1 /* Copyright (c) 2016-2021. 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 "src/kernel/EngineImpl.hpp"
8 #include "simgrid/Exception.hpp"
9 #include "simgrid/kernel/Timer.hpp"
10 #include "simgrid/kernel/routing/NetPoint.hpp"
11 #include "simgrid/kernel/routing/NetZoneImpl.hpp"
12 #include "simgrid/s4u/Host.hpp"
13 #include "simgrid/sg_config.hpp"
14 #include "src/include/surf/surf.hpp" //get_clock() and surf_solve()
15 #include "src/kernel/resource/DiskImpl.hpp"
16 #include "src/kernel/resource/profile/Profile.hpp"
17 #include "src/mc/mc_record.hpp"
18 #include "src/mc/mc_replay.hpp"
19 #include "src/simix/smx_private.hpp"
20 #include "src/smpi/include/smpi_actor.hpp"
21 #include "src/surf/network_interface.hpp"
22 #include "src/surf/xml/platf.hpp" // FIXME: KILLME. There must be a better way than mimicking XML here
24 #include <boost/algorithm/string/predicate.hpp>
30 #include "src/mc/remote/AppSide.hpp"
33 XBT_LOG_NEW_DEFAULT_CATEGORY(ker_engine, "Logging specific to Engine (kernel)");
36 EngineImpl* EngineImpl::instance_ = nullptr; /* That singleton is awful too. */
38 config::Flag<double> cfg_breakpoint{"debug/breakpoint",
39 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
40 config::Flag<bool> cfg_verbose_exit{"debug/verbose-exit", "Display the actor status at exit", true};
42 } // namespace simgrid
44 XBT_ATTRIB_NORETURN static void inthandler(int)
46 if (simgrid::kernel::cfg_verbose_exit) {
47 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option "
48 "'debug/verbose-exit').");
49 simgrid::kernel::EngineImpl::get_instance()->display_all_actor_status();
51 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'debug/verbose-exit' is set to false.");
57 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
59 if ((siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) || siginfo->si_signo == SIGBUS) {
61 "Access violation or Bus error detected.\n"
62 "This probably comes from a programming error in your code, or from a stack\n"
63 "overflow. If you are certain of your code, try increasing the stack size\n"
64 " --cfg=contexts/stack-size:XXX (current size is %u KiB).\n"
66 "If it does not help, this may have one of the following causes:\n"
67 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
68 "Failing hardware can sometimes generate such errors too.\n"
70 "If you think you've found a bug in SimGrid, please report it along with a\n"
71 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
72 "of the fault captured with gdb or valgrind.\n",
73 smx_context_stack_size / 1024);
74 } else if (siginfo->si_signo == SIGSEGV) {
75 fprintf(stderr, "Segmentation fault.\n");
77 if (smpi_enabled() && smpi_cfg_privatization() == SmpiPrivStrategies::NONE) {
78 #if HAVE_PRIVATIZATION
79 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
81 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
82 #endif /* HAVE_PRIVATIZATION */
84 #endif /* HAVE_SMPI */
90 * Install signal handler for SIGSEGV. Check that nobody has already installed
91 * its own handler. For example, the Java VM does this.
93 static void install_segvhandler()
97 if (simgrid::kernel::context::Context::install_sigsegv_stack(&old_stack, true) == -1) {
98 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
101 if (not(old_stack.ss_flags & SS_DISABLE)) {
102 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
103 old_stack.ss_size, (unsigned)old_stack.ss_flags);
104 sigaltstack(&old_stack, nullptr);
107 struct sigaction action;
108 struct sigaction old_action;
109 action.sa_sigaction = &segvhandler;
110 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
111 sigemptyset(&action.sa_mask);
113 /* Linux tend to raise only SIGSEGV where other systems also raise SIGBUS on severe error */
114 for (int sig : {SIGSEGV, SIGBUS}) {
115 if (sigaction(sig, &action, &old_action) == -1) {
116 XBT_WARN("Failed to register signal handler for signal %d: %s", sig, strerror(errno));
119 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
120 XBT_DEBUG("A signal handler was already installed for signal %d (%p). Restore it.", sig,
121 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
122 sigaction(sig, &old_action, nullptr);
132 EngineImpl::EngineImpl(int* argc, char** argv)
134 EngineImpl::instance_ = this;
137 EngineImpl::~EngineImpl()
139 /* Since hosts_ is a std::map, the hosts are destroyed in the lexicographic order, which ensures that the output is
142 while (not hosts_.empty())
143 hosts_.begin()->second->destroy();
145 /* Also delete the other data */
146 delete netzone_root_;
147 for (auto const& kv : netpoints_)
150 for (auto const& kv : links_)
152 kv.second->destroy();
154 for (auto const& kv : mailboxes_)
157 /* Free the remaining data structures */
159 xbt_dynar_free(&actors_vector_);
160 xbt_dynar_free(&dead_actors_vector_);
162 /* clear models before freeing handle, network models can use external callback defined in the handle */
163 models_prio_.clear();
166 void EngineImpl::initialize(int* argc, char** argv)
169 // The communication initialization is done ASAP, as we need to get some init parameters from the MC for different
170 // layers. But simix_global needs to be created, as we send the address of some of its fields to the MC that wants to
171 // read them directly.
