X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/39b6ed28636c4d67a0127d8526bf10dbc212cdc9..6a9d0cb70ec8497d66b6e7f95f2a08b5a2cf8da9:/src/simix/smx_global.c diff --git a/src/simix/smx_global.c b/src/simix/smx_global.c index 2cb2b39f0a..eb2a14180f 100644 --- a/src/simix/smx_global.c +++ b/src/simix/smx_global.c @@ -4,7 +4,7 @@ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ -#include "private.h" +#include "smx_private.h" #include "xbt/heap.h" #include "xbt/sysdep.h" #include "xbt/log.h" @@ -12,14 +12,7 @@ #include "xbt/ex.h" /* ex_backtrace_display */ #include "mc/mc.h" -XBT_LOG_EXTERNAL_CATEGORY(simix); -XBT_LOG_EXTERNAL_CATEGORY(simix_action); -XBT_LOG_EXTERNAL_CATEGORY(simix_deployment); -XBT_LOG_EXTERNAL_CATEGORY(simix_environment); -XBT_LOG_EXTERNAL_CATEGORY(simix_host); -XBT_LOG_EXTERNAL_CATEGORY(simix_process); -XBT_LOG_EXTERNAL_CATEGORY(simix_synchro); -XBT_LOG_EXTERNAL_CATEGORY(simix_context); +XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories"); XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)"); @@ -43,7 +36,7 @@ static void _XBT_CALL inthandler(int ignored) SIMIX_display_process_status(); } else { - XBT_INFO("CTRL-C pressed. bailing out without displaying because verbose-exit disabled"); + XBT_INFO("CTRL-C pressed. bailing out without displaying because verbose-exit is disabled"); } exit(1); } @@ -66,16 +59,6 @@ void SIMIX_global_init(int *argc, char **argv) s_smx_process_t proc; if (!simix_global) { - /* Connect our log channels: that must be done manually under windows */ - XBT_LOG_CONNECT(simix_action, simix); - XBT_LOG_CONNECT(simix_deployment, simix); - XBT_LOG_CONNECT(simix_environment, simix); - XBT_LOG_CONNECT(simix_host, simix); - XBT_LOG_CONNECT(simix_kernel, simix); - XBT_LOG_CONNECT(simix_process, simix); - XBT_LOG_CONNECT(simix_synchro, simix); - XBT_LOG_CONNECT(simix_context, simix); - simix_global = xbt_new0(s_smx_global_t, 1); simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL); @@ -86,7 +69,7 @@ void SIMIX_global_init(int *argc, char **argv) xbt_swag_new(xbt_swag_offset(proc, destroy_hookup)); simix_global->maestro_process = NULL; - simix_global->registered_functions = xbt_dict_new(); + simix_global->registered_functions = xbt_dict_new_homogeneous(NULL); simix_global->create_process_function = SIMIX_process_create; simix_global->kill_process_function = SIMIX_process_kill; @@ -135,6 +118,7 @@ void SIMIX_clean(void) SIMIX_network_exit(); xbt_heap_free(simix_timers); + simix_timers = NULL; /* Free the remaining data structures */ xbt_dynar_free(&simix_global->process_to_run); xbt_dynar_free(&simix_global->process_that_ran); @@ -181,6 +165,19 @@ XBT_INLINE double SIMIX_get_clock(void) } } +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; + } +} + void SIMIX_run(void) { double time = 0; @@ -197,19 +194,76 @@ void SIMIX_run(void) #ifdef TIME_BENCH smx_ctx_raw_new_sr(); #endif - while (xbt_dynar_length(simix_global->process_to_run)) { + 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 */ SIMIX_process_runall(); + + /* 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: + * - 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->request.call != REQ_NO_REQ) { - XBT_DEBUG("Handling request %p", &process->request); - SIMIX_request_pre(&process->request, 0); + if (process->simcall.call != SIMCALL_NONE) { + SIMIX_simcall_pre(&process->simcall, 0); } } } - time = surf_solve(SIMIX_timer_next()); + time = SIMIX_timer_next(); + if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0) + time = surf_solve(time); /* Notify all the hosts that have failed */ /* FIXME: iterate through the list of failed host and mark each of them */ @@ -223,22 +277,20 @@ void SIMIX_run(void) if (timer->func) ((void (*)(void*))timer->func)(timer->args); } - /* Wake up all process waiting for the action finish */ + /* Wake up all processes waiting for a Surf action to finish */ xbt_dynar_foreach(model_list, iter, model) { - for (set = model->states.failed_action_set; - set; - set = (set == model->states.failed_action_set) - ? model->states.done_action_set - : NULL) { - while ((action = xbt_swag_extract(set))) - SIMIX_request_post((smx_action_t) action->data); - } + 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); } /* Clean processes to destroy */ SIMIX_process_empty_trash(); - } while (time != -1.0); + } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run)); if (xbt_swag_size(simix_global->process_list) != 0) { @@ -325,6 +377,8 @@ void SIMIX_display_process_status(void) XBT_INFO("%d processes are still running, waiting for something.", nbprocess); /* List the process and their state */ + XBT_INFO + ("Legend of the following listing: \"Process (@): \""); xbt_swag_foreach(process, simix_global->process_list) { if (process->waiting_action) { @@ -332,52 +386,57 @@ void SIMIX_display_process_status(void) const char* action_description = "unknown"; switch (process->waiting_action->type) { - case SIMIX_ACTION_EXECUTE: - action_description = "execution"; - break; + case SIMIX_ACTION_EXECUTE: + action_description = "execution"; + break; - case SIMIX_ACTION_PARALLEL_EXECUTE: - action_description = "parallel execution"; - break; + case SIMIX_ACTION_PARALLEL_EXECUTE: + action_description = "parallel execution"; + break; - case SIMIX_ACTION_COMMUNICATE: - action_description = "communication"; - break; + case SIMIX_ACTION_COMMUNICATE: + action_description = "communication"; + break; - case SIMIX_ACTION_SLEEP: - action_description = "sleeping"; - break; + case SIMIX_ACTION_SLEEP: + action_description = "sleeping"; + break; - case SIMIX_ACTION_SYNCHRO: - action_description = "synchronization"; - break; + case SIMIX_ACTION_SYNCHRO: + action_description = "synchronization"; + break; - case SIMIX_ACTION_IO: - action_description = "I/O"; - break; + case SIMIX_ACTION_IO: + action_description = "I/O"; + break; } - XBT_INFO("Process %ld (%s@%s): waiting for %s action %p to finish", process->pid, process->name, process->smx_host->name, - action_description, process->waiting_action); + 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, + 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); } } } static void* SIMIX_action_mallocator_new_f(void) { smx_action_t action = xbt_new(s_smx_action_t, 1); - action->request_list = xbt_fifo_new(); + action->simcalls = xbt_fifo_new(); return action; } static void SIMIX_action_mallocator_free_f(void* action) { - xbt_fifo_free(((smx_action_t) action)->request_list); + xbt_fifo_free(((smx_action_t) action)->simcalls); xbt_free(action); } static void SIMIX_action_mallocator_reset_f(void* action) { - // we also recycle the request list - xbt_fifo_t fifo = ((smx_action_t) action)->request_list; + // we also recycle the simcall list + xbt_fifo_t fifo = ((smx_action_t) action)->simcalls; xbt_fifo_reset(fifo); memset(action, 0, sizeof(s_smx_action_t)); - ((smx_action_t) action)->request_list = fifo; + ((smx_action_t) action)->simcalls = fifo; }