Skip to contentSkip to conversation

netfilter: make net/stat/nf_conntrack procfs available again

PATCH In review Florian Westphal 2 participants 4 messages Jun 25, 2012 – Jun 27, 2012

Subsystems (2)
lore
Share & tools
lore

Florian Westphal

author maintainer
#0 lore raw

PATCH 1/1 netfilter: make net/stat/nf_conntrack procfs available again

partially reverts commit 54b07dca68557b0952585b5f4834cd0dd86eba35 (netfilter: provide config option to disable ancient procfs parts).

Problem is that this also disabled net/stat/nf_conntrack, which is useful for diagnosing certain conntrack-related issues; and there are currently no other means to obtain these statistics from userspace.

(conntrack-tools "conntrack -S" uses the proc interface, too...)

---
 net/netfilter/nf_conntrack_standalone.c |   30 +++++++++++++++++++++++++-----
 1 files changed, 25 insertions(+), 5 deletions(-)
net/netfilter/nf_conntrack_standalone.c +25 -5
index 9b39432..2363cf2 100644
@@ -269,6 +269,27 @@ static const struct file_operations ct_file_ops = {
 	.release = seq_release_net,
 };
 
+static bool nf_conntrack_procfs_create(struct net *net)
+{
+	return !!proc_net_fops_create(net, "nf_conntrack", 0440, &ct_file_ops);
+}
+
+static void nf_conntrack_procfs_remove(struct net *net)
+{
+	proc_net_remove(net, "nf_conntrack");
+}
+#else
+static bool nf_conntrack_procfs_create(struct net *net)
+{
+	return true;
+}
+
+static inline void nf_conntrack_procfs_remove(struct net *net)
+{
+}
+#endif /* CONFIG_NF_CONNTRACK_PROCFS */
+
+#ifdef CONFIG_PROC_FS
 static void *ct_cpu_seq_start(struct seq_file *seq, loff_t *pos)
 {
 	struct net *net = seq_file_net(seq);
@@ -366,8 +387,7 @@ static int nf_conntrack_standalone_init_proc(struct net *net)
 {
 	struct proc_dir_entry *pde;
 
-	pde = proc_net_fops_create(net, "nf_conntrack", 0440, &ct_file_ops);
-	if (!pde)
+	if (!nf_conntrack_procfs_create(net))
 		goto out_nf_conntrack;
 
 	pde = proc_create("nf_conntrack", S_IRUGO, net->proc_net_stat,
@@ -377,7 +397,7 @@ static int nf_conntrack_standalone_init_proc(struct net *net)
 	return 0;
 
 out_stat_nf_conntrack:
-	proc_net_remove(net, "nf_conntrack");
+	nf_conntrack_procfs_remove(net);
 out_nf_conntrack:
 	return -ENOMEM;
 }
@@ -385,7 +405,7 @@ out_nf_conntrack:
 static void nf_conntrack_standalone_fini_proc(struct net *net)
 {
 	remove_proc_entry("nf_conntrack", net->proc_net_stat);
-	proc_net_remove(net, "nf_conntrack");
+	nf_conntrack_procfs_remove(net);
 }
 #else
 static int nf_conntrack_standalone_init_proc(struct net *net)
@@ -396,7 +416,7 @@ static int nf_conntrack_standalone_init_proc(struct net *net)
 static void nf_conntrack_standalone_fini_proc(struct net *net)
 {
 }
-#endif /* CONFIG_NF_CONNTRACK_PROCFS */
+#endif /* CONFIG_PROC_FS */
 
 /* Sysctl support */
 
Signature · 1 line
1.7.3.4
#1 lore raw

Hi Florian,

Quoted 8 lines · Florian Westphal

partially reverts commit 54b07dca68557b0952585b5f4834cd0dd86eba35 (netfilter: provide config option to disable ancient procfs parts).

Problem is that this also disabled net/stat/nf_conntrack, which is useful for diagnosing certain conntrack-related issues; and there are currently no other means to obtain these statistics from userspace.

(conntrack-tools "conntrack -S" uses the proc interface, too...)

