Re: [PATCH net-next-2.6] tg3: 64bits stats

[Ben Hutchings <bhutchings@solarflare.com>]

On Mon, 2010-07-05 at 20:07 +0200, Eric Dumazet wrote:
> Le lundi 05 juillet 2010 à 18:31 +0100, Ben Hutchings a écrit :
> > On Mon, 2010-07-05 at 18:03 +0200, Eric Dumazet wrote:
> > > Le lundi 05 juillet 2010 à 11:14 +0200, Eric Dumazet a écrit :
> > > > After commit be1f3c2c027c (net: Enable 64-bit net device statistics on
> > > > 32-bit architectures), we can now provide 64bit stats, even on 32bit
> > > > arches.
> > > > 
> > > > Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
> > > > ---
> > > 
> > > Please disregard this patch.
> > > 
> > > There is small possibility a reader might read a 64bit value while
> > > another writer makes a change to it, changing high 32bit value.
> > > 
> > > A change in core network would be needed to make this 100% safe,
> > > possibly using a seqlock to protect dev->stats64
> > 
> > I really didn't want to add this overhead and complication to readers
> > when only some drivers need it.
> > 
> 
> Overhead would be minimal, only in get_stats() method and only for
> drivers that want to provide 64 bit stats...
> 
> Clearly, rx/tx path must not have any overhead.
> 
> Right now, even RTNL 64bit stats are not safe.
> 
> Should we revert prior patches or try to make progress ?

I think you should use a similar approach here as you did in the
loopback driver, i.e. update private variables in the RX and TX path and
then copy/aggregate them in the implementation ndo_get_stats64 (only
without the need for percpu stats).

If you want to include a seqlock in the driver stats interface, you can
do that but it's not going to be pretty and we're still going to need
additional seqlocks for per-queue (or percpy) stats in some drivers.

Ben.

-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.