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/internal_config.h"
10 #include "src/surf/surf_interface.hpp"
11 #include "src/surf/storage_interface.hpp"
12 #include "src/surf/xml/platf.hpp"
13 #include "smx_private.h"
14 #include "smx_private.hpp"
16 #include "xbt/ex.h" /* ex_backtrace_display */
18 #include "src/mc/mc_replay.h"
19 #include "simgrid/sg_config.h"
22 #include "src/mc/mc_private.h"
23 #include "src/mc/mc_protocol.h"
24 #include "src/mc/Client.hpp"
27 #include "src/mc/mc_protocol.h"
30 #include "src/mc/mc_record.h"
33 #include "src/smpi/private.h"
36 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
37 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
39 smx_global_t simix_global = NULL;
40 static xbt_heap_t simix_timers = NULL;
42 /** @brief Timer datatype */
43 typedef struct s_smx_timer {
49 void (*SMPI_switch_data_segment)(int) = NULL;
51 static void* SIMIX_synchro_mallocator_new_f(void);
52 static void SIMIX_synchro_mallocator_free_f(void* synchro);
53 static void SIMIX_synchro_mallocator_reset_f(void* synchro);
55 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
58 int _sg_do_verbose_exit = 1;
59 static void inthandler(int ignored)
61 if ( _sg_do_verbose_exit ) {
62 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
63 SIMIX_display_process_status();
66 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
72 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
74 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
76 "Access violation detected.\n"
77 "This can result from a programming error in your code or, although less likely,\n"
78 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
79 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
81 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
82 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
83 smx_context_stack_size / 1024);
84 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
86 "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
87 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
89 } else if (siginfo->si_signo == SIGSEGV) {
90 fprintf(stderr, "Segmentation fault.\n");
92 if (smpi_enabled() && !smpi_privatize_global_variables) {
93 #if HAVE_PRIVATIZATION
95 "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
98 "Sadly, your system does not support --cfg=smpi/privatize-global-variables:yes (yet).\n");
99 #endif /* HAVE_PRIVATIZATION */
101 #endif /* HAVE_SMPI */
106 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
109 * Install signal handler for SIGSEGV. Check that nobody has already installed
110 * its own handler. For example, the Java VM does this.
112 static void install_segvhandler(void)
114 stack_t stack, old_stack;
115 stack.ss_sp = sigsegv_stack;
116 stack.ss_size = sizeof sigsegv_stack;
119 if (sigaltstack(&stack, &old_stack) == -1) {
120 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
123 if (!(old_stack.ss_flags & SS_DISABLE)) {
124 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
125 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
126 sigaltstack(&old_stack, NULL);
129 struct sigaction action, old_action;
130 action.sa_sigaction = segvhandler;
131 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
132 sigemptyset(&action.sa_mask);
134 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
135 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
138 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
139 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
140 (old_action.sa_flags & SA_SIGINFO) ?
141 (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
142 sigaction(SIGSEGV, &old_action, NULL);
147 /********************************* SIMIX **************************************/
149 double SIMIX_timer_next(void)
151 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
154 static void kill_process(smx_process_t process)
156 SIMIX_process_kill(process, NULL);
159 static void SIMIX_storage_create_(smx_storage_t storage)
161 const char* key = xbt_dict_get_elm_key(storage);
162 SIMIX_storage_create(key, storage, NULL);
165 static std::function<void()> maestro_code;
170 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
172 maestro_code = std::move(code);
178 void SIMIX_set_maestro(void (*code)(void*), void* data)
180 maestro_code = std::bind(code, data);
185 * \brief Initialize SIMIX internal data.
190 void SIMIX_global_init(int *argc, char **argv)
193 // The communication initialization is done ASAP.
194 // We need to communicate initialization of the different layers to the model-checker.
