process hang in zfs->io_ ?

[Ganbold]

Peter Schuller wrote:
>> I'm running current as of 20080130 on an amd64 box. I can make
>> processes stuck in zfs->io_ (output truncated in ddb and top) when I
>> make some packages via ports tinderbox. The ports tinderbox access
>> local disk via nfs (I also tried nullfs).
>>
>> ddb output of ps and alltrace can be found at
>>
>> http://www.rafan.org/FreeBSD/zfs/textdump.zfs.20090130.txt
>>
>> Any ideas?
>>     
>
> A workaround is to disable the ZIL (vfs.zfs.zil_disable="1"), if you
> can afford that on the system in question. It will break the
> durability of fsync(), but retain it's write barrier semantics.
>
> Btw, does anyone have a good grasp of the status of this bug? I have
> seen vague referenced to it being a memory related deadlock for
> example, but that's about it. Is the cause known but difficult to fix,
> or just unknown?
>   

I've observed similar system hang when I tried to run svn update on
FreeBSD head repo
from 2 local directories simultaneously. I might be doing something
wrong and as
I remember I reproduced it a couple of times.

It is i386 CURRENT (r187663M) machine and loader.conf settings are:
vm.kmem_size="1024M"
vm.kmem_size_max="1024M"
vfs.zfs.arc_max="100M"

thanks,

Ganbold