Re: Killing sk->sk_callback_lock (was Re: [PATCH] net/core/sock.c remove extra wakeup)

[Eric Dumazet <dada1@cosmosbay.com>]

Herbert Xu a écrit :
> On Mon, Jun 16, 2008 at 07:38:23PM -0600, Patrick Mullaney wrote:
>> I don't follow but I wasn't part of the IRC discussion. :-) Please
>> send me a note out of band if you would like to discuss again on IRC.
>> I agree that the sk_callback_lock may be significant overhead
>> (although lockstat is not showing it to be highly contended).
> 
> Lock contention doesn't matter! The worst problem with read-write
> locks is cache-line bouncing.  That is, read locks will grab cache
> lines from other read locks thus causing performance to go down
> the drain if read locks are common (which is the only scenario
> where you use read-write locks anyway).  Moral of the story is
> that read-write locks are bad.
> 
> Cheers,

Yes, read-write locks are bad (not the concept itself, only their
implementation) but alternatives are complex to setup and maintain.

ipt_do_table() for example hits a read-write lock for every packet
handled on a chain. This is why firewalling sucks even with few rules.