X-Git-Url: http://bilbo.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/a3969091fee6a98f564720fe0be36a1d1d4d760a..2f19ac3ba6ed1cd18055078e3043b16dcaf0058e:/src/kernel/EngineImpl.cpp diff --git a/src/kernel/EngineImpl.cpp b/src/kernel/EngineImpl.cpp index b35092d028..f8bb4bcc14 100644 --- a/src/kernel/EngineImpl.cpp +++ b/src/kernel/EngineImpl.cpp @@ -10,6 +10,9 @@ #include #include +#define SIMIX_H_NO_DEPRECATED_WARNING // avoid deprecation warning on include (remove with XBT_ATTRIB_DEPRECATED_v332) +#include + #include "mc/mc.h" #include "src/kernel/EngineImpl.hpp" #include "src/kernel/resource/StandardLinkImpl.hpp" @@ -41,15 +44,6 @@ config::Flag cfg_breakpoint{"debug/breakpoint", "When non-negative, raise a SIGTRAP after given (simulated) time", -1.0}; config::Flag cfg_verbose_exit{"debug/verbose-exit", "Display the actor status at exit", true}; -xbt_dynar_t get_actors_addr() -{ -#if SIMGRID_HAVE_MC - return EngineImpl::get_instance()->get_actors_vector(); -#else - xbt_die("This function is intended to be used when compiling with MC"); -#endif -} - constexpr std::initializer_list> context_factories = { #if HAVE_RAW_CONTEXTS {"raw", &context::raw_factory}, @@ -191,7 +185,15 @@ EngineImpl::~EngineImpl() for (auto const& kv : mailboxes_) delete kv.second; - /* Free the remaining data structures */ + /* Kill all actors (but maestro) */ + maestro_->kill_all(); + run_all_actors(); + empty_trash(); + + delete maestro_; + delete context_factory_; + + /* Free the remaining data structures */ #if SIMGRID_HAVE_MC xbt_dynar_free(&actors_vector_); #endif @@ -206,9 +208,9 @@ void EngineImpl::initialize(int* argc, char** argv) EngineImpl::instance_ = this; #if SIMGRID_HAVE_MC // The communication initialization is done ASAP, as we need to get some init parameters from the MC for different - // layers. But simix_global needs to be created, as we send the address of some of its fields to the MC that wants to + // layers. But instance_ needs to be created, as we send the address of some of its fields to the MC that wants to // read them directly. - simgrid::mc::AppSide::initialize(); + simgrid::mc::AppSide::initialize(actors_vector_); #endif if (xbt_initialized == 0) { @@ -297,7 +299,7 @@ void EngineImpl::shutdown() XBT_DEBUG("EngineImpl::shutdown() called. Simulation's over."); #if HAVE_SMPI if (not instance_->actor_list_.empty()) { - if (smpi_process() && smpi_process()->initialized()) { + if (smpi_process()->initialized()) { xbt_die("Process exited without calling MPI_Finalize - Killing simulation"); } else { XBT_WARN("Process called exit when leaving - Skipping cleanups"); @@ -313,18 +315,6 @@ void EngineImpl::shutdown() xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code."); } - /* Kill all actors (but maestro) */ - instance_->maestro_->kill_all(); - instance_->run_all_actors(); - instance_->empty_trash(); - - /* Let's free maestro now */ - delete instance_->maestro_; - instance_->maestro_ = nullptr; - - /* Finish context module and SURF */ - instance_->destroy_context_factory(); - while (not timer::kernel_timers().empty()) { delete timer::kernel_timers().top().second; timer::kernel_timers().pop(); @@ -339,6 +329,12 @@ void EngineImpl::shutdown() void EngineImpl::seal_platform() const { + /* Seal only once */ + static bool sealed = false; + if (sealed) + return; + sealed = true; + /* sealing resources before run: links */ for (auto const& kv : links_) kv.second->get_iface()->seal(); @@ -409,15 +405,14 @@ void EngineImpl::add_split_duplex_link(const std::string& name, std::unique_ptr< } /** Wake up all actors waiting for a Surf action to finish */ -void EngineImpl::wake_all_waiting_actors() const +void EngineImpl::handle_ended_actions() 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() != nullptr) { // Skip vcpu actions + // Action failures are not automatically reported when the action is started by maestro (as in SimDAG) if (action->get_activity()->get_actor() == maestro_) action->get_activity()->get_iface()->complete(s4u::Activity::State::FAILED); @@ -427,11 +422,8 @@ void EngineImpl::wake_all_waiting_actors() const 