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 _sg_do_verbose_exit = 1;
31 static void _XBT_CALL inthandler(int ignored)
33 if ( _sg_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;
97 /* Prepare to display some more info when dying on Ctrl-C pressing */
98 signal(SIGINT, inthandler);
100 /* register a function to be called by SURF after the environment creation */
102 sg_platf_postparse_add_cb(SIMIX_post_create_environment);
106 simix_timers = xbt_heap_new(8, &free);
109 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
116 * \brief Clean the SIMIX simulation
118 * This functions remove the memory used by SIMIX
120 static void SIMIX_clean(void)
122 #ifdef TIME_BENCH_PER_SR
123 smx_ctx_raw_new_sr();
126 /* Kill everyone (except maestro) */
127 SIMIX_process_killall(simix_global->maestro_process, 1);
129 /* Exit the SIMIX network module */
130 SIMIX_network_exit();
132 xbt_heap_free(simix_timers);
134 /* Free the remaining data structures */
135 xbt_dynar_free(&simix_global->process_to_run);
136 xbt_dynar_free(&simix_global->process_that_ran);
137 xbt_swag_free(simix_global->process_to_destroy);
138 xbt_swag_free(simix_global->process_list);
139 simix_global->process_list = NULL;
140 simix_global->process_to_destroy = NULL;
141 xbt_dict_free(&(simix_global->registered_functions));
143 /* Let's free maestro now */
144 SIMIX_context_free(simix_global->maestro_process->context);
145 xbt_free(simix_global->maestro_process->running_ctx);
146 xbt_free(simix_global->maestro_process);
147 simix_global->maestro_process = NULL;
149 /* Restore the default exception setup */
150 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
151 __xbt_ex_terminate = &__xbt_ex_terminate_default;
153 /* Finish context module and SURF */
154 SIMIX_context_mod_exit();
158 #ifdef TIME_BENCH_AMDAHL
159 xbt_os_timer_stop(simix_global->timer_seq);
160 XBT_INFO("Amdhal timing informations. Sequential time: %lf; Parallel time: %lf",
161 xbt_os_timer_elapsed(simix_global->timer_seq),
162 xbt_os_timer_elapsed(simix_global->timer_par));
163 xbt_os_timer_free(simix_global->timer_seq);
164 xbt_os_timer_free(simix_global->timer_par);
167 xbt_mallocator_free(simix_global->action_mallocator);
168 xbt_free(simix_global);
177 * \brief A clock (in second).
179 * \return Return the clock.
181 XBT_INLINE double SIMIX_get_clock(void)
184 return MC_process_clock_get(SIMIX_process_self());
186 return surf_get_clock();
190 static int process_syscall_color(void *p)
192 switch ((*(smx_process_t *)p)->simcall.call) {
194 case SIMCALL_PROCESS_KILL:
196 case SIMCALL_PROCESS_RESUME:
205 * \brief Run the main simulation loop.
210 smx_process_t process;
212 surf_action_t action;
218 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
219 xbt_dynar_length(simix_global->process_to_run));
220 #ifdef TIME_BENCH_PER_SR
221 smx_ctx_raw_new_sr();
223 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
224 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
225 xbt_dynar_length(simix_global->process_to_run));
227 /* Run all processes that are ready to run, possibly in parallel */
228 #ifdef TIME_BENCH_AMDAHL
229 xbt_os_timer_stop(simix_global->timer_seq);
230 xbt_os_timer_resume(simix_global->timer_par);
232 SIMIX_process_runall();
233 #ifdef TIME_BENCH_AMDAHL
234 xbt_os_timer_stop(simix_global->timer_par);
235 xbt_os_timer_resume(simix_global->timer_seq);
238 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
239 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
241 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
243 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
245 /* Here, the order is ok because:
247 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
249 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
251 * - if there is no kill during the simulation, processes remain sorted according by their PID.
252 * rational: This can be proved inductively.
253 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
254 * Let's show that it is still so at the end of this round.
255 * - if a process is added when being created, that's from maestro. It can be either at startup
256 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
257 * in arbitrary order (inductive hypothesis), we are fine.
258 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
259 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
260 * of a simcall terminates, we're still good. Proof:
261 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
262 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
263 * - For communications, this function is called from SIMIX_comm_finish().
264 * This function itself don't mess with the order since simcalls are handled in FIFO order.
265 * The function is called:
266 * - before the comm starts (invalid parameters, or resource already dead or whatever).
267 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
268 * - because the communication failed or were canceled after startup. In this case, it's called from the function
269 * we are in, by the chunk:
270 * set = model->states.failed_action_set;
271 * while ((action = xbt_swag_extract(set)))
272 * SIMIX_simcall_post((smx_action_t) action->data);
273 * This order is also fixed because it depends of the order in which the surf actions were
274 * added to the system, and only maestro can add stuff this way, through simcalls.
275 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
276 * poped out of the swag does not depend on the user code's execution order.
