Re: panic when booting HEAD on i386

On Sun, Apr 21, 2013 at 7:27 PM, Konstantin Belousov <kostikbel_at_gmail.com>wrote:

> On Sun, Apr 21, 2013 at 01:57:03PM +0800, Ganbold Tsagaankhuu wrote:
> > On Sat, Apr 20, 2013 at 11:21 PM, Konstantin Belousov
> > <kostikbel_at_gmail.com>wrote:
> >
> > > On Sat, