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/smpi/include/smpi_actor.hpp"
20 #include "src/surf/network_interface.hpp"
21 #include "src/surf/xml/platf.hpp" // FIXME: KILLME. There must be a better way than mimicking XML here
23 #include <boost/algorithm/string/predicate.hpp>
29 #include "src/mc/remote/AppSide.hpp"
32 XBT_LOG_NEW_DEFAULT_CATEGORY(ker_engine, "Logging specific to Engine (kernel)");
35 EngineImpl* EngineImpl::instance_ = nullptr; /* That singleton is awful too. */
37 config::Flag<double> cfg_breakpoint{"debug/breakpoint",
38 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
39 config::Flag<bool> cfg_verbose_exit{"debug/verbose-exit", "Display the actor status at exit", true};
41 } // namespace simgrid
43 XBT_ATTRIB_NORETURN static void inthandler(int)
45 if (simgrid::kernel::cfg_verbose_exit) {
46 XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option "
47 "'debug/verbose-exit').");
48 simgrid::kernel::EngineImpl::get_instance()->display_all_actor_status();
50 XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'debug/verbose-exit' is set to false.");
56 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/)
58 if ((siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) || siginfo->si_signo == SIGBUS) {
60 "Access violation or Bus error detected.\n"
61 "This probably comes from a programming error in your code, or from a stack\n"
62 "overflow. If you are certain of your code, try increasing the stack size\n"
63 " --cfg=contexts/stack-size:XXX (current size is %u KiB).\n"
65 "If it does not help, this may have one of the following causes:\n"
66 "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
67 "Failing hardware can sometimes generate such errors too.\n"
69 "If you think you've found a bug in SimGrid, please report it along with a\n"
70 "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n"
71 "of the fault captured with gdb or valgrind.\n",
72 smx_context_stack_size / 1024);
73 } else if (siginfo->si_signo == SIGSEGV) {
74 fprintf(stderr, "Segmentation fault.\n");
76 if (smpi_enabled() && smpi_cfg_privatization() == SmpiPrivStrategies::NONE) {
77 #if HAVE_PRIVATIZATION
78 fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
80 fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
81 #endif /* HAVE_PRIVATIZATION */
83 #endif /* HAVE_SMPI */
89 * Install signal handler for SIGSEGV. Check that nobody has already installed
90 * its own handler. For example, the Java VM does this.
92 static void install_segvhandler()
96 if (simgrid::kernel::context::Context::install_sigsegv_stack(&old_stack, true) == -1) {
97 XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
100 if (not(old_stack.ss_flags & SS_DISABLE)) {
101 XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
102 old_stack.ss_size, (unsigned)old_stack.ss_flags);
103 sigaltstack(&old_stack, nullptr);
106 struct sigaction action;
107 struct sigaction old_action;
108 action.sa_sigaction = &segvhandler;
109 action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
110 sigemptyset(&action.sa_mask);
112 /* Linux tend to raise only SIGSEGV where other systems also raise SIGBUS on severe error */
113 for (int sig : {SIGSEGV, SIGBUS}) {
114 if (sigaction(sig, &action, &old_action) == -1) {
115 XBT_WARN("Failed to register signal handler for signal %d: %s", sig, strerror(errno));
118 if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
119 XBT_DEBUG("A signal handler was already installed for signal %d (%p). Restore it.", sig,
120 (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
121 sigaction(sig, &old_action, nullptr);
131 EngineImpl::~EngineImpl()
133 /* Since hosts_ is a std::map, the hosts are destroyed in the lexicographic order, which ensures that the output is
136 while (not hosts_.empty())
137 hosts_.begin()->second->destroy();
139 /* Also delete the other data */
140 delete netzone_root_;
141 for (auto const& kv : netpoints_)
144 for (auto const& kv : links_)
146 kv.second->destroy();
148 for (auto const& kv : mailboxes_)
151 /* Free the remaining data structures */
153 xbt_dynar_free(&actors_vector_);
154 xbt_dynar_free(&dead_actors_vector_);
156 /* clear models before freeing handle, network models can use external callback defined in the handle */
157 models_prio_.clear();
160 void EngineImpl::initialize(int* argc, char** argv)
162 xbt_assert(EngineImpl::instance_ == nullptr,
163 "It is currently forbidden to create more than one instance of kernel::EngineImpl");
164 EngineImpl::instance_ = this;
166 // The communication initialization is done ASAP, as we need to get some init parameters from the MC for different
167 // layers. But simix_global needs to be created, as we send the address of some of its fields to the MC that wants to
168 // read them directly.
