xfrm: destroy xfrm_state synchronously on net exit path
authorCong Wang <xiyou.wangcong@gmail.com>
Thu, 31 Jan 2019 21:05:49 +0000 (13:05 -0800)
committerSteffen Klassert <steffen.klassert@secunet.com>
Tue, 5 Feb 2019 05:29:20 +0000 (06:29 +0100)
xfrm_state_put() moves struct xfrm_state to the GC list
and schedules the GC work to clean it up. On net exit call
path, xfrm_state_flush() is called to clean up and
xfrm_flush_gc() is called to wait for the GC work to complete
before exit.

However, this doesn't work because one of the ->destructor(),
ipcomp_destroy(), schedules the same GC work again inside
the GC work. It is hard to wait for such a nested async
callback. This is also why syzbot still reports the following
warning:

 WARNING: CPU: 1 PID: 33 at net/ipv6/xfrm6_tunnel.c:351 xfrm6_tunnel_net_exit+0x2cb/0x500 net/ipv6/xfrm6_tunnel.c:351
 ...
  ops_exit_list.isra.0+0xb0/0x160 net/core/net_namespace.c:153
  cleanup_net+0x51d/0xb10 net/core/net_namespace.c:551
  process_one_work+0xd0c/0x1ce0 kernel/workqueue.c:2153
  worker_thread+0x143/0x14a0 kernel/workqueue.c:2296
  kthread+0x357/0x430 kernel/kthread.c:246
  ret_from_fork+0x3a/0x50 arch/x86/entry/entry_64.S:352

In fact, it is perfectly fine to bypass GC and destroy xfrm_state
synchronously on net exit call path, because it is in process context
and doesn't need a work struct to do any blocking work.

This patch introduces xfrm_state_put_sync() which simply bypasses
GC, and lets its callers to decide whether to use this synchronous
version. On net exit path, xfrm_state_fini() and
xfrm6_tunnel_net_exit() use it. And, as ipcomp_destroy() itself is
blocking, it can use xfrm_state_put_sync() directly too.

Also rename xfrm_state_gc_destroy() to ___xfrm_state_destroy() to
reflect this change.

Fixes: b48c05ab5d32 ("xfrm: Fix warning in xfrm6_tunnel_net_exit.")
Reported-and-tested-by: syzbot+e9aebef558e3ed673934@syzkaller.appspotmail.com
Cc: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
include/net/xfrm.h
net/ipv6/xfrm6_tunnel.c
net/key/af_key.c
net/xfrm/xfrm_state.c
net/xfrm/xfrm_user.c

index 7298a53..85386be 100644 (file)
@@ -853,7 +853,7 @@ static inline void xfrm_pols_put(struct xfrm_policy **pols, int npols)
                xfrm_pol_put(pols[i]);
 }
 
-void __xfrm_state_destroy(struct xfrm_state *);
+void __xfrm_state_destroy(struct xfrm_state *, bool);
 
 static inline void __xfrm_state_put(struct xfrm_state *x)
 {
@@ -863,7 +863,13 @@ static inline void __xfrm_state_put(struct xfrm_state *x)
 static inline void xfrm_state_put(struct xfrm_state *x)
 {
        if (refcount_dec_and_test(&x->refcnt))
-               __xfrm_state_destroy(x);
+               __xfrm_state_destroy(x, false);
+}
+
+static inline void xfrm_state_put_sync(struct xfrm_state *x)
+{
+       if (refcount_dec_and_test(&x->refcnt))
+               __xfrm_state_destroy(x, true);
 }
 
 static inline void xfrm_state_hold(struct xfrm_state *x)
@@ -1590,7 +1596,7 @@ struct xfrmk_spdinfo {
 
 struct xfrm_state *xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq);
 int xfrm_state_delete(struct xfrm_state *x);
-int xfrm_state_flush(struct net *net, u8 proto, bool task_valid);
+int xfrm_state_flush(struct net *net, u8 proto, bool task_valid, bool sync);
 int xfrm_dev_state_flush(struct net *net, struct net_device *dev, bool task_valid);
 void xfrm_sad_getinfo(struct net *net, struct xfrmk_sadinfo *si);
 void xfrm_spd_getinfo(struct net *net, struct xfrmk_spdinfo *si);
index f5b4feb..bc65db7 100644 (file)
@@ -344,8 +344,8 @@ static void __net_exit xfrm6_tunnel_net_exit(struct net *net)
        struct xfrm6_tunnel_net *xfrm6_tn = xfrm6_tunnel_pernet(net);
        unsigned int i;
 
-       xfrm_state_flush(net, IPSEC_PROTO_ANY, false);
        xfrm_flush_gc();
+       xfrm_state_flush(net, IPSEC_PROTO_ANY, false, true);
 
        for (i = 0; i < XFRM6_TUNNEL_SPI_BYADDR_HSIZE; i++)
                WARN_ON_ONCE(!hlist_empty(&xfrm6_tn->spi_byaddr[i]));
index 655c787..637030f 100644 (file)
@@ -1783,7 +1783,7 @@ static int pfkey_flush(struct sock *sk, struct sk_buff *skb, const struct sadb_m
        if (proto == 0)
                return -EINVAL;
 
-       err = xfrm_state_flush(net, proto, true);
+       err = xfrm_state_flush(net, proto, true, false);
        err2 = unicast_flush_resp(sk, hdr);
        if (err || err2) {
                if (err == -ESRCH) /* empty table - go quietly */
index 23c9289..1bb971f 100644 (file)
@@ -432,7 +432,7 @@ void xfrm_state_free(struct xfrm_state *x)
 }
 EXPORT_SYMBOL(xfrm_state_free);
 
-static void xfrm_state_gc_destroy(struct xfrm_state *x)
+static void ___xfrm_state_destroy(struct xfrm_state *x)
 {
        tasklet_hrtimer_cancel(&x->mtimer);
        del_timer_sync(&x->rtimer);
@@ -474,7 +474,7 @@ static void xfrm_state_gc_task(struct work_struct *work)
        synchronize_rcu();
 
        hlist_for_each_entry_safe(x, tmp, &gc_list, gclist)
-               xfrm_state_gc_destroy(x);
+               ___xfrm_state_destroy(x);
 }
 
 static enum hrtimer_restart xfrm_timer_handler(struct hrtimer *me)
@@ -598,14 +598,19 @@ struct xfrm_state *xfrm_state_alloc(struct net *net)
 }
 EXPORT_SYMBOL(xfrm_state_alloc);
 
-void __xfrm_state_destroy(struct xfrm_state *x)
+void __xfrm_state_destroy(struct xfrm_state *x, bool sync)
 {
        WARN_ON(x->km.state != XFRM_STATE_DEAD);
 
-       spin_lock_bh(&xfrm_state_gc_lock);
-       hlist_add_head(&x->gclist, &xfrm_state_gc_list);
-       spin_unlock_bh(&xfrm_state_gc_lock);
-       schedule_work(&xfrm_state_gc_work);
+       if (sync) {
+               synchronize_rcu();
+               ___xfrm_state_destroy(x);
+       } else {
+               spin_lock_bh(&xfrm_state_gc_lock);
+               hlist_add_head(&x->gclist, &xfrm_state_gc_list);
+               spin_unlock_bh(&xfrm_state_gc_lock);
+               schedule_work(&xfrm_state_gc_work);
+       }
 }
 EXPORT_SYMBOL(__xfrm_state_destroy);
 
@@ -708,7 +713,7 @@ xfrm_dev_state_flush_secctx_check(struct net *net, struct net_device *dev, bool
 }
 #endif
 
-int xfrm_state_flush(struct net *net, u8 proto, bool task_valid)
+int xfrm_state_flush(struct net *net, u8 proto, bool task_valid, bool sync)
 {
        int i, err = 0, cnt = 0;
 
@@ -730,7 +735,10 @@ restart:
                                err = xfrm_state_delete(x);
                                xfrm_audit_state_delete(x, err ? 0 : 1,
                                                        task_valid);
-                               xfrm_state_put(x);
+                               if (sync)
+                                       xfrm_state_put_sync(x);
+                               else
+                                       xfrm_state_put(x);
                                if (!err)
                                        cnt++;
 
@@ -2215,7 +2223,7 @@ void xfrm_state_delete_tunnel(struct xfrm_state *x)
                if (atomic_read(&t->tunnel_users) == 2)
                        xfrm_state_delete(t);
                atomic_dec(&t->tunnel_users);
-               xfrm_state_put(t);
+               xfrm_state_put_sync(t);
                x->tunnel = NULL;
        }
 }
@@ -2375,8 +2383,8 @@ void xfrm_state_fini(struct net *net)
        unsigned int sz;
 
        flush_work(&net->xfrm.state_hash_work);
-       xfrm_state_flush(net, IPSEC_PROTO_ANY, false);
        flush_work(&xfrm_state_gc_work);
+       xfrm_state_flush(net, IPSEC_PROTO_ANY, false, true);
 
        WARN_ON(!list_empty(&net->xfrm.state_all));
 
index c6d26af..a131f9f 100644 (file)
@@ -1932,7 +1932,7 @@ static int xfrm_flush_sa(struct sk_buff *skb, struct nlmsghdr *nlh,
        struct xfrm_usersa_flush *p = nlmsg_data(nlh);
        int err;
 
-       err = xfrm_state_flush(net, p->proto, true);
+       err = xfrm_state_flush(net, p->proto, true, false);
        if (err) {
                if (err == -ESRCH) /* empty table */
                        return 0;