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 /* FIXME: Yeah, I'll do it in a portable maner one day [Mt] */
28 int _surf_do_verbose_exit = 1;
29 static void _XBT_CALL inthandler(int ignored)
31 if ( _surf_do_verbose_exit ) {
32 XBT_INFO("CTRL-C pressed. Displaying status and bailing out");
33 SIMIX_display_process_status();
36 XBT_INFO("CTRL-C pressed. bailing out without displaying because verbose-exit is disabled");
41 /********************************* SIMIX **************************************/
43 XBT_INLINE double SIMIX_timer_next(void)
45 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
50 * \brief Initialize SIMIX internal data.
55 void SIMIX_global_init(int *argc, char **argv)
60 simix_global = xbt_new0(s_smx_global_t, 1);
62 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
63 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
64 simix_global->process_list =
65 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
66 simix_global->process_to_destroy =
67 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
69 simix_global->maestro_process = NULL;
70 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
72 simix_global->create_process_function = SIMIX_process_create;
73 simix_global->kill_process_function = SIMIX_process_kill;
74 simix_global->cleanup_process_function = SIMIX_process_cleanup;
75 simix_global->action_mallocator = xbt_mallocator_new(65536,
76 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
77 SIMIX_action_mallocator_reset_f);
78 simix_global->autorestart = SIMIX_host_restart_processes;
80 surf_init(argc, argv); /* Initialize SURF structures */
81 SIMIX_context_mod_init();
82 SIMIX_create_maestro_process();
84 /* context exception handlers */
85 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
86 __xbt_ex_terminate = SIMIX_process_exception_terminate;
88 /* Initialize the SIMIX network module */
91 /* Prepare to display some more info when dying on Ctrl-C pressing */
92 signal(SIGINT, inthandler);
95 simix_timers = xbt_heap_new(8, &free);
98 XBT_DEBUG("ADD SIMIX LEVELS");
99 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
104 * \brief Clean the SIMIX simulation
106 * This functions remove the memory used by SIMIX
108 void SIMIX_clean(void)
111 smx_ctx_raw_new_sr();
114 /* Kill everyone (except maestro) */
115 SIMIX_process_killall(simix_global->maestro_process);
117 /* Exit the SIMIX network module */
118 SIMIX_network_exit();
120 xbt_heap_free(simix_timers);
122 /* Free the remaining data structures */
123 xbt_dynar_free(&simix_global->process_to_run);
124 xbt_dynar_free(&simix_global->process_that_ran);
125 xbt_swag_free(simix_global->process_to_destroy);
126 xbt_swag_free(simix_global->process_list);
127 simix_global->process_list = NULL;
128 simix_global->process_to_destroy = NULL;
129 xbt_dict_free(&(simix_global->registered_functions));
131 /* Let's free maestro now */
132 SIMIX_context_free(simix_global->maestro_process->context);
133 xbt_free(simix_global->maestro_process->running_ctx);
134 xbt_free(simix_global->maestro_process);
135 simix_global->maestro_process = NULL;
137 /* Restore the default exception setup */
138 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
139 __xbt_ex_terminate = &__xbt_ex_terminate_default;
141 /* Finish context module and SURF */
142 SIMIX_context_mod_exit();
146 xbt_mallocator_free(simix_global->action_mallocator);
147 xbt_free(simix_global);
156 * \brief A clock (in second).
158 * \return Return the clock.
160 XBT_INLINE double SIMIX_get_clock(void)
163 return MC_process_clock_get(SIMIX_process_self());
165 return surf_get_clock();
169 static int process_syscall_color(void *p)
171 switch ((*(smx_process_t *)p)->simcall.call) {
173 case SIMCALL_PROCESS_KILL:
175 case SIMCALL_PROCESS_RESUME:
184 * \brief Run the main simulation loop.
189 smx_process_t process;
191 surf_action_t action;
197 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
198 xbt_dynar_length(simix_global->process_to_run));
200 smx_ctx_raw_new_sr();
202 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
203 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
204 xbt_dynar_length(simix_global->process_to_run));
206 /* Run all processes that are ready to run, possibly in parallel */
207 SIMIX_process_runall();
209 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
210 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
212 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
214 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
216 /* Here, the order is ok because:
218 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
220 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
222 * - if there is no kill during the simulation, processes remain sorted according by their PID.
223 * rational: This can be proved inductively.
224 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
225 * Let's show that it is still so at the end of this round.
226 * - if a process is added when being created, that's from maestro. It can be either at startup
227 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
228 * in arbitrary order (inductive hypothesis), we are fine.
229 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
230 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
231 * of a simcall terminates, we're still good. Proof:
232 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
233 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
234 * - For communications, this function is called from SIMIX_comm_finish().
235 * This function itself don't mess with the order since simcalls are handled in FIFO order.
236 * The function is called:
237 * - before the comm starts (invalid parameters, or resource already dead or whatever).
