[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 20 Dec 2019 16:54:46 -0800 (PST)
From: David Miller <davem@...emloft.net>
To: dust.li@...ux.alibaba.com
Cc: xiyou.wangcong@...il.com, jhs@...atatu.com,
john.fastabend@...il.com, jiri@...nulli.us,
tonylu@...ux.alibaba.com, netdev@...r.kernel.org
Subject: Re: [PATCH net-next 0/2] net: sched: unify __gnet_stats_copy_xxx()
for percpu and non-percpu
From: Dust Li <dust.li@...ux.alibaba.com>
Date: Tue, 17 Dec 2019 16:47:16 +0800
> Currently, __gnet_stats_copy_xxx() will overwrite the return value when
> percpu stats are not enabled. But when percpu stats are enabled, it will
> add the percpu stats to the result. This inconsistency brings confusion to
> its callers.
>
> This patch series unify the behaviour of __gnet_stats_copy_basic() and
> __gnet_stats_copy_queue() for percpu and non-percpu stats and fix an
> incorrect statistic for mqprio class.
>
> - Patch 1 unified __gnet_stats_copy_xxx() for both percpu and non-percpu
> - Patch 2 depending on Patch 1, fixes the problem that 'tc class show'
> for mqprio class is always 0.
I think this is going to break the estimator.
The callers of est_fetch_counters() expect continually incrementing
statistics. It does relative subtractions from previous values each
time, so if we just reset on the next statistics fetch it won't work.
Sorry I can't apply this.
Powered by blists - more mailing lists