On 3/31/23 15:06, Eelco Chaudron wrote:
>
>
> On 31 Mar 2023, at 12:38, Simon Horman wrote:
>
>> On Fri, Mar 31, 2023 at 12:05:09PM +0200, Ilya Maximets wrote:
>>> On 3/31/23 11:07, Simon Horman wrote:
>>>> On Thu, Mar 30, 2023 at 09:04:02PM +0200, Ilya Maximets wrote:
>>>>> On 3/30/23 11:45, Simon Horman wrote:
>>>>>> On Fri, Mar 17, 2023 at 09:47:36PM +0100, Eelco Chaudron wrote:
>>>>>>>> Op 17 mrt. 2023 om 21:11 heeft Marcelo Ricardo Leitner
>>>>>>>> <mleit...@redhat.com> het volgende geschreven:
>>>>>>>> Thu, Mar 16, 2023 at 09:51:34AM +0100, Eelco Chaudron wrote:
>>>>>>>>>> On 22 Dec 2022, at 13:32, Eelco Chaudron wrote:
>>>>>>>>>>> On 22 Dec 2022, at 10:26, Balazs Nemeth wrote:
>>>>>>>>>>
>>>>>>>>>>> The only way that stats->{n_packets,n_bytes} would decrease is due
>>>>>>>>>>> to an
>>>>>>
>>>>>> nit: s/way/ways/
>>
>>>>>>>>>>> overflow, or if there are bugs in how statistics are handled. In the
>>>>>>>>>>> past, there were multiple bugs that caused a jump backward. A
>>>>>>>>>>> workaround was in place to set the statistics to 0 in that case.
>>>>>>>>>>> When
>>>>>>>>>>> this happened while the revalidator was under heavy load, the
>>>>>>>>>>> workaround
>>>>>>>>>>> had an unintended side effect where should_revalidate returned false
>>>>>>>>>>> causing the flow to be removed because the metric it calculated was
>>>>>>>>>>> based on a bogus value. Since many of those bugs have now been
>>>>>>>>>>> identified and resolved, there is no need to set the statistics to
>>>>>>>>>>> 0. In
>>>>>>>>>>> addition, the (unlikely) overflow still needs to be handled
>>>>>>>>>>> appropriately.
>>>>>>
>>>>>> 1. Perhaps it would be prudent to log/count if/when this occurs
>>>>>
>>>>> +1
>>>>> We do have a coverage counter that will indicate the case where stats
>>>>> jump back. However, if we're certain that this should never happen,
>>>>> we should, probably, emit a warning or even an error log as well, so
>>>>> users are aware that something went wrong.
>>>>
>>>> I was thinking more of a counter, which seems to already be covered.
>>>> But I have no objection to your reasoning about having a warning (too).
>>>>
>>>>>
>>>>>> 2. I take it that the overflow handling would be follow-up work,
>>>>>> is that correct?
>>>>>
>>>>> The unsigned arithmetic should take case of overflowed counters,
>>>>> because the result of subtraction will still give a correct difference
>>>>> between the old and a new value, even if it overflowed and the new
>>>>> value is smaller. Unless, of course, it overflowed more than once.
>>>>
>>>> More than once between samples?
>>>> If so, I'm assuming that is not a case we can hit unless there is a bug.
>>>
>>> Right. It's actually should be practically not possible to overflow
>>> even once with a current hardware. Assuming we have a fancy 400 Gbps
>>> NIC, then it should take 11.7 years to overflow a byte counter.
>>>
>>> So, this patch is mostly removing a workaround for some bug that we
>>> hope we fixed. But it's not clear what the original bug was as the
>>> commit message for this workaround didn't specify a root cause. So,
>>> it's hard to say if it's fixed or not. And that's why I'm thinking
>>> that the error message is needed.
>>
>> Yes, I agree that is prudent.
>
> If we do add a log message, we should be careful as it could still be a wrap
> (for bytes). For packets, it’s not very likely with the current speeds 400G,
> will be around 980 years… For bytes, we can wrap easily.
>
> So I would suggest it only for packet count…
>
We already only use a packet counter for the 'ukey_invalid_stat_reset' coverage.
So, I suppose, we can just add the log under the same condition.
OTOH, Don't we check that the difference is within 3/4 of 64-bit range?
It should still take many years to overflow the byte counter.
Best regards, Ilya Maximets.
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
