[PATCH iptables-nft] iptables-nft: must withdraw PAYLOAD flag after parsing

[PATCH iptables-nft] iptables-nft: must withdraw PAYLOAD flag after parsing

[Florian Westphal]

else, next payload is stacked via 'CTX_PREV_PAYLOAD'.

Example breakage:

ip saddr 1.2.3.4 meta l4proto tcp
... is dumped as
-s 6.0.0.0 -p tcp

iptables-nft -s 1.2.3.4 -p tcp is dumped correctly, because
the expressions are ordered like:
meta l4proto tcp ip saddr 1.2.3.4

... and 'meta l4proto' will clear the PAYLOAD flag.

Fixes: 250dce876d92 ("nft-shared: support native tcp port delinearize")
Signed-off-by: Florian Westphal <fw@strlen.de>
---
 iptables/nft-shared.c                         |  2 ++
 .../ipt-restore/0018-multi-payload_0          | 27 +++++++++++++++++++
 2 files changed, 29 insertions(+)
 create mode 100755 iptables/tests/shell/testcases/ipt-restore/0018-multi-payload_0

diff --git a/iptables/nft-shared.c b/iptables/nft-shared.c
index 71e2f18dab92..66e09e8fd533 100644
--- a/iptables/nft-shared.c
+++ b/iptables/nft-shared.c
@@ -986,6 +986,8 @@ static void nft_parse_cmp(struct nft_xt_ctx *ctx, struct nftnl_expr *e)
 			nft_parse_transport(ctx, e, ctx->cs);
 			break;
 		}
+
+		ctx->flags &= ~NFT_XT_CTX_PAYLOAD;
 	}
 }
 
diff --git a/iptables/tests/shell/testcases/ipt-restore/0018-multi-payload_0 b/iptables/tests/shell/testcases/ipt-restore/0018-multi-payload_0
new file mode 100755
index 000000000000..f27577540d6e
--- /dev/null
+++ b/iptables/tests/shell/testcases/ipt-restore/0018-multi-payload_0
@@ -0,0 +1,27 @@
+#!/bin/bash
+
+# Make sure iptables-restore simply ignores
+# rules starting with -6
+
+set -e
+
+# show rules, drop uninteresting policy settings
+ipt_show() {
+	$XT_MULTI iptables-save | grep -- '-A INPUT'
+}
+
+# issue reproducer for iptables-restore
+
+$XT_MULTI iptables-restore <<EOF
+*filter
+-A INPUT -s 1.2.3.0/25 -p udp
+-A INPUT -s 1.2.3.0/26 -d 5.6.7.8/32
+-A INPUT -s 1.2.3.0/27 -d 10.2.0.0/16 -p tcp -j ACCEPT
+COMMIT
+EOF
+
+EXPECT='-A INPUT -s 1.2.3.0/25 -p udp
+-A INPUT -s 1.2.3.0/26 -d 5.6.7.8/32
+-A INPUT -s 1.2.3.0/27 -d 10.2.0.0/16 -p tcp -j ACCEPT'
+
+diff -u -Z <(echo -e "$EXPECT") <(ipt_show)
-- 
2.35.1

Re: [PATCH iptables-nft] iptables-nft: must withdraw PAYLOAD flag after parsing

[Florian Westphal]

