tvisor: a Tiny Thread-level syscall interception framework

Warning

I stopped working on this project because I realized that almost the same thing can be achieved with Kernel runtime security instrumentation very easily, though it is not userspace mechanism. Therefore, the sole purpose of this repository is to show how to build an userspace syscall interception framework by only relying on the (classic) seccomp-bpf.

The code is not complete at all but at least 95% of basic musl-libc tests are passing for "no-op" syscall interception. The one of the most difficult parts that I have not implemented is proper signal handling of blocked signals while executing interruptible syscalls.

I am not sure if I will continue this project, but I will keep this repository as a reference for the future.

tvisor is a tiny 100% userspace syscall interception framework that can be used to build a program to monitor syscalls. "T" indicates both "Tiny" and "Thread-level". Only available on Linux.

A program built with tvisor will be a single-binary, and it does not spawn guest in a separate process, but it directly runs the guest in the same process. It runs in a higher address space than the guest, and injects a tiny monitor in the process, then starts running the guest in the same process. The monitor is responsible for intercepting the syscall and forwarding it to the host while rejecting/modifying the syscall arguments if necessary. It spawns a "kernel" thread which corresponds to each guest program thread. A kernel is responsible for handling syscalls in the corresponding guest thread.

tvisor is able to run any binary either statically or dynamically(TODO but should be possible) linked, without any modification (with some exceptions).

How it works

First of all, the tvisor binaries are running in a higher address space than the guest by passing --image-base linker flag when building the tvisor binary. Tvisor users run the guest binary as ./tvisor <tvisor_args> -- <guest_binary> <guest_args>. Tvisor binary does the following things:

Parse the ELF binary and load it into the same virtual memory space.
Installs special signal handlers to handle all signals.
Spawns the "kernel thread" corresponding to each guest thread. Each kernel thread is responsible for handling target syscalls in the corresponding guest thread. It spawns the corresponding guest thread and starts running the guest thread.
Each guest thread installs sigaltstack to handle any signals raised in the guest thread. The sigalt stack contains the various data structures that allow each signal handler can communicate with the corresponding kernel thread.
Each guest thread also installs seccomp-bpf filter to intercept syscalls, which raises SIGSYS signals. Which syscall is intercepted is determined by users of tvisor library.
After all the setup is done, the guest thread start running the guest executable code.

Example

There are two tvisor usage example programs in tvisor/bin. To build them, run make build. You see the binary ./tvisor/target/aarch64-unknown-none/release/tvisor-fs which is a freestanding linux ELF binary.

tvisor-fs example program is a simple program that modifies the current working directory and root directory of the guest program by intercepting getcwd and open kind system calls. For example, for the example program that prints the current working directory getcwd, tvisor-fs runs as a file system sandbox program that changes the current working directory to /tmp:

$ ./getcwd
Current working directory: /Users/mathetake/tvisor
$ tvisor-fs --cwd /anydir -- ./getcwd
Current working directory: /anydir