On Tue, Jun 25, 2024, at 06:04, Jisheng Zhang wrote:
> This series tries to optimize riscv uaccess in the following way:
>
> patch1 implement the user_access_begin and families, I.E the unsafe
> accessors. when a function like strncpy_from_user() is called,
> the userspace access protection is disabled and enabled for every
> word read. After patch1, the protection is disabled at the beginning
> of the copy and enabled at the end.
>
> patch2 is a simple clean up
>
> patch3 uses 'asm goto' for put_user()
> patch4 uses 'asm goto output' for get_user()
>
> Both patch3 and patch4 are based on the fact -- 'asm goto' and
> 'asm goto output'generates noticeably better code since we don't need
> to test the error etc, the exception just jumps to the error handling
> directly.
Nice!
I hope that we can one day make this more straightforward
and share more of the implementation across architectures,
in particular the bits that are just wrappers around
the low-level asm.
We have something like this already on powerpc and x86,
and Linus just did the version for arm64, which I assume
you are using as a template for this. Four architectures
is probably the point at which we should try to consolidate
the code again and move as much as we can into asm-generic.
I think the only bets we really need from an architecture
here are:
- __enable_user_access()/__disable_user_access() in
the label version
- __raw_get_mem_{1,2,4,8}() and __raw_put_mem_{1,2,4,8}()
and then we can build all the normal interface functions
on those in a way that assumes we have as goto available,
with the appropriate fallback in __raw_get_mem_*() as
long as we need to support gcc-10 and older.
Arnd
