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

  • ARMv7-M ISA with Thumb2 instruction set and ARM Embedded ABI (aeabi)

  • ARCv2 ISA

For information on Kconfig configuration, see Setting Kconfig configuration values. Architectures use a Kconfig configuration scheme similar to boards.

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:

Note

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/aarch32/cortex_m/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.

Note

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.

Note

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.

Note

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/aarch32/cortex_m/Kconfig).

Thread Local Storage

To enable thread local storage on a new architecture:

  1. Implement arch_tls_stack_setup() to setup the TLS storage area in stack. Refer to the toolchain documentation on how the storage area needs to be structured. Some helper functions can be used:

    • Function z_tls_data_size() returns the size needed for thread local variables (excluding any extra data required by toolchain and architecture).

    • Function z_tls_copy() prepares the TLS storage area for thread local variables. This only copies the variable themselves and does not do architecture and/or toolchain specific data.

  2. In the context switching, grab the tls field inside the new thread’s struct k_thread and put it into an appropriate register (or some other variable) for access to the TLS storage area. Refer to toolchain and architecture documentation on which registers to use.

  3. In kconfig, add select CONFIG_ARCH_HAS_THREAD_LOCAL_STORAGE to kconfig related to the new architecture.

  4. Run the tests/kernel/threads/tls to make sure the new code works.

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. 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: arch_cpu_idle() and arch_cpu_atomic_idle().

arch_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.

arch_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

In the event of an unhandled CPU exception, the architecture code must call into z_fatal_error(). This function dumps out architecture-agnostic information and makes a policy decision on what to do next by invoking k_sys_fatal_error(). This function can be overridden to implement application-specific policies that could include locking interrupts and spinning forever (the default implementation) or even powering off the system (if supported).

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.

Memory Management

If the target platform enables paging and requires drivers to memory-map their I/O regions, CONFIG_MMU needs to be enabled and the arch_mem_map() API implemented.

Stack Objects

The presence of memory protection hardware affects how stack objects are created. All architecture ports must specify the required alignment of the stack pointer, which is some combination of CPU and ABI requirements. This is defined in architecture headers with ARCH_STACK_PTR_ALIGN and is typically something small like 4, 8, or 16 bytes.

Two types of thread stacks exist:

  • “kernel” stacks defined with K_KERNEL_STACK_DEFINE() and related APIs, which can host kernel threads running in supervisor mode or used as the stack for interrupt/exception handling. These have significantly relaxed alignment requirements and use less reserved data. No memory is reserved for prvilege elevation stacks.

  • “thread” stacks which typically use more memory, but are capable of hosting thread running in user mode, as well as any use-cases for kernel stacks.

If CONFIG_USERSPACE is not enabled, “thread” and “kernel” stacks are equivalent.

Additional macros may be defined in the architecture layer to specify the alignment of the base of stack objects, any reserved data inside the stack object not used for the thread’s stack buffer, and how to round up stack sizes to support user mode threads. In the absence of definitions some defaults are assumed:

  • ARCH_KERNEL_STACK_RESERVED: default no reserved space

  • ARCH_THREAD_STACK_RESERVED: default no reserved space

  • ARCH_KERNEL_STACK_OBJ_ALIGN: default align to ARCH_STACK_PTR_ALIGN

  • ARCH_THREAD_STACK_OBJ_ALIGN: default align to ARCH_STACK_PTR_ALIGN

  • ARCH_THREAD_STACK_SIZE_ALIGN: default round up to ARCH_STACK_PTR_ALIGN

All stack creation macros are defined in terms of these.

Stack objects all have the following layout, with some regions potentially zero-sized depending on configuration. There are always two main parts: reserved memory at the beginning, and then the stack buffer itself. The bounds of some areas can only be determined at runtime in the context of its associated thread object. Other areas are entirely computable at build time.

Some architectures may need to carve-out reserved memory at runtime from the stack buffer, instead of unconditionally reserving it at build time, or to supplement an existing reserved area (as is the case with the ARM FPU). Such carve-outs will always be tracked in thread.stack_info.start. The region specified by thread.stack_info.start and thread.stack_info.size is always fully accessible by a user mode thread. thread.stack_info.delta denotes an offset which can be used to compute the initial stack pointer from the very end of the stack object, taking into account storage for TLS and ASLR random offsets.

+---------------------+ <- thread.stack_obj
| Reserved Memory     | } K_(THREAD|KERNEL)_STACK_RESERVED
+---------------------+
| Carved-out memory   |
|.....................| <- thread.stack_info.start
| Unused stack buffer |
|                     |
|.....................| <- thread's current stack pointer
| Used stack buffer   |
|                     |
|.....................| <- Initial stack pointer. Computable
| ASLR Random offset  |      with thread.stack_info.delta
+---------------------| <- thread.userspace_local_data
| Thread-local data   |
+---------------------+ <- thread.stack_info.start +
                             thread.stack_info.size

At present, Zephyr does not support stacks that grow upward.

No Memory Protection

If no memory protection is in use, then the defaults are sufficient.

HW-based stack overflow detection

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.

Two forms of HW-based stack overflow detection are supported: dedicated CPU features for this purpose, or special read-only guard regions immediately preceding stack buffers.

CONFIG_HW_STACK_PROTECTION only catches stack overflows for supervisor threads. This is not required to catch stack overflow from user threads; CONFIG_USERSPACE is orthogonal.

This feature only detects supervisor mode stack overflows, including stack overflows when handling system calls. It doesn’t guarantee that the kernel has not been corrupted. Any stack overflow in supervisor mode should be treated as a fatal error, with no assertions about the integrity of the overall system possible.

Stack overflows in user mode are recoverable (from the kernel’s perspective) and require no special configuration; CONFIG_HW_STACK_PROTECTION only applies to catching overflows when the CPU is in sueprvisor mode.

CPU-based stack overflow detection

If we are detecting stack overflows in supervisor mode via special CPU registers (like ARM’s SPLIM), then the defaults are sufficient.

Guard-based stack overflow detection

We are detecting supervisor mode stack overflows via special memory protection region located immediately before the stack buffer that generates an exception on write. Reserved memory will be used for the guard region.

ARCH_KERNEL_STACK_RESERVED should be defined to the minimum size of a memory protection region. On most ARM CPUs this is 32 bytes. ARCH_KERNEL_STACK_OBJ_ALIGN should also be set to the required alignment for this region.

MMU-based systems should not reserve RAM for the guard region and instead simply leave an non-present virtual page below every stack when it is mapped into the address space. The stack object will still need to be properly aligned and sized to page granularity.

+-----------------------------+ <- thread.stack_obj
| Guard reserved memory       | } K_KERNEL_STACK_RESERVED
+-----------------------------+
| Guard carve-out             |
|.............................| <- thread.stack_info.start
| Stack buffer                |
.                             .

Guard carve-outs for kernel stacks are uncommon and should be avoided if possible. They tend to be needed for two situations:

  • The same stack may be re-purposed to host a user thread, in which case the guard is unnecessary and shouldn’t be unconditionally reserved. This is the case when privilege elevation stacks are not inside the stack object.

  • The required guard size is variable and depends on context. For example, some ARM CPUs have lazy floating point stacking during exceptions and may decrement the stack pointer by a large value without writing anything, completely overshooting a minimally-sized guard and corrupting adjacent memory. Rather than unconditionally reserving a larger guard, the extra memory is carved out if the thread uses floating point.

User mode enabled

Enabling user mode activates two new requirements:

  • A separate fixed-sized privilege mode stack, specified by CONFIG_PRIVILEGED_STACK_SIZE, must be allocated that the user thread cannot access. It is used as the stack by the kernel when handling system calls. If stack guards are implemented, a stack guard region must be able to be placed before it, with support for carve-outs if necessary.

  • The memory protection hardware must be able to program a region that exactly covers the thread’s stack buffer, tracked in thread.stack_info. This implies that ARCH_THREAD_STACK_SIZE_ADJUST() will need to round up the requested stack size so that a region may cover it, and that ARCH_THREAD_STACK_OBJ_ALIGN() is also specified per the granularity of the memory protection hardware.

