On Tue, 26 May 2020 15:43:52 -0700 Michael Chan wrote:
> On Tue, May 26, 2020 at 3:04 PM Jakub Kicinski <k...@kernel.org> wrote:
> > On Mon, 25 May 2020 17:41:17 -0400 Michael Chan wrote:
> > > We have logic to maintain network counters across resets by storing
> > > the counters in bp->net_stats_prev before reset. But not all resets
> > > will clear the counters. Certain resets that don't need to change
> > > the number of rings do not clear the counters. The current logic
> > > accumulates the counters before all resets, causing big jumps in
> > > the counters after some resets, such as ethtool -G.
> > >
> > > Fix it by only accumulating the counters during reset if the irq_re_init
> > > parameter is set. The parameter signifies that all rings and interrupts
> > > will be reset and that means that the counters will also be reset.
> > >
> > > Reported-by: Vijayendra Suman <vijayendra.su...@oracle.com>
> > > Fixes: b8875ca356f1 ("bnxt_en: Save ring statistics before reset.")
> > > Signed-off-by: Michael Chan <michael.c...@broadcom.com>
> >
> > Hi Michael!
> >
> > Could you explain why accumulating counters causes a jump?
>
> Yes, during chip reset, we free most hardware resources including
> possibly hardware counter resources. After freeing the hardware
> counters, the counters will go to zero. To preserve the counters, we
> take a snapshot of the hardware counters and add them to the
> bp->net_stats_prev. The counters in bp->net_stats_prev are always
> added to the current hardware counters to provide the true counters.
>
> The problem is that not all resets will free the hardware counters.
> The old code is unconditionally taking the snapshot during reset. So
> when we don't free the hardware counters, the snapshot will cause us
> to effectively double the hardware counters after the reset.
Aw! I see what you mean now, thanks for the explanation!
