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/routing/NetPoint.hpp"
10 #include "simgrid/kernel/routing/NetZoneImpl.hpp"
11 #include "simgrid/s4u/Host.hpp"
12 #include "simgrid/sg_config.hpp"
13 #include "src/include/surf/surf.hpp" //get_clock() and surf_solve()
14 #include "src/kernel/resource/DiskImpl.hpp"
15 #include "src/mc/mc_record.hpp"
16 #include "src/mc/mc_replay.hpp"
17 #include "src/simix/smx_private.hpp"
18 #include "src/surf/network_interface.hpp"
19 #include "src/surf/xml/platf.hpp" // FIXME: KILLME. There must be a better way than mimicking XML here
21 XBT_LOG_NEW_DEFAULT_CATEGORY(ker_engine, "Logging specific to Engine (kernel)");
26 config::Flag<double> cfg_breakpoint{"debug/breakpoint",
27 "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
28 EngineImpl::~EngineImpl()
30 /* Since hosts_ is a std::map, the hosts are destroyed in the lexicographic order, which ensures that the output is
33 while (not hosts_.empty())
34 hosts_.begin()->second->destroy();
36 /* Also delete the other data */
38 for (auto const& kv : netpoints_)
41 for (auto const& kv : links_)
44 actors_to_run_.clear();
45 actors_that_ran_.clear();
48 void EngineImpl::load_deployment(const std::string& file) const
53 surf_parse_open(file);
58 void EngineImpl::register_function(const std::string& name, const actor::ActorCodeFactory& code)
60 registered_functions[name] = code;
62 void EngineImpl::register_default(const actor::ActorCodeFactory& code)
64 default_function = code;
67 void EngineImpl::add_model(std::shared_ptr<resource::Model> model, const std::vector<resource::Model*>& dependencies)
69 auto model_name = model->get_name();
70 xbt_assert(models_prio_.find(model_name) == models_prio_.end(),
71 "Model %s already exists, use model.set_name() to change its name", model_name.c_str());
73 for (const auto dep : dependencies) {
74 xbt_assert(models_prio_.find(dep->get_name()) != models_prio_.end(),
75 "Model %s doesn't exists. Impossible to use it as dependency.", dep->get_name().c_str());
77 models_.push_back(model.get());
78 models_prio_[model_name] = std::move(model);
81 /** Wake up all actors waiting for a Surf action to finish */
82 void EngineImpl::wake_all_waiting_actors() const
84 for (auto const& model : models_) {
85 XBT_DEBUG("Handling the failed actions (if any)");
86 while (auto* action = model->extract_failed_action()) {
87 XBT_DEBUG(" Handling Action %p", action);
88 if (action->get_activity() != nullptr)
89 activity::ActivityImplPtr(action->get_activity())->post();
91 XBT_DEBUG("Handling the terminated actions (if any)");
92 while (auto* action = model->extract_done_action()) {
93 XBT_DEBUG(" Handling Action %p", action);
94 if (action->get_activity() == nullptr)
95 XBT_DEBUG("probably vcpu's action %p, skip", action);
97 activity::ActivityImplPtr(action->get_activity())->post();
102 * @brief Executes the actors in actors_to_run.
104 * The actors in actors_to_run are run (in parallel if possible). On exit, actors_to_run is empty, and actors_that_ran
105 * contains the list of actors that just ran. The two lists are swapped so, be careful when using them before and after
106 * a call to this function.
108 void EngineImpl::run_all_actors()
110 simix_global->context_factory->run_all();
112 actors_to_run_.swap(actors_that_ran_);
113 actors_to_run_.clear();
116 /** Execute all the tasks that are queued, e.g. `.then()` callbacks of futures. */
117 bool EngineImpl::execute_tasks()
119 xbt_assert(tasksTemp.empty());
125 // We don't want the callbacks to modify the vector we are iterating over:
126 tasks.swap(tasksTemp);
128 // Execute all the queued tasks:
129 for (auto& task : tasksTemp)
133 } while (not tasks.empty());
138 void EngineImpl::rm_daemon(actor::ActorImpl* actor)
140 auto it = daemons_.find(actor);
141 xbt_assert(it != daemons_.end(), "The dying daemon is not a daemon after all. Please report that bug.");
145 void EngineImpl::add_actor_to_run_list_no_check(actor::ActorImpl* actor)
147 XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
148 actors_to_run_.push_back(actor);
151 void EngineImpl::add_actor_to_run_list(actor::ActorImpl* actor)
153 if (std::find(begin(actors_to_run_), end(actors_to_run_), actor) != end(actors_to_run_)) {
154 XBT_DEBUG("Actor %s is already in the to_run list", actor->get_cname());
156 XBT_DEBUG("Inserting [%p] %s(%s) in the to_run list", actor, actor->get_cname(), actor->get_host()->get_cname());
157 actors_to_run_.push_back(actor);
161 void EngineImpl::run()
163 if (MC_record_replay_is_active()) {
164 mc::replay(MC_record_path());
171 XBT_DEBUG("New Schedule Round; size(queue)=%zu", actors_to_run_.size());
173 if (cfg_breakpoint >= 0.0 && surf_get_clock() >= cfg_breakpoint) {
174 XBT_DEBUG("Breakpoint reached (%g)", cfg_breakpoint.get());
175 cfg_breakpoint = -1.0;
185 while (not actors_to_run_.empty()) {
186 XBT_DEBUG("New Sub-Schedule Round; size(queue)=%zu", actors_to_run_.size());
188 /* Run all actors that are ready to run, possibly in parallel */
191 /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round */
193 /* WARNING, the order *must* be fixed or you'll jeopardize the simulation reproducibility (see RR-7653) */
195 /* Here, the order is ok because:
197 * Short proof: only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do
198 * not impact its order.
