Skip to contentSkip to conversation

netfilter: nf_tables: Implement jump limit for nft_table_validate

PATCH Superseded Shaun Brady 2 participants 4 messages Apr 22, 2025 – Apr 23, 2025

Subsystems (3)

v1 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 netfilter: nf_tables: Implement jump limit for nft_table_validate

Observing https://bugzilla.netfilter.org/show_bug.cgi?id=1665, I was able to reproduce the bug using linux-stable. Summarized, when adding large/repeated jump chains, while still staying under the NFT_JUMP_STACK_SIZE (currently 16), the kernel soons locks up.

From the bug report:
table='loop-test'
nft add table inet $table
nft add chain inet $table test0 '{type filter hook input priority 0; policy accept;}' for((i=1;i<16;i++));do nft add chain inet $table test$i;done
nft add rule inet $table test0 jump test1
for((i=1;i<15;i++));do nft add rule inet $table test$i jump test$((i+1));done nft add rule inet $table test15 tcp dport 8080 drop

  After the jump rule is added for 3 to 5 times, the system freezes and even softlockup occurs.
  for((i=1;i<15;i++));do nft add rule inet $table test$i jump test$((i+1));done
  for((i=1;i<15;i++));do nft add rule inet $table test$i jump test$((i+1));done
  for((i=1;i<15;i++));do nft add rule inet $table test$i jump test$((i+1));done

This patch is a different approach than the original proposed approach found in the bug report.

Add a new counter, total_jump_counter, to nft_ctx. On every call to nft_table_validate() (rule addition time, versus packet inspection time) start the counter at 0.

Increment said counter for every jump encountered during table validation. If the counter ever exceeds the system's jump limit *during validation*, gracefully reject the rule with -EMLINK (the same behavior as exceeding NFT_JUMP_STACK_SIZE).

This allows immediate feedback to the user about a bad chain, versus the original idea (from the bug report) of allowing the addition to the table. It keeps the in memory ruleset consistent, versus catching the failure during packet inspection at some unknown point in the future and arbitrarily denying the packet. Most importantly, it removes the original problem of a kernel crash.

The compile time limit NFT_DEFAULT_MAX_TABLE_JUMPS of 8192 was chosen to account for any normal use case, and when this value (and associated stressing loop table) was tested against a 1CPU/256MB machine, the system remained functional.

A sysctl entry net/netfilter/nf_max_table_jumps for the limit was also added for any use cases that may exceed this limit. As it is a control limit, it is only available in the root network namespace.

Example output of nft when patch is applied (and count is reached):

Error: Could not process rule: Too many links
add rule inet loop-test test14 jump test15
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

---
 Documentation/networking/netfilter-sysctl.rst |  8 ++
 include/net/netfilter/nf_tables.h             |  1 +
 include/net/netns/netfilter.h                 |  1 +
 net/netfilter/nf_tables_api.c                 | 75 ++++++++++++++++++-
 net/netfilter/nft_immediate.c                 |  1 +
 5 files changed, 85 insertions(+), 1 deletion(-)
Documentation/networking/netfilter-sysctl.rst +8 -0
index beb6d7b275d4..d957272c8ae6 100644
@@ -15,3 +15,11 @@ nf_log_all_netns - BOOLEAN
 	with LOG target; this aims to prevent containers from flooding host
 	kernel log. If enabled, this target also works in other network
 	namespaces. This variable is only accessible from init_net.
+
+nf_max_table_jumps - INTEGER (count)
+	default 8192
+
+	The maximum numbers of jumps a table can contain.  Meeting or exceeding
+	this value will cause additional rules to not be added with
+	EMLINK being return to the user. This variable is only accessible from
+	init_net.
include/net/netfilter/nf_tables.h +1 -0
index 0027beca5cd5..9c9473c37f08 100644
@@ -218,6 +218,7 @@ struct nft_ctx {
 	const struct nlattr * const 	*nla;
 	u32				portid;
 	u32				seq;
+	u32				total_jump_count;
 	u16				flags;
 	u8				family;
 	u8				level;
include/net/netns/netfilter.h +1 -0
index a6a0bf4a247e..cbb4f672d21f 100644
@@ -15,6 +15,7 @@ struct netns_nf {
 	const struct nf_logger __rcu *nf_loggers[NFPROTO_NUMPROTO];
 #ifdef CONFIG_SYSCTL
 	struct ctl_table_header *nf_log_dir_header;
+	struct ctl_table_header *nf_limit_control_dir_header;
 #ifdef CONFIG_LWTUNNEL
 	struct ctl_table_header *nf_lwtnl_dir_header;
 #endif
net/netfilter/nf_tables_api.c +74 -1

