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 static void SIMIX_clean(void);
29 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
32 int _sg_do_verbose_exit = 1;
33 static void _XBT_CALL inthandler(int ignored)
35 if ( _sg_do_verbose_exit ) {
36 XBT_INFO("CTRL-C pressed. Displaying status and bailing out");
37 SIMIX_display_process_status();
40 XBT_INFO("CTRL-C pressed. bailing out without displaying because verbose-exit is disabled");
45 /********************************* SIMIX **************************************/
47 XBT_INLINE double SIMIX_timer_next(void)
49 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
54 * \brief Initialize SIMIX internal data.
59 void SIMIX_global_init(int *argc, char **argv)
64 simix_global = xbt_new0(s_smx_global_t, 1);
66 #ifdef TIME_BENCH_AMDAHL
67 simix_global->timer_seq = xbt_os_timer_new();
68 simix_global->timer_par = xbt_os_timer_new();
69 xbt_os_cputimer_start(simix_global->timer_seq);
71 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
72 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
73 simix_global->process_list =
74 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
75 simix_global->process_to_destroy =
76 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
78 simix_global->maestro_process = NULL;
79 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
81 simix_global->create_process_function = SIMIX_process_create;
82 simix_global->kill_process_function = SIMIX_process_kill;
83 simix_global->cleanup_process_function = SIMIX_process_cleanup;
84 simix_global->action_mallocator = xbt_mallocator_new(65536,
85 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
86 SIMIX_action_mallocator_reset_f);
87 simix_global->autorestart = SIMIX_host_restart_processes;
89 surf_init(argc, argv); /* Initialize SURF structures */
90 SIMIX_context_mod_init();
91 SIMIX_create_maestro_process();
93 /* context exception handlers */
94 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
95 __xbt_ex_terminate = SIMIX_process_exception_terminate;
99 /* Prepare to display some more info when dying on Ctrl-C pressing */
100 signal(SIGINT, inthandler);
102 /* register a function to be called by SURF after the environment creation */
104 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
108 simix_timers = xbt_heap_new(8, &free);
111 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
112 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
114 if (sg_cfg_get_boolean("clean_atexit"))
117 if (_sg_cfg_exit_asap)
123 * \brief Clean the SIMIX simulation
125 * This functions remove the memory used by SIMIX
127 static void SIMIX_clean(void)
129 #ifdef TIME_BENCH_PER_SR
130 smx_ctx_raw_new_sr();
133 /* Kill everyone (except maestro) */
134 SIMIX_process_killall(simix_global->maestro_process, 1);
136 /* Exit the SIMIX network module */
137 SIMIX_network_exit();
139 xbt_heap_free(simix_timers);
141 /* Free the remaining data structures */
142 xbt_dynar_free(&simix_global->process_to_run);
143 xbt_dynar_free(&simix_global->process_that_ran);
144 xbt_swag_free(simix_global->process_to_destroy);
145 xbt_swag_free(simix_global->process_list);
146 simix_global->process_list = NULL;
147 simix_global->process_to_destroy = NULL;
148 xbt_dict_free(&(simix_global->registered_functions));
150 /* Let's free maestro now */
151 SIMIX_context_free(simix_global->maestro_process->context);
152 xbt_free(simix_global->maestro_process->running_ctx);
153 xbt_free(simix_global->maestro_process);
154 simix_global->maestro_process = NULL;
156 /* Restore the default exception setup */
157 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
158 __xbt_ex_terminate = &__xbt_ex_terminate_default;
160 /* Finish context module and SURF */
161 SIMIX_context_mod_exit();
165 #ifdef TIME_BENCH_AMDAHL
166 xbt_os_cputimer_stop(simix_global->timer_seq);
167 XBT_INFO("Amdahl timing informations. Sequential time: %f; Parallel time: %f",
168 xbt_os_timer_elapsed(simix_global->timer_seq),
169 xbt_os_timer_elapsed(simix_global->timer_par));
170 xbt_os_timer_free(simix_global->timer_seq);
171 xbt_os_timer_free(simix_global->timer_par);
174 xbt_mallocator_free(simix_global->action_mallocator);
175 xbt_free(simix_global);
184 * \brief A clock (in second).
