Re: [GIT PULL] x86/paravirt for v2.6.33

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 12/08/09 13:34, Linus Torvalds wrote:
> I do _not_ want to add any more task_pt_regs() crap, please.
>
> Why? It's wrong for at least vm86 mode (and from kernel system calls).
>    

Would the stack frame version work in these cases?

> Maybe we can't get into system calls from vm86 mode, and the kernel
> hopefully doesn't do those things anyway, but the point is, you chose the
> wrong way to go.
>    

iopl doesn't make much sense as a kernel-called syscall, unless the 
caller is intending to change the usermode iopl.  In which case, won't 
task_pt_regs() do the right thing - by pointing to the saved usermode 
register set - vs modifying the ptregs the caller may pass in?

iopl is also one of the special set of syscalls which get special 
handing in entry_*.S, so I don't think doing a direct call from within 
the kernel is ever sensible, and it should always be possible to make 
task_pt_regs return meaningful results.

I agree with you that vm86 would be a problem if its possible to call iopl.

> The old version that actually passed the stack frame was better. Why pick
> the inferior version?
>    

Mainly because it exposes the difference between the 32 and 64-bit ABIs, 
requiring separate code for each case; it seemed like an opportunity to 
remove the differences.

Anyway, I'll post a patch to revert to the pt_regs-based version shortly.

     J