Re: [PATCH] io_thread/x86: don't reset 'cs', 'ss', 'ds' and 'es' registers for io_threads

From: Peter Zijlstra
Date: Tue May 04 2021 - 04:39:59 EST



+ linux-toolchains

On Mon, May 03, 2021 at 12:14:45PM -0700, Linus Torvalds wrote:
> On Mon, May 3, 2021 at 9:05 AM Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
> >
> > Linus, what is the actual effect of allowing gdb to attach these threads? Can we instead make all the regset ops do:
> >
> > if (not actually a user thread) return -EINVAL;
>
> I don't think it matters - the end result ends up being the same, ie
> gdb gets confused about whether the (parent) thread is a 32-bit or
> 64-bit one.
>
> So the basic issue is
>
> (a) we want the IO threads to look exactly like normal user threads
> as far as the kernel is concerned, because we had way too many bugs
> due to special cases.
>
> (b) but that means that they are also visible to user space, and then
> gdb has this odd thing where it takes the 64-bit vs 32-bit data for
> the whole process from one thread, and picks the worst possible thread
> to do it (ie explicitly not even the main thread, so usually the IO
> thread!)
>
> That (a) ended up really being critical. The issues with special cases
> were just horrendous, both for security issues (ie "make them kernel
> threads but carry user credentials" just caused lots of problems) but
> also for various just random other state handling issues (signal state
> in particular).
>
> So generally, the IO threads are now 100% normal threads - it's
> literally just that they never return to user space because they are
> always just doing the IO offload on the kernel side.
>
> That part is lovely, but part of the "100% IO threads" really is that
> they share the signal struct too, which in turn means that they very
> much show up as normal threads. Again, not a problem: they really
> _are_ normal threads for all intents and purposes.
>
> But then that (b) issue means that gdb gets confused by them. I
> personally think that's just a pure gdb mis-feature, but I also think
> that "hey, if we just make the register state look like the main
> thread, and unconfuse gdb that way, problem solved".
>
> So I'd actually rather not make these non-special threads any more
> special at all. And I strongly suspect that making ptrace() not work
> on them will just confuse gdb even more - so it would make them just
> unnecessarily special in the kernel, for no actual gain.
>
> Is the right thing to do to fix gdb to not look at irrelevant thread B
> when deciding whether thread A is 64-bit or not? Yeah, that seems like
> obviously the RightThing(tm) to me.
>
> But at the same time, this is arguably about "regression", although at
> the same time it's "gdb doesn't understand new user programs that use
> new features, film at 11", so I think that argument is partly bogus
> too.
>
> So my personal preference would be:
>
> - make those threads look even more like user threads, even if that
> means giving them pointless user segment data that the threads
> themselves will never use
>
> So I think Stefan's patch is reasonable, if not pretty. Literally
> becasue of that "make these threads look even more normal"
>
> - ALSO fix gdb that is doing obviously garbage stupid things
>
> But I'm obviously not involved in that "ALSO fix gdb" part, and
> arguably the kernel hack then makes it more likely that gdb will
> continue doing its insane broken thing.

Anybody on toolchains that can help get GDB fixed?