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");
44 static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context)
46 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
48 "Access violation detected.\n"
49 "This can result from a programming error in your code or, although less likely,\n"
50 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
51 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
53 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
54 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
55 smx_context_stack_size / 1024);
56 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
58 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
59 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
65 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
68 * Install signal handler for SIGSEGV. Check that nobody has already installed
69 * its own handler. For example, the Java VM does this.
71 static void install_segvhandler(void)
73 stack_t stack, old_stack;
74 stack.ss_sp = sigsegv_stack;
75 stack.ss_size = sizeof sigsegv_stack;
78 if (sigaltstack(&stack, &old_stack) == -1) {
79 XBT_WARN("Failed to register alternate signal stack: %s",
83 if (!(old_stack.ss_flags & SS_DISABLE)) {
84 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
85 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
86 sigaltstack(&old_stack, NULL);
89 struct sigaction action, old_action;
90 action.sa_sigaction = segvhandler;
91 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
92 sigemptyset(&action.sa_mask);
94 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
95 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
99 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
100 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
101 (old_action.sa_flags & SA_SIGINFO) ?
102 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
103 sigaction(SIGSEGV, &old_action, NULL);
108 /********************************* SIMIX **************************************/
110 XBT_INLINE double SIMIX_timer_next(void)
112 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
117 * \brief Initialize SIMIX internal data.
122 void SIMIX_global_init(int *argc, char **argv)
124 s_smx_process_t proc;
127 simix_global = xbt_new0(s_smx_global_t, 1);
129 #ifdef TIME_BENCH_AMDAHL
130 simix_global->timer_seq = xbt_os_timer_new();
131 simix_global->timer_par = xbt_os_timer_new();
132 xbt_os_cputimer_start(simix_global->timer_seq);
134 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
135 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
136 simix_global->process_list =
137 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
138 simix_global->process_to_destroy =
139 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
141 simix_global->maestro_process = NULL;
142 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
144 simix_global->create_process_function = SIMIX_process_create;
145 simix_global->kill_process_function = SIMIX_process_kill;
146 simix_global->cleanup_process_function = SIMIX_process_cleanup;
147 simix_global->action_mallocator = xbt_mallocator_new(65536,
148 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
149 SIMIX_action_mallocator_reset_f);
150 simix_global->autorestart = SIMIX_host_restart_processes;
152 surf_init(argc, argv); /* Initialize SURF structures */
153 SIMIX_context_mod_init();
154 SIMIX_create_maestro_process();
156 /* context exception handlers */
157 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
158 __xbt_ex_terminate = SIMIX_process_exception_terminate;
160 SIMIX_network_init();
162 /* Prepare to display some more info when dying on Ctrl-C pressing */
163 signal(SIGINT, inthandler);
166 /* Install SEGV handler */
167 install_segvhandler();
169 /* register a function to be called by SURF after the environment creation */
171 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
175 simix_timers = xbt_heap_new(8, &free);
178 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
179 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
181 if (sg_cfg_get_boolean("clean_atexit"))
184 if (_sg_cfg_exit_asap)
190 * \brief Clean the SIMIX simulation
192 * This functions remove the memory used by SIMIX
195 void SIMIX_clean(void)
197 #ifdef TIME_BENCH_PER_SR
198 smx_ctx_raw_new_sr();
200 if (cleaned) return; // to avoid double cleaning by java and C
202 /* Kill everyone (except maestro) */
203 SIMIX_process_killall(simix_global->maestro_process, 1);
205 /* Exit the SIMIX network module */
206 SIMIX_network_exit();
208 xbt_heap_free(simix_timers);
210 /* Free the remaining data structures */
211 xbt_dynar_free(&simix_global->process_to_run);
212 xbt_dynar_free(&simix_global->process_that_ran);
213 xbt_swag_free(simix_global->process_to_destroy);
214 xbt_swag_free(simix_global->process_list);
215 simix_global->process_list = NULL;
216 simix_global->process_to_destroy = NULL;
217 xbt_dict_free(&(simix_global->registered_functions));
219 /* Let's free maestro now */
220 SIMIX_context_free(simix_global->maestro_process->context);
221 xbt_free(simix_global->maestro_process->running_ctx);
222 xbt_free(simix_global->maestro_process);
223 simix_global->maestro_process = NULL;
225 /* Restore the default exception setup */
226 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
227 __xbt_ex_terminate = &__xbt_ex_terminate_default;
229 /* Finish context module and SURF */
230 SIMIX_context_mod_exit();
234 #ifdef TIME_BENCH_AMDAHL
235 xbt_os_cputimer_stop(simix_global->timer_seq);
236 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
237 xbt_os_timer_elapsed(simix_global->timer_seq),
238 xbt_os_timer_elapsed(simix_global->timer_par));
239 xbt_os_timer_free(simix_global->timer_seq);
240 xbt_os_timer_free(simix_global->timer_par);
243 xbt_mallocator_free(simix_global->action_mallocator);
244 xbt_free(simix_global);
253 * \brief A clock (in second).
