Re: [PATCH v7 1/6] riscv: mm: dma-noncoherent: Switch using function pointers for cache management

[Arnd Bergmann]

On Tue, Apr 4, 2023, at 07:29, Christoph Hellwig wrote:
> On Thu, Mar 30, 2023 at 09:42:12PM +0100, Prabhakar wrote:
>> From: Lad Prabhakar <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx>
>> 
>> Currently, selecting which CMOs to use on a given platform is done using
>> and ALTERNATIVE_X() macro. This was manageable when there were just two
>> CMO implementations, but now that there are more and more platforms coming
>> needing custom CMOs, the use of the ALTERNATIVE_X() macro is unmanageable.
>> 
>> To avoid such issues this patch switches to use of function pointers
>> instead of ALTERNATIVE_X() macro for cache management (the only drawback
>> being performance over the previous approach).
>> 
>> void (*clean_range)(unsigned long addr, unsigned long size);
>> void (*inv_range)(unsigned long addr, unsigned long size);
>> void (*flush_range)(unsigned long addr, unsigned long size);
>> 
>
> NAK.  Function pointers for somthing high performance as cache
> maintainance is a complete no-go.

As we already discussed, this is now planned to use a direct
branch to the zicbom version when the function pointer is NULL,
which should be a fast predicted branch on all standard implementations
and only be slow on the nonstandard ones, which I think is a reasonable
compromise.

I'm also not sure I'd actually consider this a fast path, since
there is already a significant cost in accessing the invalidated
cache lines afterwards, which is likely going to be much higher than
the cost of an indirect branch.

I suppose an alternative would be to use the linux/static_call.h
infrastructure to avoid the overhead of indirect branches altogether.
Right now, this has no riscv specific implementation though, so
using it just turns into a regular indirect branch until someone
implements static_call.

       Arnd