Despite being a prototype, ARM's implementation of the CHERI ISA, Morello, is rapidly gaining traction in the research community due to the high potential for isolation and additional security it brings without sacrificing performance. I have recently ported a proof-of-concept version of Unikraft to run bare metal on the Morello machine, and I am in the process of exploring the cool security properties that combining CHERI with a modern unikernel such as Unikraft can bring.
Morello is ARM's implementation of CHERI. Capability Hardware Enhanced RISC Instructions (CHERI) extends traditional ISAs (Instruction Set Architectures) such as ARMv8 and RISC-V to enable fine-grained memory protection at the granularity of a byte, which enables memory safe version of languages like C and C++, as well as efficient compartmentalisation of both code and data. This is achieved by introducing capabilities, which can be thought of as pointers, except that they also contain bounds information and permissions. Such hardware-enforced capabilities can be used to help reduce the amount of memory-related security issues present in software by ensuring that execution and data accesses are checked against bounds and permissions, thus ensuring better memory safety. This protection can be applied at the level of an application, placing an application or subsets of it in a compartment, all the way down to encompassing individual buffers and data structures.
Morello implements CHERI capabilities by extending existing registers and pointers (via capabilities) to 128 bits, with key registers such as the PC and SP, for example, now having capability-counterparts, called PCC (Program Counter Capability) and CSP (Stack Pointer Capability). The existing ARM instruction set has been modified to allow the use of these capability registers in place of 64-bit registers in operations such as capability-aware loads and stores, in effect enforcing the bounds embedded in dereferenced capabilities. Morello also introduces a Default Data Capability (DDC) which can be used to specify bounds on every data access, which will be enforced against regular A64 load/store instructions which are capability unaware, enabling compartmentalisation of existing software.
Since Morello is an extension to a regular ARMv8 CPU - in the case of the Morello machine, the SoC is based on an existing Neoverse design - there are 3 ways in which software can be run:
I have ported Unikraft to Morello with the ability to run in hybrid mode. Running in hybrid mode has the advantage that existing libraries can be more easily adapted to work with capabilities as needed while still retaining regular A64 functionality.
Now that we've established that Morello is really cool :), let's take a look at what it takes to get a basic Unikraft system running on Morello hardware. The Morello platform leverages the work done for running Unikraft on the Raspberry Pi as a starting point. Other than different memory mappings needed to ensure proper operation (see the programmer's model), my port includes the following items:
At the start of the boot process, code runs at Exception Level 3. At this point a number of values must be set to enable access to Morello features. These are controlled by the following bits:
cptr_el3.EC (bit 9)
Bit EC controls whether access to Morello features from any exception level will cause the instruction to be trapped at EL3.
Setting this bit to
0b1 will ensure that Morello features are not trapped.
cptr_el2.CEN (bits 18 and 19) and cptr_el2.TC (bit 9)
0b0 are set to disable the trapping to EL2 of Morello instructions or instructions which access Morello system registers.
cpacr_el1, CEN (bits 18 and 19)
0b11 is set to disable the trapping of access to Morello instructions or instructions which access Morello system registers at EL0 or EL1.
With Morello now enabled, the ARM VBAR registers are now 128-bit CVBAR registers for each exception level. Therefore, to set the vector base address, we first derive a capability and then set CVBAR as normal to point to the exception vectors.
Dropping to EL1:
After setup at EL3 is complete, we perform an exception return to EL1. With Morello enabled, the exception link register requires a capability to return to. We, therefore, derive one and set CELR_EL3 to contain this capability.
The Morello machine uses a 50MHz reference clock for UART, therefore, the following calculations for BRD values are required for a baud rate of 115200:
Calculation:UART_CLK = 50MHzBaud rate divisor = (50 x (10^6)) / (16 * 115200) = 27.12673611Therefore, IBRD is 27, for the fractional part, we do thefollowing:((0.12673611 * 64) + 0.5) = 9 (as integer)So FBRD is 9
This basic proof of concept unlocks several exciting research avenues. As a first step we are porting our security-oriented library operating system, FlexOS (also built atop Unikraft), to leverage CHERI capabilities on Morello as an efficient isolation mechanism. We further plan to explore advanced use of capabilities in FlexOS and Unikraft by studying the possibility of protecting critical kernel data structures (e.g. page tables) with capabilities in hybrid mode. We also plan to investigate the incremental porting of Unikraft/FlexOS to pure capability to maximise the level of protection, hopefully without compromising on the high degree of performance brought by the unikernel OS model.
Feel free to ask questions, report issues, and meet new people.