This becomes more complicated if the memory protection hardware requires that all memory regions be sized to a power of two, and aligned to their own size. This is common on older MPUs and is known with CONFIG_MPU_REQUIRES_POWER_OF_TWO_ALIGNMENT.

thread.stack_info always tracks the user-accessible part of the stack object, it must always be correct to program a memory protection region with user access using the range stored within.

Non power-of-two memory region requirements

On systems without power-of-two region requirements, the reserved memory area for threads stacks defined by K_THREAD_STACK_RESERVED may be used to contain the privilege mode stack. The layout could be something like:

+------------------------------+ <- thread.stack_obj
| Other platform data          |
+------------------------------+
| Guard region (if enabled)    |
+------------------------------+
| Guard carve-out (if needed)  |
|..............................|
| Privilege elevation stack    |
+------------------------------| <- thread.stack_obj +
| Stack buffer                 |      K_THREAD_STACK_RESERVED =
.                              .      thread.stack_info.start

The guard region, and any carve-out (if needed) would be configured as a read-only region when the thread is created.

  • If the thread is a supervisor thread, the privilege elevation region is just extra stack memory. An overflow will eventually crash into the guard region.

  • If the thread is running in user mode, a memory protection region will be configured to allow user threads access to the stack buffer, but nothing before or after it. An overflow in user mode will crash into the privilege elevation stack, which the user thread has no access to. An overflow when handling a system call will crash into the guard region.

On an MMU system there should be no physical guards; the privilege mode stack will be mapped into kernel memory, and the stack buffer in the user part of memory, each with non-present virtual guard pages below them to catch runtime stack overflows.

Other platform data may be stored before the guard region, but this is highly discouraged if such data could be stored in thread.arch somewhere.

ARCH_THREAD_STACK_RESERVED will need to be defined to capture the size of the reserved region containing platform data, privilege elevation stacks, and guards. It must be appropriately sized such that an MPU region to grant user mode access to the stack buffer can be placed immediately after it.

Power-of-two memory region requirements

Thread stack objects must be sized and aligned to the same power of two, without any reserved memory to allow efficient packing in memory. Thus, any guards in the thread stack must be completely carved out, and the privilege elevation stack must be allocated elsewhere.

ARCH_THREAD_STACK_SIZE_ADJUST() and ARCH_THREAD_STACK_OBJ_ALIGN() should both be defined to Z_POW2_CEIL(). K_THREAD_STACK_RESERVED must be 0.

For the privilege stacks, the CONFIG_GEN_PRIV_STACKS must be, enabled. For every thread stack found in the system, a corresponding fixed- size kernel stack used for handling system calls is generated. The address of the privilege stacks can be looked up quickly at runtime based on the thread stack address using z_priv_stack_find(). These stacks are laid out the same way as other kernel-only stacks.

+-----------------------------+ <- z_priv_stack_find(thread.stack_obj)
| Reserved memory             | } K_KERNEL_STACK_RESERVED
+-----------------------------+
| Guard carve-out (if needed) |
|.............................|
| Privilege elevation stack   |
|                             |
+-----------------------------+ <- z_priv_stack_find(thread.stack_obj) +
                                     K_KERNEL_STACK_RESERVED +
                                     CONFIG_PRIVILEGED_STACK_SIZE

+-----------------------------+ <- thread.stack_obj
| MPU guard carve-out         |
| (supervisor mode only)      |
|.............................| <- thread.stack_info.start
| Stack buffer                |
.                             .

The guard carve-out in the thread stack object is only used if the thread is running in supervisor mode. If the thread drops to user mode, there is no guard and the entire object is used as the stack buffer, with full access to the associated user mode thread and thread.stack_info updated appropriately.

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.

Some architectures may need to update software memory management structures or modify hardware registers on another CPU when memory domain APIs are invoked. If so, CONFIG_ARCH_MEM_DOMAIN_SYNCHRONOUS_API must be selected by the architecture and some additional APIs must be implemented. This is common on MMU systems and uncommon on MPU systems:

  • arch_mem_domain_thread_add()

  • arch_mem_domain_thread_remove()

  • arch_mem_domain_partition_add()

  • arch_mem_domain_partition_remove()

  • arch_mem_domain_destroy()

Please see the doxygen documentation of these APIs for details.

