On Wed, Feb 13, 2019 at 02:24:34PM +0000, Julien Thierry wrote:
> On 13/02/2019 14:17, Peter Zijlstra wrote:
> > On Wed, Feb 13, 2019 at 02:00:26PM +0000, Will Deacon wrote:
> >>> This; how is getting preempted fundamentally different from scheduling
> >>> ourselves?
> >>
> >> The difference is because getting preempted in the sequence above is
> >> triggered off the back of an interrupt. On arm64, and I think also on x86,
> >> the user access state (SMAP or PAN) is saved and restored across exceptions
> >> but not across context switch. Consequently, taking an irq in a
> >> user_access_{begin,end} section and then scheduling is fine, but calling
> >> schedule directly within such a section is not.
> >
> > So how's this then:
> >
> > if (user_access_begin()) {
> >
> > preempt_disable();
> >
> > <IRQ>
> > set_need_resched();
> > </IRQ no preempt>
> >
> > preempt_enable()
> > __schedule();
> >
> > user_access_end();
> > }
> >
> > That _should_ work just fine but explodes with the proposed nonsense.
>
> AFAICT, This does not work properly because when you schedule out this
> task, you won't be saving the EFLAGS.AF/PSTATE.PAN bit on the stack, and
EFLAGS.AC, but yes.
> next time you schedule the task back in, it might no longer have the
> correct flag value (so an unsafe_get/put_user() will fail even though
> you haven't reached user_access_end()).
/me looks at __switch_to_asm() and there is indeed a distinct lack of
pushing and popping EFLAGS :/
> One solution is to deal with this in task switching code, but so far
> I've been told that calling schedule() in such a context is not expected
> to be supported.
Well, per the above it breaks the preemption model. And I hates that.
And the added WARN doesn't even begin to cover it, since you'd have to
actually hit the preempt_enable() reschedule for it trigger.
So far, all 6 in-tree users are indeed free of dodgy code, but *groan*.
