On Tue, 5 Aug 2025 07:48:41 +0200 Morten Brørup <m...@smartsharesystems.com> wrote:
> > From: fengchengwen [mailto:fengcheng...@huawei.com] > > Sent: Tuesday, 5 August 2025 02.59 > > > > On 8/4/2025 3:42 AM, Morten Brørup wrote: > > > Enabling some offload by default may conflict with a manually > > configured > > > offload. > > > Specifically, the mbuf fast release TX offload, which conflicts with > > multi > > > segment packet TX offload, was enabled by default. > > > Therefore, mbuf fast release TX offload (the only TX offload which was > > > enabled by default) is not enabled by default anymore. > > > > This may impact default performance, many vendor performance report has > > testpmd iofwd or macfwd case. > > Yes. > > IMO, any optimizations, mbuf-fast-free or mbuf-recycle, must be explicitly > enabled, and documented in the performance reports when used. > Otherwise the performance reports give a false impression of higher > performance, which is only available under certain circumstances. > I have complained about this before. > Which is also why mbuf-recycle must be explicitly enabled. > With this patch, the same finally applies to mbuf-fast-free too. > I could add a Fixes tag, but I think it is better to categorize this patch as > an improvement rather than a bug fix. > Agree that optimizations like this need to be opt-in. Maybe ethdev should automatically downgrade rather than reject bad configurations. The DPDK already has too many "benchmark special" tweaks.
