1 /* Copyright (c) 2007-2012. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include "smx_private.h"
8 #include "xbt/sysdep.h"
11 #include "xbt/ex.h" /* ex_backtrace_display */
13 #include "simgrid/sg_config.h"
15 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
16 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
17 "Logging specific to SIMIX (kernel)");
19 smx_global_t simix_global = NULL;
20 static xbt_heap_t simix_timers = NULL;
22 static void* SIMIX_action_mallocator_new_f(void);
23 static void SIMIX_action_mallocator_free_f(void* action);
24 static void SIMIX_action_mallocator_reset_f(void* action);
26 static void SIMIX_clean(void);
28 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
31 int _sg_do_verbose_exit = 1;
32 static void _XBT_CALL inthandler(int ignored)
34 if ( _sg_do_verbose_exit ) {
35 XBT_INFO("CTRL-C pressed. Displaying status and bailing out");
36 SIMIX_display_process_status();
39 XBT_INFO("CTRL-C pressed. bailing out without displaying because verbose-exit is disabled");
44 /********************************* SIMIX **************************************/
46 XBT_INLINE double SIMIX_timer_next(void)
48 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
53 * \brief Initialize SIMIX internal data.
58 void SIMIX_global_init(int *argc, char **argv)
63 simix_global = xbt_new0(s_smx_global_t, 1);
65 #ifdef TIME_BENCH_AMDAHL
66 simix_global->timer_seq = xbt_os_timer_new();
67 simix_global->timer_par = xbt_os_timer_new();
68 xbt_os_cputimer_start(simix_global->timer_seq);
70 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
71 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
72 simix_global->process_list =
73 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
74 simix_global->process_to_destroy =
75 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
77 simix_global->maestro_process = NULL;
78 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
80 simix_global->create_process_function = SIMIX_process_create;
81 simix_global->kill_process_function = SIMIX_process_kill;
82 simix_global->cleanup_process_function = SIMIX_process_cleanup;
83 simix_global->action_mallocator = xbt_mallocator_new(65536,
84 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
85 SIMIX_action_mallocator_reset_f);
86 simix_global->autorestart = SIMIX_host_restart_processes;
88 surf_init(argc, argv); /* Initialize SURF structures */
89 SIMIX_context_mod_init();
90 SIMIX_create_maestro_process();
92 /* context exception handlers */
93 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
94 __xbt_ex_terminate = SIMIX_process_exception_terminate;
98 /* Prepare to display some more info when dying on Ctrl-C pressing */
99 signal(SIGINT, inthandler);
101 /* register a function to be called by SURF after the environment creation */
103 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
107 simix_timers = xbt_heap_new(8, &free);
110 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
112 if(sg_cfg_get_boolean("clean_atexit")) atexit(SIMIX_clean);
117 * \brief Clean the SIMIX simulation
119 * This functions remove the memory used by SIMIX
121 static void SIMIX_clean(void)
123 #ifdef TIME_BENCH_PER_SR
124 smx_ctx_raw_new_sr();
127 /* Kill everyone (except maestro) */
128 SIMIX_process_killall(simix_global->maestro_process, 1);
130 /* Exit the SIMIX network module */
131 SIMIX_network_exit();
133 xbt_heap_free(simix_timers);
135 /* Free the remaining data structures */
136 xbt_dynar_free(&simix_global->process_to_run);
137 xbt_dynar_free(&simix_global->process_that_ran);
138 xbt_swag_free(simix_global->process_to_destroy);
139 xbt_swag_free(simix_global->process_list);
140 simix_global->process_list = NULL;
141 simix_global->process_to_destroy = NULL;
142 xbt_dict_free(&(simix_global->registered_functions));
144 /* Let's free maestro now */
145 SIMIX_context_free(simix_global->maestro_process->context);
146 xbt_free(simix_global->maestro_process->running_ctx);
147 xbt_free(simix_global->maestro_process);
148 simix_global->maestro_process = NULL;
150 /* Restore the default exception setup */
151 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
152 __xbt_ex_terminate = &__xbt_ex_terminate_default;
154 /* Finish context module and SURF */
155 SIMIX_context_mod_exit();
159 #ifdef TIME_BENCH_AMDAHL
160 xbt_os_cputimer_stop(simix_global->timer_seq);
161 XBT_INFO("Amdahl timing informations. Sequential time: %lf; Parallel time: %lf",
162 xbt_os_timer_elapsed(simix_global->timer_seq),
163 xbt_os_timer_elapsed(simix_global->timer_par));
164 xbt_os_timer_free(simix_global->timer_seq);
165 xbt_os_timer_free(simix_global->timer_par);
168 xbt_mallocator_free(simix_global->action_mallocator);
169 xbt_free(simix_global);
178 * \brief A clock (in second).
