X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/9c916cb125de2f1050a588d972fd67ff9aac5e29..f9e6853d2c5b7b867211700bc12f0ee57f640b30:/src/simix/smx_global.c diff --git a/src/simix/smx_global.c b/src/simix/smx_global.c index c18f9ecdbd..7df250c5c1 100644 --- a/src/simix/smx_global.c +++ b/src/simix/smx_global.c @@ -1,4 +1,4 @@ -/* Copyright (c) 2007, 2008, 2009, 2010. The SimGrid Team. +/* Copyright (c) 2007-2014. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it @@ -11,7 +11,9 @@ #include "xbt/str.h" #include "xbt/ex.h" /* ex_backtrace_display */ #include "mc/mc.h" +#include "simgrid/sg_config.h" +XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories"); XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)"); @@ -22,15 +24,13 @@ static void* SIMIX_action_mallocator_new_f(void); static void SIMIX_action_mallocator_free_f(void* action); static void SIMIX_action_mallocator_reset_f(void* action); -extern void smx_ctx_raw_new_sr(void); - /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */ #include -int _surf_do_verbose_exit = 1; +int _sg_do_verbose_exit = 1; static void _XBT_CALL inthandler(int ignored) { - if ( _surf_do_verbose_exit ) { + if ( _sg_do_verbose_exit ) { XBT_INFO("CTRL-C pressed. Displaying status and bailing out"); SIMIX_display_process_status(); } @@ -40,6 +40,64 @@ static void _XBT_CALL inthandler(int ignored) exit(1); } +static void _XBT_CALL segvhandler(int signum, siginfo_t *siginfo, void *context) +{ + if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) { + fprintf(stderr, + "Access violation detected. This can result from a stack overflow.\n" + "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n", + smx_context_stack_size / 1024); + if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) { + fprintf(stderr, + "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n", + siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr); + } + } + raise(signum); +} + +char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */ + +/** + * Install signal handler for SIGSEGV. Check that nobody has already installed + * its own handler. For example, the Java VM does this. + */ +static void install_segvhandler(void) +{ + stack_t stack, old_stack; + stack.ss_sp = sigsegv_stack; + stack.ss_size = sizeof sigsegv_stack; + stack.ss_flags = 0; + + if (sigaltstack(&stack, &old_stack) == -1) { + XBT_WARN("Failed to register alternate signal stack: %s", + strerror(errno)); + return; + } + if (!(old_stack.ss_flags & SS_DISABLE)) { + XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.", + old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags); + sigaltstack(&old_stack, NULL); + } + + struct sigaction action, old_action; + action.sa_sigaction = segvhandler; + action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO; + sigemptyset(&action.sa_mask); + + if (sigaction(SIGSEGV, &action, &old_action) == -1) { + XBT_WARN("Failed to register signal handler for SIGSEGV: %s", + strerror(errno)); + return; + } + if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) { + XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.", + (old_action.sa_flags & SA_SIGINFO) ? + (void*)old_action.sa_sigaction : (void*)old_action.sa_handler); + sigaction(SIGSEGV, &old_action, NULL); + } +} + /********************************* SIMIX **************************************/ XBT_INLINE double SIMIX_timer_next(void) @@ -48,6 +106,7 @@ XBT_INLINE double SIMIX_timer_next(void) } /** + * \ingroup SIMIX_API * \brief Initialize SIMIX internal data. * * \param argc Argc @@ -60,6 +119,11 @@ void SIMIX_global_init(int *argc, char **argv) if (!simix_global) { simix_global = xbt_new0(s_smx_global_t, 1); +#ifdef TIME_BENCH_AMDAHL + simix_global->timer_seq = xbt_os_timer_new(); + simix_global->timer_par = xbt_os_timer_new(); + xbt_os_cputimer_start(simix_global->timer_seq); +#endif simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL); simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL); simix_global->process_list = @@ -76,6 +140,7 @@ void SIMIX_global_init(int *argc, char **argv) simix_global->action_mallocator = xbt_mallocator_new(65536, SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f, SIMIX_action_mallocator_reset_f); + simix_global->autorestart = SIMIX_host_restart_processes; surf_init(argc, argv); /* Initialize SURF structures */ SIMIX_context_mod_init(); @@ -85,33 +150,49 @@ void SIMIX_global_init(int *argc, char **argv) __xbt_running_ctx_fetch = SIMIX_process_get_running_context; __xbt_ex_terminate = SIMIX_process_exception_terminate; - /* Initialize the SIMIX network module */ SIMIX_network_init(); /* Prepare to display some more info when dying on Ctrl-C pressing */ signal(SIGINT, inthandler); + + /* Install SEGV handler */ + install_segvhandler(); + + /* register a function to be called by SURF after the environment creation */ + sg_platf_init(); + sg_platf_postparse_add_cb(SIMIX_post_create_environment); + } if (!simix_timers) { simix_timers = xbt_heap_new(8, &free); } - XBT_DEBUG("ADD SIMIX