The receiving interface might have used GRO to receive more fragments than MAX_SKB_FRAGS fragments. In this case, these will not be stored in skb_shinfo(skb)->frags but merged into the frag list.
batman-adv relies on the function skb_split to split packets up into multiple smaller packets which are not larger than the MTU on the outgoing interface. But this function cannot handle frag_list entries and is only operating on skb_shinfo(skb)->frags. If it is then still trying to split such an skb and xmit'ing it on an interface without support for NETIF_F_FRAGLIST then validate_xmit_skb() will try to linearize it. But this fails due to inconsistent information and __pskb_pull_tail will trigger a BUG_ON after skb_copy_bits() returns an error.
In case of entries in frag_list, just linearize the skb before operating on it with skb_split().
Reported-by: Felix Kaechele felix@kaechele.ca Fixes: 9de347143505 ("batman-adv: layer2 unicast packet fragmentation") Signed-off-by: Sven Eckelmann sven@narfation.org --- net/batman-adv/fragmentation.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/batman-adv/fragmentation.c b/net/batman-adv/fragmentation.c index 0899a729..c120c7c6 100644 --- a/net/batman-adv/fragmentation.c +++ b/net/batman-adv/fragmentation.c @@ -475,6 +475,17 @@ int batadv_frag_send_packet(struct sk_buff *skb, goto free_skb; }
+ /* GRO might have added fragments to the fragment list instead of + * frags[]. But this is not handled by skb_split and must be + * linearized to avoid incorrect length information after all + * batman-adv fragments were created and submitted to the + * hard-interface + */ + if (skb_has_frag_list(skb) && __skb_linearize(skb)) { + ret = -ENOMEM; + goto free_skb; + } + /* Create one header to be copied to all fragments */ frag_header.packet_type = BATADV_UNICAST_FRAG; frag_header.version = BATADV_COMPAT_VERSION;