On Mon, Nov 26, 2018 at 9:27 PM Neil Horman [off-list ref] wrote:

On Mon, Nov 26, 2018 at 02:52:44PM +0800, Xin Long wrote:

quoted

I changed to count sk_wmem_alloc by skb truesize instead of 1 to
fix the sk_wmem_alloc leak caused by later truesize's change in
xfrm in Commit 02968ccf0125 ("sctp: count sk_wmem_alloc by skb
truesize in sctp_packet_transmit").

But I should have also increased sk_wmem_alloc when head->truesize
is increased in sctp_packet_gso_append() as xfrm does. Otherwise,
sctp gso packet will cause sk_wmem_alloc underflow.

Fixes: 02968ccf0125 ("sctp: count sk_wmem_alloc by skb truesize in sctp_packet_transmit")
Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 net/sctp/output.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/net/sctp/output.c b/net/sctp/output.c
index b0e74a3..025f48e 100644
--- a/net/sctp/output.c
+++ b/net/sctp/output.c

@@ -410,6 +410,7 @@ static void sctp_packet_gso_append(struct sk_buff *head, struct sk_buff *skb)
      head->truesize += skb->truesize;
      head->data_len += skb->len;
      head->len += skb->len;
+     refcount_add(skb->truesize, &head->sk->sk_wmem_alloc);

      __skb_header_release(skb);
 }

This looks to me like you are now double counting every packet that passes
through sctp_packet_transmit, once in skb_set_owner_w and again in
sctp_packet_pack=>sctp_packet_gso_append

For gso packet, the skb(head) in sctp_packet_transmit/skb_set_owner_w
is the head_skb, the skbs here are the fragments.

For non-gso packet, it will never come here.

Neil

quoted

--
2.1.0