LEVELS"); SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy); + SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy); + + if (sg_cfg_get_boolean("clean_atexit")) + atexit(SIMIX_clean); + + if (_sg_cfg_exit_asap) + exit(0); } /** + * \ingroup SIMIX_API * \brief Clean the SIMIX simulation * * This functions remove the memory used by SIMIX */ +int cleaned = 0; void SIMIX_clean(void) { -#ifdef TIME_BENCH +#ifdef TIME_BENCH_PER_SR smx_ctx_raw_new_sr(); #endif - + if (cleaned) return; // to avoid double cleaning by java and C + cleaned = 1; /* Kill everyone (except maestro) */ - SIMIX_process_killall(simix_global->maestro_process); + SIMIX_process_killall(simix_global->maestro_process, 1); /* Exit the SIMIX network module */ SIMIX_network_exit(); @@ -142,6 +223,15 @@ void SIMIX_clean(void) surf_exit(); +#ifdef TIME_BENCH_AMDAHL + xbt_os_cputimer_stop(simix_global->timer_seq); + XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f", + xbt_os_timer_elapsed(simix_global->timer_seq), + xbt_os_timer_elapsed(simix_global->timer_par)); + xbt_os_timer_free(simix_global->timer_seq); + xbt_os_timer_free(simix_global->timer_par); +#endif + xbt_mallocator_free(simix_global->action_mallocator); xbt_free(simix_global); simix_global = NULL; @@ -151,24 +241,41 @@ void SIMIX_clean(void) /** + * \ingroup SIMIX_API * \brief A clock (in second). * * \return Return the clock. */ XBT_INLINE double SIMIX_get_clock(void) { - if(MC_IS_ENABLED){ + if(MC_is_active()){ return MC_process_clock_get(SIMIX_process_self()); }else{ return surf_get_clock(); } } +static int process_syscall_color(void *p) +{ + switch ((*(smx_process_t *)p)->simcall.call) { + case SIMCALL_NONE: + case SIMCALL_PROCESS_KILL: + return 2; + case SIMCALL_PROCESS_RESUME: + return 1; + default: + return 0; + } +} + +/** + * \ingroup SIMIX_API + * \brief Run the main simulation loop. + */ void SIMIX_run(void) { double time = 0; smx_process_t process; - xbt_swag_t set; surf_action_t action; smx_timer_t timer; surf_model_t model; @@ -177,13 +284,82 @@ void SIMIX_run(void) do { XBT_DEBUG("New Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run)); -#ifdef TIME_BENCH +#ifdef TIME_BENCH_PER_SR smx_ctx_raw_new_sr(); #endif while (!xbt_dynar_is_empty(simix_global->process_to_run)) { XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run)); + + /* Run all processes that are ready to run, possibly in parallel */ +#ifdef TIME_BENCH_AMDAHL + xbt_os_cputimer_stop(simix_global->timer_seq); + xbt_os_cputimer_resume(simix_global->timer_par); +#endif SIMIX_process_runall(); +#ifdef TIME_BENCH_AMDAHL + xbt_os_cputimer_stop(simix_global->timer_par); + xbt_os_cputimer_resume(simix_global->timer_seq); +#endif + + /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */ + xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color); + + /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */ + + /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */ + + /* Here, the order is ok because: + * + * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order. + * + * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases. + * + * - if there is no kill during the simulation, processes remain sorted according by their PID. + * rational: This can be proved inductively. + * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly). + * Let's show that it is still so at the end of this round. + * - if a process is added when being created, that's from maestro. It can be either at startup + * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled + * in arbitrary order (inductive hypothesis), we are fine. + * - If a process is added because it's getting killed, its subsequent actions shouldn't matter + * - If a process gets added to process_to_run because one of their blocking action constituting the meat + * of a simcall terminates, we're still good. Proof: + * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource + * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example. + * - For communications, this function is called from SIMIX_comm_finish(). + * This function itself don't mess with the order since simcalls are handled in FIFO order. + * The function is called: + * - before the comm starts (invalid parameters, or resource already dead or whatever). + * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet + * - because the communication failed or were canceled after startup. In this case, it's called from the function + * we are in, by the chunk: + * set = model->states.failed_action_set; + * while ((action = xbt_swag_extract(set))) + * SIMIX_simcall_post((smx_action_t) action->data); + * This order is also fixed because it depends of the order in which the surf actions were + * added to the system, and only maestro can add stuff this way, through simcalls. + * We thus use the inductive hypothesis once again to conclude that the order in which actions are + * poped out of the swag does not depend on the user code's execution order. + * - because the communication terminated. In this