On 07/20/2018 10:30 AM, Florian Westphal wrote:
> Got crash report with following backtrace:
> BUG: unable to handle kernel paging request at ffff8801869daffe
> RIP: 0010:[<ffffffff816429c4>] [<ffffffff816429c4>]
> ip6_finish_output2+0x394/0x4c0
> RSP: 0018:ffff880186c83a98 EFLAGS: 00010283
> RAX: ffff8801869db00e ...
> [<ffffffff81644cdc>] ip6_finish_output+0x8c/0xf0
> [<ffffffff81644d97>] ip6_output+0x57/0x100
> [<ffffffff81643dc9>] ip6_forward+0x4b9/0x840
> [<ffffffff81645566>] ip6_rcv_finish+0x66/0xc0
> [<ffffffff81645db9>] ipv6_rcv+0x319/0x530
> [<ffffffff815892ac>] netif_receive_skb+0x1c/0x70
> [<ffffffffc0060bec>] atl1c_clean+0x1ec/0x310 [atl1c]
> ...
>
> The bad access is in neigh_hh_output(), at skb->data - 16 (HH_DATA_MOD).
> atl1c driver provided skb with no headroom, so 14 bytes (ethernet
> header) got pulled, but then 16 are copied.
>
> Reserve NET_SKB_PAD bytes headroom, like netdev_alloc_skb().
>
> Compile tested only; I lack hardware.
>
> Fixes: 7b7017642199 ("atl1c: Fix misuse of netdev_alloc_skb in refilling rx
> ring")
> Signed-off-by: Florian Westphal <f...@strlen.de>
> ---
> diff --git a/drivers/net/ethernet/atheros/atl1c/atl1c_main.c
> b/drivers/net/ethernet/atheros/atl1c/atl1c_main.c
> index 94270f654b3b..7087b88550db 100644
> --- a/drivers/net/ethernet/atheros/atl1c/atl1c_main.c
> +++ b/drivers/net/ethernet/atheros/atl1c/atl1c_main.c
> @@ -1686,6 +1686,7 @@ static struct sk_buff *atl1c_alloc_skb(struct
> atl1c_adapter *adapter)
> skb = build_skb(page_address(page) + adapter->rx_page_offset,
> adapter->rx_frag_size);
> if (likely(skb)) {
> + skb_reserve(skb, NET_SKB_PAD);
> adapter->rx_page_offset += adapter->rx_frag_size;
> if (adapter->rx_page_offset >= PAGE_SIZE)
> adapter->rx_page = NULL;
>
Yes, it is interesting IPv4 has code to deal with that( in ip_finish_output2()),
not IPv6 :/
Reviewed-by: Eric Dumazet <eduma...@google.com>
