Skip to contentSkip to conversation

Fix expectaction mask dumping, take #3

PATCH In review Pablo Neira Ayuso 4 participants 14 messages Feb 13, 2006 – Feb 23, 2006 3 messages not in the archive

Subsystems (2)
lore
Share & tools
lore

Pablo Neira Ayuso

author maintainer
#0 lore raw

PATCH 1/4 Fix expectaction mask dumping, take #3

[CTNETLINK] Fix expectaction mask dumping

The expectation mask has some particularities that make handle in a different way. The protocol number fields can be set to non-valid protocols, ie. l3num is set to 0xFFFF. Since that protocol does not exist, the mask tuple will not be dumped. Moreover, this results in a kernel panic when nf_conntrack accesses the array of protocol handlers, that is PF_MAX (0x1F) long.

This patch introduces the function ctnetlink_exp_dump_mask, that correctly dumps the expectation mask. Such function uses the l3num value from the expectation tuple that is a valid layer 3 protocol number.

Besides, this modification introduces the attribute CTA_IP_L3NUM. Although the layer 3 protocol information is sent in the nfnetlink header, if the message contains information about an expectation, it will contain information about the master conntrack (just one of the tuples), the expectation tuple and the expectation mask. In this case, the value of l3num in the expectation mask is not the same that is set in the nfnetlink message. That is why we need another field that contain the value of l3num.

Signature · 2 lines
The dawn of the fourth age of Linux firewalling is coming; a time of
great struggle and heroic deeds -- J.Kadlecsik got inspired by J.Morris

[CTNETLINK] Fix expectaction mask dumping

The expectation mask has some particularities that make handle in a different way. The protocol number fields can be set to non-valid protocols, ie. l3num is set to 0xFFFF. Since that protocol does not exist, the mask tuple will not be dumped. Moreover, this results in a kernel panic when nf_conntrack accesses the array of protocol handlers, that is PF_MAX (0x1F) long.

This patch introduces the function ctnetlink_exp_dump_mask, that correctly dumps the expectation mask. Such function uses the l3num value from the expectation tuple that is a valid layer 3 protocol number.

Besides, this modification introduces the attribute CTA_IP_L3NUM. Although the layer 3 protocol information is sent in the nfnetlink header, if the message contains information about an expectation, it will contain information about the master conntrack (just one of the tuples), the expectation tuple and the expectation mask. In this case, the value of l3num in the expectation mask is not the same that is set in the nfnetlink message. That is why we need another field that contain the value of l3num.

net/netfilter/nf_conntrack_netlink.c +57 -22
Index: net-2.6.git/net/netfilter/nf_conntrack_netlink.c
===================================================================
@@ -4,7 +4,7 @@
  * (C) 2001 by Jay Schulist <[email protected]>
  * (C) 2002-2005 by Harald Welte <[email protected]>
  * (C) 2003 by Patrick Mchardy <[email protected]>
- * (C) 2005 by Pablo Neira Ayuso <[email protected]>
+ * (C) 2005-2006 by Pablo Neira Ayuso <[email protected]>
  *
  * I've reworked this stuff to use attributes instead of conntrack 
  * structures. 5.44 am. I need more tea. --pablo 05/07/11.
@@ -55,20 +55,18 @@ static char __initdata version[] = "0.92
 
 static inline int
 ctnetlink_dump_tuples_proto(struct sk_buff *skb, 
-			    const struct nf_conntrack_tuple *tuple)
+			    const struct nf_conntrack_tuple *tuple,
+			    struct nf_conntrack_protocol *proto)
 {
-	struct nf_conntrack_protocol *proto;
 	int ret = 0;
+	struct nfattr *nest_parms = NFA_NEST(skb, CTA_TUPLE_PROTO);
 
 	NFA_PUT(skb, CTA_PROTO_NUM, sizeof(u_int8_t), &tuple->dst.protonum);
 
