ladybird/Userland/Services/SystemServer/Service.cpp
Sam Atkins 8a974ca91a SystemServer: Correct logic for services exiting successfully
WIFEXITED() returns a bool, so previously we were setting
exited_successfully to true when the service was terminated by a signal,
and false if it exited, regardless of the exit status. To test the exit
status, we have to use WEXITSTATUS() instead.

This causes us to correctly use the "3 tries then give up" logic for
services that crash, instead of infinitely attempting to respawn them.
2024-02-02 16:08:23 +00:00

387 lines
13 KiB
C++

/*
* Copyright (c) 2019-2020, Sergey Bugaev <bugaevc@serenityos.org>
*
* SPDX-License-Identifier: BSD-2-Clause
*/
#include "Service.h"
#include <AK/Debug.h>
#include <AK/HashMap.h>
#include <AK/JsonArray.h>
#include <AK/JsonObject.h>
#include <AK/String.h>
#include <AK/StringBuilder.h>
#include <LibCore/ConfigFile.h>
#include <LibCore/Directory.h>
#include <LibCore/SessionManagement.h>
#include <LibCore/SocketAddress.h>
#include <LibCore/System.h>
#include <LibFileSystem/FileSystem.h>
#include <fcntl.h>
#include <sched.h>
#include <stdio.h>
#include <sys/ioctl.h>
#include <unistd.h>
static HashMap<pid_t, Service*> s_service_map;
Service* Service::find_by_pid(pid_t pid)
{
auto it = s_service_map.find(pid);
if (it == s_service_map.end())
return nullptr;
return (*it).value;
}
ErrorOr<void> Service::setup_socket(SocketDescriptor& socket)
{
VERIFY(socket.fd == -1);
// Note: The purpose of this syscall is to remove potential left-over of previous portal.
// The return value is discarded as sockets are not always there, and unlinking a non-existent path is considered as a failure.
(void)Core::System::unlink(socket.path);
TRY(Core::Directory::create(LexicalPath(socket.path).parent(), Core::Directory::CreateDirectories::Yes));
// Note: we use SOCK_CLOEXEC here to make sure we don't leak every socket to
// all the clients. We'll make the one we do need to pass down !CLOEXEC later
// after forking off the process.
int const socket_fd = TRY(Core::System::socket(AF_LOCAL, SOCK_STREAM | SOCK_NONBLOCK | SOCK_CLOEXEC, 0));
socket.fd = socket_fd;
if (m_account.has_value()) {
auto& account = m_account.value();
TRY(Core::System::fchown(socket_fd, account.uid(), account.gid()));
}
TRY(Core::System::fchmod(socket_fd, socket.permissions));
auto socket_address = Core::SocketAddress::local(socket.path);
auto un_optional = socket_address.to_sockaddr_un();
if (!un_optional.has_value()) {
dbgln("Socket name {} is too long. BUG! This should have failed earlier!", socket.path);
VERIFY_NOT_REACHED();
}
auto un = un_optional.value();
TRY(Core::System::bind(socket_fd, (sockaddr const*)&un, sizeof(un)));
TRY(Core::System::listen(socket_fd, 16));
return {};
}
ErrorOr<void> Service::setup_sockets()
{
for (SocketDescriptor& socket : m_sockets)
TRY(setup_socket(socket));
return {};
}
void Service::setup_notifier()
{
VERIFY(m_lazy);
VERIFY(m_sockets.size() == 1);
VERIFY(!m_socket_notifier);
m_socket_notifier = Core::Notifier::construct(m_sockets[0].fd, Core::Notifier::Type::Read, this);
m_socket_notifier->on_activation = [this] {
if (auto result = handle_socket_connection(); result.is_error())
dbgln("{}", result.release_error());
};
}
ErrorOr<void> Service::handle_socket_connection()
{
VERIFY(m_sockets.size() == 1);
dbgln_if(SERVICE_DEBUG, "Ready to read on behalf of {}", name());
int socket_fd = m_sockets[0].fd;
if (m_accept_socket_connections) {
auto const accepted_fd = TRY(Core::System::accept(socket_fd, nullptr, nullptr));
TRY(determine_account(accepted_fd));
TRY(spawn(accepted_fd));
TRY(Core::System::close(accepted_fd));
} else {
remove_child(*m_socket_notifier);
m_socket_notifier = nullptr;
TRY(spawn(socket_fd));
}
return {};
}
ErrorOr<void> Service::activate()
{
VERIFY(m_pid < 0);
if (m_lazy)
setup_notifier();
else
TRY(spawn());
return {};
}
ErrorOr<void> Service::change_privileges()
{
// NOTE: Dropping privileges makes sense when SystemServer is running
// for a root session.
