Skip to contentSkip to conversation

: Add mask support to fwmark classifier

NET_SCHED Patrick McHardy 3 participants 6 messages Aug 25, 2006 – Aug 26, 2006

Subsystems (3)
Networking Network drivers Traffic control (tc)
lore
Share & tools
lore

Patrick McHardy

author
#0 lore raw

This patch adds support to mask the nfmark value before the lookup the the fw classifier. Unfortunately it has some drawbacks, so I'd be interested if anyone can think of a better way.

The problem is that in order to avoid walking through all filters contained in one instance, we need to mask the value before the lookup. This means all filters share the same mask, which is taken from the first filter created and stored in the filter head. The user interface however always refers to a single filter, not the head, so it can't be changed afterwards unless we just overwrite it whenever a new filter is installed. Both is not really perfect. The current patch doesn't allow to change the mark and enforces that all filters use the same one, which I think is better than allowing inconsistent configurations.

Any better ideas?

[NET_SCHED]: Add mask support to fwmark classifier

Support masking the nfmark value before the search. The mask value is global for all filters contained in one instance. It can only be set when a new instance is created, all filters must specify the same mask.

Signed-off-by Patrick McHardy <[email protected]>

---
commit c7dff54dc2dca206ff54f66bfce290c49f98a3c8
tree 88d48096f13674f29413dc5c9853c7d0a8c5feac
parent e5d8ce21a2261f73b078d802bd2ab3508153b177
author Patrick McHardy <[email protected]> Fri, 25 Aug 2006 12:03:19 +0200 committer Patrick McHardy <[email protected]> Fri, 25 Aug 2006 12:03:19 +0200

 include/linux/pkt_cls.h |    1 +
 net/sched/cls_fw.c      |   25 ++++++++++++++++++++++++-
 2 files changed, 25 insertions(+), 1 deletions(-)
include/linux/pkt_cls.h +1 -0
index bd2c5a2..c3f01b3 100644
@@ -305,6 +305,7 @@ enum
 	TCA_FW_POLICE,
 	TCA_FW_INDEV, /*  used by CONFIG_NET_CLS_IND */
 	TCA_FW_ACT, /* used by CONFIG_NET_CLS_ACT */
+	TCA_FW_MASK,
 	__TCA_FW_MAX
 };
net/sched/cls_fw.c +24 -1
index e6973d9..c9385dc 100644
@@ -50,6 +50,7 @@ #define HTSIZE (PAGE_SIZE/sizeof(struct 
 struct fw_head
 {
 	struct fw_filter *ht[HTSIZE];
+	u32 mask;
 };
 
 struct fw_filter

@@ -101,7 +102,7 @@ static int fw_classify(struct sk_buff *s
 	struct fw_filter *f;
 	int r;
 #ifdef CONFIG_NETFILTER
-	u32 id = skb->nfmark;
+	u32 id = skb->nfmark & head->mask;
 #else
 	u32 id = 0;
 #endif

@@ -209,7 +210,9 @@ static int
 fw_change_attrs(struct tcf_proto *tp, struct fw_filter *f,
 	struct rtattr **tb, struct rtattr **tca, unsigned long base)
 {
+	struct fw_head *head = (struct fw_head*)tp->root;
 	struct tcf_exts e;
+	u32 mask;
 	int err;
 
 	err = tcf_exts_validate(tp, tb, tca[TCA_RATE-1], &e, &fw_ext_map);

@@ -232,6 +235,15 @@ #ifdef CONFIG_NET_CLS_IND
 	}
 #endif /* CONFIG_NET_CLS_IND */
 
+	if (tb[TCA_FW_MASK-1]) {
+		if (RTA_PAYLOAD(tb[TCA_FW_MASK-1]) != sizeof(u32))
+			goto errout;
+		mask = *(u32*)RTA_DATA(tb[TCA_FW_MASK-1]);
+		if (mask != head->mask)
+			goto errout;
+	} else if (head->mask != 0xFFFFFFFF)
+		goto errout;
+
 	tcf_exts_change(tp, &f->exts, &e);
 
 	return 0;

@@ -267,9 +279,17 @@ static int fw_change(struct tcf_proto *t
 		return -EINVAL;
 
 	if (head == NULL) {
+		u32 mask = 0xFFFFFFFF;
+		if (tb[TCA_FW_MASK-1]) {
+			if (RTA_PAYLOAD(tb[TCA_FW_MASK-1]) != sizeof(u32))
+				return -EINVAL;
+			mask = *(u32*)RTA_DATA(tb[TCA_FW_MASK-1]);
+		}
+
 		head = kzalloc(sizeof(struct fw_head), GFP_KERNEL);
 		if (head == NULL)
 			return -ENOBUFS;
+		head->mask = mask;
 
 		tcf_tree_lock(tp);
 		tp->root = head;

