Architecture support

Purpose

arch/ contains all CPU-architecture-specific code: the boot stub, exception/interrupt entry, system-call entry, atomics and locking primitives, page-table types, MMU and TLB management, low-level context-switch, vDSO, kvm host backend, perf event counters, and per-platform peripherals where they don't have a driver subsystem of their own.

22 architectures live here. Anything portable between them lives elsewhere in the tree (include/asm-generic/, lib/, kernel/).

Architectures

Directory	Architecture	Status
alpha/	DEC Alpha	Maintenance only
arc/	Synopsys ARC	Active
arm/	32-bit ARM	Active (legacy hardware)
arm64/	64-bit ARM (AArch64)	First-tier; very active
csky/	C-SKY	Maintenance
hexagon/	Qualcomm Hexagon	Maintenance
loongarch/	LoongArch	Active (newer arch)
m68k/	Motorola 68000	Maintenance, hobbyist
microblaze/	Xilinx MicroBlaze	Active (FPGA SoCs)
mips/	MIPS	Active
nios2/	Altera Nios II	Maintenance
openrisc/	OpenRISC	Active (FPGA, academic)
parisc/	HP PA-RISC	Maintenance
powerpc/	POWER / PowerPC	First-tier
riscv/	RISC-V	First-tier; very active
s390/	IBM Z (System/390 / z/Architecture)	First-tier
sh/	Renesas SuperH	Maintenance
sparc/	SPARC	Maintenance
um/	User-Mode Linux	Active (test/dev)
x86/	x86 / x86_64	First-tier; the most active
xtensa/	Tensilica Xtensa	Active

The first-tier architectures (x86, arm64, riscv, powerpc, s390) get most of the new feature work. Others receive maintenance and review.

Per-arch directory shape

Each arch/<arch>/ typically contains:

arch/<arch>/
├── Kconfig              # Architecture's configuration menu
├── Makefile             # Compilation flags, default cflags
├── boot/                # Bootloader interface, decompressors
├── kernel/              # Trap entry, syscall entry, smp, time, traps, signal, vdso glue, perf
├── mm/                  # Page tables, TLB, fault handler, hugetlb
├── lib/                 # Hand-written or arch-tuned helpers (memcpy, atomics, crypto)
├── include/asm/         # asm/*.h — primitives consumed by the rest of the kernel
├── crypto/              # Hardware crypto primitives (e.g. AES-NI on x86)
├── kvm/                 # Architecture-specific KVM host code
├── platforms/, mach-*/  # Per-board / per-SoC code (mostly arm)
├── configs/             # defconfigs
├── tools/               # arch-specific tools (e.g. relocs)
└── vdso/                # virtual DSO mapped into user space

Concrete examples: arch/x86/, arch/arm64/, arch/riscv/.

What lives in arch and what doesn't

Generic abstractions live in include/asm-generic/. Architectures override only what they need. A short, partial list of what's almost always arch-specific:

• Boot path before the C entry. Decompression, MMU enable, page-table setup, jump to start_kernel.
• Trap and IRQ entry/exit (arch/<arch>/kernel/entry*.S or generic helpers in kernel/entry/).
• System-call dispatch table (varies by arch).
• Atomic operations (arch/<arch>/include/asm/atomic.h and friends).
• Memory barriers and the memory model (arch/<arch>/include/asm/barrier.h).
• Page-table walking and TLB invalidation.
• Context switch (switch_to, __switch_to).
• Per-CPU implementation (arch/<arch>/include/asm/percpu.h).
• vDSO (the small ELF mapped into user space for gettimeofday, clock_gettime, getcpu, etc.).
• Architecture-specific KVM (arch/<arch>/kvm/).
• Perf event PMU drivers.

Common subsystems with arch parts

Subsystem	Arch piece
MMU / TLB	arch/<arch>/mm/
Atomics, barriers	arch/<arch>/include/asm/{atomic,barrier}.h
Locking fast paths	arch/<arch>/include/asm/spinlock*.h
Syscall entry	arch/<arch>/kernel/entry*.S and arch/<arch>/kernel/syscall*.c
KVM	arch/<arch>/kvm/ (paired with Virtualization which covers virt/)
BPF JIT	arch/<arch>/net/bpf_jit*.c
FPU / SIMD save/restore	arch/<arch>/kernel/fpu*
ftrace / kprobes	per-arch hooks in arch/<arch>/kernel/

How it integrates with the rest of the tree

graph LR
    BL[Bootloader] --> ARCH[arch/<arch>/boot]
    ARCH --> START["start_kernel() in init/main.c"]
    START --> GENERIC[Generic kernel code]
    GENERIC -->|asm-generic shims| ASM[arch/<arch>/include/asm/]
    GENERIC -->|"function ptrs"| ARCH_OPS["arch_*() callbacks"]
    DRV[drivers/] -.-> ARCH_OPS
    KVM[virt/kvm + arch/<arch>/kvm/] --> CPUFEAT[CPU feature detection]

Most consumers never touch arch directly — they include <asm/foo.h> or <linux/foo.h>, and the build picks the right backing implementation based on CONFIG_<ARCH> and the per-arch include/asm/.

Key source files (per first-tier arch)

Arch	Bootstrap	Trap entry
x86	arch/x86/boot/, arch/x86/kernel/head_64.S	arch/x86/entry/
arm64	arch/arm64/kernel/head.S	arch/arm64/kernel/entry.S
riscv	arch/riscv/kernel/head.S	arch/riscv/kernel/entry.S
powerpc	arch/powerpc/kernel/head_64.S	arch/powerpc/kernel/
s390	arch/s390/kernel/head64.S	arch/s390/kernel/

Entry points for modification

• Adding a CPU feature flag: each arch has a feature-bits header (arch/x86/include/asm/cpufeatures.h, arch/arm64/include/asm/cpufeature.h, etc.). Wire it through cpu_feature_enabled() and the alternatives framework.
• Adding a syscall: edit arch/<arch>/include/asm/syscall_table.h (or equivalent), then the generic UAPI table in include/uapi/asm-generic/unistd.h.
• New device-tree consumer: handled by drivers/, but board-specific boot fixups for arm/arm64 may go under arch/arm//arch/arm64/.
• New BPF JIT: arch/<arch>/net/bpf_jit*.c. Examples: arch/x86/net/bpf_jit_comp.c, arch/arm64/net/bpf_jit_comp.c.

Related pages

• Init — what runs after the arch stub.
• Memory management — page-table types and TLB flush.
• Kernel core — per-CPU, locking, RCU.
• Virtualization — KVM has architecture-specific halves.