[PATCH] arm64: errata: Remove AES hwcap for COMPAT tasks on A57 and A72

[Robin Murphy]

On 2022-01-27 19:55, Arnd Bergmann wrote:
> On Thu, Jan 27, 2022 at 7:43 PM Robin Murphy <robin.murphy at arm.com> wrote:
>> On 2022-01-27 14:52, Arnd Bergmann wrote:
>>> On Thu, Jan 27, 2022 at 1:39 PM Ard Biesheuvel <ardb at kernel.org> wrote:
>>>> Cortex-A57 and Cortex-A72 have an erratum where an interrupt that
>>>> occurs between a pair of AES instructions in aarch32 mode may corrupt
>>>> the ELR. The task will subsequently produce the wrong AES result.
>>>
>>>> Fortunately, the only A72 that is known to be widely deployed with
>>>> 32-bit kernels and/or user space is the Raspberry Pi 4, which does not
>>>> implement the crypto extensions.
>>>
>>> I have previously used Amazon EC2 A1 instances (using Graviton 1, Cortex-A72)
>>> to run Debian armhf and armel user space in chroot environments. No idea how
>>> common that is for production environments, but there is a good chance that
>>> actual users might use e.g. 32-bit docker images for Alpine Linux on the same
>>> machines.
>>>
>>> There are also some early A57/A72 chips used in Android devices that may
>>> frequently run 32-bit applications, or even be limited to 32-bit
>>> kernels, see [1].
>>> Not sure how likely it is for any of them to get (a backport of) this
>>> patch though.
>>
>> Isn't ChromeOS still notoriously shipping AArch32 userspace? Between
>> Mediatek and Rockchip, both A57 and A72 are definitely playing on that
>> stage.
> 
> I wouldn't call it "notorious", but indeed that is an important user
> that I missed. For some reason I misremembered that both mt8173 and
> rk3399 are Cortex-A73 based, but they are A72 as you say.

Although it seems I misremembered about A57 - that must have been in a 
preproduction version of what became MT8173, or I'm imagining it completely.

>> I'd concur that 32-bit *kernel* on v8 is a bit of a specialist sport,
>> especially these days. IIRC Exynos 5433 might have been one of the more
>> common Android setups for that, but 2015 is now an eternity ago in phone
>> years..
> 
> Unfortunately, 32-bit kernels are still a thing on low-end Android phones,
> including some very recent Cortex-A55 based products with Android Go
> edition and 2GB of RAM or less, and also some highly popular (at the time)
> Cortex-A53 Snapdragon 410 phones started shipping before a 64-bit kernel
> port was available on Snapdragon.

Apologies, I really picked the wrong wording there, my head was buried 
in the context of this patch and larger cores in general - by "on v8" I 
think I was trying to imply shorthand for A57 up to A75 (after which it 
becomes moot anyway), and bare-metal rather than VMs at that. At the 
lower end and into weird embedded stuff AArch32 is indeed very much 
alive and well - obviously Cortex-A32 is v8, for starters. Luckily it 
isn't relevant to this erratum, although there is of course the 
implication for workarounds in general.

> A57/A72 at least disappeared from phones a long time ago because of
> both cost and power consumption reasons, so it's only the low end devices
> now. I'm not aware of any Android devices actually shipping with a 64-bit
> kernel and 32-bit user space, maybe that was never an option in the SDK.

FWIW I had the impression that that was pretty common for a while back 
before Google started actively pushing for 64-bit support, once new SoCs 
had 64-bit kernels from bringup but Android vendors still didn't like 
the storage/memory cost of having to carry both sets of runtime 
libraries around. But then my awareness of Android has always been more 
from the development end rather than what may have actually ended up 
going to market.

Cheers,
Robin.