1 /* Copyright (c) 2007-2014. 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 "smx_private.h"
9 #include "xbt/sysdep.h"
12 #include "xbt/ex.h" /* ex_backtrace_display */
14 #include "simgrid/sg_config.h"
16 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
18 "Logging specific to SIMIX (kernel)");
20 smx_global_t simix_global = NULL;
21 static xbt_heap_t simix_timers = NULL;
23 static void* SIMIX_action_mallocator_new_f(void);
24 static void SIMIX_action_mallocator_free_f(void* action);
25 static void SIMIX_action_mallocator_reset_f(void* action);
27 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
30 int _sg_do_verbose_exit = 1;
31 static void _XBT_CALL inthandler(int ignored)
33 if ( _sg_do_verbose_exit ) {
34 XBT_INFO("CTRL-C pressed. Displaying status and bailing out");
35 SIMIX_display_process_status();
38 XBT_INFO("CTRL-C pressed. bailing out without displaying because verbose-exit is disabled");
43 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
45 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
47 "Access violation detected.\n"
48 "This can result from a programming error in your code or, although less likely,\n"
49 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
50 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
52 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
53 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
54 smx_context_stack_size / 1024);
55 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
57 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
58 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
64 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
67 * Install signal handler for SIGSEGV. Check that nobody has already installed
68 * its own handler. For example, the Java VM does this.
70 static void install_segvhandler(void)
72 stack_t stack, old_stack;
73 stack.ss_sp = sigsegv_stack;
74 stack.ss_size = sizeof sigsegv_stack;
77 if (sigaltstack(&stack, &old_stack) == -1) {
78 XBT_WARN("Failed to register alternate signal stack: %s",
82 if (!(old_stack.ss_flags & SS_DISABLE)) {
83 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
84 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
85 sigaltstack(&old_stack, NULL);
88 struct sigaction action, old_action;
89 action.sa_sigaction = segvhandler;
90 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
91 sigemptyset(&action.sa_mask);
93 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
94 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
98 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
99 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
100 (old_action.sa_flags & SA_SIGINFO) ?
101 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
102 sigaction(SIGSEGV, &old_action, NULL);
106 /********************************* SIMIX **************************************/
108 XBT_INLINE double SIMIX_timer_next(void)
110 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
115 * \brief Initialize SIMIX internal data.
120 void SIMIX_global_init(int *argc, char **argv)
122 s_smx_process_t proc;
125 simix_global = xbt_new0(s_smx_global_t, 1);
127 #ifdef TIME_BENCH_AMDAHL
128 simix_global->timer_seq = xbt_os_timer_new();
129 simix_global->timer_par = xbt_os_timer_new();
130 xbt_os_cputimer_start(simix_global->timer_seq);
132 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
133 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
134 simix_global->process_list =
135 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
136 simix_global->process_to_destroy =
137 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
139 simix_global->maestro_process = NULL;
140 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
142 simix_global->create_process_function = SIMIX_process_create;
143 simix_global->kill_process_function = SIMIX_process_kill;
144 simix_global->cleanup_process_function = SIMIX_process_cleanup;
145 simix_global->action_mallocator = xbt_mallocator_new(65536,
146 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
147 SIMIX_action_mallocator_reset_f);
148 simix_global->autorestart = SIMIX_host_restart_processes;
150 surf_init(argc, argv); /* Initialize SURF structures */
151 SIMIX_context_mod_init();
152 SIMIX_create_maestro_process();
154 /* context exception handlers */
155 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
156 __xbt_ex_terminate = SIMIX_process_exception_terminate;
158 SIMIX_network_init();
160 /* Prepare to display some more info when dying on Ctrl-C pressing */
161 signal(SIGINT, inthandler);
163 /* Install SEGV handler */
164 install_segvhandler();
166 /* register a function to be called by SURF after the environment creation */
168 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
172 simix_timers = xbt_heap_new(8, &free);
175 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
176 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
178 if (sg_cfg_get_boolean("clean_atexit"))
181 if (_sg_cfg_exit_asap)
187 * \brief Clean the SIMIX simulation
189 * This functions remove the memory used by SIMIX
192 void SIMIX_clean(void)
194 #ifdef TIME_BENCH_PER_SR
195 smx_ctx_raw_new_sr();
197 if (cleaned) return; // to avoid double cleaning by java and C
199 /* Kill everyone (except maestro) */
200 SIMIX_process_killall(simix_global->maestro_process, 1);
202 /* Exit the SIMIX network module */
203 SIMIX_network_exit();
205 xbt_heap_free(simix_timers);
207 /* Free the remaining data structures */
208 xbt_dynar_free(&simix_global->process_to_run);
209 xbt_dynar_free(&simix_global->process_that_ran);
210 xbt_swag_free(simix_global->process_to_destroy);
211 xbt_swag_free(simix_global->process_list);
212 simix_global->process_list = NULL;
213 simix_global->process_to_destroy = NULL;
214 xbt_dict_free(&(simix_global->registered_functions));
216 /* Let's free maestro now */
217 SIMIX_context_free(simix_global->maestro_process->context);
218 xbt_free(simix_global->maestro_process->running_ctx);
219 xbt_free(simix_global->maestro_process);
220 simix_global->maestro_process = NULL;
222 /* Restore the default exception setup */
223 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
224 __xbt_ex_terminate = &__xbt_ex_terminate_default;
226 /* Finish context module and SURF */
227 SIMIX_context_mod_exit();
231 #ifdef TIME_BENCH_AMDAHL
232 xbt_os_cputimer_stop(simix_global->timer_seq);
233 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
234 xbt_os_timer_elapsed(simix_global->timer_seq),
235 xbt_os_timer_elapsed(simix_global->timer_par));
236 xbt_os_timer_free(simix_global->timer_seq);
237 xbt_os_timer_free(simix_global->timer_par);
240 xbt_mallocator_free(simix_global->action_mallocator);
241 xbt_free(simix_global);
250 * \brief A clock (in second).
