Re: [RFC PATCH net] net: flow_offload: simplify hw stats check handling

[Jakub Kicinski <kuba@kernel.org>]

On Thu, 7 May 2020 18:46:43 +0200 Pablo Neira Ayuso wrote:
> On Thu, May 07, 2020 at 04:49:15PM +0100, Edward Cree wrote:
> > On 07/05/2020 16:32, Pablo Neira Ayuso wrote:  
> > > On Thu, May 07, 2020 at 03:59:09PM +0100, Edward Cree wrote:  
> > >> Make FLOW_ACTION_HW_STATS_DONT_CARE be all bits, rather than none, so that
> > >>  drivers and __flow_action_hw_stats_check can use simple bitwise checks.  
> > > 
> > > You have have to explain why this makes sense in terms of semantics.
> > > 
> > > _DISABLED and _ANY are contradicting each other.  
> > No, they aren't.  The DISABLED bit means "I will accept disabled", it doesn't
> >  mean "I insist on disabled".  What _does_ mean "I insist on disabled" is if
> >  the DISABLED bit is set and no other bits are.
> > So DISABLED | ANY means "I accept disabled; I also accept immediate or
> >  delayed".  A.k.a. "I don't care, do what you like".  
> 
> Jiri said Disabled means: bail out if you cannot disable it.

That's in TC uAPI Jiri chose... doesn't mean we have to do the same
internally.

> If the driver cannot disable, then it will have to check if the
> frontend is asking for Disabled (hence, report error to the frontend)
> or if it is actually asking for Don't care.
> 
> What you propose is a context-based interpretation of the bits. So
> semantics depend on how you accumulate/combine bits.
> 
> I really think bits semantics should be interpreted on the bit alone
> itself.

These 3 paragraphs sound to me like you were arguing for Ed's approach..

> There is one exception though, that is _ANY case, where you let the
> driver pick between delayed or immediate. But if the driver does not
> support for counters, it bails out in any case, so the outcome in both
> request is basically the same.
> 
> You are asking for different outcome depending on how bits are
> combined, which can be done, but it sounds innecessarily complicated
> to me.

No, quite the opposite, the code as committed to net has magic values
which drivers have to check.

The counter-proposal is that each bit represents a configuration, and
if more than one bit is set the driver gets to choose which it prefers. 
What could be simpler?

netfilter just has to explicitly set the field to DONT_CARE rather than 
depending on 0 form zalloc() coinciding with the correct value.