180 * \return Return the clock.
182 XBT_INLINE double SIMIX_get_clock(void)
185 return MC_process_clock_get(SIMIX_process_self());
187 return surf_get_clock();
191 static int process_syscall_color(void *p)
193 switch ((*(smx_process_t *)p)->simcall.call) {
195 case SIMCALL_PROCESS_KILL:
197 case SIMCALL_PROCESS_RESUME:
206 * \brief Run the main simulation loop.
211 smx_process_t process;
213 surf_action_t action;
219 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
220 xbt_dynar_length(simix_global->process_to_run));
221 #ifdef TIME_BENCH_PER_SR
222 smx_ctx_raw_new_sr();
224 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
225 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
226 xbt_dynar_length(simix_global->process_to_run));
228 /* Run all processes that are ready to run, possibly in parallel */
229 #ifdef TIME_BENCH_AMDAHL
230 xbt_os_cputimer_stop(simix_global->timer_seq);
231 xbt_os_cputimer_resume(simix_global->timer_par);
233 SIMIX_process_runall();
234 #ifdef TIME_BENCH_AMDAHL
235 xbt_os_cputimer_stop(simix_global->timer_par);
236 xbt_os_cputimer_resume(simix_global->timer_seq);
239 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
240 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
242 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
244 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
246 /* Here, the order is ok because:
248 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
250 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
252 * - if there is no kill during the simulation, processes remain sorted according by their PID.
253 * rational: This can be proved inductively.
254 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
255 * Let's show that it is still so at the end of this round.
256 * - if a process is added when being created, that's from maestro. It can be either at startup
257 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
258 * in arbitrary order (inductive hypothesis), we are fine.
259 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
260 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
261 * of a simcall terminates, we're still good. Proof:
262 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
263 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
264 * - For communications, this function is called from SIMIX_comm_finish().
265 * This function itself don't mess with the order since simcalls are handled in FIFO order.
266 * The function is called:
267 * - before the comm starts (invalid parameters, or resource already dead or whatever).
268 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
269 * - because the communication failed or were canceled after startup. In this case, it's called from the function
270 * we are in, by the chunk:
271 * set = model->states.failed_action_set;
272 * while ((action = xbt_swag_extract(set)))
273 * SIMIX_simcall_post((smx_action_t) action->data);
274 * This order is also fixed because it depends of the order in which the surf actions were
275 * added to the system, and only maestro can add stuff this way, through simcalls.
276 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
277 * poped out of the swag does not depend on the user code's execution order.
278 * - because the communication terminated. In this case, actions are served in the order given by
279 * set = model->states.done_action_set;
280 * while ((action = xbt_swag_extract(set)))
281 * SIMIX_simcall_post((smx_action_t) action->data);
282 * and the argument is very similar to the previous one.
283 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
284 * 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.
285 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
286 * - If there is some process killings, the order is changed by this decision that comes from user-land
287 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
288 * So, even the order change induced by the process killing is perfectly reproducible.
290 * So science works, bitches [http://xkcd.com/54/].
292 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
293 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
294 * That would thus be a pure waste of time.
