X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/5ce851fd1da99bbe6b7884e8064de96248ac6b58..6fbcdfade89b3812c24152c86f8aa3be510df7f2:/src/kernel/EngineImpl.cpp diff --git a/src/kernel/EngineImpl.cpp b/src/kernel/EngineImpl.cpp index 8f0b6ca278..2d8208f144 100644 --- a/src/kernel/EngineImpl.cpp +++ b/src/kernel/EngineImpl.cpp @@ -1,64 +1,75 @@ -/* Copyright (c) 2016-2021. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2016-2023. The SimGrid Team. All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ -#include "src/kernel/EngineImpl.hpp" +#include +#include +#include +#include +#include +#include + #include "mc/mc.h" -#include "simgrid/Exception.hpp" -#include "simgrid/kernel/Timer.hpp" -#include "simgrid/kernel/routing/NetPoint.hpp" -#include "simgrid/kernel/routing/NetZoneImpl.hpp" -#include "simgrid/s4u/Host.hpp" -#include "simgrid/sg_config.hpp" -#include "src/include/surf/surf.hpp" //get_clock() and surf_solve() -#include "src/kernel/resource/DiskImpl.hpp" +#include "src/kernel/EngineImpl.hpp" +#include "src/kernel/resource/StandardLinkImpl.hpp" #include "src/kernel/resource/profile/Profile.hpp" #include "src/mc/mc_record.hpp" #include "src/mc/mc_replay.hpp" #include "src/smpi/include/smpi_actor.hpp" -#include "src/surf/network_interface.hpp" -#include "src/surf/xml/platf.hpp" // FIXME: KILLME. There must be a better way than mimicking XML here -#include "xbt/xbt_modinter.h" /* whether initialization was already done */ +#include "src/surf/xml/platf.hpp" +#include "xbt/module.h" +#include "xbt/xbt_modinter.h" /* whether initialization was already done */ #include -#ifndef _WIN32 #include -#endif /* _WIN32 */ #if SIMGRID_HAVE_MC #include "src/mc/remote/AppSide.hpp" #endif XBT_LOG_NEW_DEFAULT_CATEGORY(ker_engine, "Logging specific to Engine (kernel)"); -namespace simgrid { -namespace kernel { + +namespace simgrid::kernel { +double EngineImpl::now_ = 0.0; EngineImpl* EngineImpl::instance_ = nullptr; /* That singleton is awful too. */ 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"); +constexpr std::initializer_list> context_factories = { +#if HAVE_RAW_CONTEXTS + {"raw", &context::raw_factory}, #endif -} +#if HAVE_UCONTEXT_CONTEXTS + {"ucontext", &context::sysv_factory}, +#endif +#if HAVE_BOOST_CONTEXTS + {"boost", &context::boost_factory}, +#endif + {"thread", &context::thread_factory}, +}; + +static_assert(context_factories.size() > 0, "No context factories are enabled for this build"); -xbt_dynar_t get_dead_actors_addr() +// Create the list of possible contexts: +static inline std::string contexts_list() { -#if SIMGRID_HAVE_MC - return EngineImpl::get_instance()->get_dead_actors_vector(); -#else - xbt_die("This function is intended to be used when compiling with MC"); -#endif + std::string res; + std::string sep = ""; + for (auto const& [factory_name, _] : context_factories) { + res += sep + factory_name; + sep = ", "; + } + return res; } -} // namespace kernel -} // namespace simgrid +static config::Flag context_factory_name("contexts/factory", + ("Possible values: " + contexts_list()).c_str(), + context_factories.begin()->first); + +} // namespace simgrid::kernel XBT_ATTRIB_NORETURN static void inthandler(int) { @@ -72,7 +83,6 @@ XBT_ATTRIB_NORETURN static void inthandler(int) exit(1); } -#ifndef _WIN32 static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/) { if ((siginfo->si_signo == SIGSEGV && siginfo->si_code == SEGV_ACCERR) || siginfo->si_signo == SIGBUS) { @@ -89,11 +99,11 @@ static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/) "If you think you've found a bug in SimGrid, please report it along with a\n" "Minimal Working Example (MWE) reproducing your problem and a full backtrace\n" "of the fault captured with gdb or valgrind.\n", - smx_context_stack_size / 1024); + simgrid::kernel::context::stack_size / 1024); } else if (siginfo->si_signo == SIGSEGV) { fprintf(stderr, "Segmentation fault.