Re: [RFC][PATCH] tracing/x86: Save CR2 before tracing irqsoff on error_entry

From: Peter Zijlstra
Date: Thu Mar 21 2019 - 16:50:45 EST


On Thu, Mar 21, 2019 at 11:27:00AM -0700, Andy Lutomirski wrote:
> Well, here's pass zero at this. Untested, because it obviously
> doesn't work. Here are just a few things that are almost certainly
> wrong with it

Aah, you're proposing to simply not do TRACE_IRQS_OFF and
CALL_enter_from_user_mode and let \do_sym deal with it all.

Yes, that looks like it could almost work; esp. if you start by only
doing this for the idtentry stuff.

> - The IRQ tracing needs to be re-added.
>
> - Some real semantics need to be defined for precisely what code is
> responsible for tracing.

So we get passed \do_sym, how about we do:

call __\do_sym

And then use some CPP magic to generate the those functions such that we
have a consistent part of C glue between our asm and our 'real' C
handler.

This glue can then do the tracing in a consistent manner.

#define IDT_HANDLER(do_sym) \
asmlinkage __visible notrace void __do_sym(struct pt_regs *regs) \
{ \
trace_hardirqs_off(); \
if (user_mode(regs)) \
enter_from_user_mode(); \
do_sym(regs); \
}

Except more complicated I'm afraid, we need to handle more args etc..

> - We need some asm-callable assertions to check the following
> conditions as appropriate:
>
> (a) that IRQ flags are currently traced as off.

What do you need this for? When returning from do_sym ?

> (b) that IRQ flags are currently traced to match the IRET frame.

idem. Can't we have our C glue do that?

> (c) that our context tracking is currently in good shape. I'm not
> 100% sure how to define this.

So looking at this more; I used the %ebx games employed by
paranoid_entry to convey the state, but I didn't have to do that, the
actual condition seems to be:

regs->cs & 3

aka. user_mode(regs). In this case our C glue would need to do the
context tracking user exit.

In fact, I can change my patch to use that and reduce the ebx ugly.

> - We need to do some serious don't-instrument-me stuff to all the C
> entries, since we're now in an awful context when calling them.

Yah, but that's not new. do_page_fault(), sync_regs() at the very least
have this, so we can easily have our C glue have this too.