On Tue, Apr 18, 2023 at 02:43:02PM -0600, David Ahern wrote:
> On 4/18/23 10:41 AM, Leon Romanovsky wrote:
> > Hi,
> >
> > I came to the following diff which eliminates the kernel panics,
> > unfortunately I can explain only second hunk, but first is required
> > too.
> >
> > diff --git a/net/core/dst.c b/net/core/dst.c
> > index 3247e84045ca..750c8edfe29a 100644
> > --- a/net/core/dst.c
> > +++ b/net/core/dst.c
> > @@ -72,6 +72,8 @@ void dst_init(struct dst_entry *dst, struct dst_ops *ops,
> > dst->flags = flags;
> > if (!(flags & DST_NOCOUNT))
> > dst_entries_add(ops, 1);
> > +
> > + INIT_LIST_HEAD(&dst->rt_uncached);
>
> d288a162dd1c73507da582966f17dd226e34a0c0 moved rt_uncached from rt6_info
> and rtable to dst_entry. Only ipv4 and ipv6 usages initialize it. Since
> it is now in dst_entry, dst_init is the better place so it can be
> removed from rt_dst_alloc and rt6_info_init.
This is why I placed it there, but the rt_uncached list is initialized
in xfrm6 right before first call to rt6_uncached_list_add().
70 static int xfrm6_fill_dst(struct xfrm_dst *xdst, struct net_device *dev,
71 const struct flowi *fl)
72 {
...
92 INIT_LIST_HEAD(&xdst->u.rt6.dst.rt_uncached);
93 rt6_uncached_list_add(&xdst->u.rt6);
My silly explanation is that xfrm6_dst_destroy() can be called before xfrm6_fill_dst().
Thanks
>
