diff options
author | Jozsef Kadlecsik <kadlec@netfilter.org> | 2023-11-13 21:13:23 +0100 |
---|---|---|
committer | Pablo Neira Ayuso <pablo@netfilter.org> | 2023-11-14 16:16:21 +0100 |
commit | 28628fa952fefc7f2072ce6e8016968cc452b1ba (patch) | |
tree | 08f7350333f21b8edb207b37e435b66b7214d0c8 /net/ipv4/ipip.c | |
parent | netfilter: nf_tables: bogus ENOENT when destroying element which does not exist (diff) | |
download | linux-28628fa952fefc7f2072ce6e8016968cc452b1ba.tar.xz linux-28628fa952fefc7f2072ce6e8016968cc452b1ba.zip |
netfilter: ipset: fix race condition between swap/destroy and kernel side add/del/test
Linkui Xiao reported that there's a race condition when ipset swap and destroy is
called, which can lead to crash in add/del/test element operations. Swap then
destroy are usual operations to replace a set with another one in a production
system. The issue can in some cases be reproduced with the script:
ipset create hash_ip1 hash:net family inet hashsize 1024 maxelem 1048576
ipset add hash_ip1 172.20.0.0/16
ipset add hash_ip1 192.168.0.0/16
iptables -A INPUT -m set --match-set hash_ip1 src -j ACCEPT
while [ 1 ]
do
# ... Ongoing traffic...
ipset create hash_ip2 hash:net family inet hashsize 1024 maxelem 1048576
ipset add hash_ip2 172.20.0.0/16
ipset swap hash_ip1 hash_ip2
ipset destroy hash_ip2
sleep 0.05
done
In the race case the possible order of the operations are
CPU0 CPU1
ip_set_test
ipset swap hash_ip1 hash_ip2
ipset destroy hash_ip2
hash_net_kadt
Swap replaces hash_ip1 with hash_ip2 and then destroy removes hash_ip2 which
is the original hash_ip1. ip_set_test was called on hash_ip1 and because destroy
removed it, hash_net_kadt crashes.
The fix is to force ip_set_swap() to wait for all readers to finish accessing the
old set pointers by calling synchronize_rcu().
The first version of the patch was written by Linkui Xiao <xiaolinkui@kylinos.cn>.
v2: synchronize_rcu() is moved into ip_set_swap() in order not to burden
ip_set_destroy() unnecessarily when all sets are destroyed.
v3: Florian Westphal pointed out that all netfilter hooks run with rcu_read_lock() held
and em_ipset.c wraps the entire ip_set_test() in rcu read lock/unlock pair.
So there's no need to extend the rcu read locked area in ipset itself.
Closes: https://lore.kernel.org/all/69e7963b-e7f8-3ad0-210-7b86eebf7f78@netfilter.org/
Reported by: Linkui Xiao <xiaolinkui@kylinos.cn>
Signed-off-by: Jozsef Kadlecsik <kadlec@netfilter.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Diffstat (limited to 'net/ipv4/ipip.c')
0 files changed, 0 insertions, 0 deletions