On Thu, Jun 02, 2022 at 10:06:44AM +0200, Uladzislau Rezki (Sony) wrote:
> Currently the monitor work is scheduled with a fixed interval that
> is HZ/20 or each 50 milliseconds. The drawback of such approach is
> a low utilization of page slot in some scenarios. The page can store
> up to 512 records. For example on Android system it can look like:
>
> <snip>
> kworker/3:0-13872 [003] .... 11286.007048: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000026522604 nr_records=1
> kworker/3:0-13872 [003] .... 11286.015638: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000095ed6fca nr_records=2
> kworker/1:2-20434 [001] .... 11286.051230: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044872ffd nr_records=1
> kworker/1:2-20434 [001] .... 11286.059322: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000026522604 nr_records=2
> kworker/0:1-20052 [000] .... 11286.095295: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044872ffd nr_records=2
> kworker/0:1-20052 [000] .... 11286.103418: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000cbcf05db nr_records=1
> kworker/2:3-14372 [002] .... 11286.135155: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000095ed6fca nr_records=2
> kworker/2:3-14372 [002] .... 11286.135198: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044872ffd nr_records=1
> kworker/1:2-20434 [001] .... 11286.155377: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000cbcf05db nr_records=5
> kworker/2:3-14372 [002] .... 11286.167181: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000026522604 nr_records=5
> kworker/1:2-20434 [001] .... 11286.179202: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000008ef95e14 nr_records=1
> kworker/2:3-14372 [002] .... 11286.187398: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000c597d297 nr_records=6
> kworker/3:0-13872 [003] .... 11286.187445: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000050bf92e2 nr_records=3
> kworker/1:2-20434 [001] .... 11286.198975: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000cbcf05db nr_records=4
> kworker/1:2-20434 [001] .... 11286.207203: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000095ed6fca nr_records=4
> <snip>
>
> where a page only carries few records to reclaim a memory. In order to
> improve batching and make utilization more efficient the patch introduces
> a drain interval that can be set either to KFREE_DRAIN_JIFFIES_MAX or
> KFREE_DRAIN_JIFFIES_MIN. It is adjusted if a flood is detected, in this
> case a memory reclaim occurs more often whereas in mostly idle cases the
> interval is set to its maximum timeout that improves the utilization of
> page slots.
>
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
That does look like a problem well worth solving!
But I am missing one thing. If we are having a callback flood, why do we
need a shorter timeout? Wouldn't a check on the number of blocks queued
be simpler, more direct, and provide faster response to the start of a
callback flood?
Thanx, Paul
> ---
> kernel/rcu/tree.c | 29 +++++++++++++++++++++++++----
> 1 file changed, 25 insertions(+), 4 deletions(-)
>
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index fd16c0b46d9e..c02a64995b85 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -3249,7 +3249,8 @@ EXPORT_SYMBOL_GPL(call_rcu);
>
>
> /* Maximum number of jiffies to wait before draining a batch. */
> -#define KFREE_DRAIN_JIFFIES (HZ / 50)
> +#define KFREE_DRAIN_JIFFIES_MAX (HZ)
> +#define KFREE_DRAIN_JIFFIES_MIN (HZ / 50)
> #define KFREE_N_BATCHES 2
> #define FREE_N_CHANNELS 2
>
> @@ -3510,6 +3511,26 @@ need_offload_krc(struct kfree_rcu_cpu *krcp)
> return !!krcp->head;
> }
>
> +static void
> +schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
> +{
> + long delay, delay_left;
> +
> + delay = READ_ONCE(krcp->count) >= KVFREE_BULK_MAX_ENTR ?
> + KFREE_DRAIN_JIFFIES_MIN:KFREE_DRAIN_JIFFIES_MAX;
> +
> + if (delayed_work_pending(&krcp->monitor_work)) {
> + delay_left = krcp->monitor_work.timer.expires - jiffies;
> +
> + if (delay < delay_left)
> + mod_delayed_work(system_wq, &krcp->monitor_work, delay);
> +
> + return;
> + }
> +
> + queue_delayed_work(system_wq, &krcp->monitor_work, delay);
> +}
> +
> /*
> * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
> */
> @@ -3567,7 +3588,7 @@ static void kfree_rcu_monitor(struct work_struct *work)
> // work to repeat an attempt. Because previous batches are
> // still in progress.
> if (need_offload_krc(krcp))
> - schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
> + schedule_delayed_monitor_work(krcp);
>
> raw_spin_unlock_irqrestore(&krcp->lock, flags);
> }
> @@ -3755,7 +3776,7 @@ void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
>
> // Set timer to drain after KFREE_DRAIN_JIFFIES.
> if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
> - schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
> + schedule_delayed_monitor_work(krcp);
>
> unlock_return:
> krc_this_cpu_unlock(krcp, flags);
> @@ -3831,7 +3852,7 @@ void __init kfree_rcu_scheduler_running(void)
>
> raw_spin_lock_irqsave(&krcp->lock, flags);
> if (need_offload_krc(krcp))
> - schedule_delayed_work_on(cpu, &krcp->monitor_work, KFREE_DRAIN_JIFFIES);
> + schedule_delayed_monitor_work(krcp);
> raw_spin_unlock_irqrestore(&krcp->lock, flags);
> }
> }
> --
> 2.30.2
>
