Re: [PATCH] ipv6: mcast: use rcu-safe version of ipv6_get_lladdr()

[Ignat Korchagin]

Stupid me - forgot to reply to all and a discussion between me and
David happend off list. Below, is the transcript for posterity:

On Sun, Feb 13, 2022 at 5:53 PM David Ahern <dsahern@xxxxxxxxx> wrote:
>
> On 2/13/22 8:43 AM, Ignat Korchagin wrote:
> > On Sun, Feb 13, 2022 at 4:17 PM David Ahern <dsahern@xxxxxxxxx> wrote:
> >>
> >> On 2/12/22 1:46 PM, Ignat Korchagin wrote:
> >>> In 8965779d2c0e ("ipv6,mcast: always hold idev->lock before mca_lock")
> >>> mld_newpack() was actually migrated from "dev" to "idev' just for this
> >>> use case. It seems the most reasonable approach would be to revert
> >>> mld_newpack() back to dev and use the original code.
> >>
> >>
> >> pmc already has the reference on idev and idev->dev is the source of dev
> >> passed to mld_newpack. There is no reason to go back to the idev -> dev
> >> -> idev dance.
> >
> > I don't know. Three things which make it more reasonable in my opinion:
> >   * we're already using idev->dev in mld_newpack() - that is we're not
> > adding an extra variable here in mld_newpack() - we need it anyway, so
> > can use in multiple places
> >   * it makes the code more consistent with the same code for the same
> > reason in igmp6_send() in the same file, which uses "dev" and
> > ipv6_get_lladdr()
> >   * we're making __ipv6_get_lladdr() static again and everything in
> > the kernel is now using the public version of ipv6_get_lladdr() - I
> > think the extra indirection of idev->dev-idev is a reasonable price to
> > pay to avoid customized locking code in the caller, which may backfire
> > later again in the same way it backfired this time
>
> which is why I later said move the locking to __ipv6_get_lladdr.
> ipv6_get_lladdr takes a net_dev, looks up the idev and calls
> __ipv6_get_lladdr. __ipv6_get_lladdr handles the idev locking needs.
>
> Users of the get_lladdr API that already have the idev reference use
> __ipv6_get_lladdr. That is a common paradigm in the stack.

Ah. I see now. This does make sense as well to me.

> igmp6 code can use some modernization - but that is a net-next change.
> This is a -net change.