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.

API Reference

Timing

group arch-timing

Functions

void arch_busy_wait(uint32_t usec_to_wait)

Architecture-specific implementation of busy-waiting

Parameters
  • usec_to_wait: Wait period, in microseconds

uint32_t arch_k_cycle_get_32(void)

Obtain the current cycle count, in units that are hardware-specific

See

k_cycle_get_32()

Threads

group arch-threads

Defines

ARCH_KERNEL_STACK_RESERVED

MPU guard size for kernel-only stacks.

If MPU stack guards are used to catch stack overflows, specify the amount of space reserved in kernel stack objects. If guard sizes are context dependent, this should be in the minimum guard size, with remaining space carved out if needed.

Optional definition, defaults to 0.

See

K_KERNEL_STACK_RESERVED

ARCH_KERNEL_STACK_OBJ_ALIGN

Required alignment of the lowest address of a kernel-only stack.

Functions

void arch_new_thread(struct k_thread *thread, k_thread_stack_t *stack, char *stack_ptr, k_thread_entry_t entry, void *p1, void *p2, void *p3)

Handle arch-specific logic for setting up new threads

The stack and arch-specific thread state variables must be set up such that a later attempt to switch to this thread will succeed and we will enter z_thread_entry with the requested thread and arguments as its parameters.

At some point in this function’s implementation, z_setup_new_thread() must be called with the true bounds of the available stack buffer within the thread’s stack object.

The provided stack pointer is guaranteed to be properly aligned with respect to the CPU and ABI requirements. There may be space reserved between the stack pointer and the bounds of the stack buffer for initial stack pointer randomization and thread-local storage.

Fields in thread->base will be initialized when this is called.

Parameters
  • thread: Pointer to uninitialized struct k_thread

  • stack: Pointer to the stack object

  • stack_ptr: Aligned initial stack pointer

  • entry: Thread entry function

  • p1: 1st entry point parameter

  • p2: 2nd entry point parameter

  • p3: 3rd entry point parameter

void arch_switch(void *switch_to, void **switched_from)

Cooperatively context switch

Architectures have considerable leeway on what the specific semantics of the switch handles are, but optimal implementations should do the following if possible:

1) Push all thread state relevant to the context switch to the current stack 2) Update the switched_from parameter to contain the current stack pointer, after all context has been saved. switched_from is used as an output- only parameter and its current value is ignored (and can be NULL, see below). 3) Set the stack pointer to the value provided in switch_to 4) Pop off all thread state from the stack we switched to and return.

Some arches may implement thread->switch handle as a pointer to the thread itself, and save context somewhere in thread->arch. In this case, on initial context switch from the dummy thread, thread->switch handle for the outgoing thread is NULL. Instead of dereferencing switched_from all the way to get the thread pointer, subtract ___thread_t_switch_handle_OFFSET to obtain the thread pointer instead. That is, such a scheme would have behavior like (in C pseudocode):

void arch_switch(void *switch_to, void **switched_from) { struct k_thread *new = switch_to; struct k_thread *old = CONTAINER_OF(switched_from, struct k_thread, switch_handle);

// save old context… *switched_from = old; // restore new context… }

Note that, regardless of the underlying handle representation, the incoming switched_from pointer MUST be written through with a non-NULL value after all relevant thread state has been saved. The kernel uses this as a synchronization signal to be able to wait for switch completion from another CPU.

Parameters
  • switch_to: Incoming thread’s switch handle

  • switched_from: Pointer to outgoing thread’s switch handle storage location, which may be updated.

void arch_switch_to_main_thread(struct k_thread *main_thread, char *stack_ptr, k_thread_entry_t _main)

Custom logic for entering main thread context at early boot

Used by architectures where the typical trick of setting up a dummy thread in early boot context to “switch out” of isn’t workable.

Parameters
  • main_thread: main thread object

  • stack_ptr: Initial stack pointer

  • _main: Entry point for application main function.

int arch_float_disable(struct k_thread *thread)

Disable floating point context preservation.

The function is used to disable the preservation of floating point context information for a particular thread.

Note

For ARM architecture, disabling floating point preservation may only be requested for the current thread and cannot be requested in ISRs.

Return Value
  • 0: On success.

  • -EINVAL: If the floating point disabling could not be performed.

group arch-tls

Functions

size_t arch_tls_stack_setup(struct k_thread *new_thread, char *stack_ptr)

Setup Architecture-specific TLS area in stack.

This sets up the stack area for thread local storage. The structure inside in area is architecture specific.

Return

Number of bytes taken by the TLS area

Parameters
  • new_thread: New thread object

  • stack_ptr: Stack pointer

Power Management

group arch-pm

Functions

FUNC_NORETURN void arch_system_halt(unsigned int reason)

Halt the system, optionally propagating a reason code

void arch_cpu_idle(void)

Power save idle routine.

This function will be called by the kernel idle loop or possibly within an implementation of z_sys_power_save_idle in the kernel when the ‘_sys_power_save_flag’ variable is non-zero.

Architectures that do not implement power management instructions may immediately return, otherwise a power-saving instruction should be issued to wait for an interrupt.

Note

The function is expected to return after the interrupt that has caused the CPU to exit power-saving mode has been serviced, although this is not a firm requirement.

See

k_cpu_idle()

void arch_cpu_atomic_idle(unsigned int key)

Atomically re-enable interrupts and enter low power mode.

The requirements for arch_cpu_atomic_idle() are as follows:

  1. Enabling interrupts and entering a low-power mode needs to be atomic, i.e. there should be no period of time where interrupts are enabled before the processor enters a low-power mode. See the comments in k_lifo_get(), for example, of the race condition that occurs if this requirement is not met.

  2. After waking up from the low-power mode, the interrupt lockout state must be restored as indicated in the ‘key’ input parameter.

See

k_cpu_atomic_idle()

Parameters

Symmetric Multi-Processing

group arch-smp

Typedefs

typedef FUNC_NORETURN void (*arch_cpustart_t)(void *data)

Per-cpu entry function

Parameters
  • context: parameter, implementation specific

Functions

void arch_start_cpu(int cpu_num, k_thread_stack_t *stack, int sz, arch_cpustart_t fn, void *arg)

Start a numbered CPU on a MP-capable system.

This starts and initializes a specific CPU. The main thread on startup is running on CPU zero, other processors are numbered sequentially. On return from this function, the CPU is known to have begun operating and will enter the provided function. Its interrupts will be initialized but disabled such that irq_unlock() with the provided key will work to enable them.

Normally, in SMP mode this function will be called by the kernel initialization and should not be used as a user API. But it is defined here for special-purpose apps which want Zephyr running on one core and to use others for design-specific processing.

Parameters
  • cpu_num: Integer number of the CPU

  • stack: Stack memory for the CPU

  • sz: Stack buffer size, in bytes

  • fn: Function to begin running on the CPU.

  • arg: Untyped argument to be passed to “fn”

struct _cpu *arch_curr_cpu(void)

Return the CPU struct for the currently executing CPU

void arch_sched_ipi(void)

Broadcast an interrupt to all CPUs

This will invoke z_sched_ipi() on other CPUs in the system.

Interrupts

group arch-irq

Functions

bool arch_is_in_isr(void)

Test if the current context is in interrupt context

XXX: This is inconsistently handled among arches wrt exception context See: #17656

Return

true if we are in interrupt context

unsigned int arch_irq_lock(void)

Lock interrupts on the current CPU

See

irq_lock()

void arch_irq_unlock(unsigned int key)

Unlock interrupts on the current CPU

See

irq_unlock()

bool arch_irq_unlocked(unsigned int key)

Test if calling arch_irq_unlock() with this key would unlock irqs

Return

true if interrupts were unlocked prior to the arch_irq_lock() call that produced the key argument.

Parameters

void arch_irq_disable(unsigned int irq)

Disable the specified interrupt line

See

irq_disable()

void arch_irq_enable(unsigned int irq)

Enable the specified interrupt line

See

irq_enable()

int arch_irq_is_enabled(unsigned int irq)

Test if an interrupt line is enabled

See

irq_is_enabled()

int arch_irq_connect_dynamic(unsigned int irq, unsigned int priority, void (*routine)(const void *parameter), const void *parameter, uint32_t flags, )

Arch-specific hook to install a dynamic interrupt.

Return

The vector assigned to this interrupt

Parameters
  • irq: IRQ line number

  • priority: Interrupt priority

  • routine: Interrupt service routine

  • parameter: ISR parameter

  • flags: Arch-specific IRQ configuration flag

Userspace

group arch-userspace

Functions

uintptr_t arch_syscall_invoke0(uintptr_t call_id)

Invoke a system call with 0 arguments.

No general-purpose register state other than return value may be preserved when transitioning from supervisor mode back down to user mode for security reasons.

It is required that all arguments be stored in registers when elevating privileges from user to supervisor mode.

Processing of the syscall takes place on a separate kernel stack. Interrupts should be enabled when invoking the system call marshallers from the dispatch table. Thread preemption may occur when handling system calls.

Call ids are untrusted and must be bounds-checked, as the value is used to index the system call dispatch table, containing function pointers to the specific system call code.

Return

Return value of the system call. Void system calls return 0 here.

Parameters
  • call_id: System call ID

uintptr_t arch_syscall_invoke1(uintptr_t arg1, uintptr_t call_id)

Invoke a system call with 1 argument.

See

arch_syscall_invoke0()

Return

Return value of the system call. Void system calls return 0 here.

Parameters
  • arg1: First argument to the system call.

  • call_id: System call ID, will be bounds-checked and used to reference kernel-side dispatch table

uintptr_t arch_syscall_invoke2(uintptr_t arg1, uintptr_t arg2, uintptr_t call_id)

Invoke a system call with 2 arguments.

See

arch_syscall_invoke0()

Return

Return value of the system call. Void system calls return 0 here.

Parameters
  • arg1: First argument to the system call.

  • arg2: Second argument to the system call.

  • call_id: System call ID, will be bounds-checked and used to reference kernel-side dispatch table

uintptr_t arch_syscall_invoke3(uintptr_t arg1, uintptr_t arg2, uintptr_t arg3, uintptr_t call_id)

Invoke a system call with 3 arguments.

See

arch_syscall_invoke0()

Return

Return value of the system call. Void system calls return 0 here.

Parameters
  • arg1: First argument to the system call.

  • arg2: Second argument to the system call.

  • arg3: Third argument to the system call.

  • call_id: System call ID, will be bounds-checked and used to reference kernel-side dispatch table

uintptr_t arch_syscall_invoke4(uintptr_t arg1, uintptr_t arg2, uintptr_t arg3, uintptr_t arg4, uintptr_t call_id)

Invoke a system call with 4 arguments.

See

arch_syscall_invoke0()

Return

Return value of the system call. Void system calls return 0 here.

Parameters
  • arg1: First argument to the system call.

  • arg2: Second argument to the system call.

  • arg3: Third argument to the system call.

  • arg4: Fourth argument to the system call.

  • call_id: System call ID, will be bounds-checked and used to reference kernel-side dispatch table

uintptr_t arch_syscall_invoke5(uintptr_t arg1, uintptr_t arg2, uintptr_t arg3, uintptr_t arg4, uintptr_t arg5, uintptr_t call_id)

Invoke a system call with 5 arguments.

See

arch_syscall_invoke0()

Return

Return value of the system call. Void system calls return 0 here.

Parameters
  • arg1: First argument to the system call.

  • arg2: Second argument to the system call.

  • arg3: Third argument to the system call.

  • arg4: Fourth argument to the system call.

  • arg5: Fifth argument to the system call.

  • call_id: System call ID, will be bounds-checked and used to reference kernel-side dispatch table

uintptr_t arch_syscall_invoke6(uintptr_t arg1, uintptr_t arg2, uintptr_t arg3, uintptr_t arg4, uintptr_t arg5, uintptr_t arg6, uintptr_t call_id)

Invoke a system call with 6 arguments.

