Re: [dpdk-dev] [EXT] Re: [PATCH 1/3] mbuf: add Tx offloads for packet marking

[Nithin Dabilpuram <ndabilpuram@marvell.com>]

On Fri, May 15, 2020 at 06:52:23PM +0200, Thomas Monjalon wrote:
> 15/05/2020 18:26, Jerin Jacob:
> > On Fri, May 15, 2020 at 8:40 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> > > 15/05/2020 15:44, Nithin Dabilpuram:
> > > > On Fri, May 15, 2020 at 03:12:59PM +0200, Thomas Monjalon wrote:
> > > > > 15/05/2020 12:08, Nithin Dabilpuram:
> > > > > > On Thu, May 14, 2020 at 10:29:31PM +0200, Olivier Matz wrote:
> > > > > > > I don't see any better approach than having a mbuf flag. However, I'm
> > > > > > > still not fully convinced that a dynamic flag won't do the job. Taking
> > > > > > > 3 additional flags (among 18 remaing) for this feature also means that
> > > > > > > we have 3 flags less for dynamic flags for all applications, even for
> > > > > > > applications that will not use this feature.
> > > > > > >
> > > > > > > Would it be a problem to use a dynamic flag in this case?
> > > > > > Since packet marking feature itself is already part of spec,
> > > > > > if we move the flags to PMD specific dynamic flag, then it creates a confusion.
> > > > > >
> > > > > > It is not the case of a custom feature supported by a specific PMD.
> > > > > > I believe when other PMD's implement packet marking, the same flags will
> > > > > > suffice.
> > > > >
> > > > > A dynamic flag is not necessarily PMD-specific.
> > > > > It is just avoiding consuming bits if the feature is not used by the application.
> > > > > We must move more existing flags and fields to be dynamic.
> > > > >
> > > > > In general, all new flags and fields in mbuf should be dynamic.
> > > > > And a work must be done to move existing stuff to free more space
> > > > > for more dynamic features.
> > > >
> > > > My bad, I thought dynamic flags can only be used for PMD specific thing.
> > > >
> > > > There is however a cost of using dynamic flag which I think should be avoided
> > > > for DPDK spec defined offloads, though it's fine for PMD specific things.
> > > >
> > > > Dynamic offload flags causes application and PMD to use non constant offset
> > > > or shift which are looked up at init, instead of having a constant shift or
> > > > offset. This indirection costs some cycles due to extra loads in fast path.
> > >
> > > Yes there is a cost. We described it quite clearly last year.
> > > The default rule is now to add new flags and fields as dynamic.
> > > In case the rule was not clear, I will send a patch to insert some
> > > notes in the code and the doc.
> > 
> > Yes. Please send a patch to document the rule. That makes life easy
> > for everyone to make a boolean decision.
> 
> Yes, I will work on it.
> 
> > Here is the comment from mbuf: support dynamic fields and flags commit
> > when accepted this patch.
> > 
> > "    The typical use case is a PMD that registers space for an offload
> >     feature, when the application requests to enable this feature.  As
> >     the space in mbuf is limited, the space should only be reserved if it
> >     is going to be used (i.e when the application explicitly asks for it).
> > "
> 
> OK, there is probably a documentation gap.
> 
> > If you are pushing this feature to dynamic mbuf filed then rte_tm
> > subsystem needs to register dynamic field
> > not the PMD as the feature is part of rte_tm spec.
> 
> Is there a function in rte_tm which initializes or configure the feature?
> 
> 
> > > If you disagree with this new rule, you will have to give very good arguments.
> > 
> > What would the definition of a good argument? as the same logic can be
> > implemented with dynamic vs
> > static at the cost of dynamic indirection.
> 
> I think the only exception to add a static flag or field is to demonstrate
> how basic is the feature.
> But I think all basic features are already integrated for years.
>

Vector implementations will always have huge cost due to dynamic offload
flags. Ofcourse, I don't have a vector implementation for packet marking but,
in future if there is such an offload flag with vector implementation
then hard enforcement will not work.

>