Re: [PATCH bpf-next 4/7] bpf: implement BPF XDP link-specific introspection APIs

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Mon, Jul 13, 2020 at 7:32 AM David Ahern <dsahern@gmail.com> wrote:
>
> On 7/10/20 4:49 PM, Andrii Nakryiko wrote:
> > diff --git a/net/core/dev.c b/net/core/dev.c
> > index 025687120442..a9c634be8dd7 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -8973,6 +8973,35 @@ static void bpf_xdp_link_dealloc(struct bpf_link *link)
> >       kfree(xdp_link);
> >  }
> >
> > +static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link,
> > +                                  struct seq_file *seq)
> > +{
> > +     struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
> > +     u32 ifindex = 0;
> > +
> > +     rtnl_lock();
> > +     if (xdp_link->dev)
> > +             ifindex = xdp_link->dev->ifindex;
> > +     rtnl_unlock();
>
> Patch 2 you set dev but don't hold a refcnt on it which is why you need
> the locking here. How do you know that the dev pointer is even valid here?
>
> If xdp_link is going to have dev reference you need to take the refcnt
> and you need to handle NETDEV notifications to cleanup the bpf_link when
> the device goes away.

Here I'm following the approach taken for cgroup and netns, where we
don't want to hold cgroup with extra refcnt (as well as netns for
bpf_netns_link). The dev is guaranteed to be valid because
dev_xdp_uninstall() will be called (under rtnl_lock) before net_device
is removed/destroyed. dev_xdp_uninstall() is the only one that can set
xdp_link->dev to NULL. So if we got rtnl_lock() and see non-NULL dev
here, it means that at worst we are waiting on a rtnl lock in
dev_xdp_uninstall() in a separate thread, and until this thread
releases that lock, it's ok to query dev.

Even if we do extra refcnt, due to dev_xdp_uninstall() which sets
xdp_link->dev to NULL, any code (fill_info, show_fdinfo, update, etc)
that does something with xdp_link->dev will have to take a lock
anyways.