238 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
239 * - because the communication failed or were canceled after startup. In this case, it's called from the function
240 * we are in, by the chunk:
241 * set = model->states.failed_action_set;
242 * while ((action = xbt_swag_extract(set)))
243 * SIMIX_simcall_post((smx_action_t) action->data);
244 * This order is also fixed because it depends of the order in which the surf actions were
245 * added to the system, and only maestro can add stuff this way, through simcalls.
246 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
247 * poped out of the swag does not depend on the user code's execution order.
248 * - because the communication terminated. In this case, actions are served in the order given by
249 * set = model->states.done_action_set;
250 * while ((action = xbt_swag_extract(set)))
251 * SIMIX_simcall_post((smx_action_t) action->data);
252 * and the argument is very similar to the previous one.
253 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
254 * 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.
255 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
256 * - If there is some process killings, the order is changed by this decision that comes from user-land
257 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
258 * So, even the order change induced by the process killing is perfectly reproducible.
260 * So science works, bitches [http://xkcd.com/54/].
262 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
263 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
264 * That would thus be a pure waste of time.
267 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
268 if (process->simcall.call != SIMCALL_NONE) {
269 SIMIX_simcall_pre(&process->simcall, 0);
274 time = SIMIX_timer_next();
275 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
276 time = surf_solve(time);
278 /* Notify all the hosts that have failed */
279 /* FIXME: iterate through the list of failed host and mark each of them */
280 /* as failed. On each host, signal all the running processes with host_fail */
282 /* Handle any pending timer */
283 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
284 //FIXME: make the timers being real callbacks
285 // (i.e. provide dispatchers that read and expand the args)
286 timer = xbt_heap_pop(simix_timers);
288 ((void (*)(void*))timer->func)(timer->args);
290 /* Wake up all processes waiting for a Surf action to finish */
291 xbt_dynar_foreach(model_list, iter, model) {
292 set = model->states.failed_action_set;
293 while ((action = xbt_swag_extract(set)))
294 SIMIX_simcall_post((smx_action_t) action->data);
295 set = model->states.done_action_set;
296 while ((action = xbt_swag_extract(set)))
297 SIMIX_simcall_post((smx_action_t) action->data);
300 /* Clean processes to destroy */
301 SIMIX_process_empty_trash();
303 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
305 if (xbt_swag_size(simix_global->process_list) != 0) {
311 XBT_WARN("Oops ! Deadlock or code not perfectly clean.");
312 SIMIX_display_process_status();
318 * \brief Set the date to execute a function
320 * Set the date to execute the function on the surf.
321 * \param date Date to execute function
322 * \param function Function to be executed
323 * \param arg Parameters of the function
326 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
328 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
331 timer->func = function;
333 xbt_heap_push(simix_timers, timer, date);
337 * \brief Registers a function to create a process.
339 * This function registers a function to be called
340 * when a new process is created. The function has
341 * to call SIMIX_process_create().
342 * \param function create process function
344 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
347 simix_global->create_process_function = function;
351 * \brief Registers a function to kill a process.
353 * This function registers a function to be called when a
354 * process is killed. The function has to call the SIMIX_process_kill().
356 * \param function Kill process function
358 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
361 simix_global->kill_process_function = function;
365 * \brief Registers a function to cleanup a process.
367 * This function registers a user function to be called when
368 * a process ends properly.
370 * \param function cleanup process function
372 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
375 simix_global->cleanup_process_function = function;
379 void SIMIX_display_process_status(void)
381 if (simix_global->process_list == NULL) {
385 smx_process_t process = NULL;
386 int nbprocess = xbt_swag_size(simix_global->process_list);
388 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
389 /* List the process and their state */
391 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
392 xbt_swag_foreach(process, simix_global->process_list) {
394 if (process->waiting_action) {
396 const char* action_description = "unknown";
397 switch (process->waiting_action->type) {
399 case SIMIX_ACTION_EXECUTE:
400 action_description = "execution";
403 case SIMIX_ACTION_PARALLEL_EXECUTE:
404 action_description = "parallel execution";
407 case SIMIX_ACTION_COMMUNICATE:
408 action_description = "communication";
411 case SIMIX_ACTION_SLEEP:
412 action_description = "sleeping";
415 case SIMIX_ACTION_SYNCHRO:
416 action_description = "synchronization";
419 case SIMIX_ACTION_IO:
420 action_description = "I/O";
423 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
424 process->pid, process->name, process->smx_host->name,
425 action_description, process->waiting_action,
426 process->waiting_action->name, (int)process->waiting_action->state);
429 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, process->smx_host->name);
434 static void* SIMIX_action_mallocator_new_f(void) {
435 smx_action_t action = xbt_new(s_smx_action_t, 1);
436 action->simcalls = xbt_fifo_new();
440 static void SIMIX_action_mallocator_free_f(void* action) {
441 xbt_fifo_free(((smx_action_t) action)->simcalls);
445 static void SIMIX_action_mallocator_reset_f(void* action) {
447 // we also recycle the simcall list
448 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
449 xbt_fifo_reset(fifo);
450 memset(action, 0, sizeof(s_smx_action_t));
451 ((smx_action_t) action)->simcalls = fifo;