Re: [PATCH v2] rcu/tree: Add multiple in-flight batches of kfree_rcu work

["Paul E. McKenney" <paulmck@kernel.org>]

On Wed, Aug 28, 2019 at 10:09:52AM -0400, Joel Fernandes (Google) wrote:
> During testing, it was observed that amount of memory consumed due
> kfree_rcu() batching is 300-400MB. Previously we had only a single
> head_free pointer pointing to the list of rcu_head(s) that are to be
> freed after a grace period. Until this list is drained, we cannot queue
> any more objects on it since such objects may not be ready to be
> reclaimed when the worker thread eventually gets to drainin g the
> head_free list.
> 
> We can do better by maintaining multiple lists as done by this patch.
> Testing shows that memory consumption came down by around 100-150MB with
> just adding another list. Adding more than 1 additional list did not
> show any improvement.

Nice!  A few comments below.  Please address them and post a full v3.
(I am off the next two days, and I guarantee you that upon return I will
mix and match the wrong patches otherwise!)

> Suggested-by: Paul E. McKenney <paulmck@linux.ibm.com>
> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
> ---
>  kernel/rcu/tree.c | 61 ++++++++++++++++++++++++++++++++---------------
>  1 file changed, 42 insertions(+), 19 deletions(-)
> 
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 4f7c3096d786..5bf8f7e793ea 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -2688,28 +2688,37 @@ EXPORT_SYMBOL_GPL(call_rcu);
>  
>  /* Maximum number of jiffies to wait before draining a batch. */
>  #define KFREE_DRAIN_JIFFIES (HZ / 50)
> +#define KFREE_N_BATCHES 2
> +
> +struct kfree_rcu_work {
> +	/* The rcu_work node for queuing work with queue_rcu_work(). The work
> +	 * is done after a grace period.
> +	 */
> +	struct rcu_work rcu_work;
> +
> +	/* The list of objects that have now left ->head and are queued for
> +	 * freeing after a grace period.
> +	 */
> +	struct rcu_head *head_free;
> +
> +	struct kfree_rcu_cpu *krcp;
> +};
>  
>  /*
>   * Maximum number of kfree(s) to batch, if this limit is hit then the batch of
>   * kfree(s) is queued for freeing after a grace period, right away.
>   */
>  struct kfree_rcu_cpu {
> -	/* The rcu_work node for queuing work with queue_rcu_work(). The work
> -	 * is done after a grace period.
> -	 */
> -	struct rcu_work rcu_work;
>  
>  	/* The list of objects being queued in a batch but are not yet
>  	 * scheduled to be freed.
>  	 */
>  	struct rcu_head *head;
>  
> -	/* The list of objects that have now left ->head and are queued for
> -	 * freeing after a grace period.
> -	 */
> -	struct rcu_head *head_free;
> +	/* Pointer to the per-cpu array of kfree_rcu_work structures */
> +	struct kfree_rcu_work krw_arr[KFREE_N_BATCHES];
>  
> -	/* Protect concurrent access to this structure. */
> +	/* Protect concurrent access to this structure and kfree_rcu_work. */
>  	spinlock_t lock;
>  
>  	/* The delayed work that flushes ->head to ->head_free incase ->head
> @@ -2730,12 +2739,14 @@ static void kfree_rcu_work(struct work_struct *work)
>  {
>  	unsigned long flags;
>  	struct rcu_head *head, *next;
> -	struct kfree_rcu_cpu *krcp = container_of(to_rcu_work(work),
> -					struct kfree_rcu_cpu, rcu_work);
> +	struct kfree_rcu_work *krwp = container_of(to_rcu_work(work),
> +					struct kfree_rcu_work, rcu_work);
> +	struct kfree_rcu_cpu *krcp;
> +
> +	krcp = krwp->krcp;
>  
>  	spin_lock_irqsave(&krcp->lock, flags);
> -	head = krcp->head_free;
> -	krcp->head_free = NULL;
> +	head = xchg(&krwp->head_free, NULL);

Given that we hold the lock, why the xchg()?  Alternatively, why not
just acquire the lock in the other places you use xchg()?  This is a
per-CPU lock, so contention should not be a problem, should it?

>  	spin_unlock_irqrestore(&krcp->lock, flags);
>  
>  	/*
> @@ -2758,19 +2769,28 @@ static void kfree_rcu_work(struct work_struct *work)
>   */
>  static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp)
>  {
> +	int i = 0;
> +	struct kfree_rcu_work *krwp = NULL;
> +
>  	lockdep_assert_held(&krcp->lock);
> +	while (i < KFREE_N_BATCHES) {
> +		if (!krcp->krw_arr[i].head_free) {
> +			krwp = &(krcp->krw_arr[i]);
> +			break;
> +		}
> +		i++;
> +	}
>  
> -	/* If a previous RCU batch work is already in progress, we cannot queue
> +	/* If both RCU batches are already in progress, we cannot queue
>  	 * another one, just refuse the optimization and it will be retried
>  	 * again in KFREE_DRAIN_JIFFIES time.
>  	 */

If you are going to remove the traditional first "/*" line of a comment,
why not go all the way and cut the last one as well?  "//".

> -	if (krcp->head_free)
> +	if (!krwp)
>  		return false;
>  
> -	krcp->head_free = krcp->head;
> -	krcp->head = NULL;
> -	INIT_RCU_WORK(&krcp->rcu_work, kfree_rcu_work);
> -	queue_rcu_work(system_wq, &krcp->rcu_work);
> +	krwp->head_free = xchg(&krcp->head, NULL);

This isn't anywhere near a fastpath, so just acquiring the lock is a
better choice here.

> +	INIT_RCU_WORK(&krwp->rcu_work, kfree_rcu_work);
> +	queue_rcu_work(system_wq, &krwp->rcu_work);
>  
>  	return true;
>  }
> @@ -3736,8 +3756,11 @@ static void __init kfree_rcu_batch_init(void)
>  
>  	for_each_possible_cpu(cpu) {
>  		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
> +		int i = KFREE_N_BATCHES;
>  
>  		spin_lock_init(&krcp->lock);
> +		while (i--)
> +			krcp->krw_arr[i].krcp = krcp;

This was indeed a nice trick back in the PDP-11 days of 64-kilobyte
address spaces, so thank you for the nostalgia!  However, a straight-up
"for" loop is less vulnerable to code being added between the declaration
of "i" and the "while" loop.

>  		INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
>  	}
>  }
> -- 
> 2.23.0.187.g17f5b7556c-goog
>