Re: [PATCHSET ptrace] ptrace: implement PTRACE_SEIZE/INTERRUPT and group stop notification, take#4

[Denys Vlasenko <vda.linux@googlemail.com>]

On Friday 03 June 2011 03:34, Tejun Heo wrote:
> > This should alleviate Linus' concerns that we are suffering from unnecessary
> > featuritis - that we invent API which is significantly different
> > from existing one, and as such users will not use it.
> > 
> > Existing API doesn't use GETSIGINFO data per se to distinguish group-stop from
> > signal-delivery-stop, it uses the fact that GETSIGINFO fails on group-stop.
> > (Well, arguably it's not a "designed" API, more like "accidentally created API",
> > but nevertheless it exists right now). Oleg's proposal means that the new way
> > may be makde to work very similarly.
> 
> Okay, you want to preserve %NULL SIGINFO on all STOP traps.  Dropping
> si_pt_flags and using exit_code makes sense but I think we'll be
> better off enabling GETSIGINFO.  The affected part of code has to be
> changed anyway

Well, not exactly. The affected part of code in strace *has been changed
already*. Latest version of strace finally does show "SIGSTOP delivered"
and "group-stop happened" events differently.

Now strace, ideally, can have a minimal change

-	ptrace(PTRACE_LISTEN, pid, 0, sig);
+	if (this_is_group_stop)
+		ptrace(PTRACE_LISTEN, pid, 0, 0);
+	else
+		ptrace(PTRACE_SYSCALL, pid, 0, sig);

and boom, group-stops are handled correctly!

> and with %NULL GETSIGINFO we effectively would have 
> driven ourselves into corner if more information needs to be added
> later on.

Oleg thinks that GETSIGINFO may be not the best way to provide additional data.
Think about it. It's "get _signal_ info" op. By overloading it to mean something
else, we are creating yet another ptrace quirk: "sometimes GETSIGINFO will
return something else than signal info".

Oleg says he prefers to add a separate GETfoo op, with cleaner meaning.


> Don't current users unconditionally issue CONT(sig) anyway? 

Yes.

-- 
vda