-	/* If no protocol helper is found, this function will return the
-	 * generic protocol helper, so proto won't *ever* be NULL */
-	proto = nf_ct_proto_find_get(tuple->src.l3num, tuple->dst.protonum);
 	if (likely(proto->tuple_to_nfattr))
 		ret = proto->tuple_to_nfattr(skb, tuple);
 	
-	nf_ct_proto_put(proto);
+	NFA_NEST_END(skb, nest_parms);	
 
 	return ret;
 
@@ -77,33 +75,47 @@ nfattr_failure:
 }
 
 static inline int
-ctnetlink_dump_tuples(struct sk_buff *skb, 
-		      const struct nf_conntrack_tuple *tuple)
+ctnetlink_dump_tuples_ip(struct sk_buff *skb,
+			 const struct nf_conntrack_tuple *tuple,
+			 struct nf_conntrack_l3proto *l3proto)
 {
-	struct nfattr *nest_parms;
-	struct nf_conntrack_l3proto *l3proto;
 	int ret = 0;
-	
-	l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
-	
-	nest_parms = NFA_NEST(skb, CTA_TUPLE_IP);
+	struct nfattr *nest_parms = NFA_NEST(skb, CTA_TUPLE_IP);
+	u_int8_t l3num = (u_int8_t)tuple->src.l3num;
+
+	NFA_PUT(skb, CTA_IP_L3NUM, sizeof(u_int8_t), &l3num);
+
 	if (likely(l3proto->tuple_to_nfattr))
 		ret = l3proto->tuple_to_nfattr(skb, tuple);
+
 	NFA_NEST_END(skb, nest_parms);
 
+	return ret;
+
+nfattr_failure:
+	return -1;
+}
+
+static inline int
+ctnetlink_dump_tuples(struct sk_buff *skb, 
+		      const struct nf_conntrack_tuple *tuple)
+{
+	int ret;
+	struct nf_conntrack_l3proto *l3proto;
+	struct nf_conntrack_protocol *proto;
+
+	l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
+	ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 	nf_ct_l3proto_put(l3proto);
 
 	if (unlikely(ret < 0))
 		return ret;
 
-	nest_parms = NFA_NEST(skb, CTA_TUPLE_PROTO);
-	ret = ctnetlink_dump_tuples_proto(skb, tuple);
-	NFA_NEST_END(skb, nest_parms);
+	proto = nf_ct_proto_find_get(tuple->src.l3num, tuple->dst.protonum);
+	ret = ctnetlink_dump_tuples_proto(skb, tuple, proto);
+	nf_ct_proto_put(proto);
 
 	return ret;
-
-nfattr_failure:
-	return -1;
 }
 
 static inline int
@@ -1150,6 +1162,29 @@ nfattr_failure:
 }			
 
 static inline int
+ctnetlink_exp_dump_mask(struct sk_buff *skb, 
+			const struct nf_conntrack_tuple *tuple,
+			const struct nf_conntrack_tuple *mask)
+{
+	int ret;
+	struct nf_conntrack_l3proto *l3proto;
+	struct nf_conntrack_protocol *proto;
+
+	l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
+	ret = ctnetlink_dump_tuples_ip(skb, mask, l3proto);
+	nf_ct_l3proto_put(l3proto);
+
+	if (unlikely(ret < 0))
+		return ret;
+
+	proto = nf_ct_proto_find_get(tuple->src.l3num, tuple->dst.protonum);
+	ret = ctnetlink_dump_tuples_proto(skb, mask, proto);
+	nf_ct_proto_put(proto);
+
+	return ret;
+}
+
+static inline int
 ctnetlink_exp_dump_expect(struct sk_buff *skb,
                           const struct nf_conntrack_expect *exp)
 {
@@ -1159,7 +1194,7 @@ ctnetlink_exp_dump_expect(struct sk_buff
 
 	if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
 		goto nfattr_failure;
-	if (ctnetlink_exp_dump_tuple(skb, &exp->mask, CTA_EXPECT_MASK) < 0)
+	if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
 		goto nfattr_failure;
 	if (ctnetlink_exp_dump_tuple(skb,
 				 &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
include/linux/netfilter/nfnetlink_conntrack.h +1 -0
Index: net-2.6.git/include/linux/netfilter/nfnetlink_conntrack.h
===================================================================
@@ -52,6 +52,7 @@ enum ctattr_ip {
 	CTA_IP_V4_DST,
 	CTA_IP_V6_SRC,
 	CTA_IP_V6_DST,
+	CTA_IP_L3NUM,
 	__CTA_IP_MAX
 };
 #define CTA_IP_MAX (__CTA_IP_MAX - 1)
net/ipv4/netfilter/ip_conntrack_netlink.c +47 -17

115 lines not on this page · raw message

#1 lore raw
Quoted 1 line · Pablo Neira Ayuso

[CTNETLINK] Fix expectaction mask dumping

Thanks, Pablo. This patch looks fine to me. One cosmetic comment though:

Quoted 1 line

Besides, this modification introduces the attribute CTA_IP_L3NUM.

Please call it CTA_L3NUM or something like that. Nobody says that we will always and only deal with protocols that are called IP[vX].

If nobody else has any other comments, I suggest you resubmit with this requested change, and I'll apply and push to davem for net-2.6.17

Thanks!

Signature · 5 lines
- Harald Welte <[email protected]>                 http://netfilter.org/
============================================================================
  "Fragmentation is like classful addressing -- an interesting early
   architectural error that shows how much experimentation was going
   on while IP was being designed."                    -- Paul Vixie
#2 lore raw
In reply to Pablo Neira Ayuso
Quoted 9 lines · Pablo Neira Ayuso

[CTNETLINK] Fix expectaction mask dumping

The expectation mask has some particularities that make handle in a different way. The protocol number fields can be set to non-valid protocols, ie. l3num is set to 0xFFFF. Since that protocol does not exist, the mask tuple will not be dumped. Moreover, this results in a kernel panic when nf_conntrack accesses the array of protocol handlers, that is PF_MAX (0x1F) long.

Quoted 3 lines

This patch introduces the function ctnetlink_exp_dump_mask, that correctly dumps the expectation mask. Such function uses the l3num value from the expectation tuple that is a valid layer 3 protocol number.

This part looks fine to me, apart from one minor nitpick :)

Quoted 8 lines

Besides, this modification introduces the attribute CTA_IP_L3NUM. Although the layer 3 protocol information is sent in the nfnetlink header, if the message contains information about an expectation, it will contain information about the master conntrack (just one of the tuples), the expectation tuple and the expectation mask. In this case, the value of l3num in the expectation mask is not the same that is set in the nfnetlink message. That is why we need another field that contain the value of l3num.

I'm not sure I understand. The new attribute still contains the same value as the netlink header, doesn't it? So userspace should currently have at least two possibilities to get the correct value:

- use the value from the netlink header
- use the value from the tuple that comes with the mask, as the first part of your patch does. This seems most logically to me since the mask and the tuple belong together.

Quoted 9 lines
net/netfilter/nf_conntrack_netlink.c +3 -2
@@ -77,33 +75,47 @@ nfattr_failure:
 }
 
 static inline int
-ctnetlink_dump_tuples(struct sk_buff *skb, 
-		      const struct nf_conntrack_tuple *tuple)
+ctnetlink_dump_tuples_ip(struct sk_buff *skb,
+			 const struct nf_conntrack_tuple *tuple,
+			 struct nf_conntrack_l3proto *l3proto)

net/netfilter/nf_conntrack_netlink.c nfattr_failure: · line 84

Quoted 10 lines
net/netfilter/nf_conntrack_netlink.c +2 -6
 {
-	struct nfattr *nest_parms;
-	struct nf_conntrack_l3proto *l3proto;
 	int ret = 0;
-	
-	l3proto = nf_ct_l3proto_find_get(tuple->src.l3num);
-	
-	nest_parms = NFA_NEST(skb, CTA_TUPLE_IP);
+	struct nfattr *nest_parms = NFA_NEST(skb, CTA_TUPLE_IP);
+	u_int8_t l3num = (u_int8_t)tuple->src.l3num;

The cast is unnecessary.

