arm64: ilp32: add documentation on the ILP32 ABI for ARM64

ILP32 AARCH64 SYSCALL ABI

=========================

This document describes the ILP32 syscall ABI and where it differs

from the generic compat linux syscall interface.

ILP32 is acronym for memory model which stands for "Integers, Longs and

Pointers are 32-bit". The main purpose of ILP32 in Linux kernel is providing

compatibility with 32-bit legacy code. Also, ILP32 binaries look better in some

performance tests. ARM has AN490 document which coves ILP32 details for ARM64

platform:

http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dai0490a/ar01s01.html

AARCH64/ILP32 userspace may pass garbage in the top halve of w0-w7 registers

(syscall arguments). So top 32 bits are zeroed for them.

Comparing to AARCH32, AARCH64/ILP32 has 64-bit length of following types:

ino_t       is u64 type.

off_t       is s64 type.

blkcnt_t    is s64 type.

fsblkcnt_t  is u64 type.

fsfilcnt_t  is u64 type.

rlim_t      is u64 type.

AARCH64/ILP32 ABI uses standard syscall table which can be found at

include/uapi/asm-generic/unistd.h, with the exceptions listed below.

Syscalls which pass 64-bit values are handled by the code shared from

AARCH32 and pass that value as a pair. Following syscalls are affected:

fadvise64_64()

fallocate()

ftruncate64()

pread64()

pwrite64()

readahead()

sync_file_range()

truncate64()

ptrace() syscall is handled by compat version.

shmat() syscall is handled by non-compat handler as aarch64/ilp32 has no

limitation on 4-pages alignment for shared memory.

statfs() and fstatfs() take the size of struct statfs as an argument.

It is calculated differently in kernel and user spaces. So AARCH32 handlers

are taken to handle it.

struct rt_sigframe is redefined and contains struct compat_siginfo,

as compat syscalls expect, and struct ilp32_ucontext, to handle

AARCH64 register set and 32-bit userspace register representation.

elf_gregset_t is taken from lp64 to handle registers properly.