\n"); #if HAVE_SMPI - if (smpi_enabled() && smpi_cfg_privatization() == SmpiPrivStrategies::NONE) { + if (SMPI_is_inited() && smpi_cfg_privatization() == SmpiPrivStrategies::NONE) { #if HAVE_PRIVATIZATION fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n"); #else @@ -106,8 +116,8 @@ static void segvhandler(int signum, siginfo_t* siginfo, void* /*context*/) } /** - * Install signal handler for SIGSEGV. Check that nobody has already installed - * its own handler. For example, the Java VM does this. + * Install signal handler for SIGSEGV. Check that nobody has already installed its own handler. + * Historically, the Java VM did that but this could maybe removed now that Java is gone (TODO) */ static void install_segvhandler() { @@ -143,36 +153,26 @@ static void install_segvhandler() } } -#endif /* _WIN32 */ - -namespace simgrid { -namespace kernel { +namespace simgrid::kernel { EngineImpl::~EngineImpl() { - /* Since hosts_ is a std::map, the hosts are destroyed in the lexicographic order, which ensures that the output is - * reproducible. - */ - while (not hosts_.empty()) - hosts_.begin()->second->destroy(); - /* Also delete the other data */ delete netzone_root_; - for (auto const& kv : netpoints_) - delete kv.second; + for (auto const& [_, netpoint] : netpoints_) + delete netpoint; - for (auto const& kv : links_) - if (kv.second) - kv.second->destroy(); + for (auto const& [_, mailbox] : mailboxes_) + delete mailbox; - for (auto const& kv : mailboxes_) - delete kv.second; + /* 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_); - xbt_dynar_free(&dead_actors_vector_); -#endif /* clear models before freeing handle, network models can use external callback defined in the handle */ models_prio_.clear(); } @@ -184,7 +184,7 @@ 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(); #endif @@ -199,17 +199,65 @@ void EngineImpl::initialize(int* argc, char** argv) /* Prepare to display some more info when dying on Ctrl-C pressing */ std::signal(SIGINT, inthandler); - -#ifndef _WIN32 install_segvhandler(); -#endif /* register a function to be called by SURF after the environment creation */ sg_platf_init(); - s4u::Engine::on_platform_created.connect(surf_presolve); + s4u::Engine::on_platform_created_cb([this]() { this->presolve(); }); if (config::get_value("debug/clean-atexit")) - atexit(s4u::Engine::shutdown); + atexit(shutdown); +} + +void EngineImpl::context_mod_init() const +{ + xbt_assert(not instance_->has_context_factory()); + +#if HAVE_SMPI && defined(__NetBSD__) + smpi_init_options_internal(false); + std::string priv = config::get_value("smpi/privatization"); + if (context_factory_name == "thread" && (priv == "dlopen" || priv == "yes" || priv == "default" || priv == "1")) { + XBT_WARN("dlopen+thread broken on Apple and BSD. Switching to raw contexts."); + context_factory_name = "raw"; + } +#endif + +#if HAVE_SMPI && defined(__FreeBSD__) + smpi_init_options_internal(false); + if (context_factory_name == "thread" && config::get_value("smpi/privatization") != "no") { + XBT_WARN("mmap broken on FreeBSD, but dlopen+thread broken too. Switching to dlopen+raw contexts."); + context_factory_name = "raw"; + } +#endif + + /* use the factory specified by --cfg=contexts/factory:value */ + for (auto const& [factory_name, factory] : context_factories) + if (context_factory_name == factory_name) { + instance_->set_context_factory(factory()); + break; + } + + if (not instance_->has_context_factory()) { + XBT_ERROR("Invalid context factory specified. Valid factories on this machine:"); +#if HAVE_RAW_CONTEXTS + XBT_ERROR(" raw: high performance context factory implemented specifically for SimGrid"); +#else + XBT_ERROR(" (raw contexts were disabled at compilation time on this machine -- check configure logs for details)"); +#endif +#if HAVE_UCONTEXT_CONTEXTS + XBT_ERROR(" ucontext: classical system V contexts (implemented with makecontext, swapcontext and friends)"); +#else + XBT_ERROR(" (ucontext was disabled at compilation time on this machine -- check configure logs for details)"); +#endif +#if HAVE_BOOST_CONTEXTS + XBT_ERROR(" boost: this uses the boost libraries context implementation"); +#else + XBT_ERROR(" (boost was disabled at compilation time on this machine -- check configure logs for details. Did you " + "install the libboost-context-dev package?)"); +#endif + XBT_ERROR(" thread: slow portability layer using standard threads as provided by libstdc"); + xbt_die("Please use a valid factory."); + } } void EngineImpl::shutdown() @@ -217,16 +265,9 @@ void EngineImpl::shutdown() if (EngineImpl::instance_ == nullptr) return; XBT_DEBUG("EngineImpl::shutdown() called. Simulation's over."); - if (instance_->has_actors_to_run() && simgrid_get_clock() <= 0.0) { - XBT_CRITICAL(" "); - XBT_CRITICAL("The time is still 0, and you still have processes ready to run."); - XBT_CRITICAL("It seems that you forgot to run the simulation that you setup."); - xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code."); - } - #if HAVE_SMPI if (not instance_->actor_list_.empty()) { - if (smpi_process()->initialized()) { + if (smpi_process() && smpi_process()->initialized()) { xbt_die("Process exited without calling MPI_Finalize - Killing simulation"); } else { XBT_WARN("Process called exit when leaving - Skipping cleanups"); @@ -235,16 +276,12 @@ void EngineImpl::shutdown() } #endif - /* Kill all actors (but maestro) */ - instance_->maestro_->kill_all(); - instance_->run_all_actors(); - instance_->empty_trash(); - - /* Let's free maestro now */ - instance_->destroy_maestro(); - - /* Finish context module and SURF */ - instance_->destroy_context_factory(); + if (instance_->has_actors_to_run() && simgrid_get_clock() <= 0.0) { + XBT_CRITICAL(" "); + XBT_CRITICAL("The time is still 0, and you still have processes ready to run."); + XBT_CRITICAL("It seems that you forgot to run the simulation that you setup."); + xbt_die("Bailing out to avoid that stop-before-start madness. Please fix your code."); + } while (not timer::kernel_timers().empty()) { delete timer::kernel_timers().top().second; @@ -258,13 +295,22 @@ void EngineImpl::shutdown() instance_ = nullptr; } +void EngineImpl::seal_platform() const +{ + /* Seal only once */ + static bool sealed = false; + if (sealed) + return; + sealed = true; + + /* seal netzone root, recursively seal children netzones, hosts and disks */ + netzone_root_->seal(); +} + void EngineImpl::load_platform(const std::string& platf) { double start = xbt_os_time(); - if (boost::algorithm::ends_with(platf, ".so") or boost::algorithm::ends_with(platf, ".dylib")) { -#ifdef _WIN32 - xbt_die("loading platform through shared library isn't supported on windows"); -#else + if (boost::algorithm::ends_with(platf, ".so") || boost::algorithm::ends_with(platf, ".dylib")) { void* handle = dlopen(platf.c_str(), RTLD_LAZY); xbt_assert(handle, "Impossible to open platform file: %s", platf.c_str()); platf_handle_ = std::unique_ptr>(handle, dlclose); @@ -273,7 +319,6 @@ void EngineImpl::load_platform(const std::string& platf) const char* dlsym_error = dlerror(); xbt_assert(not dlsym_error, "Error: %s", dlsym_error); callable(*simgrid::s4u::Engine::get_instance()); -#endif /* _WIN32 */ } else { parse_platform_file(platf); } @@ -307,7 +352,7 @@ void EngineImpl::add_model(std::shared_ptr model, const std::ve 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) { + 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()); } @@ -315,28 +360,33 @@ void EngineImpl::add_model(std::shared_ptr model, const std::ve models_prio_[model_name] = std::move(model); } -void EngineImpl::add_split_duplex_link(const std::string& name, std::unique_ptr 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 +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 (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); + 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 (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_) + action->get_activity()->get_iface()->complete(s4u::Activity::State::FINISHED); + activity::ActivityImplPtr(action->get_activity())->post(); + } } } } @@ -349,7 +399,11 @@ 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(); @@ -358,35 +412,7 @@ void EngineImpl::run_all_actors() 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> 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 item == actor_list_.end() ? nullptr : item->second; } void EngineImpl::remove_daemon(actor::ActorImpl* actor) @@ -419,9 +445,6 @@ void EngineImpl::empty_trash() 