Re: [PATCH v2 03/12] rcu: Support runtime NOCB initialization and dynamic offloading

[Frederic Weisbecker]

Le Mon, Apr 13, 2026 at 03:43:09PM +0800, Qiliang Yuan a écrit :
> Context:
> The RCU Non-Callback (NOCB) infrastructure traditionally requires
> boot-time parameters (e.g., rcu_nocbs) to allocate masks and spawn
> management kthreads (rcuog/rcuo). This prevents systems from activating
> offloading on-demand without a reboot.
> 
> Problem:
> Dynamic Housekeeping Management requires CPUs to transition to
> NOCB mode at runtime when they are newly isolated. Without boot-time
> setup, the NOCB masks are unallocated, and critical kthreads are missing,
> preventing effective tick suppression and isolation.
> 
> Solution:
> Refactor RCU initialization to support dynamic on-demand setup.
> - Introduce rcu_init_nocb_dynamic() to allocate masks and organize
>   kthreads if the system wasn't initially configured for NOCB.
> - Introduce rcu_housekeeping_reconfigure() to iterate over CPUs and
>   perform safe offload/deoffload transitions via hotplug sequences
>   (cpu_down -> offload -> cpu_up) when a housekeeping cpuset triggers
>   a notifier event.
> - Remove __init from rcu_organize_nocb_kthreads to allow runtime
>   reconfiguration of the callback management hierarchy.
> 
> This enables a true "Zero-Conf" isolation experience where any CPU
> can be fully isolated at runtime regardless of boot parameters.
> 
> Signed-off-by: Qiliang Yuan <realwujing@xxxxxxxxx>
> ---
>  kernel/rcu/rcu.h       |  4 +++
>  kernel/rcu/tree.c      | 75 ++++++++++++++++++++++++++++++++++++++++++++++++++
>  kernel/rcu/tree.h      |  2 +-
>  kernel/rcu/tree_nocb.h | 31 +++++++++++++--------
>  4 files changed, 100 insertions(+), 12 deletions(-)
> 
> diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
> index 9b10b57b79ada..282874443c96b 100644
> --- a/kernel/rcu/rcu.h
> +++ b/kernel/rcu/rcu.h
> @@ -663,8 +663,12 @@ unsigned long srcu_batches_completed(struct srcu_struct *sp);
>  #endif // #else // #ifdef CONFIG_TINY_SRCU
>  
>  #ifdef CONFIG_RCU_NOCB_CPU
> +void rcu_init_nocb_dynamic(void);
> +void rcu_spawn_cpu_nocb_kthread(int cpu);
>  void rcu_bind_current_to_nocb(void);
>  #else
> +static inline void rcu_init_nocb_dynamic(void) { }
> +static inline void rcu_spawn_cpu_nocb_kthread(int cpu) { }
>  static inline void rcu_bind_current_to_nocb(void) { }
>  #endif
>  
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 55df6d37145e8..84c8388cf89a1 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -4928,4 +4928,79 @@ void __init rcu_init(void)
>  #include "tree_stall.h"
>  #include "tree_exp.h"
>  #include "tree_nocb.h"
> +
> +#ifdef CONFIG_SMP
> +static int rcu_housekeeping_reconfigure(struct notifier_block *nb,
> +					unsigned long action, void *data)
> +{
> +	struct housekeeping_update *upd = data;
> +	struct task_struct *t;
> +	int cpu;
> +
> +	if (action != HK_UPDATE_MASK || upd->type != HK_TYPE_RCU)
> +		return NOTIFY_OK;
> +
> +	rcu_init_nocb_dynamic();
> +
> +	for_each_possible_cpu(cpu) {
> +		struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
> +		bool isolated = !cpumask_test_cpu(cpu, upd->new_mask);
> +		bool offloaded = rcu_rdp_is_offloaded(rdp);
> +
> +		if (isolated && !offloaded) {
> +			/* Transition to NOCB */
> +			pr_info("rcu: CPU %d transitioning to NOCB mode\n", cpu);
> +			if (cpu_online(cpu)) {
> +				remove_cpu(cpu);

We plan to assume that the CPU is offline while updating HK_TYPE_KERNEL_NOISE
through cpusets. So you shouldn't need to care about offlining here.


> +				rcu_spawn_cpu_nocb_kthread(cpu);
> +				rcu_nocb_cpu_offload(cpu);
> +				add_cpu(cpu);
> +			} else {
> +				rcu_spawn_cpu_nocb_kthread(cpu);
> +				rcu_nocb_cpu_offload(cpu);
> +			}
> +		} else if (!isolated && offloaded) {
> +			/* Transition to CB */
> +			pr_info("rcu: CPU %d transitioning to CB mode\n", cpu);
> +			if (cpu_online(cpu)) {
> +				remove_cpu(cpu);
> +				rcu_nocb_cpu_deoffload(cpu);
> +				add_cpu(cpu);
> +			} else {
> +				rcu_nocb_cpu_deoffload(cpu);
> +			}
> +		}
> +	}
> +
> +	t = READ_ONCE(rcu_state.gp_kthread);
> +	if (t)
> +		housekeeping_affine(t, HK_TYPE_RCU);
> +
> +#ifdef CONFIG_TASKS_RCU
> +	t = get_rcu_tasks_gp_kthread();
> +	if (t)
> +		housekeeping_affine(t, HK_TYPE_RCU);
> +#endif
> +
> +#ifdef CONFIG_TASKS_RUDE_RCU
> +	t = get_rcu_tasks_rude_gp_kthread();
> +	if (t)
> +		housekeeping_affine(t, HK_TYPE_RCU);
> +#endif

No need to handle kthreads affinities. This is already taken care of by isolated
cpuset partitions.

Thanks.

-- 
Frederic Weisbecker
SUSE Labs