1 /* Copyright (c) 2007-2015. 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. */
10 #include <signal.h> /* Signal handling */
12 #include "src/internal_config.h"
14 #include <xbt/functional.hpp>
16 #include <simgrid/s4u/host.hpp>
18 #include "src/surf/surf_interface.hpp"
19 #include "src/surf/storage_interface.hpp"
20 #include "src/surf/xml/platf.hpp"
21 #include "smx_private.h"
23 #include "xbt/ex.h" /* ex_backtrace_display */
25 #include "src/mc/mc_replay.h"
26 #include "simgrid/sg_config.h"
28 #include "src/kernel/activity/SynchroExec.hpp"
29 #include "src/kernel/activity/SynchroComm.hpp"
30 #include "src/kernel/activity/SynchroSleep.hpp"
31 #include "src/kernel/activity/SynchroIo.hpp"
32 #include "src/kernel/activity/SynchroRaw.hpp"
35 #include "src/mc/mc_private.h"
36 #include "src/mc/mc_protocol.h"
37 #include "src/mc/Client.hpp"
40 #include "src/mc/mc_protocol.h"
43 #include "src/mc/mc_record.h"
46 #include "src/smpi/private.h"
49 XBT_LOG_NEW_CATEGORY(simix, "All SIMIX categories");
50 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_kernel, simix, "Logging specific to SIMIX (kernel)");
52 std::unique_ptr<simgrid::simix::Global> simix_global;
53 static xbt_heap_t simix_timers = nullptr;
55 /** @brief Timer datatype */
56 typedef struct s_smx_timer {
58 simgrid::xbt::Task<void()> callback;
60 s_smx_timer()=default;
61 s_smx_timer(double date, simgrid::xbt::Task<void()> callback) : date(date), callback(std::move(callback)) {}
64 void (*SMPI_switch_data_segment)(int) = nullptr;
66 int _sg_do_verbose_exit = 1;
67 static void inthandler(int ignored)
69 if ( _sg_do_verbose_exit ) {
70 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option 'verbose-exit').");
71 SIMIX_display_process_status();
74 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'verbose-exit' is set to false.");
80 static void segvhandler(int signum, siginfo_t *siginfo, void *context)
82 if (siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) {
84 "Access violation detected.\n"
85 "This can result from a programming error in your code or, although less likely,\n"
86 "from a bug in SimGrid itself. This can also be the sign of a bug in the OS or\n"
87 "in third-party libraries. Failing hardware can sometimes generate such errors\n"
89 "Finally, if nothing of the above applies, this can result from a stack overflow.\n"
90 "Try to increase stack size with --cfg=contexts/stack_size (current size is %d KiB).\n",
91 smx_context_stack_size / 1024);
92 if (XBT_LOG_ISENABLED(simix_kernel, xbt_log_priority_debug)) {
93 fprintf(stderr, "siginfo = {si_signo = %d, si_errno = %d, si_code = %d, si_addr = %p}\n",
94 siginfo->si_signo, siginfo->si_errno, siginfo->si_code, siginfo->si_addr);
96 } else if (siginfo->si_signo == SIGSEGV) {
97 fprintf(stderr, "Segmentation fault.\n");
99 if (smpi_enabled() && !smpi_privatize_global_variables) {
100 #if HAVE_PRIVATIZATION
101 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatize-global-variables:yes.\n");
103 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatize-global-variables:yes (yet).\n");
104 #endif /* HAVE_PRIVATIZATION */
106 #endif /* HAVE_SMPI */
111 char sigsegv_stack[SIGSTKSZ]; /* alternate stack for SIGSEGV handler */
114 * Install signal handler for SIGSEGV. Check that nobody has already installed
115 * its own handler. For example, the Java VM does this.
