Re: [PATCH -next v2 10/11] rcu: Latch normal synchronize_rcu() path on flood

[Frederic Weisbecker]

Le Tue, May 19, 2026 at 09:45:23PM +0200, Uladzislau Rezki (Sony) a écrit :
> Currently, rcu_normal_wake_from_gp is only enabled by default
> on small systems(<= 16 CPUs) or when a user explicitly set it
> enabled.
> 
> Introduce an adaptive latching mechanism:
>  * Track the number of in-flight synchronize_rcu() requests
>    using a new rcu_sr_normal_count counter;
> 
>  * If the count reaches/exceeds RCU_SR_NORMAL_LATCH_THR(64),
>    it sets the rcu_sr_normal_latched, reverting new requests
>    onto the scaled wait_rcu_gp() path;
> 
>  * The latch is cleared only when the pending requests are fully
>    drained(nr == 0);
> 
>  * Enables rcu_normal_wake_from_gp by default for all systems,
>    relying on this dynamic throttling instead of static CPU
>    limits.
> 
> Testing(synthetic flood workload):
>   * Kernel version: 6.19.0-rc6
>   * Number of CPUs: 1536
>   * 60K concurrent synchronize_rcu() calls
> 
> Perf(cycles, system-wide):
>   total cycles: 932020263832
>   rcu_sr_normal_add_req(): 2650282811 cycles(~0.28%)
> 
> Perf report excerpt:
>   0.01%  0.01%  sync_test/...  [k] rcu_sr_normal_add_req
> 
> Measured overhead of rcu_sr_normal_add_req() remained ~0.28%
> of total CPU cycles in this synthetic stress test.
> 
> Tested-by: Samir M <samir@xxxxxxxxxxxxx>
> Suggested-by: Joel Fernandes <joelagnelf@xxxxxxxxxx>
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
> ---
>  .../admin-guide/kernel-parameters.txt         | 10 ++--
>  kernel/rcu/tree.c                             | 52 ++++++++++++++-----
>  2 files changed, 44 insertions(+), 18 deletions(-)
> 
> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> index 4d0f545fb3ec..d5db2e85d551 100644
> --- a/Documentation/admin-guide/kernel-parameters.txt
> +++ b/Documentation/admin-guide/kernel-parameters.txt
> @@ -5862,13 +5862,13 @@ Kernel parameters
>  			use a call_rcu[_hurry]() path. Please note, this is for a
>  			normal grace period.
>  
> -			How to enable it:
> +			How to disable it:
>  
> -			echo 1 > /sys/module/rcutree/parameters/rcu_normal_wake_from_gp
> -			or pass a boot parameter "rcutree.rcu_normal_wake_from_gp=1"
> +			echo 0 > /sys/module/rcutree/parameters/rcu_normal_wake_from_gp
> +			or pass a boot parameter "rcutree.rcu_normal_wake_from_gp=0"
>  
> -			Default is 1 if num_possible_cpus() <= 16 and it is not explicitly
> -			disabled by the boot parameter passing 0.
> +			Default is 1 if it is not explicitly disabled by the boot parameter
> +			passing 0.
>  
>  	rcuscale.gp_async= [KNL]
>  			Measure performance of asynchronous
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 09f0cef5014c..94274330d1db 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -1632,17 +1632,21 @@ static void rcu_sr_put_wait_head(struct llist_node *node)
>  	atomic_set_release(&sr_wn->inuse, 0);
>  }
>  
> -/* Enable rcu_normal_wake_from_gp automatically on small systems. */
> -#define WAKE_FROM_GP_CPU_THRESHOLD 16
> -
> -static int rcu_normal_wake_from_gp = -1;
> +static int rcu_normal_wake_from_gp = 1;
>  module_param(rcu_normal_wake_from_gp, int, 0644);
>  static struct workqueue_struct *sync_wq;
>  
> +#define RCU_SR_NORMAL_LATCH_THR 64
> +
> +/* Number of in-flight synchronize_rcu() calls queued on srs_next. */
> +static atomic_long_t rcu_sr_normal_count;
> +static int rcu_sr_normal_latched; /* 0/1 */
> +
>  static void rcu_sr_normal_complete(struct llist_node *node)
>  {
>  	struct rcu_synchronize *rs = container_of(
>  		(struct rcu_head *) node, struct rcu_synchronize, head);
> +	long nr;
>  
>  	WARN_ONCE(IS_ENABLED(CONFIG_PROVE_RCU) &&
>  		!poll_state_synchronize_rcu_full(&rs->oldstate),
> @@ -1650,6 +1654,15 @@ static void rcu_sr_normal_complete(struct llist_node *node)
>  
>  	/* Finally. */
>  	complete(&rs->completion);
> +	nr = atomic_long_dec_return(&rcu_sr_normal_count);
> +	WARN_ON_ONCE(nr < 0);
> +
> +	/*
> +	 * Unlatch: switch back to normal path when fully
> +	 * drained and if it has been latched.
> +	 */
> +	if (nr == 0)
> +		(void)cmpxchg(&rcu_sr_normal_latched, 1, 0);

Given that it's already ordered by the llist add / del and the
atomic_long_inc/dec_return, there should be no chance for bad
things happening such as negative returned dec.

So it could be cmpxchg_relaxed(). But anyway, just an optimization.

In any case,

Reviewed-by: Frederic Weisbecker <frederic@xxxxxxxxxx>

-- 
Frederic Weisbecker
SUSE Labs