On Wed, Feb 13, 2019 at 02:17:20PM +0100, Peter Zijlstra wrote:
> On Wed, Feb 13, 2019 at 10:50:21AM +0000, Julien Thierry wrote:
> > On 13/02/2019 10:35, Peter Zijlstra wrote:
> > > On Tue, Feb 12, 2019 at 09:15:13AM +0000, Julien Thierry wrote:
> > >
> > >>>>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > >>>>> index a674c7db..b1bb7e9 100644
> > >>>>> --- a/kernel/sched/core.c
> > >>>>> +++ b/kernel/sched/core.c
> > >>>>> @@ -3289,6 +3289,14 @@ static inline void schedule_debug(struct
> > >>>>> task_struct *prev)
> > >>>>> __schedule_bug(prev);
> > >>>>> preempt_count_set(PREEMPT_DISABLED);
> > >>>>> }
> > >>>>> +
> > >>>>> + if (IS_ENABLED(CONFIG_DEBUG_UACCESS_SLEEP) &&
> > >>>>> + unlikely(unsafe_user_region_active())) {
> > >>>>> + printk(KERN_ERR "BUG: scheduling while user_access
> > >>>>> enabled: %s/%d/0x%08x\n",
> > >>>>> + prev->comm, prev->pid, preempt_count());
> > >>>>> + dump_stack();
> > >>>>> + }
> > >>>>> +
> > >>>>> rcu_sleep_check();
> > >>>>>
> > >>>>> profile_hit(SCHED_PROFILING, __builtin_return_address(0));
> > >
> > >> I guess I'll drop the might_resched() part of this patch if that sounds
> > >> alright.
> > >
> > > I'm still confused by the schedule_debug() part. How is that not broken?
> >
> > Hmmm, I am not exactly sure which part you expect to be broken, I guess
> > it's because of the nature of the uaccess unsafe accessor usage.
> >
> > Basically, the following is a definite no:
> > if (user_access_begin(ptr, size)) {
> >
> > [...]
> >
> > //something that calls schedule
> >
> > [...]
> >
> > user_access_end();
> > }
> >
> >
> > However the following is fine:
> >
> > - user_access_begin(ptr, size)
> > - taking irq/exception
> > - get preempted
>
> This; how is getting preempted fundamentally different from scheduling
> ourselves?
This is also the thing that PREEMPT_VOLUNTARY hinges on; it inserts
'random' reschedule points through might_sleep() and cond_resched().
If you're preemptible; you must be able to schedule and vice-versa.
You're breaking that.
> > - get resumed at some point in time
> > - restore state + eret
> > - user_access_end()
> >
> > That's because exceptions/irq implicitly "suspend" the user access
> > region. (That's what I'm trying to clarify with the comment)
> > So, unsafe_user_region_active() should return false in a irq/exception
> > context.
> >
> > Is this what you were concerned about? Or there still something that
> > might be broken?
>
> I really hate the asymetry introduced between preemptible and being able
> to schedule. Both end up calling __schedule() and there really should
> not be a difference.
