On 11/07/2018 06:05 PM, 배석진 wrote:
> --------- Original Message ---------
> Sender : Eric Dumazet <eric.duma...@gmail.com>
> Date : 2018-11-08 10:44 (GMT+9)
> Title : Re: [Kernel][NET] Bug report on packet defragmenting
>
>> On 11/07/2018 05:29 PM, 배석진 wrote:
>>
>>> If ipv6_defrag hook is not excuted simultaneously, then it's ok.
>>> ipv6_defrag hook can handle that. [exam 3]
>>
>> This seems wrong.
>>
>> This is the root cause, we should not try to work around it but fix it.
>>
>> There is no guarantee that RSS/RPS/RFS can help here, packets can sit in
>> per-cpu
>> backlogs long enough to reproduce the issue, if RX queues interrupts are
>> spread
>> over many cpus.
>
>
> Dear Dumazet,
>
> Even if rx irq be spread to overal cpu, hash will be made by src/des address.
> then they'll have a same hash and cpu. is it not enough?
> Did you mean that we need total solution for all steering method? not just
> only RPS?
>
IPv6 defrag unit should work all the times, even if 10 cpus have to feed
fragments for the same
datagram at the same time.
RPS is just a hint to spread packets on different cpus.
Basically we could have the following patch and everything must still work
properly
(presumably at lower performance, if RPS/RFS is any good, of course)
diff --git a/net/core/dev.c b/net/core/dev.c
index
0ffcbdd55fa9ee545c807f2ed3fc178830e3075a..c1269bb0d6c86b097cfff2d8395d8cbf2d596537
100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4036,7 +4036,7 @@ static int get_rps_cpu(struct net_device *dev, struct
sk_buff *skb,
goto done;
skb_reset_network_header(skb);
- hash = skb_get_hash(skb);
+ hash = prandom_u32();
if (!hash)
goto done;
Sure, it is better if RPS is smarter, but if there is a bug in IPv6 defrag unit
we must investigate and root-cause it.
