On 2021/11/15 5:01, Eric Dumazet wrote:
On 11/13/21 10:02 PM, Tetsuo Handa wrote:
quoted
sk_clone_lock() needs to call get_net() and sock_inuse_inc() together, or
s/sock_inuse_inc/sock_inuse_add/
quoted
socket_seq_show() will underflow when __sk_free() from sk_free() from
sk_free_unlock_clone() is called.
IMO, a "sock_inuse_get() underflow" is a very different problem,
Yes, a different problem. I found this problem while trying to examine
https://syzkaller.appspot.com/bug?extid=694120e1002c117747ed where
somebody might be failing to call get_net() or we might be failing to
make sure that all timers are synchronously stopped before put_net().
I suspect this should be fixed with the following patch.
My patch addresses a permanent underflow problem which remains as long as
that namespace exists. Your patch addresses a temporal underflow problem
which happens due to calculating the sum without locks. Therefore, we can
apply both patches if we want.
quoted hunk ↗ jump to hunk
diff --git a/net/core/sock.c b/net/core/sock.c
index c57d9883f62c75f522b7f6bc68451aaf8429dc83..bac8e2b62521301ce897728fff9622c4c05419a3 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -3573,7 +3573,7 @@ int sock_inuse_get(struct net *net)
for_each_possible_cpu(cpu)
res += *per_cpu_ptr(net->core.sock_inuse, cpu);
- return res;
+ return max(res, 0);
}
EXPORT_SYMBOL_GPL(sock_inuse_get);
Bug added in commit 648845ab7e200993dccd3948c719c858368c91e7
Author: Tonghao Zhang [off-list ref]
Date: Thu Dec 14 05:51:58 2017 -0800
sock: Move the socket inuse to namespace.