Re: [PATCH v3 net-next 3/3] net/sched: sch_frag: add generic packet fragment support.

[Jakub Kicinski <kuba@kernel.org>]

On Fri, 20 Nov 2020 07:38:36 +0800 wenxu@ucloud.cn wrote:
> +int tcf_dev_queue_xmit(struct sk_buff *skb, int (*xmit)(struct sk_buff *skb))
> +{
> +	xmit_hook_func *xmit_hook;
> +
> +	xmit_hook = rcu_dereference(tcf_xmit_hook);
> +	if (xmit_hook)
> +		return xmit_hook(skb, xmit);
> +	else
> +		return xmit(skb);
> +}
> +EXPORT_SYMBOL_GPL(tcf_dev_queue_xmit);

I'm concerned about the performance impact of these indirect calls.

Did you check what code compiler will generate? What the impact with
retpolines enabled is going to be?

Now that sch_frag is no longer a module this could be simplified.

First of all - xmit_hook can only be sch_frag_xmit_hook, so please use
that directly. 

	if (READ_ONCE(tcf_xmit_hook_count)) 
		sch_frag_xmit_hook(...
	else
		dev_queue_xmit(...

The abstraction is costly and not necessary right now IMO.

Then probably the counter should be:

	u32 __read_mostly tcf_xmit_hook_count;

To avoid byte loads and having it be places in an unlucky cache line.

You could also make the helper a static inline in a header.


Unless I'm not giving the compiler enough credit and the performance
impact of this patch with retpolines on is indiscernible, but that'd
need to be proven by testing...