case, actions are served in the order given by + * set = model->states.done_action_set; + * while ((action = xbt_swag_extract(set))) + * SIMIX_simcall_post((smx_action_t) action->data); + * and the argument is very similar to the previous one. + * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed. + * 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. + * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized. + * - If there is some process killings, the order is changed by this decision that comes from user-land + * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible. + * So, even the order change induced by the process killing is perfectly reproducible. + * + * So science works, bitches [http://xkcd.com/54/]. + * + * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled + * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly). + * That would thus be a pure waste of time. + */ + xbt_dynar_foreach(simix_global->process_that_ran, iter, process) { if (process->simcall.call != SIMCALL_NONE) { SIMIX_simcall_pre(&process->simcall, 0); @@ -202,24 +378,43 @@ void SIMIX_run(void) /* Handle any pending timer */ while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) { //FIXME: make the timers being real callbacks - // (i.e. provide dispatchers that read and expand the args) + // (i.e. provide dispatchers that read and expand the args) timer = xbt_heap_pop(simix_timers); if (timer->func) ((void (*)(void*))timer->func)(timer->args); + xbt_free(timer); } + /* Wake up all processes waiting for a Surf action to finish */ xbt_dynar_foreach(model_list, iter, model) { - set = model->states.failed_action_set; - while ((action = xbt_swag_extract(set))) - SIMIX_simcall_post((smx_action_t) action->data); - set = model->states.done_action_set; - while ((action = xbt_swag_extract(set))) - SIMIX_simcall_post((smx_action_t) action->data); + while ((action = surf_model_extract_failed_action_set(model))) + SIMIX_simcall_post((smx_action_t) surf_action_get_data(action)); + + while ((action = surf_model_extract_done_action_set(model))) + if (surf_action_get_data(action) == NULL) + XBT_DEBUG("probably vcpu's action %p, skip", action); + else + SIMIX_simcall_post((smx_action_t) surf_action_get_data(action)); + } + + /* Autorestart all process */ + if(host_that_restart) { + char *hostname = NULL; + xbt_dynar_foreach(host_that_restart,iter,hostname) { + XBT_INFO("Restart processes on host: %s",hostname); + SIMIX_host_autorestart(SIMIX_host_get_by_name(hostname)); + } + xbt_dynar_reset(host_that_restart); } /* Clean processes to destroy */ SIMIX_process_empty_trash(); + + XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run)); + // !(time == -1.0 && xbt_dynar_is_empty()) + + } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run)); if (xbt_swag_size(simix_global->process_list) != 0) { @@ -228,19 +423,19 @@ void SIMIX_run(void) TRACE_end(); #endif - XBT_WARN("Oops ! Deadlock or code not perfectly clean."); + XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean."); SIMIX_display_process_status(); xbt_abort(); } } /** - * \brief Set the date to execute a function + * \brief Set the date to execute a function * * Set the date to execute the function on the surf. - * \param date Date to execute function - * \param function Function to be executed - * \param arg Parameters of the function + * \param date Date to execute function + * \param function Function to be executed + * \param arg Parameters of the function * */ XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg) @@ -275,7 +470,7 @@ XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t * * \param function Kill process function */ -XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t +XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t function) { simix_global->kill_process_function = function; @@ -338,15 +533,22 @@ void SIMIX_display_process_status(void) case SIMIX_ACTION_IO: action_description = "I/O"; + break; + /* **************************************/ + /* TUTORIAL: New API */ + case SIMIX_ACTION_NEW_API: + action_description = "NEW API"; + /* **************************************/ + break; } XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish", - process->pid, process->name, process->smx_host->name, + process->pid, process->name, sg_host_name(process->smx_host), action_description, process->waiting_action, process->waiting_action->name, (int)process->waiting_action->state); } else { - XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, process->smx_host->name); + XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host)); } } } @@ -370,3 +572,12 @@ static void SIMIX_action_mallocator_reset_f(void* action) { memset(action, 0, sizeof(s_smx_action_t)); ((smx_action_t) action)->simcalls = fifo; } + +xbt_dict_t SIMIX_pre_asr_get_properties(smx_simcall_t simcall, const char *name){ + return SIMIX_asr_get_properties(name); +} +xbt_dict_t SIMIX_asr_get_properties(const char *name) +{ + return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL); +} +