200 * Long proof: actors remain sorted through an arbitrary (implicit, complex but fixed) order in all cases.
202 * - if there is no kill during the simulation, actors remain sorted according by their PID.
203 * Rationale: This can be proved inductively.
204 * Assume that actors_to_run is sorted at a beginning of one round (it is at round 0: the deployment file
205 * is parsed linearly).
206 * Let's show that it is still so at the end of this round.
207 * - if an actor is added when being created, that's from maestro. It can be either at startup
208 * time (and then in PID order), or in response to a process_create simcall. Since simcalls are handled
209 * in arbitrary order (inductive hypothesis), we are fine.
210 * - If an actor is added because it's getting killed, its subsequent actions shouldn't matter
211 * - If an actor gets added to actors_to_run because one of their blocking action constituting the meat
212 * of a simcall terminates, we're still good. Proof:
213 * - You are added from ActorImpl::simcall_answer() only. When this function is called depends on the
214 * resource kind (network, cpu, disk, whatever), but the same arguments hold. Let's take communications
216 * - For communications, this function is called from SIMIX_comm_finish().
217 * This function itself don't mess with the order since simcalls are handled in FIFO order.
218 * The function is called:
219 * - before the comm starts (invalid parameters, or resource already dead or whatever).
220 * The order then trivial holds since maestro didn't interrupt its handling of the simcall yet
221 * - because the communication failed or were canceled after startup. In this case, it's called from
222 * the function we are in, by the chunk:
223 * set = model->states.failed_action_set;
224 * while ((synchro = extract(set)))
225 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
226 * This order is also fixed because it depends of the order in which the surf actions were
227 * added to the system, and only maestro can add stuff this way, through simcalls.
228 * We thus use the inductive hypothesis once again to conclude that the order in which synchros are
229 * popped out of the set does not depend on the user code's execution order.
230 * - because the communication terminated. In this case, synchros are served in the order given by
231 * set = model->states.done_action_set;
232 * while ((synchro = extract(set)))
233 * SIMIX_simcall_post((smx_synchro_t) synchro->data);
234 * and the argument is very similar to the previous one.
235 * So, in any case, the orders of calls to CommImpl::finish() do not depend on the order in which user
236 * actors are executed.
237 * So, in any cases, the orders of actors within actors_to_run do not depend on the order in which
238 * user actors were executed previously.
239 * So, if there is no killing in the simulation, the simulation reproducibility is not jeopardized.
240 * - If there is some actor killings, the order is changed by this decision that comes from user-land
241 * But this decision may not have been motivated by a situation that were different because the simulation is
243 * So, even the order change induced by the actor killing is perfectly reproducible.
245 * So science works, bitches [http://xkcd.com/54/].
247 * We could sort the actors_that_ran array completely so that we can describe the order in which simcalls are
248 * handled (like "according to the PID of issuer"), but it's not mandatory (order is fixed already even if
250 * That would thus be a pure waste of time.
253 for (auto const& actor : actors_that_ran_) {
254 if (actor->simcall_.call_ != simix::Simcall::NONE) {
255 actor->simcall_handle(0);
261 wake_all_waiting_actors();
262 } while (execute_tasks());
264 /* If only daemon actors remain, cancel their actions, mark them to die and reschedule them */
265 if (simix_global->process_list.size() == daemons_.size())
266 for (auto const& dmon : daemons_) {
267 XBT_DEBUG("Kill %s", dmon->get_cname());
268 simix_global->maestro_->kill(dmon);
272 time = timer::Timer::next();
273 if (time > -1.0 || not simix_global->process_list.empty()) {
274 XBT_DEBUG("Calling surf_solve");
275 time = surf_solve(time);
276 XBT_DEBUG("Moving time ahead : %g", time);
279 /* Notify all the hosts that have failed */
280 /* FIXME: iterate through the list of failed host and mark each of them */
281 /* as failed. On each host, signal all the running actors with host_fail */
283 // Execute timers and tasks until there isn't anything to be done:
286 again = timer::Timer::execute_all();
289 wake_all_waiting_actors();
292 /* Clean actors to destroy */
293 simix_global->empty_trash();
295 XBT_DEBUG("### time %f, #actors %zu, #to_run %zu", time, simix_global->process_list.size(), actors_to_run_.size());
297 if (time < 0. && actors_to_run_.empty() && not simix_global->process_list.empty()) {
298 if (simix_global->process_list.size() <= daemons_.size()) {
299 XBT_CRITICAL("Oops! Daemon actors cannot do any blocking activity (communications, synchronization, etc) "
300 "once the simulation is over. Please fix your on_exit() functions.");
302 XBT_CRITICAL("Oops! Deadlock or code not perfectly clean.");
304 simix_global->display_all_actor_status();
305 simgrid::s4u::Engine::on_deadlock();
306 for (auto const& kv : simix_global->process_list) {
307 XBT_DEBUG("Kill %s", kv.second->get_cname());
308 simix_global->maestro_->kill(kv.second);
311 } while (time > -1.0 || has_actors_to_run());
313 if (not simix_global->process_list.empty())
316 simgrid::s4u::Engine::on_simulation_end();
318 } // namespace kernel
319 } // namespace simgrid