Re: [PATCH] powerpc/kprobes: Fix trap address when trap happened in real mode

[Christophe Leroy]

Le 18/02/2020 Ã 13:33, Masami Hiramatsu a ÃcritÂ:
> On Tue, 18 Feb 2020 12:04:41 +0100
> Christophe Leroy <christophe.leroy@xxxxxx> wrote:
>
> > > > Nevertheless, if one symbol has been forgotten in the blacklist, I think
> > > > it is a problem if it generate Oopses.
> > >
> > > There is a long history also on x86 to make a blacklist. Anyway, how did
> > > you get this error on PPC32? Somewhere would you like to probe and
> > > it is a real mode function? Or, it happened unexpectedly?
> >
> > The first Oops I got was triggered by a WARN_ON() kind of trap in real
> > mode. The trap exception handler called kprobe_handler() which tried to
> > read the instruction at the trap address (which was a real-mode address)
> > so it triggered a Bad Access Fault.
> >
> > This was initially the purpose of my patch.
>
> OK, then filtering the trap reason in kprobe handler is a bit strange.
> It should be done in the previous stage (maybe in trap.c)

See commit 
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v5.6-rc2&id=6cc89bad60a673a24386f1ada83de8a068a78909

> Can we filter it by exception flag or only by checking the instruction
> which causes the exception, or needs get_kprobe()...?

The trap instruction used by kprobe is also used for other purposes like 
BUG_ON() or WARN_ON(), so needs get_kprobe()



> > After discussion with you, I started looking at what would be the effect
> > of setting a kprobe event in a function which runs in real mode.
>
> If the kprobe single-stepping (or emulation) works in real mode, just
> ignore the kprobes pre/post_handlers and increment nmissed count.
>
> If that doesn't work, we have to call a BUG_ON, because we can not
> continue the code execution. And also, you have to find a way to make
> a blacklist for real mode code.

Yes, it has to be done function by function (hoppefully there's not more 
than a dozen).
But I'd like something which can fails gracefully for the functions we 
will forget to mark noprobe.

But as a first step I'd really like a bug fix in 5.6 to avoid Oopsing in 
kprobe_handler() at a non-kprobe trap.

Christophe