Florian Westphal <fw@strlen.de> wrote:
> else, next payload is stacked via 'CTX_PREV_PAYLOAD'.
> 
> Example breakage:
> 
> ip saddr 1.2.3.4 meta l4proto tcp
> ... is dumped as
> -s 6.0.0.0 -p tcp
> 
> iptables-nft -s 1.2.3.4 -p tcp is dumped correctly, because
> the expressions are ordered like:
> meta l4proto tcp ip saddr 1.2.3.4
> 
> ... and 'meta l4proto' will clear the PAYLOAD flag.
> 
> Fixes: 250dce876d92 ("nft-shared: support native tcp port delinearize")
> Signed-off-by: Florian Westphal <fw@strlen.de>
> ---
>  iptables/nft-shared.c                         |  2 ++
>  .../ipt-restore/0018-multi-payload_0          | 27 +++++++++++++++++++
>  2 files changed, 29 insertions(+)
>  create mode 100755 iptables/tests/shell/testcases/ipt-restore/0018-multi-payload_0
> 
> diff --git a/iptables/nft-shared.c b/iptables/nft-shared.c
> index 71e2f18dab92..66e09e8fd533 100644
> --- a/iptables/nft-shared.c
> +++ b/iptables/nft-shared.c
> @@ -986,6 +986,8 @@ static void nft_parse_cmp(struct nft_xt_ctx *ctx, struct nftnl_expr *e)
>  			nft_parse_transport(ctx, e, ctx->cs);
>  			break;
>  		}
> +
> +		ctx->flags &= ~NFT_XT_CTX_PAYLOAD;
>  	}

This isn't ideal either since this breaks dissection of '1-42' ranges
that use two compare operands, i.e.:

cmp reg1 gte 1
cmp reg1 lte 42

...as first cmp 'hides' reg1 again.

I'd propose to rework this context stuff:
no more payload/meta/whatever flags, instead 'mirror' the raw data
registers.

Other ideas/suggestions?

Re: [PATCH iptables-nft] iptables-nft: must withdraw PAYLOAD flag after parsing

[Phil Sutter]

On Mon, Sep 19, 2022 at 11:31:09PM +0200, Florian Westphal wrote:
> Florian Westphal <fw@strlen.de> wrote:
> > else, next payload is stacked via 'CTX_PREV_PAYLOAD'.
> > 
> > Example breakage:
> > 
> > ip saddr 1.2.3.4 meta l4proto tcp
> > ... is dumped as
> > -s 6.0.0.0 -p tcp
> > 
> > iptables-nft -s 1.2.3.4 -p tcp is dumped correctly, because
> > the expressions are ordered like:
> > meta l4proto tcp ip saddr 1.2.3.4
> > 
> > ... and 'meta l4proto' will clear the PAYLOAD flag.
> > 
> > Fixes: 250dce876d92 ("nft-shared: support native tcp port delinearize")
> > Signed-off-by: Florian Westphal <fw@strlen.de>
> > ---
> >  iptables/nft-shared.c                         |  2 ++
> >  .../ipt-restore/0018-multi-payload_0          | 27 +++++++++++++++++++
> >  2 files changed, 29 insertions(+)
> >  create mode 100755 iptables/tests/shell/testcases/ipt-restore/0018-multi-payload_0
> > 
> > diff --git a/iptables/nft-shared.c b/iptables/nft-shared.c
> > index 71e2f18dab92..66e09e8fd533 100644
> > --- a/iptables/nft-shared.c
> > +++ b/iptables/nft-shared.c
> > @@ -986,6 +986,8 @@ static void nft_parse_cmp(struct nft_xt_ctx *ctx, struct nftnl_expr *e)
> >  			nft_parse_transport(ctx, e, ctx->cs);
> >  			break;
> >  		}
> > +
> > +		ctx->flags &= ~NFT_XT_CTX_PAYLOAD;
> >  	}
> 
> This isn't ideal either since this breaks dissection of '1-42' ranges
> that use two compare operands, i.e.:
> 
> cmp reg1 gte 1
> cmp reg1 lte 42
> 
> ...as first cmp 'hides' reg1 again.
> 
> I'd propose to rework this context stuff:
> no more payload/meta/whatever flags, instead 'mirror' the raw data
> registers.
> 
> Other ideas/suggestions?

When do we use multiple flags? I see we need NFT_XT_CTX_BITWISE in
addition to NFT_XT_CTX_META. Do we need e.g. NFT_XT_CTX_META and
NFT_XT_CTX_PAYLOAD at the same time? My idea would be to use an enum for
the LHS expression which is overwritten by each consecutive LHS
expression and a bitfield for the "on top" stuff.

