#include "src/mc/remote/CheckerSide.hpp"
#include "src/mc/explo/Exploration.hpp"
+#include "src/mc/mc_environ.h"
+#include "xbt/config.hpp"
+#include "xbt/system_error.hpp"
+
+#if SIMGRID_HAVE_STATEFUL_MC
#include "src/mc/explo/LivenessChecker.hpp"
#include "src/mc/sosp/RemoteProcessMemory.hpp"
-#include "xbt/system_error.hpp"
+#endif
#ifdef __linux__
-#include <sys/personality.h>
#include <sys/prctl.h>
#endif
namespace simgrid::mc {
-XBT_ATTRIB_NORETURN 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, bool need_ptrace)
{
/* 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
sigemptyset(&mask);
xbt_assert(sigprocmask(SIG_SETMASK, &mask, nullptr) >= 0, "Could not unblock signals");
xbt_assert(prctl(PR_SET_PDEATHSIG, SIGHUP) == 0, "Could not PR_SET_PDEATHSIG");
-
- // Make sure that the application process layout is not randomized, so that the info we gather is stable over re-execs
- if (personality(ADDR_NO_RANDOMIZE) == -1) {
- XBT_ERROR("Could not set the NO_RANDOMIZE personality");
- throw xbt::errno_error();
- }
#endif
- // Remove CLOEXEC to pass the socket to the application
- int fdflags = fcntl(socket, F_GETFD, 0);
- xbt_assert(fdflags != -1 && fcntl(socket, F_SETFD, fdflags & ~FD_CLOEXEC) != -1,
- "Could not remove CLOEXEC for socket");
-
setenv(MC_ENV_SOCKET_FD, std::to_string(socket).c_str(), 1);
+ if (need_ptrace)
+ setenv(MC_ENV_NEED_PTRACE, "1", 1);
/* Setup the tokenizer that parses the cfg:model-check/setenv parameter */
using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
#elif defined BSD
ptrace(PT_CONTINUE, pid, (caddr_t)1, 0);
#else
-#error "no ptrace equivalent coded for this platform"
+ xbt_die("no ptrace equivalent coded for this platform, stateful model-checking is impossible.");
#endif
xbt_assert(errno == 0,
"Ptrace does not seem to be usable in your setup (errno: %d). "
"If you run from within a docker, adding `--cap-add SYS_PTRACE` to the docker line may help. "
"If it does not help, please report this bug.",
errno);
+ XBT_DEBUG("%d ptrace correctly setup.", getpid());
}
-void CheckerSide::setup_events()
+void CheckerSide::setup_events(bool socket_only)
{
- if (base_ != nullptr)
- event_base_free(base_.get());
auto* base = event_base_new();
base_.reset(base);
[](evutil_socket_t, short events, void* arg) {
auto checker = static_cast<simgrid::mc::CheckerSide*>(arg);
if (events == EV_READ) {
- std::array<char, MC_MESSAGE_LENGTH> buffer;
- ssize_t size = recv(checker->get_channel().get_socket(), buffer.data(), buffer.size(), MSG_DONTWAIT);
- if (size == -1) {
- XBT_ERROR("Channel::receive failure: %s", strerror(errno));
- if (errno != EAGAIN)
- throw simgrid::xbt::errno_error();
- }
-
- if (not checker->handle_message(buffer.data(), size))
- checker->break_loop();
+ do {
+ std::array<char, MC_MESSAGE_LENGTH> buffer;
+ ssize_t size = checker->get_channel().receive(buffer.data(), buffer.size(), MSG_DONTWAIT);
+ if (size == -1) {
+ XBT_ERROR("Channel::receive failure: %s", strerror(errno));
+ if (errno != EAGAIN)
+ throw simgrid::xbt::errno_error();
+ }
+
+ if (size == 0) // The app closed the socket. It must be dead by now.
