Re: [PATCH v2 1/2] tcp: add generic netlink support for tcp_metrics

[Eric Dumazet <eric.dumazet@gmail.com>]

On Sun, 2012-09-02 at 08:36 +0300, Julian Anastasov wrote:
> +
> +static int tcp_metrics_flush_all(struct net *net)
> +{
> +	unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
> +	struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
> +	struct tcp_metrics_block *tm;
> +	unsigned int sync_count = 0;
> +	unsigned int row;
> +
> +	for (row = 0; row < max_rows; row++, hb++) {
> +		spin_lock_bh(&tcp_metrics_lock);
> +		tm = deref_locked_genl(hb->chain);
> +		if (tm)
> +			hb->chain = NULL;
> +		spin_unlock_bh(&tcp_metrics_lock);
> +		while (tm) {
> +			struct tcp_metrics_block *next;
> +
> +			next = deref_genl(tm->tcpm_next);
> +			kfree_rcu(tm, rcu_head);
> +			if (!((++sync_count) & 2047))
> +				synchronize_rcu();
> +			tm = next;
> +		}
> +	}
> +	return 0;
> +}

It looks like the synchronize_rcu() call is not exactly what you wanted,
but then net/ipv4/fib_trie.c has the same mistake.

What we want here is to force pending call_rcu() calls to complete, so
that we dont consume too much memory. So it would probably better to
call rcu_barrier() instead.

If other cpus are idle or outside of rcu read lock sections,
synchronize_rcu() should basically do nothing at all.

But I am not sure its worth the trouble ?

Commit c3059477fce2d956a0bb3e04357324780c5d8eeb (ipv4: Use
synchronize_rcu() during trie_rebalance()) was needed because FIB TRIE
can really use huge amounts of memory, thats hardly the case with
tcp_metrics.