Re: [PATCH 1/1] rcu/tree: Drop the lock before entering to page allocator

[Uladzislau Rezki]

On Thu, Jul 16, 2020 at 02:27:07PM -0400, Joel Fernandes wrote:
> On Thu, Jul 16, 2020 at 04:37:14PM +0200, Uladzislau Rezki wrote:
> > On Thu, Jul 16, 2020 at 09:36:47AM -0400, Joel Fernandes wrote:
> > > On Thu, Jul 16, 2020 at 11:19:13AM +0200, Uladzislau Rezki wrote:
> > > > On Wed, Jul 15, 2020 at 07:13:33PM -0400, Joel Fernandes wrote:
> > > > > On Wed, Jul 15, 2020 at 2:56 PM Sebastian Andrzej Siewior
> > > > > <bigeasy@xxxxxxxxxxxxx> wrote:
> > > > > >
> > > > > > On 2020-07-15 20:35:37 [+0200], Uladzislau Rezki (Sony) wrote:
> > > > > > > @@ -3306,6 +3307,9 @@ kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
> > > > > > >                       if (IS_ENABLED(CONFIG_PREEMPT_RT))
> > > > > > >                               return false;
> > > > > > >
> > > > > > > +                     preempt_disable();
> > > > > > > +                     krc_this_cpu_unlock(*krcp, *flags);
> > > > > >
> > > > > > Now you enter memory allocator with disabled preemption. This isn't any
> > > > > > better but we don't have a warning for this yet.
> > > > > > What happened to the part where I asked for a spinlock_t?
> > > > > 
> > > > > Ulad,
> > > > > Wouldn't the replacing of preempt_disable() with migrate_disable()
> > > > > above resolve Sebastian's issue?
> > > > >
> > > > This for regular kernel only. That means that migrate_disable() is
> > > > equal to preempt_disable(). So, no difference.
> > > 
> > > But this will force preempt_disable() context into the low-level page
> > > allocator on -RT kernels which I believe is not what Sebastian wants. The
> > > whole reason why the spinlock vs raw-spinlock ordering matters is, because on
> > > RT, the spinlock is sleeping. So if you have:
> > > 
> > > raw_spin_lock(..);
> > > spin_lock(..);   <-- can sleep on RT, so Sleep while atomic (SWA) violation.
> > > 
> > > That's the main reason you are dropping the lock before calling the
> > > allocator.
> > > 
> > No. Please read the commit message of this patch. This is for regular kernel.
> 
> Wait, so what is the hesitation to put migrate_disable() here? It is even
> further documentation (annotation) that the goal here is to stay on the same
> CPU - as you indicated in later emails.
> 
Actually preempt_disable() does the same for !RT. I agree that
migrate_disable() annotation looks better from the point you
mentioned.

> And the documentation aspect is also something Sebastian brought. A plain
> preempt_disable() is frowned up if there are alternative API that document
> the usage.
> 
> > You did a patch:
> > 
> > <snip>
> >    if (IS_ENABLED(CONFIG_PREEMPT_RT))
> >        return false;
> > <snip>
> 
> I know, that's what we're discussing.
> 
> So again, why the hatred for migrate_disable() ? :)
> 
Let's do migrate_disable(), i do not mind :)

--
Vlad Rezki