117 static void install_segvhandler()
119 stack_t stack, old_stack;
120 stack.ss_sp = sigsegv_stack;
121 stack.ss_size = sizeof sigsegv_stack;
124 if (sigaltstack(&stack, &old_stack) == -1) {
125 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
128 if (!(old_stack.ss_flags & SS_DISABLE)) {
129 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zd, flags=%x). Restore it.",
130 old_stack.ss_sp, old_stack.ss_size, old_stack.ss_flags);
131 sigaltstack(&old_stack, nullptr);
134 struct sigaction action, old_action;
135 action.sa_sigaction = &segvhandler;
136 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
137 sigemptyset(&action.sa_mask);
139 if (sigaction(SIGSEGV, &action, &old_action) == -1) {
140 XBT_WARN("Failed to register signal handler for SIGSEGV: %s", strerror(errno));
143 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
144 XBT_DEBUG("A signal handler was already installed for SIGSEGV (%p). Restore it.",
145 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
146 sigaction(SIGSEGV, &old_action, nullptr);
152 /********************************* SIMIX **************************************/
153 double SIMIX_timer_next()
155 return xbt_heap_size(simix_timers) > 0 ? xbt_heap_maxkey(simix_timers) : -1.0;
158 static void kill_process(smx_actor_t process)
160 SIMIX_process_kill(process, nullptr);
163 static std::function<void()> maestro_code;
168 simgrid::xbt::signal<void()> onDeadlock;
170 XBT_PUBLIC(void) set_maestro(std::function<void()> code)
172 maestro_code = std::move(code);
178 void SIMIX_set_maestro(void (*code)(void*), void* data)
180 maestro_code = std::bind(code, data);
185 * \brief Initialize SIMIX internal data.
190 void SIMIX_global_init(int *argc, char **argv)
193 // The communication initialization is done ASAP.
194 // We need to communicate initialization of the different layers to the model-checker.
195 simgrid::mc::Client::initialize();
199 simix_global = std::unique_ptr<simgrid::simix::Global>(new simgrid::simix::Global());
201 simgrid::simix::ActorImpl proc;
202 simix_global->process_to_run = xbt_dynar_new(sizeof(smx_actor_t), nullptr);
203 simix_global->process_that_ran = xbt_dynar_new(sizeof(smx_actor_t), nullptr);
204 simix_global->process_list = xbt_swag_new(xbt_swag_offset(proc, process_hookup));
205 simix_global->process_to_destroy = xbt_swag_new(xbt_swag_offset(proc, destroy_hookup));
206 simix_global->maestro_process = nullptr;
207 simix_global->create_process_function = &SIMIX_process_create;
208 simix_global->kill_process_function = &kill_process;
209 simix_global->cleanup_process_function = &SIMIX_process_cleanup;
210 simix_global->mutex = xbt_os_mutex_init();
212 surf_init(argc, argv); /* Initialize SURF structures */
213 SIMIX_context_mod_init();
215 // Either create a new context with maestro or create
216 // a context object with the current context mestro):
217 simgrid::simix::create_maestro(maestro_code);
219 /* Prepare to display some more info when dying on Ctrl-C pressing */
220 signal(SIGINT, inthandler);
223 install_segvhandler();
225 /* register a function to be called by SURF after the environment creation */
227 simgrid::surf::on_postparse.connect(SIMIX_post_create_environment);
228 simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) {
229 SIMIX_host_create(&host);
232 simgrid::surf::storageCreatedCallbacks.connect([](simgrid::surf::Storage* storage) {
233 const char* name = storage->getName();
234 // TODO, create sg_storage_by_name
235 sg_storage_t s = xbt_lib_get_elm_or_null(storage_lib, name);
236 xbt_assert(s != nullptr, "Storage not found for name %s", name);
238 SIMIX_storage_create(name, s, nullptr);
241 SIMIX_STORAGE_LEVEL = xbt_lib_add_level(storage_lib, SIMIX_storage_destroy);
244 simix_timers = xbt_heap_new(8, [](void* p) {
245 delete static_cast<smx_timer_t>(p);
248 if (xbt_cfg_get_boolean("clean-atexit"))
251 if (_sg_cfg_exit_asap)
258 * \brief Clean the SIMIX simulation
260 * This functions remove the memory used by SIMIX
265 return; // to avoid double cleaning by java and C
268 if (SIMIX_process_count()>0){
269 if(smpi_process_initialized()){
270 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
272 XBT_WARN("Process called exit when leaving - Skipping cleanups");
279 XBT_DEBUG("SIMIX_clean called. Simulation's over.");
280 if (!xbt_dynar_is_empty(simix_global->process_to_run) && SIMIX_get_clock() <= 0.0) {
282 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
283 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
284 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
286 /* Kill all processes (but maestro) */
287 SIMIX_process_killall(simix_global->maestro_process, 1);
289 /* Exit the SIMIX network module */
290 SIMIX_mailbox_exit();
292 xbt_heap_free(simix_timers);
293 simix_timers = nullptr;
294 /* Free the remaining data structures */
295 xbt_dynar_free(&simix_global->process_to_run);
296 xbt_dynar_free(&simix_global->process_that_ran);
297 xbt_swag_free(simix_global->process_to_destroy);
298 xbt_swag_free(simix_global->process_list);
299 simix_global->process_list = nullptr;
300 simix_global->process_to_destroy = nullptr;
302 xbt_os_mutex_destroy(simix_global->mutex);
303 simix_global->mutex = nullptr;
305 /* Let's free maestro now */
306 delete simix_global->maestro_process->context;
307 simix_global->maestro_process->context = nullptr;
308 delete simix_global->maestro_process;
309 simix_global->maestro_process = nullptr;
311 /* Finish context module and SURF */
312 SIMIX_context_mod_exit();
316 simix_global = nullptr;
323 * \brief A clock (in second).
