If you need urgent consulting help click here

Arm Cortex-M Developer Guide

Overview

This page contains detailed information about the status of the Arm Cortex-M architecture porting in the Zephyr RTOS and describes key aspects when developing Zephyr applications for Arm Cortex-M-based platforms.

Key supported features

The table below summarizes the status of key OS features in the different Arm Cortex-M implementation variants.

Processor families

Architecture variant

Arm v6-M

Arm v7-M

Arm v8-M

Arm v8.1-M

M0/M1

M0+

M3

M4

M7

M23

M33

M55

OS Features

Programmable fault IRQ priorities

Y

N

Y

Y

Y

N

Y

Y

Single-thread kernel support

Y

Y

Y

Y

Y

Y

Y

Y

Thread local storage support

Y

Y

Y

Y

Y

Y

Y

Y

Interrupt handling

Regular interrupts

Y

Y

Y

Y

Y

Y

Y

Y

Dynamic interrupts

Y

Y

Y

Y

Y

Y

Y

Y

Direct interrupts

Y

Y

Y

Y

Y

Y

Y

Y

Zero Latency interrupts

N

N

Y

Y

Y

Y

Y

Y

CPU idling

Y

Y

Y

Y

Y

Y

Y

Y

Native system timer (SysTick)

N 1

Y

Y

Y

Y

Y

Y

Y

Memory protection

User mode

N

Y

Y

Y

Y

Y

Y

Y

HW stack protection (MPU)

N

N

Y

Y

Y

Y

Y

Y

HW-assisted stack limit checking

N

N

N

N

N

Y 2

Y

Y

HW-assisted null-pointer dereference detection

N

N

Y

Y

Y

Y

Y

Y

HW-assisted atomic operations

N

N

Y

Y

Y

N

Y

Y

Support for non-cacheable regions

N

N

Y

Y

Y

N

Y

Y

Execute SRAM functions

N

N

Y

Y

Y

N

Y

Y

Floating Point Services

N

N

N

Y

Y

N

Y

Y

DSP ISA

N

N

N

Y

Y

N

Y

Y

Trusted-Execution

Native TrustZone-M support

N

N

N

N

N

Y

Y

Y

TF-M integration

N

N

N

N

N

N

Y

N

Code relocation

Y

Y

Y

Y

Y

Y

Y

Y

SW-based vector table relaying

Y

Y

Y

Y

Y

Y

Y

Y

HW-assisted timing functions

N

N

Y

Y

Y

N

Y

Y

Notes

1

SysTick is optional in Cortex-M1

2

Stack limit checking only in Secure builds in Cortex-M23

OS features

Threads

Thread stack alignment

Each Zephyr thread is defined with its own stack memory. By default, Cortex-M enforces a double word thread stack alignment, see CONFIG_STACK_ALIGN_DOUBLE_WORD. If MPU-based HW-assisted stack overflow detection (CONFIG_MPU_STACK_GUARD) is enabled, thread stacks need to be aligned with a larger value, reflected by CONFIG_ARM_MPU_REGION_MIN_ALIGN_AND_SIZE. In Arm v6-M and Arm v7-M architecture variants, thread stacks are additionally required to be align with a value equal to their size, in applications that need to support user mode (CONFIG_USERSPACE). The thread stack sizes in that case need to be a power of two. This is all reflected by CONFIG_MPU_REQUIRES_POWER_OF_TWO_ALIGNMENT, that is enforced in Arm v6-M and Arm v7-M builds with user mode support.

Stack pointers

While executing in thread mode the processor is using the Process Stack Pointer (PSP). The processor uses the Main Stack Pointer (MSP) while executing in handler mode, that is, while servicing exceptions and HW interrupts. Using PSP in thread mode facilitates thread stack pointer manipulation during thread context switching, without affecting the current execution context flow in handler mode.

In Arm Cortex-M builds a single interrupt stack memory is shared among exceptions and interrupts. The size of the interrupt stack needs to be selected taking into consideration nested interrupts, each pushing an additional stack frame. Developers can modify the interrupt stack size using CONFIG_ISR_STACK_SIZE.

The interrupt stack is also used during early boot so the kernel can initialize the main thread’s stack before switching to the main thread.

Thread context switching

In Arm Cortex-M builds, the PendSV exception is used in order to trigger a context switch to a different thread. PendSV exception is always present in Cortex-M implementations. PendSV is configured with the lowest possible interrupt priority level, in all Cortex-M variants. The main reasons for that design are

  • to utilize the tail chaining feature of Cortex-M processors, and thus limit the number of context switch operations that occur.

  • to not impact the interrupt latency observed by HW interrupts.

As a result, context switch in Cortex-M is non-atomic, i.e. it may be preempted by HW interrupts, however, a context-switch operation must be completed before a new thread context-switch may start.

Typically a thread context-switch will perform the following operations

  • When switching-out the current thread, the processor stores

    • the callee-saved registers (R4 - R11) in the thread’s container for callee-saved registers, which is located in kernel memory

    • the thread’s current operation mode

      • user or privileged execution mode

      • presence of an active floating point context

      • the EXC_RETURN value of the current handler context (PendSV)

    • the floating point callee-saved registers (S16 - S31) in the thread’s container for FP callee-saved registers, if the current thread has an active FP context

    • the PSP of the current thread which points to the beginning of the current thread’s exception stack frame. The latter contains the caller-saved context and the return address of the switched-out thread.

  • When switching-in a new thread the processor

    • restores the new thread’s callee-saved registers from the thread’s container for callee-saved registers

    • restores the new thread’s operation mode

    • restores the FP callee-saved registers if the switched-in thread had an active FP context before being switched-out

    • re-programs the dynamic MPU regions to allow a user thread access its stack and application memories, and/or programs a stack-overflow MPU guard at the bottom of the thread’s privileged stack

    • restores the PSP for the incoming thread and re-programs the stack pointer limit register (if applicable, see CONFIG_BUILTIN_STACK_GUARD)

    • optionally does a stack limit checking for the switched-in thread, if sentinel-based stack limit checking is enabled (see CONFIG_STACK_SENTINEL).

PendSV exception return sequence restores the new thread’s caller-saved registers and the return address, as part of unstacking the exception stack frame.

The implementation of the context-switch mechanism is present in arch/arm/core/aarch32/swap_helper.S.

Stack limit checking (Arm v8-M)

Armv8-M and Armv8.1-M variants support stack limit checking using the MSPLIM and PSPLIM core registers. The feature is enabled when CONFIG_BUILTIN_STACK_GUARD is set. When stack limit checking is enabled, both the thread’s privileged or user stack, as well as the interrupt stack are guarded by PSPLIM and MSPLIM registers, respectively. MSPLIM is configured once during kernel boot, while PSLIM is re-programmed during every thread context-switch or during system calls, when the thread switches from using its default stack to using its privileged stack, and vice versa. PSPLIM re-programming

  • has a relatively low runtime overhead (programming is done with MSR instructions)

  • does not impact interrupt latency

  • does not require any memory areas to be reserved for stack guards

  • does not make use of MPU regions

It is, therefore, considered as a lightweight but very efficient stack overflow detection mechanism in Cortex-M applications.

Stack overflows trigger the dedicated UsageFault exception provided by Arm v8-M.

Interrupt handling features

This section describes certain aspects around exception and interrupt handling in Arm Cortex-M.

Interrupt priority levels

The number of available (configurable) interrupt priority levels is determined by the number of implemented interrupt priority bits in NVIC; this needs to be described for each Cortex-M platform using DeviceTree:

&nvic {
        arm,num-irq-priority-bits = <#priority-bits>;
};

Reserved priority levels

A number of interrupt priority levels are reserved for the OS.

By design, system fault exceptions have the highest priority level. In Baseline Cortex-M, this is actually enforced by hardware, as HardFault is the only available processor fault exception, and its priority is higher than any configurable exception priority.

In Mainline Cortex-M, the available fault exceptions (e.g. MemManageFault, UsageFault, etc.) are assigned the highest configurable priority level. (CONFIG_CPU_CORTEX_M_HAS_PROGRAMMABLE_FAULT_PRIOS signifies explicitly that the Cortex-M implementation supports configurable fault priorities.)

This priority level is never shared with HW interrupts (an exception to this rule is described below). As a result, processor faults occurring in regular ISRs will be handled by the corresponding fault handler and will not escalate to a HardFault, similar to processor faults occurring in thread mode.

SVC exception is normally configured with the highest configurable priority level (an exception to this rule will be described below). SVCs are used by the Zephyr kernel to dispatch system calls, trigger runtime system errors (e.g. Kernel oops or panic), or implement IRQ offloading.

In Baseline Cortex-M the priority level of SVC may be shared with other exceptions or HW interrupts that are also given the highest configurable priority level (As a result of this, kernel runtime errors during interrupt handling will escalate to HardFault. Additional logic in the fault handling routines ensures that such runtime errors are detected successfully).

In Mainline Cortex-M, however, the SVC priority level is reserved, thus normally it is only shared with the fault exceptions of configurable priority. This simplifies the fault handling routines in Mainline Cortex-M architecture, since runtime kernel errors are serviced by the SVC handler (i.e no HardFault escalation, even if the kernel errors occur in ISR context).

HW interrupts in Mainline Cortex-M builds are allocated a priority level lower than the SVC.

One exception to the above rules is when Zephyr applications support Zero Latency Interrupts (ZLIs). Such interrupts are designed to have a priority level higher than any HW or system interrupt. If the ZLI feature is enabled in Mainline Cortex-M builds (see CONFIG_ZERO_LATENCY_IRQS), then

  • ZLIs are assigned the highest configurable priority level

  • SVCs are assigned the second highest configurable priority level

  • Regular HW interrupts are assigned priority levels lower than SVC.

The priority level configuration in Cortex-M is implemented in include/arch/arm/aarch32/exc.h.

Locking and unlocking IRQs

In Baseline Cortex-M locking interrupts is implemented using the PRIMASK register.

arch_irq_lock()

will set the PRIMASK register to 1, eventually, masking all IRQs with configurable priority. While this fulfils the OS requirement of locking interrupts, the consequence is that kernel runtime errors (triggering SVCs) will escalate to HardFault.

In Mainline Cortex-M locking interrupts is implemented using the BASEPRI register (Mainline Cortex-M builds select CONFIG_CPU_CORTEX_M_HAS_BASEPRI to signify that BASEPRI register is implemented.). By modifying BASEPRI (or BASEPRI_MAX) arch_irq_lock() masks all system and HW interrupts with the exception of

  • SVCs

  • processor faults

  • ZLIs

This allows zero latency interrupts to be triggered inside OS critical sections. Additionally, this allows system (processor and kernel) faults to be handled by Zephyr in exactly the same way, regardless of whether IRQs have been locked or not when the error occurs. It also allows for system calls to be dispatched while IRQs are locked.

Note

Mainline Cortex-M fault handling is designed and configured in a way that all processor and kernel faults are handled by the corresponding exception handlers and never result in HardFault escalation. In other words, a HardFault may only occur in Zephyr applications that have modified the default fault handling configurations. The main reason for this design was to reserve the HardFault exception for handling exceptional error conditions in safety critical applications.

Dynamic direct interrupts

Cortex-M builds support the installation of direct interrupt service routines during runtime. Direct interrupts are designed for performance-critical interrupt handling and do not go through all of the common Zephyr interrupt handling code.

Direct dynamic interrupts are enabled via switching on CONFIG_DYNAMIC_DIRECT_INTERRUPTS.

Note that enabling direct dynamic interrupts requires enabling support for dynamic interrupts in the kernel, as well (see CONFIG_DYNAMIC_INTERRUPTS).

Zero Latency interrupts

As described above, in Mainline Cortex-M applications, the Zephyr kernel reserves the highest configurable interrupt priority level for its own use (SVC). SVCs will not be masked by interrupt locking. Zero-latency interrupt can be used to set up an interrupt at the highest interrupt priority which will not be blocked by interrupt locking. To use the ZLI feature CONFIG_ZERO_LATENCY_IRQS needs to be enabled.

Zero latency IRQs have minimal interrupt latency, as they will always preempt regular HW or system interrupts.

Note, however, that since ZLI ISRs will run at a priority level higher than the kernel exceptions they cannot use any kernel functionality. Additionally, since the ZLI interrupt priority level is equal to processor fault priority level, faults occurring in ZLI ISRs will escalate to HardFault and will not be handled in the same way as regular processor faults. Developers need to be aware of this limitation.

CPU Idling

The Cortex-M architecture port implements both k_cpu_idle() and k_cpu_atomic_idle(). The implementation is present in arch/arm/core/aarch32/cpu_idle.S.

In both implementations, the processor will attempt to put the core to low power mode. In k_cpu_idle() the processor ends up executing WFI (Wait For Interrupt) instruction, while in k_cpu_atomic_idle() the processor will execute a WFE (Wait For Event) instruction.

