Re: [PATCH net-next v2] rps: selective flow shedding during softnet overflow

[Willem de Bruijn <willemb@google.com>]

On Mon, Apr 22, 2013 at 6:30 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> On Mon, 2013-04-22 at 16:46 -0400, Willem de Bruijn wrote:
>
>> +             len = sizeof(*cur) + netdev_flow_limit_table_len;
> [1]
>> +             mutex_lock(&flow_limit_update_mutex);
> [2]
>> +             for_each_possible_cpu(i) {
>> +                     sd = &per_cpu(softnet_data, i);
>> +                     cur = rcu_dereference_protected(sd->flow_limit,
>> +                                   lockdep_is_held(flow_limit_update_mutex));
>> +                     if (cur && !cpumask_test_cpu(i, mask)) {
>> +                             RCU_INIT_POINTER(sd->flow_limit, NULL);
>> +                             synchronize_rcu();
>> +                             kfree(cur);
>> +                     } else if (!cur && cpumask_test_cpu(i, mask)) {
>> +                             cur = kzalloc(len, GFP_KERNEL);
>> +                             cur->num_buckets = netdev_flow_limit_table_len;
> [3]
>
> Its a bit tricky, but the value of netdev_flow_limit_table_len could
> change between [1] and [3]
>
> So you should read its value once, or protect the whole thing using
> mutex_lock(&flow_limit_update_mutex) in sysctl code ( and move [1] after
> [2])

Thanks for the detailed explanation. I implemented the second
solution: make writes to the two sysctls mutually exclusive.
>
>