On Thu, 24 Oct 2019, Thomas Gleixner wrote:
> Whatever we decide it is, leaving it completely inconsistent is not a
> solution at all. The options are:
Actually there is also:
0) Always do unconditional trace_irqs_on().
But that does not allow to actually trace the real return flags
state which might be useful to diagnose crap which results from user
space CLI.
> 1) Always do conditional tracing depending on the user_regs->eflags.IF
> state.
>
> 2) #1 + warn once when syscalls and exceptions (except NMI/MCE) happen
> and user_regs->eflags.IF is cleared.
>
> 3a) #2 + enforce signal handling to run with interrupts enabled.
>
> 3b) #2 + set regs->eflags.IF. So the state is always correct from the
> kernel POV. Of course that changes existing behaviour, but its
> changing undefined and inconsistent behaviour.
>
> 4) Let iopl(level) return -EPERM if level == 3.
>
> Yeah, I know it's not possible due to regressions (DPKD uses iopl(3)),
> but TBH that'd be the sanest option of all.
>
> Of course the infinite wisdom of hardware designers tied IN, INS, OUT,
> OUTS and CLI/STI together on IOPL so we cannot even distangle them in
> any way.
>
> The only way out would be to actually use a full 8K sized I/O bitmap,
> but that's a massive pain as it has to be copied on every context
> switch.
>
> Really pretty options to chose from ...
>
> Thanks,
>
> tglx
>
