1 /* Copyright (c) 2007, 2008, 2009, 2010. 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 */
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 extern void smx_ctx_raw_new_sr(void);
28 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
31 int _surf_do_verbose_exit = 1;
32 static void _XBT_CALL inthandler(int ignored)
34 if ( _surf_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 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
66 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
67 simix_global->process_list =
68 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
69 simix_global->process_to_destroy =
70 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
72 simix_global->maestro_process = NULL;
73 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
75 simix_global->create_process_function = SIMIX_process_create;
76 simix_global->kill_process_function = SIMIX_process_kill;
77 simix_global->cleanup_process_function = SIMIX_process_cleanup;
78 simix_global->action_mallocator = xbt_mallocator_new(65536,
79 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
80 SIMIX_action_mallocator_reset_f);
81 simix_global->autorestart = SIMIX_host_restart_processes;
83 surf_init(argc, argv); /* Initialize SURF structures */
84 SIMIX_context_mod_init();
85 SIMIX_create_maestro_process();
87 /* context exception handlers */
88 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
89 __xbt_ex_terminate = SIMIX_process_exception_terminate;
91 /* Initialize the SIMIX network module */
94 /* Prepare to display some more info when dying on Ctrl-C pressing */
95 signal(SIGINT, inthandler);
98 simix_timers = xbt_heap_new(8, &free);
101 XBT_DEBUG("ADD SIMIX LEVELS");
102 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
107 * \brief Clean the SIMIX simulation
109 * This functions remove the memory used by SIMIX
111 void SIMIX_clean(void)
114 smx_ctx_raw_new_sr();
117 /* Kill everyone (except maestro) */
118 SIMIX_process_killall(simix_global->maestro_process);
120 /* Exit the SIMIX network module */
121 SIMIX_network_exit();
123 xbt_heap_free(simix_timers);
125 /* Free the remaining data structures */
126 xbt_dynar_free(&simix_global->process_to_run);
127 xbt_dynar_free(&simix_global->process_that_ran);
128 xbt_swag_free(simix_global->process_to_destroy);
129 xbt_swag_free(simix_global->process_list);
130 simix_global->process_list = NULL;
131 simix_global->process_to_destroy = NULL;
132 xbt_dict_free(&(simix_global->registered_functions));
134 /* Let's free maestro now */
135 SIMIX_context_free(simix_global->maestro_process->context);
136 xbt_free(simix_global->maestro_process->running_ctx);
137 xbt_free(simix_global->maestro_process);
138 simix_global->maestro_process = NULL;
140 /* Restore the default exception setup */
141 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
142 __xbt_ex_terminate = &__xbt_ex_terminate_default;
144 /* Finish context module and SURF */
145 SIMIX_context_mod_exit();
149 xbt_mallocator_free(simix_global->action_mallocator);
150 xbt_free(simix_global);
159 * \brief A clock (in second).
161 * \return Return the clock.
163 XBT_INLINE double SIMIX_get_clock(void)
166 return MC_process_clock_get(SIMIX_process_self());
168 return surf_get_clock();
172 static int process_syscall_color(void *p)
174 switch ((*(smx_process_t *)p)->simcall.call) {
176 case SIMCALL_PROCESS_KILL:
178 case SIMCALL_PROCESS_RESUME:
187 * \brief Run the main simulation loop.
192 smx_process_t process;
194 surf_action_t action;
200 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
201 xbt_dynar_length(simix_global->process_to_run));
203 smx_ctx_raw_new_sr();
205 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
206 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
207 xbt_dynar_length(simix_global->process_to_run));
209 /* Run all processes that are ready to run, possibly in parallel */
210 SIMIX_process_runall();
212 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
213 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
215 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
217 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
219 /* Here, the order is ok because:
221 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
223 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
225 * - if there is no kill during the simulation, processes remain sorted according by their PID.
226 * rational: This can be proved inductively.
227 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
228 * Let's show that it is still so at the end of this round.
229 * - if a process is added when being created, that's from maestro. It can be either at startup
230 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
231 * in arbitrary order (inductive hypothesis), we are fine.
232 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
233 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
234 * of a simcall terminates, we're still good. Proof:
235 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
236 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
237 * - For communications, this function is called from SIMIX_comm_finish().
238 * This function itself don't mess with the order since simcalls are handled in FIFO order.
239 * The function is called:
240 * - before the comm starts (invalid parameters, or resource already dead or whatever).
241 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
242 * - because the communication failed or were canceled after startup. In this case, it's called from the function
243 * we are in, by the chunk:
244 * set = model->states.failed_action_set;
245 * while ((action = xbt_swag_extract(set)))
246 * SIMIX_simcall_post((smx_action_t) action->data);
247 * This order is also fixed because it depends of the order in which the surf actions were
248 * added to the system, and only maestro can add stuff this way, through simcalls.