186 * \return Return the clock.
188 XBT_INLINE double SIMIX_get_clock(void)
191 return MC_process_clock_get(SIMIX_process_self());
193 return surf_get_clock();
197 static int process_syscall_color(void *p)
199 switch ((*(smx_process_t *)p)->simcall.call) {
201 case SIMCALL_PROCESS_KILL:
203 case SIMCALL_PROCESS_RESUME:
212 * \brief Run the main simulation loop.
217 smx_process_t process;
218 surf_action_t action;
224 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
225 xbt_dynar_length(simix_global->process_to_run));
226 #ifdef TIME_BENCH_PER_SR
227 smx_ctx_raw_new_sr();
229 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
230 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
231 xbt_dynar_length(simix_global->process_to_run));
233 /* Run all processes that are ready to run, possibly in parallel */
234 #ifdef TIME_BENCH_AMDAHL
235 xbt_os_cputimer_stop(simix_global->timer_seq);
236 xbt_os_cputimer_resume(simix_global->timer_par);
238 SIMIX_process_runall();
239 #ifdef TIME_BENCH_AMDAHL
240 xbt_os_cputimer_stop(simix_global->timer_par);
241 xbt_os_cputimer_resume(simix_global->timer_seq);
244 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
245 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
247 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
249 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
251 /* Here, the order is ok because:
253 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
255 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
257 * - if there is no kill during the simulation, processes remain sorted according by their PID.
258 * rational: This can be proved inductively.
259 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
260 * Let's show that it is still so at the end of this round.
261 * - if a process is added when being created, that's from maestro. It can be either at startup
262 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
263 * in arbitrary order (inductive hypothesis), we are fine.
264 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
265 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
266 * of a simcall terminates, we're still good. Proof:
267 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
268 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
269 * - For communications, this function is called from SIMIX_comm_finish().
270 * This function itself don't mess with the order since simcalls are handled in FIFO order.
271 * The function is called:
272 * - before the comm starts (invalid parameters, or resource already dead or whatever).
273 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
274 * - because the communication failed or were canceled after startup. In this case, it's called from the function
275 * we are in, by the chunk:
276 * set = model->states.failed_action_set;
277 * while ((action = xbt_swag_extract(set)))
278 * SIMIX_simcall_post((smx_action_t) action->data);
279 * This order is also fixed because it depends of the order in which the surf actions were
280 * added to the system, and only maestro can add stuff this way, through simcalls.
281 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
282 * poped out of the swag does not depend on the user code's execution order.
283 * - because the communication terminated. In this case, actions are served in the order given by
284 * set = model->states.done_action_set;
285 * while ((action = xbt_swag_extract(set)))
286 * SIMIX_simcall_post((smx_action_t) action->data);
287 * and the argument is very similar to the previous one.
288 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
289 * 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.
290 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
291 * - If there is some process killings, the order is changed by this decision that comes from user-land
292 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
293 * So, even the order change induced by the process killing is perfectly reproducible.
295 * So science works, bitches [http://xkcd.com/54/].
297 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
298 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
299 * That would thus be a pure waste of time.
