On Fri, Aug 15, 2025 at 11:31 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
>
> On Fri, Aug 15, 2025 at 04:02:14PM +0300, Alexei Starovoitov wrote:
> > On Fri, Aug 15, 2025 at 9:18 AM Menglong Dong <menglong8.dong@xxxxxxxxx> wrote:
> > >
> > > migrate_disable() is called to disable migration in the kernel, and it is
> > > used togather with rcu_read_lock() oftenly.
> > >
> > > However, with PREEMPT_RCU disabled, it's unnecessary, as rcu_read_lock()
> > > will disable preemption, which will also disable migration.
> > >
> > > Introduce rcu_migrate_enable() and rcu_migrate_disable(), which will do
> > > the migration enable and disable only when the rcu_read_lock() can't do
> > > it.
> > >
> > > Signed-off-by: Menglong Dong <dongml2@xxxxxxxxxxxxxxx>
> > > ---
> > > include/linux/rcupdate.h | 18 ++++++++++++++++++
> > > 1 file changed, 18 insertions(+)
> > >
> > > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > > index 120536f4c6eb..0d9dbd90d025 100644
> > > --- a/include/linux/rcupdate.h
> > > +++ b/include/linux/rcupdate.h
> > > @@ -72,6 +72,16 @@ static inline bool same_state_synchronize_rcu(unsigned long oldstate1, unsigned
> > > void __rcu_read_lock(void);
> > > void __rcu_read_unlock(void);
> > >
> > > +static inline void rcu_migrate_enable(void)
> > > +{
> > > + migrate_enable();
> > > +}
> >
> > Interesting idea.
> > I think it has to be combined with rcu_read_lock(), since this api
> > makes sense only when used together.
> >
> > rcu_read_lock_dont_migrate() ?
> >
> > It will do rcu_read_lock() + migrate_disalbe() in PREEMPT_RCU
> > and rcu_read_lock() + preempt_disable() otherwise?
>
> That could easily be provided. Or just make one, and if it starts
> having enough use cases, it could be pulled into RCU proper.
Hi, do you mean that we should start with a single
use case? In this series, I started it with the BPF
subsystem. Most of the situations are similar, which will
call rcu_read_lock+migrate_disable and run bpf prog.
>
> > Also I'm not sure we can rely on rcu_read_lock()
> > disabling preemption in all !PREEMPT_RCU cases.
> > iirc it's more nuanced than that.
>
> For once, something about RCU is non-nuanced. But don't worry, it won't
> happen again. ;-)
>
> In all !PREEMPT_RCU, preemption must be disabled across all RCU read-side
> critical sections in order for RCU to work correctly.
Great! I worried about this part too.
Thanks!
Menglong Dong
>
> Thanx, Paul