See

arch_syscall_invoke0()

Return

Return value of the system call. Void system calls return 0 here.

Parameters
  • arg1: First argument to the system call.

  • arg2: Second argument to the system call.

  • arg3: Third argument to the system call.

  • arg4: Fourth argument to the system call.

  • arg5: Fifth argument to the system call.

  • arg6: Sixth argument to the system call.

  • call_id: System call ID, will be bounds-checked and used to reference kernel-side dispatch table

bool arch_is_user_context(void)

Indicate whether we are currently running in user mode

Return

true if the CPU is currently running with user permissions

int arch_mem_domain_max_partitions_get(void)

Get the maximum number of partitions for a memory domain.

Return

Max number of partitions, or -1 if there is no limit

int arch_buffer_validate(void *addr, size_t size, int write)

Check memory region permissions.

Given a memory region, return whether the current memory management hardware configuration would allow a user thread to read/write that region. Used by system calls to validate buffers coming in from userspace.

Notes: The function is guaranteed to never return validation success, if the entire buffer area is not user accessible.

The function is guaranteed to correctly validate the permissions of the supplied buffer, if the user access permissions of the entire buffer are enforced by a single, enabled memory management region.

In some architectures the validation will always return failure if the supplied memory buffer spans multiple enabled memory management regions (even if all such regions permit user access).

Warning

0 size buffer has undefined behavior.

Return

nonzero if the permissions don’t match.

Parameters
  • addr: start address of the buffer

  • size: the size of the buffer

  • write: If nonzero, additionally check if the area is writable. Otherwise, just check if the memory can be read.

FUNC_NORETURN void arch_user_mode_enter(k_thread_entry_t user_entry, void *p1, void *p2, void *p3)

Perform a one-way transition from supervisor to kernel mode.

Implementations of this function must do the following:

  • Reset the thread’s stack pointer to a suitable initial value. We do not need any prior context since this is a one-way operation.

  • Set up any kernel stack region for the CPU to use during privilege elevation

  • Put the CPU in whatever its equivalent of user mode is

  • Transfer execution to arch_new_thread() passing along all the supplied arguments, in user mode.

Parameters
  • user_entry: Entry point to start executing as a user thread

  • p1: 1st parameter to user thread

  • p2: 2nd parameter to user thread

  • p3: 3rd parameter to user thread

FUNC_NORETURN void arch_syscall_oops(void *ssf)

Induce a kernel oops that appears to come from a specific location.

Normally, k_oops() generates an exception that appears to come from the call site of the k_oops() itself.

However, when validating arguments to a system call, if there are problems we want the oops to appear to come from where the system call was invoked and not inside the validation function.

Parameters
  • ssf: System call stack frame pointer. This gets passed as an argument to _k_syscall_handler_t functions and its contents are completely architecture specific.

size_t arch_user_string_nlen(const char *s, size_t maxsize, int *err)

Safely take the length of a potentially bad string.

This must not fault, instead the err parameter must have -1 written to it. This function otherwise should work exactly like libc strnlen(). On success *err should be set to 0.

Return

Length of the string, not counting NULL byte, up to maxsize

Parameters
  • s: String to measure

  • maxsize: Max length of the string

  • err: Error value to write

bool arch_mem_coherent(void *ptr)

Detect memory coherence type.

Required when ARCH_HAS_COHERENCE is true. This function returns true if the byte pointed to lies within an architecture-defined “coherence region” (typically implemented with uncached memory) and can safely be used in multiprocessor code without explicit flush or invalidate operations.

Note

The result is for only the single byte at the specified address, this API is not required to check region boundaries or to expect aligned pointers. The expectation is that the code above will have queried the appropriate address(es).

void arch_cohere_stacks(struct k_thread *old_thread, void *old_switch_handle, struct k_thread *new_thread)

Ensure cache coherence prior to context switch.

Required when ARCH_HAS_COHERENCE is true. On cache-incoherent multiprocessor architectures, thread stacks are cached by default for performance reasons. They must therefore be flushed appropriately on context switch. The rules are:

  1. The region containing live data in the old stack (generally the bytes between the current stack pointer and the top of the stack memory) must be flushed to underlying storage so a new CPU that runs the same thread sees the correct data. This must happen before the assignment of the switch_handle field in the thread struct which signals the completion of context switch.

