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. This can result from a stack overflow.\n"
48 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
49 smx_context_stack_size / 1024);
50 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
52 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
53 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
59 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
62 * Install signal handler for SIGSEGV. Check that nobody has already installed
63 * its own handler. For example, the Java VM does this.
65 static void install_segvhandler(void)
67 stack_t stack, old_stack;
68 stack.ss_sp = sigsegv_stack;
69 stack.ss_size = sizeof sigsegv_stack;
72 if (sigaltstack(&stack, &old_stack) == -1) {
73 XBT_WARN("Failed to register alternate signal stack: %s",
77 if (!(old_stack.ss_flags & SS_DISABLE)) {
78 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
79 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
80 sigaltstack(&old_stack, NULL);
83 struct sigaction action, old_action;
84 action.sa_sigaction = segvhandler;
85 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
86 sigemptyset(&action.sa_mask);
88 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
89 XBT_WARN("Failed to register signal handler for SIGSEGV: %s",
93 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
94 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
95 (old_action.sa_flags & SA_SIGINFO) ?
96 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
97 sigaction(SIGSEGV, &old_action, NULL);
101 /********************************* SIMIX **************************************/
103 XBT_INLINE double SIMIX_timer_next(void)
105 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
110 * \brief Initialize SIMIX internal data.
115 void SIMIX_global_init(int *argc, char **argv)
117 s_smx_process_t proc;
120 simix_global = xbt_new0(s_smx_global_t, 1);
122 #ifdef TIME_BENCH_AMDAHL
123 simix_global->timer_seq = xbt_os_timer_new();
124 simix_global->timer_par = xbt_os_timer_new();
125 xbt_os_cputimer_start(simix_global->timer_seq);
127 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
128 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
129 simix_global->process_list =
130 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
131 simix_global->process_to_destroy =
132 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
134 simix_global->maestro_process = NULL;
135 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
137 simix_global->create_process_function = SIMIX_process_create;
138 simix_global->kill_process_function = SIMIX_process_kill;
139 simix_global->cleanup_process_function = SIMIX_process_cleanup;
140 simix_global->action_mallocator = xbt_mallocator_new(65536,
141 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
142 SIMIX_action_mallocator_reset_f);
143 simix_global->autorestart = SIMIX_host_restart_processes;
145 surf_init(argc, argv); /* Initialize SURF structures */
146 SIMIX_context_mod_init();
147 SIMIX_create_maestro_process();
149 /* context exception handlers */
150 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
151 __xbt_ex_terminate = SIMIX_process_exception_terminate;
153 SIMIX_network_init();
155 /* Prepare to display some more info when dying on Ctrl-C pressing */
156 signal(SIGINT, inthandler);
158 /* Install SEGV handler */
159 install_segvhandler();
161 /* register a function to be called by SURF after the environment creation */
163 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
167 simix_timers = xbt_heap_new(8, &free);
170 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
171 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
173 if (sg_cfg_get_boolean("clean_atexit"))
176 if (_sg_cfg_exit_asap)
182 * \brief Clean the SIMIX simulation
184 * This functions remove the memory used by SIMIX
187 void SIMIX_clean(void)
189 #ifdef TIME_BENCH_PER_SR
190 smx_ctx_raw_new_sr();
192 if (cleaned) return; // to avoid double cleaning by java and C
194 /* Kill everyone (except maestro) */
195 SIMIX_process_killall(simix_global->maestro_process, 1);
197 /* Exit the SIMIX network module */
198 SIMIX_network_exit();
200 xbt_heap_free(simix_timers);
202 /* Free the remaining data structures */
203 xbt_dynar_free(&simix_global->process_to_run);
204 xbt_dynar_free(&simix_global->process_that_ran);
205 xbt_swag_free(simix_global->process_to_destroy);
206 xbt_swag_free(simix_global->process_list);
207 simix_global->process_list = NULL;
208 simix_global->process_to_destroy = NULL;
209 xbt_dict_free(&(simix_global->registered_functions));
211 /* Let's free maestro now */
212 SIMIX_context_free(simix_global->maestro_process->context);
213 xbt_free(simix_global->maestro_process->running_ctx);
214 xbt_free(simix_global->maestro_process);
215 simix_global->maestro_process = NULL;
217 /* Restore the default exception setup */
218 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
219 __xbt_ex_terminate = &__xbt_ex_terminate_default;
221 /* Finish context module and SURF */
222 SIMIX_context_mod_exit();
226 #ifdef TIME_BENCH_AMDAHL
227 xbt_os_cputimer_stop(simix_global->timer_seq);
228 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
229 xbt_os_timer_elapsed(simix_global->timer_seq),
230 xbt_os_timer_elapsed(simix_global->timer_par));
231 xbt_os_timer_free(simix_global->timer_seq);
232 xbt_os_timer_free(simix_global->timer_par);
235 xbt_mallocator_free(simix_global->action_mallocator);
236 xbt_free(simix_global);
245 * \brief A clock (in second).
