Re: [PATCH v2 net-next 0/7] Cleanup to main DSA structures

[Vladimir Oltean <vladimir.oltean@nxp.com>]

On Wed, Jan 05, 2022 at 11:04:24AM -0800, Florian Fainelli wrote:
> On 1/5/22 10:59 AM, Vladimir Oltean wrote:
> > On Wed, Jan 05, 2022 at 10:39:04AM -0800, Florian Fainelli wrote:
> >> On 1/5/22 5:21 AM, Vladimir Oltean wrote:
> >>> This series contains changes that do the following:
> >>>
> >>> - struct dsa_port reduced from 576 to 544 bytes, and first cache line a
> >>>   bit better organized
> >>> - struct dsa_switch from 160 to 136 bytes, and first cache line a bit
> >>>   better organized
> >>> - struct dsa_switch_tree from 112 to 104 bytes, and first cache line a
> >>>   bit better organized
> >>>
> >>> No changes compared to v1, just split into a separate patch set.
> >>
> >> This is all looking good to me. I suppose we could possibly swap the
> >> 'nh' and 'tag_ops' member since dst->tag_ops is used in
> >> dsa_tag_generic_flow_dissect() which is a fast path, what do you think?
> > 
> > pahole is telling me that dst->tag_ops is in the first cache line on
> > both arm64 and arm32. Are you saying that it's better for it to take
> > dst->nh's place?
> 
> Sorry it stuck in my head somehow based upon patch 7's pahole output
> that tag_ops was still in the second cache line when the after clearly
> shows that it moved to the first cache line. No need to add additional
> changes then, thanks!

Yes, that output is a bit verbose and small details are easy to miss
(adding --expand_types makes it even worse). Happened to me quite a few
times during the session...

I'll prepare a v3 once we figure out what would suffice in terms of
warning other developers about the lack of bit field atomicity.