vmbase

This directory contains a Rust crate and static library which can be used to write no_std Rust binaries to run in an aarch64 VM under crosvm (via the VirtualizationService), such as for pVM firmware, a VM bootloader or kernel.

In particular it provides:

• An entry point that initialises the MMU with a hard-coded identity mapping, enables the cache, prepares the image and allocates a stack.
• An exception vector to call your exception handlers.
• A UART driver and println! macro for early console logging.
• Functions to shutdown or reboot the VM.

Libraries are also available for heap allocation, page table manipulation and PSCI calls.

Usage

The example subdirectory contains an example of how to use it for a VM bootloader.

Build file

Start by creating a rust_ffi_static rule containing your main module:

rust_ffi_static {
    name: "libvmbase_example",
    crate_name: "vmbase_example",
    srcs: ["src/main.rs"],
    edition: "2021",
    no_stdlibs: true,
    stdlibs: [
        "libcompiler_builtins.rust_sysroot",
        "libcore.rust_sysroot",
    ],
    rustlibs: [
        "libvmbase",
    ],
    enabled: false,
    target: {
        android_arm64: {
            enabled: true,
        },
    },
}

Note that stdlibs must be explicitly specified, as we don't want the normal set of libraries used for a C++ binary intended to run in Android userspace.

Entry point

Your main module needs to specify a couple of special attributes:

#![no_main]
#![no_std]

This tells rustc that it doesn‘t depend on std, and won’t have the usual main function as an entry point. Instead, vmbase provides a macro to specify your main function:

use vmbase::{main, println};

main!(main);

pub fn main(arg0: u64, arg1: u64, arg2: u64, arg3: u64) {
    println!("Hello world");
}

vmbase adds a wrapper around your main function to initialise the console driver first (with the UART at base address 0x3f8, the first UART allocated by crosvm), and make a PSCI SYSTEM_OFF call to shutdown the VM if your main function ever returns.

You can also shutdown the VM by calling vmbase::power::shutdown or ‘reboot’ by calling vmbase::power::reboot. Either will cause crosvm to terminate the VM, but by convention we use shutdown to indicate that the VM has finished cleanly, and reboot to indicate an error condition.

Exception handlers

You must provide handlers for each of the 8 types of exceptions which can occur on aarch64. These must use the C ABI, and have the expected names. For example, to log sync exceptions and reboot:

use vmbase::{console::emergency_write_str, power::reboot};

extern "C" fn sync_exception_current() {
    emergency_write_str("sync_exception_current\n");

    let mut esr: u64;
    unsafe {
        asm!("mrs {esr}, esr_el1", esr = out(reg) esr);
    }
    eprintln!("esr={:#08x}", esr);

    reboot();
}

The println! macro shouldn't be used in exception handlers, because it relies on a global instance of the UART driver which might be locked when the exception happens, which would result in deadlock. Instead you can use emergency_write_str and eprintln!, which will re-initialise the UART every time to ensure that it can be used. This should still be used with care, as it may interfere with whatever the rest of the program is doing with the UART.

Note also that in some cases when the system is in a bad state resulting in the stack not working properly, eprintln! may hang. emergency_write_str may be more reliable as it seems to avoid any stack allocation. This is why the example above uses emergency_write_str first to ensure that at least something is logged, before trying eprintln! to print more details.

See example/src/exceptions.rs for a complete example.

Linker script and initial idmap

The entry point code expects to be provided a hardcoded identity-mapped page table to use initially. This must contain at least the region where the image itself is loaded, some writable DRAM to use for the .bss and .data sections and stack, and a device mapping for the UART MMIO region. See the example/idmap.S for an example of how this can be constructed.

The addresses in the pagetable must map the addresses the image is linked at, as we don't support relocation. This can be achieved with a linker script, like the one in example/image.ld. The key part is the regions provided to be used for the image and writable data:

MEMORY
{
	image		: ORIGIN = 0x80200000, LENGTH = 2M
	writable_data	: ORIGIN = 0x80400000, LENGTH = 2M
}

Building a binary

To link your Rust code together with the entry point code and idmap into a static binary, you need to use a cc_binary rule:

cc_binary {
    name: "vmbase_example_elf",
    stem: "vmbase_example",
    srcs: [
        "idmap.S",
    ],
    static_libs: [
        "libvmbase_entry",
        "libvmbase_example",
    ],
    static_executable: true,
    nocrt: true,
    system_shared_libs: ["libc"],
    stl: "none",
    linker_scripts: [
        "image.ld",
        ":vmbase_sections",
    ],
    installable: false,
    enabled: false,
    target: {
        android_arm64: {
            enabled: true,
        },
    },
}

This takes your Rust library (libvmbase_example), the vmbase library entry point and exception vector (libvmbase_entry) and your initial idmap (idmap.S) and builds a static binary with your linker script (image.ld) and the one provided by vmbase (sections.ld). This is an ELF binary, but to run it as a VM bootloader you need to objcopy it to a raw binary image instead, which you can do with a raw_binary rule:

raw_binary {
    name: "vmbase_example",
    src: ":vmbase_example_elf",
    stem: "vmbase_example.bin",
    enabled: false,
    target: {
        android_arm64: {
            enabled: true,
        },
    },
}

The resulting binary can then be used to start a VM by passing it as the bootloader in a VirtualMachineRawConfig.