Re: [PATCH] uprobe: Add uprobe_pre/post_sstep_notifier to NOKPROBE_SYMBOL

[Pratyush Anand]

Hi Oleg,

On 02/08/2016:10:30:35 PM, Oleg Nesterov wrote:
> On 08/02, Pratyush Anand wrote:
> >
> > uprobe_pre_sstep_notifier and uprobe_post_sstep_notifier are called from
> > debug exception handler, so blacklist them for kprobing.
> 
> Let me add kprobes maintainers, I am a bit confused...
> 
> > @@ -1997,6 +1998,7 @@ int uprobe_pre_sstep_notifier(struct pt_regs *regs)
> >  	set_thread_flag(TIF_UPROBE);
> >  	return 1;
> >  }
> > +NOKPROBE_SYMBOL(uprobe_pre_sstep_notifier);
> >
> >  /*
> >   * uprobe_post_sstep_notifier gets called in interrupt context as part of notifier
> > @@ -2014,6 +2016,7 @@ int uprobe_post_sstep_notifier(struct pt_regs *regs)
> >  	set_thread_flag(TIF_UPROBE);
> >  	return 1;
> >  }
> > +NOKPROBE_SYMBOL(uprobe_post_sstep_notifier);
> 
> but if we need to blacklist uprobe_pre/post_sstep_notifier then we
> also need to blacklist their caller, arch_uprobe_exception_notify() ?

I think yes, in ARM64 I have done that. However, arm64 does not use notifier
method, so arch_uprobe_exception_notify() is just a dummy function for it.

> 
> and every .notifier_call used in register_die_notifier() ?

I tried to look into x86 notify path related to uprobe_pre/post_sstep_notifier().
I see that calling sequence is like do_int3()-> notify_die() ->
atomic_notifier_call_chain() -> __atomic_notifier_call_chain() ->
notifier_call_chain() -> arch_uprobe_exception_notify().

In this sequence, every function is blacklisted for kprobe except
arch_uprobe_exception_notify(). So, I am unable to understand, if
notifier_call_chain() is not safe for kprobe then how can it be safe for a
function it calls.

~Pratyush