252 * \return Return the clock.
254 XBT_INLINE double SIMIX_get_clock(void)
257 return MC_process_clock_get(SIMIX_process_self());
259 return surf_get_clock();
263 static int process_syscall_color(void *p)
265 switch ((*(smx_process_t *)p)->simcall.call) {
267 case SIMCALL_PROCESS_KILL:
269 case SIMCALL_PROCESS_RESUME:
278 * \brief Run the main simulation loop.
283 smx_process_t process;
284 surf_action_t action;
290 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
291 xbt_dynar_length(simix_global->process_to_run));
292 #ifdef TIME_BENCH_PER_SR
293 smx_ctx_raw_new_sr();
295 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
296 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
297 xbt_dynar_length(simix_global->process_to_run));
299 /* Run all processes that are ready to run, possibly in parallel */
300 #ifdef TIME_BENCH_AMDAHL
301 xbt_os_cputimer_stop(simix_global->timer_seq);
302 xbt_os_cputimer_resume(simix_global->timer_par);
304 SIMIX_process_runall();
305 #ifdef TIME_BENCH_AMDAHL
306 xbt_os_cputimer_stop(simix_global->timer_par);
307 xbt_os_cputimer_resume(simix_global->timer_seq);
310 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
311 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
313 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
315 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
317 /* Here, the order is ok because:
319 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
321 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
323 * - if there is no kill during the simulation, processes remain sorted according by their PID.
324 * rational: This can be proved inductively.
325 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
326 * Let's show that it is still so at the end of this round.
327 * - if a process is added when being created, that's from maestro. It can be either at startup
328 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
329 * in arbitrary order (inductive hypothesis), we are fine.
330 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
331 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
332 * of a simcall terminates, we're still good. Proof:
333 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
334 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
335 * - For communications, this function is called from SIMIX_comm_finish().
336 * This function itself don't mess with the order since simcalls are handled in FIFO order.
337 * The function is called:
338 * - before the comm starts (invalid parameters, or resource already dead or whatever).
339 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
340 * - because the communication failed or were canceled after startup. In this case, it's called from the function
341 * we are in, by the chunk:
342 * set = model->states.failed_action_set;
343 * while ((action = xbt_swag_extract(set)))
344 * SIMIX_simcall_post((smx_action_t) action->data);
345 * This order is also fixed because it depends of the order in which the surf actions were
346 * added to the system, and only maestro can add stuff this way, through simcalls.
347 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
348 * poped out of the swag does not depend on the user code's execution order.
349 * - because the communication terminated. In this case, actions are served in the order given by
350 * set = model->states.done_action_set;
351 * while ((action = xbt_swag_extract(set)))
352 * SIMIX_simcall_post((smx_action_t) action->data);
353 * and the argument is very similar to the previous one.
354 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
355 * 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.
356 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
357 * - If there is some process killings, the order is changed by this decision that comes from user-land
358 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
359 * So, even the order change induced by the process killing is perfectly reproducible.
361 * So science works, bitches [http://xkcd.com/54/].
363 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
364 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
365 * That would thus be a pure waste of time.