#3 lore raw
Quoted 8 lines · Patrick McHardy

I'm not sure I understand. The new attribute still contains the same value as the netlink header, doesn't it? So userspace should currently have at least two possibilities to get the correct value:

- use the value from the netlink header
- use the value from the tuple that comes with the mask, as the first part of your patch does. This seems most logically to me since the mask and the tuple belong together.

I still dream about a connection with one end in ipv4 space and the other in ipv6. If we ever want to cover such a model, then the tuple needs to include l3num.

Signature · 5 lines
- Harald Welte <[email protected]>                 http://netfilter.org/
============================================================================
  "Fragmentation is like classful addressing -- an interesting early
   architectural error that shows how much experimentation was going
   on while IP was being designed."                    -- Paul Vixie
#4 lore raw
Quoted 13 lines · Harald Welte
Quoted 8 lines · Patrick McHardy

I'm not sure I understand. The new attribute still contains the same value as the netlink header, doesn't it? So userspace should currently have at least two possibilities to get the correct value:

- use the value from the netlink header
- use the value from the tuple that comes with the mask, as the first part of your patch does. This seems most logically to me since the mask and the tuple belong together.

Quoted 3 lines

I still dream about a connection with one end in ipv4 space and the other in ipv6. If we ever want to cover such a model, then the tuple needs to include l3num.

Yes, but until then it looks totally redundant. Since the bandwidth of netlink is limited, I think we shouldn't add new attributes without really needing them.

#5 lore raw

From: Patrick McHardy <[email protected]>
Date: Thu, 16 Feb 2006 21:11:01 +0100

Quoted 14 lines
Quoted 13 lines · Harald Welte
Quoted 8 lines · Patrick McHardy

I'm not sure I understand. The new attribute still contains the same value as the netlink header, doesn't it? So userspace should currently have at least two possibilities to get the correct value:

- use the value from the netlink header
- use the value from the tuple that comes with the mask, as the first part of your patch does. This seems most logically to me since the mask and the tuple belong together.

Quoted 7 lines
Quoted 3 lines

I still dream about a connection with one end in ipv4 space and the other in ipv6. If we ever want to cover such a model, then the tuple needs to include l3num.

Yes, but until then it looks totally redundant. Since the bandwidth of netlink is limited, I think we shouldn't add new attributes without really needing them.

'l3num' field in expectation mask may be 0xff. Then the new field is necessary so that kernel can pass the "exact value" in it to userspace.

But, I don't know whether userspace really wants to know the exact value in it or not. I assumed, yes, but if it is not ture, I'll agree to Patrick.

-- Yasuyuki Kozakai

A message that is not in the archive (sent off-list or not yet received): [email protected]

#7 lore raw
Quoted 8 lines · Yasuyuki KOZAKAI

From: Patrick McHardy <[email protected]>
Date: Thu, 16 Feb 2006 21:11:01 +0100

Quoted 3 lines

Yes, but until then it looks totally redundant. Since the bandwidth of netlink is limited, I think we shouldn't add new attributes without really needing them.

Quoted 5 lines

'l3num' field in expectation mask may be 0xff. Then the new field is necessary so that kernel can pass the "exact value" in it to userspace.

But, I don't know whether userspace really wants to know the exact value in it or not. I assumed, yes, but if it is not ture, I'll agree to Patrick.

My point was that a mask is pretty meaningless without the thing it masks, which is the tuple (except maybe a mask of all-zeros). The tuple itself already contains the correct protocol number.