297 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
298 if (process->simcall.call != SIMCALL_NONE) {
299 SIMIX_simcall_pre(&process->simcall, 0);
304 time = SIMIX_timer_next();
305 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
306 time = surf_solve(time);
308 /* Notify all the hosts that have failed */
309 /* FIXME: iterate through the list of failed host and mark each of them */
310 /* as failed. On each host, signal all the running processes with host_fail */
312 /* Handle any pending timer */
313 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
314 //FIXME: make the timers being real callbacks
315 // (i.e. provide dispatchers that read and expand the args)
316 timer = xbt_heap_pop(simix_timers);
318 ((void (*)(void*))timer->func)(timer->args);
321 /* Wake up all processes waiting for a Surf action to finish */
322 xbt_dynar_foreach(model_list, iter, model) {
323 set = model->states.failed_action_set;
324 while ((action = xbt_swag_extract(set)))
325 SIMIX_simcall_post((smx_action_t) action->data);
326 set = model->states.done_action_set;
327 while ((action = xbt_swag_extract(set)))
328 SIMIX_simcall_post((smx_action_t) action->data);
331 /* Autorestart all process */
332 if(host_that_restart) {
333 char *hostname = NULL;
334 xbt_dynar_foreach(host_that_restart,iter,hostname) {
335 XBT_INFO("Restart processes on host: %s",hostname);
336 SIMIX_host_autorestart(SIMIX_host_get_by_name(hostname));
338 xbt_dynar_reset(host_that_restart);
341 /* Clean processes to destroy */
342 SIMIX_process_empty_trash();
344 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
346 if (xbt_swag_size(simix_global->process_list) != 0) {
352 XBT_WARN("Oops ! Deadlock or code not perfectly clean.");
353 SIMIX_display_process_status();
359 * \brief Set the date to execute a function
361 * Set the date to execute the function on the surf.
362 * \param date Date to execute function
363 * \param function Function to be executed
364 * \param arg Parameters of the function
367 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
369 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
372 timer->func = function;
374 xbt_heap_push(simix_timers, timer, date);
378 * \brief Registers a function to create a process.
380 * This function registers a function to be called
381 * when a new process is created. The function has
382 * to call SIMIX_process_create().
383 * \param function create process function
385 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
388 simix_global->create_process_function = function;
392 * \brief Registers a function to kill a process.
394 * This function registers a function to be called when a
395 * process is killed. The function has to call the SIMIX_process_kill().
397 * \param function Kill process function
399 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
402 simix_global->kill_process_function = function;
406 * \brief Registers a function to cleanup a process.
408 * This function registers a user function to be called when
409 * a process ends properly.
411 * \param function cleanup process function
413 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
416 simix_global->cleanup_process_function = function;
420 void SIMIX_display_process_status(void)
422 if (simix_global->process_list == NULL) {
426 smx_process_t process = NULL;
427 int nbprocess = xbt_swag_size(simix_global->process_list);
429 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
430 /* List the process and their state */
432 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
433 xbt_swag_foreach(process, simix_global->process_list) {
435 if (process->waiting_action) {
437 const char* action_description = "unknown";
438 switch (process->waiting_action->type) {
440 case SIMIX_ACTION_EXECUTE:
441 action_description = "execution";
444 case SIMIX_ACTION_PARALLEL_EXECUTE:
445 action_description = "parallel execution";
448 case SIMIX_ACTION_COMMUNICATE:
449 action_description = "communication";
452 case SIMIX_ACTION_SLEEP:
453 action_description = "sleeping";
456 case SIMIX_ACTION_SYNCHRO:
457 action_description = "synchronization";
460 case SIMIX_ACTION_IO:
461 action_description = "I/O";
463 /* **************************************/
464 /* TUTORIAL: New API */
465 case SIMIX_ACTION_NEW_API:
466 action_description = "NEW API";
467 /* **************************************/
471 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
472 process->pid, process->name, sg_host_name(process->smx_host),
473 action_description, process->waiting_action,
474 process->waiting_action->name, (int)process->waiting_action->state);
477 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host));
482 static void* SIMIX_action_mallocator_new_f(void) {
483 smx_action_t action = xbt_new(s_smx_action_t, 1);
484 action->simcalls = xbt_fifo_new();
488 static void SIMIX_action_mallocator_free_f(void* action) {
489 xbt_fifo_free(((smx_action_t) action)->simcalls);
493 static void SIMIX_action_mallocator_reset_f(void* action) {
495 // we also recycle the simcall list
496 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
497 xbt_fifo_reset(fifo);
498 memset(action, 0, sizeof(s_smx_action_t));
499 ((smx_action_t) action)->simcalls = fifo;
502 xbt_dict_t SIMIX_pre_asr_get_properties(smx_simcall_t simcall, const char *name){
503 return SIMIX_asr_get_properties(name);
505 xbt_dict_t SIMIX_asr_get_properties(const char *name)
507 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);