* under the terms of the license (GNU LGPL) which comes with this package. */
#include "src/mc/api/RemoteApp.hpp"
-#include "src/internal_config.h" // HAVE_SMPI
#include "src/mc/explo/Exploration.hpp"
#include "src/mc/mc_config.hpp"
#include "xbt/asserts.h"
-#if HAVE_SMPI
-#include "smpi/smpi.h"
-#include "src/smpi/include/private.hpp"
-#endif
-#include "signal.h"
#include "src/mc/api/State.hpp"
#include "src/mc/mc_config.hpp"
#include "src/mc/mc_exit.hpp"
#include "src/mc/mc_private.hpp"
#include "xbt/log.h"
#include "xbt/system_error.hpp"
+#include <signal.h>
+#include <algorithm>
#include <array>
#include <boost/tokenizer.hpp>
#include <memory>
+#include <numeric>
#include <string>
#include <fcntl.h>
namespace simgrid::mc {
-static void run_child_process(int socket, const std::vector<char*>& args)
+XBT_ATTRIB_NORETURN static void run_child_process(int socket, const std::vector<char*>& args)
{
/* On startup, simix_global_init() calls simgrid::mc::Client::initialize(), which checks whether the MC_ENV_SOCKET_FD
* env variable is set. If so, MC mode is assumed, and the client is setup from its side
RemoteApp::RemoteApp(const std::vector<char*>& args)
{
-#if HAVE_SMPI
- smpi_init_options(); // only performed once
- xbt_assert(smpi_cfg_privatization() != SmpiPrivStrategies::MMAP,
- "Please use the dlopen privatization schema when model-checking SMPI code");
-#endif
-
// Create an AF_LOCAL socketpair used for exchanging messages
// between the model-checker process (ourselves) and the model-checked
// process:
xbt_assert(mc_model_checker == nullptr, "Did you manage to start the MC twice in this process?");
- auto process = std::make_unique<simgrid::mc::RemoteProcess>(pid);
+ auto process = std::make_unique<simgrid::mc::RemoteProcessMemory>(pid);
model_checker_ = std::make_unique<simgrid::mc::ModelChecker>(std::move(process), sockets[1]);
mc_model_checker = model_checker_.get();
model_checker_->start();
/* Take the initial snapshot */
- model_checker_->wait_for_requests();
- initial_snapshot_ = std::make_shared<simgrid::mc::Snapshot>(0);
+ wait_for_requests();
+ initial_snapshot_ =
+ std::make_shared<simgrid::mc::Snapshot>(0, page_store_, model_checker_->get_remote_process_memory());
}
RemoteApp::~RemoteApp()
{
initial_snapshot_ = nullptr;
if (model_checker_) {
- model_checker_->shutdown();
+ shutdown();
model_checker_ = nullptr;
mc_model_checker = nullptr;
}
void RemoteApp::restore_initial_state() const
{
- this->initial_snapshot_->restore(&model_checker_->get_remote_process());
+ this->initial_snapshot_->restore(model_checker_->get_remote_process_memory());
}
unsigned long RemoteApp::get_maxpid() const
{
- return model_checker_->get_remote_process().get_maxpid();
+ // note: we could maybe cache it and count the actor creation on checker side too.
+ // But counting correctly accross state checkpoint/restore would be annoying.
