On Mon, Nov 28, 2016 at 10:40 PM, David Miller [off-list ref] wrote:

From: Eric Dumazet <redacted>
Date: Sat, 26 Nov 2016 18:16:00 -0800

quoted

On Sun, 2016-11-27 at 00:47 +0200, Saeed Mahameed wrote:

quoted

On Fri, Nov 25, 2016 at 5:46 PM, Eric Dumazet [off-list ref] wrote:

quoted

As you see here in SRIOV mode (PF only) reads   sw stats from FW.
Tariq, I think we need to fix this.

Sure, my patch does not change this at all.

If mlx4_is_master() is false, then we aggregate the software states and
only the software stats.

My patch makes this aggregation possible at the time ethtool or
ndo_get_stat64() are called, since this absolutely not depend on the 250
ms timer fetching hardware stats.

Saeed please provide counter arguments or ACK this patch, thank you.

I have nothing against this patch, I just wanted to point out that
this patch is just fixing the symptom.
We keep ignoring the root cause that dev_get_stats is called under a
spin_lock which really ties our hands "us device drivers developers"
and push us towards those fragile solutions like deferred work for
caching statistics.
I am not saying that Eric should fix this in his patch, i just wanted
to raise the awareness of the root cause.

Regarding the SRIOV PF stats, i will take it with Tariq internally.

Bottom line, I ACK this patch, I might even do the same myself for
mlx5 :), but there are some follow ups that need to be done.