Linux Ethernet Bridging - Dec 2019 - [Bridge] [PATCH net-next] net: bridge: add STP xstats

On 10/12/2019 22:10, Vivien Didelot wrote:
> Hi Nikolay,
> 
> On Tue, 10 Dec 2019 21:50:10 +0200, Nikolay Aleksandrov <nikolay at
cumulusnetworks.com> wrote:
>>>> Why do you need percpu ? All of these seem to be incremented
with the
>>>> bridge lock held. A few more comments below.
>>>
>>> All other xstats are incremented percpu, I simply followed the
pattern.
>>>
>>
>> We have already a lock, we can use it and avoid the whole per-cpu
memory handling.
>> It seems to be acquired in all cases where these counters need to be
changed.
> 
> Since the other xstats counters are currently implemented this way, I
prefer
> to keep the code as is, until we eventually change them all if percpu is in
> fact not needed anymore.
> 
> The new series is ready and I can submit it now if there's no
objection.
> 
> 
> Thanks,
> 
> 	Vivien
> 
There is a reason other counters use per-cpu - they're incremented without
any locking from fast-path.
The bridge STP code already has a lock which is acquired in all of these paths
and we don't need
this overhead and the per-cpu memory allocations. Unless you can find a STP
codepath which actually
needs per-cpu, I'd prefer you drop it.

Thank you,
 Nik

On Tue, 10 Dec 2019 22:15:26 +0200, Nikolay Aleksandrov <nikolay at
cumulusnetworks.com> wrote:
> >>>> Why do you need percpu ? All of these seem to be
incremented with the
> >>>> bridge lock held. A few more comments below.
> >>>
> >>> All other xstats are incremented percpu, I simply followed the
pattern.
> >>>
> >>
> >> We have already a lock, we can use it and avoid the whole per-cpu
memory handling.
> >> It seems to be acquired in all cases where these counters need to
be changed.
> > 
> > Since the other xstats counters are currently implemented this way, I
prefer
> > to keep the code as is, until we eventually change them all if percpu
is in
> > fact not needed anymore.
> > 
> > The new series is ready and I can submit it now if there's no
objection.
> 
> There is a reason other counters use per-cpu - they're incremented
without any locking from fast-path.
> The bridge STP code already has a lock which is acquired in all of these
paths and we don't need
> this overhead and the per-cpu memory allocations. Unless you can find a STP
codepath which actually
> needs per-cpu, I'd prefer you drop it.
Ho ok I understand what you mean now. I'll drop the percpu attribute.


Thanks,

	Vivien