Re: [RFC PATCH 1/7] net: skbuff: Add tun_vnet_hash flag

[Akihiko Odaki <akihiko.odaki@daynix.com>]

On 2023/10/09 3:39, Willem de Bruijn wrote:
> On Sun, Oct 8, 2023 at 7:22 AM Akihiko Odaki <akihiko.odaki@daynix.com> wrote:
>>
>> tun_vnet_hash can use this flag to indicate it stored virtio-net hash
>> cache to cb.
>>
>> Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com>
>> ---
>>   include/linux/skbuff.h | 2 ++
>>   1 file changed, 2 insertions(+)
>>
>> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
>> index 4174c4b82d13..e638f157c13c 100644
>> --- a/include/linux/skbuff.h
>> +++ b/include/linux/skbuff.h
>> @@ -837,6 +837,7 @@ typedef unsigned char *sk_buff_data_t;
>>    *     @truesize: Buffer size
>>    *     @users: User count - see {datagram,tcp}.c
>>    *     @extensions: allocated extensions, valid if active_extensions is nonzero
>> + *     @tun_vnet_hash: tun stored virtio-net hash cache to cb
>>    */
>>
>>   struct sk_buff {
>> @@ -989,6 +990,7 @@ struct sk_buff {
>>   #if IS_ENABLED(CONFIG_IP_SCTP)
>>          __u8                    csum_not_inet:1;
>>   #endif
>> +       __u8                    tun_vnet_hash:1;
> 
> sk_buff space is very limited.
> 
> No need to extend it, especially for code that stays within a single
> subsystem (tun).
> 
> To a lesser extent the same point applies to the qdisc_skb_cb.

I had to extend sk_buff because it does not stay in tun but moves back 
and forth between qdisc and tun.

The new members of sk_buff and qdisc_skb_cb are stored by tun's 
ndo_select_queue(). The control will go back to qdisc after 
ndo_select_queue() function finishes. Eventually tun's ndo_start_xmit() 
will be called by qdisc and consumes the stored members. qdisc is 
required to keep the stored members intact.

tun_vnet_hash is a bit special. It is put into sk_buff because 
ndo_select_queue() is not always called and it may be left 
uninitialized. ndo_start_xmit() may read some garbage from cb's old user 
if it is put into cb.