172 simgrid::mc::AppSide::initialize();
175 surf_init(argc, argv); /* Initialize SURF structures */
177 /* Prepare to display some more info when dying on Ctrl-C pressing */
178 std::signal(SIGINT, inthandler);
181 install_segvhandler();
184 /* register a function to be called by SURF after the environment creation */
186 s4u::Engine::on_platform_created.connect(surf_presolve);
188 if (config::get_value<bool>("debug/clean-atexit"))
191 void EngineImpl::shutdown()
193 static bool already_cleaned_up = false;
194 if (already_cleaned_up)
195 return; // to avoid double cleaning by java and C
196 already_cleaned_up = true;
198 simix_global->destroy_context_factory();
199 return; // Nothing more to shutdown
201 XBT_DEBUG("EngineImpl::shutdown() called. Simulation's over.");
202 if (instance_->has_actors_to_run() && simgrid_get_clock() <= 0.0) {
204 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
205 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
206 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
209 /* Kill all actors (but maestro) */
210 instance_->maestro_->kill_all();
211 instance_->run_all_actors();
212 instance_->empty_trash();
215 if (not instance_->actor_list_.empty()) {
216 if (smpi_process()->initialized()) {
217 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
219 XBT_WARN("Process called exit when leaving - Skipping cleanups");
225 /* Let's free maestro now */
226 instance_->destroy_maestro();
228 /* Finish context module and SURF */
229 simix_global->destroy_context_factory();
231 while (not timer::kernel_timers().empty()) {
232 delete timer::kernel_timers().top().second;
233 timer::kernel_timers().pop();
239 simix_global = nullptr;
244 void EngineImpl::load_platform(const std::string& platf)
246 double start = xbt_os_time();
247 if (boost::algorithm::ends_with(platf, ".so") or boost::algorithm::ends_with(platf, ".dylib")) {
249 xbt_die("loading platform through shared library isn't supported on windows");
251 void* handle = dlopen(platf.c_str(), RTLD_LAZY);
252 xbt_assert(handle, "Impossible to open platform file: %s", platf.c_str());
253 platf_handle_ = std::unique_ptr<void, std::function<int(void*)>>(handle, dlclose);
254 using load_fct_t = void (*)(const simgrid::s4u::Engine&);
255 auto callable = (load_fct_t)dlsym(platf_handle_.get(), "load_platform");
256 const char* dlsym_error = dlerror();
257 xbt_assert(not dlsym_error, "Error: %s", dlsym_error);
258 callable(*simgrid::s4u::Engine::get_instance());
261 parse_platform_file(platf);
264 double end = xbt_os_time();
265 XBT_DEBUG("PARSE TIME: %g", (end - start));
268 void EngineImpl::load_deployment(const std::string& file) const
273 surf_parse_open(file);
278 void EngineImpl::register_function(const std::string& name, const actor::ActorCodeFactory& code)
280 registered_functions[name] = code;
282 void EngineImpl::register_default(const actor::ActorCodeFactory& code)
284 default_function = code;
287 void EngineImpl::add_model(std::shared_ptr<resource::Model> model, const std::vector<resource::Model*>& dependencies)
289 auto model_name = model->get_name();
290 xbt_assert(models_prio_.find(model_name) == models_prio_.end(),
291 "Model %s already exists, use model.set_name() to change its name", model_name.c_str());
293 for (const auto dep : dependencies) {
294 xbt_assert(models_prio_.find(dep->get_name()) != models_prio_.end(),
295 "Model %s doesn't exists. Impossible to use it as dependency.", dep->get_name().c_str());
297 models_.push_back(model.get());
298 models_prio_[model_name] = std::move(model);
301 void EngineImpl::add_split_duplex_link(const std::string& name, std::unique_ptr<resource::SplitDuplexLinkImpl> link)
303 split_duplex_links_[name] = std::move(link);
306 /** Wake up all actors waiting for a Surf action to finish */
307 void EngineImpl::wake_all_waiting_actors() const
309 for (auto const& model : models_) {
310 XBT_DEBUG("Handling the failed actions (if any)");
311 while (auto* action = model->extract_failed_action()) {
312 XBT_DEBUG(" Handling Action %p", action);
313 if (action->get_activity() != nullptr)
314 activity::ActivityImplPtr(action->get_activity())->post();
316 XBT_DEBUG("Handling the terminated actions (if any)");
317 while (auto* action = model->extract_done_action()) {
318 XBT_DEBUG(" Handling Action %p", action);
319 if (action->get_activity() == nullptr)
320 XBT_DEBUG("probably vcpu's action %p, skip", action);
322 activity::ActivityImplPtr(action->get_activity())->post();
327 * @brief Executes the actors in actors_to_run.
