286 行
6.8 KiB
C++
286 行
6.8 KiB
C++
#include <sys/prctl.h>
|
|
#include <seccomp.h>
|
|
#include <vector>
|
|
#include <initializer_list>
|
|
#include <string.h>
|
|
#include <sched.h>
|
|
#include <unistd.h>
|
|
#include <fstream>
|
|
#include <sys/types.h>
|
|
#include <sys/stat.h>
|
|
#include <fcntl.h>
|
|
#include <unistd.h>
|
|
#include <string.h>
|
|
#include <filesystem>
|
|
#include <sys/mount.h>
|
|
#include <sys/capability.h>
|
|
#include "../logger.h"
|
|
#include "../utils.h"
|
|
#include "../random.h"
|
|
|
|
#include "sandbox-linux.h"
|
|
|
|
/* TODO: make a whitelist approach. So far we simply blacklist
|
|
* obvious systemcalls. To whitelist, we need to analyse our
|
|
* dependencies (http library, sqlite wrapper, sqlite lib etc.) */
|
|
|
|
/* TODO: cleanup our sandboxing directory (unmount and delete folders) after exit */
|
|
bool SandboxLinux::seccomp_blacklist(std::initializer_list<int> syscalls)
|
|
{
|
|
scmp_filter_ctx ctx;
|
|
ctx = seccomp_init(SCMP_ACT_ALLOW);
|
|
if(ctx == NULL)
|
|
{
|
|
Logger::error() << "failed to init seccomp_init";
|
|
return false;
|
|
}
|
|
|
|
for(auto sc : syscalls)
|
|
{
|
|
if(seccomp_rule_add(ctx, SCMP_ACT_KILL_PROCESS, sc, 0) < 0)
|
|
{
|
|
Logger::error() << "Failed to add a seccomp rule";
|
|
return false;
|
|
}
|
|
}
|
|
|
|
bool success = seccomp_load(ctx) == 0;
|
|
if(!success)
|
|
{
|
|
Logger::error() << "Failed to load seccomp filter";
|
|
return false;
|
|
}
|
|
seccomp_release(ctx);
|
|
return success;
|
|
}
|
|
|
|
bool SandboxLinux::bindMountPaths(std::string target_root, const std::vector<std::string> &paths)
|
|
{
|
|
for(const std::string &path : paths)
|
|
{
|
|
std::string chroot_target_path = target_root + path;
|
|
if(std::filesystem::exists(chroot_target_path))
|
|
{
|
|
continue;
|
|
}
|
|
if(std::filesystem::is_regular_file(path))
|
|
{
|
|
std::fstream f1;
|
|
f1.open(chroot_target_path, std::ios::out);
|
|
f1.close();
|
|
}
|
|
else
|
|
{
|
|
std::error_code ec;
|
|
//TODO: fails if the stuff already exists, but it shouldn't according to doc?
|
|
if(!std::filesystem::create_directories(chroot_target_path, ec))
|
|
{
|
|
Logger::error() << "Error while trying to duplicate structure for sandbox. Dir creation failed. Path: " << chroot_target_path << " Error: " << ec.message();
|
|
return false;
|
|
}
|
|
|
|
if(mount(path.c_str(), chroot_target_path.c_str(), NULL, MS_BIND, NULL) == -1)
|
|
{
|
|
Logger::error() << "Bind mount for " << path << " -> " << chroot_target_path << " failed! " << strerror(errno);
|
|
return false;
|
|
}
|
|
}
|
|
|
|
}
|
|
return true;
|
|
}
|
|
|
|
bool SandboxLinux::isolateNamespaces(std::vector<std::string> fsPaths)
|
|
{
|
|
std::sort(fsPaths.begin(), fsPaths.end(), [](const std::string &a, const std::string &b){ return a.length() < b.length(); });
|
|
|
|
auto current_uid = getuid();
|
|
auto current_gid = getgid();
|
|
if(unshare(CLONE_NEWUSER) == -1)
|
|
{
|
|
Logger::error() << "Failed to unshare user namespace: " << strerror(errno);
|
|
return false;
|
|
}
|
|
|
|
std::fstream setgroups;
|
|
setgroups.open("/proc/self/setgroups", std::ios::out | std::ios::app);
|
|
setgroups << "deny";
|
|
setgroups.flush();
|
|
setgroups.close();
|
|
|
|
std::fstream uid_map;
|
|
uid_map.open("/proc/self/uid_map", std::ios::out | std::ios::app);
|
|
uid_map << "0 " << current_uid << " 1\n";
|
|
uid_map.flush();
|
|
uid_map.close();
|
|
|
|
std::fstream gid_map;
|
|
gid_map.open("/proc/self/gid_map", std::ios::out);
|
|
uid_map << "0 " << current_gid << " 1\n";
|
|
gid_map.flush();
|
|
gid_map.close();
|
|
|
|
if(unshare(CLONE_NEWNS) == -1)
|
|
{
|
|
Logger::error() << "Failed to unshare mount namespace: " << strerror(errno);
|
|
return false;
|
|
}
|
|
|
|
/*
|
|
* TODO: breaks server.
|
|
* TODO: fork?