// This could happen when we need to spawn a Service to serve a client with non-user UID/GID.
// However, in case the user explicitly specified a username via the User= option, then we must
// try to login as at that user, so we can't ignore the failure when it was requested to change
// privileges.
if (auto current_uid = getuid(); m_account.has_value() && m_account.value().uid() != current_uid) {
if (current_uid != 0 && !m_must_login)
return {};
auto& account = m_account.value();
if (auto error_or_void = account.login(); error_or_void.is_error()) {
dbgln("Failed to drop privileges (tried to change to GID={}, UID={}), due to {}\n", account.gid(), account.uid(), error_or_void.error());
exit(1);
}
TRY(Core::System::setenv("HOME"sv, account.home_directory(), true));
}
return {};
}
ErrorOr<void> Service::spawn(int socket_fd)
{
if (!FileSystem::exists(m_executable_path)) {
dbgln("{}: binary \"{}\" does not exist, skipping service.", name(), m_executable_path);
return Error::from_errno(ENOENT);
}
dbgln_if(SERVICE_DEBUG, "Spawning {}", name());
m_run_timer.start();
pid_t pid = TRY(Core::System::fork());
if (pid == 0) {
// We are the child.
if (m_working_directory.has_value())
TRY(Core::System::chdir(*m_working_directory));
struct sched_param p;
p.sched_priority = m_priority;
int rc = sched_setparam(0, &p);
if (rc < 0) {
perror("sched_setparam");
VERIFY_NOT_REACHED();
}
if (m_stdio_file_path.has_value()) {
close(STDIN_FILENO);
auto const fd = TRY(Core::System::open(*m_stdio_file_path, O_RDWR, 0));
VERIFY(fd == 0);
dup2(STDIN_FILENO, STDOUT_FILENO);
dup2(STDIN_FILENO, STDERR_FILENO);
if (isatty(STDIN_FILENO)) {
ioctl(STDIN_FILENO, TIOCSCTTY);
}
} else {
if (isatty(STDIN_FILENO)) {
ioctl(STDIN_FILENO, TIOCNOTTY);
}
close(STDIN_FILENO);
close(STDOUT_FILENO);
close(STDERR_FILENO);
auto const fd = TRY(Core::System::open("/dev/null"sv, O_RDWR));
VERIFY(fd == STDIN_FILENO);
dup2(STDIN_FILENO, STDOUT_FILENO);
dup2(STDIN_FILENO, STDERR_FILENO);
}
StringBuilder socket_takeover_builder;
if (socket_fd >= 0) {
// We were spawned by socket activation. We currently only support
// single sockets for socket activation, so make sure that's the case.
VERIFY(m_sockets.size() == 1);
int fd = dup(socket_fd);
TRY(socket_takeover_builder.try_appendff("{}:{}", m_sockets[0].path, fd));
} else {
// We were spawned as a regular process, so dup every socket for this
// service and let the service know via SOCKET_TAKEOVER.
for (unsigned i = 0; i < m_sockets.size(); i++) {
SocketDescriptor& socket = m_sockets.at(i);
int new_fd = dup(socket.fd);
if (i != 0)
TRY(socket_takeover_builder.try_append(';'));
TRY(socket_takeover_builder.try_appendff("{}:{}", socket.path, new_fd));
}
}
if (!m_sockets.is_empty()) {
// The new descriptor is !CLOEXEC here.
TRY(Core::System::setenv("SOCKET_TAKEOVER"sv, socket_takeover_builder.string_view(), true));
}
TRY(change_privileges());
TRY(m_environment.view().for_each_split_view(' ', SplitBehavior::Nothing, [&](auto env) {
return Core::System::putenv(env);
}));
Vector<StringView, 10> arguments;
TRY(arguments.try_append(m_executable_path));
TRY(m_extra_arguments.view().for_each_split_view(' ', SplitBehavior::Nothing, [&](auto arg) {
return arguments.try_append(arg);
}));
TRY(Core::System::exec(m_executable_path, arguments, Core::System::SearchInPath::No));
} else if (!m_multi_instance) {
// We are the parent.
