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);
82 surf_init(argc, argv); /* Initialize SURF structures */
83 SIMIX_context_mod_init();
84 SIMIX_create_maestro_process();
86 /* context exception handlers */
87 __xbt_running_ctx_fetch = SIMIX_process_get_running_context;
88 __xbt_ex_terminate = SIMIX_process_exception_terminate;
90 /* Initialize the SIMIX network module */
93 /* Prepare to display some more info when dying on Ctrl-C pressing */
94 signal(SIGINT, inthandler);
97 simix_timers = xbt_heap_new(8, &free);
100 XBT_DEBUG("ADD SIMIX LEVELS");
101 SIMIX_HOST_LEVEL = xbt_lib_add_level(host_lib,SIMIX_host_destroy);
106 * \brief Clean the SIMIX simulation
108 * This functions remove the memory used by SIMIX
110 void SIMIX_clean(void)
113 smx_ctx_raw_new_sr();
116 /* Kill everyone (except maestro) */
117 SIMIX_process_killall(simix_global->maestro_process);
119 /* Exit the SIMIX network module */
120 SIMIX_network_exit();
122 xbt_heap_free(simix_timers);
124 /* Free the remaining data structures */
125 xbt_dynar_free(&simix_global->process_to_run);
126 xbt_dynar_free(&simix_global->process_that_ran);
127 xbt_swag_free(simix_global->process_to_destroy);
128 xbt_swag_free(simix_global->process_list);
129 simix_global->process_list = NULL;
130 simix_global->process_to_destroy = NULL;
131 xbt_dict_free(&(simix_global->registered_functions));
133 /* Let's free maestro now */
134 SIMIX_context_free(simix_global->maestro_process->context);
135 xbt_free(simix_global->maestro_process->running_ctx);
136 xbt_free(simix_global->maestro_process);
137 simix_global->maestro_process = NULL;
139 /* Restore the default exception setup */
140 __xbt_running_ctx_fetch = &__xbt_ex_ctx_default;
141 __xbt_ex_terminate = &__xbt_ex_terminate_default;
143 /* Finish context module and SURF */
144 SIMIX_context_mod_exit();
148 xbt_mallocator_free(simix_global->action_mallocator);
149 xbt_free(simix_global);
158 * \brief A clock (in second).
160 * \return Return the clock.
162 XBT_INLINE double SIMIX_get_clock(void)
165 return MC_process_clock_get(SIMIX_process_self());
167 return surf_get_clock();
171 static int process_syscall_color(void *p)
173 switch ((*(smx_process_t *)p)->simcall.call) {
175 case SIMCALL_PROCESS_KILL:
177 case SIMCALL_PROCESS_RESUME:
186 * \brief Run the main simulation loop.
191 smx_process_t process;
193 surf_action_t action;
199 XBT_DEBUG("New Schedule Round; size(queue)=%lu",
200 xbt_dynar_length(simix_global->process_to_run));
202 smx_ctx_raw_new_sr();
204 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
205 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu",
206 xbt_dynar_length(simix_global->process_to_run));
208 /* Run all processes that are ready to run, possibly in parallel */
209 SIMIX_process_runall();
211 /* Move all killing processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
212 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
214 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
216 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
218 /* Here, the order is ok because:
220 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
222 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
224 * - if there is no kill during the simulation, processes remain sorted according by their PID.
225 * rational: This can be proved inductively.
226 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
227 * Let's show that it is still so at the end of this round.
228 * - if a process is added when being created, that's from maestro. It can be either at startup
229 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
230 * in arbitrary order (inductive hypothesis), we are fine.
231 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
232 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
233 * of a simcall terminates, we're still good. Proof:
234 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
235 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
236 * - For communications, this function is called from SIMIX_comm_finish().
237 * This function itself don't mess with the order since simcalls are handled in FIFO order.
238 * The function is called:
239 * - before the comm starts (invalid parameters, or resource already dead or whatever).
240 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
241 * - because the communication failed or were canceled after startup. In this case, it's called from the function
242 * we are in, by the chunk:
243 * set = model->states.failed_action_set;
244 * while ((action = xbt_swag_extract(set)))
245 * SIMIX_simcall_post((smx_action_t) action->data);
246 * This order is also fixed because it depends of the order in which the surf actions were
247 * added to the system, and only maestro can add stuff this way, through simcalls.
248 * We thus use the inductive hypothesis once again to conclude that the order in which actions are
249 * poped out of the swag does not depend on the user code's execution order.
250 * - because the communication terminated. In this case, actions are served in the order given by
251 * set = model->states.done_action_set;
252 * while ((action = xbt_swag_extract(set)))
253 * SIMIX_simcall_post((smx_action_t) action->data);
254 * and the argument is very similar to the previous one.
255 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
256 * 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.
257 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
258 * - If there is some process killings, the order is changed by this decision that comes from user-land
259 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
260 * So, even the order change induced by the process killing is perfectly reproducible.
