Re: [PATCH] bnx2x: fix occasional statistics off-by-4GB error

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

On Sun, 2013-03-17 at 21:37 +0200, Eilon Greenstein wrote:
> On Sun, 2013-03-17 at 14:24 -0400, David Miller wrote:
> > From: "Dmitry Kravkov" <dmitry@broadcom.com>
> > Date: Sun, 17 Mar 2013 13:10:37 +0000
> > 
> > > Probably this commit resolved the issue:
> > > 
> > > commit bef05406ac0ea6f468e1e25e9934f3011ea9259b
> > > Author: Dmitry Kravkov <dmitry@broadcom.com>
> > > Date:   Tue Sep 11 04:34:08 2012 +0000
> > > 
> > >     bnx2x: Avoid sending multiple statistics queries
> > > 
> > > Can you try it pls?
> > 
> > These are completely seperate bugs.
> > 
> > The macro in question does not handle rollover of the lower 32-bits of
> > the statistic properly at all, and therefore Maciej's patch should be
> > applied and queud up for -stable.
> > 
> > Please give it your ACK unless you can find a bug in his change.
> 
> Both the high value and the low value are read from the chip - so this
> patch will increment the higher 32 bits twice (well, more than once).
> Not taking the HW/FW high counter at all is also not acceptable since
> the reading frequency is not high enough without those.
> 
> So the original patch is nacked, but we are trying to figure out what is
> causing the statistics to misbehave and it might be related to sending
> statistics query twice.

This looks like the typical problem of updating a 64bit value in non
atomic way.

Its guaranteed to happen on 32bit hosts.
We had to introduce include/linux/u64_stats_sync.h to help to solve this
without adding extra cost on 64bit arches.

In bnx2x case, we perform 32bit operations even on 64bit host, so we
probably need to add a seqcount_t, so that a stat consumer can detect it
read a stale value.