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>
29 XBT_LOG_NEW_DEFAULT_CATEGORY(ker_engine, "Logging specific to Engine (kernel)");
33 EngineImpl* EngineImpl::instance_ = nullptr; /* That singleton is awful too. */
35 config::Flag<double> cfg_breakpoint{"debug/breakpoint",
36 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
37 void EngineImpl::shutdown()
39 static bool already_cleaned_up = false;
40 if (already_cleaned_up)
41 return; // to avoid double cleaning by java and C
42 already_cleaned_up = true;
43 if (not EngineImpl::instance_) {
44 simix_global->destroy_maestro();
45 simix_global->destroy_context_factory();
46 return; // Nothing more to shutdown
48 XBT_DEBUG("EngineImpl::shutdown() called. Simulation's over.");
50 if (instance_->has_actors_to_run() && simgrid_get_clock() <= 0.0) {
52 XBT_CRITICAL("The time is still 0, and you still have processes ready to run.");
53 XBT_CRITICAL("It seems that you forgot to run the simulation that you setup.");
54 xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code.");
58 if (not instance_->actor_list_.empty()) {
59 if (smpi_process()->initialized()) {
60 xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
62 XBT_WARN("Process called exit when leaving - Skipping cleanups");
68 /* Kill all actors (but maestro) */
69 simix_global->get_maestro()->kill_all();
70 instance_->run_all_actors();
71 instance_->empty_trash();
73 /* Let's free maestro now */
74 simix_global->destroy_maestro();
76 /* Finish context module and SURF */
77 simix_global->destroy_context_factory();
79 while (not timer::kernel_timers().empty()) {
80 delete timer::kernel_timers().top().second;
81 timer::kernel_timers().pop();
87 simgrid::s4u::Engine::shutdown();
89 simix_global = nullptr;
92 EngineImpl::~EngineImpl()
94 /* Since hosts_ is a std::map, the hosts are destroyed in the lexicographic order, which ensures that the output is
97 while (not hosts_.empty())
98 hosts_.begin()->second->destroy();
100 /* Also delete the other data */
101 delete netzone_root_;
102 for (auto const& kv : netpoints_)
105 for (auto const& kv : links_)
107 kv.second->destroy();
109 for (auto const& kv : mailboxes_)
112 /* Free the remaining data structures */
114 xbt_dynar_free(&actors_vector_);
115 xbt_dynar_free(&dead_actors_vector_);
117 /* clear models before freeing handle, network models can use external callback defined in the handle */
118 models_prio_.clear();
121 void EngineImpl::load_platform(const std::string& platf)
123 double start = xbt_os_time();
124 if (boost::algorithm::ends_with(platf, ".so") or boost::algorithm::ends_with(platf, ".dylib")) {
126 xbt_die("loading platform through shared library isn't supported on windows");
128 void* handle = dlopen(platf.c_str(), RTLD_LAZY);
129 xbt_assert(handle, "Impossible to open platform file: %s", platf.c_str());
130 platf_handle_ = std::unique_ptr<void, std::function<int(void*)>>(handle, dlclose);
131 using load_fct_t = void (*)(const simgrid::s4u::Engine&);
132 auto callable = (load_fct_t)dlsym(platf_handle_.get(), "load_platform");
133 const char* dlsym_error = dlerror();
134 xbt_assert(not dlsym_error, "Error: %s", dlsym_error);
135 callable(*simgrid::s4u::Engine::get_instance());
138 parse_platform_file(platf);
141 double end = xbt_os_time();
142 XBT_DEBUG("PARSE TIME: %g", (end - start));
145 void EngineImpl::load_deployment(const std::string& file) const
150 surf_parse_open(file);
155 void EngineImpl::register_function(const std::string& name, const actor::ActorCodeFactory& code)
157 registered_functions[name] = code;
159 void EngineImpl::register_default(const actor::ActorCodeFactory& code)
161 default_function = code;
164 void EngineImpl::add_model(std::shared_ptr<resource::Model> model, const std::vector<resource::Model*>& dependencies)
166 auto model_name = model->get_name();
167 xbt_assert(models_prio_.find(model_name) == models_prio_.end(),
168 "Model %s already exists, use model.set_name() to change its name", model_name.c_str());
170 for (const auto dep : dependencies) {
171 xbt_assert(models_prio_.find(dep->get_name()) != models_prio_.end(),
172 "Model %s doesn't exists. Impossible to use it as dependency.", dep->get_name().c_str());
174 models_.push_back(model.get());
175 models_prio_[model_name] = std::move(model);
178 void EngineImpl::add_split_duplex_link(const std::string& name, std::unique_ptr<resource::SplitDuplexLinkImpl> link)
180 split_duplex_links_[name] = std::move(link);
183 /** Wake up all actors waiting for a Surf action to finish */
184 void EngineImpl::wake_all_waiting_actors() const
186 for (auto const& model : models_) {
187 XBT_DEBUG("Handling the failed actions (if any)");
188 while (auto* action = model->extract_failed_action()) {
189 XBT_DEBUG(" Handling Action %p", action);
190 if (action->get_activity() != nullptr)
191 activity::ActivityImplPtr(action->get_activity())->post();
193 XBT_DEBUG("Handling the terminated actions (if any)");
194 while (auto* action = model->extract_done_action()) {
195 XBT_DEBUG(" Handling Action %p", action);
196 if (action->get_activity() == nullptr)
197 XBT_DEBUG("probably vcpu's action %p, skip", action);
199 activity::ActivityImplPtr(action->get_activity())->post();
204 * @brief Executes the actors in actors_to_run.
206 * The actors in actors_to_run are run (in parallel if possible). On exit, actors_to_run is empty, and actors_that_ran
207 * contains the list of actors that just ran. The two lists are swapped so, be careful when using them before and after
208 * a call to this function.
210 void EngineImpl::run_all_actors()
212 simix_global->get_context_factory()->run_all();
214 actors_to_run_.swap(actors_that_ran_);
215 actors_to_run_.clear();
218 actor::ActorImpl* EngineImpl::get_actor_by_pid(aid_t pid)
220 auto item = actor_list_.find(pid);
221 if (item != actor_list_.end())
225 for (auto& a : actors_to_destroy_)
226 if (a.get_pid() == pid)
228 return nullptr; // Not found, even in the trash
230 /** Execute all the tasks that are queued, e.g. `.then()` callbacks of futures. */
231 bool EngineImpl::execute_tasks()
236 std::vector<xbt::Task<void()>> tasksTemp;
238 // We don't want the callbacks to modify the vector we are iterating over:
239 tasks.swap(tasksTemp);
241 // Execute all the queued tasks:
242 for (auto& task : tasksTemp)
246 } while (not tasks.empty());
251 void EngineImpl::remove_daemon(actor::ActorImpl* actor)
253 auto it = daemons_.find(actor);
254 xbt_assert(it != daemons_.end(), "The dying daemon is not a daemon after all. Please report that bug.");
258 void EngineImpl::add_actor_to_run_list_no_check(actor::ActorImpl* actor)
260 XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
261 actors_to_run_.push_back(actor);
264 void EngineImpl::add_actor_to_run_list(actor::ActorImpl* actor)
266 if (std::find(begin(actors_to_run_), end(actors_to_run_), actor) != end(actors_to_run_)) {
267 XBT_DEBUG("Actor %s is already in the to_run list", actor->get_cname());
269 XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
270 actors_to_run_.push_back(actor);
273 void EngineImpl::empty_trash()
275 while (not actors_to_destroy_.empty()) {
276 actor::ActorImpl* actor = &actors_to_destroy_.front();
277 actors_to_destroy_.pop_front();
278 XBT_DEBUG("Getting rid of %s (refcount: %d)", actor->get_cname(), actor->get_refcount());
279 intrusive_ptr_release(actor);
282 xbt_dynar_reset(dead_actors_vector_);
286 void EngineImpl::display_all_actor_status() const
288 XBT_INFO("%zu actors are still running, waiting for something.", actor_list_.size());
289 /* List the actors and their state */
290 XBT_INFO("Legend of the following listing: \"Actor <pid> (<name>@<host>): <status>\"");
291 for (auto const& kv : actor_list_) {
292 actor::ActorImpl* actor = kv.second;
294 if (actor->waiting_synchro_) {
295 const char* synchro_description = "unknown";
297 if (boost::dynamic_pointer_cast<kernel::activity::ExecImpl>(actor->waiting_synchro_) != nullptr)
298 synchro_description = "execution";
300 if (boost::dynamic_pointer_cast<kernel::activity::CommImpl>(actor->waiting_synchro_) != nullptr)
301 synchro_description = "communication";
303 if (boost::dynamic_pointer_cast<kernel::activity::SleepImpl>(actor->waiting_synchro_) != nullptr)
304 synchro_description = "sleeping";
306 if (boost::dynamic_pointer_cast<kernel::activity::RawImpl>(actor->waiting_synchro_) != nullptr)
307 synchro_description = "synchronization";
309 if (boost::dynamic_pointer_cast<kernel::activity::IoImpl>(actor->waiting_synchro_) != nullptr)
310 synchro_description = "I/O";
312 XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %d to finish", actor->get_pid(),
313 actor->get_cname(), actor->get_host()->get_cname(), synchro_description,
314 (xbt_log_no_loc ? (size_t)0xDEADBEEF : (size_t)actor->waiting_synchro_.get()),
315 actor->waiting_synchro_->get_cname(), (int)actor->waiting_synchro_->state_);
317 XBT_INFO("Actor %ld (%s@%s) simcall %s", actor->get_pid(), actor->get_cname(), actor->get_host()->get_cname(),
318 SIMIX_simcall_name(actor->simcall_));
323 void EngineImpl::run()
325 if (MC_record_replay_is_active()) {
326 mc::replay(MC_record_path());
334 XBT_DEBUG("New Schedule Round; size(queue)=%zu", actors_to_run_.size());
336 if (cfg_breakpoint >= 0.0 && surf_get_clock() >= cfg_breakpoint) {
337 XBT_DEBUG("Breakpoint reached (%g)", cfg_breakpoint.get());
338 cfg_breakpoint = -1.0;
348 while (not actors_to_run_.empty()) {
349 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", actors_to_run_.size());
351 /* Run all actors that are ready to run, possibly in parallel */
354 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
356 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
358 /* Here, the order is ok because:
360 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
361 * not impact its order.
