Last yr we wrote about how shifting native code in Android from C++ to Rust has resulted in fewer safety vulnerabilities. Most of the elements we talked about then had been system companies in userspace (operating below Linux), however these will not be the one elements sometimes written in memory-unsafe languages. Many security-critical elements of an Android system run in a “bare-metal” setting, exterior of the Linux kernel, and these are traditionally written in C. As a part of our efforts to harden firmware on Android gadgets, we’re more and more utilizing Rust in these bare-metal environments too.
To that finish, we’ve rewritten the Android Virtualization Framework’s protected VM (pVM) firmware in Rust to offer a reminiscence secure basis for the pVM root of belief. This firmware performs the same operate to a bootloader, and was initially constructed on prime of U-Boot, a extensively used open supply bootloader. However, U-Boot was not designed with safety in a hostile setting in thoughts, and there have been quite a few safety vulnerabilities present in it because of out of bounds reminiscence entry, integer underflow and reminiscence corruption. Its VirtIO drivers particularly had various lacking or problematic bounds checks. We mounted the precise points we present in U-Boot, however by leveraging Rust we are able to keep away from these kinds of memory-safety vulnerabilities in future. The new Rust pVM firmware was launched in Android 14.
As a part of this effort, we contributed again to the Rust neighborhood through the use of and contributing to current crates the place attainable, and publishing various new crates as properly. For instance, for VirtIO in pVM firmware we’ve frolicked fixing bugs and soundness points within the current virtio-drivers crate, in addition to including new performance, and at the moment are serving to preserve this crate. We’ve printed crates for making PSCI and different Arm SMCCC calls, and for managing web page tables. These are only a begin; we plan to launch extra Rust crates to assist bare-metal programming on a spread of platforms. These crates are additionally getting used exterior of Android, resembling in Project Oak and the bare-metal part of our Comprehensive Rust course.
Training engineers
Many engineers have been positively stunned by how productive and nice Rust is to work with, offering good high-level options even in low-level environments. The engineers engaged on these tasks come from a spread of backgrounds. Our complete Rust course has helped skilled and novice programmers rapidly come in control. Anecdotally the Rust sort system (together with the borrow checker and lifetimes) helps keep away from making errors which might be simply made in C or C++, resembling leaking tips that could stack-allocated values out of scope.
One of our bare-metal Rust course attendees had this to say:
"sorts will be constructed that herald all of Rust's niceties and safeties and but nonetheless compile all the way down to extraordinarily environment friendly code like writes of constants to memory-mapped IO."
97% of attendees that accomplished a survey agreed the course was price their time.
Advantages and challenges
Device drivers are sometimes written in an object-oriented style for flexibility, even in C. Rust traits, which will be seen as a type of compile-time polymorphism, present a helpful high-level abstraction for this. In many instances this may be resolved totally at compile time, with no runtime overhead of dynamic dispatch by way of vtables or structs of operate pointers.
There have been some challenges. Safe Rust’s sort system is designed with an implicit assumption that the one reminiscence this system must care about is allotted by this system (be it on the stack, the heap, or statically), and solely utilized by this system. Bare-metal applications usually need to take care of MMIO and shared reminiscence, which break this assumption. This tends to require numerous unsafe code and uncooked pointers, with restricted instruments for encapsulation. There is a few disagreement within the Rust neighborhood concerning the soundness of references to MMIO area, and the amenities for working with uncooked pointers in steady Rust are presently considerably restricted. The stabilisation of offset_of, slice_ptr_get, slice_ptr_len, offset_of and different nightly options will enhance this, however it’s nonetheless difficult to encapsulate cleanly. Better syntax for accessing struct fields and array indices by way of uncooked pointers with out creating references would even be useful.
The concurrency launched by interrupt and exception handlers will also be awkward, as they usually must entry shared mutable state however can’t depend on with the ability to take locks. Better abstractions for vital sections will assist considerably, however there are some exceptions that may’t virtually be disabled, resembling web page faults used to implement copy-on-write or different on-demand web page mapping methods.
Another situation we’ve had is that some unsafe operations, resembling manipulating the web page desk, can’t be encapsulated cleanly as they’ve security implications for the entire program. Usually in Rust we’re capable of encapsulate unsafe operations (operations which can trigger undefined behaviour in some circumstances, as a result of they’ve contracts which the compiler can’t examine) in secure wrappers the place we guarantee the required preconditions in order that it’s not attainable for any caller to trigger undefined behaviour. However, mapping or unmapping pages in a single a part of this system could make different components of this system invalid, so we haven’t discovered a means to offer a totally common secure interface to this. It ought to be famous that the identical issues apply to a program written in C, the place the programmer all the time has to purpose concerning the security of the entire program.
Some individuals adopting Rust for bare-metal use instances have raised issues about binary measurement. We have seen this in some instances; for instance our Rust pVM firmware binary is round 460 kB in comparison with 220 kB for the sooner C model. However, this isn’t a good comparability as we additionally added extra performance which allowed us to take away different elements from the boot chain, so the general measurement of all VM boot chain elements was comparable. We additionally weren’t notably optimizing for binary measurement on this case; velocity and correctness had been extra vital. In instances the place binary measurement is vital, compiling with measurement optimization, being cautious about dependencies, and avoiding Rust’s string formatting equipment in launch builds often permits comparable outcomes to C.
Architectural assist is one other concern. Rust is mostly properly supported on the Arm and RISC-V cores that we see most frequently, however assist for extra esoteric architectures (for instance, the Qualcomm Hexagon DSP included in lots of Qualcomm SoCs utilized in Android telephones) will be missing in comparison with C.
The way forward for bare-metal Rust
Overall, regardless of these challenges and limitations, we’ve nonetheless discovered Rust to be a big enchancment over C (or C++), each by way of security and productiveness, in all of the bare-metal use instances the place we’ve tried it thus far. We plan to make use of it wherever sensible.
As properly because the work within the Android Virtualization Framework, the workforce engaged on Trusty (the open-source Trusted Execution Environment used on Pixel telephones, amongst others) have been exhausting at work including assist for Trusted Applications written in Rust. For instance, the reference KeyMint Trusted Application implementation is now in Rust. And there’s extra to come back in future Android gadgets, as we proceed to make use of Rust to enhance safety of the gadgets you belief.