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 */
14 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
15 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix,
16 "Logging specific to SIMIX (kernel)");
18 smx_global_t simix_global = NULL;
19 static xbt_heap_t simix_timers = NULL;
21 static void* SIMIX_action_mallocator_new_f(void);
22 static void SIMIX_action_mallocator_free_f(void* action);
23 static void SIMIX_action_mallocator_reset_f(void* action);
25 static void SIMIX_clean(void);
27 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
30 int _surf_do_verbose_exit = 1;
31 static void _XBT_CALL inthandler(int ignored)
33 if ( _surf_do_verbose_exit ) {
34 XBT_INFO("CTRL-C pressed. Displaying status and bailing out");
35 SIMIX_display_process_status();
38 XBT_INFO("CTRL-C pressed. bailing out without displaying because verbose-exit is disabled");
43 /********************************* SIMIX **************************************/
45 XBT_INLINE double SIMIX_timer_next(void)
47 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
52 * \brief Initialize SIMIX internal data.
57 void SIMIX_global_init(int *argc, char **argv)
62 simix_global = xbt_new0(s_smx_global_t, 1);
64 #ifdef TIME_BENCH_AMDAHL
65 simix_global->timer_seq = xbt_os_timer_new();
66 simix_global->timer_par = xbt_os_timer_new();
67 xbt_os_timer_start(simix_global->timer_seq);
69 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
70 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
71 simix_global->process_list =
72 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
73 simix_global->process_to_destroy =
74 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
76 simix_global->maestro_process = NULL;
77 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
79 simix_global->create_process_function = SIMIX_process_create;
80 simix_global->kill_process_function = SIMIX_process_kill;
81 simix_global->cleanup_process_function = SIMIX_process_cleanup;
82 simix_global->action_mallocator = xbt_mallocator_new(65536,
83 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
84 SIMIX_action_mallocator_reset_f);
85 simix_global->autorestart = SIMIX_host_restart_processes;
87 surf_init(argc, argv); /* Initialize SURF structures */
88 SIMIX_context_mod_init();
89 SIMIX_create_maestro_process();
91 /* context exception handlers */
92 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
93 __xbt_ex_terminate = SIMIX_process_exception_terminate;
95 /* Initialize the SIMIX network module */
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 XBT_DEBUG("ADD SIMIX LEVELS");
111 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
118 * \brief Clean the SIMIX simulation
120 * This functions remove the memory used by SIMIX
122 static void SIMIX_clean(void)
124 #ifdef TIME_BENCH_PER_SR
125 smx_ctx_raw_new_sr();
128 /* Kill everyone (except maestro) */
129 SIMIX_process_killall(simix_global->maestro_process);
131 /* Exit the SIMIX network module */
132 SIMIX_network_exit();
134 xbt_heap_free(simix_timers);
136 /* Free the remaining data structures */
137 xbt_dynar_free(&simix_global->process_to_run);
138 xbt_dynar_free(&simix_global->process_that_ran);
139 xbt_swag_free(simix_global->process_to_destroy);
140 xbt_swag_free(simix_global->process_list);
141 simix_global->process_list = NULL;
142 simix_global->process_to_destroy = NULL;
143 xbt_dict_free(&(simix_global->registered_functions));
145 /* Let's free maestro now */
146 SIMIX_context_free(simix_global->maestro_process->context);
147 xbt_free(simix_global->maestro_process->running_ctx);
148 xbt_free(simix_global->maestro_process);
149 simix_global->maestro_process = NULL;
151 /* Restore the default exception setup */
152 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
153 __xbt_ex_terminate = &__xbt_ex_terminate_default;
155 /* Finish context module and SURF */
156 SIMIX_context_mod_exit();
160 #ifdef TIME_BENCH_AMDAHL
161 xbt_os_timer_stop(simix_global->timer_seq);
162 XBT_INFO("Amdhal timing informations. Sequential time: %lf; Parallel time: %lf",
163 xbt_os_timer_elapsed(simix_global->timer_seq),
164 xbt_os_timer_elapsed(simix_global->timer_par));
165 xbt_os_timer_free(simix_global->timer_seq);
166 xbt_os_timer_free(simix_global->timer_par);
169 xbt_mallocator_free(simix_global->action_mallocator);
170 xbt_free(simix_global);
179 * \brief A clock (in second).
181 * \return Return the clock.
183 XBT_INLINE double SIMIX_get_clock(void)
186 return MC_process_clock_get(SIMIX_process_self());
188 return surf_get_clock();
192 static int process_syscall_color(void *p)
194 switch ((*(smx_process_t *)p)->simcall.call) {
196 case SIMCALL_PROCESS_KILL:
198 case SIMCALL_PROCESS_RESUME:
207 * \brief Run the main simulation loop.
212 smx_process_t process;
214 surf_action_t action;
220 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
221 xbt_dynar_length(simix_global->process_to_run));
222 #ifdef TIME_BENCH_PER_SR
223 smx_ctx_raw_new_sr();
225 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
226 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
227 xbt_dynar_length(simix_global->process_to_run));
229 /* Run all processes that are ready to run, possibly in parallel */
230 #ifdef TIME_BENCH_AMDAHL
231 xbt_os_timer_stop(simix_global->timer_seq);
232 xbt_os_timer_resume(simix_global->timer_par);
234 SIMIX_process_runall();
235 #ifdef TIME_BENCH_AMDAHL
236 xbt_os_timer_stop(simix_global->timer_par);
237 xbt_os_timer_resume(simix_global->timer_seq);
240 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
241 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
243 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
245 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
247 /* Here, the order is ok because:
249 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
251 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
253 * - if there is no kill during the simulation, processes remain sorted according by their PID.