277 * - because the communication terminated. In this case, actions are served in the order given by
278 * set = model->states.done_action_set;
279 * while ((action = xbt_swag_extract(set)))
280 * SIMIX_simcall_post((smx_action_t) action->data);
281 * and the argument is very similar to the previous one.
282 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
283 * 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.
284 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
285 * - If there is some process killings, the order is changed by this decision that comes from user-land
286 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
287 * So, even the order change induced by the process killing is perfectly reproducible.
289 * So science works, bitches [http://xkcd.com/54/].
291 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
292 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
293 * That would thus be a pure waste of time.
296 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
297 if (process->simcall.call != SIMCALL_NONE) {
298 SIMIX_simcall_pre(&process->simcall, 0);
303 time = SIMIX_timer_next();
304 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
305 time = surf_solve(time);
307 /* Notify all the hosts that have failed */
308 /* FIXME: iterate through the list of failed host and mark each of them */
309 /* as failed. On each host, signal all the running processes with host_fail */
311 /* Handle any pending timer */
312 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
313 //FIXME: make the timers being real callbacks
314 // (i.e. provide dispatchers that read and expand the args)
315 timer = xbt_heap_pop(simix_timers);
317 ((void (*)(void*))timer->func)(timer->args);
319 /* Wake up all processes waiting for a Surf action to finish */
320 xbt_dynar_foreach(model_list, iter, model) {
321 set = model->states.failed_action_set;
322 while ((action = xbt_swag_extract(set)))
323 SIMIX_simcall_post((smx_action_t) action->data);
324 set = model->states.done_action_set;
325 while ((action = xbt_swag_extract(set)))
326 SIMIX_simcall_post((smx_action_t) action->data);
329 /* Clean processes to destroy */
330 SIMIX_process_empty_trash();
332 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
334 if (xbt_swag_size(simix_global->process_list) != 0) {
340 XBT_WARN("Oops ! Deadlock or code not perfectly clean.");
341 SIMIX_display_process_status();
347 * \brief Set the date to execute a function
349 * Set the date to execute the function on the surf.
350 * \param date Date to execute function
351 * \param function Function to be executed
352 * \param arg Parameters of the function
355 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
357 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
360 timer->func = function;
362 xbt_heap_push(simix_timers, timer, date);
366 * \brief Registers a function to create a process.
368 * This function registers a function to be called
369 * when a new process is created. The function has
370 * to call SIMIX_process_create().
371 * \param function create process function
373 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
376 simix_global->create_process_function = function;
380 * \brief Registers a function to kill a process.
382 * This function registers a function to be called when a
383 * process is killed. The function has to call the SIMIX_process_kill().
385 * \param function Kill process function
387 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
390 simix_global->kill_process_function = function;
394 * \brief Registers a function to cleanup a process.
396 * This function registers a user function to be called when
397 * a process ends properly.
399 * \param function cleanup process function
401 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
404 simix_global->cleanup_process_function = function;
408 void SIMIX_display_process_status(void)
410 if (simix_global->process_list == NULL) {
414 smx_process_t process = NULL;
415 int nbprocess = xbt_swag_size(simix_global->process_list);
417 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
418 /* List the process and their state */
420 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
421 xbt_swag_foreach(process, simix_global->process_list) {
423 if (process->waiting_action) {
425 const char* action_description = "unknown";
426 switch (process->waiting_action->type) {
428 case SIMIX_ACTION_EXECUTE:
429 action_description = "execution";
432 case SIMIX_ACTION_PARALLEL_EXECUTE:
433 action_description = "parallel execution";
436 case SIMIX_ACTION_COMMUNICATE:
437 action_description = "communication";
440 case SIMIX_ACTION_SLEEP:
441 action_description = "sleeping";
444 case SIMIX_ACTION_SYNCHRO:
445 action_description = "synchronization";
448 case SIMIX_ACTION_IO:
449 action_description = "I/O";
451 /* **************************************/
452 /* TUTORIAL: New API */
453 case SIMIX_ACTION_NEW_API:
454 action_description = "NEW API";
455 /* **************************************/
459 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
460 process->pid, process->name, sg_host_name(process->smx_host),
461 action_description, process->waiting_action,
462 process->waiting_action->name, (int)process->waiting_action->state);
465 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, sg_host_name(process->smx_host));
470 static void* SIMIX_action_mallocator_new_f(void) {
471 smx_action_t action = xbt_new(s_smx_action_t, 1);
472 action->simcalls = xbt_fifo_new();
476 static void SIMIX_action_mallocator_free_f(void* action) {
477 xbt_fifo_free(((smx_action_t) action)->simcalls);
481 static void SIMIX_action_mallocator_reset_f(void* action) {
483 // we also recycle the simcall list
484 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
485 xbt_fifo_reset(fifo);
486 memset(action, 0, sizeof(s_smx_action_t));
487 ((smx_action_t) action)->simcalls = fifo;
490 xbt_dict_t SIMIX_pre_asr_get_properties(smx_simcall_t simcall, const char *name){
491 return SIMIX_asr_get_properties(name);
493 xbt_dict_t SIMIX_asr_get_properties(const char *name)
495 return xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL);