[ovs-dev] [PATCH] datapath: Update inner offsets when expanding headroom.

[Jesse Gross]

On Tue, Mar 31, 2015 at 11:02 AM, Pravin Shelar <pshelar at nicira.com> wrote:
> On Tue, Mar 31, 2015 at 9:34 AM, Jesse Gross <jesse at nicira.com> wrote:
>> skb protocol offsets are relative to the beginning of the
>> buffer and therefore must be updated if the buffer size is
>> expanded. Kernel functions do this automatically for existing
>> fields but obviously not for anything that we backport. This
>> introduces a wrapper for pskb_expand_head() to update the
>> inner protocol fields that we have backported.
>>
>> Without this, a kernel crash can be triggered with tunnel
>> packets that do not have enough headroom and need to be
>> segmented.
>>
> which pskb_expand head actually causes this issue for vxlan?

This is the result of skb_cow_head() in vxlan_xmit_skb().

>> diff --git a/datapath/linux/compat/gso.h b/datapath/linux/compat/gso.h
>> index 337d13a..ba94eae 100644
>> --- a/datapath/linux/compat/gso.h
>> +++ b/datapath/linux/compat/gso.h
>> @@ -125,6 +125,27 @@ static inline void skb_reset_inner_headers(struct sk_buff *skb)
>>         skb_set_inner_network_header(skb, skb_network_offset(skb));
>>         skb_set_inner_transport_header(skb, skb_transport_offset(skb));
>>  }
>> +
>> +static inline int rpl_pskb_expand_head(struct sk_buff *skb, int nhead,
>> +                                      int ntail, gfp_t gfp_mask)
>> +{
>> +       int err;
>> +
>> +       err = pskb_expand_head(skb, nhead, ntail, gfp_mask);
>> +       if (err)
>> +               return err;
>> +
>> +#if LINUX_VERSION_CODE < KERNEL_VERSION(3,10,0)
>> +       OVS_GSO_CB(skb)->inner_mac_header += nhead;
>> +#endif
>> +#if LINUX_VERSION_CODE < KERNEL_VERSION(3,8,0)
>> +       OVS_GSO_CB(skb)->inner_network_header += nhead;
>> +#endif
>> +
> Can you add API for this to gso.h so that we can use existing kernel
> version checks in gso.h.

OK, I'll send a v2.