exile.h provides an API for processes on Linux to easily isolate themselves in order
to mitigate the effect of exploited vulnerabilities, i. e. when attacker has achieved
arbitrary code execution. exile.h makes it simpler for developers to use existing technologies such as Seccomp and Linux Namespaces. Those generally require knowledge of details and are not trivial for developers to employ, which prevents a more widespread adoption.
The following section offers small examples. Then the motivation is explained in more detail. Proper API documentation will be maintained in other files.
This section quickly demonstrates the simplicity of the API. It serves as an overview to get a first impression.
system() is used to keep the example C code short. It also demonstrates that subprocesses are also subject to restrictions imposed by exile.h.
While the examples show different features separately, it is generally possible to combine those.
#include "exile.h"
#include <assert.h>
int main(void)
{
system("echo test > /home/user/testfile");
struct exile_policy *policy = exile_init_policy();
exile_append_path_policies(policy, EXILE_FS_ALLOW_ALL_READ, "/home/user");
exile_append_path_policies(policy, EXILE_FS_ALLOW_ALL_READ | EXILE_FS_ALLOW_ALL_WRITE, "/tmp");
int ret = exile_enable_policy(policy);
if(ret != 0)
{
exit(EXIT_FAILURE);
}
int fd = open("/home/user/test", O_CREAT | O_WRONLY | O_TRUNC, 0600);
assert(fd == -1);
fd = open("/home/user/testfile", O_RDONLY);
//use fd
assert(fd != -1);
fd = open("/tmp/testfile", O_CREAT | O_WRONLY | O_TRUNC, 0600);
//use fd
assert(fd != -1);
return 0;
}The assert() calls won't be fired, consistent with the policy that allows only reading from /home/user. We can write to /tmp/ though as it was specified in the policy.
exile.h allows specifying which syscalls are permitted or denied. In the following example, 'ls' is never executed, as the specified "vows" do not allow the execve() system call. The process will be killed.
#include "exile.h"
int main(void)
{
struct exile_policy *policy = exile_init_policy();
policy->vow_promises = exile_vows_from_str("stdio rpath wpath cpath");
exile_enable_policy(policy);
printf("Trying to execute...");
execlp("/bin/ls", "ls", "/", NULL);
}exile offers a quick way to isolate a process from the default network namespace.
#include "exile.h"
int main(void)
{
struct exile_policy *policy = exile_init_policy();
policy->namespace_options |= EXILE_UNSHARE_NETWORK;
int ret = exile_enable_policy(policy);
if(ret != 0)
{
exit(EXIT_FAILURE);
}
system("curl -I https://evil.tld");
}Produces curl: (6) Could not resolve host: evil.tld. For example, this is useful for subprocesses which do not need
network access, but perform tasks such as parsing user-supplied file formats.
Currently, work is being done that hopefully will allow isolation of individual function calls in a mostly pain-free manner.
Consider the following C++ code:
#include <iostream>
#include <fstream>
#include "exile.hpp"
std::string cat(std::string path)
{
std::fstream f1;
f1.open(path.c_str(), std::ios::in);
std::string content;
std::string line;
while(getline(f1, line)) {
content += line + "\n";
}
return content;
}
int main(void)
{
struct exile_policy *policy = exile_init_policy();
policy->vow_promises = exile_vows_from_str("stdio rpath");
std::string content = exile_launch<std::string>(policy, cat, "/etc/hosts");
std::cout << content;
policy = exile_init_policy();
policy->vow_promises = exile_vows_from_str("stdio");
try
{
content = exile_launch<std::string>(policy, cat, "/etc/hosts");
std::cout << content;
}
catch(std::exception &e)
{
std::cout << "launch failure: " << e.what() << std::endl;
}
}We execute "cat()". The first call succeeds. In the second, we get an exception, because the subprocess "cat()" was launched in violated the policy (missing "rpath" vow).
