On Mon, Feb 3, 2025, at 11:21, Julian Vetter wrote:
> On 1/31/25 08:28, Arnd Bergmann wrote:
>> Those functions have been unchanged since before the git
>> history and there are some comments that I don't find helpful.
>> One thing they do is to deal with unaligned memory buffers,
>> which the generic ones don't, but that could be easily added
>> using get_unaligned/put_unaligned, expecting the compiler
>> to optimize the memory side of the transfer.
>
> I'm not sure what you suggest. Do you mean adding
> get_unaligned/put_unaligned in the generic include/asm-generic/io.h
> functions, or just adding get_unaligned/put_unaligned to the body of the
> parisc specific functions? Instead of all the "switch... case 0x2..." code?
I meant adding put_unaligned/get_unaligned like this:
--- a/include/asm-generic/io.h
+++ b/include/asm-generic/io.h
@@ -422,7 +422,8 @@ static inline void readsw(const volatile void __iomem *addr, void *buffer,
do {
u16 x = __raw_readw(addr);
- *buf++ = x;
+ put_unaligned(x, buf);
+ buf++;
} while (--count);
}
}
This has no effect on architectures that can handle unaligned
load/store on memory, and should behave the same way as the
parisc code otherwise. On powerpc and possibly others, there
needs to be a barrier (eieio() on powerpc) inside of the loop,
according to Segher's earlier comments.
Arnd
