On 2020-09-18 10:22:32 [+0200], pet...@infradead.org wrote:
> > > One reason for not allowing migrate_disable() to sleep was: FPU code.
> > >
> > > Could it be it does something like:
> > >
> > > preempt_disable();
> > > spin_lock();
> > >
> > > spin_unlock();
> > > preempt_enable();
> > >
> > > Where we'll never get preempted while migrate_disable()'d and thus never
> > > trigger any of the sleep paths?
> >
> > I try to get rid of something like that. This doesn't work either way
> > because the spin_lock() may block which it can't with disabled
> > preemption.
>
> Yeah, that obviously should have been migrate_disable/enable instead of
> spin_lock/unlock :/
Ah. Me stupid. fpregs_lock() does
preempt_disable();
local_bh_disable();
which is more or less the "official" pattern. As of today
local_bh_disable() does migrate_disable() / spin_lock(). Not sure what
we end up with for local_bh_disable() in the end.
We used not have a BLK here on RT but ended up in all kind of locking
problems because vanilla treats local_bh_disable() as a BLK and uses it
for locking.
Today we have a per-CPU spinlock_t in local_bh_disable() to emulate the
BKL. But this pattern above isn't working due to the atomic part…
Sebastian