121 lines not on this page · raw message

net/netfilter/nft_immediate.c +1 -0
index 02ee5fb69871..b21736e389a4 100644
@@ -260,6 +260,7 @@ static int nft_immediate_validate(const struct nft_ctx *ctx,
 	case NFT_JUMP:
 	case NFT_GOTO:
 		pctx->level++;
+		pctx->total_jump_count++;
 		err = nft_chain_validate(ctx, data->verdict.chain);
 		if (err < 0)
 			return err;
Signature · 1 line
2.49.0
maintainer
#1 lore raw
Quoted 4 lines · Shaun Brady

The compile time limit NFT_DEFAULT_MAX_TABLE_JUMPS of 8192 was chosen to account for any normal use case, and when this value (and associated stressing loop table) was tested against a 1CPU/256MB machine, the system remained functional.

Keep in mind that one can register 1024 base chains, and that we have at least 5 hook points (ingress -> prerouting -> forward -> postrouting -> egress), so one could create a ruleset where a packet visits 41943040 chain jumps while in softirq context.

Furthermore, the largest ruleset I have archived here (iptables-save kubernetes ruleset dump) has 27k jumps (many who are mutually exclusive and user-defined chains that are always terminal), but nf_tables_api.c lacks the ability to detect either of these cases).

With the proposed change, the ruleset won't load anymore.

net/netfilter/nf_tables_api.c line 35

Quoted 2 lines
net/netfilter/nf_tables_api.c +2 -0
+u32 sysctl_nf_max_table_jumps __read_mostly = NFT_DEFAULT_MAX_TABLE_JUMPS;
+EXPORT_SYMBOL(sysctl_nf_max_table_jumps);

Why is this exported?

net/netfilter/nf_tables_api.c static struct notifier_block nft_nl_notifier = { · line 11907

Quoted 11 lines
net/netfilter/nf_tables_api.c +11 -0
+static int netfilter_limit_control_sysctl_init(struct net *net)
+{
+	if (net_eq(net, &init_net)) {
+		net->nf.nf_limit_control_dir_header = register_net_sysctl(
+				net,
+				"net/netfilter",
+				nf_limit_control_sysctl_table);
+		if (!net->nf.nf_limit_control_dir_header)
+			goto err_alloc;
+	}
+	return 0;

I think you can just make this a global variable.
Or, thats the alternative, make this a pernet tunable as long as the owning user_ns is the initial user namespace.

I think the idea is fine, but I'm not sure its going to work as-is.

Possible solutions to soften the impact/breakage potential:
- make the sysctl only affect non-init-net namespaces.
- make the sysctl only affect non-init-user-ns owned namespaces.

- Add the obseved total jump count to the table structure Then, when validating, do not start from 0 but from the sum of the total jump count of all registered tables in the same family.
netdev family will need to be counted unconditionally.

This will allow to reject ruleset that create 1k base chains for each family:hook point combination, which in turn would allow to increase the default upper limit.

#2 lore raw

Thanks for the feedback! Clarifications/questions inline.

