On Fri, Sep 11, 2026 at 09:50:09AM +0800, KunWu Chan wrote:
> On Fri, Sep 11, 2026 at 12:58 AM Paul E. McKenney <[email protected]> wrote:
> >
> > On Thu, Sep 10, 2026 at 04:46:29PM +0800, Kunwu Chan wrote:
> > > Atomic SRCU remain in the small size state, so add a
> > > WARN_ON_ONCE() before the transition check to catch any future
> > > code path that might violate this invariant.
> > >
> > > Signed-off-by: Kunwu Chan <[email protected]>
> > > ---
> > > Changes since v1:
> > > - Drop the !is_atomic guard from v1 and add WARN_ON_ONCE() instead,
> > > per Paul McKenney's suggestion.
> > > - v1:
> > > https://lore.kernel.org/rcu/[email protected]/
> > > ---
> > > kernel/rcu/srcutree.c | 2 ++
> > > 1 file changed, 2 insertions(+)
> > >
> > > diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
> > > index 84c022ec8e09..68441c32cebf 100644
> > > --- a/kernel/rcu/srcutree.c
> > > +++ b/kernel/rcu/srcutree.c
> > > @@ -1074,6 +1074,8 @@ static void srcu_gp_end(struct srcu_struct *ssp,
> > > bool is_atomic)
> > > }
> > >
> > > /* Transition to big if needed. */
> > > + WARN_ON_ONCE(ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC &&
> > > + ss_state != SRCU_SIZE_SMALL);
> >
> > Also much better, but we also need to avoid doing the transition in this
> > case, because doing that transition breaks atomic SRCU. On way to make
> > this happen is something like this:
> >
> > if (ssp->srcu_reader_flavor == SRCU_READ_FLAVOR_ATOMIC && ss_state
> > != SRCU_SIZE_SMALL) {
> > WARN_ON_ONCE(1);
> > } else if (ss_state != SRCU_SIZE_SMALL && ss_state !=
> > SRCU_SIZE_BIG) {
> > if (ss_state == SRCU_SIZE_ALLOC)
> > init_srcu_struct_nodes(ssp, GFP_KERNEL);
> >
>
> Thanks, Paul. Good point. The original !is_atomic guard in v1
> already prevented the transition for atomic SRCU. I was mainly
> thinking of the WARN as a diagnostic, since this state should not
> occur with the current code. I agree that retaining the protection
> against the transition is safer if the invariant is ever violated.
>
> > Other approaches might take advantage of the fact that WARN_ON_ONCE()
> > returns the value of its argument, allowing WARN_ON_ONCE() to be used
> > as a condition in an "if" statement.
> >
>
> I prefer the first approach. Using !WARN_ON_ONCE() directly in the
> existing condition would make it rather long and less readable,
> while the explicit if/else keeps the control flow clear.
Very good, looking forward to seeing what you come up with.
Thanx, Paul
> Thanks,
> Kunwu
>
> > Thanx, Paul
> >
> > > if (ss_state != SRCU_SIZE_SMALL && ss_state != SRCU_SIZE_BIG) {
> > > if (ss_state == SRCU_SIZE_ALLOC)
> > > init_srcu_struct_nodes(ssp, GFP_KERNEL);
> > > --
> > > 2.43.0
> > >
