Re: [PATCH net v2] ip6mr: Fix lockdep and sparse RCU warnings

[ericnetdev dumazet]

Le mar. 1 oct. 2024 à 16:36, Stefan Wiehler <stefan.wiehler@xxxxxxxxx> a écrit :
>
> > This could be a lockdep annotation error then, at least for
> > RT6_TABLE_DFLT, oh well.
>
> As you have already explained, we can ignore the ip6mr_vif_seq_start() error
> path, so the issue boils down to ip6mr_get_table() being called without
> entering a RCU read-side critical section from these 4 functions:
> ipmr_vif_seq_start(), ip6mr_ioctl(), ip6mr_compat_ioctl() and
> ip6mr_get_route(). It is my understanding that in none of these four cases the
> RTNL lock is held either; at least according to the RCU-lockdep splat we
> clearly see that this is not the case in ip6mr_ioctl()  – but please correct me
> if I'm wrong.
>
> > Note that net/ipv4/ipmr.c would have a similar issue.
>
> Yes, looks indeed like that :-/
>
> > Please split your patch in small units, their Fixes: tags are likely
> > different, and if some code breaks something,
> > fixing the issue will be easier.
> >
> > The changelog seemed to only address the first ip6mr_vif_seq_start() part.
>
> If you prefer that I can split the change into 4 commits addressing each of the
> 4 functions mentioned before.

Yes please. Extending rcu_read_lock() sections needs inspection,
because we can not sleep there.