Le 20/12/2017 à 23:15, teor a écrit :
> 
>> On 21 Dec 2017, at 08:51, teor <[email protected]> wrote:
>>
>>>
>>> 1) Why didn't we see this abuse wave coming ? We kept replying to reporters 
>>> of the dreaded "Failing because we have XXX connections already. Please 
>>> read doc/TUNING for guidance" about how they could amend their config to 
>>> accept more connections. Although the 'global scale' of those events should 
>>> have been detected, without most of use assuming it was due to nodes' bad 
>>> config.
>>
>> Load spikes are normal, particularly with the HSDir flag, because HSDir
>> usage is not bandwidth-weighted.
>>
>> Allowing more connections *is* the right thing to do with this attack,
>> if your OS has the resources. Several of my relays never went down,
>> because they were over-provisioned with RAM and CPU.
>>
>> Others only went down temporarily, during the most intense phases.
>> (And then their excessive bandwidth weight was redistributed, and they
>> have been coping well.)
>>
>> If you don't have the resources to handle that many connections, then
>> limiting connections is the right thing to do. If you can't do it
>> using tor, then a firewall is the way to go.

This has been put in place and relay is now able to sustain the still
ongoing flood.

>>
>> (There are some bugs in Tor that make the attack more effective than
>> it should be. We're working on fixing them.)
> 
> To mitigate this attack, we recommend setting MaxMemInQueues to the amount
> of RAM you have available per tor instance (or maybe a few hundred MB less).
> 
> Tor estimates it, but the estimate isn't very good.
> 

This has been added about 12 hours ago (and relay SIGHUPed) and I still
cannot see any trace of circuit OOM kills in relay logs.

And the 2 most recent heartbeat reports show a 'normal' circuit count

Thanks for all the fish :)


-- 
Frédéric CORNU
http://wardsback.org
_______________________________________________
tor-relays mailing list
[email protected]
https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays

Reply via email to