329 * The actors in actors_to_run are run (in parallel if possible). On exit, actors_to_run is empty, and actors_that_ran
330 * contains the list of actors that just ran. The two lists are swapped so, be careful when using them before and after
331 * a call to this function.
333 void EngineImpl::run_all_actors()
335 simix_global->get_context_factory()->run_all();
337 actors_to_run_.swap(actors_that_ran_);
338 actors_to_run_.clear();
341 actor::ActorImpl* EngineImpl::get_actor_by_pid(aid_t pid)
343 auto item = actor_list_.find(pid);
344 if (item != actor_list_.end())
348 for (auto& a : actors_to_destroy_)
349 if (a.get_pid() == pid)
351 return nullptr; // Not found, even in the trash
354 /** Execute all the tasks that are queued, e.g. `.then()` callbacks of futures. */
355 bool EngineImpl::execute_tasks()
360 std::vector<xbt::Task<void()>> tasksTemp;
362 // We don't want the callbacks to modify the vector we are iterating over:
363 tasks.swap(tasksTemp);
365 // Execute all the queued tasks:
366 for (auto& task : tasksTemp)
370 } while (not tasks.empty());
375 void EngineImpl::remove_daemon(actor::ActorImpl* actor)
377 auto it = daemons_.find(actor);
378 xbt_assert(it != daemons_.end(), "The dying daemon is not a daemon after all. Please report that bug.");
382 void EngineImpl::add_actor_to_run_list_no_check(actor::ActorImpl* actor)
384 XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
385 actors_to_run_.push_back(actor);
388 void EngineImpl::add_actor_to_run_list(actor::ActorImpl* actor)
390 if (std::find(begin(actors_to_run_), end(actors_to_run_), actor) != end(actors_to_run_)) {
391 XBT_DEBUG("Actor %s is already in the to_run list", actor->get_cname());
393 XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
394 actors_to_run_.push_back(actor);
397 void EngineImpl::empty_trash()
399 while (not actors_to_destroy_.empty()) {
400 actor::ActorImpl* actor = &actors_to_destroy_.front();
401 actors_to_destroy_.pop_front();
402 XBT_DEBUG("Getting rid of %s (refcount: %d)", actor->get_cname(), actor->get_refcount());
403 intrusive_ptr_release(actor);
406 xbt_dynar_reset(dead_actors_vector_);
410 void EngineImpl::display_all_actor_status() const
412 XBT_INFO("%zu actors are still running, waiting for something.", actor_list_.size());
413 /* List the actors and their state */
414 XBT_INFO("Legend of the following listing: \"Actor <pid> (<name>@<host>): <status>\"");
415 for (auto const& kv : actor_list_) {
416 actor::ActorImpl* actor = kv.second;
418 if (actor->waiting_synchro_) {
419 const char* synchro_description = "unknown";
421 if (boost::dynamic_pointer_cast<kernel::activity::ExecImpl>(actor->waiting_synchro_) != nullptr)
422 synchro_description = "execution";
424 if (boost::dynamic_pointer_cast<kernel::activity::CommImpl>(actor->waiting_synchro_) != nullptr)
425 synchro_description = "communication";
427 if (boost::dynamic_pointer_cast<kernel::activity::SleepImpl>(actor->waiting_synchro_) != nullptr)
428 synchro_description = "sleeping";
430 if (boost::dynamic_pointer_cast<kernel::activity::RawImpl>(actor->waiting_synchro_) != nullptr)
431 synchro_description = "synchronization";
433 if (boost::dynamic_pointer_cast<kernel::activity::IoImpl>(actor->waiting_synchro_) != nullptr)
434 synchro_description = "I/O";
436 XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %d to finish", actor->get_pid(),
437 actor->get_cname(), actor->get_host()->get_cname(), synchro_description,
438 (xbt_log_no_loc ? (size_t)0xDEADBEEF : (size_t)actor->waiting_synchro_.get()),
439 actor->waiting_synchro_->get_cname(), (int)actor->waiting_synchro_->state_);
441 XBT_INFO("Actor %ld (%s@%s) simcall %s", actor->get_pid(), actor->get_cname(), actor->get_host()->get_cname(),
442 SIMIX_simcall_name(actor->simcall_));
447 void EngineImpl::run()
449 if (MC_record_replay_is_active()) {
450 mc::replay(MC_record_path());
458 XBT_DEBUG("New Schedule Round; size(queue)=%zu", actors_to_run_.size());
460 if (cfg_breakpoint >= 0.0 && surf_get_clock() >= cfg_breakpoint) {
461 XBT_DEBUG("Breakpoint reached (%g)", cfg_breakpoint.get());
462 cfg_breakpoint = -1.0;
472 while (not actors_to_run_.empty()) {
473 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", actors_to_run_.size());
475 /* Run all actors that are ready to run, possibly in parallel */
478 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
480 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
482 /* Here, the order is ok because:
484 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
485 * not impact its order.