169 simgrid::mc::AppSide::initialize();
172 surf_init(argc, argv); /* Initialize SURF structures */
174 instance_->context_mod_init();
176 /* Prepare to display some more info when dying on Ctrl-C pressing */
177 std::signal(SIGINT, inthandler);
180 install_segvhandler();
183 /* register a function to be called by SURF after the environment creation */
185 s4u::Engine::on_platform_created.connect(surf_presolve);
187 if (config::get_value<bool>("debug/clean-atexit"))
191 void EngineImpl::shutdown()
193 if (EngineImpl::instance_ == nullptr)
195 XBT_DEBUG("EngineImpl::shutdown() called. Simulation's over.");
196 if (instance_->has_actors_to_run() && simgrid_get_clock() <= 0.0) {
198 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
199 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
200 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
203 /* Kill all actors (but maestro) */
204 instance_->maestro_->kill_all();
205 instance_->run_all_actors();
206 instance_->empty_trash();
209 if (not instance_->actor_list_.empty()) {
210 if (smpi_process()->initialized()) {
211 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
213 XBT_WARN("Process called exit when leaving - Skipping cleanups");
219 /* Let's free maestro now */
220 instance_->destroy_maestro();
222 /* Finish context module and SURF */
223 instance_->destroy_context_factory();
225 while (not timer::kernel_timers().empty()) {
226 delete timer::kernel_timers().top().second;
227 timer::kernel_timers().pop();
237 void EngineImpl::load_platform(const std::string& platf)
239 double start = xbt_os_time();
240 if (boost::algorithm::ends_with(platf, ".so") or boost::algorithm::ends_with(platf, ".dylib")) {
242 xbt_die("loading platform through shared library isn't supported on windows");
244 void* handle = dlopen(platf.c_str(), RTLD_LAZY);
245 xbt_assert(handle, "Impossible to open platform file: %s", platf.c_str());
246 platf_handle_ = std::unique_ptr<void, std::function<int(void*)>>(handle, dlclose);
247 using load_fct_t = void (*)(const simgrid::s4u::Engine&);
248 auto callable = (load_fct_t)dlsym(platf_handle_.get(), "load_platform");
249 const char* dlsym_error = dlerror();
250 xbt_assert(not dlsym_error, "Error: %s", dlsym_error);
251 callable(*simgrid::s4u::Engine::get_instance());
254 parse_platform_file(platf);
257 double end = xbt_os_time();
258 XBT_DEBUG("PARSE TIME: %g", (end - start));
261 void EngineImpl::load_deployment(const std::string& file) const
266 surf_parse_open(file);
271 void EngineImpl::register_function(const std::string& name, const actor::ActorCodeFactory& code)
273 registered_functions[name] = code;
275 void EngineImpl::register_default(const actor::ActorCodeFactory& code)
277 default_function = code;
280 void EngineImpl::add_model(std::shared_ptr<resource::Model> model, const std::vector<resource::Model*>& dependencies)
282 auto model_name = model->get_name();
283 xbt_assert(models_prio_.find(model_name) == models_prio_.end(),
284 "Model %s already exists, use model.set_name() to change its name", model_name.c_str());
286 for (const auto dep : dependencies) {
287 xbt_assert(models_prio_.find(dep->get_name()) != models_prio_.end(),
288 "Model %s doesn't exists. Impossible to use it as dependency.", dep->get_name().c_str());
290 models_.push_back(model.get());
291 models_prio_[model_name] = std::move(model);
294 void EngineImpl::add_split_duplex_link(const std::string& name, std::unique_ptr<resource::SplitDuplexLinkImpl> link)
296 split_duplex_links_[name] = std::move(link);
299 /** Wake up all actors waiting for a Surf action to finish */
300 void EngineImpl::wake_all_waiting_actors() const
302 for (auto const& model : models_) {
303 XBT_DEBUG("Handling the failed actions (if any)");
304 while (auto* action = model->extract_failed_action()) {
305 XBT_DEBUG(" Handling Action %p", action);
306 if (action->get_activity() != nullptr)
307 activity::ActivityImplPtr(action->get_activity())->post();
309 XBT_DEBUG("Handling the terminated actions (if any)");
310 while (auto* action = model->extract_done_action()) {
311 XBT_DEBUG(" Handling Action %p", action);
312 if (action->get_activity() == nullptr)
313 XBT_DEBUG("probably vcpu's action %p, skip", action);
315 activity::ActivityImplPtr(action->get_activity())->post();
320 * @brief Executes the actors in actors_to_run.
