Re: [PATCH net-next 1/3] ppp: convert rlock to rwlock to improve RX concurrency

[Eric Dumazet]

On Thu, Jun 26, 2025 at 9:00 PM Qingfang Deng <dqfext@xxxxxxxxx> wrote:
>
> On Fri, Jun 27, 2025 at 12:23 AM Guillaume Nault <gnault@xxxxxxxxxx> wrote:
> > That doesn't look right. Several PPP Rx features are stateful
> > (multilink, compression, etc.) and the current implementations
> > currently don't take any precaution when updating the shared states.
> >
> > For example, see how bsd_decompress() (in bsd_comp.c) updates db->*
> > fields all over the place. This db variable comes from ppp->rc_state,
> > which is passed as parameter of the ppp->rcomp->decompress() call in
> > ppp_decompress_frame().
> >
> > I think a lot of work would be needed before we could allow
> > ppp_do_recv() to run concurrently on the same struct ppp.
>
> Right. I think we can grab a write lock where it updates struct ppp.

tldr: network maintainers do not want rwlock back.

If you really care about concurrency, do not use rwlock, because it is
more expensive than a spinlock and very problematic.

Instead use RCU for readers, and spinlock for the parts needing exclusion.

Adding rwlock in network fast path is almost always a very bad choice.

Take a look at commit 0daf07e527095e6 for gory details.