On Mon, Mar 06, 2006 at 02:01:11PM -0800, Benjamin LaHaise ([EMAIL PROTECTED])
wrote:
> Hello again,
>
> This patch introduces the use of rcu for the ipv4 established connections
> hashtable, as well as the timewait table since they are closely intertwined.
> This removes 4 atomic operations per packet from the tcp_v4_rcv codepath,
> which helps quite a bit when the other performance barriers in the system
> are removed. Eliminating the rwlock cache bouncing should also help on SMP
> systems.
>
> By itself, this improves local netperf performance on a P4/HT by ~260Mbit/s
> on average. With smaller packets (say, ethernet size) the difference should
> be larger.
>
> Note that this patch changes the semantics of __inet_lookup() and
> __inet_lookup_established() to *not* perform a sock_hold(). This way
> we can avoid the atomic inc and dec of the reference count and rely on
> rcu pinning the socket over the scope of the rcu_read_lock_bh() region.
RCU introduces noticeble latencies for writing part compared to rwlock,
so workloads when a lot of new short connections are created and destroyed,
like busy web server within LAN with small latencies, will suffer from
this, doesn't it?
> Only minimal fixes for ipv6 and dccp to use the renamed wlock, as I am not
> setup to test them. I have stared at this patch for a while, and I think
> it's correct, but more eyes are definately warranted. Most of the issues
> regarding the state and lifespan of a struct sock are already handled by
> the network layer as the socket can be locked after being retrieved from
> the connection hashtable, we just have to be careful about reinitializing
> fields that the rcu hash list depends on.
>
> -ben
--
Evgeniy Polyakov
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
