Stop leaking mbuf tags during LRO.

[np]

Summary

tcp_lro_rx must free the tags that might be associated with the header
mbuf before it is demoted to an ordinary (not packet header) mbuf.

[adrian]

I'd rather we had an m_ function that did the downgrade of pkthdr->no-pkthdr and did the tidyup, like deleting mbuf chain.

That way we can use it everywhere consistently, add assertions, etc.

[np]

In D2708#51299, @adrian wrote:

I'd rather we had an m_ function that did the downgrade of pkthdr->no-pkthdr and did the tidyup, like deleting mbuf chain.

That way we can use it everywhere consistently, add assertions, etc.

There is an m_demote() but I thought it would be a bit heavyweight for tcp_lro_rx. We already know m in tcp_lro_rx is a header mbuf, and we don't want to spend time looking at any m->m_next linked mbufs. Maybe I should add this in mbuf.h next to m_demote?

void
m_demote_hdr(struct mbuf *m)
{
M_ASSERTPKTHDR(m);
m_tag_delete_chain(m, NULL);
m->m_flags &= ~M_PKTHDR;
}

[np]

Add a generic m_demote_hdr.

I'll build and test with this updated patch. The bzero in
m_demote/m_demote_hdr should perhaps be for kernel with INVARIANTS only.
I'll leave that alone for now.

[np]

Use _pkthdr as the suffix as there are existing routines (m_copy_pkthdr,
m_dup_pkthdr, more?) that use _pkthdr and not _hdr.

Add a few assertions in tcp_lro_rx.

[np]

Final version. Now I'll wait for reviews.

[hiren]

Good catch and fix.