Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 27 Jun 2012 21:35:20 -0400
From:      "Dieter BSD" <dieterbsd@engineer.com>
To:        To: freebsd-hackers@freebsd.org,freebsd-geom@freebsd.org
Subject:   Re: Freeze when running freebsd-update
Message-ID:  <20120628013522.298400@gmx.com>

next in thread | raw e-mail | index | archive | help
>>>> Robert writes:
>>>>> 3) the box is responsive to hitting enter at the console (it produces
>>>>> another login: prompt)
>>>>
>>>> Getty is in memory and can run.
>>>>
>>>>> 5) if I try to login to the console, it lets me enter a username then
>>>>> locks up totally, it does not present me with a password: prompt.
>>>>
>>>> Login(1) is not in memory, and the kernel cannot read it from disk
>>>> for some reason.
>>>>
>>>> I can get this symptom by writing a large file to a disk on a
>>>> controller that FreeBSD doesn't support NCQ on. I assume there
>>>> is a logjam in the buffer cache. Something trivial like reading
>>>> login in from disk that would normally happen in well under a
>>>> second can take many minutes.
>>>>
>>>> Perhaps geli is causing a similar logjam? Does it hang forever or
>>>> is it just obscenely slow? If it truely hangs forever it is
>>>> probably something else. Is there disk activity after it hangs?
>>>> Can you try it without geli? systat -vmstat might provide a clue.
>>>
>>> Well, it is geli. I'm unable to reproduce the freeze on the same
>>> exact system with everything else the same except for no geli. I'm
>>> going to move this thread over to geom, and continue it there. Thanks
>>> for your help!
>>
>> It occurs to me that it will need twice as much memory for disk i/o.
>> 1 buffer for encrypted and 1 for unencrypted. I know nothing about geli,
>> so I don't know if it uses the buffer cache for both, or what.
>> Could it be that the kernel isn't keeping enough memory free and
>> manages to paint itself into a corner and not have space to store
>> the unencrypted version of disk reads, and can't page/swap anything
>> out to make space because it doesn't have space to store the encrypted
>> version to write?
>
> I think that's probably about what is happening. I'm still waiting
> for an answer on the geom mailing list, but I will do some testing
> with increasing memory sizes and see where the problem stops
> occurring.

Some of the vfs.*buf sysctls might be useful?



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20120628013522.298400>