302 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
303 if (process->simcall.call != SIMCALL_NONE) {
304 SIMIX_simcall_pre(&process->simcall, 0);
309 time = SIMIX_timer_next();
310 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
311 time = surf_solve(time);
313 /* Notify all the hosts that have failed */
314 /* FIXME: iterate through the list of failed host and mark each of them */
315 /* as failed. On each host, signal all the running processes with host_fail */
317 /* Handle any pending timer */
318 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
319 //FIXME: make the timers being real callbacks
320 // (i.e. provide dispatchers that read and expand the args)
321 timer = xbt_heap_pop(simix_timers);
323 ((void (*)(void*))timer->func)(timer->args);
327 /* Wake up all processes waiting for a Surf action to finish */
328 xbt_dynar_foreach(model_list, iter, model) {
329 while ((action = surf_model_extract_failed_action_set(model)))
331 SIMIX_simcall_post((smx_action_t) surf_action_get_data(action));
333 while ((action = surf_model_extract_done_action_set(model)))
334 if (surf_action_get_data(action) == NULL)
335 XBT_DEBUG("probably vcpu's action %p, skip", action);
337 SIMIX_simcall_post((smx_action_t) surf_action_get_data(action));
340 /* Autorestart all process */
341 if(host_that_restart) {
342 char *hostname = NULL;
343 xbt_dynar_foreach(host_that_restart,iter,hostname) {
344 XBT_INFO("Restart processes on host: %s",hostname);
345 SIMIX_host_autorestart(SIMIX_host_get_by_name(hostname));
347 xbt_dynar_reset(host_that_restart);
350 /* Clean processes to destroy */
351 SIMIX_process_empty_trash();
354 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
355 // !(time == -1.0 && xbt_dynar_is_empty())
358 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
360 if (xbt_swag_size(simix_global->process_list) != 0) {
366 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
367 SIMIX_display_process_status();
373 * \brief Set the date to execute a function
375 * Set the date to execute the function on the surf.
376 * \param date Date to execute function
377 * \param function Function to be executed
378 * \param arg Parameters of the function
381 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
383 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
386 timer->func = function;
388 xbt_heap_push(simix_timers, timer, date);
392 * \brief Registers a function to create a process.
394 * This function registers a function to be called
395 * when a new process is created. The function has
396 * to call SIMIX_process_create().
397 * \param function create process function
399 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
402 simix_global->create_process_function = function;
406 * \brief Registers a function to kill a process.
408 * This function registers a function to be called when a
409 * process is killed. The function has to call the SIMIX_process_kill().
411 * \param function Kill process function
413 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
416 simix_global->kill_process_function = function;
420 * \brief Registers a function to cleanup a process.
422 * This function registers a user function to be called when
423 * a process ends properly.
425 * \param function cleanup process function
427 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
430 simix_global->cleanup_process_function = function;
434 void SIMIX_display_process_status(void)
436 if (simix_global->process_list == NULL) {
440 smx_process_t process = NULL;
441 int nbprocess = xbt_swag_size(simix_global->process_list);
443 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
444 /* List the process and their state */
446 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
447 xbt_swag_foreach(process, simix_global->process_list) {
449 if (process->waiting_action) {
451 const char* action_description = "unknown";
452 switch (process->waiting_action->type) {
454 case SIMIX_ACTION_EXECUTE:
455 action_description = "execution";
458 case SIMIX_ACTION_PARALLEL_EXECUTE:
459 action_description = "parallel execution";
462 case SIMIX_ACTION_COMMUNICATE:
463 action_description = "communication";
466 case SIMIX_ACTION_SLEEP:
467 action_description = "sleeping";
470 case SIMIX_ACTION_SYNCHRO:
471 action_description = "synchronization";
474 case SIMIX_ACTION_IO:
475 action_description = "I/O";
477 /* **************************************/
478 /* TUTORIAL: New API */
479 case SIMIX_ACTION_NEW_API:
480 action_description = "NEW API";
481 /* **************************************/
485 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
486 process->pid, process->name, sg_host_name(process->smx_host),
487 action_description, process->waiting_action,
488 process->waiting_action->name, (int)process->waiting_action->state);
491 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host));
496 static void* SIMIX_action_mallocator_new_f(void) {
497 smx_action_t action = xbt_new(s_smx_action_t, 1);
498 action->simcalls = xbt_fifo_new();
502 static void SIMIX_action_mallocator_free_f(void* action) {
503 xbt_fifo_free(((smx_action_t) action)->simcalls);
507 static void SIMIX_action_mallocator_reset_f(void* action) {
509 // we also recycle the simcall list
510 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
511 xbt_fifo_reset(fifo);
512 memset(action, 0, sizeof(s_smx_action_t));
513 ((smx_action_t) action)->simcalls = fifo;
516 xbt_dict_t SIMIX_pre_asr_get_properties(smx_simcall_t simcall, const char *name){
517 return SIMIX_asr_get_properties(name);
519 xbt_dict_t SIMIX_asr_get_properties(const char *name)
521 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);