[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sat, 16 Apr 2016 09:49:10 +0200
From: Thomas Graf <tgraf@...g.ch>
To: Roopa Prabhu <roopa@...ulusnetworks.com>
Cc: netdev@...r.kernel.org, jhs@...atatu.com, davem@...emloft.net
Subject: Re: [PATCH net-next v3 1/2] rtnetlink: add new RTM_GETSTATS message
to query stats
On 04/15/16 at 08:28pm, Roopa Prabhu wrote:
> +static u16 rtnl_stats_calcit(struct sk_buff *skb, struct nlmsghdr *nlh)
> +{
> + struct net *net = sock_net(skb->sk);
> + struct net_device *dev;
> + u16 min_ifinfo_dump_size = 0;
> + struct if_stats_msg *ifsm;
> + u32 filter_mask;
> +
> + ifsm = nlmsg_data(nlh);
> + filter_mask = ifsm->filter_mask;
> +
> + /* traverse the list of net devices and compute the minimum
> + * buffer size based upon the filter mask.
> + */
> + list_for_each_entry(dev, &net->dev_base_head, dev_list) {
> + min_ifinfo_dump_size = max_t(u16, min_ifinfo_dump_size,
> + if_nlmsg_stats_size(dev,
> + filter_mask));
> + }
Iterating over all net_devices in the namespace is quite an expensive
operation and it would now be done twice.
I understand that this code is taken over from rtnl_calcit() but there
the cost is at least only paid if ext_filter_mask is actually set and
the user opts into additional statistics.
I wonder if we can reduce the cost for the stats interface as its
purpose is to be minimal cost. I suggest we only add the loop once we
have an extension which actually depends on it. We can then try and
figure out to not require it.
Powered by blists - more mailing lists