Re: [PATCH 13/X] uprobes: introduce UTASK_SSTEP_TRAPPED logic

[Ananth N Mavinakayanahalli <ananth@in.ibm.com>]

On Mon, Oct 24, 2011 at 04:41:27PM +0200, Oleg Nesterov wrote:
> On 10/22, Ananth N Mavinakayanahalli wrote:
> >
> > On Wed, Oct 19, 2011 at 11:53:44PM +0200, Oleg Nesterov wrote:
> > > Finally, add UTASK_SSTEP_TRAPPED state/code to handle the case when
> > > xol insn itself triggers the signal.
> > >
> > > In this case we should restart the original insn even if the task is
> > > already SIGKILL'ed (say, the coredump should report the correct ip).
> > > This is even more important if the task has a handler for SIGSEGV/etc,
> > > The _same_ instruction should be repeated again after return from the
> > > signal handler, and SSTEP can never finish in this case.
> >
> > Oleg,
> >
> > Not sure I understand this completely...
> 
> I hope you do not think I do ;)

I think you understand it better than you think you do :-)

> > When you say 'correct ip' you mean the original vaddr where we now have
> > a uprobe breakpoint and not the xol copy, right?
> 
> Yes,
> 
> > Coredump needs to report the correct ip, but should it also not report
> > correctly the instruction that caused the signal? Ergo, shouldn't we
> > put the original instruction back at the uprobed vaddr?
> 
> OK, now I see what you mean. I was confused by the "restore the original
> instruction before _restart_" suggestion.
> 
> Agreed! it would be nice to "hide" these int3's if we dump the core, but
> I think this is a bit off-topic. It makes sense to do this in any case,
> even if the core-dumping was triggered by another thread/insn. It makes
> sense to remove all int3's, not only at regs->ip location. But how can
> we do this? This is nontrivial.

I don't think that is a problem.. see below...

> And. Even worse. Suppose that you do "gdb probed_application". Now you
> see int3's in the disassemble output. What can we do?

In this case, nothing.

> I think we can do nothing, at least currently. This just reflects the
> fact that uprobe connects to inode, not to process/mm/etc.
> 
> What do you think?

Thinking further on this, in the normal 'running gdb on a core' case, we
won't have this problem, as the binary that we point gdb to, will be a
pristine one, without the uprobe int3s, right?

Ananth