On Wed, Jul 05, 2023 at 03:03:36PM +0200, Florian Westphal wrote:
> Thadeu Lima de Souza Cascardo <cascardo@xxxxxxxxxxxxx> wrote:
> > When evaluating byteorder expressions with size 2, a union with 32-bit and
> > 16-bit members is used. Since the 16-bit members are aligned to 32-bit,
> > the array accesses will be out-of-bounds.
> >
> > It may lead to a stack-out-of-bounds access like the one below:
>
> Yes, this is broken.
>
> > Using simple s32 and s16 pointers for each of these accesses fixes the
> > problem.
>
> I'm not sure this is correct. Its certainly less wrong of course.
>
> > Fixes: 96518518cc41 ("netfilter: add nftables")
> > Cc: stable@xxxxxxxxxxxxxxx
> > Reported-by: Tanguy DUBROCA (@SidewayRE) from @Synacktiv working with ZDI
> > Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@xxxxxxxxxxxxx>
> > ---
> > net/netfilter/nft_byteorder.c | 17 ++++++++++-------
> > 1 file changed, 10 insertions(+), 7 deletions(-)
> >
> > diff --git a/net/netfilter/nft_byteorder.c b/net/netfilter/nft_byteorder.c
> > index 9a85e797ed58..aa16bd2e92e2 100644
> > --- a/net/netfilter/nft_byteorder.c
> > +++ b/net/netfilter/nft_byteorder.c
> > @@ -30,11 +30,14 @@ void nft_byteorder_eval(const struct nft_expr *expr,
> > const struct nft_byteorder *priv = nft_expr_priv(expr);
> > u32 *src = ®s->data[priv->sreg];
> > u32 *dst = ®s->data[priv->dreg];
> > - union { u32 u32; u16 u16; } *s, *d;
> > + u32 *s32, *d32;
> > + u16 *s16, *d16;
> > unsigned int i;
> >
> > - s = (void *)src;
> > - d = (void *)dst;
> > + s32 = (void *)src;
> > + d32 = (void *)dst;
> > + s16 = (void *)src;
> > + d16 = (void *)dst;
> >
> > switch (priv->size) {
> > case 8: {
> > @@ -62,11 +65,11 @@ void nft_byteorder_eval(const struct nft_expr *expr,
> > switch (priv->op) {
> > case NFT_BYTEORDER_NTOH:
> > for (i = 0; i < priv->len / 4; i++)
> > - d[i].u32 = ntohl((__force __be32)s[i].u32);
> > + d32[i] = ntohl((__force __be32)s32[i]);
> > break;
> > case NFT_BYTEORDER_HTON:
> > for (i = 0; i < priv->len / 4; i++)
> > - d[i].u32 = (__force __u32)htonl(s[i].u32);
> > + d32[i] = (__force __u32)htonl(s32[i]);
> > break;
>
> Ack, this looks better, but I'd just use src[i] and dst[i] rather than
> the weird union pointers the original has.
>
> > @@ -74,11 +77,11 @@ void nft_byteorder_eval(const struct nft_expr *expr,
> > switch (priv->op) {
> > case NFT_BYTEORDER_NTOH:
> > for (i = 0; i < priv->len / 2; i++)
> > - d[i].u16 = ntohs((__force __be16)s[i].u16);
> > + d16[i] = ntohs((__force __be16)s16[i]);
>
> This on the other hand... I'd say this should mimic what the 64bit
> case is doing and use nft_reg_store16() nft_reg_load16() helpers for
> the register accesses.
>
> something like:
>
> for (i = 0; i < priv->len / 2; i++) {
> v16 = nft_reg_load16(&src[i]);
> nft_reg_store16(&dst[i], + ntohs((__force __be16)v16));
> }
>
The problem here is that we cannot index the 32-bit dst and src pointers as if
they were 16-bit pointers. We will end up with the exact same problem we are
trying to fix here.
I can change the code to use the accessors, but they use u32 pointers, so it
would end up looking like:
diff --git a/net/netfilter/nft_byteorder.c b/net/netfilter/nft_byteorder.c
index 9a85e797ed58..fd8ce6426b2b 100644
--- a/net/netfilter/nft_byteorder.c
+++ b/net/netfilter/nft_byteorder.c
@@ -30,11 +30,10 @@ void nft_byteorder_eval(const struct nft_expr *expr,
const struct nft_byteorder *priv = nft_expr_priv(expr);
u32 *src = ®s->data[priv->sreg];
u32 *dst = ®s->data[priv->dreg];
- union { u32 u32; u16 u16; } *s, *d;
unsigned int i;
- s = (void *)src;
- d = (void *)dst;
+ u16 *s16 = (void *)src;
+ u16 *d16 = (void *)dst;
switch (priv->size) {
case 8: {
@@ -62,23 +61,29 @@ void nft_byteorder_eval(const struct nft_expr *expr,
switch (priv->op) {
case NFT_BYTEORDER_NTOH:
for (i = 0; i < priv->len / 4; i++)
- d[i].u32 = ntohl((__force __be32)s[i].u32);
+ dst[i] = ntohl((__force __be32)src[i]);
break;
case NFT_BYTEORDER_HTON:
for (i = 0; i < priv->len / 4; i++)
- d[i].u32 = (__force __u32)htonl(s[i].u32);
+ dst[i] = (__force __u32)htonl(src[i]);
break;
}
break;
case 2:
switch (priv->op) {
case NFT_BYTEORDER_NTOH:
- for (i = 0; i < priv->len / 2; i++)
- d[i].u16 = ntohs((__force __be16)s[i].u16);
+ for (i = 0; i < priv->len / 2; i++) {
+ __be16 src16;
+ src16 = nft_reg_load_be16((u32 *)&s16[i]);
+ nft_reg_store_be16((u32 *)&d16[i], ntohs(src16));
+ }
break;
case NFT_BYTEORDER_HTON:
- for (i = 0; i < priv->len / 2; i++)
- d[i].u16 = (__force __u16)htons(s[i].u16);
+ for (i = 0; i < priv->len / 2; i++) {
+ u16 src16;
+ src16 = nft_reg_load16((u32 *)&s16[i]);
+ nft_reg_store16((u32 *)&d16[i], (__force __u16)htons(src16));
+ }
break;
}
break;
> [ not even compile tested ]
>
> Same for the htons case.
>
> On a slightly related note, some of the nftables test cases create bogus
> conversions, e.g.:
>
> # src/nft --debug=netlink add rule ip6 t c 'ct mark set ip6 dscp << 2 |
> # 0x10'
> ip6 t c
> [ payload load 2b @ network header + 0 => reg 1 ]
> [ bitwise reg 1 = ( reg 1 & 0x0000c00f ) ^ 0x00000000 ]
> [ bitwise reg 1 = ( reg 1 >> 0x00000006 ) ]
> [ byteorder reg 1 = ntoh(reg 1, 2, 1) ] // NO-OP! should be reg 1, 2, 2) I presume?
>
> I'd suggest to add a patch for nf-next that rejects such crap.
