Re: [PATCH 17/24] x86/entry: Introduce struct ist_regs

[Lai Jiangshan]

On 2021/9/10 08:18, Lai Jiangshan wrote:
> On 2021/9/1 01:50, Lai Jiangshan wrote:
> > From: Lai Jiangshan <laijs@xxxxxxxxxxxxxxxxx>
> >
> > struct ist_regs is the upmost stack frame for IST interrupt, it
> > consists of struct pt_regs and other fields which will be added in
> > later patch.
> >
> > Make vc_switch_off_ist() take ist_regs as its argument and it will switch
> > the whole ist_regs if needed.
> >
> > Make the frame before calling paranoid_entry() and paranoid_exit() be
> > struct ist_regs.
> >
> > This patch is prepared for converting paranoid_entry() and paranoid_exit()
> > into C code which will need the additinal fields to store the results in
> > paranoid_entry() and to use them in paranoid_exit().
>
> This patch was over designed.
>
> In ASM code, we can easily save results in the callee-saved registers.
> For example, rc3 is saved in %r14, gsbase info is saved in %rbx.
>
> And in C code, we can't save results in registers.  And I thought there was
> no place to save the results because the CR3 and gsbase are not kernel's.
> So I extended the pt_regs to ist_regs to save the results.
>
> But it was incorrect.  The results can be saved in percpu data at the end of
> paranoid_entry() after the CR3/gsbase are settled down.  And the results
> can be read at the beginning of paranoid_exit() before the CR3/gsbase are
> switched to the interrupted context's.
>
> sigh.

Sigh again. IST interrupts can be nested.  We can't save the results in percpu
data unless we make it stack-like which is not a good idea.  The changelog
failed to express it.

> > The C interrupt handlers don't use struct ist_regs due to they don't need
> > the additional fields in the struct ist_regs, and they still use pt_regs.