[Samba] Strange Bind freezing

[Nikita Druba]

Interesting fact:
My server, where work jail with Samba AD DC, have 40 CPUs(20 cores and 
40 threads). Therefore, when I starts named, it is creates 40 workers 
for every listen ip, i.e. 40 tcp and 40 udp for every ip.
Because its too much for my configuration, I intuitively made a decision 
to try to decrease number of named workers to 10 by "-n 10".
And all works without freezing with correct resolv.conf during last 4 days.
Where better to tell about this bug - to Bind or FreeBSD developers?



04.12.2021 9:22, Daniel O'Connor via samba пишет:
>
>> On 4 Dec 2021, at 17:51, Nikita Druba <admin at npo-lencor.ru> wrote:
>>
>> 02.12.2021 7:50, Daniel O'Connor пишет:
>>>> On 2 Dec 2021, at 16:16, Nikita Druba via samba <samba at lists.samba.org> wrote:
>>>>
>>>> I  forgot to add, that config of new DC jail, zfs, named and samba fully the same as old DC and very similar for several other my samba DCs. I tried to switch on Internal DNS and back. I also tried to disable all Bind options, that no refers in samba wiki. I do not understand, where else I can found some information, what here is wrong.
>>> I would try ktrace'ing the bind process, eg...
>>> sudo -u bind ktrace -f /tmp/named.ktr named -g <rest of options>
>>>
>>> Then reproduce and sift through the trace looking for bad things..
>>> sudo kdump -f /tmp/named.ktr
>>>
>>> Unfortunately ktrace is pretty low level (since it traces syscalls) but you might get a hint.
>>>
>> I tried to collect some logs by ktrace and catched freeze moment. After last from usual log(when Bind freezing), in kdump starts many times repeating the next records:
>>
>>   36460 named    CALL  nanosleep(0x7fffffffea30,0)
>>   36460 named    RET   nanosleep 0
>>
>> What can it means?  What do you thing - its a question of Bind or FreeBSD?
> That just means it called nanosleep system call which is what usleep etc are implemented as.
> Unfortunately ktrace is very low level so it can be hard to determine the context..
>
> Something like truss or strace would provide a bit more context although they have a higher performance impact (probably not an issue for you).
>
>
> --
> Daniel O'Connor
> "The nice thing about standards is that there
> are so many of them to choose from."
>   -- Andrew Tanenbaum
>
>