[Devel] [PATCH RHEL7 COMMIT] ms/netfilter: ipset: Add list flush to cancel_gc
Konstantin Khorenko
khorenko at virtuozzo.com
Mon Sep 30 15:21:54 MSK 2024
The commit is pushed to "branch-rh7-3.10.0-1160.119.1.vz7.224.x-ovz" and will appear at git at bitbucket.org:openvz/vzkernel.git
after rh7-3.10.0-1160.119.1.vz7.224.3
------>
commit 916bbfebb3fc1b677dbc1529716b00c7b3af3635
Author: Alexander Maltsev <keltar.gw at gmail.com>
Date: Wed Sep 25 17:35:43 2024 +0800
ms/netfilter: ipset: Add list flush to cancel_gc
Flushing list in cancel_gc drops references to other lists right away,
without waiting for RCU to destroy list. Fixes race when referenced
ipsets can't be destroyed while referring list is scheduled for destroy.
Fixes: 97f7cf1cd80e ("netfilter: ipset: fix performance regression in swap operation")
Signed-off-by: Alexander Maltsev <keltar.gw at gmail.com>
Acked-by: Jozsef Kadlecsik <kadlec at netfilter.org>
Signed-off-by: Pablo Neira Ayuso <pablo at netfilter.org>
https://virtuozzo.atlassian.net/browse/PSBM-155867
(cherry picked from commit c1193d9bbbd379defe9be3c6de566de684de8a6f)
Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
=================
Patchset description:
netfilter: ipset: Fix possible cause of memory corruption
Patch [1] fixes possible race between swap/destroy and add/del/test.
Here is is possible order of events when this race can lead to double
free with kfree_rcu on already freed hash bucket:
# Thread 1
+-> ip_set_add
+-> set = ip_set_rcu_get(xt_net(par), index)
< pause >
# Thread 2
ipset swap
ipset destroy
+-> mtype_destroy
+-> mtype_ahash_destroy
+-> n = __ipset_dereference(hbucket(t, i))
+-> kfree(n)
# Thread 1
< unpause >
+-> ip_set_lock(set)
+-> hash_net4_kadt
+-> mtype_variant->adt[adt]
+-> net4_kadt_add (mtype_add)
+-> n = rcu_dereference_bh(hbucket(t, key))
+-> if (n->pos >= n->size)
+-> old = n
+-> if (old != ERR_PTR(-ENOENT))
+-> kfree_rcu(old, rcu)
That can in it's turn lead to possible rcu free list corruption if this
double fried memory is reused just after rcu_free and before actual rcu
callback.
note1: The patch [1] has a reproducer but, sadly, I was unable to
reproduce the situation, even adding mdelay(100) to ip_set_add/del/test.
note2: All other patches are fixups to the original fixing patch.
note3: We don't have proof that this a fix to original issue, we only
know that ipset hash buckets were several times seen double freed just
before crash, which might indirectly indicate that the original problem
is related to ipset hash buckets.
Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
https://virtuozzo.atlassian.net/browse/PSBM-155867
Alexander Maltsev (1):
netfilter: ipset: Add list flush to cancel_gc
Eric Dumazet (1):
netns: add pre_exit method to struct pernet_operations
Jozsef Kadlecsik (5):
netfilter: ipset: fix race condition between swap/destroy and kernel
side add/del/test [1]
netfilter: ipset: fix performance regression in swap operation
netfilter: ipset: Missing gc cancellations fixed
netfilter: ipset: Fix race between namespace cleanup and gc in the
list:set type
netfilter: ipset: Fix suspicious rcu_dereference_protected()
---
net/netfilter/ipset/ip_set_list_set.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/net/netfilter/ipset/ip_set_list_set.c b/net/netfilter/ipset/ip_set_list_set.c
index b9dc3063e0d4..dd0fa9ada492 100644
--- a/net/netfilter/ipset/ip_set_list_set.c
+++ b/net/netfilter/ipset/ip_set_list_set.c
@@ -553,6 +553,9 @@ list_set_cancel_gc(struct ip_set *set)
if (SET_WITH_TIMEOUT(set))
del_timer_sync(&map->gc);
+
+ /* Flush list to drop references to other ipsets */
+ list_set_flush(set);
}
static const struct ip_set_type_variant set_variant = {
More information about the Devel
mailing list