195 simgrid::mc::Client::initialize();
198 s_smx_process_t proc;
201 simix_global = xbt_new0(s_smx_global_t, 1);
203 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
204 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
205 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
206 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
208 simix_global->maestro_process = NULL;
209 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
211 simix_global->create_process_function = SIMIX_process_create;
212 simix_global->kill_process_function = kill_process;
213 simix_global->cleanup_process_function = SIMIX_process_cleanup;
214 simix_global->synchro_mallocator = xbt_mallocator_new(65536,
215 SIMIX_synchro_mallocator_new_f, SIMIX_synchro_mallocator_free_f,
216 SIMIX_synchro_mallocator_reset_f);
217 simix_global->mutex = xbt_os_mutex_init();
219 surf_init(argc, argv); /* Initialize SURF structures */
220 SIMIX_context_mod_init();
222 // Either create a new context with maestro or create
223 // a context object with the current context mestro):
224 simgrid::simix::create_maestro(maestro_code);
226 /* context exception handlers */
227 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
228 __xbt_ex_terminate = SIMIX_process_exception_terminate;
230 /* Prepare to display some more info when dying on Ctrl-C pressing */
231 signal(SIGINT, inthandler);
234 install_segvhandler();
236 /* register a function to be called by SURF after the environment creation */
238 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
239 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
240 SIMIX_host_create(&host);
242 SIMIX_HOST_LEVEL = simgrid::s4u::Host::extension_create(SIMIX_host_destroy);
244 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
245 const char* id = storage->getName();
246 // TODO, create sg_storage_by_name
247 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, id);
248 xbt_assert(s != NULL, "Storage not found for name %s", id);
249 SIMIX_storage_create_(s);
252 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
255 simix_timers = xbt_heap_new(8, &free);
258 if (xbt_cfg_get_boolean("clean-atexit"))
261 if (_sg_cfg_exit_asap)
268 * \brief Clean the SIMIX simulation
270 * This functions remove the memory used by SIMIX
272 void SIMIX_clean(void)
274 #ifdef TIME_BENCH_PER_SR
275 smx_ctx_raw_new_sr();
277 if (smx_cleaned) return; // to avoid double cleaning by java and C
279 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
280 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() == 0.0) {
282 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
283 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
284 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
286 /* Kill all processes (but maestro) */
287 SIMIX_process_killall(simix_global->maestro_process, 1);
289 /* Exit the SIMIX network module */
290 SIMIX_mailbox_exit();
292 xbt_heap_free(simix_timers);
294 /* Free the remaining data structures */
295 xbt_dynar_free(&simix_global->process_to_run);
296 xbt_dynar_free(&simix_global->process_that_ran);
297 xbt_swag_free(simix_global->process_to_destroy);
298 xbt_swag_free(simix_global->process_list);
299 simix_global->process_list = NULL;
300 simix_global->process_to_destroy = NULL;
301 xbt_dict_free(&(simix_global->registered_functions));
303 xbt_os_mutex_destroy(simix_global->mutex);
304 simix_global->mutex = NULL;
306 /* Let's free maestro now */
307 SIMIX_context_free(simix_global->maestro_process->context);
308 xbt_free(simix_global->maestro_process->running_ctx);
309 xbt_free(simix_global->maestro_process);
310 simix_global->maestro_process = NULL;
312 /* Restore the default exception setup */
313 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
314 __xbt_ex_terminate = &__xbt_ex_terminate_default;
316 /* Finish context module and SURF */
317 SIMIX_context_mod_exit();
321 xbt_mallocator_free(simix_global->synchro_mallocator);
322 xbt_free(simix_global);
331 * \brief A clock (in second).
333 * \return Return the clock.
335 double SIMIX_get_clock(void)
337 if(MC_is_active() || MC_record_replay_is_active()){
338 return MC_process_clock_get(SIMIX_process_self());
340 return surf_get_clock();
344 static int process_syscall_color(void *p)
346 switch ((*(smx_process_t *)p)->simcall.call) {
348 case SIMCALL_PROCESS_KILL:
350 case SIMCALL_PROCESS_RESUME:
359 * \brief Run the main simulation loop.