487 * Long proof: actors remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
489 * - if there is no kill during the simulation, actors remain sorted according by their PID.
490 * Rationale: This can be proved inductively.
491 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
492 * is parsed linearly).
493 * Let's show that it is still so at the end of this round.
494 * - if an actor is added when being created, that's from maestro. It can be either at startup
495 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
496 * in arbitrary order (inductive hypothesis), we are fine.
497 * - If an actor is added because it's getting killed, its subsequent actions shouldn't matter
498 * - If an actor gets added to actors_to_run because one of their blocking action constituting the meat
499 * of a simcall terminates, we're still good. Proof:
500 * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the
501 * resource kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications
503 * - For communications, this function is called from SIMIX_comm_finish().
504 * This function itself don't mess with the order since simcalls are handled in FIFO order.
505 * The function is called:
506 * - before the comm starts (invalid parameters, or resource already dead or whatever).
507 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
508 * - because the communication failed or were canceled after startup. In this case, it's called from
509 * the function we are in, by the chunk:
510 * set = model->states.failed_action_set;
511 * while ((synchro = extract(set)))
512 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
513 * This order is also fixed because it depends of the order in which the surf actions were
514 * added to the system, and only maestro can add stuff this way, through simcalls.
515 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
516 * popped out of the set does not depend on the user code's execution order.
517 * - because the communication terminated. In this case, synchros are served in the order given by
518 * set = model->states.done_action_set;
519 * while ((synchro = extract(set)))
520 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
521 * and the argument is very similar to the previous one.
522 * So, in any case, the orders of calls to CommImpl::finish() do not depend on the order in which user
523 * actors are executed.
524 * So, in any cases, the orders of actors within actors_to_run do not depend on the order in which
525 * user actors were executed previously.
526 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
527 * - If there is some actor killings, the order is changed by this decision that comes from user-land
528 * But this decision may not have been motivated by a situation that were different because the simulation is
530 * So, even the order change induced by the actor killing is perfectly reproducible.
532 * So science works, bitches [http://xkcd.com/54/].
534 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
535 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
537 * That would thus be a pure waste of time.
540 for (auto const& actor : actors_that_ran_) {
541 if (actor->simcall_.call_ != simix::Simcall::NONE) {
542 actor->simcall_handle(0);
548 wake_all_waiting_actors();
549 } while (execute_tasks());
551 /* If only daemon actors remain, cancel their actions, mark them to die and reschedule them */
552 if (actor_list_.size() == daemons_.size())
553 for (auto const& dmon : daemons_) {
554 XBT_DEBUG("Kill %s", dmon->get_cname());
555 maestro_->kill(dmon);
559 time = timer::Timer::next();
560 if (time > -1.0 || not actor_list_.empty()) {
561 XBT_DEBUG("Calling surf_solve");
562 time = surf_solve(time);
563 XBT_DEBUG("Moving time ahead : %g", time);
566 /* Notify all the hosts that have failed */
567 /* FIXME: iterate through the list of failed host and mark each of them */
568 /* as failed. On each host, signal all the running actors with host_fail */
570 // Execute timers and tasks until there isn't anything to be done:
573 again = timer::Timer::execute_all();
576 wake_all_waiting_actors();
579 /* Clean actors to destroy */
582 XBT_DEBUG("### time %f, #actors %zu, #to_run %zu", time, actor_list_.size(), actors_to_run_.size());
584 if (time < 0. && actors_to_run_.empty() && not actor_list_.empty()) {
585 if (actor_list_.size() <= daemons_.size()) {
586 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
587 "once the simulation is over. Please fix your on_exit() functions.");
589 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
591 display_all_actor_status();
592 simgrid::s4u::Engine::on_deadlock();
593 for (auto const& kv : actor_list_) {
594 XBT_DEBUG("Kill %s", kv.second->get_cname());
595 maestro_->kill(kv.second);
598 } while (time > -1.0 || has_actors_to_run());
600 if (not actor_list_.empty())
603 simgrid::s4u::Engine::on_simulation_end();
605 } // namespace kernel
606 } // namespace simgrid