Architecture Porting Guide

An architecture port is needed to enable Zephyr to run on an ISA or an ABI that is not currently supported.

The following are examples of ISAs and ABIs that Zephyr supports:

  • x86_32 ISA with System V ABI
  • x86_32 ISA with IAMCU ABI
  • ARMv7-M ISA with Thumb2 instruction set and ARM Embedded ABI (aeabi)
  • ARCv2 ISA

For information on Kconfig configuration, see the Setting configuration values section in the Board Porting Guide. Architectures use a similar Kconfig configuration scheme. The Kconfig - Tips and Best Practices page has some general recommendations and tips for writing Kconfig files as well.

An architecture port can be divided in several parts; most are required and some are optional:

  • The early boot sequence: each architecture has different steps it must take when the CPU comes out of reset (required).
  • Interrupt and exception handling: each architecture handles asynchronous and unrequested events in a specific manner (required).
  • Thread context switching: the Zephyr context switch is dependent on the ABI and each ISA has a different set of registers to save (required).
  • Thread creation and termination: A thread’s initial stack frame is ABI and architecture-dependent, and thread abortion possibly as well (required).
  • Device drivers: most often, the system clock timer and the interrupt controller are tied to the architecture (some required, some optional).
  • Utility libraries: some common kernel APIs rely on a architecture-specific implementation for performance reasons (required).
  • CPU idling/power management: most architectures implement instructions for putting the CPU to sleep (partly optional, most likely very desired).
  • Fault management: for implementing architecture-specific debug help and handling of fatal error in threads (partly optional).
  • Linker scripts and toolchains: architecture-specific details will most likely be needed in the build system and when linking the image (required).

Early Boot Sequence

The goal of the early boot sequence is to take the system from the state it is after reset to a state where is can run C code and thus the common kernel initialization sequence. Most of the time, very few steps are needed, while some architectures require a bit more work to be performed.

Common steps for all architectures:

  • Setup an initial stack.
  • If running an XIP kernel, copy initialized data
  • from ROM to RAM.
  • If not using an ELF loader, zero the BSS section.
  • Jump to _Cstart(), the early kernel initialization
    • _Cstart() is responsible for context switching out of the fake context running at startup into the main thread.

Some examples of architecture-specific steps that have to be taken:

  • If given control in real mode on x86_32, switch to 32-bit protected mode.
  • Setup the segment registers on x86_32 to handle boot loaders that leave them in an unknown or broken state.
  • Initialize a board-specific watchdog on Cortex-M3/4.
  • Switch stacks from MSP to PSP on Cortex-M.
  • Use a different approach than calling into _Swap() on Cortex-M to prevent race conditions.
  • Setup FIRQ and regular IRQ handling on ARCv2.

Interrupt and Exception Handling

Each architecture defines interrupt and exception handling differently.

When a device wants to signal the processor that there is some work to be done on its behalf, it raises an interrupt. When a thread does an operation that is not handled by the serial flow of the software itself, it raises an exception. Both, interrupts and exceptions, pass control to a handler. The handler is known as an ISR in the case of interrupts. The handler perform the work required the exception or the interrupt. For interrupts, that work is device-specific. For exceptions, it depends on the exception, but most often the core kernel itself is responsible for providing the handler.

The kernel has to perform some work in addition to the work the handler itself performs. For example:

  • Prior to handing control to the handler:
    • Save the currently executing context.
    • Possibly getting out of power saving mode, which includes waking up devices.
    • Updating the kernel uptime if getting out of tickless idle mode.
  • After getting control back from the handler:
    • Decide whether to perform a context switch.
    • When performing a context switch, restore the context being context switched in.

This work is conceptually the same across architectures, but the details are completely different:

  • The registers to save and restore.
  • The processor instructions to perform the work.
  • The numbering of the exceptions.
  • etc.

It thus needs an architecture-specific implementation, called the interrupt/exception stub.

Another issue is that the kernel defines the signature of ISRs as:

void (*isr)(void *parameter)

Architectures do not have a consistent or native way of handling parameters to an ISR. As such there are two commonly used methods for handling the parameter.

  • Using some architecture defined mechanism, the parameter value is forced in the stub. This is commonly found in X86-based architectures.
  • The parameters to the ISR are inserted and tracked via a separate table requiring the architecture to discover at runtime which interrupt is executing. A common interrupt handler demuxer is installed for all entries of the real interrupt vector table, which then fetches the device’s ISR and parameter from the separate table. This approach is commonly used in the ARC and ARM architectures via the CONFIG_GEN_ISR_TABLES implementation. You can find examples of the stubs by looking at _interrupt_enter() in x86, _IntExit() in ARM, _isr_wrapper() in ARM, or the full implementation description for ARC in arch/arc/core/isr_wrapper.S.

Each architecture also has to implement primitives for interrupt control:


IRQ_CONNECT is a macro that uses assembler and/or linker script tricks to connect interrupts at build time, saving boot time and text size.

The vector table should contain a handler for each interrupt and exception that can possibly occur. The handler can be as simple as a spinning loop. However, we strongly suggest that handlers at least print some debug information. The information helps figuring out what went wrong when hitting an exception that is a fault, like divide-by-zero or invalid memory access, or an interrupt that is not expected (spurious interrupt). See the ARM implementation in arch/arm/core/fault.c for an example.

Thread Context Switching

Multi-threading is the basic purpose to have a kernel at all. Zephyr supports two types of threads: preemptible and cooperative.

Two crucial concepts when writing an architecture port are the following:

  • Cooperative threads run at a higher priority than preemptible ones, and always preempt them.
  • After handling an interrupt, if a cooperative thread was interrupted, the kernel always goes back to running that thread, since it is not preemptible.

A context switch can happen in several circumstances:

  • When a thread executes a blocking operation, such as taking a semaphore that is currently unavailable.
  • When a preemptible thread unblocks a thread of higher priority by releasing the object on which it was blocked.
  • When an interrupt unblocks a thread of higher priority than the one currently executing, if the currently executing thread is preemptible.
  • When a thread runs to completion.
  • When a thread causes a fatal exception and is removed from the running threads. For example, referencing invalid memory,

Therefore, the context switching must thus be able to handle all these cases.

The kernel keeps the next thread to run in a “cache”, and thus the context switching code only has to fetch from that cache to select which thread to run.

There are two types of context switches: cooperative and preemptive.

  • A cooperative context switch happens when a thread willfully gives the control to another thread. There are two cases where this happens
    • When a thread explicitly yields.
    • When a thread tries to take an object that is currently unavailable and is willing to wait until the object becomes available.
  • A preemptive context switch happens either because an ISR or a thread causes an operation that schedules a thread of higher priority than the one currently running, if the currently running thread is preemptible. An example of such an operation is releasing an object on which the thread of higher priority was waiting.


Control is never taken from cooperative thread when one of them is the running thread.

A cooperative context switch is always done by having a thread call the _Swap() kernel internal symbol. When _Swap is called, the kernel logic knows that a context switch has to happen: _Swap does not check to see if a context switch must happen. Rather, _Swap decides what thread to context switch in. _Swap is called by the kernel logic when an object being operated on is unavailable, and some thread yielding/sleeping primitives.


On x86 and Nios2, _Swap is generic enough and the architecture flexible enough that _Swap can be called when exiting an interrupt to provoke the context switch. This should not be taken as a rule, since neither the ARM Cortex-M or ARCv2 port do this.

Since _Swap is cooperative, the caller-saved registers from the ABI are already on the stack. There is no need to save them in the k_thread structure.

A context switch can also be performed preemptively. This happens upon exiting an ISR, in the kernel interrupt exit stub:

  • _interrupt_enter on x86 after the handler is called.
  • _IntExit on ARM.
  • _firq_exit and _rirq_exit on ARCv2.

In this case, the context switch must only be invoked when the interrupted thread was preemptible, not when it was a cooperative one, and only when the current interrupt is not nested.

The kernel also has the concept of “locking the scheduler”. This is a concept similar to locking the interrupts, but lighter-weight since interrupts can still occur. If a thread has locked the scheduler, is it temporarily non-preemptible.

So, the decision logic to invoke the context switch when exiting an interrupt is simple:

  • If the interrupted thread is not preemptible, do not invoke it.
  • Else, fetch the cached thread from the ready queue, and:
    • If the cached thread is not the current thread, invoke the context switch.
    • Else, do not invoke it.

This is simple, but crucial: if this is not implemented correctly, the kernel will not function as intended and will experience bizarre crashes, mostly due to stack corruption.


If running a coop-only system, i.e. if CONFIG_NUM_PREEMPT_PRIORITIES is 0, no preemptive context switch ever happens. The interrupt code can be optimized to not take any scheduling decision when this is the case.

Thread Creation and Termination

To start a new thread, a stack frame must be constructed so that the context switch can pop it the same way it would pop one from a thread that had been context switched out. This is to be implemented in an architecture-specific _new_thread internal routine.

The thread entry point is also not to be called directly, i.e. it should not be set as the PC for the new thread. Rather it must be wrapped in _thread_entry. This means that the PC in the stack frame shall be set to _thread_entry, and the thread entry point shall be passed as the first parameter to _thread_entry. The specifics of this depend on the ABI.

The need for an architecture-specific thread termination implementation depends on the architecture. There is a generic implementation, but it might not work for a given architecture.

One reason that has been encountered for having an architecture-specific implementation of thread termination is that aborting a thread might be different if aborting because of a graceful exit or because of an exception. This is the case for ARM Cortex-M, where the CPU has to be taken out of handler mode if the thread triggered a fatal exception, but not if the thread gracefully exits its entry point function.

This means implementing an architecture-specific version of k_thread_abort(), and setting the Kconfig option CONFIG_ARCH_HAS_THREAD_ABORT as needed for the architecture (e.g. see arch/arm//core/cortex_m/Kconfig).

Device Drivers

The kernel requires very few hardware devices to function. In theory, the only required device is the interrupt controller, since the kernel can run without a system clock. In practice, to get access to most, if not all, of the sanity check test suite, a system clock is needed as well. Since these two are usually tied to the architecture, they are part of the architecture port.

Interrupt Controllers

There can be significant differences between the interrupt controllers and the interrupt concepts across architectures.

For example, x86 has the concept of an IDT and different interrupt controllers. Although modern systems mostly standardized on the APIC, some small Quark-based systems use the MVIC. Also, the position of an interrupt in the IDT determines its priority.

On the other hand, the ARM Cortex-M has the NVIC as part of the architecture definition. There is no need for an IDT-like table that is separate from the NVIC vector table. The position in the table has nothing to do with priority of an IRQ: priorities are programmable per-entry.

The ARCv2 has its interrupt unit as part of the architecture definition, which is somewhat similar to the NVIC. However, where ARC defines interrupts as having a one-to-one mapping between exception and interrupt numbers (i.e. exception 1 is IRQ1, and device IRQs start at 16), ARM has IRQ0 being equivalent to exception 16 (and weirdly enough, exception 1 can be seen as IRQ-15).

All these differences mean that very little, if anything, can be shared between architectures with regards to interrupt controllers.

System Clock

x86 has APIC timers and the HPET as part of its architecture definition. ARM Cortex-M has the SYSTICK exception. Finally, ARCv2 has the timer0/1 device.

Kernel timeouts are handled in the context of the system clock timer driver’s interrupt handler.

Tickless Idle

The kernel has support for tickless idle. Tickless idle is the concept where no system clock timer interrupt is to be delivered to the CPU when the kernel is about to go idle and the closest timeout expiry is passed a certain threshold. When this condition happens, the system clock is reprogrammed far in the future instead of for a periodic tick. For this to work, the system clock timer driver must support it.

Tickless idle is optional but strongly recommended to achieve low-power consumption.

The kernel has built-in support for going into tickless idle.

The system clock timer driver must implement some hooks to support tickless idle. See existing drivers for examples.

The interrupt entry stub (_interrupt_enter, _isr_wrapper) needs to be adapted to handle exiting tickless idle. See examples in the code for existing architectures.

Console Over Serial Line

There is one other device that is almost a requirement for an architecture port, since it is so useful for debugging. It is a simple polling, output-only, serial port driver on which to send the console (printk, printf) output.

It is not required, and a RAM console (CONFIG_RAM_CONSOLE) can be used to send all output to a circular buffer that can be read by a debugger instead.

Utility Libraries

The kernel depends on a few functions that can be implemented with very few instructions or in a lock-less manner in modern processors. Those are thus expected to be implemented as part of an architecture port.

  • Atomic operators.
  • Find-least-significant-bit-set and find-most-significant-bit-set.
    • If instructions do not exist for a given architecture, it is always possible to implement these functions as generic C functions.

It is possible to use compiler built-ins to implement these, but be careful they use the required compiler barriers.

CPU Idling/Power Management

The kernel provides support for CPU power management with two functions: k_cpu_idle() and k_cpu_atomic_idle().

k_cpu_idle() can be as simple as calling the power saving instruction for the architecture with interrupts unlocked, for example hlt on x86, wfi or wfe on ARM, sleep on ARC. This function can be called in a loop within a context that does not care if it get interrupted or not by an interrupt before going to sleep. There are basically two scenarios when it is correct to use this function:

  • In a single-threaded system, in the only thread when the thread is not used for doing real work after initialization, i.e. it is sitting in a loop doing nothing for the duration of the application.
  • In the idle thread.

k_cpu_atomic_idle(), on the other hand, must be able to atomically re-enable interrupts and invoke the power saving instruction. It can thus be used in real application code, again in single-threaded systems.

Normally, idling the CPU should be left to the idle thread, but in some very special scenarios, these APIs can be used by applications.

Both functions must exist for a given architecture. However, the implementation can be simply the following steps, if desired:

  1. unlock interrupts
  2. NOP

However, a real implementation is strongly recommended.

Fault Management

Each architecture provides two fatal error handlers:

  • _NanoFatalErrorHandler, called by software for unrecoverable errors.
  • _SysFatalErrorHandler, which makes the decision on how to handle the thread where the error is generated, most likely by terminating it.

See the current architecture implementations for examples.

Toolchain and Linking

Toolchain support has to be added to the build system.

Some architecture-specific definitions are needed in include/toolchain/gcc.h. See what exists in that file for currently supported architectures.

Each architecture also needs its own linker script, even if most sections can be derived from the linker scripts of other architectures. Some sections might be specific to the new architecture, for example the SCB section on ARM and the IDT section on x86.

Hardware Stack Protection

This option uses hardware features to generate a fatal error if a thread in supervisor mode overflows its stack. This is useful for debugging, although for a couple reasons, you can’t reliably make any assertions about the state of the system after this happens:

  • The kernel could have been inside a critical section when the overflow occurs, leaving important global data structures in a corrupted state.
  • For systems that implement stack protection using a guard memory region, it’s possible to overshoot the guard and corrupt adjacent data structures before the hardware detects this situation.

To enable the CONFIG_HW_STACK_PROTECTION feature, the system must provide some kind of hardware-based stack overflow protection, and enable the CONFIG_ARCH_HAS_STACK_PROTECTION option.

There are no C APIs that need to be implemented to support stack protection, and it’s entirely implemented within the arch/ code. However in most cases (such as if a guard region needs to be defined) the architecture will need to declare its own versions of the K_THREAD_STACK macros in arch/cpu.h:


For systems that implement stack protection using a Memory Protection Unit (MPU) or Memory Management Unit (MMU), this is typically done by declaring a guard memory region immediately before the stack area.

  • On MMU systems, this guard area is an entire page whose permissions in the page table will generate a fault on writes. This page needs to be configured in the arch’s _new_thread() function.
  • On MPU systems, one of the MPU regions needs to be reserved for the thread stack guard area, whose size should be minimized. The region in the MPU should be reconfigured on context switch such that the guard region for the incoming thread is not writable.

User Mode Threads

To support user mode threads, several kernel-to-arch APIs need to be implemented, and the system must enable the CONFIG_ARCH_HAS_USERSPACE option. Please see the documentation for each of these functions for more details:

  • _arch_buffer_validate() to test whether the current thread has access permissions to a particular memory region
  • _arch_user_mode_enter() which will irreversibly drop a supervisor thread to user mode privileges. The stack must be wiped.
  • _arch_syscall_oops() which generates a kernel oops when system call parameters can’t be validated, in such a way that the oops appears to be generated from where the system call was invoked in the user thread
  • _arch_syscall_invoke0() through _arch_syscall_invoke6() invoke a system call with the appropriate number of arguments which must all be passed in during the privilege elevation via registers.
  • _arch_is_user_context() return nonzero if the CPU is currently running in user mode
  • _arch_mem_domain_max_partitions_get() which indicates the max number of regions for a memory domain. MMU systems have an unlimited amount, MPU systems have constraints on this.
  • _arch_mem_domain_partition_remove() Remove a partition from a memory domain if the currently executing thread was part of that domain.
  • _arch_mem_domain_destroy() Reset the thread’s memory domain configuration

In addition to implementing these APIs, there are some other tasks as well:

  • _new_thread() needs to spawn threads with K_USER in user mode
  • On context switch, the outgoing thread’s stack memory should be marked inaccessible to user mode by making the appropriate configuration changes in the memory management hardware.. The incoming thread’s stack memory should likewise be marked as accessible. This ensures that threads can’t mess with other thread stacks.
  • On context switch, the system needs to switch between memory domains for the incoming and outgoing threads.
  • Thread stack areas must include a kernel stack region. This should be inaccessible to user threads at all times. This stack will be used when system calls are made. This should be fixed size for all threads, and must be large enough to handle any system call.
  • A software interrupt or some kind of privilege elevation mechanism needs to be established. This is closely tied to how the _arch_syscall_invoke macros are implemented. On system call, the appropriate handler function needs to be looked up in _k_syscall_table. Bad system call IDs should jump to the K_SYSCALL_BAD handler. Upon completion of the system call, care must be taken not to leak any register state back to user mode.