[Devel] Re: [PATCH net-2.6.26 2/6][NETNS][SOCK]: Introduce per-net inuse counters.

[Pavel Emelyanov]

Eric Dumazet wrote:
> Pavel Emelyanov a écrit :
>> This is probably the most controversial part of the set.
>>
>> The counters are stored in a per-cpu array on a struct net. To 
>> index in this array the prot->inuse is declared as int and used. 
>>
>> Numbers (indices) to protos are generated with the appropriate 
>> enum. I though about using some existing IPPROTO_XXX numbers for
>> protocols but they were too large (IPPROTO_RAW is 255) and did 
>> not differ for ipv4 and ipv6 (there's no IP6PROTO_RAW, etc).
>>
>> The sock_prot_inuse_(add|get) now use the net argument to
>> get the counter, but this all hides under CONFIG_NET_NS.
>>
>> The sock_prot_inuse_(init|fini) are no-ops. DEFINE_PROTO_INUSE
>> is empty and REF_PROTO_INUSE assigns an index to a proto.
>>
>>   
> 
> Given that :
> 
> 1) pcounter should really go away from kernel, since Andrew disagree 
> with the implementation.

Does this and ... (below)

> 2) the need to enumerate all protocols in your enum, it seems ... ugly :)

Yup :(

> 3) alloc_percpu(struct net_prot_inuse) per net is nice because we dont 
> waste memory (if we had to use percpu_counters for each proto for example)
> 
> I suggest to :
> 
> 1) not use pcounter anymore

... this mean that I can rework the inuse accounting in order not
to use pcounters at all even with CONFIG_NET_NS=n? :)

> 2) change 'inuse' field to 'inuse_idx' or 'prot_num' that is 
> automatically allocated at proto_register time, instead statically at 
> compile time.

Hm... I like this approach. Will do.

> Just provide a big enough NET_INUSE_NR (might depend on IPV6 present or 
> not, static or module) to take into account all possible protocols.

Well, I though about this, but wasn't sure whether such heuristics
would be accepted. 

> struct net_prot_inuse {
> 	int val[NET_INUSE_NR];
> };
> 
>