vm_thread_new: kstack allocation failed with many ZFS FS and NFSD

[Kris Kennaway]

Thomas Vogt wrote:
> Hi Kris
> 
> Am 11.03.2008 um 01:30 schrieb Kris Kennaway:
>> Thomas Vogt wrote:
>>> Hi List(s)
>>> I try to simulate real workload for our environment in my lab. The idea
>>> was to create 10k+ ZFS fs with several thousand files on each fs and
>>> then measure daily workload performance. Maybe 10k fs sounds silly but
>>> if you need individual quota for every user on a system, 5-10k fs are
>>> not unusual for ZFS
>>> My script to cerate zfs fs
>>> #!/bin/sh
>>> i=0; while [ $i != 10000 ]; do zfs create tank/script$i; i=`expr $i +
>>> 1`; done
>>> My script stopped after creating ~4850 FS with:
>>> vm_thread_new: kstack allocation failed
>>> vm_thread_new: kstack allocation failed
>>> vm_thread_new: kstack allocation failed
>>> vm_thread_new: kstack allocation failed
>>> vm_thread_new: kstack allocation failed
>>> vm_thread_new: kstack allocation failed
>>
>> Your kernel has run out of memory.  If you cannot tune kmem_size 
>> further then it cannot handle this many ZFS filesystems.
> 
> Are there no limitation for vm.kmem_size* sysctls? I tried to increase 
> vm.kmem_size* with larger values than 1500M but the system paniced in 
> the boot process.

Yes, there is an upper bound somewhere around this point with the 
default kernel layout.

> Mark Tinguely told me maybe i can edit sys/amd64/include/pmap.h and 
> change the line:
> 
> - #define KPDPI         (NPDPEPG-2)    /* kernbase at -2GB */
> + #define KPDPI         (NPDPEPG-4)    /* kernbase at -4GB */
> 
> I will try this. Any idea if this is save?

I don't know, sorry.

Kris