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 "smx_private.h"
16 #include "smx_private.hpp"
18 #include "xbt/sysdep.h"
21 #include "xbt/ex.h" /* ex_backtrace_display */
23 #include "src/mc/mc_replay.h"
24 #include "simgrid/sg_config.h"
26 #include "src/surf/callbacks.h"
29 #include "src/mc/mc_private.h"
30 #include "src/mc/mc_protocol.h"
31 #include "src/mc/mc_client.h"
36 #include "src/mc/mc_protocol.h"
39 #include "src/mc/mc_record.h"
42 #include "src/smpi/private.h"
45 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
46 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
47 "Logging specific to SIMIX (kernel)");
49 smx_global_t simix_global = NULL;
50 static xbt_heap_t simix_timers = NULL;
52 /** @brief Timer datatype */
53 typedef struct s_smx_timer {
59 void (*SMPI_switch_data_segment)(int) = NULL;
61 static void* SIMIX_synchro_mallocator_new_f(void);
62 static void SIMIX_synchro_mallocator_free_f(void* synchro);
63 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
65 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
68 int _sg_do_verbose_exit = 1;
69 static void _XBT_CALL inthandler(int ignored)
71 if ( _sg_do_verbose_exit ) {
72 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
73 SIMIX_display_process_status();
76 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
82 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
84 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
86 "Access violation detected.\n"
87 "This can result from a programming error in your code or, although less likely,\n"
88 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
89 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
91 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
92 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
93 smx_context_stack_size / 1024);
94 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
96 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
97 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
99 } else if (siginfo->si_signo == SIGSEGV) {
100 fprintf(stderr, "Segmentation fault.\n");
102 if (smpi_enabled() && !smpi_privatize_global_variables) {
104 "Try to enable SMPI variable privatization with --cfg=smpi/privatize_global_variables:yes.\n");
111 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
114 * Install signal handler for SIGSEGV. Check that nobody has already installed
115 * its own handler. For example, the Java VM does this.
117 static void install_segvhandler(void)
119 stack_t stack, old_stack;
120 stack.ss_sp = sigsegv_stack;
121 stack.ss_size = sizeof sigsegv_stack;
124 if (sigaltstack(&stack, &old_stack) == -1) {
125 XBT_WARN("Failed to register alternate signal stack: %s",
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",
145 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
146 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
147 (old_action.sa_flags & SA_SIGINFO) ?
148 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
149 sigaction(SIGSEGV, &old_action, NULL);
154 /********************************* SIMIX **************************************/
156 double SIMIX_timer_next(void)
158 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
161 static void kill_process(smx_process_t process)
163 SIMIX_process_kill(process, NULL);
166 static void SIMIX_storage_create_(smx_storage_t storage)
168 const char* key = xbt_dict_get_elm_key(storage);
169 SIMIX_storage_create(key, storage, NULL);
174 * \brief Initialize SIMIX internal data.
179 void SIMIX_global_init(int *argc, char **argv)
182 _sg_do_model_check = getenv(MC_ENV_VARIABLE) != NULL;
185 s_smx_process_t proc;
188 simix_global = xbt_new0(s_smx_global_t, 1);
190 #ifdef TIME_BENCH_AMDAHL
191 simix_global->timer_seq = xbt_os_timer_new();
192 simix_global->timer_par = xbt_os_timer_new();
193 xbt_os_cputimer_start(simix_global->timer_seq);
195 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
196 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
197 simix_global->process_list =
198 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
199 simix_global->process_to_destroy =
200 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
202 simix_global->maestro_process = NULL;
203 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
205 simix_global->create_process_function = SIMIX_process_create;
206 simix_global->kill_process_function = kill_process;
207 simix_global->cleanup_process_function = SIMIX_process_cleanup;
208 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
209 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
210 SIMIX_synchro_mallocator_reset_f);
211 simix_global->autorestart = SIMIX_host_restart_processes;
212 simix_global->mutex = xbt_os_mutex_init();
214 surf_init(argc, argv); /* Initialize SURF structures */
215 SIMIX_context_mod_init();
216 SIMIX_create_maestro_process();
218 /* context exception handlers */
219 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
220 __xbt_ex_terminate = SIMIX_process_exception_terminate;
222 SIMIX_network_init();
224 /* Prepare to display some more info when dying on Ctrl-C pressing */
225 signal(SIGINT, inthandler);
228 /* Install SEGV handler */
229 install_segvhandler();
231 /* register a function to be called by SURF after the environment creation */
233 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
234 surf_on_host_created(SIMIX_host_create);
235 surf_on_storage_created(SIMIX_storage_create_);
239 simix_timers = xbt_heap_new(8, &free);
242 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
244 if (sg_cfg_get_boolean("clean_atexit"))
248 // The communication initialization is done ASAP.
