Re: [RFC][PATCH 1/4] skbtrace: core feature

[Li Yu <raise.sail@gmail.com>]

于 2012年07月11日 12:03, Eric Dumazet 写道:
> On Wed, 2012-07-11 at 10:17 +0800, Li Yu wrote:
>> From: Li Yu <bingtian.ly@taobao.com>
>>
>> This implements core feature of skbtrace, which contains glue code of
>> tracepoints subsystem and relay file system, and provide skbtrace API
>> for particular networking traces.
>>
>> Thanks
>
> Hi Li
>
> This seems a huge amount of code, on an already complex stack.
>
> I am not convinced its needed. It looks like a debugging aid you had to
> write in order to understand better linux network stack.
>
> Lets see if you manage to maintain this for a while before considering
> upstreaming it.
>

Indeed, They are not toy patches and need some time to verify their
practicability. I approximately started this project on February of
this year since I am asked to repeatedly solve some similar performance
problems or explain surprised exceptional behaviors of networking stack.
Some hard investigation works also are duplicated again and again. I
hope that skbtrace such like is able to improve this problem-solve
process.

> You said that some 'buggy' drivers set rxhash to zero, but its a valid
> operation.
>
> You said 'it seems that RPS hashing can not work well for some corner
> cases', but its a known fact.
>

Em, we really are able to verify RPS imbalance by checking the last
column of /proc/net/softnet_stat, but skbtrace can give us more details
of RSS/RPS hashing. For improper RPS hashing case, it can provide more
details of what really happen in real time.

Thanks.

> Thanks
>
>
>