Naturally, there is a performance overhead. Certain challenges remain, such as the fact that being executed in a subproces, we operate on copies, so handling references is not something that has been given much thought. There is also the fact that clone()ing from threads opens a can of worms, particularly with locks. Hence, exile_launch() is best avoided in multi-threaded contexts.
No release yet, experimental, API is unstable, builds will break on updates of this library.
Currently, it's mainly evolving from the needs of my other projects which use exile.h.
exile.h unlocks existing Linux mechanisms to facilitate isolation of processes from resources. Limiting the scope of what programs can do helps defending the rest of the system when a process gets under attacker's control (when classic mitigations such as ASLR etc. failed). To this end, OpenBSD has the pledge() and unveil() functions available. Those functions are helpful mitigation mechanisms, but such accessible ways are unfortunately not readily available on Linux. This is where exile.h steps in.
Seccomp allows restricting the system calls available to a process and thus decrease the systems attack surface, but it generally is not easy to use. Requiring BPF filter instructions, you generally just can't make use of it right away without learning about BPF. exile.h provides an API inspired by pledge(), building on top of seccomp. It also provides an interface to manually restrict the system calls that can be issued.
Traditional methods employed to restrict file system access, like different uids/gids, chroot, bind-mounts, namespaces etc. may require administrator intervention, are perhaps only suitable for daemons and not desktop applications, or are generally rather involved. As a positive example, Landlock since 5.13 is a vast improvement to limit file system access of processes. It also greatly simplifies exile.h' implementation of fs isolation.
Abstracting those details may help developers bring sandboxing into their applications.
- Restricting file system access (using Landlock or Namespaces/chroot as fallback)
- Systemcall filtering (using seccomp-bpf). An interface inspired by OpenBSD's pledge() is available
- Dropping privileges in general, such as capabilities
- Isolating the application from the network, etc. through Namespaces
- Helpers to isolate single functions
A way for end users/administrators to restrict processes. In the future, a wrapper binary may be available to achieve this, but it generally aims for developers to bring sandboxing/isolation into their software. This allows a more fine-grained approach, as the developers are more familiar with their software. Applying restrictions with solutions like AppArmor requires them to be present and installed on the system and it's easy to break things this way.
Therefore, software should ideally be written with sandboxing in mind from the beginning.
Will be available once the interface stabilizes.
It's recommended to start with [README.usage.md] to get a feeling for exile.h. API-Documentation: [README.api.md]
Built upon kernel technologies, exile.h naturally inherits their limitations:
- New syscalls can be introduced by new kernel versions. exile.h must keep in sync, and users must keep the library up to date.
- seccomp has no deep argument inspection (yet), particularly new syscalls cannot be reasonably filtered, such as clone3(), or io_uring.
- You can't know what syscalls libraries will issue. An update to existing libraries may cause them to use different syscalls not allowed by a policy. However, using vows and keeping up to date with exile.h should cover that.
- Landlock, currently, does not apply to syscalls such as stat().
TODO:
- ioctl does not know the fd, so checking values is kind of strange
- redundancies: some things are handled by capabilties, other by seccomp or both
- no magic, be reasonable, devs should not get sloppy, restrict IPC.
Kernel >=3.17
While mostly transparent to users of this API, kernel >= 5.13 is required to take advantage of Landlock. Furthermore, it depends on distro-provided kernels being reasonable and enabling it by default. In practise, Landlock maybe won't be used in some cases so exile.h will use a combination of namespaces, bind mounts and chroot as fallbacks.
No.
You can thank a Debian-specific kernel patch for that. Execute
echo 1 > /proc/sys/kernel/unprivileged_userns_clone to disable that patch for now.
Note that newer releases should not cause this problem any longer, as explained in the Debian release notes.
pledge() cannot be properly implemented using seccomp. The "vow" concept here may look similiar, and it is, but it's not pledge().
Contributions are very welcome. Options:
- Pull-Request on github
- Mail to
exile at quitesimple.orgwith instructions on where to pull the changes from. - Mailing a classic patch/diff to the same address.
ISC