249 // We need to communicate initialization of the different layers to the model-checker.
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 #ifdef TIME_BENCH_PER_SR
267 smx_ctx_raw_new_sr();
269 if (smx_cleaned) return; // to avoid double cleaning by java and C
271 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
272 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
274 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
275 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
276 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
278 /* Kill all processes (but maestro) */
279 SIMIX_process_killall(simix_global->maestro_process, 1);
281 /* Exit the SIMIX network module */
282 SIMIX_network_exit();
284 xbt_heap_free(simix_timers);
286 /* Free the remaining data structures */
287 xbt_dynar_free(&simix_global->process_to_run);
288 xbt_dynar_free(&simix_global->process_that_ran);
289 xbt_swag_free(simix_global->process_to_destroy);
290 xbt_swag_free(simix_global->process_list);
291 simix_global->process_list = NULL;
292 simix_global->process_to_destroy = NULL;
293 xbt_dict_free(&(simix_global->registered_functions));
295 xbt_os_mutex_destroy(simix_global->mutex);
296 simix_global->mutex = NULL;
298 /* Let's free maestro now */
299 SIMIX_context_free(simix_global->maestro_process->context);
300 xbt_free(simix_global->maestro_process->running_ctx);
301 xbt_free(simix_global->maestro_process);
302 simix_global->maestro_process = NULL;
304 /* Restore the default exception setup */
305 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
306 __xbt_ex_terminate = &__xbt_ex_terminate_default;
308 /* Finish context module and SURF */
309 SIMIX_context_mod_exit();
313 #ifdef TIME_BENCH_AMDAHL
314 xbt_os_cputimer_stop(simix_global->timer_seq);
315 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
316 xbt_os_timer_elapsed(simix_global->timer_seq),
317 xbt_os_timer_elapsed(simix_global->timer_par));
318 xbt_os_timer_free(simix_global->timer_seq);
319 xbt_os_timer_free(simix_global->timer_par);
322 xbt_mallocator_free(simix_global->synchro_mallocator);
323 xbt_free(simix_global);
332 * \brief A clock (in second).
334 * \return Return the clock.
336 double SIMIX_get_clock(void)
338 if(MC_is_active() || MC_record_replay_is_active()){
339 return MC_process_clock_get(SIMIX_process_self());
341 return surf_get_clock();
345 static int process_syscall_color(void *p)
347 switch ((*(smx_process_t *)p)->simcall.call) {
349 case SIMCALL_PROCESS_KILL:
351 case SIMCALL_PROCESS_RESUME:
360 * \brief Run the main simulation loop.
365 MC_record_replay_init();
366 MC_record_replay_from_string(MC_record_path);
371 smx_process_t process;
372 surf_action_t action;
378 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
379 xbt_dynar_length(simix_global->process_to_run));
380 #ifdef TIME_BENCH_PER_SR
381 smx_ctx_raw_new_sr();
383 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
384 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
385 xbt_dynar_length(simix_global->process_to_run));
387 /* Run all processes that are ready to run, possibly in parallel */
388 #ifdef TIME_BENCH_AMDAHL
389 xbt_os_cputimer_stop(simix_global->timer_seq);
390 xbt_os_cputimer_resume(simix_global->timer_par);
392 SIMIX_process_runall();
393 #ifdef TIME_BENCH_AMDAHL
394 xbt_os_cputimer_stop(simix_global->timer_par);
395 xbt_os_cputimer_resume(simix_global->timer_seq);
398 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
399 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
401 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
403 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
405 /* Here, the order is ok because:
407 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
409 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
411 * - if there is no kill during the simulation, processes remain sorted according by their PID.
