/* 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 <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 "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/profile/Profile.hpp"
#include "src/mc/mc_record.hpp"
#include "src/mc/mc_replay.hpp"
-#include "src/simix/smx_private.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 "src/surf/xml/platf.hpp"
+#include "surf/surf.hpp" //surf_presolve() and surf_solve()
+#include "xbt/xbt_modinter.h" /* whether initialization was already done */
#include <boost/algorithm/string/predicate.hpp>
#ifndef _WIN32
#include <dlfcn.h>
#endif /* _WIN32 */
-XBT_LOG_NEW_DEFAULT_CATEGORY(ker_engine, "Logging specific to Engine (kernel)");
+#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 {
+EngineImpl* EngineImpl::instance_ = nullptr; /* That singleton is awful too. */
config::Flag<double> cfg_breakpoint{"debug/breakpoint",
"When non-negative, raise a SIGTRAP after given (simulated) time", -1.0};
-EngineImpl::~EngineImpl()
+config::Flag<bool> 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
+}
- while (not timer::kernel_timers().empty()) {
- delete timer::kernel_timers().top().second;
- timer::kernel_timers().pop();
+xbt_dynar_t get_dead_actors_addr()
+{
+#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
+}
+
+constexpr std::initializer_list<std::pair<const char*, context::ContextFactoryInitializer>> 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");
+
+// Create the list of possible contexts:
+static inline std::string contexts_list()
+{
+ std::string res;
+ std::string sep = "";
+ for (auto const& factory : context_factories) {
+ res += sep + factory.first;
+ sep = ", ";
+ }
+ return res;
+}
+
+static config::Flag<std::string> context_factory_name("contexts/factory",
+ (std::string("Possible values: ") + contexts_list()).c_str(),
+ context_factories.begin()->first);
+
+} // namespace kernel
+} // namespace simgrid
+
+XBT_ATTRIB_NORETURN static void inthandler(int)
+{
+ if (simgrid::kernel::cfg_verbose_exit) {
+ XBT_INFO("CTRL-C pressed. The current status will be displayed before exit (disable that behavior with option "
+ "'debug/verbose-exit').");
+ simgrid::kernel::EngineImpl::get_instance()->display_all_actor_status();
+ } else {
+ XBT_INFO("CTRL-C pressed, exiting. Hiding the current process status since 'debug/verbose-exit' is set to false.");
+ }
+ 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) {
+ fprintf(stderr,
+ "Access violation or Bus error detected.\n"
+ "This probably comes from a programming error in your code, or from a stack\n"
+ "overflow. If you are certain of your code, try increasing the stack size\n"
+ " --cfg=contexts/stack-size:XXX (current size is %u KiB).\n"
+ "\n"
+ "If it does not help, this may have one of the following causes:\n"
+ "a bug in SimGrid, a bug in the OS or a bug in a third-party libraries.\n"
+ "Failing hardware can sometimes generate such errors too.\n"
+ "\n"
+ "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",
+ 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 HAVE_PRIVATIZATION
+ fprintf(stderr, "Try to enable SMPI variable privatization with --cfg=smpi/privatization:yes.\n");
+#else
+ fprintf(stderr, "Sadly, your system does not support --cfg=smpi/privatization:yes (yet).\n");
+#endif /* HAVE_PRIVATIZATION */
+ }
+#endif /* HAVE_SMPI */
+ }
+ std::raise(signum);
+}
+
+/**
+ * Install signal handler for SIGSEGV. Check that nobody has already installed
+ * its own handler. For example, the Java VM does this.
+ */
+static void install_segvhandler()
+{
+ stack_t old_stack;
+
+ if (simgrid::kernel::context::Context::install_sigsegv_stack(&old_stack, true) == -1) {
+ XBT_WARN("Failed to register alternate signal stack: %s", strerror(errno));
+ return;
+ }
+ if (not(old_stack.ss_flags & SS_DISABLE)) {
+ XBT_DEBUG("An alternate stack was already installed (sp=%p, size=%zu, flags=%x). Restore it.", old_stack.ss_sp,
+ old_stack.ss_size, (unsigned)old_stack.ss_flags);
+ sigaltstack(&old_stack, nullptr);
+ }
+
+ struct sigaction action;
+ struct sigaction old_action;
+ action.sa_sigaction = &segvhandler;
+ action.sa_flags = SA_ONSTACK | SA_RESETHAND | SA_SIGINFO;
+ sigemptyset(&action.sa_mask);
+
+ /* Linux tend to raise only SIGSEGV where other systems also raise SIGBUS on severe error */
+ for (int sig : {SIGSEGV, SIGBUS}) {
+ if (sigaction(sig, &action, &old_action) == -1) {
+ XBT_WARN("Failed to register signal handler for signal %d: %s", sig, strerror(errno));
+ continue;
+ }
+ if ((old_action.sa_flags & SA_SIGINFO) || old_action.sa_handler != SIG_DFL) {
+ XBT_DEBUG("A signal handler was already installed for signal %d (%p). Restore it.", sig,
+ (old_action.sa_flags & SA_SIGINFO) ? (void*)old_action.sa_sigaction : (void*)old_action.sa_handler);
+ sigaction(sig, &old_action, nullptr);
+ }
}
+}
+
+#endif /* _WIN32 */
+namespace simgrid {
+namespace kernel {
+
+EngineImpl::~EngineImpl()
+{
/* Since hosts_ is a std::map, the hosts are destroyed in the lexicographic order, which ensures that the output is
* reproducible.
