Linux namespaces

Detaches the calling VM/actor from the role of supervisor for the process/actor represented by self. After this operation is done, the process/actor represented by self is allowed to outlive the calling process.

Sends signo to the process. On SIGKILL, it’ll also close the channel.

Emilua follows the advice from WireGuard developers to avoid parsing bugs by avoiding object serialization altogether. Sequenced-packet sockets with builtin framing are used so we always receive/send whole messages in one API call.

There is a hard-limit (configurable at build time) on the maximum number of members you can send per message. This limit would need to exist anyway to avoid DoS from bad clients.

Another limitation is that no nesting is allowed. You can either send a single non-nil value or a non-empty dictionary where every member in it is a leaf from the root tree. The messaging API is part of the attack surface that bad clients can exploit. We cannot afford a single bug here, so the code must be simple. By forbidding subtrees we can ignore recursion complexities and simplify the code a lot.

The struct used to receive messages follows:

A variant class is needed to send the messages. Given a variant is needed anyway, we just adopt NaN-tagging for its implementation as that will make the struct members packed together and no memory from the host process hidden among paddings will leak to the containers.

The code assumes that no signaling NaNs are ever produced by the Lua VM to simplify the NaN-tagging scheme^[2][3]. The type is stored in the mantissa bits of a signaling NaN.

If the first member is nil, then we have a non-dictionary value stored in members[1]. Otherwise, a nil will act as a sentinel to the end of the dictionary. No sentinel will exist when the dictionary is fully filled.

read() calls will write to objects of this type directly (i.e. no intermediate char[N] buffer is used) so we avoid any complexity with code related to alignment adjustments.

memset(buf, 0, s) is used to clear any unused member of the struct before a call to write() so we avoid leaking memory from the process to any container.

Strings are preceded by a single byte that contains the size of the string that follows. Therefore, strings are limited to 255 characters. Following from this scheme, a buffer sufficiently large to hold the largest message is declared to avoid any buffer overflow. However, we still perform bounds checking to make sure no uninitialized data from the code stack is propagated back to Lua code to avoid leaking any memory. The bounds checking function in the code has a simple implementation that doesn’t make the code much more complex and it’s easy to follow.

To send file descriptors over, SCM_RIGHTS is used. There are a lot of quirks involved with SCM_RIGHTS (e.g. extra file descriptors could be stuffed into the buffer even if you didn’t expect them). The encoding scheme for the network buffer is far simpler to use than SCM_RIGHTS' ancillary data. Complexity-wise, there’s far greater chance to introduce a bug in code related to SCM_RIGHTS than a bug in the code that parses the network buffer.

Code could be simpler if we only supported messaging strings over, but that would just defer the problem of secure serialization on the user’s back. Code should be simple, but not simpler. By throwing all complexity on the user’s back, the implementation would offer no security. At least we centralized the sensitive object serialization so only one block of code need to be reviewed and audited.

UNIX systems allow the userspace to spawn new processes by a fork() followed by an exec(). exec() really means an executable will be available in the container, but this assumption doesn’t play nice with our idea of spawning new actors in an empty container.

What we really want is to to perform a fork followed by no exec() call. This approach in itself also has its own problems. exec() is the only call that will flush the address space of the running process. If we don’t exec() then the new process that was supposed to run untrusted code with no access to system resources will be able to read all previous memory — memory that will most likely contain sensitive information that we didn’t want leaked. Other problems such as threads (supported by the Emilua runtime) would also hinder our ability to use fork() without exec()ing.

One simple approach to solve all these problems is to fork() at the beginning of the program so we fork() before any sensitive information is loaded in the process' memory. Forking at a well known point also brings other benefits. We know that no thread has been created yet, so resources such as locks and the global memory allocator stay in a well defined state. By creating this extra process before much more extra virtual memory or file descriptor slots in our process table have been requested, we also make sure that further processes creation will be cheaper.

Every time the main process wants to create an actor in a new process, it’ll defer its job onto the supervisor that was fork()ed near main(). An AF_UNIX+SOCK_SEQPACKET socket is used to orchestrate this process. Given the supervisor is only used to create new processes, it can use blocking APIs that will simplify the code a lot. The blocking read() on the socket also means that it won’t be draining any CPU resources when it’s not needed. Also important is the threat model here. The main process is not trying to attack the supervisor process. The supervisor is also trusted and it doesn’t need to run inside a container. SCM_RIGHTS handling between the main process and the supervisor is simplified a lot due to these constraints.