412 * rational: This can be proved inductively.
413 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
414 * Let's show that it is still so at the end of this round.
415 * - if a process is added when being created, that's from maestro. It can be either at startup
416 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
417 * in arbitrary order (inductive hypothesis), we are fine.
418 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
419 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
420 * of a simcall terminates, we're still good. Proof:
421 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
422 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
423 * - For communications, this function is called from SIMIX_comm_finish().
424 * This function itself don't mess with the order since simcalls are handled in FIFO order.
425 * The function is called:
426 * - before the comm starts (invalid parameters, or resource already dead or whatever).
427 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
428 * - because the communication failed or were canceled after startup. In this case, it's called from the function
429 * we are in, by the chunk:
430 * set = model->states.failed_action_set;
431 * while ((synchro = xbt_swag_extract(set)))
432 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
433 * This order is also fixed because it depends of the order in which the surf actions were
434 * added to the system, and only maestro can add stuff this way, through simcalls.
435 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
436 * poped out of the swag does not depend on the user code's execution order.
437 * - because the communication terminated. In this case, synchros are served in the order given by
438 * set = model->states.done_action_set;
439 * while ((synchro = xbt_swag_extract(set)))
440 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
441 * and the argument is very similar to the previous one.
442 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
443 * 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.
444 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
445 * - If there is some process killings, the order is changed by this decision that comes from user-land
446 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
447 * So, even the order change induced by the process killing is perfectly reproducible.
449 * So science works, bitches [http://xkcd.com/54/].
451 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
452 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
453 * That would thus be a pure waste of time.
456 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
457 if (process->simcall.call != SIMCALL_NONE) {
458 SIMIX_simcall_handle(&process->simcall, 0);
461 /* Wake up all processes waiting for a Surf action to finish */
462 xbt_dynar_foreach(all_existing_models, iter, model) {
463 XBT_DEBUG("Handling process whose action failed");
464 while ((action = surf_model_extract_failed_action_set(model))) {
465 XBT_DEBUG(" Handling Action %p",action);
466 SIMIX_simcall_exit((smx_synchro_t) action->getData());
468 XBT_DEBUG("Handling process whose action terminated normally");
469 while ((action = surf_model_extract_done_action_set(model))) {
470 XBT_DEBUG(" Handling Action %p",action);
471 if (action->getData() == NULL)
472 XBT_DEBUG("probably vcpu's action %p, skip", action);
474 SIMIX_simcall_exit((smx_synchro_t) action->getData());
479 time = SIMIX_timer_next();
480 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
481 XBT_DEBUG("Calling surf_solve");
482 time = surf_solve(time);
483 XBT_DEBUG("Moving time ahead : %g", time);
485 /* Notify all the hosts that have failed */
486 /* FIXME: iterate through the list of failed host and mark each of them */
487 /* as failed. On each host, signal all the running processes with host_fail */
489 /* Handle any pending timer */
490 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
491 //FIXME: make the timers being real callbacks
492 // (i.e. provide dispatchers that read and expand the args)
493 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
495 timer->func(timer->args);
499 /* Wake up all processes waiting for a Surf action to finish */
500 xbt_dynar_foreach(all_existing_models, iter, model) {
501 XBT_DEBUG("Handling process whose action failed");
502 while ((action = surf_model_extract_failed_action_set(model))) {
503 XBT_DEBUG(" Handling Action %p",action);
504 SIMIX_simcall_exit((smx_synchro_t) action->getData());
506 XBT_DEBUG("Handling process whose action terminated normally");
507 while ((action = surf_model_extract_done_action_set(model))) {
508 XBT_DEBUG(" Handling Action %p",action);
509 if (action->getData() == NULL)
510 XBT_DEBUG("probably vcpu's action %p, skip", action);
512 SIMIX_simcall_exit((smx_synchro_t) action->getData());
516 /* Autorestart all process */
517 if(host_that_restart) {
518 char *hostname = NULL;
519 xbt_dynar_foreach(host_that_restart,iter,hostname) {
520 XBT_INFO("Restart processes on host: %s",hostname);
521 SIMIX_host_autorestart(sg_host_by_name(hostname));
523 xbt_dynar_reset(host_that_restart);
526 /* Clean processes to destroy */
527 SIMIX_process_empty_trash();
530 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
532 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
534 if (xbt_swag_size(simix_global->process_list) != 0) {
538 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
539 SIMIX_display_process_status();
545 * \brief Set the date to execute a function
547 * Set the date to execute the function on the surf.