322 * The actors in actors_to_run are run (in parallel if possible). On exit, actors_to_run is empty, and actors_that_ran
323 * contains the list of actors that just ran. The two lists are swapped so, be careful when using them before and after
324 * a call to this function.
326 void EngineImpl::run_all_actors()
328 instance_->get_context_factory()->run_all();
330 actors_to_run_.swap(actors_that_ran_);
331 actors_to_run_.clear();
334 actor::ActorImpl* EngineImpl::get_actor_by_pid(aid_t pid)
336 auto item = actor_list_.find(pid);
337 if (item != actor_list_.end())
341 for (auto& a : actors_to_destroy_)
342 if (a.get_pid() == pid)
344 return nullptr; // Not found, even in the trash
347 /** Execute all the tasks that are queued, e.g. `.then()` callbacks of futures. */
348 bool EngineImpl::execute_tasks()
353 std::vector<xbt::Task<void()>> tasksTemp;
355 // We don't want the callbacks to modify the vector we are iterating over:
356 tasks.swap(tasksTemp);
358 // Execute all the queued tasks:
359 for (auto& task : tasksTemp)
363 } while (not tasks.empty());
368 void EngineImpl::remove_daemon(actor::ActorImpl* actor)
370 auto it = daemons_.find(actor);
371 xbt_assert(it != daemons_.end(), "The dying daemon is not a daemon after all. Please report that bug.");
375 void EngineImpl::add_actor_to_run_list_no_check(actor::ActorImpl* actor)
377 XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
378 actors_to_run_.push_back(actor);
381 void EngineImpl::add_actor_to_run_list(actor::ActorImpl* actor)
383 if (std::find(begin(actors_to_run_), end(actors_to_run_), actor) != end(actors_to_run_)) {
384 XBT_DEBUG("Actor %s is already in the to_run list", actor->get_cname());
386 XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
387 actors_to_run_.push_back(actor);
390 void EngineImpl::empty_trash()
392 while (not actors_to_destroy_.empty()) {
393 actor::ActorImpl* actor = &actors_to_destroy_.front();
394 actors_to_destroy_.pop_front();
395 XBT_DEBUG("Getting rid of %s (refcount: %d)", actor->get_cname(), actor->get_refcount());
396 intrusive_ptr_release(actor);
399 xbt_dynar_reset(dead_actors_vector_);
403 void EngineImpl::display_all_actor_status() const
405 XBT_INFO("%zu actors are still running, waiting for something.", actor_list_.size());
406 /* List the actors and their state */
407 XBT_INFO("Legend of the following listing: \"Actor <pid> (<name>@<host>): <status>\"");
408 for (auto const& kv : actor_list_) {
409 actor::ActorImpl* actor = kv.second;
411 if (actor->waiting_synchro_) {
412 const char* synchro_description = "unknown";
414 if (boost::dynamic_pointer_cast<kernel::activity::ExecImpl>(actor->waiting_synchro_) != nullptr)
415 synchro_description = "execution";
417 if (boost::dynamic_pointer_cast<kernel::activity::CommImpl>(actor->waiting_synchro_) != nullptr)
418 synchro_description = "communication";
420 if (boost::dynamic_pointer_cast<kernel::activity::SleepImpl>(actor->waiting_synchro_) != nullptr)
421 synchro_description = "sleeping";
423 if (boost::dynamic_pointer_cast<kernel::activity::RawImpl>(actor->waiting_synchro_) != nullptr)
424 synchro_description = "synchronization";
426 if (boost::dynamic_pointer_cast<kernel::activity::IoImpl>(actor->waiting_synchro_) != nullptr)
427 synchro_description = "I/O";
429 XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %d to finish", actor->get_pid(),
430 actor->get_cname(), actor->get_host()->get_cname(), synchro_description,
431 (xbt_log_no_loc ? (size_t)0xDEADBEEF : (size_t)actor->waiting_synchro_.get()),
432 actor->waiting_synchro_->get_cname(), (int)actor->waiting_synchro_->state_);
434 XBT_INFO("Actor %ld (%s@%s) simcall %s", actor->get_pid(), actor->get_cname(), actor->get_host()->get_cname(),
435 SIMIX_simcall_name(actor->simcall_));
440 void EngineImpl::run()
442 if (MC_record_replay_is_active()) {
443 mc::replay(MC_record_path());
451 XBT_DEBUG("New Schedule Round; size(queue)=%zu", actors_to_run_.size());
453 if (cfg_breakpoint >= 0.0 && surf_get_clock() >= cfg_breakpoint) {
454 XBT_DEBUG("Breakpoint reached (%g)", cfg_breakpoint.get());
455 cfg_breakpoint = -1.0;
465 while (not actors_to_run_.empty()) {
466 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", actors_to_run_.size());
468 /* Run all actors that are ready to run, possibly in parallel */
471 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
473 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
475 /* Here, the order is ok because:
477 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
478 * not impact its order.
