2016-12-02, 14:09:25 -0500, David Miller wrote:
> From: Sabrina Dubroca <s...@queasysnail.net>
> Date: Fri, 2 Dec 2016 16:49:29 +0100
>
> > geneve{,6}_build_skb can end up doing a pskb_expand_head(), which
> > makes the ip_hdr(skb) reference we stashed earlier stale. Since it's
> > only needed as an argument to ip_tunnel_ecn_encap(), move this
> > directly in the function call.
> >
> > Fixes: 08399efc6319 ("geneve: ensure ECN info is handled properly in all
> > tx/rx paths")
> > Signed-off-by: Sabrina Dubroca <s...@queasysnail.net>
>
> Applied and queued up for -stable, thanks.
>
> This bug happens so many times that I think it might be time for
> a debugging mode for pskb_expand_head() that unconditionally
> reallocates the skb->data buffer regardless of whether it's
> necessary or not and somehow unmaps the previous buffer to
> force a trap on stale pointers.
The problem with that is you'd need to enable the "debugging mode" in
all wrappers, so that they don't bypass the actual call to
pskb_expand_head(). And that still leaves all the direct calls to
pskb_expand_head() that are guarded by some kind of check (just two
random hits without even looking very hard:
net/core/pktgen.c:process_ipsec, net/ipv4/ip_gre.c:gre_fb_xmit).
Then I think we could just rely on KASAN (that's how I noticed this
bug).
> Better ideas welcome, of course :)
May not be better ;) but at least another idea:
I'd like to try something based on static analysis. We'd need a way to
tag cached pointers to skb->data (via ip_hdr() or whatever), and
propagate the notion that pskb_expand_head() makes these cached
pointers stale through layers of function calls. I don't know how
feasible this is with the tools we have.
--
Sabrina
