+ seal_platform();
+
+ if (MC_record_replay_is_active()) {
+ mc::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();