Guest Support
RVM paravirtualizes RTOS kernels with drop-in port layers: the guest library mirrors the relative layout of each official RTOS tree and contains only the files that were added — official kernel files are never included or modified, so a plain overlay is always safe.
Support matrix
| Guest | Version | Status | ARMv6-M | ARMv7-M | RV32G | Toolchains |
|---|---|---|---|---|---|---|
| RMP | latest | Practical | ✅ | ✅ | ✅ | Keil, IAR, GCC |
| FreeRTOS | V9.0.0 | Practical | ✅ | ✅ | ✅ | Keil, IAR, GCC |
| uC/OS-II | V2.93 | Practical | ✅ | ✅ | ✅ | Keil, IAR, GCC |
| uC/OS-III | V3.08.02 | Practical | ✅ | ✅ | ✅ | Keil, IAR, GCC |
| RT-Thread | V4.1.1 | Practical | ✅ | ✅ | ✅ | Keil, IAR, GCC |
| ChibiOS | 21.11.6 | Practical | ✅ | ✅ | ✅ | Keil, IAR, GCC |
| RTEMS | 7.0-rc1 | Practical | — | ✅ | — | GCC |
| Zephyr | 3.7.0 LTS | Practical | — | ✅ | — | GCC |
| NuttX | 13.1.0 | Practical | — | ✅ | — | GCC |
| DjyOS | V1.1.0 | Commemorative | — | ✅ | — | GCC |
| KLite | V5.2.1 | Commemorative | — | ✅ | — | GCC |
- Practical ports are production ports; the ✅ columns show each port's verified architectures and toolchains, and every practical port is measured in the benchmark tables. RTEMS, Zephyr and NuttX target ARMv7-M with GCC, verified on the H743 and F429 boards; their benchmark rows are on the H743. ChibiOS covers the same three architectures and three toolchains as the classic guests, verified on the G071, H743, F429 and CH32V307 boards.
- Commemorative ports are memorial single-combination ports (Cortex-M7 + GCC + STM32H743ZIT6). They demonstrate the porting technique on kernels with a fixed single combination.
- RMP is not distributed in the guest library: it carries its RVM platform layers natively (
Source/Platform/*_RVM/in the RMP tree) and needs no guest-side packaging.
Evaluation boards
Cortex-M0+: STM32G071RBT6 · Cortex-M4: STM32F429ZIT6 · Cortex-M7: STM32H743ZIT6 · RV32IMAC: CH32V307VCT6
Per-guest notes
RMP
EDI's own real-time kernel. It defines the original RVM switch template that every other port transplants: locked-state yield requests become a kernel pend flag replayed at unlock, so the port only ever sees unmasked, immediate switch requests.
FreeRTOS (FRT)
A software deferred-yield that mimics the official PendSV pend via a nesting counter. configUSE_RVM_OPTIMISED_YIELD selects between the task-stack fast path and the pure-hypercall yield (which costs ~4x Yield latency with identical semantics); on RV32P, configUSE_PORT_RVF/configUSE_PORT_RVD select the FPU frame width.
uC/OS-II (UO2)
Self-contained (no uC-CPU/uC-LIB). OS_CPU_SR_Save/Restore map onto the RVM virtual interrupt mask snapshot; clean frames take the assembly fast path, dirty ones the yield hypercall. Kernel objects must be created from a task context, so applications boot from an init task.
uC/OS-III (UO3)
OS_CFG_RVM_FAST_YIELD_EN toggles the same fast/hypercall trade-off as FreeRTOS. The direct switch always lands the incoming task unmasked, preserving the kernel's Disable; Yield; Enable contract.
RT-Thread (RTT)
Fast path for clean frames, deferred hypercall path for dirty ICI/IT frames. Requires RT_USING_LIBC. On RV32P the port is GCC-only.
DjyOS (DJY)
DjyOS schedules events; every switch runs inside the kernel's async-signal mask. The port hosts all threads in one RVM user thread and masquerades contexts as 22-word exception frames. The guest must be compiled with -mfloat-abi=soft.
KLite (KLT)
The always-deferred kernel shape: cpu_contex_switch() only pends, and the outermost cpu_leave_critical() consumes the pend through a yield dispatcher with a masquerade fast path. Selected in the project generator with the -klt flag.
ChibiOS (CHB)
ChibiOS 21.11.6 LTS runs as an unprivileged RVM user thread with its own scheduler and chconf.h. The port adds two architecture directories, os/common/ports/ARMv7-M-RVM and RV32P-RVM (plus ARMv6-M-RVM for M0+), without touching the ChibiOS kernel. Clean task frames take a direct task-stack fast path; dirty or vector-captured frames go through the virtual context handler, and idle parks in RVM_Hyp_Vct_Wait(). ARMv7-M supports GCC, IAR EWARM and Keil MDK; RV32P is GCC-only. Verified with FPU frames on H743 and F429, basic frames on G071, and RVF frames on CH32V307. CHB_FAST_PATH=0 builds the all-vector variant.
RTEMS (RTE)
RTEMS 7.0-rc1 runs as an unprivileged RVM user thread. The port is distributed as the RTEMS_V701 BSP/spec/sample overlay for an official RTEMS tree, not a standalone tree. The task-level switch uses a 22-word basic frame (56-word extended frame with FPU) and a direct task-stack fast path; dirty or ISR-captured frames are handed to the Vct context handler. Setting RTE_FAST_PATH=False compiles out the fast path and routes every switch through the vector thread. The BSP keeps its fixed -O2 optimisation level. RTE_CPU_FLAGS selects Cortex-M4F or Cortex-M7; FPU frames are implemented and the two-task FPU test passes on H743. The full 15-item benchmark lives in the RTE chip test headers.
Zephyr (ZPR)
Zephyr 3.7.0 LTS runs as an unprivileged RVM user thread through the arch/arm/core/cortex_m/v7m_rvm/ arch port; upstream Cortex-M sources are untouched and the build selects the new port. The generator writes a per-VM zpr_board.dts for the single virtual armv7m board and passes it as DTS_SOURCE; no STM32 dtsi, RCC/PLL or peripheral nodes are involved. Normal ZPR images use the native benchmark process as the TIM4 heartbeat, so a normal guest never registers or rewrites TIM4. The context layer supports 22-word basic and 56-word FPU-extended frames; FPU preemption passes on H743 with both fast-path and all-Vct builds, and on F429 with the fast path. The port requires !SMP, !USERSPACE, !ARM_MPU, !CORTEX_M_SYSTICK, !TICKLESS_KERNEL and !GEN_ISR_TABLES. The benchmark exercises the common 15-item subset.
NuttX (NTX)
NuttX 13.1.0 runs as an unprivileged RVM user thread in a flat build. The port lives in arch/arm/src/armv7-m-rvm/ behind the chip-agnostic armv7-m-rvm/ntx board (virt for Cortex-M7, virt-m4 for Cortex-M4). A yield only posts the switch flag; the masquerade frame (22 words, 56 with FPU) is built on the task stack, and CONFIG_ARMV7M_RVM_FAST_PATH selects the direct stack-swap path for clean frames. No userspace, MPU, SMP, tickless mode or physical NVIC. The benchmark covers the nine NuttX-native rows — FIFO and event notification are absent because NuttX has no such primitives.
Port contracts
Every port implements the same five-part job — critical sections, yield, context switch, interrupt delivery hooks and startup — while each kernel guarantees something different. The full contract/strategy analysis for each guest (including the interrupt model's three-layer gate and frame interchangeability) lives in the GUEST.md document in the repository.
For installation (tree overlay) and project generation, see Getting Started.
