Re: [PATCH 0/7] Add support to handle misaligned accesses in S-mode

[Clément Léger]

On 04/10/2023 10:26, David Laight wrote:
> From: Clément Léger
>> Sent: 02 October 2023 08:40
>>
>> On 30/09/2023 11:23, Conor Dooley wrote:
>>> On Tue, Sep 26, 2023 at 05:03:09PM +0200, Clément Léger wrote:
>>>> Since commit 61cadb9 ("Provide new description of misaligned load/store
>>>> behavior compatible with privileged architecture.") in the RISC-V ISA
>>>> manual, it is stated that misaligned load/store might not be supported.
>>>> However, the RISC-V kernel uABI describes that misaligned accesses are
>>>> supported.
> ...
> 
> That it just really horrid.
> If the cpu is going to trap misaligned accesses then you want
> The compiler generated code (ie packed data) not to generate
> misaligned accesses.

Hi David,

Saying that you support misaligned access does not mean that they are
going to be efficient, just that they are supported (in fact, the uABI
state that they may perform poorly). The compiler is actually not so
stupid and will try to do as much aligned access as possible in what it
generates (unless forced by some assembly, cast or whatever that can
screw up alignment accesses). This is already the case and it will most
probably not change.

> So you have to change the kernel uABI.

I don't think so. Rule N°1 for kernel development is "Don't break the
userspace". So if changing the RISC-V uABI to say "misaligned accesses
are not supported", that is unlikely to happen. We stated that
misaligned access are supported and thus, they will continue to be
supported.

> 
> OTOH if you known that such accesses won't fault and will be
> not really slower than aligned accesses then optimised versions
> of some functions (like memcpy and checksums) can use misaligned
> accesses.

Yes, this is selected by HAVE_EFFICIENT_UNALIGNED_ACCESS. On RISC-V,
since the specification says nothing about the efficiency of such
access, we can't select it like that. Some RISC-V based SoC/CPUs might
want to select it manually in their config. In order to support that
dynamically and in a generic way, some future work could involve using
static keys for such alternatives and enable it based on the speed that
was detected.

Thanks,

Clément

> 
> 	David
> 
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)