+ checker->handle_waitpid();
+ else if (not checker->handle_message(buffer.data(), size)) {
+ checker->break_loop();
+ break;
+ }
+ } while (checker->get_channel().has_pending_data());
} else {
xbt_die("Unexpected event");
}
this);
event_add(socket_event_, nullptr);
- signal_event_ = event_new(
- base, SIGCHLD, EV_SIGNAL | EV_PERSIST,
- [](evutil_socket_t sig, short events, void* arg) {
- auto checker = static_cast<simgrid::mc::CheckerSide*>(arg);
- if (events == EV_SIGNAL) {
- if (sig == SIGCHLD)
- checker->handle_waitpid();
- else
- xbt_die("Unexpected signal: %d", sig);
- } else {
- xbt_die("Unexpected event");
- }
- },
- this);
- event_add(signal_event_, nullptr);
+ if (socket_only) {
+ signal_event_ = nullptr;
+ } else {
+ signal_event_ = event_new(
+ base, SIGCHLD, EV_SIGNAL | EV_PERSIST,
+ [](evutil_socket_t sig, short events, void* arg) {
+ auto checker = static_cast<simgrid::mc::CheckerSide*>(arg);
+ if (events == EV_SIGNAL) {
+ if (sig == SIGCHLD)
+ checker->handle_waitpid();
+ else
+ xbt_die("Unexpected signal: %d", sig);
+ } else {
+ xbt_die("Unexpected event");
+ }
+ },
+ this);
+ event_add(signal_event_, nullptr);
+ }
}
-CheckerSide::CheckerSide(const std::vector<char*>& args, bool need_memory_introspection) : running_(true)
+/* When this constructor is called, no other checkerside exists */
+CheckerSide::CheckerSide(const std::vector<char*>& args, bool need_memory_info) : running_(true)
{
- // Create an AF_LOCAL socketpair used for exchanging messages between the model-checker process (ancestor)
+ XBT_DEBUG("Create a CheckerSide. Needs_meminfo: %s", need_memory_info ? "YES" : "no");
+
+ // Create an AF_UNIX socketpair used for exchanging messages between the model-checker process (ancestor)
// and the application process (child)
int sockets[2];
- xbt_assert(socketpair(AF_LOCAL, SOCK_SEQPACKET | SOCK_CLOEXEC, 0, sockets) != -1, "Could not create socketpair");
+ xbt_assert(socketpair(AF_UNIX,
+#ifdef __APPLE__
+ SOCK_STREAM, /* Mac OSX does not have AF_UNIX + SOCK_SEQPACKET, even if that's faster */
+#else
+ SOCK_SEQPACKET,
+#endif
+ 0, sockets) != -1,
+ "Could not create socketpair: %s", strerror(errno));
pid_ = fork();
- xbt_assert(pid_ >= 0, "Could not fork model-checked process");
+ xbt_assert(pid_ >= 0, "Could not fork application process");
if (pid_ == 0) { // Child
::close(sockets[1]);
- run_child_process(sockets[0], args);
+ run_child_process(sockets[0], args, need_memory_info); // We need ptrace if we need the mem info
DIE_IMPOSSIBLE;
}
::close(sockets[0]);
channel_.reset_socket(sockets[1]);
- setup_events();
- wait_application_process(pid_);
+ setup_events(false); /* we need a signal handler too */
+ if (need_memory_info) {
+#if SIMGRID_HAVE_STATEFUL_MC
+ // setup ptrace and sync with the app
+ wait_application_process(pid_);
- // Request the initial memory on need
- if (need_memory_introspection) {
- channel_.send(MessageType::INITIAL_ADDRESSES);
- s_mc_message_initial_addresses_reply_t answer;
+ // Request the initial memory on need
+ channel_.send(MessageType::NEED_MEMINFO);
+ s_mc_message_need_meminfo_reply_t answer;
ssize_t answer_size = channel_.receive(answer);
xbt_assert(answer_size != -1, "Could not receive message");
- xbt_assert(answer.type == MessageType::INITIAL_ADDRESSES_REPLY,
- "The received message is not the INITIAL_ADDRESS_REPLY I was expecting but of type %s",
+ xbt_assert(answer.type == MessageType::NEED_MEMINFO_REPLY,
+ "The received message is not the NEED_MEMINFO_REPLY I was expecting but of type %s",
to_c_str(answer.type));
xbt_assert(answer_size == sizeof answer, "Broken message (size=%zd; expected %zu)", answer_size, sizeof answer);