+
+ model_checker_->get_channel().send(MessageType::ACTORS_MAXPID);
+ s_mc_message_int_t answer;
+ ssize_t answer_size = model_checker_->get_channel().receive(answer);
+ xbt_assert(answer_size != -1, "Could not receive message");
+ xbt_assert(answer_size == sizeof answer, "Broken message (size=%zd; expected %zu)", answer_size, sizeof answer);
+ xbt_assert(answer.type == MessageType::ACTORS_MAXPID_REPLY,
+ "Received unexpected message %s (%i); expected MessageType::ACTORS_MAXPID_REPLY (%i)",
+ to_c_str(answer.type), (int)answer.type, (int)MessageType::ACTORS_MAXPID_REPLY);
+
+ return answer.value;
}
void RemoteApp::get_actors_status(std::map<aid_t, ActorState>& whereto) const
{
- s_mc_message_t msg;
- memset(&msg, 0, sizeof msg);
- msg.type = simgrid::mc::MessageType::ACTORS_STATUS;
- model_checker_->channel().send(msg);
+ // The messaging happens as follows:
+ //
+ // CheckerSide AppSide
+ // send ACTORS_STATUS ---->
+ // <----- send ACTORS_STATUS_REPLY
+ // <----- send `N` `s_mc_message_actors_status_one_t` structs
+ // <----- send `M` `s_mc_message_simcall_probe_one_t` structs
+ model_checker_->get_channel().send(MessageType::ACTORS_STATUS);
s_mc_message_actors_status_answer_t answer;
- ssize_t received = model_checker_->channel().receive(answer);
- xbt_assert(received != -1, "Could not receive message");
- xbt_assert(received == sizeof(answer) && answer.type == MessageType::ACTORS_STATUS_REPLY,
- "Received unexpected message %s (%i, size=%i) "
- "expected MessageType::ACTORS_STATUS_REPLY (%i, size=%i)",
- to_c_str(answer.type), (int)answer.type, (int)received, (int)MessageType::ACTORS_STATUS_REPLY,
- (int)sizeof(answer));
+ ssize_t answer_size = model_checker_->get_channel().receive(answer);
+ xbt_assert(answer_size != -1, "Could not receive message");
+ xbt_assert(answer_size == sizeof answer, "Broken message (size=%zd; expected %zu)", answer_size, sizeof answer);
+ xbt_assert(answer.type == MessageType::ACTORS_STATUS_REPLY,
+ "Received unexpected message %s (%i); expected MessageType::ACTORS_STATUS_REPLY (%i)",
+ to_c_str(answer.type), (int)answer.type, (int)MessageType::ACTORS_STATUS_REPLY);
+
+ // Message sanity checks
+ xbt_assert(answer.count >= 0, "Received an ACTOR_STATUS_REPLY message with an actor count of '%d' < 0", answer.count);
+ xbt_assert(answer.transition_count >= 0, "Received an ACTOR_STATUS_REPLY message with transition_count '%d' < 0",
+ answer.transition_count);
+ xbt_assert(answer.transition_count == 0 || answer.count >= 0,
+ "Received an ACTOR_STATUS_REPLY message with no actor data "
+ "but with transition data nonetheless");
std::vector<s_mc_message_actors_status_one_t> status(answer.count);
if (answer.count > 0) {
- size_t size = status.size() * sizeof(s_mc_message_actors_status_one_t);
- received = model_checker_->channel().receive(status.data(), size);
+ size_t size = status.size() * sizeof(s_mc_message_actors_status_one_t);
+ ssize_t received = model_checker_->get_channel().receive(status.data(), size);
xbt_assert(static_cast<size_t>(received) == size);
}
+ // Ensures that each actor sends precisely `answer.transition_count` transitions. While technically
+ // this doesn't catch the edge case where actor A sends 3 instead of 2 and actor B sends 2 instead
+ // of 3 transitions, that is ignored here since that invariant needs to be enforced on the AppSide
+ const auto expected_transitions = std::accumulate(
+ status.begin(), status.end(), 0, [](int total, const auto& actor) { return total + actor.n_transitions; });
+ xbt_assert(expected_transitions == answer.transition_count,
+ "Expected to receive %d transition(s) but was only notified of %d by the app side", expected_transitions,
+ answer.transition_count);
+
+ std::vector<s_mc_message_simcall_probe_one_t> probes(answer.transition_count);
+ if (answer.transition_count > 0) {
+ for (auto& probe : probes) {
+ ssize_t received = model_checker_->get_channel().receive(probe);
+ xbt_assert(received >= 0, "Could not receive response to ACTORS_PROBE message (%s)", strerror(errno));
+ xbt_assert(static_cast<size_t>(received) == sizeof probe,
+ "Could not receive response to ACTORS_PROBE message (%zd bytes received != %zu bytes expected",
+ received, sizeof probe);
+ }
+ }
+
whereto.clear();
- for (auto const& actor : status)
- whereto.try_emplace(actor.aid, actor.aid, actor.enabled, actor.max_considered);
+ std::move_iterator probes_iter(probes.begin());
+
+ for (const auto& actor : status) {
+ xbt_assert(actor.n_transitions == 0 || actor.n_transitions == actor.max_considered,
+ "If any transitions are serialized for an actor, it must match the "