363 if (MC_record_path) {
364 simgrid::mc::replay(MC_record_path);
369 smx_process_t process;
370 surf_action_t action;
376 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
377 xbt_dynar_length(simix_global->process_to_run));
378 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
379 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
380 xbt_dynar_length(simix_global->process_to_run));
382 /* Run all processes that are ready to run, possibly in parallel */
383 SIMIX_process_runall();
385 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
386 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
388 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
390 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
392 /* Here, the order is ok because:
394 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
396 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
398 * - if there is no kill during the simulation, processes remain sorted according by their PID.
399 * rational: This can be proved inductively.
400 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
401 * Let's show that it is still so at the end of this round.
402 * - if a process is added when being created, that's from maestro. It can be either at startup
403 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
404 * in arbitrary order (inductive hypothesis), we are fine.
405 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
406 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
407 * of a simcall terminates, we're still good. Proof:
408 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
409 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
410 * - For communications, this function is called from SIMIX_comm_finish().
411 * This function itself don't mess with the order since simcalls are handled in FIFO order.
412 * The function is called:
413 * - before the comm starts (invalid parameters, or resource already dead or whatever).
414 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
415 * - because the communication failed or were canceled after startup. In this case, it's called from the function
416 * we are in, by the chunk:
417 * set = model->states.failed_action_set;
418 * while ((synchro = xbt_swag_extract(set)))
419 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
420 * This order is also fixed because it depends of the order in which the surf actions were
421 * added to the system, and only maestro can add stuff this way, through simcalls.
422 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
423 * poped out of the swag does not depend on the user code's execution order.
424 * - because the communication terminated. In this case, synchros are served in the order given by
425 * set = model->states.done_action_set;
426 * while ((synchro = xbt_swag_extract(set)))
427 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
428 * and the argument is very similar to the previous one.
429 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
430 * 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.
431 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
432 * - If there is some process killings, the order is changed by this decision that comes from user-land
433 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
434 * So, even the order change induced by the process killing is perfectly reproducible.
436 * So science works, bitches [http://xkcd.com/54/].
438 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
439 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
440 * That would thus be a pure waste of time.
443 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
444 if (process->simcall.call != SIMCALL_NONE) {
445 SIMIX_simcall_handle(&process->simcall, 0);
448 /* Wake up all processes waiting for a Surf action to finish */
449 xbt_dynar_foreach(all_existing_models, iter, model) {
450 XBT_DEBUG("Handling process whose action failed");
451 while ((action = surf_model_extract_failed_action_set(model))) {
452 XBT_DEBUG(" Handling Action %p",action);
453 SIMIX_simcall_exit((smx_synchro_t) action->getData());
455 XBT_DEBUG("Handling process whose action terminated normally");
456 while ((action = surf_model_extract_done_action_set(model))) {
457 XBT_DEBUG(" Handling Action %p",action);
458 if (action->getData() == NULL)
459 XBT_DEBUG("probably vcpu's action %p, skip", action);
461 SIMIX_simcall_exit((smx_synchro_t) action->getData());
466 time = SIMIX_timer_next();
467 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
468 XBT_DEBUG("Calling surf_solve");
469 time = surf_solve(time);
470 XBT_DEBUG("Moving time ahead : %g", time);
472 /* Notify all the hosts that have failed */
473 /* FIXME: iterate through the list of failed host and mark each of them */
474 /* as failed. On each host, signal all the running processes with host_fail */
476 /* Handle any pending timer */
477 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
478 //FIXME: make the timers being real callbacks
479 // (i.e. provide dispatchers that read and expand the args)
480 timer = (smx_timer_t) xbt_heap_pop(simix_timers);
482 timer->func(timer->args);
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());
503 /* Autorestart all process */
504 char *hostname = NULL;
505 xbt_dynar_foreach(host_that_restart,iter,hostname) {
506 XBT_INFO("Restart processes on host: %s",hostname);
507 SIMIX_host_autorestart(sg_host_by_name(hostname));
509 xbt_dynar_reset(host_that_restart);
511 /* Clean processes to destroy */
512 SIMIX_process_empty_trash();
515 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
517 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
519 if (xbt_swag_size(simix_global->process_list) != 0) {
523 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
524 SIMIX_display_process_status();
530 * \brief Set the date to execute a function
532 * Set the date to execute the function on the surf.
