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 #ifdef TIME_BENCH_AMDAHL
63 simix_global->timer_seq = xbt_os_timer_new();
64 simix_global->timer_par = xbt_os_timer_new();
65 xbt_os_timer_start(simix_global->timer_seq);
67 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_process_t), NULL);
68 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_process_t), NULL);
69 simix_global->process_list =
70 xbt_swag_new(xbt_swag_offset(proc, process_hookup));
71 simix_global->process_to_destroy =
72 xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
74 simix_global->maestro_process = NULL;
75 simix_global->registered_functions = xbt_dict_new_homogeneous(NULL);
77 simix_global->create_process_function = SIMIX_process_create;
78 simix_global->kill_process_function = SIMIX_process_kill;
79 simix_global->cleanup_process_function = SIMIX_process_cleanup;
80 simix_global->action_mallocator = xbt_mallocator_new(65536,
81 SIMIX_action_mallocator_new_f, SIMIX_action_mallocator_free_f,
82 SIMIX_action_mallocator_reset_f);
83 simix_global->autorestart = SIMIX_host_restart_processes;
85 surf_init(argc, argv); /* Initialize SURF structures */
86 SIMIX_context_mod_init();
87 SIMIX_create_maestro_process();
89 /* context exception handlers */
90 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
91 __xbt_ex_terminate = SIMIX_process_exception_terminate;
93 /* Initialize the SIMIX network module */
96 /* Prepare to display some more info when dying on Ctrl-C pressing */
97 signal(SIGINT, inthandler);
100 simix_timers = xbt_heap_new(8, &free);
103 XBT_DEBUG("ADD SIMIX LEVELS");
104 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
109 * \brief Clean the SIMIX simulation
111 * This functions remove the memory used by SIMIX
113 void SIMIX_clean(void)
115 #ifdef TIME_BENCH_PER_SR
116 smx_ctx_raw_new_sr();
119 /* Kill everyone (except maestro) */
120 SIMIX_process_killall(simix_global->maestro_process);
122 /* Exit the SIMIX network module */
123 SIMIX_network_exit();
125 xbt_heap_free(simix_timers);
127 /* Free the remaining data structures */
128 xbt_dynar_free(&simix_global->process_to_run);
129 xbt_dynar_free(&simix_global->process_that_ran);
130 xbt_swag_free(simix_global->process_to_destroy);
131 xbt_swag_free(simix_global->process_list);
132 simix_global->process_list = NULL;
133 simix_global->process_to_destroy = NULL;
134 xbt_dict_free(&(simix_global->registered_functions));
136 /* Let's free maestro now */
137 SIMIX_context_free(simix_global->maestro_process->context);
138 xbt_free(simix_global->maestro_process->running_ctx);
139 xbt_free(simix_global->maestro_process);
140 simix_global->maestro_process = NULL;
142 /* Restore the default exception setup */
143 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
144 __xbt_ex_terminate = &__xbt_ex_terminate_default;
146 /* Finish context module and SURF */
147 SIMIX_context_mod_exit();
151 #ifdef TIME_BENCH_AMDAHL
152 xbt_os_timer_stop(simix_global->timer_seq);
153 XBT_INFO("Amdhal timing informations. Sequential time: %lf; Parallel time: %lf",
154 xbt_os_timer_elapsed(simix_global->timer_seq),
155 xbt_os_timer_elapsed(simix_global->timer_par));
156 xbt_os_timer_free(simix_global->timer_seq);
157 xbt_os_timer_free(simix_global->timer_par);
160 xbt_mallocator_free(simix_global->action_mallocator);
161 xbt_free(simix_global);
170 * \brief A clock (in second).
172 * \return Return the clock.
174 XBT_INLINE double SIMIX_get_clock(void)
177 return MC_process_clock_get(SIMIX_process_self());
179 return surf_get_clock();
183 static int process_syscall_color(void *p)
185 switch ((*(smx_process_t *)p)->simcall.call) {
187 case SIMCALL_PROCESS_KILL:
189 case SIMCALL_PROCESS_RESUME:
198 * \brief Run the main simulation loop.
203 smx_process_t process;
205 surf_action_t action;
211 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
212 xbt_dynar_length(simix_global->process_to_run));
213 #ifdef TIME_BENCH_PER_SR
214 smx_ctx_raw_new_sr();
216 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
217 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
218 xbt_dynar_length(simix_global->process_to_run));
220 /* Run all processes that are ready to run, possibly in parallel */
221 #ifdef TIME_BENCH_AMDAHL
222 xbt_os_timer_stop(simix_global->timer_seq);
223 xbt_os_timer_resume(simix_global->timer_par);
225 SIMIX_process_runall();
226 #ifdef TIME_BENCH_AMDAHL
227 xbt_os_timer_stop(simix_global->timer_par);
228 xbt_os_timer_resume(simix_global->timer_seq);
231 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
232 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
234 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
236 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
238 /* Here, the order is ok because:
240 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
242 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
244 * - if there is no kill during the simulation, processes remain sorted according by their PID.
245 * rational: This can be proved inductively.
246 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
247 * Let's show that it is still so at the end of this round.
248 * - if a process is added when being created, that's from maestro. It can be either at startup
249 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
250 * in arbitrary order (inductive hypothesis), we are fine.
251 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
252 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
253 * of a simcall terminates, we're still good. Proof:
254 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
255 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
256 * - For communications, this function is called from SIMIX_comm_finish().