548 * \param date Date to execute function
549 * \param function Function to be executed
550 * \param arg Parameters of the function
553 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
555 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
558 timer->func = function;
560 xbt_heap_push(simix_timers, timer, date);
563 /** @brief cancels a timer that was added earlier */
564 void SIMIX_timer_remove(smx_timer_t timer) {
565 xbt_heap_rm_elm(simix_timers, timer, timer->date);
568 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
569 double SIMIX_timer_get_date(smx_timer_t timer) {
570 return timer?timer->date:0;
574 * \brief Registers a function to create a process.
576 * This function registers a function to be called
577 * when a new process is created. The function has
578 * to call SIMIX_process_create().
579 * \param function create process function
581 void SIMIX_function_register_process_create(smx_creation_func_t
584 simix_global->create_process_function = function;
588 * \brief Registers a function to kill a process.
590 * This function registers a function to be called when a
591 * process is killed. The function has to call the SIMIX_process_kill().
593 * \param function Kill process function
595 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
598 simix_global->kill_process_function = function;
602 * \brief Registers a function to cleanup a process.
604 * This function registers a user function to be called when
605 * a process ends properly.
607 * \param function cleanup process function
609 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
612 simix_global->cleanup_process_function = function;
616 void SIMIX_display_process_status(void)
618 if (simix_global->process_list == NULL) {
622 smx_process_t process = NULL;
623 int nbprocess = xbt_swag_size(simix_global->process_list);
625 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
626 /* List the process and their state */
628 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
629 xbt_swag_foreach(process, simix_global->process_list) {
631 if (process->waiting_synchro) {
633 const char* synchro_description = "unknown";
634 switch (process->waiting_synchro->type) {
636 case SIMIX_SYNC_EXECUTE:
637 synchro_description = "execution";
640 case SIMIX_SYNC_PARALLEL_EXECUTE:
641 synchro_description = "parallel execution";
644 case SIMIX_SYNC_COMMUNICATE:
645 synchro_description = "communication";
648 case SIMIX_SYNC_SLEEP:
649 synchro_description = "sleeping";
652 case SIMIX_SYNC_JOIN:
653 synchro_description = "joining";
656 case SIMIX_SYNC_SYNCHRO:
657 synchro_description = "synchronization";
661 synchro_description = "I/O";
664 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
665 process->pid, process->name, sg_host_get_name(process->host),
666 synchro_description, process->waiting_synchro,
667 process->waiting_synchro->name, (int)process->waiting_synchro->state);
670 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
675 static void* SIMIX_synchro_mallocator_new_f(void) {
676 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
677 synchro->simcalls = xbt_fifo_new();
681 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
682 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
686 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
688 // we also recycle the simcall list
689 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
690 xbt_fifo_reset(fifo);
691 memset(synchro, 0, sizeof(s_smx_synchro_t));
692 ((smx_synchro_t) synchro)->simcalls = fifo;
695 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
696 return SIMIX_asr_get_properties(name);
698 xbt_dict_t SIMIX_asr_get_properties(const char *name)
700 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
703 int SIMIX_is_maestro()
705 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;