/* We now have enough info to create the memory address space */
remote_memory_ = std::make_unique<simgrid::mc::RemoteProcessMemory>(pid_, answer.mmalloc_default_mdp);
+#else
+ xbt_die("Cannot introspect memory without MC support");
+#endif
}
wait_for_requests();
{
event_del(socket_event_);
event_free(socket_event_);
- event_del(signal_event_);
- event_free(signal_event_);
-
- if (running()) {
- XBT_DEBUG("Killing process");
- kill(get_pid(), SIGKILL);
- while (waitpid(-1, nullptr, WNOHANG) > 0) {
- /* we don't really care about errors here, as we are shutting things down anyway */
- /* The child will get ripped by the next waitpid anyway */
- }
+ if (signal_event_ != nullptr) {
+ event_del(signal_event_);
+ event_free(signal_event_);
}
- // usleep(500); // Try to reduce the load on my system. Commented because it's not even enough :(
+}
+
+/* This constructor is called when cloning a checkerside to get its application to fork away */
+CheckerSide::CheckerSide(int socket, CheckerSide* child_checker)
+ : channel_(socket, child_checker->channel_), running_(true), child_checker_(child_checker)
+{
+ setup_events(true); // We already have a signal handled in that case
+
+ s_mc_message_int_t answer;
+ ssize_t s = get_channel().receive(answer);
+ xbt_assert(s != -1, "Could not receive answer to FORK_REPLY");
+ xbt_assert(s == sizeof answer, "Broken message (size=%zd; expected %zu)", s, sizeof answer);
+ xbt_assert(answer.type == MessageType::FORK_REPLY,
+ "Received unexpected message %s (%i); expected MessageType::FORK_REPLY (%i)", to_c_str(answer.type),
+ (int)answer.type, (int)MessageType::FORK_REPLY);
+ pid_ = answer.value;
+
+ wait_for_requests();
+}
+
+std::unique_ptr<CheckerSide> CheckerSide::clone(int master_socket, const std::string& master_socket_name)
+{
+ s_mc_message_fork_t m = {};
+ m.type = MessageType::FORK;
+ xbt_assert(master_socket_name.size() == MC_SOCKET_NAME_LEN);
+ std::copy_n(begin(master_socket_name), MC_SOCKET_NAME_LEN, begin(m.socket_name));
+ xbt_assert(get_channel().send(m) == 0, "Could not ask the app to fork on need.");
+
+ int sock = accept(master_socket, nullptr /* I know who's connecting*/, nullptr);
+ xbt_assert(sock > 0, "Cannot accept the incomming connection of the forked app: %s.", strerror(errno));
+
+ return std::make_unique<CheckerSide>(sock, this);
}
void CheckerSide::finalize(bool terminate_asap)
bool CheckerSide::handle_message(const char* buffer, ssize_t size)
{
s_mc_message_t base_message;
- xbt_assert(size >= (ssize_t)sizeof(base_message), "Broken message");
+ ssize_t consumed;
+ xbt_assert(size >= (ssize_t)sizeof(base_message), "Broken message. Got only %ld bytes.", size);
memcpy(&base_message, buffer, sizeof(base_message));
switch (base_message.type) {
case MessageType::IGNORE_HEAP: {
+ consumed = sizeof(s_mc_message_ignore_heap_t);
+#if SIMGRID_HAVE_STATEFUL_MC
if (remote_memory_ != nullptr) {
s_mc_message_ignore_heap_t message;
- xbt_assert(size == sizeof(message), "Broken message");
+ xbt_assert(size >= static_cast<ssize_t>(sizeof(message)), "Broken message");
memcpy(&message, buffer, sizeof(message));
IgnoredHeapRegion region;
region.address = message.address;
region.size = message.size;
get_remote_memory()->ignore_heap(region);
- } else {
+ } else
+#endif
XBT_INFO("Ignoring a IGNORE_HEAP message because we don't need to introspect memory.");
- }
break;
}
case MessageType::UNIGNORE_HEAP: {
+ consumed = sizeof(s_mc_message_ignore_memory_t);
+#if SIMGRID_HAVE_STATEFUL_MC
if (remote_memory_ != nullptr) {
s_mc_message_ignore_memory_t message;