  2. Any data areas to be read from the new stack (generally the same as the live region when it was saved) should be invalidated (and NOT flushed!) in the data cache. This is because another CPU may have run or re-initialized the thread since this CPU suspended it, and any data present in cache will be stale.

  • old_thread The old thread to be flushed before being allowed to run on other CPUs.

  • old_switch_handle The switch handle to be stored into old_thread (it will not be valid until the cache is flushed so is not present yet). This will be NULL if inside z_swap() (because the arch_switch() has not saved it yet).

  • new_thread The new thread to be invalidated before it runs locally.

Note

The kernel will call this function during interrupt exit when a new thread has been chosen to run, and also immediately before entering arch_switch() to effect a code-driven context switch. In the latter case, it is very likely that more data will be written to the old_thread stack region after this function returns but before the completion of the switch. Simply flushing naively here is not sufficient on many architectures and coordination with the arch_switch() implementation is likely required.

Memory Management

group arch-mmu

Functions

int arch_mem_map(void *dest, uintptr_t addr, size_t size, uint32_t flags)

Map physical memory into the virtual address space

This is a low-level interface to mapping pages into the address space. Behavior when providing unaligned addresses/sizes is undefined, these are assumed to be aligned to CONFIG_MMU_PAGE_SIZE.

The core kernel handles all management of the virtual address space; by the time we invoke this function, we know exactly where this mapping will be established. If the page tables already had mappings installed for the virtual memory region, these will be overwritten.

If the target architecture supports multiple page sizes, currently only the smallest page size will be used.

The memory range itself is never accessed by this operation.

This API must be safe to call in ISRs or exception handlers. Calls to this API are assumed to be serialized, and indeed all usage will originate from kernel/mm.c which handles virtual memory management.

This API is part of infrastructure still under development and may change.

See

z_mem_map()

Parameters
  • dest: Page-aligned Destination virtual address to map

  • addr: Page-aligned Source physical address to map

  • size: Page-aligned size of the mapped memory region in bytes

  • flags: Caching, access and control flags, see K_MAP_* macros

Return Value
  • 0: Success

  • -ENOTSUP: Unsupported cache mode with no suitable fallback, or unsupported flags

  • -ENOMEM: Memory for additional paging structures unavailable

void arch_mem_unmap(void *addr, size_t size)

Remove mappings for a provided virtual address range

This is a low-level interface for un-mapping pages from the address space. When this completes, the relevant page table entries will be updated as if no mapping was ever made for that memory range. No previous context needs to be preserved. This function must update mappings in all active page tables.

Behavior when providing unaligned addresses/sizes is undefined, these are assumed to be aligned to CONFIG_MMU_PAGE_SIZE.

Behavior when providing an address range that is not already mapped is undefined.

This function should never require memory allocations for paging structures, and it is not necessary to free any paging structures. Empty page tables due to all contained entries being un-mapped may remain in place.

Implementations must invalidate TLBs as necessary.

This API is part of infrastructure still under development and may change.

Parameters
  • addr: Page-aligned base virtual address to un-map

  • size: Page-aligned region size

Miscellaneous Architecture APIs

group arch-misc

Functions

int arch_printk_char_out(int c)

Early boot console output hook

Definition of this function is optional. If implemented, any invocation of printk() (or logging calls with CONFIG_LOG_MINIMAL which are backed by printk) will default to sending characters to this function. It is useful for early boot debugging before main serial or console drivers come up.

This can be overridden at runtime with __printk_hook_install().

The default __weak implementation of this does nothing.

Return

The character printed

Parameters
  • c: Character to print

void arch_kernel_init(void)

Architecture-specific kernel initialization hook

This function is invoked near the top of _Cstart, for additional architecture-specific setup before the rest of the kernel is brought up.

TODO: Deprecate, most arches are using a prep_c() function to do the same thing in a simpler way

void arch_nop(void)

Do nothing and return. Yawn.