Re: [RFC v5.1 9/9] [DON'T APPLY] cache: sifive-ccache: add cache flushing capability

[Arnd Bergmann]

On Wed, Jan 4, 2023, at 12:56, Conor Dooley wrote:
> On Wed, Jan 04, 2023 at 11:19:44AM +0100, Arnd Bergmann wrote:
>> On Wed, Jan 4, 2023, at 10:23, Conor Dooley wrote:
>> I would try to replace both of these indirections and instead
>> handle it all from C code in arch_sync_dma_for_device() directly,
>> for the purpose of readability and maintainability.
>> static inline void dma_cache_clean(void *vaddr, size_t size)
>> {
>>         if (!cache_maint_ops.clean)
>>                zicbom_cache_clean(vaddr, size, riscv_cbom_block_size);
>
> And I figure that this function is effectively a wrapper around ALT_CMO_OP()?
>
>>         else
>>                cache_maint_ops.clean(vaddr, size, riscv_cbom_block_size);
>
> And this one gets registered by the driver using an interface like the
> one I already proposed, just with the cache_maint_ops struct expanded?

Yes, exactly.

> Extrapolating, with these changes having an errata would not even be
> needed in order to do cache maintenance.
> Since the ALT_CMO_OP() version would only be used inside
> zicbom_cache_clean(), assuming I understood correctly, a driver could
> just register cache_maint_ops for a given platform without having to
> muck around with errata.

That is the idea, and ALT_CMO_OP() itself can just go away
as by just putting the inline asm without the alternative into
the zicbom_cache_clean() version, making the THEAD branch yet
another cache_maint_ops instance.

>> which then makes it very clear what the actual code path
>> is, while leaving the zicbom case free of indirect function
>> calls. You can still use a static_branch() to optimize the
>> conditional, but I would try to avoid any extra indirection
>> levels or errata checks.
>
> The other thing that I like about this is we can then remove the various
> calls to ALT_CMO_OP() that are scattered around arch/riscv now & replace
> them with functions that have more understandable names.

I only see them in arch/riscv/mm/dma-noncoherent.c and arch/riscv/mm/pmem.c,
but yes, both of these should just call the new functions, whatever the
calling conventions end up being.

    Arnd