When using the CPU idling API in Cortex-M it is important to note the following:

  • Both k_cpu_idle() and k_cpu_atomic_idle() are assumed to be invoked with interrupts locked. This is taken care of by the kernel if the APIs are called by the idle thread.

  • After waking up from low power mode, both functions will restore interrupts unconditionally, that is, regardless of the interrupt lock status before the CPU idle API was called.

The Zephyr CPU Idling mechanism is detailed in CPU Idling.

Memory protection features

This section describes certain aspects around memory protection features in Arm Cortex-M applications.

User mode system calls

User mode is supported in Cortex-M platforms that implement the standard (Arm) MPU or a similar core peripheral logic for memory access policy configuration and control, such as the NXP MPU for Kinetis platforms. (Currently, CONFIG_ARCH_HAS_USERSPACE is selected if CONFIG_ARM_MPU is enabled by the user in the board default Kconfig settings).

A thread performs a system call by triggering a (synchronous) SVC exception, where

  • up to 5 arguments are placed on registers R1 - R5

  • system call ID is placed on register R6.

The SVC Handler will branch to the system call preparation logic, which will perform the following operations

  • switch the thread’s PSP to point to the beginning of the thread’s privileged stack area, optionally reprogramming the PSPLIM if stack limit checking is enabled

  • modify CONTROL register to switch to privileged mode

  • modify the return address in the SVC exception stack frame, so that after exception return the system call dispatcher is executed (in thread privileged mode)

Once the system call execution is completed the system call dispatcher will restore the user’s original PSP and PSPLIM and switch the CONTROL register back to unprivileged mode before returning back to the caller of the system call.

System calls execute in thread mode and can be preempted by interrupts at any time. A thread may also be context-switched-out while doing a system call; the system call will resume as soon as the thread is switched-in again.

The system call dispatcher executes at SVC priority, therefore it cannot be preempted by HW interrupts (with the exception of ZLIs), which may observe some additional interrupt latency if they occur during a system call preparation.

MPU-assisted stack overflow detection

Cortex-M platforms with MPU may enable CONFIG_MPU_STACK_GUARD to enable the MPU-based stack overflow detection mechanism. The following points need to be considered when enabling the MPU stack guards

  • stack overflows are triggering processor faults as soon as they occur

  • the mechanism is essential for detecting stack overflows in supervisor threads, or user threads in privileged mode; stack overflows in threads in user mode will always be detected regardless of CONFIG_MPU_STACK_GUARD being set.

  • stack overflows are always detected, however, the mechanism does not guarantee that no memory corruption occurs when supervisor threads overflow their stack memory

  • CONFIG_MPU_STACK_GUARD will normally reserve one MPU region for programming the stack guard (in certain Arm v8-M configurations with CONFIG_MPU_GAP_FILLING enabled 2 MPU regions are required to implement the guard feature)

  • MPU guards are re-programmed at every context-switch, adding a small overhead to the thread swap routine. Compared, however, to the CONFIG_BUILTIN_STACK_GUARD feature, no re-programming occurs during system calls.

  • When CONFIG_HW_STACK_PROTECTION is enabled on Arm v8-M platforms the native stack limit checking mechanism is used by default instead of the MPU-based stack overflow detection mechanism; users may override this setting by manually enabling CONFIG_MPU_STACK_GUARD in these scenarios.

Memory map and MPU considerations

Fixed MPU regions

By default, when CONFIG_ARM_MPU is enabled a set of fixed MPU regions are programmed during system boot.

  • One MPU region programs the entire flash area as read-execute. User can override this setting by enabling CONFIG_MPU_ALLOW_FLASH_WRITE, which programs the flash with RWX permissions. If CONFIG_USERSPACE is enabled unprivileged access on the entire flash area is allowed.

  • One MPU region programs the entire SRAM area with privileged-only RW permissions. That is, an MPU region is utilized to disallow execute permissions on SRAM. (An exception to this setting is when CONFIG_MPU_GAP_FILLING is disabled (Arm v8-M only); in that case no SRAM MPU programming is done so the access is determined by the default Arm memory map policies, allowing for privileged-only RWX permissions on SRAM).

  • All the memory regions defined in the devicetree with the compatible zephyr,memory-region and at least the property zephyr,memory-region-mpu defining the MPU permissions for the memory region. See the next section for more details.

The above MPU regions are defined in soc/arm/common/cortex_m/arm_mpu_regions.c. Alternative MPU configurations are allowed by enabling CONFIG_CPU_HAS_CUSTOM_FIXED_SOC_MPU_REGIONS. When enabled, this option signifies that the Cortex-M SoC will define and configure its own fixed MPU regions in the SoC definition.

Fixed MPU regions defined in devicetree

The user can define memory regions to be allocated and created in the linker script using nodes with the zephyr,memory-region devicetree compatible. When the property zephyr,memory-region-mpu is present in such a node, a new MPU region will be allocated and programmed during system boot.

The property zephyr,memory-region-mpu is a string carrying the attributes for the MPU region. It is converted to a C token for use defining the attributes of the MPU region.

For example, to define a new non-cacheable memory region in devicetree:

sram_no_cache: memory@20300000 {
     compatible = "zephyr,memory-region", "mmio-sram";
     reg = <0x20300000 0x100000>;
     zephyr,memory-region = "SRAM_NO_CACHE";
     zephyr,memory-region-mpu = "RAM_NOCACHE";
};

This will automatically create a new MPU entry in soc/arm/common/cortex_m/arm_mpu_regions.c with the correct name, base, size and attributes gathered directly from the devicetree. See include/zephyr/linker/devicetree_regions.h for more details.

Static MPU regions

Additional static MPU regions may be programmed once during system boot. These regions are required to enable certain features

  • a RX region to allow execution from SRAM, when CONFIG_ARCH_HAS_RAMFUNC_SUPPORT is enabled and users have defined functions to execute from SRAM.

  • a RX region for relocating text sections to SRAM, when CONFIG_CODE_DATA_RELOCATION_SRAM is enabled

  • a no-cache region to allow for a none-cacheable SRAM area, when CONFIG_NOCACHE_MEMORY is enabled

  • a possibly unprivileged RW region for GCOV code coverage accounting area, when CONFIG_COVERAGE_GCOV is enabled

  • a no-access region to implement null pointer dereference detection, when CONFIG_NULL_POINTER_EXCEPTION_DETECTION_MPU is enabled

The boundaries of these static MPU regions are derived from symbols exposed by the linker, in include/linker/linker-defs.h.

Dynamic MPU regions

Certain thread-specific MPU regions may be re-programmed dynamically, at each thread context switch:

  • an unprivileged RW region for the current thread’s stack area (for user threads)

  • a read-only region for the MPU stack guard

  • unprivileged RW regions for the partitions of the current thread’s application memory domain.

Considerations

The number of available MPU regions for a Cortex-M platform is a limited resource. Most platforms have 8 MPU regions, while some Cortex-M33 or Cortex-M7 platforms may have up to 16 MPU regions. Therefore there is a relatively strict limitation on how many fixed, static and dynamic MPU regions may be programmed simultaneously. For platforms with 8 available MPU regions it might not be possible to enable all the aforementioned features that require MPU region programming. In most practical applications, however, only a certain set of features is required and 8 MPU regions are, in many cases, sufficient.

In Arm v8-M processors the MPU architecture does not allow programmed MPU regions to overlap. CONFIG_MPU_GAP_FILLING controls whether the fixed MPU region covering the entire SRAM is programmed. When it does, a full SRAM area partitioning is required, in order to program the static and the dynamic MPU regions. This increases the total number of required MPU regions. When CONFIG_MPU_GAP_FILLING is not enabled the fixed MPU region covering the entire SRAM is not programmed, thus, the static and dynamic regions are simply programmed on top of the always-existing background region (full-SRAM partitioning is not required). Note, however, that the background SRAM region allows execution from SRAM, so when CONFIG_MPU_GAP_FILLING is not set Zephyr is not protected against attacks that attempt to execute malicious code from SRAM.

Floating point Services

Both unshared and shared FP registers mode are supported in Cortex-M (see Floating Point Services for more details).

When FPU support is enabled in the build (CONFIG_FPU is enabled), the sharing FP registers mode (CONFIG_FPU_SHARING) is enabled by default. This is done as some compiler configurations may activate a floating point context by generating FP instructions for any thread, regardless of whether floating point calculations are performed, and that context must be preserved when switching such threads in and out.

The developers can still disable the FP sharing mode in their application projects, and switch to Unshared FP registers mode, if it is guaranteed that the image code does not generate FP instructions outside the single thread context that is allowed (and supposed) to do so.

Under FPU sharing mode, the callee-saved FPU registers are saved and restored in context-switch, if the corresponding threads have an active FP context. This adds some runtime overhead on the swap routine. In addition to the runtime overhead, the sharing FPU mode

  • requires additional memory for each thread to save the callee-saved FP registers

  • requires additional stack memory for each thread, to stack the caller-saved FP registers, upon exception entry, if an FP context is active. Note, however, that since lazy stacking is enabled, there is no runtime overhead of FP context stacking in regular interrupts (FP state preservation is only activated in the swap routine in PendSV interrupt).

Misc

Chain-loadable images

Cortex-M applications may either be standalone images or chain-loadable, for instance, by a bootloader. Application images chain-loadable by bootloaders (or other applications) normally occupy a specific area in the flash denoted as their code partition. CONFIG_USE_DT_CODE_PARTITION will ensure that a Zephyr chain-loadable image will be linked into its code partition, specified in DeviceTree.

HW initialization at boot

In order to boot properly, chain-loaded applications may require that the core Arm hardware registers and peripherals are initialized in their reset values. Enabling CONFIG_INIT_ARCH_HW_AT_BOOT Zephyr to force the initialization of the internal Cortex-M architectural state during boot to the reset values as specified by the corresponding Arm architecture manual.

Software vector relaying

In Cortex-M platforms that implement the VTOR register (see CONFIG_CPU_CORTEX_M_HAS_VTOR), chain-loadable images relocate the Cortex-M vector table by updating the VTOR register with the offset of the image vector table.

Baseline Cortex-M platforms without VTOR register might not be able to relocate their vector table which remains at a fixed location. Therefore, a chain-loadable image will require an alternative way to route HW interrupts and system exceptions to its own vector table; this is achieved with software vector relaying.

When a bootloader image enables CONFIG_SW_VECTOR_RELAY it is able to relay exceptions and interrupts based on a vector table pointer that is set by the chain-loadable application. The latter sets the CONFIG_SW_VECTOR_RELAY_CLIENT option to instruct the boot sequence to set the vector table pointer in SRAM so that the bootloader can forward the exceptions and interrupts to the chain-loadable image’s software vector table.

While this feature is intended for processors without VTOR register, it may also be used in Mainline Cortex-M platforms.

Code relocation

Cortex-M support the code relocation feature. When CONFIG_CODE_DATA_RELOCATION_SRAM is selected, Zephyr will relocate .text, data and .bss sections from the specified files and place it in SRAM. It is possible to relocate only parts of the code sections into SRAM, without relocating the whole image text and data sections. More details on the code relocation feature can be found in Code And Data Relocation.

Linking Cortex-M applications

Most Cortex-M platforms make use of the default Cortex-M GCC linker script in include/arch/arm/aarch32/cortex-m/scripts/linked.ld, although it is possible for platforms to use a custom linker script as well.

CMSIS

Cortex-M CMSIS headers are hosted in a standalone module repository: zephyrproject-rtos/cmsis.

CONFIG_CPU_CORTEX_M selects CONFIG_HAS_CMSIS_CORE to signify that CMSIS headers are available for all supported Cortex-M variants.

Testing

A list of unit tests for the Cortex-M porting and miscellaneous features is present in tests/arch/arm/. The tests suites are continuously extended and new test suites are added, in an effort to increase the coverage of the Cortex-M architecture support in Zephyr.

QEMU

We use QEMU to verify the implemented features of the Cortex-M architecture port in Zephyr. Adequate coverage is achieved by defining and utilizing a list of QEMU targets, each with a specific architecture variant and Arm peripheral support list.

The table below lists the QEMU platform targets defined in Zephyr along with the corresponding Cortex-M implementation variant and the peripherals these targets emulate.

QEMU target

Architecture variant

Arm v6-M

Arm v7-M

Arm v8-M

Arm v8.1-M

qemu_cortex_m0

qemu_cortex_m3

mps2_an385

mps2_an521

mps3_an547

Emulated features

NVIC

Y

Y

Y

Y

Y

BASEPRI

N

Y

Y

Y

Y

SysTick

N

Y

Y

Y

Y

MPU

N

N

Y

Y

Y

FPU

N

N

N

Y

N

SPLIM

N

N

N

Y

Y

TrustZone-M

N

N

N

Y

N

Maintainers & Collaborators

The status of the Arm Cortex-M architecture port in Zephyr is: maintained. The updated list of maintainers and collaborators for Cortex-M can be found in MAINTAINERS.yml.