Re: [PATCH 2/3] x86: Pass in pt_regs pointer for syscalls that need it

[Brian Gerst <brgerst@gmail.com>]

On Wed, Feb 11, 2009 at 2:50 PM, H. Peter Anvin <hpa@zytor.com> wrote:
> Brian Gerst wrote:
>>
>> I guess I could go back to extracting the args from the pt_regs struct
>> given just the pointer.  How do you intend to handle system calls in
>> your changes (normal ones, not needing pt_regs)?
>>
>
> My plan was to by default load up the three first arguments in (%eax, %edx,
> %ecx) followed by the remaining arguments on the stack... I currently have
> it as a reorganized struct pt_regs, but I'm still trying to figure out if it
> would make more sense from a correctness and performance perspective to
> instead have duplicates of these entries.
>
> For the pt_regs-using registers, they would need a tiny trampoline, looking
> like:
>
>        leal 16(%esp),%eax
>        jmp <real function>
>
>        -hpa
>
>

IMHO, copying the 4th-6th args to a new stack frame is the only way to
guarantee that gcc won't trash any part of pt_regs.  The question is
whether to do it unconditionally, or try to be clever and only copy
them for the syscalls that actually need them.

--
Brian Gerst