Re: [ethtool PATCH 2/2] Add RX packet classification interface

[Dimitris Michailidis <dm@chelsio.com>]

Ben Hutchings wrote:
> On Mon, 2011-03-07 at 10:43 -0800, Alexander Duyck wrote:
>> On 3/7/2011 10:28 AM, Ben Hutchings wrote:
>>> On Mon, 2011-03-07 at 10:22 -0800, Dimitris Michailidis wrote:
>>>> Ben Hutchings wrote:
>>>>> On Mon, 2011-03-07 at 09:04 -0800, Alexander Duyck wrote:
>>>>>> The only time where location really matters is if you are attempting to
>>>>>> overwrite an existing rule and I am not sure how that would be handled
>>>>>> in ntuple anyway since right now adding additional rules via ntuple for
>>>>>> ixgbe just results in duplicate rules being defined.
>>>>> As I understand it, the location also determines the *priority* for the
>>>>> rule.
>>>> This is true, at least for TCAMs.  But it's relevant only when multiple
>>>> filters would match a packet.  People often use non-overlapping filters, for
>>>> these adding the filter at any available slot is OK.
>>> Right.  But ethtool would have to determine that the filter was non-
>>> overlapping, before ignoring the location.  Also it cannot allow
>>> deletion by location if it ever ignores the location on insertion.  We
>>> should make the location optional at both the command-line and API
>>> level, but never ignore it.
>>>
>> I wasn't implying that we ignore it for rules inserted via the nfc 
>> interface.  Only for those inserted via the ntuple interface.
> 
> We should never fall back to the ntuple interface if a location is
> specified!
> 
>> My reasoning for that was because it had occurred to me that what my 
>> patch series had done is allow for ntuples to be displayed via the 
>> get_rx_nfc interface.  As such you would end up with a location being 
>> implied when displaying the rules since it would give you a list of n 
>> entities.
> 
> We need to sort that out then.
> 
>> If you attempted to restore the rules you would probably end up with the 
>> location information for filters 0..(n-1), and that could be dropped 
>> since it would just be extra information.
>>
>>>>> Which is why I wrote that "@fs.@location specifies the index to
>>>>> use and must not be ignored."
>>>>>
>>>>> To support hardware where the filter table is hash-based rather than a
>>>>> TCAM, we would need some kind of flag or special value of location that
>>>>> means 'wherever'.
>>>> I'd find the 'wherever' option useful for TCAMs too.  Maybe even have a few
>>>> of those, like 'first available', 'any', and 'last available'.  The last one
>>>> is quite useful for catch-all rules without requiring one to know the TCAM size.
>>> Agreed.
>>>
>>> Ben.
>> The first and last options make a lot of sense to me.  The one I'm not 
>> sure about would be the "any" option.  It seems like it would be 
>> redundant with the "first available" option or is there something I'm 
>> missing?
>>
>> Also the code I have currently for the user space is just starting at 0 
>> and filling in the rules on a first available basis for location not 
>> specified.  Is this going to work for most cases or should I look at 
>> changing it to something like a "last available" approach for the nfc 
>> based filters?
> 
> My *guess* (and this is just a guess) is that users are more likely to
> want to specify explicit priorities for the high-priority rules and not
> for the low-priority rules.  So if the location is not explicitly set
> then we should choose the last available (lowest-priority) location in a
> TCAM, possibly excluding the very last location so that 'last' will
> still work.

I agree with this (and I misspoke in my previous mail about defaulting to 
first available).  I'll just reiterate that if a user doesn't specify a 
location the driver, rather than ethtool, needs to select one in order to 
accommodate any device restrictions.