On Thu, Oct 8, 2015 at 7:35 PM, Alexei Starovoitov <a...@plumgrid.com> wrote:
> On 10/8/15 5:50 PM, Devon H. O'Dell wrote:
>>>
>>> with the amount of skb_sender_cpu_clear() all over the code base
>>> >I wonder whether there is a better solution to all of these.
>>
>> I think there is. We found that splitting the union of sender_cpu and
>> napi_id solved the issue for us. In general, I think this is an OK
>> solution as long as the following hold:
>>
>> * skbs are always allocated via kzalloc
>> * out -> out cloned skbs are always cloned on the same CPU
>> * an extra four bytes in skbuff isn't a bad thing
>
>
> I'm pretty sure extending sk_buff for this is not acceptable.
That's unfortunate.
> I was thinking may be we can use sign bit to distinguish between
> napi_id and sender_cpu.
> Like:
> if ((int)skb->sender_cpu >= 0)
> skb->sender_cpu = - (raw_smp_processor_id() + 1);
> and inside get_xps_queue() use it only if it's negative.
I like the idea, but it seems unnecessarily magical. What about using
a bitfield? Then there's just an option bit that is either
OPTION_NAPI_ID or OPTION_SENDER_CPU. Then the check to set sender_cpu
in netdev_pick_tx becomes
if (skb->sender_napi_option == OPTION_NAPI_ID || skb->sender_cpu == 0) ...
> Then we can remove skb_sender_cpu_clear() from everywhere.
> Adding a check to napi_hash_add() to make sure that napi_id is not
> negative is probably ok too.
We could change this to check that sender_napi_option would be
OPTION_NAPI_ID with the bitfield idea.
My names are probably bad, but I think the idea is less magical (and
is effectively the same thing you are proposing).
> Thoughts?
--dho
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