Cheers, Phil

Re: [PATCH iptables-nft] iptables-nft: must withdraw PAYLOAD flag after parsing

[Florian Westphal]

Phil Sutter <phil@nwl.cc> wrote:
> On Mon, Sep 19, 2022 at 11:31:09PM +0200, Florian Westphal wrote:
> > Florian Westphal <fw@strlen.de> wrote:
> > > else, next payload is stacked via 'CTX_PREV_PAYLOAD'.
> > > 
> > > Example breakage:
> > > 
> > > ip saddr 1.2.3.4 meta l4proto tcp
> > > ... is dumped as
> > > -s 6.0.0.0 -p tcp
> > > 
> > > iptables-nft -s 1.2.3.4 -p tcp is dumped correctly, because
> > > the expressions are ordered like:
> > > meta l4proto tcp ip saddr 1.2.3.4
> > > 
> > > ... and 'meta l4proto' will clear the PAYLOAD flag.
> > > 
> > > Fixes: 250dce876d92 ("nft-shared: support native tcp port delinearize")
> > > Signed-off-by: Florian Westphal <fw@strlen.de>
> > > ---
> > >  iptables/nft-shared.c                         |  2 ++
> > >  .../ipt-restore/0018-multi-payload_0          | 27 +++++++++++++++++++
> > >  2 files changed, 29 insertions(+)
> > >  create mode 100755 iptables/tests/shell/testcases/ipt-restore/0018-multi-payload_0
> > > 
> > > diff --git a/iptables/nft-shared.c b/iptables/nft-shared.c
> > > index 71e2f18dab92..66e09e8fd533 100644
> > > --- a/iptables/nft-shared.c
> > > +++ b/iptables/nft-shared.c
> > > @@ -986,6 +986,8 @@ static void nft_parse_cmp(struct nft_xt_ctx *ctx, struct nftnl_expr *e)
> > >  			nft_parse_transport(ctx, e, ctx->cs);
> > >  			break;
> > >  		}
> > > +
> > > +		ctx->flags &= ~NFT_XT_CTX_PAYLOAD;
> > >  	}
> > 
> > This isn't ideal either since this breaks dissection of '1-42' ranges
> > that use two compare operands, i.e.:
> > 
> > cmp reg1 gte 1
> > cmp reg1 lte 42
> > 
> > ...as first cmp 'hides' reg1 again.
> > 
> > I'd propose to rework this context stuff:
> > no more payload/meta/whatever flags, instead 'mirror' the raw data
> > registers.
> > 
> > Other ideas/suggestions?
> 
> When do we use multiple flags? I see we need NFT_XT_CTX_BITWISE in
> addition to NFT_XT_CTX_META. Do we need e.g. NFT_XT_CTX_META and
> NFT_XT_CTX_PAYLOAD at the same time?

No, that makes no sense. I'm working on a new decoder that handles
each dreg individually.

Then, only one of meta/immediate/payload can be active per register.
bitwise can be set in addition.

I hope to have a rough RFC draft ready by tomorrw.
One big advantage is that we would no longer have to clear the flags
in the individual nft_expr -> xt_match/target dissectors.

Instead, those are always auto-cleared when a nft expressio writes
to the register.

> the LHS expression which is overwritten by each consecutive LHS
> expression and a bitfield for the "on top" stuff.

Yes, thats the same concept that I'm aiming for:

+enum nft_ctx_reg_type {
+       NFT_XT_REG_UNDEF,
+       NFT_XT_REG_PAYLOAD,
+       NFT_XT_REG_IMMEDIATE,
+       NFT_XT_REG_META,
+};
+
+struct nft_xt_ctx_reg {
+       enum nft_ctx_reg_type type:8;
+
+       union {
+               struct {
+                       uint32_t base;
+                       uint32_t offset;
+                       uint32_t len;
+               } payload;
+               struct {
+                       uint32_t data[4];
+                       uint8_t len;
+               } immediate;
+               struct {
+                       uint32_t key;
+               } meta;
+       };
+
+       struct {
+               uint32_t mask[4];
+               uint32_t xor[4];
+               bool set;
+       } bitwise;
+};