254 * rational: This can be proved inductively.
255 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
256 * Let's show that it is still so at the end of this round.
257 * - if a process is added when being created, that's from maestro. It can be either at startup
258 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
259 * in arbitrary order (inductive hypothesis), we are fine.
260 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
261 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
262 * of a simcall terminates, we're still good. Proof:
263 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
264 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
265 * - For communications, this function is called from SIMIX_comm_finish().
266 * This function itself don't mess with the order since simcalls are handled in FIFO order.
267 * The function is called:
268 * - before the comm starts (invalid parameters, or resource already dead or whatever).
269 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
270 * - because the communication failed or were canceled after startup. In this case, it's called from the function
271 * we are in, by the chunk:
272 * set = model->states.failed_action_set;
273 * while ((action = xbt_swag_extract(set)))
274 * SIMIX_simcall_post((smx_action_t) action->data);
275 * This order is also fixed because it depends of the order in which the surf actions were
276 * added to the system, and only maestro can add stuff this way, through simcalls.
277 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
278 * poped out of the swag does not depend on the user code's execution order.
279 * - because the communication terminated. In this case, actions are served in the order given by
280 * set = model->states.done_action_set;
281 * while ((action = xbt_swag_extract(set)))
282 * SIMIX_simcall_post((smx_action_t) action->data);
283 * and the argument is very similar to the previous one.
284 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
285 * 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.
286 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
287 * - If there is some process killings, the order is changed by this decision that comes from user-land
288 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
289 * So, even the order change induced by the process killing is perfectly reproducible.
291 * So science works, bitches [http://xkcd.com/54/].
293 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
294 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
295 * That would thus be a pure waste of time.
298 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
299 if (process->simcall.call != SIMCALL_NONE) {
300 SIMIX_simcall_pre(&process->simcall, 0);
305 time = SIMIX_timer_next();
306 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
307 time = surf_solve(time);
309 /* Notify all the hosts that have failed */
310 /* FIXME: iterate through the list of failed host and mark each of them */
311 /* as failed. On each host, signal all the running processes with host_fail */
313 /* Handle any pending timer */
314 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
315 //FIXME: make the timers being real callbacks
316 // (i.e. provide dispatchers that read and expand the args)
317 timer = xbt_heap_pop(simix_timers);
319 ((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 /* Clean processes to destroy */
332 SIMIX_process_empty_trash();
334 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
336 if (xbt_swag_size(simix_global->process_list) != 0) {
342 XBT_WARN("Oops ! Deadlock or code not perfectly clean.");
343 SIMIX_display_process_status();
349 * \brief Set the date to execute a function
351 * Set the date to execute the function on the surf.
352 * \param date Date to execute function
353 * \param function Function to be executed
354 * \param arg Parameters of the function
357 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
359 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
362 timer->func = function;
364 xbt_heap_push(simix_timers, timer, date);
368 * \brief Registers a function to create a process.
370 * This function registers a function to be called
371 * when a new process is created. The function has
372 * to call SIMIX_process_create().
373 * \param function create process function
375 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
378 simix_global->create_process_function = function;
382 * \brief Registers a function to kill a process.
384 * This function registers a function to be called when a
385 * process is killed. The function has to call the SIMIX_process_kill().
387 * \param function Kill process function
389 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
392 simix_global->kill_process_function = function;
396 * \brief Registers a function to cleanup a process.
398 * This function registers a user function to be called when
399 * a process ends properly.
401 * \param function cleanup process function
403 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
406 simix_global->cleanup_process_function = function;
410 void SIMIX_display_process_status(void)
412 if (simix_global->process_list == NULL) {
416 smx_process_t process = NULL;
417 int nbprocess = xbt_swag_size(simix_global->process_list);
419 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
420 /* List the process and their state */
422 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
423 xbt_swag_foreach(process, simix_global->process_list) {
425 if (process->waiting_action) {
427 const char* action_description = "unknown";
428 switch (process->waiting_action->type) {
430 case SIMIX_ACTION_EXECUTE:
431 action_description = "execution";
434 case SIMIX_ACTION_PARALLEL_EXECUTE:
435 action_description = "parallel execution";
438 case SIMIX_ACTION_COMMUNICATE:
439 action_description = "communication";
442 case SIMIX_ACTION_SLEEP:
443 action_description = "sleeping";
446 case SIMIX_ACTION_SYNCHRO:
447 action_description = "synchronization";
450 case SIMIX_ACTION_IO:
451 action_description = "I/O";
454 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
455 process->pid, process->name, process->smx_host->name,
456 action_description, process->waiting_action,
457 process->waiting_action->name, (int)process->waiting_action->state);
460 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, process->smx_host->name);
465 static void* SIMIX_action_mallocator_new_f(void) {
466 smx_action_t action = xbt_new(s_smx_action_t, 1);
467 action->simcalls = xbt_fifo_new();
471 static void SIMIX_action_mallocator_free_f(void* action) {
472 xbt_fifo_free(((smx_action_t) action)->simcalls);
476 static void SIMIX_action_mallocator_reset_f(void* action) {
478 // we also recycle the simcall list
479 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
480 xbt_fifo_reset(fifo);
481 memset(action, 0, sizeof(s_smx_action_t));
482 ((smx_action_t) action)->simcalls = fifo;
485 xbt_dict_t SIMIX_asr_get_properties(const char *name)
487 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);