#8 lore raw

From: Patrick McHardy <[email protected]>
Date: Thu, 23 Feb 2006 10:43:32 +0100

Quoted 12 lines
Quoted 11 lines · Yasuyuki KOZAKAI

From: Patrick McHardy <[email protected]>
Date: Thu, 16 Feb 2006 21:11:01 +0100

Quoted 3 lines

Yes, but until then it looks totally redundant. Since the bandwidth of netlink is limited, I think we shouldn't add new attributes without really needing them.

'l3num' field in expectation mask may be 0xff. Then the new field is necessary so that kernel can pass the "exact value" in it to userspace.

Quoted 6 lines
Quoted 2 lines

But, I don't know whether userspace really wants to know the exact value in it or not. I assumed, yes, but if it is not ture, I'll agree to Patrick.

My point was that a mask is pretty meaningless without the thing it masks, which is the tuple (except maybe a mask of all-zeros). The tuple itself already contains the correct protocol number.

What I want to say in the just previous mail I sent is same. :) I just wanted Pablo to confirm that.

-- Yasuyuki Kozakai

author maintainer
#9 lore raw
In reply to Patrick McHardy

Hi Patrick,

Quoted 9 lines · Patrick McHardy
Quoted 8 lines

Besides, this modification introduces the attribute CTA_IP_L3NUM. Although the layer 3 protocol information is sent in the nfnetlink header, if the message contains information about an expectation, it will contain information about the master conntrack (just one of the tuples), the expectation tuple and the expectation mask. In this case, the value of l3num in the expectation mask is not the same that is set in the nfnetlink message. That is why we need another field that contain the value of l3num.

Quoted 8 lines

I'm not sure I understand. The new attribute still contains the same value as the netlink header, doesn't it? So userspace should currently have at least two possibilities to get the correct value:

- use the value from the netlink header
- use the value from the tuple that comes with the mask, as the first part of your patch does. This seems most logically to me since the mask and the tuple belong together.

The problem is that currently the expectation mask is not dumped. find_l3proto returns the generic protocol handler for 0xFF, and that doesn't dump any layer 3 information. Moreover, the expectation mask has l3num value that is different from the l3num in the nfnetlink header, that's why I introduced this field.

I can send a patch to remove the expectation mask dumping but I'm not sure if this information could be useful for userspace helpers. Harald?

Signature · 2 lines
The dawn of the fourth age of Linux firewalling is coming; a time of
great struggle and heroic deeds -- J.Kadlecsik got inspired by J.Morris
author maintainer
#10 lore raw
Quoted 20 lines · Pablo Neira Ayuso
Quoted 9 lines · Patrick McHardy
Quoted 8 lines

Besides, this modification introduces the attribute CTA_IP_L3NUM. Although the layer 3 protocol information is sent in the nfnetlink header, if the message contains information about an expectation, it will contain information about the master conntrack (just one of the tuples), the expectation tuple and the expectation mask. In this case, the value of l3num in the expectation mask is not the same that is set in the nfnetlink message. That is why we need another field that contain the value of l3num.

Quoted 8 lines

I'm not sure I understand. The new attribute still contains the same value as the netlink header, doesn't it? So userspace should currently have at least two possibilities to get the correct value:

- use the value from the netlink header
- use the value from the tuple that comes with the mask, as the first part of your patch does. This seems most logically to me since the mask and the tuple belong together.

Quoted 8 lines

The problem is that currently the expectation mask is not dumped. find_l3proto returns the generic protocol handler for 0xFF, and that doesn't dump any layer 3 information. Moreover, the expectation mask has l3num value that is different from the l3num in the nfnetlink header, that's why I introduced this field.

I can send a patch to remove the expectation mask dumping but I'm not sure if this information could be useful for userspace helpers. Harald?

