On Thu, Nov 19, 2020 at 05:42:03PM +0100, Peter Zijlstra wrote:
> On Thu, Nov 19, 2020 at 04:28:23PM +0000, Will Deacon wrote:
> > On Thu, Nov 19, 2020 at 05:14:48PM +0100, Peter Zijlstra wrote:
> > > On Fri, Nov 13, 2020 at 09:37:13AM +0000, Will Deacon wrote:
> > > > When exec'ing a 32-bit task on a system with mismatched support for
> > > > 32-bit EL0, try to ensure that it starts life on a CPU that can actually
> > > > run it.
> > > >
> > > > Signed-off-by: Will Deacon <will@xxxxxxxxxx>
> > > > ---
> > > > arch/arm64/kernel/process.c | 12 +++++++++++-
> > > > 1 file changed, 11 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/arch/arm64/kernel/process.c b/arch/arm64/kernel/process.c
> > > > index 1540ab0fbf23..17b94007fed4 100644
> > > > --- a/arch/arm64/kernel/process.c
> > > > +++ b/arch/arm64/kernel/process.c
> > > > @@ -625,6 +625,16 @@ unsigned long arch_align_stack(unsigned long sp)
> > > > return sp & ~0xf;
> > > > }
> > > >
> > > > +static void adjust_compat_task_affinity(struct task_struct *p)
> > > > +{
> > > > + const struct cpumask *mask = system_32bit_el0_cpumask();
> > > > +
> > > > + if (restrict_cpus_allowed_ptr(p, mask))
> > > > + set_cpus_allowed_ptr(p, mask);
> > >
> > > This silently destroys user state, at the very least that ought to go
> > > with a WARN or something. Ideally SIGKILL though. What's to stop someone
> > > from doing a sched_setaffinity() right after the execve, same problem.
> > > So why bother..
> >
> > It's no different to CPU hot-unplug though, is it? From the perspective of
> > the 32-bit task, the 64-bit-only cores were hot-unplugged at the point of
> > execve(). Calls to sched_setaffinity() for 32-bit tasks will reject attempts
> > to include 64-bit-only cores.
>
> select_fallback_rq() has a printk() in to at least notify things went
> bad. But I don't particularly like the current hotplug semantics; I've
> wanted to disallow the hotplug when it would result in this case, but
> computing that is tricky. It's one of those things that's forever on the
> todo list ... :/
I know that feeling...
I can add a printk() in the case where we override the mask (I think taking
the subset is ok), since I agree that it would be better if userspace had
had the foresight to avoid the situation in the first place.
> > I initially wanted to punt this all to userspace, but one of the big
> > problems with that is when a 64-bit task is running on a CPU only capable
> > of running 64-bit tasks and it execve()s a 32-bit task. At the point, we
> > have to do something because we can't even run the new task for it to do
> > a sched_affinity() call (and we also can't deliver SIGILL).
>
> Userspace can see that one coming though... I suppose you can simply
> make the execve fail before the point of no return.
If we could open up all the 32-bit apps out there and fix them, then I'd be
more sympathetic, but the reality is that we need to run existing binaries
on these stupid systems and exec'ing 32-bit payloads from 64-bit tasks is
something that we need to continue to support.
If it makes things any better, all of this stuff is off by default and gated
on a cmdline option.
Will
