>
> On 17 October 2026 00:41:00 BST, Zqiang <[email protected]> wrote:
>
> >
> > >
> > > On 15 October 2026 10:46:19 BST, Zqiang <[email protected]> wrote:
> > >
> > > >
> > > > The RCU Tasks Rude grace period is implemented by invoking the
> > > > schedule_on_each_cpu() with an empty work function, which forces
> > > > a context switch on every online CPU and guarantees that all
> > > > rude-tasks read-side critical sections have completed. However,
> > > > schedule_on_each_cpu() can fail (via alloc_percpu() returning
> > > > -ENOMEM) and returns immediately, in which case those context
> > > > switches never happen, too-short grace period happen.
> > > >
> > > > This commit therefore wraps the schedule_on_each_cpu() call in
> > > > WARN_ONCE() so that this failure is reported as a "Too-short
> > > > grace period" warning.
> > > >
> > > NAK, this seems like a edge case, and potentially panicking the system
> > > with panic on oops doesn't seem right to me, because we would reserve
> > >
> > that
> >
> > >
> > > for when something definitely goes really wrong and we need to notify
> > > the user.
> > >
> > > I don't believe this patch should go in for this reason, it's a heavy
> > >
> > edge
> >
> > >
> > > case, that doesn't *have* nessesarily need to be accounted for.
> > >
> > Indeed, this is a rare event, so I use WARN_ONCE() to mark it, the
> > WARN_ONCE()
> > not panicking system unless you set warn_on_oops true, this can help us
> > troubleshoot
> > UAF issues caused by a too short tasks-rude grace period.
> >
> > or maybe we can use DEFINE_PER_CPU(struct work_struct, rude_work) to
> > replace
> > alloc_percpu(struct work_struct) queue to all online CPUs.
> >
> > Anyway, Thank you for your reply. I would also like to hear the opinions
> > of others.
> >
> Fair enough, but me and Paul done some discussions:
>
> Here is the TL;DR
>
> My ^W^Wai slop^W^W tool helped me with this: if the too short grace
> happens,
> it falls back.
>
> root@p10-linux-brads:~/linux# mailto:root@p10-linux-brads:~/linux# git diff
> kernel/rcu
> diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
> index 627295396cd9..65bb7d4ed632 100644
> --- a/kernel/rcu/tasks.h
> +++ b/kernel/rcu/tasks.h
> @@ -1344,15 +1344,23 @@ void exit_tasks_rcu_finish(void) { }
> // Ordering is provided by the scheduler's context-switch code.
>
> // Empty function to allow workqueues to force a context switch.
> -static void rcu_tasks_be_rude(struct work_struct *work)
> +static void __maybe_unused rcu_tasks_be_rude(struct work_struct *work)
> {
> }
>
> // Wait for one rude RCU-tasks grace period.
> +static void rcu_tasks_rude_ipi(void *info)
> +{
> +}
> +
> static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
> {
> + int cpu;
> rtp->n_ipis += cpumask_weight(cpu_online_mask);
> - schedule_on_each_cpu(rcu_tasks_be_rude);
> + cpus_read_lock();
> + for_each_online_cpu(cpu)
> + smp_call_function_single(cpu, rcu_tasks_rude_ipi, NULL, 1);
> + cpus_read_unlock();
> }
RCU Tasks Rude use preempt_disable/enable as the read critical section,
you use smp_call_function_single() to send ipi to each online CPUs,
it's not forces a context sched switch, and not wait the preempt_disable/enable
read critical section to exit, this is the wrong mechanism. ;-(
Obviously, we only need to not apply alloc_percpu dynamic allocation:
diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index fcac7361ec51..49b629b0010b 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -1357,11 +1357,25 @@ static void rcu_tasks_be_rude(struct work_struct *work)
{
}
+static DEFINE_PER_CPU(struct work_struct, tasks_rude_work);
+
// Wait for one rude RCU-tasks grace period.
static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
{
+ int cpu;
+
+ guard(cpus_read_lock)();
rtp->n_ipis += cpumask_weight(cpu_online_mask);
- schedule_on_each_cpu(rcu_tasks_be_rude);
+
+ for_each_online_cpu(cpu) {
+ struct work_struct *work = &per_cpu(tasks_rude_work, cpu);
+
+ INIT_WORK(work, rcu_tasks_be_rude);
+ schedule_work_on(cpu, work);
+ }
+
+ for_each_online_cpu(cpu)
+ flush_work(&per_cpu(tasks_rude_work, cpu));
}
static void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
Thanks
Zqiang
>
> static void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
> root@p10-linux-brads:~/linux# mailto:root@p10-linux-brads:~/linux#
>
> On POWERPC, this passes 12 hours of rcutorture on the RUDE01 config.
>
> If you plan on submitting this into a patch, please add my authorship, or
> my c-d-b.
>
> Feel free to rcutorture test it ;d
>
> BTW: have you reproduced this on your own system?
>
> >
> > Thanks
> > Zqiang
> >
> > >
> > > Zqiang, have you reproduced this OOM on your own system, if so, could
> > >
> > you
> >
> > >
> > > tell me more details, you may turn my frown into a smiley face ;)
> > >
> > > >
> > > > Signed-off-by: Zqiang <[email protected]>
> > > > ---
> > > > kernel/rcu/tasks.h | 3 ++-
> > > > 1 file changed, 2 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
> > > > index fcac7361ec51..91c25717402e 100644
> > > > --- a/kernel/rcu/tasks.h
> > > > +++ b/kernel/rcu/tasks.h
> > > > @@ -1361,7 +1361,8 @@ static void rcu_tasks_be_rude(struct work_struct
> > >
> > *work)
> >
> > >
> > > > static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
> > > > {
> > > > rtp->n_ipis += cpumask_weight(cpu_online_mask);
> > > > - schedule_on_each_cpu(rcu_tasks_be_rude);
> > > > + WARN_ONCE(schedule_on_each_cpu(rcu_tasks_be_rude),
> > > > + "%s: Too-short grace period, OOM is now expected behavior!!!\n",
> > >
> > rtp->name);
> >
> > >
> > > >
> > > Good message, makes sense. But I don't think this is right, see above.
> > >
> > > >
> > > > }
> > > >
> > > > static void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t
> > > > func);
> > > >
> > > --- Thanks!
> > > "I'm not a very positive person" - Linus torvalds
> > >
> >
> --- Thanks!
> "I'm not a very positive person" - Linus torvalds
>
