Re: CONFIG_RETHUNK int3 filling prevents kprobes in function body

[Google]

On Mon, 5 Sep 2022 17:20:30 +0200
Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> On Mon, Sep 05, 2022 at 05:10:27PM +0200, Peter Zijlstra wrote:
> > On Mon, Sep 05, 2022 at 10:57:58AM -0400, Steven Rostedt wrote:
> > > On Sun, 4 Sep 2022 23:07:13 +0900
> > > Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx> wrote:
> > > 
> > > > Can we use another instruction for padding instead of INT3? (e.g. NOP or UD2)
> > > > 
> > > > Or, can I expect the instruction length in __return_sites[] are always 5?
> > > > If so, I can just skip 5 bytes if the address is in __return_sites[].
> > > 
> > > Perhaps another option is to have a table of where the padding is placed
> > > (tagged), and that kprobes could check to see if the int3 is due to this
> > > padding or not?
> > 
> > I don't see need for that. If you want to be strict you can simply
> > follow the branches found earlier, if you want to be lazy, you can
> > decode until you run out of the symbol size.
> 
> Another lazy option is to teach the thing that 'ret' is followed by 0,1
> or 4 'int3' instructions depending on CONFIG_SLS, CONFIG_RETHUNK, but
> that'll get you into trouble with future SLS compiler options, like the
> aforementioned JMP-SLS option.

Yes, I agree... OK, let me try to decode branches to find the
nearest earlier target instruction.

Thank you!

-- 
Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>