Re: nf_flow on big-endian (was: Re: linux-next: build warning after merge of the net-next tree)

[Pablo Neira Ayuso <pablo@netfilter.org>]

[-- Attachment #1: Type: text/plain, Size: 1621 bytes --]

Hi Geert,

On Tue, Nov 26, 2019 at 12:06:03PM +0100, Geert Uytterhoeven wrote:
> On Thu, Nov 21, 2019 at 8:36 AM Stephen Rothwell <sfr@canb.auug.org.au> wrote:
> > After merging the net-next tree, today's linux-next build (powerpc
> > allyesconfig) produced this warning:
> >
> > net/netfilter/nf_flow_table_offload.c: In function 'nf_flow_rule_match':
> > net/netfilter/nf_flow_table_offload.c:80:21: warning: unsigned conversion from 'int' to '__be16' {aka 'short unsigned int'} changes value from '327680' to '0' [-Woverflow]
> >    80 |   mask->tcp.flags = TCP_FLAG_RST | TCP_FLAG_FIN;
> >       |                     ^~~~~~~~~~~~
> >
> > Introduced by commit
> >
> >   c29f74e0df7a ("netfilter: nf_flow_table: hardware offload support")
> 
> This is now upstream, and must be completely broken on big-endian
> platforms.
> 
> The other user of the flags field looks buggy, too
> (net/core/flow_dissector.c:__skb_flow_dissect_tcp()[*]):
> 
>      key_tcp->flags = (*(__be16 *) &tcp_flag_word(th) & htons(0x0FFF));
> 
> Disclaimer: I'm not familiar with the code or protocol, so below are just
> my gut feelings.
> 
>      struct flow_dissector_key_tcp {
>             __be16 flags;
>     };
> 
> Does this have to be __be16, i.e. does it go over the wire?
> If not, this should probably be __u16, and set using
> "be32_to_cpu(flags) >> 16"?
> If yes, "cpu_to_be16(be32_to_cpu(flags) >> 16)"?
> (Ugh, needs convenience macros)
> 
> [*] ac4bb5de27010e41 ("net: flow_dissector: add support for dissection
> of tcp flags")

I'm attaching a tentative patch, please let me know this is fixing up
this issue there.

Thanks.

[-- Attachment #2: x.patch --]
[-- Type: text/x-diff, Size: 1674 bytes --]

diff --git a/include/net/flow_dissector.h b/include/net/flow_dissector.h
index b8c20e9f343e..30ad4e07ff52 100644
--- a/include/net/flow_dissector.h
+++ b/include/net/flow_dissector.h
@@ -189,10 +189,17 @@ struct flow_dissector_key_eth_addrs {
 
 /**
  * struct flow_dissector_key_tcp:
- * @flags: flags
+ * @flags: TCP flags (16-bit, including the initial Data offset field bits)
+ * @word: Data offset + reserved bits + TCP flags + window
  */
 struct flow_dissector_key_tcp {
-	__be16 flags;
+	union {
+		struct {
+			__be16 flags;
+			__be16 __pad;
+		};
+		__be32	flag_word;
+	};
 };
 
 /**
diff --git a/net/core/flow_dissector.c b/net/core/flow_dissector.c
index ca871657a4c4..83af4633f306 100644
--- a/net/core/flow_dissector.c
+++ b/net/core/flow_dissector.c
@@ -756,7 +756,7 @@ __skb_flow_dissect_tcp(const struct sk_buff *skb,
 	key_tcp = skb_flow_dissector_target(flow_dissector,
 					    FLOW_DISSECTOR_KEY_TCP,
 					    target_container);
-	key_tcp->flags = (*(__be16 *) &tcp_flag_word(th) & htons(0x0FFF));
+	key_tcp->flag_word = tcp_flag_word(th);
 }
 
 static void
diff --git a/net/netfilter/nf_flow_table_offload.c b/net/netfilter/nf_flow_table_offload.c
index c94ebad78c5c..30205d57226d 100644
--- a/net/netfilter/nf_flow_table_offload.c
+++ b/net/netfilter/nf_flow_table_offload.c
@@ -87,8 +87,8 @@ static int nf_flow_rule_match(struct nf_flow_match *match,
 
 	switch (tuple->l4proto) {
 	case IPPROTO_TCP:
-		key->tcp.flags = 0;
-		mask->tcp.flags = TCP_FLAG_RST | TCP_FLAG_FIN;
+		key->tcp.flag_word = 0;
+		mask->tcp.flag_word = TCP_FLAG_RST | TCP_FLAG_FIN;
 		match->dissector.used_keys |= BIT(FLOW_DISSECTOR_KEY_TCP);
 		break;
 	case IPPROTO_UDP: