+
+void EngineImpl::add_model(std::shared_ptr<resource::Model> model, const std::vector<resource::Model*>& dependencies)
+{
+ auto model_name = model->get_name();
+ xbt_assert(models_prio_.find(model_name) == models_prio_.end(),
+ "Model %s already exists, use model.set_name() to change its name", model_name.c_str());
+
+ for (const auto dep : dependencies) {
+ xbt_assert(models_prio_.find(dep->get_name()) != models_prio_.end(),
+ "Model %s doesn't exists. Impossible to use it as dependency.", dep->get_name().c_str());
+ }
+ models_.push_back(model.get());
+ models_prio_[model_name] = std::move(model);
+}
+
+void EngineImpl::add_split_duplex_link(const std::string& name, std::unique_ptr<resource::SplitDuplexLinkImpl> link)
+{
+ split_duplex_links_[name] = std::move(link);
+}
+
+/** Wake up all actors waiting for a Surf action to finish */
+void EngineImpl::wake_all_waiting_actors() const
+{
+ for (auto const& model : models_) {
+ XBT_DEBUG("Handling the failed actions (if any)");
+ while (auto* action = model->extract_failed_action()) {
+ XBT_DEBUG(" Handling Action %p", action);
+ if (action->get_activity() != nullptr) {
+ // If nobody told the interface that the activity has failed, that's because no actor waits on it (maestro
+ // started it). SimDAG I see you!
+ if (action->get_activity()->get_actor() == maestro_)
+ action->get_activity()->get_iface()->complete(s4u::Activity::State::FAILED);
+
+ activity::ActivityImplPtr(action->get_activity())->post();
+ }
+ }
+ XBT_DEBUG("Handling the terminated actions (if any)");
+ while (auto* action = model->extract_done_action()) {
+ XBT_DEBUG(" Handling Action %p", action);
+ if (action->get_activity() == nullptr)
+ XBT_DEBUG("probably vcpu's action %p, skip", action);
+ else {
+ // If nobody told the interface that the activity is finished, that's because no actor waits on it (maestro
+ // started it). SimDAG I see you!
+ action->get_activity()->set_finish_time(action->get_finish_time());
+
+ if (action->get_activity()->get_actor() == maestro_)
+ action->get_activity()->get_iface()->complete(s4u::Activity::State::FINISHED);
+
+ activity::ActivityImplPtr(action->get_activity())->post();
+ }
+ }
+ }
+}
+/**
+ * @brief Executes the actors in actors_to_run.
+ *
+ * The actors in actors_to_run are run (in parallel if possible). On exit, actors_to_run is empty, and actors_that_ran
+ * contains the list of actors that just ran. The two lists are swapped so, be careful when using them before and after
+ * a call to this function.
+ */
+void EngineImpl::run_all_actors()
+{
+ instance_->get_context_factory()->run_all();
+
+ actors_to_run_.swap(actors_that_ran_);
+ actors_to_run_.clear();
+}
+
+actor::ActorImpl* EngineImpl::get_actor_by_pid(aid_t pid)
+{
+ auto item = actor_list_.find(pid);
+ if (item != actor_list_.end())
+ return item->second;
+
+ // Search the trash
+ for (auto& a : actors_to_destroy_)
+ if (a.get_pid() == pid)
+ return &a;
+ return nullptr; // Not found, even in the trash
+}
+
+/** Execute all the tasks that are queued, e.g. `.then()` callbacks of futures. */
+bool EngineImpl::execute_tasks()
+{
+ if (tasks.empty())
+ return false;
+
+ std::vector<xbt::Task<void()>> tasksTemp;
+ do {
+ // We don't want the callbacks to modify the vector we are iterating over:
+ tasks.swap(tasksTemp);
+
+ // Execute all the queued tasks:
+ for (auto& task : tasksTemp)
+ task();
+
+ tasksTemp.clear();
+ } while (not tasks.empty());
+
+ return true;
+}
+
+void EngineImpl::remove_daemon(actor::ActorImpl* actor)
+{
+ auto it = daemons_.find(actor);
+ xbt_assert(it != daemons_.end(), "The dying daemon is not a daemon after all. Please report that bug.");
+ daemons_.erase(it);
+}
+
+void EngineImpl::add_actor_to_run_list_no_check(actor::ActorImpl* actor)
+{
+ XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
+ actors_to_run_.push_back(actor);
+}
+
+void EngineImpl::add_actor_to_run_list(actor::ActorImpl* actor)
+{
+ if (std::find(begin(actors_to_run_), end(actors_to_run_), actor) != end(actors_to_run_)) {
+ XBT_DEBUG("Actor %s is already in the to_run list", actor->get_cname());
+ } else {
+ XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
+ actors_to_run_.push_back(actor);
+ }
+}
+void EngineImpl::empty_trash()
+{
+ while (not actors_to_destroy_.empty()) {
+ actor::ActorImpl* actor = &actors_to_destroy_.front();
+ actors_to_destroy_.pop_front();
+ XBT_DEBUG("Getting rid of %s (refcount: %d)", actor->get_cname(), actor->get_refcount());
+ intrusive_ptr_release(actor);
+ }
+#if SIMGRID_HAVE_MC
+ xbt_dynar_reset(dead_actors_vector_);
+#endif
+}
+
+void EngineImpl::display_all_actor_status() const
+{
+ XBT_INFO("%zu actors are still running, waiting for something.", actor_list_.size());
+ /* List the actors and their state */
+ XBT_INFO("Legend of the following listing: \"Actor <pid> (<name>@<host>): <status>\"");
+ for (auto const& kv : actor_list_) {
+ const actor::ActorImpl* actor = kv.second;
+
+ if (actor->waiting_synchro_) {
+ const char* synchro_description = "unknown";
+
+ if (boost::dynamic_pointer_cast<kernel::activity::ExecImpl>(actor->waiting_synchro_) != nullptr)
+ synchro_description = "execution";
+
+ if (boost::dynamic_pointer_cast<kernel::activity::CommImpl>(actor->waiting_synchro_) != nullptr)
+ synchro_description = "communication";
+
+ if (boost::dynamic_pointer_cast<kernel::activity::SleepImpl>(actor->waiting_synchro_) != nullptr)
+ synchro_description = "sleeping";
+
+ if (boost::dynamic_pointer_cast<kernel::activity::RawImpl>(actor->waiting_synchro_) != nullptr)
+ synchro_description = "synchronization";
+
+ if (boost::dynamic_pointer_cast<kernel::activity::IoImpl>(actor->waiting_synchro_) != nullptr)
+ synchro_description = "I/O";
+
+ XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %s to finish", actor->get_pid(),
+ actor->get_cname(), actor->get_host()->get_cname(), synchro_description,
+ (xbt_log_no_loc ? (size_t)0xDEADBEEF : (size_t)actor->waiting_synchro_.get()),
+ actor->waiting_synchro_->get_cname(), actor->waiting_synchro_->get_state_str());
+ } else {
+ XBT_INFO("Actor %ld (%s@%s) simcall %s", actor->get_pid(), actor->get_cname(), actor->get_host()->get_cname(),
+ SIMIX_simcall_name(actor->simcall_));
+ }
+ }
+}
+
+void EngineImpl::presolve() const
+{
+ XBT_DEBUG("Consume all trace events occurring before the starting time.");
+ double next_event_date;
+ while ((next_event_date = profile::future_evt_set.next_date()) != -1.0) {
+ if (next_event_date > NOW)
+ break;
+
+ double value = -1.0;
+ resource::Resource* resource = nullptr;
+ while (auto* event = profile::future_evt_set.pop_leq(next_event_date, &value, &resource)) {
+ if (value >= 0)
+ resource->apply_event(event, value);
+ }
+ }
+
+ XBT_DEBUG("Set every models in the right state by updating them to 0.");
+ for (auto const& model : models_)
+ model->update_actions_state(NOW, 0.0);
+}
+
+double EngineImpl::solve(double max_date) const
+{
+ double time_delta = -1.0; /* duration */
+ double value = -1.0;
+ resource::Resource* resource = nullptr;
+
+ if (max_date != -1.0) {
+ xbt_assert(max_date >= NOW, "You asked to simulate up to %f, but that's in the past already", max_date);
+
+ time_delta = max_date - NOW;
+ }
+
+ XBT_DEBUG("Looking for next event in all models");
+ for (auto model : models_) {
+ if (not model->next_occurring_event_is_idempotent()) {
+ continue;
+ }
+ double next_event = model->next_occurring_event(NOW);
+ if ((time_delta < 0.0 || next_event < time_delta) && next_event >= 0.0) {
+ time_delta = next_event;
+ }
+ }
+
+ XBT_DEBUG("Min for resources (remember that NS3 don't update that value): %f", time_delta);
+
+ XBT_DEBUG("Looking for next trace event");
+
+ while (true) { // Handle next occurring events until none remains
+ double next_event_date = profile::future_evt_set.next_date();
+ XBT_DEBUG("Next TRACE event: %f", next_event_date);
+
+ for (auto model : models_) {
+ /* Skip all idempotent models, they were already treated above
+ * NS3 is the one to handled here */
+ if (model->next_occurring_event_is_idempotent())
+ continue;
+
+ if (next_event_date != -1.0) {
+ time_delta = std::min(next_event_date - NOW, time_delta);
+ } else {
+ time_delta = std::max(next_event_date - NOW, time_delta); // Get the positive component
+ }
+
+ XBT_DEBUG("Run the NS3 network at most %fs", time_delta);
+ // run until min or next flow
+ double model_next_action_end = model->next_occurring_event(time_delta);
+
+ XBT_DEBUG("Min for network : %f", model_next_action_end);
+ if (model_next_action_end >= 0.0)
+ time_delta = model_next_action_end;
+ }
+
+ if (next_event_date < 0.0 || (next_event_date > NOW + time_delta)) {
+ // next event may have already occurred or will after the next resource change, then bail out
+ XBT_DEBUG("no next usable TRACE event. Stop searching for it");
+ break;
+ }
+
+ XBT_DEBUG("Updating models (min = %g, NOW = %g, next_event_date = %g)", time_delta, NOW, next_event_date);
+
+ while (auto* event = profile::future_evt_set.pop_leq(next_event_date, &value, &resource)) {
+ if (resource->is_used() || (watched_hosts().find(resource->get_cname()) != watched_hosts().end())) {
+ time_delta = next_event_date - NOW;
+ XBT_DEBUG("This event invalidates the next_occurring_event() computation of models. Next event set to %f",
+ time_delta);
+ }
+ // FIXME: I'm too lame to update NOW live, so I change it and restore it so that the real update with surf_min
+ // will work
+ double round_start = NOW;
+ NOW = next_event_date;
+ /* update state of the corresponding resource to the new value. Does not touch lmm.
+ It will be modified if needed when updating actions */
+ XBT_DEBUG("Calling update_resource_state for resource %s", resource->get_cname());
+ resource->apply_event(event, value);
+ NOW = round_start;
+ }
+ }
+
+ /* FIXME: Moved this test to here to avoid stopping simulation if there are actions running on cpus and all cpus are
+ * with availability = 0. This may cause an infinite loop if one cpu has a trace with periodicity = 0 and the other a
+ * trace with periodicity > 0.
+ * The options are: all traces with same periodicity(0 or >0) or we need to change the way how the events are managed
+ */
+ if (time_delta < 0) {
+ XBT_DEBUG("No next event at all. Bail out now.");
+ return -1.0;
+ }
+
+ XBT_DEBUG("Duration set to %f", time_delta);
+
+ // Bump the time: jump into the future
+ NOW = NOW + time_delta;
+
+ // Inform the models of the date change
+ for (auto const& model : models_)
+ model->update_actions_state(NOW, time_delta);
+
+ s4u::Engine::on_time_advance(time_delta);
+
+ return time_delta;
+}
+
+void EngineImpl::run(double max_date)
+{
+ seal_platform();
+
+ if (MC_record_replay_is_active()) {
+ mc::RecordTrace::replay(MC_record_path());
+ empty_trash();
+ return;
+ }
+
+ double elapsed_time = -1;
+ const std::set<s4u::Activity*>* vetoed_activities = s4u::Activity::get_vetoed_activities();
+
+ do {
+ XBT_DEBUG("New Schedule Round; size(queue)=%zu", actors_to_run_.size());
+
+ if (cfg_breakpoint >= 0.0 && simgrid_get_clock() >= cfg_breakpoint) {
+ XBT_DEBUG("Breakpoint reached (%g)", cfg_breakpoint.get());
+ cfg_breakpoint = -1.0;
+#ifdef SIGTRAP
+ std::raise(SIGTRAP);
+#else
+ std::raise(SIGABRT);
+#endif
+ }
+
+ execute_tasks();
+
+ while (not actors_to_run_.empty()) {
+ XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", actors_to_run_.size());
+
+ /* Run all actors that are ready to run, possibly in parallel */
+ run_all_actors();
+
+ /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
+
+ /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
+
+ /* Here, the order is ok because:
+ *
+ * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
+ * not impact its order.
+ *
+ * Long proof: actors remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
+ *
+ * - if there is no kill during the simulation, actors remain sorted according by their PID.
+ * Rationale: This can be proved inductively.
+ * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
+ * is parsed linearly).
+ * Let's show that it is still so at the end of this round.
+ * - if an actor is added when being created, that's from maestro. It can be either at startup
+ * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
+ * in arbitrary order (inductive hypothesis), we are fine.
+ * - If an actor is added because it's getting killed, its subsequent actions shouldn't matter
+ * - If an actor gets added to actors_to_run because one of their blocking action constituting the meat
+ * of a simcall terminates, we're still good. Proof:
+ * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the
+ * resource kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications
+ * as an example.
+ * - For communications, this function is called from CommImpl::finish().
+ * This function itself don't mess with the order since simcalls are handled in FIFO order.
+ * The function is called:
+ * - before the comm starts (invalid parameters, or resource already dead or whatever).
+ * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
+ * - because the communication failed or were canceled after startup. In this case, it's called from
+ * the function we are in, by the chunk:
+ * set = model->states.failed_action_set;
+ * while ((synchro = extract(set)))
+ * SIMIX_simcall_post((smx_synchro_t) synchro->data);
+ * This order is also fixed because it depends of the order in which the surf actions were
+ * added to the system, and only maestro can add stuff this way, through simcalls.
+ * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
+ * popped out of the set does not depend on the user code's execution order.
+ * - because the communication terminated. In this case, synchros are served in the order given by
+ * set = model->states.done_action_set;
+ * while ((synchro = extract(set)))
+ * SIMIX_simcall_post((smx_synchro_t) synchro->data);
+ * and the argument is very similar to the previous one.
+ * So, in any case, the orders of calls to CommImpl::finish() do not depend on the order in which user
+ * actors are executed.
+ * So, in any cases, the orders of actors within actors_to_run do not depend on the order in which
+ * user actors were executed previously.
+ * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
+ * - If there is some actor killings, the order is changed by this decision that comes from user-land
+ * But this decision may not have been motivated by a situation that were different because the simulation is
+ * not reproducible.
+ * So, even the order change induced by the actor killing is perfectly reproducible.
+ *
+ * So science works, bitches [http://xkcd.com/54/].
+ *
+ * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
+ * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
+ * unfriendly).
+ * That would thus be a pure waste of time.
+ */
+
+ for (auto const& actor : actors_that_ran_) {
+ if (actor->simcall_.call_ != simix::Simcall::NONE) {
+ actor->simcall_handle(0);
+ }
+ }
+
+ execute_tasks();
+ do {
+ wake_all_waiting_actors();
+ } while (execute_tasks());
+
+ /* If only daemon actors remain, cancel their actions, mark them to die and reschedule them */
+ if (actor_list_.size() == daemons_.size())
+ for (auto const& dmon : daemons_) {
+ XBT_DEBUG("Kill %s", dmon->get_cname());
+ maestro_->kill(dmon);
+ }
+ }
+
+ // Compute the max_date of the next solve.
+ // It's either when a timer occurs, or when user-specified deadline is reached, or -1 if none is given
+ double next_time = timer::Timer::next();
+ if (next_time < 0 && max_date > -1) {
+ next_time = max_date;
+ } else if (next_time > -1 && max_date > -1) { // either both <0, or both >0
+ next_time = std::min(next_time, max_date);
+ }
+
+ XBT_DEBUG("Calling solve(%g) %g", next_time, NOW);
+ elapsed_time = solve(next_time);
+ XBT_DEBUG("Moving time ahead. NOW=%g; elapsed: %g", NOW, elapsed_time);
+
+ /* Notify all the hosts that have failed */
+ /* FIXME: iterate through the list of failed host and mark each of them */
+ /* as failed. On each host, signal all the running actors with host_fail */
+
+ // Execute timers and tasks until there isn't anything to be done:
+ bool again = false;
+ do {
+ again = timer::Timer::execute_all();
+ if (execute_tasks())
+ again = true;
+ wake_all_waiting_actors();
+ } while (again);
+
+ /* Clean actors to destroy */
+ empty_trash();
+
+ XBT_DEBUG("### elapsed time %f, #actors %zu, #to_run %zu, #vetoed %d", elapsed_time, actor_list_.size(),
+ actors_to_run_.size(), (vetoed_activities == nullptr ? -1 : static_cast<int>(vetoed_activities->size())));
+
+ if (elapsed_time < 0. && actors_to_run_.empty() && not actor_list_.empty()) {
+ if (actor_list_.size() <= daemons_.size()) {
+ XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
+ "once the simulation is over. Please fix your on_exit() functions.");
+ } else {
+ XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
+ }
+ display_all_actor_status();
+ simgrid::s4u::Engine::on_deadlock();
+ for (auto const& kv : actor_list_) {
+ XBT_DEBUG("Kill %s", kv.second->get_cname());
+ maestro_->kill(kv.second);
+ }
+ }
+ } while ((vetoed_activities == nullptr || vetoed_activities->empty()) &&
+ ((elapsed_time > -1.0 && not double_equals(max_date, NOW, 0.00001)) || has_actors_to_run()));
+
+ if (not actor_list_.empty() && max_date < 0 && not(vetoed_activities == nullptr || vetoed_activities->empty()))
+ THROW_IMPOSSIBLE;
+
+ simgrid::s4u::Engine::on_simulation_end();
+}
+
+double EngineImpl::get_clock()
+{
+ return NOW;
+}
+} // namespace kernel
+} // namespace simgrid
+
+void SIMIX_run() // XBT_ATTRIB_DEPRECATED_v332
+{
+ simgrid::kernel::EngineImpl::get_instance()->run(-1);