+ "total number of transitions that can be considered for the actor "
+ "(currently %d), but only %d transition(s) was/were said to be encoded",
+ actor.max_considered, actor.n_transitions);
+
+ std::vector<std::shared_ptr<Transition>> actor_transitions;
+ for (int times_considered = 0; times_considered < actor.n_transitions; times_considered++, probes_iter++) {
+ std::stringstream stream((*probes_iter).buffer.data());
+ actor_transitions.emplace_back(deserialize_transition(actor.aid, times_considered, stream));
+ }
+
+ XBT_DEBUG("Received %zu transitions for actor %ld", actor_transitions.size(), actor.aid);
+ whereto.try_emplace(actor.aid, actor.aid, actor.enabled, actor.max_considered, std::move(actor_transitions));
+ }
}
void RemoteApp::check_deadlock() const
{
- xbt_assert(model_checker_->channel().send(MessageType::DEADLOCK_CHECK) == 0, "Could not check deadlock state");
+ xbt_assert(model_checker_->get_channel().send(MessageType::DEADLOCK_CHECK) == 0, "Could not check deadlock state");
s_mc_message_int_t message;
- ssize_t s = model_checker_->channel().receive(message);
- xbt_assert(s != -1, "Could not receive message");
- xbt_assert(s == sizeof(message) && message.type == MessageType::DEADLOCK_CHECK_REPLY,
- "Received unexpected message %s (%i, size=%i) "
- "expected MessageType::DEADLOCK_CHECK_REPLY (%i, size=%i)",
- to_c_str(message.type), (int)message.type, (int)s, (int)MessageType::DEADLOCK_CHECK_REPLY,
- (int)sizeof(message));
+ ssize_t received = model_checker_->get_channel().receive(message);
+ xbt_assert(received != -1, "Could not receive message");
+ xbt_assert(received == sizeof message, "Broken message (size=%zd; expected %zu)", received, sizeof message);
+ xbt_assert(message.type == MessageType::DEADLOCK_CHECK_REPLY,
+ "Received unexpected message %s (%i); expected MessageType::DEADLOCK_CHECK_REPLY (%i)",
+ to_c_str(message.type), (int)message.type, (int)MessageType::DEADLOCK_CHECK_REPLY);
if (message.value != 0) {
XBT_CINFO(mc_global, "Counter-example execution trace:");
throw DeadlockError();
}
}
+
+void RemoteApp::wait_for_requests()
+{
+ /* Resume the application */
+ if (model_checker_->get_channel().send(MessageType::CONTINUE) != 0)
+ throw xbt::errno_error();
+ get_remote_process_memory().clear_cache();
+
+ if (this->get_remote_process_memory().running())
+ model_checker_->channel_handle_events();
+}
+
+void RemoteApp::shutdown()
+{
+ XBT_DEBUG("Shutting down model-checker");
+
+ RemoteProcessMemory& process = get_remote_process_memory();
+ if (process.running()) {
+ XBT_DEBUG("Killing process");
+ finalize_app(true);
+ kill(process.pid(), SIGKILL);
+ process.terminate();
+ }
+}
+
+Transition* RemoteApp::handle_simcall(aid_t aid, int times_considered, bool new_transition)
+{
+ s_mc_message_simcall_execute_t m = {};
+ m.type = MessageType::SIMCALL_EXECUTE;
+ m.aid_ = aid;
+ m.times_considered_ = times_considered;
+ model_checker_->get_channel().send(m);
+
+ get_remote_process_memory().clear_cache();
+ if (this->get_remote_process_memory().running())
+ model_checker_->channel_handle_events(); // The app may send messages while processing the transition
+
+ s_mc_message_simcall_execute_answer_t answer;
+ ssize_t s = model_checker_->get_channel().receive(answer);
+ xbt_assert(s != -1, "Could not receive message");
+ xbt_assert(s == sizeof answer, "Broken message (size=%zd; expected %zu)", s, sizeof answer);
+ xbt_assert(answer.type == MessageType::SIMCALL_EXECUTE_ANSWER,
+ "Received unexpected message %s (%i); expected MessageType::SIMCALL_EXECUTE_ANSWER (%i)",
+ to_c_str(answer.type), (int)answer.type, (int)MessageType::SIMCALL_EXECUTE_ANSWER);
+
+ if (new_transition) {
+ std::stringstream stream(answer.buffer.data());
+ return deserialize_transition(aid, times_considered, stream);
+ } else
+ return nullptr;
+}
+
+void RemoteApp::finalize_app(bool terminate_asap)
+{
+ s_mc_message_int_t m = {};
+ m.type = MessageType::FINALIZE;
+ m.value = terminate_asap;
+ xbt_assert(model_checker_->get_channel().send(m) == 0, "Could not ask the app to finalize on need");
+
+ s_mc_message_t answer;
+ ssize_t s = model_checker_->get_channel().receive(answer);
+ xbt_assert(s != -1, "Could not receive answer to FINALIZE");
+ xbt_assert(s == sizeof answer, "Broken message (size=%zd; expected %zu)", s, sizeof answer);
+ xbt_assert(answer.type == MessageType::FINALIZE_REPLY,
+ "Received unexpected message %s (%i); expected MessageType::FINALIZE_REPLY (%i)", to_c_str(answer.type),
+ (int)answer.type, (int)MessageType::FINALIZE_REPLY);
+}
+
} // namespace simgrid::mc