480 * Long proof: actors remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
482 * - if there is no kill during the simulation, actors remain sorted according by their PID.
483 * Rationale: This can be proved inductively.
484 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
485 * is parsed linearly).
486 * Let's show that it is still so at the end of this round.
487 * - if an actor is added when being created, that's from maestro. It can be either at startup
488 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
489 * in arbitrary order (inductive hypothesis), we are fine.
490 * - If an actor is added because it's getting killed, its subsequent actions shouldn't matter
491 * - If an actor gets added to actors_to_run because one of their blocking action constituting the meat
492 * of a simcall terminates, we're still good. Proof:
493 * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the
494 * resource kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications
496 * - For communications, this function is called from SIMIX_comm_finish().
497 * This function itself don't mess with the order since simcalls are handled in FIFO order.
498 * The function is called:
499 * - before the comm starts (invalid parameters, or resource already dead or whatever).
500 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
501 * - because the communication failed or were canceled after startup. In this case, it's called from
502 * the function we are in, by the chunk:
503 * set = model->states.failed_action_set;
504 * while ((synchro = extract(set)))
505 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
506 * This order is also fixed because it depends of the order in which the surf actions were
507 * added to the system, and only maestro can add stuff this way, through simcalls.
508 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
509 * popped out of the set does not depend on the user code's execution order.
510 * - because the communication terminated. In this case, synchros are served in the order given by
511 * set = model->states.done_action_set;
512 * while ((synchro = extract(set)))
513 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
514 * and the argument is very similar to the previous one.
515 * So, in any case, the orders of calls to CommImpl::finish() do not depend on the order in which user
516 * actors are executed.
517 * So, in any cases, the orders of actors within actors_to_run do not depend on the order in which
518 * user actors were executed previously.
519 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
520 * - If there is some actor killings, the order is changed by this decision that comes from user-land
521 * But this decision may not have been motivated by a situation that were different because the simulation is
523 * So, even the order change induced by the actor killing is perfectly reproducible.
525 * So science works, bitches [http://xkcd.com/54/].
527 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
528 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
530 * That would thus be a pure waste of time.
533 for (auto const& actor : actors_that_ran_) {
534 if (actor->simcall_.call_ != simix::Simcall::NONE) {
535 actor->simcall_handle(0);
541 wake_all_waiting_actors();
542 } while (execute_tasks());
544 /* If only daemon actors remain, cancel their actions, mark them to die and reschedule them */
545 if (actor_list_.size() == daemons_.size())
546 for (auto const& dmon : daemons_) {
547 XBT_DEBUG("Kill %s", dmon->get_cname());
548 maestro_->kill(dmon);
552 time = timer::Timer::next();
553 if (time > -1.0 || not actor_list_.empty()) {
554 XBT_DEBUG("Calling surf_solve");
555 time = surf_solve(time);
556 XBT_DEBUG("Moving time ahead : %g", time);
559 /* Notify all the hosts that have failed */
560 /* FIXME: iterate through the list of failed host and mark each of them */
561 /* as failed. On each host, signal all the running actors with host_fail */
563 // Execute timers and tasks until there isn't anything to be done:
566 again = timer::Timer::execute_all();
569 wake_all_waiting_actors();
572 /* Clean actors to destroy */
575 XBT_DEBUG("### time %f, #actors %zu, #to_run %zu", time, actor_list_.size(), actors_to_run_.size());
577 if (time < 0. && actors_to_run_.empty() && not actor_list_.empty()) {
578 if (actor_list_.size() <= daemons_.size()) {
579 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
580 "once the simulation is over. Please fix your on_exit() functions.");
582 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
584 display_all_actor_status();
585 simgrid::s4u::Engine::on_deadlock();
586 for (auto const& kv : actor_list_) {
587 XBT_DEBUG("Kill %s", kv.second->get_cname());
588 maestro_->kill(kv.second);
591 } while (time > -1.0 || has_actors_to_run());
593 if (not actor_list_.empty())
596 simgrid::s4u::Engine::on_simulation_end();
598 } // namespace kernel
599 } // namespace simgrid