*/
for (auto const& kv : netpoints_)
delete kv.second;
- for (auto const& kv : links_)
- if (kv.second)
- kv.second->destroy();
+ while (not links_.empty())
+ links_.begin()->second->destroy();
for (auto const& kv : mailboxes_)
delete kv.second;
#endif
/* clear models before freeing handle, network models can use external callback defined in the handle */
models_prio_.clear();
+}
+
+void EngineImpl::initialize(int* argc, char** argv)
+{
+ xbt_assert(EngineImpl::instance_ == nullptr,
+ "It is currently forbidden to create more than one instance of kernel::EngineImpl");
+ 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
+ // read them directly.
+ simgrid::mc::AppSide::initialize();
+#endif
+
+ if (xbt_initialized == 0) {
+ xbt_init(argc, argv);
+
+ sg_config_init(argc, argv);
+ }
+
+ instance_->context_mod_init();
+
+ /* Prepare to display some more info when dying on Ctrl-C pressing */
+ std::signal(SIGINT, inthandler);
+
#ifndef _WIN32
- if (platf_handle_)
- dlclose(platf_handle_);
+ 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);
+
+ if (config::get_value<bool>("debug/clean-atexit"))
+ atexit(shutdown);
+}
+
+void EngineImpl::context_mod_init() const
+{
+ xbt_assert(not instance_->has_context_factory());
+
+#if HAVE_SMPI && (defined(__APPLE__) || defined(__NetBSD__))
+ smpi_init_options_internal(false);
+ std::string priv = config::get_value<std::string>("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<std::string>("smpi/privatization") != "no") {
+ XBT_WARN("mmap broken on FreeBSD, but dlopen+thread broken too. Switching to dlopen+raw contexts.");
+ context_factory_name = "raw";
+ }
+#endif
+
+ /* select the context factory to use to create the contexts */
+ if (context::factory_initializer != nullptr) { // Give Java a chance to hijack the factory mechanism
+ instance_->set_context_factory(context::factory_initializer());
+ return;
+ }
+ /* use the factory specified by --cfg=contexts/factory:value */
+ for (auto const& factory : context_factories)
+ if (context_factory_name == factory.first) {
+ instance_->set_context_factory(factory.second());
+ 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 pthreads as provided by gcc");
+ xbt_die("Please use a valid factory.");
+ }
+}
+
+void EngineImpl::shutdown()
+{
+ if (EngineImpl::instance_ == nullptr)
+ return;
+ XBT_DEBUG("EngineImpl::shutdown() called. Simulation's over.");
+#if HAVE_SMPI
+ if (not instance_->actor_list_.empty()) {
+ if (smpi_process()->initialized()) {
+ xbt_die("Process exited without calling MPI_Finalize - Killing simulation");
+ } else {
+ XBT_WARN("Process called exit when leaving - Skipping cleanups");
+ return;
+ }
+ }
+#endif
+
+ 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.");
+ }
+
+ /* 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();
+
+ while (not timer::kernel_timers().empty()) {
+ delete timer::kernel_timers().top().second;
+ timer::kernel_timers().pop();
+ }
+
+ tmgr_finalize();
+ sg_platf_exit();
+
+ delete instance_;
+ instance_ = nullptr;
}
void EngineImpl::load_platform(const std::string& platf)
#ifdef _WIN32
xbt_die("loading platform through shared library isn't supported on windows");
#else
- platf_handle_ = dlopen(platf.c_str(), RTLD_LAZY);
- xbt_assert(platf_handle_, "Impossible to open platform file: %s", platf.c_str());
+ 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<void, std::function<int(void*)>>(handle, dlclose);
using load_fct_t = void (*)(const simgrid::s4u::Engine&);
- dlerror();
- auto callable = (load_fct_t)dlsym(platf_handle_, "load_platform");
+ auto callable = (load_fct_t)dlsym(platf_handle_.get(), "load_platform");
const char* dlsym_error = dlerror();
xbt_assert(not dlsym_error, "Error: %s", dlsym_error);
callable(*simgrid::s4u::Engine::get_instance());
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
{
*/
void EngineImpl::run_all_actors()
{
- simix_global->get_context_factory()->run_all();
+ instance_->get_context_factory()->run_all();
actors_to_run_.swap(actors_that_ran_);
actors_to_run_.clear();
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()
{
- xbt_assert(tasksTemp.empty());
-
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);
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;
#ifdef SIGTRAP
* - 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().
+ * - 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).
if (actor_list_.size() == daemons_.size())
for (auto const& dmon : daemons_) {
XBT_DEBUG("Kill %s", dmon->get_cname());
- simix_global->get_maestro()->kill(dmon);
+ maestro_->kill(dmon);
}
}
simgrid::s4u::Engine::on_deadlock();
for (auto const& kv : actor_list_) {
XBT_DEBUG("Kill %s", kv.second->get_cname());
- simix_global->get_maestro()->kill(kv.second);
+ maestro_->kill(kv.second);
}
}
} while (time > -1.0 || has_actors_to_run());
}
} // namespace kernel
} // namespace simgrid
+
+void SIMIX_run() // XBT_ATTRIB_DEPRECATED_v332
+{
+ simgrid::kernel::EngineImpl::get_instance()->run();
+}