Re: [PATCH net-next V3] net: dynamic ingress_queue allocation

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

Le dimanche 03 octobre 2010 à 11:42 +0200, Jarek Poplawski a écrit :

> I'd consider rcu_dereference_rtnl(). Btw, technically qdisc_lookup()
> doesn't require rtnl, and there was time it was used without it
> (on xmit path).


Hmm, for me, rcu_dereference_rtnl() is a bit lazy.

Either we are a reader and should use rcu_dereference(), or a writer and
RTNL should be held.

Mixing two conditions in a "super macro" is a workaround that we used to
promptly shutup some lockdep splats. Real fix would be to use strict
lockdep conditions, because this better documents the code and the
locking invariants.

BTW, rtnl_dereference() should be changed to use
rcu_dereference_protected() instead of rcu_dereference_check() :
If RTBL is held, there is no need to force a barrier.


> I think you should also add a comment here why this rcu is used, and
> that it changes only once in dev's liftime.
> 

This comment was needed in the previous version of the patch, with the
smb_wmb() barrier. Now I switched to regular rcu use, nothing prevents
us to change dev->ingress_queue in flight. Of course there is no current
interest doing so.