- xbt_assert(size == sizeof(message), "Broken message");
+ xbt_assert(size == static_cast<ssize_t>(sizeof(message)), "Broken message");
memcpy(&message, buffer, sizeof(message));
get_remote_memory()->unignore_heap((void*)message.addr, message.size);
- } else {
+ } else
+#endif
XBT_INFO("Ignoring an UNIGNORE_HEAP message because we don't need to introspect memory.");
- }
break;
}
case MessageType::IGNORE_MEMORY: {
+ consumed = sizeof(s_mc_message_ignore_memory_t);
+#if SIMGRID_HAVE_STATEFUL_MC
if (remote_memory_ != nullptr) {
s_mc_message_ignore_memory_t message;
- xbt_assert(size == sizeof(message), "Broken message");
+ xbt_assert(size >= static_cast<ssize_t>(sizeof(message)), "Broken message");
memcpy(&message, buffer, sizeof(message));
get_remote_memory()->ignore_region(message.addr, message.size);
- } else {
+ } else
+#endif
XBT_INFO("Ignoring an IGNORE_MEMORY message because we don't need to introspect memory.");
- }
break;
}
case MessageType::STACK_REGION: {
+ consumed = sizeof(s_mc_message_stack_region_t);
+#if SIMGRID_HAVE_STATEFUL_MC
if (remote_memory_ != nullptr) {
s_mc_message_stack_region_t message;
- xbt_assert(size == sizeof(message), "Broken message");
+ xbt_assert(size >= static_cast<ssize_t>(sizeof(message)), "Broken message");
memcpy(&message, buffer, sizeof(message));
get_remote_memory()->stack_areas().push_back(message.stack_region);
- } else {
+ } else
+#endif
XBT_INFO("Ignoring an STACK_REGION message because we don't need to introspect memory.");
- }
break;
}
case MessageType::REGISTER_SYMBOL: {
+ consumed = sizeof(s_mc_message_register_symbol_t);
+#if SIMGRID_HAVE_STATEFUL_MC
s_mc_message_register_symbol_t message;
- xbt_assert(size == sizeof(message), "Broken message");
+ xbt_assert(size >= static_cast<ssize_t>(sizeof(message)), "Broken message");
memcpy(&message, buffer, sizeof(message));
xbt_assert(not message.callback, "Support for client-side function proposition is not implemented.");
XBT_DEBUG("Received symbol: %s", message.name.data());
LivenessChecker::automaton_register_symbol(*get_remote_memory(), message.name.data(), remote((int*)message.data));
+#else
+ xbt_die("Please don't use liveness properties when MC is compiled out.");
+#endif
break;
}
case MessageType::WAITING:
+ consumed = sizeof(s_mc_message_t);
+ if (size > consumed) {
+ XBT_DEBUG("%d reinject %d bytes after a %s message", getpid(), (int)(size - consumed),
+ to_c_str(base_message.type));
+ channel_.reinject(&buffer[consumed], size - consumed);
+ }
+
return false;
case MessageType::ASSERTION_FAILED:
+ // report_assertion_failure() is NORETURN, but it may change when we report more than one error per run,
+ // so please keep the consumed computation even if clang-static detects it as a dead affectation.
+ consumed = sizeof(s_mc_message_t);
Exploration::get_instance()->report_assertion_failure();
break;
default:
- xbt_die("Unexpected message from model-checked application");
+ xbt_die("Unexpected message from the application");
+ }
+ if (size > consumed) {
+ XBT_DEBUG("%d reinject %d bytes after a %s message", getpid(), (int)(size - consumed), to_c_str(base_message.type));
+ channel_.reinject(&buffer[consumed], size - consumed);
}
return true;
}
void CheckerSide::wait_for_requests()
{
- /* Resume the application */
+ XBT_DEBUG("Resume the application");
if (get_channel().send(MessageType::CONTINUE) != 0)
throw xbt::errno_error();
clear_memory_cache();
void CheckerSide::clear_memory_cache()
{
+#if SIMGRID_HAVE_STATEFUL_MC
if (remote_memory_)
remote_memory_->clear_cache();
+#endif
}
-void CheckerSide::handle_waitpid()
+void CheckerSide::handle_dead_child(int status)
{