325 * \return Return the clock.
327 double SIMIX_get_clock()
329 if(MC_is_active() || MC_record_replay_is_active()){
330 return MC_process_clock_get(SIMIX_process_self());
332 return surf_get_clock();
336 static int process_syscall_color(void *p)
338 switch ((*(smx_actor_t *)p)->simcall.call) {
340 case SIMCALL_PROCESS_KILL:
342 case SIMCALL_PROCESS_RESUME:
349 /** Wake up all processes waiting for a Surf action to finish */
350 static void SIMIX_wake_processes()
352 surf_action_t action;
354 for(auto model : *all_existing_models) {
355 XBT_DEBUG("Handling the processes whose action failed (if any)");
356 while ((action = surf_model_extract_failed_action_set(model))) {
357 XBT_DEBUG(" Handling Action %p",action);
358 SIMIX_simcall_exit((smx_activity_t) action->getData());
360 XBT_DEBUG("Handling the processes whose action terminated normally (if any)");
361 while ((action = surf_model_extract_done_action_set(model))) {
362 XBT_DEBUG(" Handling Action %p",action);
363 if (action->getData() == nullptr)
364 XBT_DEBUG("probably vcpu's action %p, skip", action);
366 SIMIX_simcall_exit((smx_activity_t) action->getData());
371 /** Handle any pending timer */
372 static bool SIMIX_execute_timers()
375 while (xbt_heap_size(simix_timers) > 0 && SIMIX_get_clock() >= SIMIX_timer_next()) {
377 //FIXME: make the timers being real callbacks
378 // (i.e. provide dispatchers that read and expand the args)
379 smx_timer_t timer = (smx_timer_t) xbt_heap_pop(simix_timers);
384 xbt_die("Exception throwed ouf of timer callback");
391 /** Execute all the tasks that are queued
393 * e.g. `.then()` callbacks of futures.
395 static bool SIMIX_execute_tasks()
397 xbt_assert(simix_global->tasksTemp.empty());
399 if (simix_global->tasks.empty())
404 // We don't want the callbacks to modify the vector we are iterating over:
405 swap(simix_global->tasks, simix_global->tasksTemp);
407 // Execute all the queued tasks:
408 for (auto& task : simix_global->tasksTemp)
411 simix_global->tasksTemp.clear();
412 } while (!simix_global->tasks.empty());
419 * \brief Run the main simulation loop.
423 if (MC_record_path) {
424 simgrid::mc::replay(MC_record_path);
432 XBT_DEBUG("New Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
434 SIMIX_execute_tasks();
436 while (!xbt_dynar_is_empty(simix_global->process_to_run)) {
437 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%lu", xbt_dynar_length(simix_global->process_to_run));
439 /* Run all processes that are ready to run, possibly in parallel */
440 SIMIX_process_runall();
442 /* Move all killer processes to the end of the list, because killing a process that have an ongoing simcall is a bad idea */
443 xbt_dynar_three_way_partition(simix_global->process_that_ran, process_syscall_color);
445 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
447 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
449 /* Here, the order is ok because:
451 * Short proof: only maestro adds stuff to the process_to_run array, so the execution order of user contexts do not impact its order.
453 * Long proof: processes remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
455 * - if there is no kill during the simulation, processes remain sorted according by their PID.
456 * rational: This can be proved inductively.