257 * This function itself don't mess with the order since simcalls are handled in FIFO order.
258 * The function is called:
259 * - before the comm starts (invalid parameters, or resource already dead or whatever).
260 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
261 * - because the communication failed or were canceled after startup. In this case, it's called from the function
262 * we are in, by the chunk:
263 * set = model->states.failed_action_set;
264 * while ((action = xbt_swag_extract(set)))
265 * SIMIX_simcall_post((smx_action_t) action->data);
266 * This order is also fixed because it depends of the order in which the surf actions were
267 * added to the system, and only maestro can add stuff this way, through simcalls.
268 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
269 * poped out of the swag does not depend on the user code's execution order.
270 * - because the communication terminated. In this case, actions are served in the order given by
271 * set = model->states.done_action_set;
272 * while ((action = xbt_swag_extract(set)))
273 * SIMIX_simcall_post((smx_action_t) action->data);
274 * and the argument is very similar to the previous one.
275 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
276 * 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.
277 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
278 * - If there is some process killings, the order is changed by this decision that comes from user-land
279 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
280 * So, even the order change induced by the process killing is perfectly reproducible.
282 * So science works, bitches [http://xkcd.com/54/].
284 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
285 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
286 * That would thus be a pure waste of time.
289 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
290 if (process->simcall.call != SIMCALL_NONE) {
291 SIMIX_simcall_pre(&process->simcall, 0);
296 time = SIMIX_timer_next();
297 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
298 time = surf_solve(time);
300 /* Notify all the hosts that have failed */
301 /* FIXME: iterate through the list of failed host and mark each of them */
302 /* as failed. On each host, signal all the running processes with host_fail */
304 /* Handle any pending timer */
305 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
306 //FIXME: make the timers being real callbacks
307 // (i.e. provide dispatchers that read and expand the args)
308 timer = xbt_heap_pop(simix_timers);
310 ((void (*)(void*))timer->func)(timer->args);
312 /* Wake up all processes waiting for a Surf action to finish */
313 xbt_dynar_foreach(model_list, iter, model) {
314 set = model->states.failed_action_set;
315 while ((action = xbt_swag_extract(set)))
316 SIMIX_simcall_post((smx_action_t) action->data);
317 set = model->states.done_action_set;
318 while ((action = xbt_swag_extract(set)))
319 SIMIX_simcall_post((smx_action_t) action->data);
322 /* Clean processes to destroy */
323 SIMIX_process_empty_trash();
325 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
327 if (xbt_swag_size(simix_global->process_list) != 0) {
333 XBT_WARN("Oops ! Deadlock or code not perfectly clean.");
334 SIMIX_display_process_status();
340 * \brief Set the date to execute a function
342 * Set the date to execute the function on the surf.
343 * \param date Date to execute function
344 * \param function Function to be executed
345 * \param arg Parameters of the function
348 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
350 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
353 timer->func = function;
355 xbt_heap_push(simix_timers, timer, date);
359 * \brief Registers a function to create a process.
361 * This function registers a function to be called
362 * when a new process is created. The function has
363 * to call SIMIX_process_create().
364 * \param function create process function
366 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
369 simix_global->create_process_function = function;
373 * \brief Registers a function to kill a process.
375 * This function registers a function to be called when a
376 * process is killed. The function has to call the SIMIX_process_kill().
378 * \param function Kill process function
380 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t_smxprocess_t
383 simix_global->kill_process_function = function;
387 * \brief Registers a function to cleanup a process.
389 * This function registers a user function to be called when
390 * a process ends properly.
392 * \param function cleanup process function
394 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
397 simix_global->cleanup_process_function = function;
401 void SIMIX_display_process_status(void)
403 if (simix_global->process_list == NULL) {
407 smx_process_t process = NULL;
408 int nbprocess = xbt_swag_size(simix_global->process_list);
410 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
411 /* List the process and their state */
413 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
414 xbt_swag_foreach(process, simix_global->process_list) {
416 if (process->waiting_action) {
418 const char* action_description = "unknown";
419 switch (process->waiting_action->type) {
421 case SIMIX_ACTION_EXECUTE:
422 action_description = "execution";
425 case SIMIX_ACTION_PARALLEL_EXECUTE:
426 action_description = "parallel execution";
429 case SIMIX_ACTION_COMMUNICATE:
430 action_description = "communication";
433 case SIMIX_ACTION_SLEEP:
434 action_description = "sleeping";
437 case SIMIX_ACTION_SYNCHRO:
438 action_description = "synchronization";
441 case SIMIX_ACTION_IO:
442 action_description = "I/O";
445 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
446 process->pid, process->name, process->smx_host->name,
447 action_description, process->waiting_action,
448 process->waiting_action->name, (int)process->waiting_action->state);
451 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, process->smx_host->name);
456 static void* SIMIX_action_mallocator_new_f(void) {
457 smx_action_t action = xbt_new(s_smx_action_t, 1);
458 action->simcalls = xbt_fifo_new();
462 static void SIMIX_action_mallocator_free_f(void* action) {
463 xbt_fifo_free(((smx_action_t) action)->simcalls);
467 static void SIMIX_action_mallocator_reset_f(void* action) {
469 // we also recycle the simcall list
470 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
471 xbt_fifo_reset(fifo);
472 memset(action, 0, sizeof(s_smx_action_t));
473 ((smx_action_t) action)->simcalls = fifo;