255 * \return Return the clock.
257 XBT_INLINE double SIMIX_get_clock(void)
260 return MC_process_clock_get(SIMIX_process_self());
262 return surf_get_clock();
266 static int process_syscall_color(void *p)
268 switch ((*(smx_process_t *)p)->simcall.call) {
270 case SIMCALL_PROCESS_KILL:
272 case SIMCALL_PROCESS_RESUME:
281 * \brief Run the main simulation loop.
286 smx_process_t process;
287 surf_action_t action;
293 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
294 xbt_dynar_length(simix_global->process_to_run));
295 #ifdef TIME_BENCH_PER_SR
296 smx_ctx_raw_new_sr();
298 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
299 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
300 xbt_dynar_length(simix_global->process_to_run));
302 /* Run all processes that are ready to run, possibly in parallel */
303 #ifdef TIME_BENCH_AMDAHL
304 xbt_os_cputimer_stop(simix_global->timer_seq);
305 xbt_os_cputimer_resume(simix_global->timer_par);
307 SIMIX_process_runall();
308 #ifdef TIME_BENCH_AMDAHL
309 xbt_os_cputimer_stop(simix_global->timer_par);
310 xbt_os_cputimer_resume(simix_global->timer_seq);
313 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
314 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
316 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
318 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
320 /* Here, the order is ok because:
322 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
324 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
326 * - if there is no kill during the simulation, processes remain sorted according by their PID.
327 * rational: This can be proved inductively.
328 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
329 * Let's show that it is still so at the end of this round.
330 * - if a process is added when being created, that's from maestro. It can be either at startup
331 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
332 * in arbitrary order (inductive hypothesis), we are fine.
333 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
334 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
335 * of a simcall terminates, we're still good. Proof:
336 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
337 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
338 * - For communications, this function is called from SIMIX_comm_finish().
339 * This function itself don't mess with the order since simcalls are handled in FIFO order.
340 * The function is called:
341 * - before the comm starts (invalid parameters, or resource already dead or whatever).
342 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
343 * - because the communication failed or were canceled after startup. In this case, it's called from the function
344 * we are in, by the chunk:
345 * set = model->states.failed_action_set;
346 * while ((action = xbt_swag_extract(set)))
347 * SIMIX_simcall_post((smx_action_t) action->data);
348 * This order is also fixed because it depends of the order in which the surf actions were
349 * added to the system, and only maestro can add stuff this way, through simcalls.
350 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
351 * poped out of the swag does not depend on the user code's execution order.
352 * - because the communication terminated. In this case, actions are served in the order given by
353 * set = model->states.done_action_set;
354 * while ((action = xbt_swag_extract(set)))
355 * SIMIX_simcall_post((smx_action_t) action->data);
356 * and the argument is very similar to the previous one.
357 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
358 * 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.
359 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
360 * - If there is some process killings, the order is changed by this decision that comes from user-land
361 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
362 * So, even the order change induced by the process killing is perfectly reproducible.
364 * So science works, bitches [http://xkcd.com/54/].
366 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
367 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
368 * That would thus be a pure waste of time.
