[RFC PATCH 2/3] arm64: Add support for asymmetric AArch32 EL0 configurations

[Will Deacon]

On Fri, Oct 09, 2020 at 08:13:56AM +0200, Greg Kroah-Hartman wrote:
> On Thu, Oct 08, 2020 at 07:16:40PM +0100, Qais Yousef wrote:
> > When the CONFIG_ASYMMETRIC_AARCH32 option is enabled (EXPERT), the type
> > of the ARM64_HAS_32BIT_EL0 capability becomes WEAK_LOCAL_CPU_FEATURE.
> > The kernel will now return true for system_supports_32bit_el0() and
> > checks 32-bit tasks are affined to AArch32 capable CPUs only in
> > do_notify_resume(). If the affinity contains a non-capable AArch32 CPU,
> > the tasks will get SIGKILLed. If the last CPU supporting 32-bit is
> > offlined, the kernel will SIGKILL any scheduled 32-bit tasks (the
> > alternative is to prevent offlining through a new .cpu_disable feature
> > entry).
> > 
> > In addition to the relaxation of the ARM64_HAS_32BIT_EL0 capability,
> > this patch factors out the 32-bit cpuinfo and features setting into
> > separate functions: __cpuinfo_store_cpu_32bit(),
> > init_cpu_32bit_features(). The cpuinfo of the booting CPU
> > (boot_cpu_data) is now updated on the first 32-bit capable CPU even if
> > it is a secondary one. The ID_AA64PFR0_EL0_64BIT_ONLY feature is relaxed
> > to FTR_NONSTRICT and FTR_HIGHER_SAFE when the asymmetric AArch32 support
> > is enabled. The compat_elf_hwcaps are only verified for the
> > AArch32-capable CPUs to still allow hotplugging AArch64-only CPUs.
> > 
> > Make sure that KVM never sees the asymmetric 32bit system. Guest can
> > still ignore ID registers and force run 32bit at EL0.
> > 
> > Co-developed-by: Qais Yousef <qais.yousef at arm.com>
> > Signed-off-by: Catalin Marinas <catalin.marinas at arm.com>
> > Signed-off-by: Qais Yousef <qais.yousef at arm.com>
> > ---
> >  arch/arm64/Kconfig                   | 14 ++++++
> >  arch/arm64/include/asm/cpu.h         |  2 +
> >  arch/arm64/include/asm/cpucaps.h     |  3 +-
> >  arch/arm64/include/asm/cpufeature.h  | 20 +++++++-
> >  arch/arm64/include/asm/thread_info.h |  5 +-
> >  arch/arm64/kernel/cpufeature.c       | 66 +++++++++++++++-----------
> >  arch/arm64/kernel/cpuinfo.c          | 71 ++++++++++++++++++----------
> >  arch/arm64/kernel/process.c          | 17 +++++++
> >  arch/arm64/kernel/signal.c           | 18 +++++++
> >  arch/arm64/kvm/arm.c                 |  5 +-
> >  arch/arm64/kvm/guest.c               |  2 +-
> >  arch/arm64/kvm/sys_regs.c            | 14 +++++-
> >  12 files changed, 176 insertions(+), 61 deletions(-)
> > 
> > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > index 6d232837cbee..591853504dc4 100644
> > --- a/arch/arm64/Kconfig
> > +++ b/arch/arm64/Kconfig
> > @@ -1868,6 +1868,20 @@ config DMI
> >  
> >  endmenu
> >  
> > +config ASYMMETRIC_AARCH32
> > +	bool "Allow support for asymmetric AArch32 support"
> > +	depends on COMPAT && EXPERT
> 
> Why EXPERT?  You don't want this able to be enabled by anyone?

TBH, I'd be inclined to drop the Kconfig option altogether. We're not
looking at a lot of code here, so all it does is further fragment the
build testing we get from CI (or it just ends up being enabled all of the
time).

A cmdline option, on the other hand, makes a tonne of sense to me, as it
acts as an "opt-in" that the distribution is ready to handle the madness
(because userspace will need to care about this even with the scheduler
hacks proposed here).

Will