Date:   Fri, 21 Apr 2017 11:35:27 +1000
From:   Gavin Shan <gwshan@...ux.vnet.ibm.com>
To:     Florian Fainelli <f.fainelli@...il.com>
Cc:     Gavin Shan <gwshan@...ux.vnet.ibm.com>,
        David Miller <davem@...emloft.net>, netdev@...r.kernel.org,
        joe@...ches.com, kubakici@...pl
Subject: Re: [PATCH v3 net-next 5/8] net/ncsi: Dump NCSI packet statistics

On Thu, Apr 20, 2017 at 05:59:27PM -0700, Florian Fainelli wrote:
>On 04/20/2017 05:44 PM, Gavin Shan wrote:
>> On Thu, Apr 20, 2017 at 05:26:14PM -0700, Florian Fainelli wrote:
>>> On 04/20/2017 04:38 PM, Gavin Shan wrote:
>>>> On Thu, Apr 20, 2017 at 01:21:03PM -0400, David Miller wrote:
>>>>> From: Gavin Shan <gwshan@...ux.vnet.ibm.com>
>>>>> Date: Tue, 18 Apr 2017 16:51:32 +1000
>>>>>
>>>>>> This creates /sys/kernel/debug/ncsi/<eth0>/stats to dump the NCSI
>>>>>> packets sent and received over all packages and channels. It's useful
>>>>>> to diagnose NCSI problems, especially when NCSI packages and channels
>>>>>> aren't probed properly. The statistics can be gained from debugfs file
>>>>>> as below:
>>>>>>
>>>>>>  # cat /sys/kernel/debug/ncsi/eth0/stats
>>>>>
>>>>> There is no reason you cannot use ethtool statistics to provide this
>>>>> information to the user.
>>>>>
>>>>
>>>> It can be dumped by ethtool, but it's more reasonable to dump them
>>>> through debugfs for couple of reasons: (1) ethtool usually dumps
>>>> statistics collected by hardware, but this debugfs file dumps the
>>>> statistics of packets seen (collected) by software. They are different
>>>> things. Note that NCSI channel collects statistics in hardware and it's
>>>> not exposed or dumped by this patchset. They are candidates for ethtool.
>>>> (2) To expose this through ethtool relies on the availability of the tool.
>>>> It's nicer not to depend on it. (3) This interface can be used to check
>>>> the debug packet has been sent successfully through /sys/kernel/debug/ncsi/eth0/pkt,
>>>> dumping the statistics through debugfs make this (debugging) mechanism
>>>> consistent.
>>>
>>> Can't you create a ncsi folder under /sys/class/net/eth0/nsci/ and then
>>> put your stats in there? That would at least look slightly consistent
>>> with what is already existing for the non-NC-SI networking stack.
>> 
>> Do you think it's good place to have /sys/class/net/eth0/ncsi/pkt?
>> Note this file accepts commands to send NCSI comands and dumps the
>> corresponding response on read. Also, it's a debugging interface
>> and disabled (invisible) for the most time. I think all directories
>> and files in /sys/class/net/eth0/ should be visible and the structure
>> is stable all the time.
>
>Statistics should not be debugging features having them all the time is
>invaluable, so in that regard they could probably be part of a sysfs
>interface of some kind. If this "pkt" file is special, then yes, maybe
>debugfs is appropriate here.
>
>It sounds like you may also want to consider doing a NC-SI generic
>netlink family at some point, because chances are that are you going to
>export more and more information over time, just like there could be
>additional commands/actions to be performed as well as asynchronous
>events. Netlink is great for that, downside is that you have to write a
>custom user-space tool (or extend iproute2).

Yes, it's definitely good idea to cover everything using one interface.
It's for sure that the statistics collected by hardware should be dumped
by ethtool in future. I will move functionalities introduced by this
patchset to ethtool as well, as Dave suggested in another thread. Hope
Joe has no objection on this.

Cheers,
Gavin 

Powered by blists - more mailing lists