457 * Assume that process_to_run is sorted at a beginning of one round (it is at round 0: the deployment file is parsed linearly).
458 * Let's show that it is still so at the end of this round.
459 * - if a process is added when being created, that's from maestro. It can be either at startup
460 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
461 * in arbitrary order (inductive hypothesis), we are fine.
462 * - If a process is added because it's getting killed, its subsequent actions shouldn't matter
463 * - If a process gets added to process_to_run because one of their blocking action constituting the meat
464 * of a simcall terminates, we're still good. Proof:
465 * - You are added from SIMIX_simcall_answer() only. When this function is called depends on the resource
466 * kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications as an example.
467 * - For communications, this function is called from SIMIX_comm_finish().
468 * This function itself don't mess with the order since simcalls are handled in FIFO order.
469 * The function is called:
470 * - before the comm starts (invalid parameters, or resource already dead or whatever).
471 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
472 * - because the communication failed or were canceled after startup. In this case, it's called from the function
473 * we are in, by the chunk:
474 * set = model->states.failed_action_set;
475 * while ((synchro = xbt_swag_extract(set)))
476 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
477 * This order is also fixed because it depends of the order in which the surf actions were
478 * added to the system, and only maestro can add stuff this way, through simcalls.
479 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
480 * poped out of the swag does not depend on the user code's execution order.
481 * - because the communication terminated. In this case, synchros are served in the order given by
482 * set = model->states.done_action_set;
483 * while ((synchro = xbt_swag_extract(set)))
484 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
485 * and the argument is very similar to the previous one.
486 * So, in any case, the orders of calls to SIMIX_comm_finish() do not depend on the order in which user processes are executed.
487 * 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.
488 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
489 * - If there is some process killings, the order is changed by this decision that comes from user-land
490 * But this decision may not have been motivated by a situation that were different because the simulation is not reproducible.
491 * So, even the order change induced by the process killing is perfectly reproducible.
493 * So science works, bitches [http://xkcd.com/54/].
495 * We could sort the process_that_ran array completely so that we can describe the order in which simcalls are handled
496 * (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if unfriendly).
497 * That would thus be a pure waste of time.
501 xbt_dynar_foreach(simix_global->process_that_ran, iter, process) {
502 if (process->simcall.call != SIMCALL_NONE) {
503 SIMIX_simcall_handle(&process->simcall, 0);
507 SIMIX_execute_tasks();
509 SIMIX_wake_processes();
510 } while (SIMIX_execute_tasks());
514 time = SIMIX_timer_next();
515 if (time > -1.0 || xbt_swag_size(simix_global->process_list) != 0) {
516 XBT_DEBUG("Calling surf_solve");
517 time = surf_solve(time);
518 XBT_DEBUG("Moving time ahead : %g", time);
521 /* Notify all the hosts that have failed */
522 /* FIXME: iterate through the list of failed host and mark each of them */
523 /* as failed. On each host, signal all the running processes with host_fail */
525 // Execute timers and tasks until there isn't anything to be done:
528 again = SIMIX_execute_timers();
529 if (SIMIX_execute_tasks())
531 SIMIX_wake_processes();
534 /* Autorestart all process */
535 for (auto host: host_that_restart) {
536 XBT_INFO("Restart processes on host: %s", sg_host_get_name(host));
537 SIMIX_host_autorestart(host);
539 host_that_restart.clear();
541 /* Clean processes to destroy */
542 SIMIX_process_empty_trash();
544 XBT_DEBUG("### time %f, empty %d", time, xbt_dynar_is_empty(simix_global->process_to_run));
546 if (xbt_dynar_is_empty(simix_global->process_to_run) &&
547 xbt_swag_size(simix_global->process_list) != 0)
548 simgrid::simix::onDeadlock();
550 } while (time > -1.0 || !xbt_dynar_is_empty(simix_global->process_to_run));
552 if (xbt_swag_size(simix_global->process_list) != 0) {
556 XBT_CRITICAL("Oops ! Deadlock or code not perfectly clean.");
557 SIMIX_display_process_status();
563 * \brief Set the date to execute a function
565 * Set the date to execute the function on the surf.
