark each of them */
512 /* as failed. On each host, signal all the running processes with host_fail */
514 /* Handle any pending timer */
515 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
516 //FIXME: make the timers being real callbacks
517 // (i.e. provide dispatchers that read and expand the args)
518 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
520 timer->func(timer->args);
524 /* Wake up all processes waiting for a Surf action to finish */
525 xbt_dynar_foreach(all_existing_models, iter, model) {
526 XBT_DEBUG("Handling process whose action failed");
527 while ((action = surf_model_extract_failed_action_set(model))) {
528 XBT_DEBUG(" Handling Action %p",action);
529 SIMIX_simcall_exit((smx_synchro_t) action->getData());
531 XBT_DEBUG("Handling process whose action terminated normally");
532 while ((action = surf_model_extract_done_action_set(model))) {
533 XBT_DEBUG(" Handling Action %p",action);
534 if (action->getData() == NULL)
535 XBT_DEBUG("probably vcpu's action %p, skip", action);
537 SIMIX_simcall_exit((smx_synchro_t) action->getData());
541 /* Autorestart all process */
542 char *hostname = NULL;
543 xbt_dynar_foreach(host_that_restart,iter,hostname) {
544 XBT_INFO("Restart processes on host: %s",hostname);
545 SIMIX_host_autorestart(sg_host_by_name(hostname));
547 xbt_dynar_reset(host_that_restart);
549 /* Clean processes to destroy */
550 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 (*function)(void*), void *arg)
578 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
581 timer->func = function;
583 xbt_heap_push(simix_timers, timer, date);
586 /** @brief cancels a timer that was added earlier */
587 void SIMIX_timer_remove(smx_timer_t timer) {
588 xbt_heap_rm_elm(simix_timers, timer, timer->date);
591 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
592 double SIMIX_timer_get_date(smx_timer_t timer) {
593 return timer?timer->date:0;
597 * \brief Registers a function to create a process.
599 * This function registers a function to be called
600 * when a new process is created. The function has
601 * to call SIMIX_process_create().
602 * \param function create process function
604 void SIMIX_function_register_process_create(smx_creation_func_t
607 simix_global->create_process_function = function;
611 * \brief Registers a function to kill a process.
613 * This function registers a function to be called when a
614 * process is killed. The function has to call the SIMIX_process_kill().
616 * \param function Kill process function
618 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
621 simix_global->kill_process_function = function;
625 * \brief Registers a function to cleanup a process.
627 * This function registers a user function to be called when
628 * a process ends properly.
630 * \param function cleanup process function
632 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
635 simix_global->cleanup_process_function = function;
639 void SIMIX_display_process_status(void)
641 if (simix_global->process_list == NULL) {
645 smx_process_t process = NULL;
646 int nbprocess = xbt_swag_size(simix_global->process_list);
648 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
649 /* List the process and their state */
651 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
652 xbt_swag_foreach(process, simix_global->process_list) {
654 if (process->waiting_synchro) {
656 const char* synchro_description = "unknown";
657 switch (process->waiting_synchro->type) {
659 case SIMIX_SYNC_EXECUTE:
660 synchro_description = "execution";
663 case SIMIX_SYNC_PARALLEL_EXECUTE:
664 synchro_description = "parallel execution";
667 case SIMIX_SYNC_COMMUNICATE:
668 synchro_description = "communication";
671 case SIMIX_SYNC_SLEEP:
672 synchro_description = "sleeping";
675 case SIMIX_SYNC_JOIN:
676 synchro_description = "joining";
679 case SIMIX_SYNC_SYNCHRO:
680 synchro_description = "synchronization";
684 synchro_description = "I/O";
687 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
688 process->pid, process->name, sg_host_get_name(process->host),
689 synchro_description, process->waiting_synchro,
690 process->waiting_synchro->name, (int)process->waiting_synchro->state);
693 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
698 static void* SIMIX_synchro_mallocator_new_f(void) {
699 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
700 synchro->simcalls = xbt_fifo_new();
704 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
705 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
709 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
711 // we also recycle the simcall list
712 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
713 xbt_fifo_reset(fifo);
714 memset(synchro, 0, sizeof(s_smx_synchro_t));
715 ((smx_synchro_t) synchro)->simcalls = fifo;
718 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
719 return SIMIX_asr_get_properties(name);
721 xbt_dict_t SIMIX_asr_get_properties(const char *name)
723 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
726 int SIMIX_is_maestro()
728 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;