Quoted 10 lines · Florian Westphal
Quoted 1 line · Shaun Brady

limit of 8192 was chosen to account for any normal use case

Furthermore, the largest ruleset I have archived here (iptables-save kubernetes ruleset dump) has 27k jumps (many who are mutually exclusive and user-defined chains that are always terminal), but nf_tables_api.c lacks the ability to detect either of these cases).

With the proposed change, the ruleset won't load anymore.

Much of my testing was omitted from the commit message. 8192 was chosen as to what seemed significantly above normal usage; I was way off. What I did observe was that machines (both big and small) start to act up around 16M. Would it ease minds to simply increase this to something like 4M or 8M? This would cover the largest case you have but keep us below the dangerous threshold. The tunable was in the event someone had an extreme use case we didn't think of, or if they wanted to be extra cautious.

Quoted 4 lines
Quoted 2 lines

+EXPORT_SYMBOL(sysctl_nf_max_table_jumps);

Why is this exported?

I believe I was initing at a different location that required this, and did not back this out. I will remove.

Quoted 4 lines

Possible solutions to soften the impact/breakage potential:
- make the sysctl only affect non-init-net namespaces.
- make the sysctl only affect non-init-user-ns owned namespaces.

I may be misunderstanding how limiting control to (only) non-init-* namespaces would help. It certainly would keep a namespace from taking the whole system down, but it would leave the original problem of being able to create the deadly jump configuration purely in the init-net. Maybe protecting from a namespace is more fruitful than an operator making mistakes (the initial revisions intent).

Quoted 5 lines

- Add the obseved total jump count to the table structure Then, when validating, do not start from 0 but from the sum of the total jump count of all registered tables in the same family.
netdev family will need to be counted unconditionally.

I had not considered one could spread the problem across multiple tables (even if you can't jump between them). This is a good insight, and I will account for this.

Quoted 1 line

I think the idea is fine, but I'm not sure its going to work as-is.

Glad to hear, and I would like to try for a rev2, if my question/clarifications make sense.

Thanks again!

SB

maintainer
#3 lore raw
Quoted 10 lines · Shaun Brady
Quoted 6 lines

Furthermore, the largest ruleset I have archived here (iptables-save kubernetes ruleset dump) has 27k jumps (many who are mutually exclusive and user-defined chains that are always terminal), but nf_tables_api.c lacks the ability to detect either of these cases).

With the proposed change, the ruleset won't load anymore.

Much of my testing was omitted from the commit message. 8192 was chosen as to what seemed significantly above normal usage; I was way off.

8k is brain damaged^W^W very high for nftables thanks to the existence of verdict maps. The problem is iptables(-nft) and its linear rules.

Quoted 3 lines

What I did observe was that machines (both big and small) start to act up around 16M. Would it ease minds to simply increase this to something like 4M or 8M?

What about going with 64k and NOT applying that limit in the init_netns?

The rationale would be that if you have the priviliges to ramp up the limitation threshold that limit doesn't exist in practice anyway.

Quoted 9 lines
Quoted 3 lines

Possible solutions to soften the impact/breakage potential:
- make the sysctl only affect non-init-net namespaces.
- make the sysctl only affect non-init-user-ns owned namespaces.

I may be misunderstanding how limiting control to (only) non-init-* namespaces would help. It certainly would keep a namespace from taking the whole system down, but it would leave the original problem of being able to create the deadly jump configuration purely in the init-net.

Sure, but why do you need to protect init_net?

Quoted 2 lines

Maybe protecting from a namespace is more fruitful than an operator making mistakes (the initial revisions intent).

I don't see how you can make such rulesets on accident.

Quoted 8 lines
Quoted 4 lines

- Add the obseved total jump count to the table structure Then, when validating, do not start from 0 but from the sum of the total jump count of all registered tables in the same family.
netdev family will need to be counted unconditionally.

I had not considered one could spread the problem across multiple tables (even if you can't jump between them). This is a good insight, and I will account for this.

Thanks!