368 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
369 if (process->simcall.call != SIMCALL_NONE) {
370 SIMIX_simcall_pre(&process->simcall, 0);
375 time = SIMIX_timer_next();
376 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
377 time = surf_solve(time);
379 /* Notify all the hosts that have failed */
380 /* FIXME: iterate through the list of failed host and mark each of them */
381 /* as failed. On each host, signal all the running processes with host_fail */
383 /* Handle any pending timer */
384 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
385 //FIXME: make the timers being real callbacks
386 // (i.e. provide dispatchers that read and expand the args)
387 timer = xbt_heap_pop(simix_timers);
389 ((void (*)(void*))timer->func)(timer->args);
393 /* Wake up all processes waiting for a Surf action to finish */
394 xbt_dynar_foreach(model_list, iter, model) {
395 while ((action = surf_model_extract_failed_action_set(model)))
396 SIMIX_simcall_post((smx_action_t) surf_action_get_data(action));
398 while ((action = surf_model_extract_done_action_set(model)))
399 if (surf_action_get_data(action) == NULL)
400 XBT_DEBUG("probably vcpu's action %p, skip", action);
402 SIMIX_simcall_post((smx_action_t) surf_action_get_data(action));
405 /* Autorestart all process */
406 if(host_that_restart) {
407 char *hostname = NULL;
408 xbt_dynar_foreach(host_that_restart,iter,hostname) {
409 XBT_INFO("Restart processes on host: %s",hostname);
410 SIMIX_host_autorestart(SIMIX_host_get_by_name(hostname));
412 xbt_dynar_reset(host_that_restart);
415 /* Clean processes to destroy */
416 SIMIX_process_empty_trash();
419 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
420 // !(time == -1.0 && xbt_dynar_is_empty())
423 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
425 if (xbt_swag_size(simix_global->process_list) != 0) {
431 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
432 SIMIX_display_process_status();
438 * \brief Set the date to execute a function
440 * Set the date to execute the function on the surf.
441 * \param date Date to execute function
442 * \param function Function to be executed
443 * \param arg Parameters of the function
446 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
448 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
451 timer->func = function;
453 xbt_heap_push(simix_timers, timer, date);
457 * \brief Registers a function to create a process.
459 * This function registers a function to be called
460 * when a new process is created. The function has
461 * to call SIMIX_process_create().
462 * \param function create process function
464 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
467 simix_global->create_process_function = function;
471 * \brief Registers a function to kill a process.
473 * This function registers a function to be called when a
474 * process is killed. The function has to call the SIMIX_process_kill().
476 * \param function Kill process function
478 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
481 simix_global->kill_process_function = function;
485 * \brief Registers a function to cleanup a process.
487 * This function registers a user function to be called when
488 * a process ends properly.
490 * \param function cleanup process function
492 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
495 simix_global->cleanup_process_function = function;
499 void SIMIX_display_process_status(void)
501 if (simix_global->process_list == NULL) {
505 smx_process_t process = NULL;
506 int nbprocess = xbt_swag_size(simix_global->process_list);
508 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
509 /* List the process and their state */
511 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
512 xbt_swag_foreach(process, simix_global->process_list) {
514 if (process->waiting_action) {
516 const char* action_description = "unknown";
517 switch (process->waiting_action->type) {
519 case SIMIX_ACTION_EXECUTE:
520 action_description = "execution";
523 case SIMIX_ACTION_PARALLEL_EXECUTE:
524 action_description = "parallel execution";
527 case SIMIX_ACTION_COMMUNICATE:
528 action_description = "communication";
531 case SIMIX_ACTION_SLEEP:
532 action_description = "sleeping";
535 case SIMIX_ACTION_SYNCHRO:
536 action_description = "synchronization";
539 case SIMIX_ACTION_IO:
540 action_description = "I/O";
542 /* **************************************/
543 /* TUTORIAL: New API */
544 case SIMIX_ACTION_NEW_API:
545 action_description = "NEW API";
546 /* **************************************/
550 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
551 process->pid, process->name, sg_host_name(process->smx_host),
552 action_description, process->waiting_action,
553 process->waiting_action->name, (int)process->waiting_action->state);
556 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host));
561 static void* SIMIX_action_mallocator_new_f(void) {
562 smx_action_t action = xbt_new(s_smx_action_t, 1);
563 action->simcalls = xbt_fifo_new();
567 static void SIMIX_action_mallocator_free_f(void* action) {
568 xbt_fifo_free(((smx_action_t) action)->simcalls);
572 static void SIMIX_action_mallocator_reset_f(void* action) {
574 // we also recycle the simcall list
575 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
576 xbt_fifo_reset(fifo);
577 memset(action, 0, sizeof(s_smx_action_t));
578 ((smx_action_t) action)->simcalls = fifo;
581 xbt_dict_t SIMIX_pre_asr_get_properties(smx_simcall_t simcall, const char *name){
582 return SIMIX_asr_get_properties(name);
584 xbt_dict_t SIMIX_asr_get_properties(const char *name)
586 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);