566 * \param date Date to execute function
567 * \param function Function to be executed
568 * \param arg Parameters of the function
571 smx_timer_t SIMIX_timer_set(double date, void (*callback)(void*), void *arg)
573 smx_timer_t timer = new s_smx_timer_t(date, [=](){ callback(arg); });
574 xbt_heap_push(simix_timers, timer, date);
578 smx_timer_t SIMIX_timer_set(double date, simgrid::xbt::Task<void()> callback)
580 smx_timer_t timer = new s_smx_timer_t(date, std::move(callback));
581 xbt_heap_push(simix_timers, timer, date);
585 /** @brief cancels a timer that was added earlier */
586 void SIMIX_timer_remove(smx_timer_t timer) {
587 xbt_heap_rm_elm(simix_timers, timer, timer->date);
590 /** @brief Returns the date at which the timer will trigger (or 0 if nullptr timer) */
591 double SIMIX_timer_get_date(smx_timer_t timer) {
592 return timer?timer->date:0;
596 * \brief Registers a function to create a process.
598 * This function registers a function to be called
599 * when a new process is created. The function has
600 * to call SIMIX_process_create().
601 * \param function create process function
603 void SIMIX_function_register_process_create(smx_creation_func_t function)
605 simix_global->create_process_function = function;
609 * \brief Registers a function to kill a process.
611 * This function registers a function to be called when a process is killed. The function has to call the
612 * SIMIX_process_kill().
614 * \param function Kill process function
616 void SIMIX_function_register_process_kill(void_pfn_smxprocess_t function)
618 simix_global->kill_process_function = function;
622 * \brief Registers a function to cleanup a process.
624 * This function registers a user function to be called when a process ends properly.
626 * \param function cleanup process function
628 void SIMIX_function_register_process_cleanup(void_pfn_smxprocess_t function)
630 simix_global->cleanup_process_function = function;
634 void SIMIX_display_process_status()
636 if (simix_global->process_list == nullptr) {
640 smx_actor_t process = nullptr;
641 int nbprocess = xbt_swag_size(simix_global->process_list);
643 XBT_INFO("%d processes are still running, waiting for something.", nbprocess);
644 /* List the process and their state */
645 XBT_INFO("Legend of the following listing: \"Process <pid> (<name>@<host>): <status>\"");
646 xbt_swag_foreach(process, simix_global->process_list) {
648 if (process->waiting_synchro) {
650 const char* synchro_description = "unknown";
652 if (dynamic_cast<simgrid::kernel::activity::Exec*>(process->waiting_synchro) != nullptr)
653 synchro_description = "execution";
655 if (dynamic_cast<simgrid::kernel::activity::Comm*>(process->waiting_synchro) != nullptr)
656 synchro_description = "communication";
658 if (dynamic_cast<simgrid::kernel::activity::Sleep*>(process->waiting_synchro) != nullptr)
659 synchro_description = "sleeping";
661 if (dynamic_cast<simgrid::kernel::activity::Raw*>(process->waiting_synchro) != nullptr)
662 synchro_description = "synchronization";
664 if (dynamic_cast<simgrid::kernel::activity::Io*>(process->waiting_synchro) != nullptr)
665 synchro_description = "I/O";
669 switch (process->waiting_synchro->type) {
670 case SIMIX_SYNC_PARALLEL_EXECUTE:
671 synchro_description = "parallel execution";
674 case SIMIX_SYNC_JOIN:
675 synchro_description = "joining";
679 XBT_INFO("Process %lu (%s@%s): waiting for %s synchro %p (%s) in state %d to finish",
680 process->pid, process->name.c_str(), sg_host_get_name(process->host),
681 synchro_description, process->waiting_synchro,
682 process->waiting_synchro->name.c_str(), (int)process->waiting_synchro->state);
685 XBT_INFO("Process %lu (%s@%s)", process->pid, process->name.c_str(), sg_host_get_name(process->host));
690 xbt_dict_t simcall_HANDLER_asr_get_properties(smx_simcall_t simcall, const char *name){
691 return SIMIX_asr_get_properties(name);
693 xbt_dict_t SIMIX_asr_get_properties(const char *name)
695 return static_cast<xbt_dict_t>(xbt_lib_get_or_null(as_router_lib, name, ROUTING_PROP_ASR_LEVEL));
698 int SIMIX_is_maestro()
700 return simix_global==nullptr /*SimDag*/|| SIMIX_process_self() == simix_global->maestro_process;