@@ -330,6 +350,7 @@ static void fw_walk(struct tcf_proto *tp
 static int fw_dump(struct tcf_proto *tp, unsigned long fh,
 		   struct sk_buff *skb, struct tcmsg *t)
 {
+	struct fw_head *head = (struct fw_head *)tp->root;
 	struct fw_filter *f = (struct fw_filter*)fh;
 	unsigned char	 *b = skb->tail;
 	struct rtattr *rta;

@@ -351,6 +372,8 @@ #ifdef CONFIG_NET_CLS_IND
 	if (strlen(f->indev))
 		RTA_PUT(skb, TCA_FW_INDEV, IFNAMSIZ, f->indev);
 #endif /* CONFIG_NET_CLS_IND */
+	if (head->mask != 0xFFFFFFFF)
+		RTA_PUT(skb, TCA_FW_MASK, 4, &head->mask);
 
 	if (tcf_exts_dump(skb, &f->exts, &fw_ext_map) < 0)
 		goto rtattr_failure;
#1 lore raw

* Patrick McHardy <[email protected]> 2006-08-25 12:29

Quoted 4 lines

This patch adds support to mask the nfmark value before the lookup the the fw classifier. Unfortunately it has some drawbacks, so I'd be interested if anyone can think of a better way.

Quoted 10 lines

The problem is that in order to avoid walking through all filters contained in one instance, we need to mask the value before the lookup. This means all filters share the same mask, which is taken from the first filter created and stored in the filter head. The user interface however always refers to a single filter, not the head, so it can't be changed afterwards unless we just overwrite it whenever a new filter is installed. Both is not really perfect. The current patch doesn't allow to change the mark and enforces that all filters use the same one, which I think is better than allowing inconsistent configurations.

The other option gets down to replacing the hash table with a list and that's not an option in my opinion. This looks very good to me.

author
#2 lore raw
Quoted 14 lines · Thomas Graf

* Patrick McHardy <[email protected]> 2006-08-25 12:29

Quoted 10 lines

The problem is that in order to avoid walking through all filters contained in one instance, we need to mask the value before the lookup. This means all filters share the same mask, which is taken from the first filter created and stored in the filter head. The user interface however always refers to a single filter, not the head, so it can't be changed afterwards unless we just overwrite it whenever a new filter is installed. Both is not really perfect. The current patch doesn't allow to change the mark and enforces that all filters use the same one, which I think is better than allowing inconsistent configurations.

Quoted 3 lines

The other option gets down to replacing the hash table with a list and that's not an option in my opinion. This looks very good to me.

Great, thanks. I'll send it off to Dave with two similar patches for IPv4 and DecNET routing rules.

#3 lore raw
Quoted 15 lines · Patrick McHardy
Quoted 14 lines · Thomas Graf

* Patrick McHardy <[email protected]> 2006-08-25 12:29

Quoted 10 lines

The problem is that in order to avoid walking through all filters contained in one instance, we need to mask the value before the lookup. This means all filters share the same mask, which is taken from the first filter created and stored in the filter head. The user interface however always refers to a single filter, not the head, so it can't be changed afterwards unless we just overwrite it whenever a new filter is installed. Both is not really perfect. The current patch doesn't allow to change the mark and enforces that all filters use the same one, which I think is better than allowing inconsistent configurations.

Quoted 7 lines
Quoted 3 lines

The other option gets down to replacing the hash table with a list and that's not an option in my opinion. This looks very good to me.

Great, thanks. I'll send it off to Dave with two similar patches for IPv4 and DecNET routing rules.

ACKed by me as well.

cheers,
jamal

#4 lore raw
In reply to Patrick McHardy
Quoted 15 lines · Patrick McHardy
Quoted 14 lines · Thomas Graf

* Patrick McHardy <[email protected]> 2006-08-25 12:29

Quoted 10 lines

The problem is that in order to avoid walking through all filters contained in one instance, we need to mask the value before the lookup. This means all filters share the same mask, which is taken from the first filter created and stored in the filter head. The user interface however always refers to a single filter, not the head, so it can't be changed afterwards unless we just overwrite it whenever a new filter is installed. Both is not really perfect. The current patch doesn't allow to change the mark and enforces that all filters use the same one, which I think is better than allowing inconsistent configurations.

Quoted 4 lines
Quoted 3 lines

The other option gets down to replacing the hash table with a list and that's not an option in my opinion. This looks very good to me.

This doesnt obsolete my previous ack, but:

Another approach could have been to add the mask as part of the hashing. and you add the new hash field not in the head rather in the filter. At runtime, you hash - walk the bucket and compare the mask as well as the index.

The above could be a future improvement.

cheers,
jamal

author
#5 lore raw
Quoted 4 lines · jamal

Another approach could have been to add the mask as part of the hashing. and you add the new hash field not in the head rather in the filter. At runtime, you hash - walk the bucket and compare the mask as well as the index.

That doesn't work. To what do you compare it? We have a mark from the packet .. but no mask. And we don't want to compare the mask but use it to mask the mark value.