363 * Long proof: actors remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
365 * - if there is no kill during the simulation, actors remain sorted according by their PID.
366 * Rationale: This can be proved inductively.
367 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
368 * is parsed linearly).
369 * Let's show that it is still so at the end of this round.
370 * - if an actor is added when being created, that's from maestro. It can be either at startup
371 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
372 * in arbitrary order (inductive hypothesis), we are fine.
373 * - If an actor is added because it's getting killed, its subsequent actions shouldn't matter
374 * - If an actor gets added to actors_to_run because one of their blocking action constituting the meat
375 * of a simcall terminates, we're still good. Proof:
376 * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the
377 * resource kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications
379 * - For communications, this function is called from SIMIX_comm_finish().
380 * This function itself don't mess with the order since simcalls are handled in FIFO order.
381 * The function is called:
382 * - before the comm starts (invalid parameters, or resource already dead or whatever).
383 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
384 * - because the communication failed or were canceled after startup. In this case, it's called from
385 * the function we are in, by the chunk:
386 * set = model->states.failed_action_set;
387 * while ((synchro = extract(set)))
388 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
389 * This order is also fixed because it depends of the order in which the surf actions were
390 * added to the system, and only maestro can add stuff this way, through simcalls.
391 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
392 * popped out of the set does not depend on the user code's execution order.
393 * - because the communication terminated. In this case, synchros are served in the order given by
394 * set = model->states.done_action_set;
395 * while ((synchro = extract(set)))
396 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
397 * and the argument is very similar to the previous one.
398 * So, in any case, the orders of calls to CommImpl::finish() do not depend on the order in which user
399 * actors are executed.
400 * So, in any cases, the orders of actors within actors_to_run do not depend on the order in which
401 * user actors were executed previously.
402 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
403 * - If there is some actor killings, the order is changed by this decision that comes from user-land
404 * But this decision may not have been motivated by a situation that were different because the simulation is
406 * So, even the order change induced by the actor killing is perfectly reproducible.
408 * So science works, bitches [http://xkcd.com/54/].
410 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
411 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
413 * That would thus be a pure waste of time.
416 for (auto const& actor : actors_that_ran_) {
417 if (actor->simcall_.call_ != simix::Simcall::NONE) {
418 actor->simcall_handle(0);
424 wake_all_waiting_actors();
425 } while (execute_tasks());
427 /* If only daemon actors remain, cancel their actions, mark them to die and reschedule them */
428 if (actor_list_.size() == daemons_.size())
429 for (auto const& dmon : daemons_) {
430 XBT_DEBUG("Kill %s", dmon->get_cname());
431 simix_global->get_maestro()->kill(dmon);
435 time = timer::Timer::next();
436 if (time > -1.0 || not actor_list_.empty()) {
437 XBT_DEBUG("Calling surf_solve");
438 time = surf_solve(time);
439 XBT_DEBUG("Moving time ahead : %g", time);
442 /* Notify all the hosts that have failed */
443 /* FIXME: iterate through the list of failed host and mark each of them */
444 /* as failed. On each host, signal all the running actors with host_fail */
446 // Execute timers and tasks until there isn't anything to be done:
449 again = timer::Timer::execute_all();
452 wake_all_waiting_actors();
455 /* Clean actors to destroy */
458 XBT_DEBUG("### time %f, #actors %zu, #to_run %zu", time, actor_list_.size(), actors_to_run_.size());
460 if (time < 0. && actors_to_run_.empty() && not actor_list_.empty()) {
461 if (actor_list_.size() <= daemons_.size()) {
462 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
463 "once the simulation is over. Please fix your on_exit() functions.");
465 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
467 display_all_actor_status();
468 simgrid::s4u::Engine::on_deadlock();
469 for (auto const& kv : actor_list_) {
470 XBT_DEBUG("Kill %s", kv.second->get_cname());
471 simix_global->get_maestro()->kill(kv.second);
474 } while (time > -1.0 || has_actors_to_run());
476 if (not actor_list_.empty())
479 simgrid::s4u::Engine::on_simulation_end();
481 } // namespace kernel
482 } // namespace simgrid