However some care is still needed to setup the supervisor. Each actor will initially be an exact copy of the supervisor process memory and we want to make sure that no sensitive data is leaked there. The first thing we do right after creating the supervisor is collecting any sensitive information that might still exist in the main process (e.g. argv and envp) and instructing the supervisor process to explicit_bzero() them. This compromise is not as good as exec() would offer, but it’s the best we can do while we limit ourselves to reasonably portable C code with few assumptions about dynamic/static linkage against system libraries, and other settings from the host environment.

This problem doesn’t end here. Now that we assume the process memory from the supervisor contains no sensitive data, we want to keep it that way. It may be true that every container is assumed as a container that some hacker already took over (that’s why we’re isolating them, after all), but one container shouldn’t leak information to another one. In other words, we don’t even want to load sensitive information regarding the setup of any container from the supervisor process as that could leak into future containers. The solution here is to serialize such information (e.g. the init.script) such that it is only sent directly to the final process. Another AF_UNIX+SOCK_SEQPACKET socket is used.

Now to the assumptions on the container process. We do assume that it’ll run code that is potentially dangerous and some hacker might own the container at some point. However the initial setup does not run arbitrary dangerous code and it still is part of the trusted computing base. The problem is that we don’t know whether the init.script will need to load sensitive information at any point to perform its job. That’s why we setup the Lua VM that runs init.script to use a custom allocator that will explicit_bzero() all allocated memory at the end. Allocations done by external libraries such as libcap lie outside of our control, but they rarely matter anyway.

That’s mostly the bulk of our problems and how we handle them. Other problems are summarized in the short list below.

The output from tools such as top start to become rather cool when you play with nested containers:

PID1 eases our life a lot. As soon as any container starts to act suspiciously we can safely kill the whole subtree of processes by sending SIGKILL to the PID1 that started it.

AF_UNIX+SOCK_SEQPACKET sockets are connection-oriented and simplify our work even further. We shutdown() the ends of each pair such that they’ll act unidirectionally just like pipes. When all copies of one end die, the operation on the other end will abort. The actor API translates to MPSC channels, so we never ever send the reading end to any container (we only make copies of the sending end). The kernel will take care of any tricky reference counting necessary (and SIGKILLing PID1 will make sure no unwanted end survives).

The only work left for us to do is pretty much to just orchestrate the internal concurrency architecture of the runtime (e.g. watch out for blocking reads). Given that we want to abort reads when all the copies of the sending end are destroyed, we don’t keep any copy to the sending end in our own process. Everytime we need to send our address over, we create a new pair of sockets to send the newly created sending end over. inbox will unify the receipt of messages coming from any of these sockets. You can think of each newly created socket as a new capability. If one capability is revoked, others remain unaffected.

One good actor could send our address further to a bad actor, and there is no way to revoke access to the bad actor without also revoking access to the good actor, but that is in line with capability-based security systems. Access rights are transitive. In fact, a bad actor could write 0-sized messages over the AF_UNIX+SOCK_SEQPACKET socket to trick us into thinking the channel was already closed. We’ll happily close the channel and there is no problem here. The system can happily recover later on (and only this capability is revoked anyway).

The runtime doesn’t schedule any read on the socket unless the user calls inbox:receive(). Upon reading a new message the runtime will either wake the receiving fiber directly, or enqueue the result in a buffer if no receiving fiber exists at the time (this can happen if the user interrupted the fiber, or another result arrived and woke the fiber up already). inbox:receive() won’t schedule any read on the socket if there’s some result already enqueued in the buffer.

We don’t offer any helper to spawn a program (i.e. system.spawn()) within an existing PID namespace. That’s intentional (although one could still do it through init.script). setns(fd, CLONE_NEWPID) is dangerous. Only exec() will flush the address space for the process. The window of time that exists until exec() is called means that any memory from the previous process could be read by a compromised container (cf. ptrace(2)).

A mix of approaches is used to test the implementation.

There’s an unit test for every class of good inputs. There are unit tests for accidental bad inputs that one might try to perform through the Lua API. The unit tests always try to create one scenario for buffered messages and another for immediate delivery of the result.

When support for plugins is enabled, fuzz tests are built as well. The fuzzers are generation-based. One fuzzer will generate good input and test if the program will accept all of them. Another fuzzer will mutate a good input into a bad one (e.g. truncate the message size to attempt a buffer overflow), and check if the program rejects all of them.

There are some other tests as well (e.g. ensure no padding exists between the members of the C struct we send over the wire).

These functions live inside the global table C. errno (or 0 on success) is returned as the second result.

Other exported functions work as usual (except that errno or 0 is returned as the second result):

These constants live inside the global table C.