On 6/27/2018 3:47 AM, Willem de Bruijn wrote:
>>>> +static int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
>>>> {
>>>> #ifdef CONFIG_XPS
>>>> struct xps_dev_maps *dev_maps;
>>>> - struct xps_map *map;
>>>> + struct sock *sk = skb->sk;
>>>> int queue_index = -1;
>>>>
>>>> if (!static_key_false(&xps_needed))
>>>> return -1;
>>>>
>>>> rcu_read_lock();
>>>> - dev_maps = rcu_dereference(dev->xps_cpus_map);
>>>> + if (!static_key_false(&xps_rxqs_needed))
>>>> + goto get_cpus_map;
>>>> +
>>>> + dev_maps = rcu_dereference(dev->xps_rxqs_map);
>>>> if (dev_maps) {
>>>> - unsigned int tci = skb->sender_cpu - 1;
>>>> + int tci = sk_rx_queue_get(sk);
>>>
>>> What if the rx device differs from the tx device?
>>>
>> I think I have 3 options here:
>> 1. Cache the ifindex in sock_common which will introduce a new
>> additional field in sock_common.
>> 2. Use dev_get_by_napi_id to get the device id. This could be expensive,
>> if the rxqs_map is set, this will be done on every packet and involves
>> walking through the hashlist for napi_id lookup.
>
> The tx queue mapping is cached in the sk for connected sockets, but
> indeed this would be expensive for many workloads.
>
>> 3. Remove validating device id, similar to how it is in skb_tx_hash
>> where rx_queue recorded is used and if not, fall through to flow hash
>> calculation.
>> What do you think is suitable here?
>
> Alternatively, just accept the misprediction in this rare case. But do
> make the caveat explicit in the documentation.
>
Okay, I will add this in the documentation.
