openvswitch: fix skb_panic due to the incorrect actions attrlen
authorLiping Zhang <zlpnobody@gmail.com>
Wed, 16 Aug 2017 05:30:07 +0000 (13:30 +0800)
committerDavid S. Miller <davem@davemloft.net>
Wed, 16 Aug 2017 21:12:37 +0000 (14:12 -0700)
For sw_flow_actions, the actions_len only represents the kernel part's
size, and when we dump the actions to the userspace, we will do the
convertions, so it's true size may become bigger than the actions_len.

But unfortunately, for OVS_PACKET_ATTR_ACTIONS, we use the actions_len
to alloc the skbuff, so the user_skb's size may become insufficient and
oops will happen like this:
  skbuff: skb_over_panic: text:ffffffff8148fabf len:1749 put:157 head:
  ffff881300f39000 data:ffff881300f39000 tail:0x6d5 end:0x6c0 dev:<NULL>
  ------------[ cut here ]------------
  kernel BUG at net/core/skbuff.c:129!
  [...]
  Call Trace:
   <IRQ>
   [<ffffffff8148be82>] skb_put+0x43/0x44
   [<ffffffff8148fabf>] skb_zerocopy+0x6c/0x1f4
   [<ffffffffa0290d36>] queue_userspace_packet+0x3a3/0x448 [openvswitch]
   [<ffffffffa0292023>] ovs_dp_upcall+0x30/0x5c [openvswitch]
   [<ffffffffa028d435>] output_userspace+0x132/0x158 [openvswitch]
   [<ffffffffa01e6890>] ? ip6_rcv_finish+0x74/0x77 [ipv6]
   [<ffffffffa028e277>] do_execute_actions+0xcc1/0xdc8 [openvswitch]
   [<ffffffffa028e3f2>] ovs_execute_actions+0x74/0x106 [openvswitch]
   [<ffffffffa0292130>] ovs_dp_process_packet+0xe1/0xfd [openvswitch]
   [<ffffffffa0292b77>] ? key_extract+0x63c/0x8d5 [openvswitch]
   [<ffffffffa029848b>] ovs_vport_receive+0xa1/0xc3 [openvswitch]
  [...]

Also we can find that the actions_len is much little than the orig_len:
  crash> struct sw_flow_actions 0xffff8812f539d000
  struct sw_flow_actions {
    rcu = {
      next = 0xffff8812f5398800,
      func = 0xffffe3b00035db32
    },
    orig_len = 1384,
    actions_len = 592,
    actions = 0xffff8812f539d01c
  }

So as a quick fix, use the orig_len instead of the actions_len to alloc
the user_skb.

Last, this oops happened on our system running a relative old kernel, but
the same risk still exists on the mainline, since we use the wrong
actions_len from the beginning.

Fixes: ccea74457bbd ("openvswitch: include datapath actions with sampled-packet upcall to userspace")
Cc: Neil McKee <neil.mckee@inmon.com>
Signed-off-by: Liping Zhang <zlpnobody@gmail.com>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/openvswitch/actions.c
net/openvswitch/datapath.c
net/openvswitch/datapath.h

index e4610676299bcdac626db1a30cd4da44ccc62c0b..a54a556fcdb57d95b4a4a6606016ead527b93d64 100644 (file)
@@ -1337,6 +1337,7 @@ int ovs_execute_actions(struct datapath *dp, struct sk_buff *skb,
                goto out;
        }
 
+       OVS_CB(skb)->acts_origlen = acts->orig_len;
        err = do_execute_actions(dp, skb, key,
                                 acts->actions, acts->actions_len);
 
index 45fe8c8a884df36100bc0cf26d279cb92e51bdf8..6b44fe4052825a87b373bafb58ca014e3ec99015 100644 (file)
@@ -381,7 +381,7 @@ static int queue_gso_packets(struct datapath *dp, struct sk_buff *skb,
 }
 
 static size_t upcall_msg_size(const struct dp_upcall_info *upcall_info,
-                             unsigned int hdrlen)
+                             unsigned int hdrlen, int actions_attrlen)
 {
        size_t size = NLMSG_ALIGN(sizeof(struct ovs_header))
                + nla_total_size(hdrlen) /* OVS_PACKET_ATTR_PACKET */
@@ -398,7 +398,7 @@ static size_t upcall_msg_size(const struct dp_upcall_info *upcall_info,
 
        /* OVS_PACKET_ATTR_ACTIONS */
        if (upcall_info->actions_len)
-               size += nla_total_size(upcall_info->actions_len);
+               size += nla_total_size(actions_attrlen);
 
        /* OVS_PACKET_ATTR_MRU */
        if (upcall_info->mru)
@@ -465,7 +465,8 @@ static int queue_userspace_packet(struct datapath *dp, struct sk_buff *skb,
        else
                hlen = skb->len;
 
-       len = upcall_msg_size(upcall_info, hlen - cutlen);
+       len = upcall_msg_size(upcall_info, hlen - cutlen,
+                             OVS_CB(skb)->acts_origlen);
        user_skb = genlmsg_new(len, GFP_ATOMIC);
        if (!user_skb) {
                err = -ENOMEM;
index 5d8dcd88815f0622a48a54f14eaf5b5b9c53b18c..480600649d0b03a185f73b587181a3136d75e9b2 100644 (file)
@@ -99,11 +99,13 @@ struct datapath {
  * when a packet is received by OVS.
  * @mru: The maximum received fragement size; 0 if the packet is not
  * fragmented.
+ * @acts_origlen: The netlink size of the flow actions applied to this skb.
  * @cutlen: The number of bytes from the packet end to be removed.
  */
 struct ovs_skb_cb {
        struct vport            *input_vport;
        u16                     mru;
+       u16                     acts_origlen;
        u32                     cutlen;
 };
 #define OVS_CB(skb) ((struct ovs_skb_cb *)(skb)->cb)