Re: [PATCH net] net/mlx5e: limit log messages due to (ovs) probing to _once

[Roi Dayan <roid@mellanox.com>]

On 2020-04-09 1:42 AM, marcelo.leitner@gmail.com wrote:
> On Wed, Apr 08, 2020 at 10:38:52PM +0000, Saeed Mahameed wrote:
>> On Wed, 2020-04-08 at 18:54 -0300, marcelo.leitner@gmail.com wrote:
>>> On Wed, Apr 08, 2020 at 07:51:22PM +0000, Saeed Mahameed wrote:
>>> ...
>>>>>> I understand it is for debug only but i strongly suggest to not
>>>>>> totally
>>>>>> suppress these messages and maybe just move them to tracepoints
>>>>>> buffer
>>>>>> ? for those who would want to really debug .. 
>>>>>>
>>>>>> we already have some tracepoints implemented for en_tc.c 
>>>>>> mlx5/core/diag/en_tc_tracepoints.c, maybe we should define a
>>>>>> tracepoint
>>>>>> for error reporting .. 
>>>>>
>>>>> That, or s/netdev_warn/netdev_dbg/, but both are more hidden to
>>>>> the
>>>>> user than the _once.
>>>>>
>>>>
>>>> i don't see any reason to pollute kernel log with debug messages
>>>> when
>>>> we have tracepoint buffer for en_tc .. 
>>>
>>> So we're agreeing that these need to be changed. Good.
>>
>> I would like to wait for the feedback from the CC'ed mlnx TC
>> developers..
>>
>> I just pinged them, lets see what they think.
>>
>> but i totally agree, TC can support 100k offloads requests per seconds,
>> dumping every possible issue to the kernel log shouldn't be an option,this is not a boot or a fatal error/warning .. 
>>
>>>
>>> I don't think a sysadmin would be using tracepoints for
>>> troubleshooting this, but okay. My only objective here is exactly
>>> what
>>> you said, to not pollute kernel log too much with these potentially
>>> repetitive messages.
>>
>> these types of errors are easily reproduce-able, a sysadmin can see and
>> report the errno and the extack message, and in case it is really
>> required, the support or development team can ask to turn on trace-
>> points or debug and reproduce .. 
> 
> Roger that, thanks Saeed.
> 

Hi Marcelo,

I was somewhat in favor for *_once when first read it, without starting to
enable probe stuff but I guess *_once will become redundant pretty quick.
another option is debugging with tc verbose flag but ovs doesn't support
logging extack errors today.
checking ovs issues on a large system with many bridges/ports/rules without
the tc/driver errors in the log will be very difficult.
in some places we already only use extack without netdev_warn.
so currently in favor in removing the other logs if extack error exists to
avoid flooding the log each time ovs age out an unsupported rule and re-adds it.
i'm also in favor for the trace points to ease debug.

Thanks,
Roi