- XBT_DEBUG("Check for wait event");
- int status;
- pid_t pid;
- while ((pid = waitpid(-1, &status, WNOHANG)) != 0) {
- if (pid == -1) {
- if (errno == ECHILD) { // No more children:
- xbt_assert(not this->running(), "Inconsistent state");
- break;
- } else {
- xbt_die("Could not wait for pid: %s", strerror(errno));
- }
- }
-
- if (pid == get_pid()) {
- // From PTRACE_O_TRACEEXIT:
+ // From PTRACE_O_TRACEEXIT:
#ifdef __linux__
- if (status >> 8 == (SIGTRAP | (PTRACE_EVENT_EXIT << 8))) {
- unsigned long eventmsg;
- xbt_assert(ptrace(PTRACE_GETEVENTMSG, pid, 0, &eventmsg) != -1, "Could not get exit status");
- status = static_cast<int>(eventmsg);
- if (WIFSIGNALED(status)) {
- this->terminate();
- Exploration::get_instance()->report_crash(status);
- }
- }
+ if (status >> 8 == (SIGTRAP | (PTRACE_EVENT_EXIT << 8))) {
+ unsigned long eventmsg;
+ xbt_assert(ptrace(PTRACE_GETEVENTMSG, pid_, 0, &eventmsg) != -1, "Could not get exit status");
+ status = static_cast<int>(eventmsg);
+ if (WIFSIGNALED(status)) {
+ this->terminate();
+ Exploration::get_instance()->report_crash(status);
+ }
+ }
#endif
- // We don't care about non-lethal signals, just reinject them:
- if (WIFSTOPPED(status)) {
- XBT_DEBUG("Stopped with signal %i", (int)WSTOPSIG(status));
- errno = 0;
+ // We don't care about non-lethal signals, just reinject them:
+ if (WIFSTOPPED(status)) {
+ XBT_DEBUG("Stopped with signal %i", (int)WSTOPSIG(status));
+ errno = 0;
#ifdef __linux__
- ptrace(PTRACE_CONT, pid, 0, WSTOPSIG(status));
+ ptrace(PTRACE_CONT, pid_, 0, WSTOPSIG(status));
#elif defined BSD
- ptrace(PT_CONTINUE, pid, (caddr_t)1, WSTOPSIG(status));
+ ptrace(PT_CONTINUE, pid_, (caddr_t)1, WSTOPSIG(status));
#endif
- xbt_assert(errno == 0, "Could not PTRACE_CONT");
- }
+ xbt_assert(errno == 0, "Could not PTRACE_CONT: %s", strerror(errno));
+ }
- else if (WIFSIGNALED(status)) {
- this->terminate();
- Exploration::get_instance()->report_crash(status);
- } else if (WIFEXITED(status)) {
- XBT_DEBUG("Child process is over");
- this->terminate();
+ else if (WIFSIGNALED(status)) {
+ this->terminate();
+ Exploration::get_instance()->report_crash(status);
+ } else if (WIFEXITED(status)) {
+ XBT_DEBUG("Child process is over");
+ this->terminate();
+ }
+}
+
+void CheckerSide::handle_waitpid()
+{
+ XBT_DEBUG("%d checks for wait event. %s", getpid(),
+ child_checker_ == nullptr ? "Wait directly." : "Ask our proxy to wait for its child.");
+
+ if (child_checker_ == nullptr) { // Wait directly
+ int status;
+ pid_t pid;
+ while ((pid = waitpid(-1, &status, WNOHANG)) != 0) {
+ if (pid == -1) {
+ if (errno == ECHILD) { // No more children:
+ xbt_assert(not this->running(), "Inconsistent state");
+ break;
+ } else {
+ xbt_die("Could not wait for pid: %s", strerror(errno));
+ }
}
+
+ if (pid == get_pid())
+ handle_dead_child(status);
}
+
+ } else { // Ask our proxy to wait for us
+ s_mc_message_int_t request = {};
+ request.type = MessageType::WAIT_CHILD;
+ request.value = pid_;
+ xbt_assert(child_checker_->get_channel().send(request) == 0,
+ "Could not ask my child to waitpid its child for me: %s", strerror(errno));
+
+ s_mc_message_int_t answer;
+ ssize_t answer_size = child_checker_->get_channel().receive(answer);
+ xbt_assert(answer_size != -1, "Could not receive message");
+ xbt_assert(answer.type == MessageType::WAIT_CHILD_REPLY,
+ "The received message is not the WAIT_CHILD_REPLY I was expecting but of type %s",
+ to_c_str(answer.type));
+ xbt_assert(answer_size == sizeof answer, "Broken message (size=%zd; expected %zu)", answer_size, sizeof answer);
+ handle_dead_child(answer.value);
}
}
-
} // namespace simgrid::mc
