On Wed, Nov 20, 2019 at 02:58:27AM +0800, Herbert Xu wrote:
> /* Replace the internal control structure with a new one. */
> -static void padata_replace(struct padata_instance *pinst,
> - struct parallel_data *pd_new)
> +static int padata_replace_one(struct padata_shell *ps)
> {
> - struct parallel_data *pd_old = pinst->pd;
> + struct parallel_data *pd_old = rcu_dereference_protected(ps->pd, 1);
> + struct parallel_data *pd_new;
> int notification_mask = 0;
>
> - pinst->flags |= PADATA_RESET;
> -
> - rcu_assign_pointer(pinst->pd, pd_new);
> + pd_new = padata_alloc_pd(ps);
> + if (!pd_new)
> + return -ENOMEM;
>
> - synchronize_rcu();
> + rcu_assign_pointer(ps->pd, pd_new);
>
> if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
> notification_mask |= PADATA_CPU_PARALLEL;
> @@ -510,10 +515,25 @@ static void padata_replace(struct padata_instance *pinst,
> if (atomic_dec_and_test(&pd_old->refcnt))
> padata_free_pd(pd_old);
>
> + return notification_mask;
> +}
> +
> +static void padata_replace(struct padata_instance *pinst)
> +{
> + int notification_mask = 0;
> + struct padata_shell *ps;
> +
> + pinst->flags |= PADATA_RESET;
> +
> + list_for_each_entry(ps, &pinst->pslist, list)
> + notification_mask |= padata_replace_one(ps);
> +
> + synchronize_rcu();
> +
> if (notification_mask)
> blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
> notification_mask,
> - &pd_new->cpumask);
> + &pinst->cpumask);
>
> pinst->flags &= ~PADATA_RESET;
> }
I think it's possible for a task in padata_do_parallel() racing with another in
padata_replace() to use a pd after free. The synchronize_rcu() comes after the
pd_old->refcnt's are dec'd.
padata_do_parallel()
rcu_dereference_bh(ps->pd)
// doesn't see PADATA_RESET set
padata_replace()
pinst->flags |= PADATA_RESET
padata_replace_one()
rcu_assign_pointer(ps->pd, pd_new)
atomic_dec_and_test(&pd_old->refcnt)
padata_free_pd()
atomic_inc(&pd->refcnt) // too late
If I'm not missing something, one way out is adding a list_head to
parallel_data for remembering the old pd's on a local list in padata_replace()
so that this function can loop over it and drop the refs after
synchronize_rcu(). A padata_do_parallel() call will have then had a chance to
take a ref on the pd it's using.
And, not this patch, but with the removal of flushing it seems there's no need
for PADATA_RESET, so it and its EBUSY error can go away.