I can pass the following patch to David. It implements the missing code in ctnetlink to dump the statistics. Thus, conntrack doesn't use any /proc interface anymore (the changes to conntrack still pending).

You can test them by means of:

http://git.netfilter.org/cgi-bin/gitweb.cgi?p=libmnl.git;a=shortlog;h=refs/heads/nfct-stats

If you're OK with it, I'll integrate this in a backward compatible way (first try to use netlink, if not available, use /proc).

Still, I think that passing this to current may be useful. Although you can workaround this by enable that option. What do you prefer?

Quoted 1 line

From f1bb652b62671bc87d01ab2d94d267a481f795c5 Mon Sep 17 00:00:00 2001

From: Pablo Neira Ayuso <[email protected]>
Date: Tue, 26 Jun 2012 20:27:09 +0200
Subject: [PATCH] ctnetlink: export conntrack statistic via netlink

This patch adds the following messages:

IPCTNL_MSG_CT_GET_STATS_CPU
IPCTNL_MSG_CT_GET_STATS
IPCTNL_MSG_EXP_GET_STATS_CPU

To display connection tracking system per-cpu and global statistics.

This provides a replacement for the following /proc interfaces:

/proc/net/stat/nf_conntrack
/proc/sys/net/netfilter/nf_conntrack_count

---
 include/linux/netfilter/nfnetlink_conntrack.h |   38 +++++
 net/netfilter/nf_conntrack_netlink.c          |  227 ++++++++++++++++++++++++-
 2 files changed, 264 insertions(+), 1 deletion(-)
include/linux/netfilter/nfnetlink_conntrack.h +38 -0
index 7688833..f649f74 100644
@@ -7,6 +7,8 @@ enum cntl_msg_types {
 	IPCTNL_MSG_CT_GET,
 	IPCTNL_MSG_CT_DELETE,
 	IPCTNL_MSG_CT_GET_CTRZERO,
+	IPCTNL_MSG_CT_GET_STATS_CPU,
+	IPCTNL_MSG_CT_GET_STATS,
 
 	IPCTNL_MSG_MAX
 };
@@ -15,6 +17,7 @@ enum ctnl_exp_msg_types {
 	IPCTNL_MSG_EXP_NEW,
 	IPCTNL_MSG_EXP_GET,
 	IPCTNL_MSG_EXP_DELETE,
+	IPCTNL_MSG_EXP_GET_STATS_CPU,
 
 	IPCTNL_MSG_EXP_MAX
 };
@@ -203,4 +206,39 @@ enum ctattr_secctx {
 };
 #define CTA_SECCTX_MAX (__CTA_SECCTX_MAX - 1)
 
+enum ctattr_stats_cpu {
+	CTA_STATS_UNSPEC,
+	CTA_STATS_SEARCHED,
+	CTA_STATS_FOUND,
+	CTA_STATS_NEW,
+	CTA_STATS_INVALID,
+	CTA_STATS_IGNORE,
+	CTA_STATS_DELETE,
+	CTA_STATS_DELETE_LIST,
+	CTA_STATS_INSERT,
+	CTA_STATS_INSERT_FAILED,
+	CTA_STATS_DROP,
+	CTA_STATS_EARLY_DROP,
+	CTA_STATS_ERROR,
+	CTA_STATS_SEARCH_RESTART,
+	__CTA_STATS_MAX,
+};
+#define CTA_STATS_MAX (__CTA_STATS_MAX - 1)
+
+enum ctattr_stats_global {
+	CTA_STATS_GLOBAL_UNSPEC,
+	CTA_STATS_GLOBAL_ENTRIES,
+	__CTA_STATS_GLOBAL_MAX,
+};
+#define CTA_STATS_GLOBAL_MAX (__CTA_STATS_GLOBAL_MAX - 1)
+
+enum ctattr_expect_stats {
+	CTA_STATS_EXP_UNSPEC,
+	CTA_STATS_EXP_NEW,
+	CTA_STATS_EXP_CREATE,
+	CTA_STATS_EXP_DELETE,
+	__CTA_STATS_EXP_MAX,
+};
+#define CTA_STATS_EXP_MAX (__CTA_STATS_EXP_MAX - 1)
+
 #endif /* _IPCONNTRACK_NETLINK_H */
net/netfilter/nf_conntrack_netlink.c +226 -1

257 lines not on this page · raw message

Signature · 1 line
1.7.10
author maintainer
#2 lore raw
Quoted 8 lines · Pablo Neira Ayuso
Quoted 7 lines · Florian Westphal

partially reverts commit 54b07dca68557b0952585b5f4834cd0dd86eba35 (netfilter: provide config option to disable ancient procfs parts).

