Re: [PATCH 1/3] inetpeer: Support ipv6 addresses.

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

Le lundi 29 novembre 2010 à 21:53 -0800, David Miller a écrit :
> From: Eric Dumazet <eric.dumazet@gmail.com>
> Date: Tue, 30 Nov 2010 06:42:16 +0100
> 
> > Its a bit early in the morning here, I must confess I dont yet
> > understand your patch David :)
> > 
> > As we use a tree, why not using two different trees for ipv4 / ipv6 ?
> 
> The "key" just creates a natural ordering in the tree, it's
> almost arbitrary except that it must distribute well amongst
> the entries.

Hmm. AVL search must take a decision, take the left or the right path.

if current key is equal, which path do you take ?


@@ -165,9 +208,9 @@ static void unlink_from_unused(struct inet_peer *p)
>  	for (u = rcu_dereference_protected(peers.root,		\
>  			lockdep_is_held(&peers.lock));		\
>  	     u != peer_avl_empty; ) {				\
> -		if (_daddr == u->v4daddr)			\
> +		if (inet_peer_addr_equal(_daddr, &u->daddr))	\
>  			break;					\
> -		if ((__force __u32)_daddr < (__force __u32)u->v4daddr)	\
> +		if (key < inet_peer_key(&u->daddr))		\
>  			v = &u->avl_left;			\
>  		else						\
>  			v = &u->avl_right;			\


Apparently you take the right one, you may miss the target if its on the
left path ?

> 
> I currently don't see any reason to make two trees right now.
> 

Cost of a tree is one pointer, and ipv4 search would be faster if we use
different search functions.

> > I dont understand how computing a 32bit key (sort of hash key) is going
> > to help when hash collision happens, with an avl tree.
> > Either version of tree (AVL/rbtree) will be expensive to use if depth is
> > big (With 2 millions entries, depth is going to be very big). I
> > understand you want to get rid of route cache ?
> 
> Do we plan to talk to 2 million unique destinations and have active
> non-default metrics for each one of them very often?
> 
> inet_peer entries will only get created when we need to make
> non-default metric settings for a specific destination address.
> 
> See that's the thing, it's scope is so much smaller than the existing
> routing cache.  It's only going to be used in limited if not
> controlled cases.
> 

OK good :)

I have no idea how many addresses have non default metric settings.

Do you know how to make an estimation on a server ?