Re: [PATCH v9 17/20] ebpf: Fix RSS error handling

[Akihiko Odaki <akihiko.odaki@daynix.com>]

On 2024/04/13 21:16, Yuri Benditovich wrote:
> On Wed, Apr 3, 2024 at 2:12 PM Akihiko Odaki <akihiko.odaki@daynix.com> wrote:
>>
>> calculate_rss_hash() was using hash value 0 to tell if it calculated
>> a hash, but the hash value may be 0 on a rare occasion. Have a
>> distinct bool value for correctness.
> 
> This is interesting question whether in reality the hash value might
> be 0 or not.
> On one hand - this seems like a kind of fix.
> On another hard - this adds computation cycles for each packet, and the
> corner case that this targets to fix seems hardly reachable if at all.
> Optimistic estimation is 2.5*10^-8 percent of address:address:port triplets.
> I would suggest at least to find some proof of the fact that the calculated
> hash might be 0 in real case where source addresses are not random.

Your estimation, which suggests the low probability of error, and the 
fact that an error in RSS is not fatal and only regresses the 
performance, should be sufficient to determine how this patch should be 
handled; this patch is nice to have, but is not worth to backport to 
stable or to merge before the 9.0 release.

Regards,
Akihiko Odaki