Problem is that this also disabled net/stat/nf_conntrack, which is useful for diagnosing certain conntrack-related issues; and there are currently no other means to obtain these statistics from userspace.

Quoted 5 lines
Quoted 1 line

(conntrack-tools "conntrack -S" uses the proc interface, too...)

I can pass the following patch to David. It implements the missing code in ctnetlink to dump the statistics. Thus, conntrack doesn't use any /proc interface anymore (the changes to conntrack still pending).

Thanks, that would be fine, too.

Quoted 2 lines

If you're OK with it, I'll integrate this in a backward compatible way (first try to use netlink, if not available, use /proc).

Sounds good.

Quoted 2 lines

Still, I think that passing this to current may be useful. Although you can workaround this by enable that option. What do you prefer?

Ignore my patch. In the meantime people can NF_CONNTRACK_PROCFS=y; we just have to wait a bit (e.g. a year) before killing the nfct proc code completely.

#3 lore raw
Quoted 15 lines · Florian Westphal
Quoted 8 lines · Pablo Neira Ayuso
Quoted 7 lines · Florian Westphal

partially reverts commit 54b07dca68557b0952585b5f4834cd0dd86eba35 (netfilter: provide config option to disable ancient procfs parts).

Problem is that this also disabled net/stat/nf_conntrack, which is useful for diagnosing certain conntrack-related issues; and there are currently no other means to obtain these statistics from userspace.

Quoted 5 lines
Quoted 1 line

(conntrack-tools "conntrack -S" uses the proc interface, too...)

I can pass the following patch to David. It implements the missing code in ctnetlink to dump the statistics. Thus, conntrack doesn't use any /proc interface anymore (the changes to conntrack still pending).

Quoted 6 lines

Thanks, that would be fine, too.

Quoted 2 lines

If you're OK with it, I'll integrate this in a backward compatible way (first try to use netlink, if not available, use /proc).

Sounds good.

JFYI:

http://git.netfilter.org/cgi-bin/gitweb.cgi?p=conntrack-tools.git;a=commit;h=8062d7fa6e0744a47c33ef0d3e17cc80ed005486

http://git.netfilter.org/cgi-bin/gitweb.cgi?p=conntrack-tools.git;a=commit;h=d3fa4f391fb40414a5c1cea16faac65d2c66a75c

I've pushed those two patches to the ct-stats branch to support dumping statistics via ctnetlink. I'll merge them once we hit 3.6-rc1.

Part of that code can be moved to the libraries, we can do that later.

BTW, you require git tree snapshot from libnetfilter_conntrack for this.

And the -S output format is not backward compatible to previous:

cpu=0         searched=9367 found=428234 new=287508 invalid=1 ignore=4 delete=291546 delete_list=6045 insert=2007 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0
cpu=1         searched=394 found=40745 new=1667 invalid=0 ignore=0 delete=378 delete_list=378 insert=1667 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0
cpu=2         searched=336 found=77297 new=2127 invalid=0 ignore=1 delete=661 delete_list=661 insert=2128 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0
cpu=3         searched=71 found=15016 new=1391 invalid=0 ignore=0 delete=87 delete_list=87 insert=1391 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0
Quoted 6 lines
Quoted 2 lines

Still, I think that passing this to current may be useful. Although you can workaround this by enable that option. What do you prefer?

Ignore my patch. In the meantime people can NF_CONNTRACK_PROCFS=y; we just have to wait a bit (e.g. a year) before killing the nfct proc code completely.

Makes sense. I'll pass my patch to David, drop this and extend the time to kill that /proc code for some time.