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 @@ -429,9 +452,7 @@ 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 (@): \""); - for (auto const& kv : actor_list_) { - actor::ActorImpl* actor = kv.second; - + for (auto const& [_, actor] : actor_list_) { if (actor->waiting_synchro_) { const char* synchro_description = "unknown"; @@ -444,39 +465,180 @@ void EngineImpl::display_all_actor_status() const if (boost::dynamic_pointer_cast(actor->waiting_synchro_) != nullptr) synchro_description = "sleeping"; - if (boost::dynamic_pointer_cast(actor->waiting_synchro_) != nullptr) + if (boost::dynamic_pointer_cast(actor->waiting_synchro_) != nullptr) synchro_description = "synchronization"; if (boost::dynamic_pointer_cast(actor->waiting_synchro_) != nullptr) synchro_description = "I/O"; - XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %d to finish", actor->get_pid(), + XBT_INFO("Actor %ld (%s@%s): waiting for %s activity %#zx (%s) in state %s to finish %s", 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(), (int)actor->waiting_synchro_->state_); + actor->waiting_synchro_->get_cname(), actor->waiting_synchro_->get_state_str(), + (actor->simcall_.observer_ != nullptr && not xbt_log_no_loc + ? actor->simcall_.observer_->to_string().c_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_.observer_ != nullptr ? actor->simcall_.observer_->to_string().c_str() + : actor->simcall_.get_cname())); + } + } +} + +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(value<0) + continue; + 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_ += 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() +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::replay(MC_record_path()); + mc::RecordTrace::replay(MC_record_path()); empty_trash(); return; } - double time = 0; + double elapsed_time = -1; + const std::set* 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 && surf_get_clock() >= cfg_breakpoint) { + if (cfg_breakpoint >= 0.0 && simgrid_get_clock() >= cfg_breakpoint) { XBT_DEBUG("Breakpoint reached (%g)", cfg_breakpoint.get()); - cfg_breakpoint = -1.0; + cfg_breakpoint = -1.0; // Let the simulation continue without hiting the breakpoint again and again #ifdef SIGTRAP std::raise(SIGTRAP); #else @@ -484,86 +646,22 @@ void EngineImpl::run() #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 SIMIX_comm_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. + /* 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()) @@ -573,56 +671,58 @@ void EngineImpl::run() } } - time = timer::Timer::next(); - if (time > -1.0 || not actor_list_.empty()) { - XBT_DEBUG("Calling surf_solve"); - time = surf_solve(time); - XBT_DEBUG("Moving time ahead : %g", time); + // 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); } - /* 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 */ + 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); - // 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 */ empty_trash(); - XBT_DEBUG("### time %f, #actors %zu, #to_run %zu", time, actor_list_.size(), actors_to_run_.size()); + 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(vetoed_activities->size()))); - if (time < 0. && actors_to_run_.empty() && not actor_list_.empty()) { + 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."); + XBT_CRITICAL("Oops! Deadlock detected, some activities are still around but will never complete. " + "This usually happens when the user code is 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); + for (auto const& [_, actor] : actor_list_) { + XBT_DEBUG("Kill %s", actor->get_cname()); + maestro_->kill(actor); } } - } while (time > -1.0 || has_actors_to_run()); + } 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()) + if (not actor_list_.empty() && max_date < 0 && not(vetoed_activities == nullptr || vetoed_activities->empty())) THROW_IMPOSSIBLE; simgrid::s4u::Engine::on_simulation_end(); } -} // namespace kernel -} // namespace simgrid -void SIMIX_run() // XBT_ATTRIB_DEPRECATED_v332 +double EngineImpl::get_clock() { - simgrid::kernel::EngineImpl::get_instance()->run(); + return now_; } +} // namespace simgrid::kernel