I think I can answer myself after some thinking: in order to create an expectation from userspace we will need to set the value of l3num of the expectation mask. Such value will be different from the value in the nfnetlink header, so I still think that we need that new CTA_L3NUM attribute.

Signature · 2 lines
The dawn of the fourth age of Linux firewalling is coming; a time of
great struggle and heroic deeds -- J.Kadlecsik got inspired by J.Morris
#11 lore raw

From: Pablo Neira Ayuso <[email protected]>
Date: Wed, 22 Feb 2006 04:20:39 +0100

Quoted 3 lines
Quoted 3 lines

The problem is that currently the expectation mask is not dumped. find_l3proto returns the generic protocol handler for 0xFF, and that doesn't dump any layer 3 information.

Yes, I want this bug to go away before 2.6.16 is shipped.

Quoted 3 lines
Quoted 3 lines

Moreover, the expectation mask has
l3num value that is different from the l3num in the nfnetlink header, that's why I introduced this field.

Yes, but in the current code l3num in expectation mask is always 0xff. And even if we port all helpers of ip_conntrack to nf_conntrack, they will set 0xff to l3num.

Which situation kernel wants to set the value except of 0xff to l3num in expectation mask ?

Quoted 8 lines
Quoted 2 lines

I can send a patch to remove the expectation mask dumping but I'm not sure if this information could be useful for userspace helpers. Harald?

I think I can answer myself after some thinking: in order to create an expectation from userspace we will need to set the value of l3num of the expectation mask. Such value will be different from the value in the nfnetlink header, so I still think that we need that new CTA_L3NUM attribute.

The similar question can arise. Which situation userspace wants to set the value except of 0xff in l3num in expectation mask ?

-- Yasuyuki Kozakai

A message that is not in the archive (sent off-list or not yet received): [email protected]

#13 lore raw
Quoted 9 lines · Yasuyuki KOZAKAI

From: Pablo Neira Ayuso <[email protected]>
Date: Wed, 22 Feb 2006 04:20:39 +0100

Quoted 3 lines
Quoted 3 lines

Moreover, the expectation mask has
l3num value that is different from the l3num in the nfnetlink header, that's why I introduced this field.

Quoted 6 lines

Yes, but in the current code l3num in expectation mask is always 0xff. And even if we port all helpers of ip_conntrack to nf_conntrack, they will set 0xff to l3num.

Which situation kernel wants to set the value except of 0xff to l3num in expectation mask ?

Harald and me were actually talking about simplifying the helper lookup by making the dst part of an expectation fixed (for ip_conntrack). With nf_conntrack l3num is in the src-part of the tuple, but thats just for memory-efficiency, it can be treated equally to protonum and the mask can be eliminated.

#14 lore raw

From: Patrick McHardy <[email protected]>
Date: Thu, 23 Feb 2006 10:48:57 +0100

Quoted 2 lines

Harald and me were actually talking about simplifying the helper lookup by making the dst part of an expectation fixed (for ip_conntrack). With

I got time to read deeply the patch from him yesterday midnight. (but that was for nf_conntrack). Simplifying the helper lookup is fine to me.

Quoted 3 lines

nf_conntrack l3num is in the src-part of the tuple, but thats just for memory-efficiency, it can be treated equally to protonum and the mask can be eliminated.

Yes, that patch actually eliminates it.

-- Yasuyuki Kozakai

A message that is not in the archive (sent off-list or not yet received): [email protected]

#16 lore raw
Quoted 11 lines · Yasuyuki KOZAKAI

From: Patrick McHardy <[email protected]>
Date: Thu, 23 Feb 2006 10:48:57 +0100

Quoted 2 lines

Harald and me were actually talking about simplifying the helper lookup by making the dst part of an expectation fixed (for ip_conntrack). With

I got time to read deeply the patch from him yesterday midnight. (but that was for nf_conntrack). Simplifying the helper lookup is fine to me.

Oops, I actually meant the expectation lookup. But the idea is pretty similar I think.