Re: Minutes of techboard meeting, 2025-10-01

[Stephen Hemminger <stephen@networkplumber.org>]

[-- Attachment #1: Type: text/plain, Size: 2232 bytes --]

To help with definition of fast free flag, I would like the documentation
to be better.

Something like:
Fast free flag allows driver to avoid expensive atomic operators on ref
count and assume all mbufs in a burst are from the same pool.

Should also add debug asserts in drivers implenting fast free.

On Mon, Oct 6, 2025, 16:40 Morten Brørup <mb@smartsharesystems.com> wrote:

> > From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> > Sent: Friday, 3 October 2025 11.18
> > Subject: Minutes of techboard meeting, 2025-10-01
>
> > * Use of FAST_FREE and multi-buffer/scattered mbuf flags
> >   - The flags for enabling fast-free and supporting multi-mbuf packets
> > are
> >     now documented incompatible
> >   - Previously they were not defined as incompatible, but that seems to
> >     have been assumed for some usages.
> >   - Techboard discussed how best to resolve this incompatibility with
> >     regards to:
> >     - ensuring correctness
> >     - avoiding major churn to DPDK code
> >     - avoiding churn to end-user code
> >   - Options discussed:
> >     1 change definition back to not have the settings incompatible:
> > this
> >       necessitates checking drivers for correctness
> >     2 keep as explicitly incompatible and report error if both
> > specified:
> >       this could break end-user apps, and requires changes to example
> > apps
> >     3 drop the fast-free flag if multi-segment mbufs are also
> > specified:
> >       "hides" the issue, but probably minimises changes. Would need to
> >       decide whether the dropping of flag done in drivers vs ethdev
> > level.
> >       Pros and cons to both options. Needs clear documenting.
> >   - No firm decision reached, will discuss more over email.
>
> IMO, the patch [1] making MBUF_FAST_FREE and MULTI_SEGS explicitly
> incompatible should be reverted, at least for RC1.
> That will take the project back to the state it was in before we started
> this discussion.
> And all the examples broken by the patch (because they use both TX
> offloads) will not need fixing.
>
> [1]:
> https://patchwork.dpdk.org/project/dpdk/patch/20250803194218.683318-3-mb@smartsharesystems.com/
>
>

[-- Attachment #2: Type: text/html, Size: 3093 bytes --]