|
|
* if(unshare(CLONE_NEWPID) == -1)
|
|
{
|
|
Logger::error() << "Failed to unshare pid namespace: " << strerror(errno);
|
|
return false;
|
|
}*/
|
|
|
|
|
|
//The purpose is to start with a clean sandbox dir.
|
|
//We maybe could work with mkdirat, and check whether it exists alrady, to avoid
|
|
//some attacks where an attacker gueses the dir, but in that case the system is already compromised
|
|
//TODO: still, check, whether this is something we must consider here or not...
|
|
Random random;
|
|
std::string rootpath = "/tmp/qswiki_sandbox_" + random.getRandomHexString(10) + "/";
|
|
if(!std::filesystem::create_directory(rootpath))
|
|
{
|
|
Logger::error() << "Failed to create chroot directory for sandbox";
|
|
return false;
|
|
}
|
|
|
|
|
|
if(!bindMountPaths(rootpath, fsPaths ))
|
|
{
|
|
Logger::error() << "Bind mounting paths failed!";
|
|
return false;
|
|
}
|
|
|
|
|
|
|
|
if(chroot(rootpath.c_str()) == -1)
|
|
{
|
|
Logger::error() << "chroot to sandbox failed!";
|
|
return false;
|
|
}
|
|
if(chdir("/") == -1)
|
|
{
|
|
Logger::error() << "chdir to sandbox failed!";
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
bool SandboxLinux::enableForInit()
|
|
{
|
|
umask(0027);
|
|
|
|
//TODO. there is execv for SPARC. Sigh...
|
|
if(!seccomp_blacklist({ SCMP_SYS(execveat), SCMP_SYS(execve) }))
|
|
{
|
|
Logger::error() << "Failed to install blacklisting seccomp filter";
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
|
|
|
|
}
|
|
|
|
bool SandboxLinux::enablePreWorker(std::vector<std::string> fsPaths)
|
|
{
|
|
if(!isolateNamespaces(fsPaths))
|
|
{
|
|
Logger::error() << "Failed to isolate namespaces";
|
|
return false;
|
|
}
|
|
return true;
|
|
|
|
}
|
|
|
|
bool SandboxLinux::supported()
|
|
{
|
|
std::fstream stream;
|
|
stream.open("/proc/sys/kernel/unprivileged_userns_clone");
|
|
if(stream.is_open())
|
|
{
|
|
std::string str;
|
|
stream >> str;
|
|
if(str[0] == '0')
|
|
{
|
|
Logger::error() << "Please write '1' to /proc/sys/kernel/unprivileged_userns_clone in order to enable sandboxing support on this system";
|
|
return false;
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
bool SandboxLinux::enableForWorker()
|
|
{
|
|
int cap = 0;
|
|
int res = 0;
|
|
while((res = prctl(PR_CAPBSET_DROP, cap++, 0, 0, 0)) == 0);
|
|
if(res == -1 && errno != EINVAL)
|
|
{
|
|
Logger::error() << "Failed to drop the capability bounding set!";
|
|
return false;
|
|
}
|
|
|
|
__user_cap_header_struct h = { 0 };
|
|
h.pid = 0;
|
|
h.version = _LINUX_CAPABILITY_VERSION_3;
|
|
__user_cap_data_struct drop[2];
|
|
drop[0].effective = 0;
|
|
drop[0].permitted = 0;
|
|
drop[0].inheritable = 0;
|
|
drop[1].effective = 0;
|
|
drop[1].permitted = 0;
|
|
drop[1].inheritable = 0;
|
|
if(capset(&h, drop) == -1)
|
|
{
|
|
Logger::error() << "Failed to drop capabilities: " << strerror(errno);
|
|
return false;
|
|
}
|
|
|
|
|
|
|
|
|
|
/* TODO: as said, a whitelist approach is better. As such, this list is bound to be incomplete in the
|
|
* sense that more could be listed here and some critical ones are probably missing */
|
|
if(! seccomp_blacklist({
|
|
SCMP_SYS(setuid),
|
|
SCMP_SYS(setuid32),
|
|
SCMP_SYS(connect),
|
|
SCMP_SYS(chroot),
|
|
SCMP_SYS(pivot_root),
|
|
SCMP_SYS(mount),
|
|
SCMP_SYS(setns),
|
|
SCMP_SYS(unshare),
|
|
SCMP_SYS(ptrace),
|
|
SCMP_SYS(personality)
|
|
|
|
}))
|
|
{
|
|
Logger::error() << "Sandbox: Activation of seccomp blacklist failed!";
|
|
return false;
|
|
}
|
|
if(prctl(PR_SET_DUMPABLE, 0) == -1)
|
|
{
|
|
Logger::error() << "prctl: PR_SET_DUMPABLE failed";
|
|
return false;
|
|
}
|
|
if(prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0) == -1)
|
|
{
|
|
Logger::error() << "prctl: PR_SET_NO_NEW_PRIVS failed: " << strerror(errno);
|
|
return false;
|
|
}
|
|
|
|
if(! seccomp_blacklist({ SCMP_SYS(prctl) }))
|
|
{
|
|
Logger::error() << "Sandbox: Activation of seccomp blacklist failed!";
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|