Re: [PATCH] net: reduce net_rx_action() latency to 2 HZ

[Paul Gortmaker <paul.gortmaker@windriver.com>]

On 13-03-21 11:27 AM, Eric Dumazet wrote:
> On Thu, 2013-03-21 at 11:03 -0400, Paul Gortmaker wrote:
>> [CC'ing stable & Willy - for the older releases not fed by
>> http://patchwork.ozlabs.org/bundle/davem/stable/ ]
>>
>> On Tue, Mar 5, 2013 at 12:15 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>>> From: Eric Dumazet <edumazt@google.com>
>>>
>>> We should use time_after_eq() to get maximum latency of two ticks,
>>> instead of three.
>>>
>>> Bug added in commit 24f8b2385 (net: increase receive packet quantum)
>>
>> I'm not sure what applications would notice the extra tick, but 24f8b takes
>> us back to 2.6.29.  It cherry picks cleanly onto 2.6.34, so it probably also
>> does the same for Willy's 2.6.32 longterm too.
>>
>> Commit is now mainline d114a3338747255518 - v3.9-rc3~36^2~34.
> 
> BQL (Bytes Queue Limit) relies on TX completion being run often, and
> Qdisc being serviced often as well. If net_rx_action() hogs the cpu,
> net_tx_action() is delayed and NIC can stall.
> 
> I wrote this patch because I was investigating a regression when a
> Google application began using BQL enabled kernels.
> 
> About the latency in itself, following commit is way more interesting.
> 
> commit c10d73671ad30f5 (softirq: reduce latencies)
> 
> As without it, I could trigger more than 50ms latencies for the poor
> user thread interrupted by softirq processing.

That is also reasonably portable back to 2.6.34.  And it is more
interesting too -- it will be interesting in a preempt_rt context
too, once RT moves ahead off the current 3.6 baseline, which still
has the old count-limit of 10 vs the new 2ms time limit.

RT (3.4 and 3.6 based) currently has this patch from Steven:
http://git.kernel.org/cgit/linux/kernel/git/paulg/3.6-rt-patches.git/tree/net-tx-action-avoid-livelock-on-rt.patch

Anyway, thanks for the heads up on this commit.
Paul.