On Tue, 2004-09-28 at 09:33, Thomas Graf wrote:
> > Speaking of generic stats; i have a patch netlink ready which may need
> > some extensions. I did post it a while back on netdev but didnt get
> > feedback.
>
> The code looks good and I couldn't spot any errors but I'm not
> sure if the locking in gen_copy_[x]stats is a good thing.
> Shouldn't that be done earlier by the caller?
In the netsched code that became a portability issue; Dave fixed it
there, so i just replicated here. If you feel like doing something
clever you are welcome to submit a patch.
> This prevents
> corruption but it allows duplicated TLVs in an skb. I suggest
> to make the caller have a lock on his data and only allow one
> dumper at the same time until the dump is complete, or at least
> provide a lockless variant for callers doing the locking on
> their own.
>
Reminds me:
gnet_stats needs to have TLVs embedded in it.
bytes,drops, packets are generic enough; others are not.
So if we add a length field then we can add TLVs for things like
QSTATS = { qlen, backlog} etc.
This means we could then allow for adding a lot of different
stats. A big lesson from current tc_stats.
cheers,
jamal
|