A proc_remove() can sleep. so that it can't be inside of spin_lock.
Hence proc_remove() is moved to outside of spin_lock. and it also
adds mutex to sync create and remove of proc entry(config->pde).
test commands:
SHELL#1
%while :; do iptables -A INPUT -p udp -i enp2s0 -d 192.168.1.100 \
--dport 9000 -j CLUSTERIP --new --hashmode sourceip \
--clustermac 01:00:5e:00:00:21 --total-nodes 3 --local-node 3; \
iptables -F; done
SHELL#2
%while :; do echo +1 > /proc/net/ipt_CLUSTERIP/192.168.1.100; \
echo -1 > /proc/net/ipt_CLUSTERIP/192.168.1.100; done
[ 2949.569864] BUG: sleeping function called from invalid context at kernel/sched/completion.c:99
[ 2949.579944] in_atomic(): 1, irqs_disabled(): 0, pid: 5472, name: iptables
[ 2949.587920] 1 lock held by iptables/5472:
[ 2949.592711] #0:
000000008f0ebcf2 (&(&cn->lock)->rlock){+...}, at: refcount_dec_and_lock+0x24/0x50
[ 2949.603307] CPU: 1 PID: 5472 Comm: iptables Tainted: G W 4.19.0-rc5+ #16
[ 2949.604212] Hardware name: To be filled by O.E.M. To be filled by O.E.M./Aptio CRB, BIOS 5.6.5 07/08/2015
[ 2949.604212] Call Trace:
[ 2949.604212] dump_stack+0xc9/0x16b
[ 2949.604212] ? show_regs_print_info+0x5/0x5
[ 2949.604212] ___might_sleep+0x2eb/0x420
[ 2949.604212] ? set_rq_offline.part.87+0x140/0x140
[ 2949.604212] ? _rcu_barrier_trace+0x400/0x400
[ 2949.604212] wait_for_completion+0x94/0x710
[ 2949.604212] ? wait_for_completion_interruptible+0x780/0x780
[ 2949.604212] ? __kernel_text_address+0xe/0x30
[ 2949.604212] ? __lockdep_init_map+0x10e/0x5c0
[ 2949.604212] ? __lockdep_init_map+0x10e/0x5c0
[ 2949.604212] ? __init_waitqueue_head+0x86/0x130
[ 2949.604212] ? init_wait_entry+0x1a0/0x1a0
[ 2949.604212] proc_entry_rundown+0x208/0x270
[ 2949.604212] ? proc_reg_get_unmapped_area+0x370/0x370
[ 2949.604212] ? __lock_acquire+0x4500/0x4500
[ 2949.604212] ? complete+0x18/0x70
[ 2949.604212] remove_proc_subtree+0x143/0x2a0
[ 2949.708655] ? remove_proc_entry+0x390/0x390
[ 2949.708655] clusterip_tg_destroy+0x27a/0x630 [ipt_CLUSTERIP]
[ ... ]
Fixes: b3e456fce9f5 ("netfilter: ipt_CLUSTERIP: fix a race condition of proc file creation")
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
#endif
enum clusterip_hashmode hash_mode; /* which hashing mode */
u_int32_t hash_initval; /* hash initialization */
- struct rcu_head rcu;
+ struct rcu_head rcu; /* for call_rcu_bh */
struct net *net; /* netns for pernet list */
char ifname[IFNAMSIZ]; /* device ifname */
};
#ifdef CONFIG_PROC_FS
struct proc_dir_entry *procdir;
+ /* mutex protects the config->pde*/
+ struct mutex mutex;
#endif
};
local_bh_disable();
if (refcount_dec_and_lock(&c->entries, &cn->lock)) {
+ list_del_rcu(&c->list);
+ spin_unlock(&cn->lock);
+ local_bh_enable();
/* In case anyone still accesses the file, the open/close
* functions are also incrementing the refcount on their own,
* so it's safe to remove the entry even if it's in use. */
#ifdef CONFIG_PROC_FS
+ mutex_lock(&cn->mutex);
if (cn->procdir)
proc_remove(c->pde);
+ mutex_unlock(&cn->mutex);
#endif
- list_del_rcu(&c->list);
- spin_unlock(&cn->lock);
- local_bh_enable();
-
return;
}
local_bh_enable();
/* create proc dir entry */
sprintf(buffer, "%pI4", &ip);
+ mutex_lock(&cn->mutex);
c->pde = proc_create_data(buffer, 0600,
cn->procdir,
&clusterip_proc_fops, c);
+ mutex_unlock(&cn->mutex);
if (!c->pde) {
err = -ENOMEM;
goto err;
pr_err("Unable to proc dir entry\n");
return -ENOMEM;
}
+ mutex_init(&cn->mutex);
#endif /* CONFIG_PROC_FS */
return 0;
static void clusterip_net_exit(struct net *net)
{
struct clusterip_net *cn = clusterip_pernet(net);
+
#ifdef CONFIG_PROC_FS
+ mutex_lock(&cn->mutex);
proc_remove(cn->procdir);
cn->procdir = NULL;
+ mutex_unlock(&cn->mutex);
#endif
nf_unregister_net_hook(net, &cip_arp_ops);
}