262 * So science works, bitches [http://xkcd.com/54/].
264 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
265 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
266 * That would thus be a pure waste of time.
269 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
270 if (process->simcall.call != SIMCALL_NONE) {
271 SIMIX_simcall_pre(&process->simcall, 0);
276 time = SIMIX_timer_next();
277 if (time != -1.0 || xbt_swag_size(simix_global->process_list) != 0)
278 time = surf_solve(time);
280 /* Notify all the hosts that have failed */
281 /* FIXME: iterate through the list of failed host and mark each of them */
282 /* as failed. On each host, signal all the running processes with host_fail */
284 /* Handle any pending timer */
285 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
286 //FIXME: make the timers being real callbacks
287 // (i.e. provide dispatchers that read and expand the args)
288 timer = xbt_heap_pop(simix_timers);
290 ((void (*)(void*))timer->func)(timer->args);
292 /* Wake up all processes waiting for a Surf action to finish */
293 xbt_dynar_foreach(model_list, iter, model) {
294 set = model->states.failed_action_set;
295 while ((action = xbt_swag_extract(set)))
296 SIMIX_simcall_post((smx_action_t) action->data);
297 set = model->states.done_action_set;
298 while ((action = xbt_swag_extract(set)))
299 SIMIX_simcall_post((smx_action_t) action->data);
302 /* Clean processes to destroy */
303 SIMIX_process_empty_trash();
305 } while (time != -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
307 if (xbt_swag_size(simix_global->process_list) != 0) {
313 XBT_WARN("Oops ! Deadlock or code not perfectly clean.");
314 SIMIX_display_process_status();
320 * \brief Set the date to execute a function
322 * Set the date to execute the function on the surf.
323 * \param date Date to execute function
324 * \param function Function to be executed
325 * \param arg Parameters of the function
328 XBT_INLINE void SIMIX_timer_set(double date, void *function, void *arg)
330 smx_timer_t timer = xbt_new0(s_smx_timer_t, 1);
333 timer->func = function;
335 xbt_heap_push(simix_timers, timer, date);
339 * \brief Registers a function to create a process.
341 * This function registers a function to be called
342 * when a new process is created. The function has
343 * to call SIMIX_process_create().
344 * \param function create process function
346 XBT_INLINE void SIMIX_function_register_process_create(smx_creation_func_t
349 simix_global->create_process_function = function;
353 * \brief Registers a function to kill a process.
355 * This function registers a function to be called when a
356 * process is killed. The function has to call the SIMIX_process_kill().
358 * \param function Kill process function
360 XBT_INLINE void SIMIX_function_register_process_kill(void_pfn_smxprocess_t
363 simix_global->kill_process_function = function;
367 * \brief Registers a function to cleanup a process.
369 * This function registers a user function to be called when
370 * a process ends properly.
372 * \param function cleanup process function
374 XBT_INLINE void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t
377 simix_global->cleanup_process_function = function;
381 void SIMIX_display_process_status(void)
383 if (simix_global->process_list == NULL) {
387 smx_process_t process = NULL;
388 int nbprocess = xbt_swag_size(simix_global->process_list);
390 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
391 /* List the process and their state */
393 ("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
394 xbt_swag_foreach(process, simix_global->process_list) {
396 if (process->waiting_action) {
398 const char* action_description = "unknown";
399 switch (process->waiting_action->type) {
401 case SIMIX_ACTION_EXECUTE:
402 action_description = "execution";
405 case SIMIX_ACTION_PARALLEL_EXECUTE:
406 action_description = "parallel execution";
409 case SIMIX_ACTION_COMMUNICATE:
410 action_description = "communication";
413 case SIMIX_ACTION_SLEEP:
414 action_description = "sleeping";
417 case SIMIX_ACTION_SYNCHRO:
418 action_description = "synchronization";
421 case SIMIX_ACTION_IO:
422 action_description = "I/O";
425 XBT_INFO("Process %lu (%s@%s): waiting for %s action %p (%s) in state %d to finish",
426 process->pid, process->name, process->smx_host->name,
427 action_description, process->waiting_action,
428 process->waiting_action->name, (int)process->waiting_action->state);
431 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name, process->smx_host->name);
436 static void* SIMIX_action_mallocator_new_f(void) {
437 smx_action_t action = xbt_new(s_smx_action_t, 1);
438 action->simcalls = xbt_fifo_new();
442 static void SIMIX_action_mallocator_free_f(void* action) {
443 xbt_fifo_free(((smx_action_t) action)->simcalls);
447 static void SIMIX_action_mallocator_reset_f(void* action) {
449 // we also recycle the simcall list
450 xbt_fifo_t fifo = ((smx_action_t) action)->simcalls;
451 xbt_fifo_reset(fifo);
452 memset(action, 0, sizeof(s_smx_action_t));
453 ((smx_action_t) action)->simcalls = fifo;