Re: [dpdk-dev] [PATCH] mbuf: support dynamic fields and flags

[Olivier Matz <olivier.matz@6wind.com>]

Hi Keith,

On Sat, Sep 21, 2019 at 08:28:32AM +0000, Wiles, Keith wrote:
> 
> 
> > On Sep 18, 2019, at 6:54 PM, Olivier Matz <olivier.matz@6wind.com> wrote:
> > 
> > Many features require to store data inside the mbuf. As the room in mbuf
> > structure is limited, it is not possible to have a field for each
> > feature. Also, changing fields in the mbuf structure can break the API
> > or ABI.
> > 
> > This commit addresses these issues, by enabling the dynamic registration
> > of fields or flags:
> > 
> > - a dynamic field is a named area in the rte_mbuf structure, with a
> >  given size (>= 1 byte) and alignment constraint.
> > - a dynamic flag is a named bit in the rte_mbuf structure.
> > 
> > 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).
> > 
> > The registration can be done at any moment, but it is not possible
> > to unregister fields or flags for now.
> > 
> > Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
> > Acked-by: Thomas Monjalon <thomas@monjalon.net>
> > —
> > 
> 

> The idea of registration for space in the mbuf I am not a big fan. I did like
> Konstantin’s suggestion of having the compiler help with optimizing the code,
> but with a slight difference. Maybe I misunderstand, but now with this design
> you have to pass the offsets to different parts of the application or place in
> global memory or have each section request the offsets. It seems great if the
> application is one big application or an appliance model application having
> control of the whole design not so good for service chains like designs where
> different parts of the whole application is design by different teams.

If the global variable storing the offset is defined in the mbuf layer, what
would be the problem?

The only things you would have to do is:

1/ ensure the offset is registered
   rte_mbuf_dyn_timestamp_register()

2/ use helpers
   rte_mbuf_dyn_timestamp_get(), rte_mbuf_dyn_timestamp_set(), ...

> Konstantin’s suggest if I understand it was to use structures to allow the
> compiler to optimize the access to the mbuf and I like that idea, but with one
> change we add a field in the mbuf to define the mbuf structure type.
>
> Say 0 is the standard rte_mbuf type then type 1 could be the IPSec offset type
> mbuf, type 2 could be something else, … The type 0 looks just like the mbuf we
> have today with maybe the optional fields set to reserved or some type of
> filler variables to reserve the holes in the structure. Then type 1 is the
> IPSec mbuf and in the reserved sections of the mbuf contain the IPSec related
> data with the standard mbuf fields still matching the type 0 version.

This very look like the "selective layout" in our presentation [1], page 14.

Your example talks about IPsec, but someone else will want to use a
sequence number, another one a timestamp, and another one will want to
use this space for its own application. There are a lot of use cases,
and it does not scale to have a layout for each of them. Worst, if
someone wants IPsec + a sequence number, how can it work?

One of the problem to solve is to avoid mutually exclusive feature (i.e.
union of fields that cannot be used together in the mbuf).

> This allows the mbuf to be used by the developer and the compiler now knows
> exactly where the fields are located in the structure and does not have to
> deal with any of the macros and offsets and registration suggested here. Just
> cast the mbuf pointer into the new type mbuf structure. We just have to make
> sure the code that needs to use a given mbuf type has access to the structure
> definitions.

With the current proposal, we can imagine an API to ask to register a
field at a specific offset. It can then be used in the application, so
that accesses are done at no cost compared to a static field, because
the offset would be const.

In the driver, the same logic could be used, but dynamically:

  if (offset == PREFERRED_OFFSET) {
    /* code with static offset */
  } else {
    /* generic code */
  }

But I'm not sure it would scale a lot if there are several features
using dynamic fields.

> If the mbufs it going to be translated from one type mbuf to another mbuf
> type, we just have to define that type and then cast the mbuf pointer to that
> structure. When an mbuf is received from IPSec PMD then the application needs
> to forward that mbuf to the next stage it can reset the type to 0 or to
> another type filling in the reserved fields to be used by the next stage in
> the pipeline.

What you describe is one use case.

What could be done with the API mentionned above (but I think it is
dangerous), is to allow a user to register 2 different fields at the
same offset, using a specific flag. This could work if the user knows
that these 2 fields are never used at the same time.

> The mbuf now contains the type and every point in the application can look at
> the type to determine how that mbuf is defined. I am sure there are some holes
> here, but I think it is a better solution then using all of these macros,
> offset values and registration APIs.

I'm not convinced having selective layouts is doable. The layouts cannot
fit all possible use cases, and managing the different layouts in the
driver looks difficult to me. Additionnaly, it does not solve the
problem of mutually exclusive features.


Thanks for the feedback.
Olivier

[1] https://static.sched.com/hosted_files/dpdkbordeaux2019/2b/dpdk-201909-dyn-mbuf.pdf