m_move_pkthdr leaves m_nextpkt 'dangling'

[Karim Fodil-Lemelin]

On 2017-10-13 5:10 PM, Gleb Smirnoff wrote:
> On Fri, Oct 13, 2017 at 12:59:47AM -0700, Adrian Chadd wrote:
> A> >>>> When doing so m_move_pkthdr is called to copy the current PKTHDR fields
> A> >>>> (tags and flags) to the mbuf that was prepended. The function also does:
> A> >>>>
> A> >>>> to->m_pkthdr = from->m_pkthdr;
> A> >>>>
> A> >>>> This, for the case I am interested in, essentially leaves the 'from'
> A> >>>> mbuf
> A> >>>> with a dangling pointer m_nextpkt pointing to the next fragment. While
> A> >>>> this
> A> >>>> is mostly harmless because only mbufs of pkthdr types are supposed to
> A> >>>> have
> A> >>>> m_nextpkt it triggers some panics when running with INVARIANTS in
> A> >>>> NetGraph
> A> >>>> (see ng_base.c :: CHECK_DATA_MBUF(m)):
> A> >>>>
> A> >>>> ...
> A> >>>>                          if (n->m_nextpkt != NULL)
> A> >>>> \
> A> >>>>                                  panic("%s: m_nextpkt", __func__);
> A> >>>> \
> A> >>>>                  }
> A> >>>> ...
> A> >>>>
> A> >>>> So I would like to propose the following patch:
> A> >>>>
> A> >>>> @@ -442,10 +442,11 @@ m_move_pkthdr(struct mbuf *to, struct mbuf *from)
> A> >>>>          if ((to->m_flags & M_EXT) == 0)
> A> >>>>                  to->m_data = to->m_pktdat;
> A> >>>>          to->m_pkthdr = from->m_pkthdr;          /* especially tags */
> A> >>>>          SLIST_INIT(&from->m_pkthdr.tags);       /* purge tags from src
> A> >>>> */
> A> >>>>          from->m_flags &= ~M_PKTHDR;
> A> >>>> +       from->m_nextpkt = NULL;
> A> >>>>   }
>
> Not only mbufs of M_PKTHDR may have m_nextpkt set. However, I tend to agree
> with the patch. But shouldn't we first copy the m_nextpkt to the new mbuf:
>
> +	to->m_nextpkt = from->m_nextpkt;
> +	from->m_nextpkt = NULL;
>
> Same way as we deal with tags.
>
>
Hi Gleb,

I think you are correct. If we look at the 'spirit' of m_move_pkthdr(); 
In my mind, it is to deep copy all fields related to a packet header and 
since m_nextpkt should only be carried by packet headers, it makes sense 
to copy it within m_move_pkthdr().

This also raises the question (my apologies in advance from bringing 
this up...) of weather or not m_nextpkt belongs in struct m_hdr and not 
in struct pkthdr.

In our case we are copying it explicitly outside the function as most of 
users of m_move_pkthdr() do.

Thanks for looking in to this.

Karim.