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! + if (action->get_activity() != nullptr) { + // Action termination are not automatically reported when the action is started by maestro (as in SimDAG) action->get_activity()->set_finish_time(action->get_finish_time()); if (action->get_activity()->get_actor() == maestro_) @@ -451,7 +443,11 @@ void EngineImpl::wake_all_waiting_actors() const */ void EngineImpl::run_all_actors() { - instance_->get_context_factory()->run_all(); + instance_->get_context_factory()->run_all(actors_to_run_); + + for (auto const& actor : actors_to_run_) + if (actor->to_be_freed()) + actor->cleanup_from_kernel(); actors_to_run_.swap(actors_that_ran_); actors_to_run_.clear(); @@ -463,32 +459,7 @@ actor::ActorImpl* EngineImpl::get_actor_by_pid(aid_t 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> 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; + return nullptr; // Not found } void EngineImpl::remove_daemon(actor::ActorImpl* actor) @@ -555,7 +526,7 @@ void EngineImpl::display_all_actor_status() const 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_)); + actor->simcall_.get_cname()); } } } @@ -687,6 +658,15 @@ void EngineImpl::run(double max_date) { seal_platform(); + if (MC_is_active()) { +#if SIMGRID_HAVE_MC + mc::AppSide::get()->main_loop(); +#else + xbt_die("MC_is_active() is not supposed to return true in non-MC settings"); +#endif + THROW_IMPOSSIBLE; // main_loop never returns + } + if (MC_record_replay_is_active()) { mc::RecordTrace::replay(MC_record_path()); empty_trash(); @@ -701,7 +681,6 @@ void EngineImpl::run(double max_date) 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 @@ -709,47 +688,22 @@ void EngineImpl::run(double max_date) #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: - * - * Only maestro adds stuff to the actors_to_run array, so the execution order of user contexts do not impact its order. - * - * In addition, actors remain sorted through an arbitrary but fixed order in all cases: - * - * - If there is no killing during the simulation, actors remain sorted according by their PID. - * - Killer actors are moved to the end of the scheduling round (to let victims finish their simcall before dying), but - * (1) this decision of killing is reproducible because the simulation was reproducible until then - * (2) this reordering introduces no reproducibility hazard in the subsequent simulation. - * Even the order change induced by the actor killing is perfectly reproducible. - * - * So the array order is implicit and somewhat complex, but fixed and reproducible (science works, http://xkcd.com/54/). - * - * We could manually sort the actors_that_ran array so that simcalls are handled in an easy to predict order - * (e.g. "according to the PID of issuer"), but it's not mandatory for the simulation soundness and reproducibility, - * and would thus be a pure waste of time. + /* answer sequentially and in a fixed arbitrary order all the simcalls that were issued during that sub-round. + * The order must be fixed for the simulation to be reproducible (see RR-7653). It's OK here because only maestro + * changes the list. Killer actors are moved to the end to let victims finish their simcall before dying, but + * the order remains reproducible (even if arbitrarily). No need to sort the vector for sake of reproducibility. */ - - for (auto const& actor : actors_that_ran_) { - if (actor->simcall_.call_ != simix::Simcall::NONE) { + for (auto const& actor : actors_that_ran_) + if (actor->simcall_.call_ != actor::Simcall::Type::NONE) actor->simcall_handle(0); - } - } - execute_tasks(); - do { - wake_all_waiting_actors(); - } while (execute_tasks()); + handle_ended_actions(); /* If only daemon actors remain, cancel their actions, mark them to die and reschedule them */ if (actor_list_.size() == daemons_.size()) @@ -772,17 +726,11 @@ void EngineImpl::run(double max_date) 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: + // Execute timers 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(); + handle_ended_actions(); } while (again); /* Clean actors to destroy */