On 4/18/16, 8:41 PM, David Miller wrote:
> From: Roopa Prabhu <ro...@cumulusnetworks.com>
> Date: Mon, 18 Apr 2016 14:10:19 -0700
>
>> This patch adds a new RTM_GETSTATS message to query link stats via
>> netlink from the kernel. RTM_NEWLINK also dumps stats today, but
>> RTM_NEWLINK returns a lot more than just stats and is expensive in
>> some cases when frequent polling for stats from userspace is a
>> common operation.
> I'm holding off on this until we sort out the 64-bit netlink
> attribute alignment issue.
sure,
>
> Meanwhile, I'll some kind of a fix into the tree for the
> rtnl_fill_stats() change so that it doesn't cause unaligned
> accesses.
>
> I just tested out a clever idea, where for architectures where
> unaligned accesses is a problem, we insert a zero length NOP attribute
> before the 64-bit stats. This makes it properly aligned. A quick
> hack patch just passed testing on my sparc64 box, but I'll go over it
> some more.
>
> diff --git a/include/uapi/linux/if_link.h b/include/uapi/linux/if_link.h
> index bb3a90b..5ffdcb3 100644
> --- a/include/uapi/linux/if_link.h
> +++ b/include/uapi/linux/if_link.h
> @@ -155,6 +155,7 @@ enum {
> IFLA_PROTO_DOWN,
> IFLA_GSO_MAX_SEGS,
> IFLA_GSO_MAX_SIZE,
> + IFLA_PAD,
> __IFLA_MAX
> };
>
> diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
> index a7a3d34..b192576 100644
> --- a/net/core/rtnetlink.c
> +++ b/net/core/rtnetlink.c
> @@ -1052,6 +1052,15 @@ static noinline_for_stack int rtnl_fill_stats(struct
> sk_buff *skb,
> struct rtnl_link_stats64 *sp;
> struct nlattr *attr;
>
> + /* Add a zero length NOP attribute so that the nla_data()
> + * of the IFLA_STATS64 will be 64-bit aligned.
> + */
> +#ifndef HAVE_EFFICIENT_UNALIGNED_ACCESS
> + attr = nla_reserve(skb, IFLA_PAD, 0);
> + if (!attr)
> + return -EMSGSIZE;
> +#endif
> +
> attr = nla_reserve(skb, IFLA_STATS64,
> sizeof(struct rtnl_link_stats64));
> if (!attr)
>
that is cleaver :)
thanks!
