Re: [PATCH net-next v2 2/2] unix: Show number of pending scm files of receive queue in fdinfo

[Kirill Tkhai <ktkhai@virtuozzo.com>]

Hi,

On 24.02.2020 13:15, Paolo Abeni wrote:
> Hi,
> 
> On Mon, 2019-12-09 at 10:03 +0000, Kirill Tkhai wrote:
>> diff --git a/include/net/af_unix.h b/include/net/af_unix.h
>> index 3426d6dacc45..17e10fba2152 100644
>> --- a/include/net/af_unix.h
>> +++ b/include/net/af_unix.h
>> @@ -41,6 +41,10 @@ struct unix_skb_parms {
>>  	u32			consumed;
>>  } __randomize_layout;
>>  
>> +struct scm_stat {
>> +	u32 nr_fds;
>> +};
>> +
> 
> I'd like to drop the 'destructor' argument from
> __skb_try_recv_datagram() and friends - that will both clean-up the
> datagram code a bit and will avoid an indirect call in fast-path.
> 
> unix_dgram_recvmsg() needs special care: with the proposed change
> scm_stat_del() will be called explicitly after _skb_try_recv_datagram()
> while 'nr_fds' must to be updated under the receive queue lock.
> 
> Any of the following should work:
> - change 'nr_fds' to an atomic type, and drop all lockdep stuff
> - acquire again the receive queue spinlock before calling
> scm_stat_del(), ev doing that only 'if UNIXCB(skb).fp'
> - open code a variant of __skb_try_recv_datagram() which will take care
> of scm_stat_del() under the receive queue lock.
> 
> Do you have any preferences? If you don't plan to add more fields to
> 'struct scm_stat' I would go for the first option.

The first option looks the best in my opinion.

Kirill