371 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
372 if (process->simcall.call != SIMCALL_NONE) {
373 SIMIX_simcall_pre(&process->simcall, 0);
378 time = SIMIX_timer_next();
379 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
380 time = surf_solve(time);
382 /* Notify all the hosts that have failed */
383 /* FIXME: iterate through the list of failed host and mark each of them */
384 /* as failed. On each host, signal all the running processes with host_fail */
386 /* Handle any pending timer */
387 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
388 //FIXME: make the timers being real callbacks
389 // (i.e. provide dispatchers that read and expand the args)
390 timer = xbt_heap_pop(simix_timers);
392 ((void (*)(void*))timer->func)(timer->args);
396 /* Wake up all processes waiting for a Surf action to finish */
397 xbt_dynar_foreach(model_list, iter, model) {
398 while ((action = surf_model_extract_failed_action_set(model)))
399 SIMIX_simcall_post((smx_action_t) surf_action_get_data(action));
401 while ((action = surf_model_extract_done_action_set(model)))
402 if (surf_action_get_data(action) == NULL)
403 XBT_DEBUG("probably vcpu's action %p, skip", action);
405 SIMIX_simcall_post((smx_action_t) surf_action_get_data(action));
408 /* Autorestart all process */
409 if(host_that_restart) {
410 char *hostname = NULL;
411 xbt_dynar_foreach(host_that_restart,iter,hostname) {
412 XBT_INFO("Restart processes on host: %s",hostname);
413 SIMIX_host_autorestart(SIMIX_host_get_by_name(hostname));
415 xbt_dynar_reset(host_that_restart);
418 /* Clean processes to destroy */
419 SIMIX_process_empty_trash();
422 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
423 // !(time == -1.0 && xbt_dynar_is_empty())
426 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
428 if (xbt_swag_size(simix_global->process_list) != 0) {
434 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
435 SIMIX_display_process_status();
441 * \brief Set the date to execute a function
443 * Set the date to execute the function on the surf.
444 * \param date Date to execute function
445 * \param function Function to be executed
446 * \param arg Parameters of the function
449 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
451 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
454 timer->func = function;
456 xbt_heap_push(simix_timers, timer, date);
460 * \brief Registers a function to create a process.
462 * This function registers a function to be called
463 * when a new process is created. The function has
464 * to call SIMIX_process_create().
465 * \param function create process function
467 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
470 simix_global->create_process_function = function;
474 * \brief Registers a function to kill a process.
476 * This function registers a function to be called when a
477 * process is killed. The function has to call the SIMIX_process_kill().
479 * \param function Kill process function
481 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
484 simix_global->kill_process_function = function;
488 * \brief Registers a function to cleanup a process.
490 * This function registers a user function to be called when
491 * a process ends properly.
493 * \param function cleanup process function
495 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
498 simix_global->cleanup_process_function = function;
502 void SIMIX_display_process_status(void)
504 if (simix_global->process_list == NULL) {
508 smx_process_t process = NULL;
509 int nbprocess = xbt_swag_size(simix_global->process_list);
511 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
512 /* List the process and their state */
514 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
515 xbt_swag_foreach(process, simix_global->process_list) {
517 if (process->waiting_action) {
519 const char* action_description = "unknown";
520 switch (process->waiting_action->type) {
522 case SIMIX_ACTION_EXECUTE:
523 action_description = "execution";
526 case SIMIX_ACTION_PARALLEL_EXECUTE:
527 action_description = "parallel execution";
530 case SIMIX_ACTION_COMMUNICATE:
531 action_description = "communication";
534 case SIMIX_ACTION_SLEEP:
535 action_description = "sleeping";
538 case SIMIX_ACTION_SYNCHRO:
539 action_description = "synchronization";
542 case SIMIX_ACTION_IO:
543 action_description = "I/O";
545 /* **************************************/
546 /* TUTORIAL: New API */
547 case SIMIX_ACTION_NEW_API:
548 action_description = "NEW API";
549 /* **************************************/
553 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
554 process->pid, process->name, sg_host_name(process->smx_host),
555 action_description, process->waiting_action,
556 process->waiting_action->name, (int)process->waiting_action->state);
559 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host));
564 static void* SIMIX_action_mallocator_new_f(void) {
565 smx_action_t action = xbt_new(s_smx_action_t, 1);
566 action->simcalls = xbt_fifo_new();
570 static void SIMIX_action_mallocator_free_f(void* action) {
571 xbt_fifo_free(((smx_action_t) action)->simcalls);
575 static void SIMIX_action_mallocator_reset_f(void* action) {
577 // we also recycle the simcall list
578 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
579 xbt_fifo_reset(fifo);
580 memset(action, 0, sizeof(s_smx_action_t));
581 ((smx_action_t) action)->simcalls = fifo;
584 xbt_dict_t SIMIX_pre_asr_get_properties(smx_simcall_t simcall, const char *name){
585 return SIMIX_asr_get_properties(name);
587 xbt_dict_t SIMIX_asr_get_properties(const char *name)
589 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);