Re: [PATCH -tip v4 10/12] x86/kprobes: Push a fake return address at kretprobe_trampoline

[Masami Hiramatsu]

On Wed, 24 Mar 2021 11:01:43 -0500
Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:

> On Wed, Mar 24, 2021 at 10:40:58AM +0900, Masami Hiramatsu wrote:
> > On Tue, 23 Mar 2021 23:30:07 +0100
> > Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > 
> > > On Mon, Mar 22, 2021 at 03:41:40PM +0900, Masami Hiramatsu wrote:
> > > >  	".global kretprobe_trampoline\n"
> > > >  	".type kretprobe_trampoline, @function\n"
> > > >  	"kretprobe_trampoline:\n"
> > > >  #ifdef CONFIG_X86_64
> > > 
> > > So what happens if we get an NMI here? That is, after the RET but before
> > > the push? Then our IP points into the trampoline but we've not done that
> > > push yet.
> > 
> > Not only NMI, but also interrupts can happen. There is no cli/sti here.
> > 
> > Anyway, thanks for pointing!
> > I think in UNWIND_HINT_TYPE_REGS and UNWIND_HINT_TYPE_REGS_PARTIAL cases
> > ORC unwinder also has to check the state->ip and if it is kretprobe_trampoline,
> > it should be recovered.
> > What about this?
> 
> I think the REGS and REGS_PARTIAL cases can also be affected by function
> graph tracing.  So should they use the generic unwind_recover_ret_addr()
> instead of unwind_recover_kretprobe()?

Yes, but I'm not sure this parameter can be applied.
For example, it passed "state->sp - sizeof(unsigned long)" as where the
return address stored address. Is that same on ftrace graph too?

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>