Documentation/arm64/asymmetric-32bit.rst

======================
Asymmetric 32-bit SoCs
======================

Author: Will Deacon <will@kernel.org>

This document describes the impact of asymmetric 32-bit SoCs on the
execution of 32-bit (``AArch32``) applications.

Date: 2021-05-17

Introduction
============

Some Armv9 SoCs suffer from a big.LITTLE misfeature where only a subset
of the CPUs are capable of executing 32-bit user applications. On such
a system, Linux by default treats the asymmetry as a "mismatch" and
disables support for both the ``PER_LINUX32`` personality and
``execve(2)`` of 32-bit ELF binaries, with the latter returning
``-ENOEXEC``. If the mismatch is detected during late onlining of a
64-bit-only CPU, then the onlining operation fails and the new CPU is
unavailable for scheduling.

Surprisingly, these SoCs have been produced with the intention of
running legacy 32-bit binaries. Unsurprisingly, that doesn't work very
well with the default behaviour of Linux.

It seems inevitable that future SoCs will drop 32-bit support
altogether, so if you're stuck in the unenviable position of needing to
run 32-bit code on one of these transitionary platforms then you would
be wise to consider alternatives such as recompilation, emulation or
retirement. If neither of those options are practical, then read on.

Enabling kernel support
=======================

Since the kernel support is not completely transparent to userspace,
allowing 32-bit tasks to run on an asymmetric 32-bit system requires an
explicit "opt-in" and can be enabled by passing the
``allow_mismatched_32bit_el0`` parameter on the kernel command-line.

For the remainder of this document we will refer to an *asymmetric
system* to mean an asymmetric 32-bit SoC running Linux with this kernel
command-line option enabled.

Userspace impact
================

32-bit tasks running on an asymmetric system behave in mostly the same
way as on a homogeneous system, with a few key differences relating to
CPU affinity.

sysfs
-----

The subset of CPUs capable of running 32-bit tasks is described in
``/sys/devices/system/cpu/aarch32_el0`` and is documented further in
``Documentation/ABI/testing/sysfs-devices-system-cpu``.

**Note:** CPUs are advertised by this file as they are detected and so
late-onlining of 32-bit-capable CPUs can result in the file contents
being modified by the kernel at runtime. Once advertised, CPUs are never
removed from the file.

``execve(2)``
-------------

On a homogeneous system, the CPU affinity of a task is preserved across
``execve(2)``. This is not always possible on an asymmetric system,
specifically when the new program being executed is 32-bit yet the
affinity mask contains 64-bit-only CPUs. In this situation, the kernel
determines the new affinity mask as follows:

  1. If the 32-bit-capable subset of the affinity mask is not empty,
     then the affinity is restricted to that subset and the old affinity
     mask is saved. This saved mask is inherited over ``fork(2)`` and
     preserved across ``execve(2)`` of 32-bit programs.

     **Note:** This step does not apply to ``SCHED_DEADLINE`` tasks.
     See `SCHED_DEADLINE`_.

  2. Otherwise, the cpuset hierarchy of the task is walked until an
     ancestor is found containing at least one 32-bit-capable CPU. The
     affinity of the task is then changed to match the 32-bit-capable
     subset of the cpuset determined by the walk.

  3. On failure (i.e. out of memory), the affinity is changed to the set
     of all 32-bit-capable CPUs of which the kernel is aware.

A subsequent ``execve(2)`` of a 64-bit program by the 32-bit task will
invalidate the affinity mask saved in (1) and attempt to restore the CPU
affinity of the task using the saved mask if it was previously valid.
This restoration may fail due to intervening changes to the deadline
policy or cpuset hierarchy, in which case the ``execve(2)`` continues
with the affinity unchanged.

Calls to ``sched_setaffinity(2)`` for a 32-bit task will consider only
the 32-bit-capable CPUs of the requested affinity mask. On success, the
affinity for the task is updated and any saved mask from a prior
``execve(2)`` is invalidated.

``SCHED_DEADLINE``
------------------

Explicit admission of a 32-bit deadline task to the default root domain
(e.g. by calling ``sched_setattr(2)``) is rejected on an asymmetric
32-bit system unless admission control is disabled by writing -1 to
``/proc/sys/kernel/sched_rt_runtime_us``.

``execve(2)`` of a 32-bit program from a 64-bit deadline task will
return ``-ENOEXEC`` if the root domain for the task contains any
64-bit-only CPUs and admission control is enabled. Concurrent offlining
of 32-bit-capable CPUs may still necessitate the procedure described in
`execve(2)`_, in which case step (1) is skipped and a warning is
emitted on the console.

**Note:** It is recommended that a set of 32-bit-capable CPUs are placed
into a separate root domain if ``SCHED_DEADLINE`` is to be used with
32-bit tasks on an asymmetric system. Failure to do so is likely to
result in missed deadlines.

Cpusets
-------

The affinity of a 32-bit task on an asymmetric system may include CPUs
that are not explicitly allowed by the cpuset to which it is attached.
This can occur as a result of the following two situations:

  - A 64-bit task attached to a cpuset which allows only 64-bit CPUs
    executes a 32-bit program.

  - All of the 32-bit-capable CPUs allowed by a cpuset containing a
    32-bit task are offlined.

In both of these cases, the new affinity is calculated according to step
(2) of the process described in `execve(2)`_ and the cpuset hierarchy is
unchanged irrespective of the cgroup version.

CPU hotplug
-----------

On an asymmetric system, the first detected 32-bit-capable CPU is
prevented from being offlined by userspace and any such attempt will
return ``-EPERM``. Note that suspend is still permitted even if the
primary CPU (i.e. CPU 0) is 64-bit-only.

KVM
---

Although KVM will not advertise 32-bit EL0 support to any vCPUs on an
asymmetric system, a broken guest at EL1 could still attempt to execute
32-bit code at EL0. In this case, an exit from a vCPU thread in 32-bit
mode will return to host userspace with an ``exit_reason`` of
``KVM_EXIT_FAIL_ENTRY`` and will remain non-runnable until successfully
re-initialised by a subsequent ``KVM_ARM_VCPU_INIT`` operation.