249 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
250 * poped out of the swag does not depend on the user code's execution order.
251 * - because the communication terminated. In this case, actions are served in the order given by
252 * set = model->states.done_action_set;
253 * while ((action = xbt_swag_extract(set)))
254 * SIMIX_simcall_post((smx_action_t) action->data);
255 * and the argument is very similar to the previous one.
256 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
257 * 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.
258 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
259 * - If there is some process killings, the order is changed by this decision that comes from user-land
260 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
261 * So, even the order change induced by the process killing is perfectly reproducible.
263 * So science works, bitches [http://xkcd.com/54/].
265 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
266 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
267 * That would thus be a pure waste of time.
270 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
271 if (process->simcall.call != SIMCALL_NONE) {
272 SIMIX_simcall_pre(&process->simcall, 0);
277 time = SIMIX_timer_next();
278 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
279 time = surf_solve(time);
281 /* Notify all the hosts that have failed */
282 /* FIXME: iterate through the list of failed host and mark each of them */
283 /* as failed. On each host, signal all the running processes with host_fail */
285 /* Handle any pending timer */
286 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
287 //FIXME: make the timers being real callbacks
288 // (i.e. provide dispatchers that read and expand the args)
289 timer = xbt_heap_pop(simix_timers);
291 ((void (*)(void*))timer->func)(timer->args);
293 /* Wake up all processes waiting for a Surf action to finish */
294 xbt_dynar_foreach(model_list, iter, model) {
295 set = model->states.failed_action_set;
296 while ((action = xbt_swag_extract(set)))
297 SIMIX_simcall_post((smx_action_t) action->data);
298 set = model->states.done_action_set;
299 while ((action = xbt_swag_extract(set)))
300 SIMIX_simcall_post((smx_action_t) action->data);
303 /* Clean processes to destroy */
304 SIMIX_process_empty_trash();
306 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
308 if (xbt_swag_size(simix_global->process_list) != 0) {
314 XBT_WARN("Oops ! Deadlock or code not perfectly clean.");
315 SIMIX_display_process_status();
321 * \brief Set the date to execute a function
323 * Set the date to execute the function on the surf.
324 * \param date Date to execute function
325 * \param function Function to be executed
326 * \param arg Parameters of the function
329 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
331 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
334 timer->func = function;
336 xbt_heap_push(simix_timers, timer, date);
340 * \brief Registers a function to create a process.
342 * This function registers a function to be called
343 * when a new process is created. The function has
344 * to call SIMIX_process_create().
345 * \param function create process function
347 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
350 simix_global->create_process_function = function;
354 * \brief Registers a function to kill a process.
356 * This function registers a function to be called when a
357 * process is killed. The function has to call the SIMIX_process_kill().
359 * \param function Kill process function
361 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
364 simix_global->kill_process_function = function;
368 * \brief Registers a function to cleanup a process.
370 * This function registers a user function to be called when
371 * a process ends properly.
373 * \param function cleanup process function
375 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
378 simix_global->cleanup_process_function = function;
382 void SIMIX_display_process_status(void)
384 if (simix_global->process_list == NULL) {
388 smx_process_t process = NULL;
389 int nbprocess = xbt_swag_size(simix_global->process_list);
391 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
392 /* List the process and their state */
394 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
395 xbt_swag_foreach(process, simix_global->process_list) {
397 if (process->waiting_action) {
399 const char* action_description = "unknown";
400 switch (process->waiting_action->type) {
402 case SIMIX_ACTION_EXECUTE:
403 action_description = "execution";
406 case SIMIX_ACTION_PARALLEL_EXECUTE:
407 action_description = "parallel execution";
410 case SIMIX_ACTION_COMMUNICATE:
411 action_description = "communication";
414 case SIMIX_ACTION_SLEEP:
415 action_description = "sleeping";
418 case SIMIX_ACTION_SYNCHRO:
419 action_description = "synchronization";
422 case SIMIX_ACTION_IO:
423 action_description = "I/O";
426 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
427 process->pid, process->name, process->smx_host->name,
428 action_description, process->waiting_action,
429 process->waiting_action->name, (int)process->waiting_action->state);
432 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, process->smx_host->name);
437 static void* SIMIX_action_mallocator_new_f(void) {
438 smx_action_t action = xbt_new(s_smx_action_t, 1);
439 action->simcalls = xbt_fifo_new();
443 static void SIMIX_action_mallocator_free_f(void* action) {
444 xbt_fifo_free(((smx_action_t) action)->simcalls);
448 static void SIMIX_action_mallocator_reset_f(void* action) {
450 // we also recycle the simcall list
451 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
452 xbt_fifo_reset(fifo);
453 memset(action, 0, sizeof(s_smx_action_t));
454 ((smx_action_t) action)->simcalls = fifo;