247 * \return Return the clock.
249 XBT_INLINE double SIMIX_get_clock(void)
252 return MC_process_clock_get(SIMIX_process_self());
254 return surf_get_clock();
258 static int process_syscall_color(void *p)
260 switch ((*(smx_process_t *)p)->simcall.call) {
262 case SIMCALL_PROCESS_KILL:
264 case SIMCALL_PROCESS_RESUME:
273 * \brief Run the main simulation loop.
278 smx_process_t process;
279 surf_action_t action;
285 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
286 xbt_dynar_length(simix_global->process_to_run));
287 #ifdef TIME_BENCH_PER_SR
288 smx_ctx_raw_new_sr();
290 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
291 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
292 xbt_dynar_length(simix_global->process_to_run));
294 /* Run all processes that are ready to run, possibly in parallel */
295 #ifdef TIME_BENCH_AMDAHL
296 xbt_os_cputimer_stop(simix_global->timer_seq);
297 xbt_os_cputimer_resume(simix_global->timer_par);
299 SIMIX_process_runall();
300 #ifdef TIME_BENCH_AMDAHL
301 xbt_os_cputimer_stop(simix_global->timer_par);
302 xbt_os_cputimer_resume(simix_global->timer_seq);
305 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
306 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
308 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
310 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
312 /* Here, the order is ok because:
314 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
316 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
318 * - if there is no kill during the simulation, processes remain sorted according by their PID.
319 * rational: This can be proved inductively.
320 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
321 * Let's show that it is still so at the end of this round.
322 * - if a process is added when being created, that's from maestro. It can be either at startup
323 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
324 * in arbitrary order (inductive hypothesis), we are fine.
325 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
326 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
327 * of a simcall terminates, we're still good. Proof:
328 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
329 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
330 * - For communications, this function is called from SIMIX_comm_finish().
331 * This function itself don't mess with the order since simcalls are handled in FIFO order.
332 * The function is called:
333 * - before the comm starts (invalid parameters, or resource already dead or whatever).
334 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
335 * - because the communication failed or were canceled after startup. In this case, it's called from the function
336 * we are in, by the chunk:
337 * set = model->states.failed_action_set;
338 * while ((action = xbt_swag_extract(set)))
339 * SIMIX_simcall_post((smx_action_t) action->data);
340 * This order is also fixed because it depends of the order in which the surf actions were
341 * added to the system, and only maestro can add stuff this way, through simcalls.
342 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
343 * poped out of the swag does not depend on the user code's execution order.
344 * - because the communication terminated. In this case, actions are served in the order given by
345 * set = model->states.done_action_set;
346 * while ((action = xbt_swag_extract(set)))
347 * SIMIX_simcall_post((smx_action_t) action->data);
348 * and the argument is very similar to the previous one.
349 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
350 * 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.
351 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
352 * - If there is some process killings, the order is changed by this decision that comes from user-land
353 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
354 * So, even the order change induced by the process killing is perfectly reproducible.
356 * So science works, bitches [http://xkcd.com/54/].
358 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
359 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
360 * That would thus be a pure waste of time.