533 * \param date Date to execute function
534 * \param function Function to be executed
535 * \param arg Parameters of the function
538 smx_timer_t SIMIX_timer_set(double date, void (*function)(void*), void *arg)
540 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
543 timer->func = function;
545 xbt_heap_push(simix_timers, timer, date);
548 /** @brief cancels a timer that was added earlier */
549 void SIMIX_timer_remove(smx_timer_t timer) {
550 xbt_heap_rm_elm(simix_timers, timer, timer->date);
553 /** @brief Returns the date at which the timer will trigger (or 0 if NULL timer) */
554 double SIMIX_timer_get_date(smx_timer_t timer) {
555 return timer?timer->date:0;
559 * \brief Registers a function to create a process.
561 * This function registers a function to be called
562 * when a new process is created. The function has
563 * to call SIMIX_process_create().
564 * \param function create process function
566 void SIMIX_function_register_process_create(smx_creation_func_t
569 simix_global->create_process_function = function;
573 * \brief Registers a function to kill a process.
575 * This function registers a function to be called when a
576 * process is killed. The function has to call the SIMIX_process_kill().
578 * \param function Kill process function
580 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
583 simix_global->kill_process_function = function;
587 * \brief Registers a function to cleanup a process.
589 * This function registers a user function to be called when
590 * a process ends properly.
592 * \param function cleanup process function
594 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
597 simix_global->cleanup_process_function = function;
601 void SIMIX_display_process_status(void)
603 if (simix_global->process_list == NULL) {
607 smx_process_t process = NULL;
608 int nbprocess = xbt_swag_size(simix_global->process_list);
610 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
611 /* List the process and their state */
613 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
614 xbt_swag_foreach(process, simix_global->process_list) {
616 if (process->waiting_synchro) {
618 const char* synchro_description = "unknown";
619 switch (process->waiting_synchro->type) {
621 case SIMIX_SYNC_EXECUTE:
622 synchro_description = "execution";
625 case SIMIX_SYNC_PARALLEL_EXECUTE:
626 synchro_description = "parallel execution";
629 case SIMIX_SYNC_COMMUNICATE:
630 synchro_description = "communication";
633 case SIMIX_SYNC_SLEEP:
634 synchro_description = "sleeping";
637 case SIMIX_SYNC_JOIN:
638 synchro_description = "joining";
641 case SIMIX_SYNC_SYNCHRO:
642 synchro_description = "synchronization";
646 synchro_description = "I/O";
649 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
650 process->pid, process->name, sg_host_get_name(process->host),
651 synchro_description, process->waiting_synchro,
652 process->waiting_synchro->name, (int)process->waiting_synchro->state);
655 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_get_name(process->host));
660 static void* SIMIX_synchro_mallocator_new_f(void) {
661 smx_synchro_t synchro = xbt_new(s_smx_synchro_t, 1);
662 synchro->simcalls = xbt_fifo_new();
666 static void SIMIX_synchro_mallocator_free_f(void* synchro) {
667 xbt_fifo_free(((smx_synchro_t) synchro)->simcalls);
671 static void SIMIX_synchro_mallocator_reset_f(void* synchro) {
673 // we also recycle the simcall list
674 xbt_fifo_t fifo = ((smx_synchro_t) synchro)->simcalls;
675 xbt_fifo_reset(fifo);
676 memset(synchro, 0, sizeof(s_smx_synchro_t));
677 ((smx_synchro_t) synchro)->simcalls = fifo;
680 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
681 return SIMIX_asr_get_properties(name);
683 xbt_dict_t SIMIX_asr_get_properties(const char *name)
685 return (xbt_dict_t) xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);
688 int SIMIX_is_maestro()
690 return simix_global==NULL /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;