m_pid = pid;
s_service_map.set(pid, this);
}
return {};
}
ErrorOr<void> Service::did_exit(int status)
{
using namespace AK::TimeLiterals;
VERIFY(m_pid > 0);
VERIFY(!m_multi_instance);
if (WIFEXITED(status))
dbgln("Service {} has exited with exit code {}", name(), WEXITSTATUS(status));
if (WIFSIGNALED(status))
dbgln("Service {} terminated due to signal {}", name(), WTERMSIG(status));
s_service_map.remove(m_pid);
m_pid = -1;
if (!m_keep_alive)
return {};
auto run_time = m_run_timer.elapsed_time();
bool exited_successfully = WIFEXITED(status) && WEXITSTATUS(status) == 0;
if (!exited_successfully && run_time < 1_sec) {
switch (m_restart_attempts) {
case 0:
dbgln("Trying again");
break;
case 1:
dbgln("Third time's the charm?");
break;
default:
dbgln("Giving up on {}. Good luck!", name());
return {};
}
m_restart_attempts++;
}
TRY(activate());
return {};
}
Service::Service(Core::ConfigFile const& config, StringView name)
: Core::EventReceiver(nullptr)
{
VERIFY(config.has_group(name));
set_name(name);
m_executable_path = config.read_entry(name, "Executable", ByteString::formatted("/bin/{}", this->name()));
m_extra_arguments = config.read_entry(name, "Arguments");
m_stdio_file_path = config.read_entry_optional(name, "StdIO");
auto prio = config.read_entry_optional(name, "Priority");
if (prio == "low")
m_priority = 10;
else if (prio == "normal" || !prio.has_value())
m_priority = 30;
else if (prio == "high")
m_priority = 50;
else
VERIFY_NOT_REACHED();
m_keep_alive = config.read_bool_entry(name, "KeepAlive");
m_lazy = config.read_bool_entry(name, "Lazy");
m_user = config.read_entry_optional(name, "User");
if (m_user.has_value()) {
auto result = Core::Account::from_name(*m_user, Core::Account::Read::PasswdOnly);
if (result.is_error()) {
warnln("Failed to resolve user {}: {}", m_user, result.error());
} else {
m_must_login = true;
m_account = result.value();
}
}
m_working_directory = config.read_entry_optional(name, "WorkingDirectory");
m_environment = config.read_entry(name, "Environment");
m_system_modes = config.read_entry(name, "SystemModes", "graphical").split(',');
m_multi_instance = config.read_bool_entry(name, "MultiInstance");
m_accept_socket_connections = config.read_bool_entry(name, "AcceptSocketConnections");
ByteString socket_entry = config.read_entry(name, "Socket");
ByteString socket_permissions_entry = config.read_entry(name, "SocketPermissions", "0600");
if (!socket_entry.is_empty()) {
Vector<ByteString> socket_paths = socket_entry.split(',');
Vector<ByteString> socket_perms = socket_permissions_entry.split(',');
m_sockets.ensure_capacity(socket_paths.size());
// Need i here to iterate along with all other vectors.
for (unsigned i = 0; i < socket_paths.size(); i++) {
auto const path = Core::SessionManagement::parse_path_with_sid(socket_paths.at(i));
if (path.is_error()) {
// FIXME: better error handling for this case.
TODO();
}
// Socket path (plus NUL) must fit into the structs sent to the Kernel.
VERIFY(path.value().length() < UNIX_PATH_MAX);
// This is done so that the last permission repeats for every other
// socket. So you can define a single permission, and have it
// be applied for every socket.
mode_t permissions = strtol(socket_perms.at(min(socket_perms.size() - 1, (long unsigned)i)).characters(), nullptr, 8) & 0777;
m_sockets.empend(path.value(), -1, permissions);
}
}
// Lazy requires Socket, but only one.
VERIFY(!m_lazy || m_sockets.size() == 1);
// AcceptSocketConnections always requires Socket (single), Lazy, and MultiInstance.
VERIFY(!m_accept_socket_connections || (m_sockets.size() == 1 && m_lazy && m_multi_instance));
// MultiInstance doesn't work with KeepAlive.
VERIFY(!m_multi_instance || !m_keep_alive);
}
ErrorOr<NonnullRefPtr<Service>> Service::try_create(Core::ConfigFile const& config, StringView name)
{
return TRY(adopt_nonnull_ref_or_enomem(new (nothrow) Service(config, name)));
}
bool Service::is_enabled_for_system_mode(StringView mode) const
{
return m_system_modes.contains_slow(mode);
}
ErrorOr<void> Service::determine_account(int fd)
{
struct ucred creds = {};
socklen_t creds_size = sizeof(creds);
TRY(Core::System::getsockopt(fd, SOL_SOCKET, SO_PEERCRED, &creds, &creds_size));
m_account = TRY(Core::Account::from_uid(creds.uid, Core::Account::Read::PasswdOnly));
return {};
}
Service::~Service()
{
for (auto& socket : m_sockets) {
if (auto rc = remove(socket.path.characters()); rc != 0)
dbgln("{}", Error::from_errno(errno));
}
}