363 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
364 if (process->simcall.call != SIMCALL_NONE) {
365 SIMIX_simcall_pre(&process->simcall, 0);
370 time = SIMIX_timer_next();
371 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
372 time = surf_solve(time);
374 /* Notify all the hosts that have failed */
375 /* FIXME: iterate through the list of failed host and mark each of them */
376 /* as failed. On each host, signal all the running processes with host_fail */
378 /* Handle any pending timer */
379 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
380 //FIXME: make the timers being real callbacks
381 // (i.e. provide dispatchers that read and expand the args)
382 timer = xbt_heap_pop(simix_timers);
384 ((void (*)(void*))timer->func)(timer->args);
388 /* Wake up all processes waiting for a Surf action to finish */
389 xbt_dynar_foreach(model_list, iter, model) {
390 while ((action = surf_model_extract_failed_action_set(model)))
391 SIMIX_simcall_post((smx_action_t) surf_action_get_data(action));
393 while ((action = surf_model_extract_done_action_set(model)))
394 if (surf_action_get_data(action) == NULL)
395 XBT_DEBUG("probably vcpu's action %p, skip", action);
397 SIMIX_simcall_post((smx_action_t) surf_action_get_data(action));
400 /* Autorestart all process */
401 if(host_that_restart) {
402 char *hostname = NULL;
403 xbt_dynar_foreach(host_that_restart,iter,hostname) {
404 XBT_INFO("Restart processes on host: %s",hostname);
405 SIMIX_host_autorestart(SIMIX_host_get_by_name(hostname));
407 xbt_dynar_reset(host_that_restart);
410 /* Clean processes to destroy */
411 SIMIX_process_empty_trash();
414 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
415 // !(time == -1.0 && xbt_dynar_is_empty())
418 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
420 if (xbt_swag_size(simix_global->process_list) != 0) {
426 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
427 SIMIX_display_process_status();
433 * \brief Set the date to execute a function
435 * Set the date to execute the function on the surf.
436 * \param date Date to execute function
437 * \param function Function to be executed
438 * \param arg Parameters of the function
441 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
443 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
446 timer->func = function;
448 xbt_heap_push(simix_timers, timer, date);
452 * \brief Registers a function to create a process.
454 * This function registers a function to be called
455 * when a new process is created. The function has
456 * to call SIMIX_process_create().
457 * \param function create process function
459 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
462 simix_global->create_process_function = function;
466 * \brief Registers a function to kill a process.
468 * This function registers a function to be called when a
469 * process is killed. The function has to call the SIMIX_process_kill().
471 * \param function Kill process function
473 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
476 simix_global->kill_process_function = function;
480 * \brief Registers a function to cleanup a process.
482 * This function registers a user function to be called when
483 * a process ends properly.
485 * \param function cleanup process function
487 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
490 simix_global->cleanup_process_function = function;
494 void SIMIX_display_process_status(void)
496 if (simix_global->process_list == NULL) {
500 smx_process_t process = NULL;
501 int nbprocess = xbt_swag_size(simix_global->process_list);
503 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
504 /* List the process and their state */
506 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
507 xbt_swag_foreach(process, simix_global->process_list) {
509 if (process->waiting_action) {
511 const char* action_description = "unknown";
512 switch (process->waiting_action->type) {
514 case SIMIX_ACTION_EXECUTE:
515 action_description = "execution";
518 case SIMIX_ACTION_PARALLEL_EXECUTE:
519 action_description = "parallel execution";
522 case SIMIX_ACTION_COMMUNICATE:
523 action_description = "communication";
526 case SIMIX_ACTION_SLEEP:
527 action_description = "sleeping";
530 case SIMIX_ACTION_SYNCHRO:
531 action_description = "synchronization";
534 case SIMIX_ACTION_IO:
535 action_description = "I/O";
537 /* **************************************/
538 /* TUTORIAL: New API */
539 case SIMIX_ACTION_NEW_API:
540 action_description = "NEW API";
541 /* **************************************/
545 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
546 process->pid, process->name, sg_host_name(process->smx_host),
547 action_description, process->waiting_action,
548 process->waiting_action->name, (int)process->waiting_action->state);
551 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host));
556 static void* SIMIX_action_mallocator_new_f(void) {
557 smx_action_t action = xbt_new(s_smx_action_t, 1);
558 action->simcalls = xbt_fifo_new();
562 static void SIMIX_action_mallocator_free_f(void* action) {
563 xbt_fifo_free(((smx_action_t) action)->simcalls);
567 static void SIMIX_action_mallocator_reset_f(void* action) {
569 // we also recycle the simcall list
570 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
571 xbt_fifo_reset(fifo);
572 memset(action, 0, sizeof(s_smx_action_t));
573 ((smx_action_t) action)->simcalls = fifo;
576 xbt_dict_t SIMIX_pre_asr_get_properties(smx_simcall_t simcall, const char *name){
577 return SIMIX_asr_get_properties(name);
579 xbt_dict_t SIMIX_asr_get_properties(const char *name)
581 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);