Re: Base GDB tracing 32bit applications on amd64 kernel (with 64bit debugger) part 1)

[Maxime Villard <max%m00nbsd.net@localhost>]

Le 20/07/2019 à 10:58, Manuel Bouyer a écrit :
> On Sat, Jul 20, 2019 at 09:40:38AM +0200, Maxime Villard wrote:
> > Le 17/07/2019 à 11:08, Manuel Bouyer a écrit :
> > > On Wed, Jul 17, 2019 at 11:00:04AM +0200, Martin Husemann wrote:
> > > > >    - Xen is causing some problems here (as usual...). See process_write_regs(),
> > > > >      we are resetting the ss/cs registers to the default values. Yet in
> > > > >      cpu_setmcontext32 we suddenly don't care about resetting ss/cs to the
> > > > >      default values. It's not clear to me which side is incorrect.
> > > >
> > > > Cc'ing Manuel - can you check this, please?
> > >
> > > I'm probably missing something. there's no Xen-specific code in
> > > process_write_regs()
> >
> > ? There is a #ifdef XENPV in this very function.
> >
> > The point was: there is an asymmetry between process_write_regs() and
> > cpu_setmcontext32(), because the former resets ss/cs, and the latter
> > does not. One side has to be incorrect, but it's not clear which.
>
> But process_write_regs() is for 64bit processes, while cpu_setmcontext32()
> is for 32bits, isn't it ?

The point of this thread is precisely to make process_write_regs() handle
32bit processes as well.

Beyond that, the current asymmetry between cpu_setmcontext and cpu_setmcontext32
is already highly suspicious. Yes they are dedicated to 64bit and 32bit
respectively, but even then, I don't understand how Xen itself can make the
difference.

It is possible that the #ifdef XENPV in cpu_setmcontext is just wrong, and
we just don't need to reset ss/cs anywhere.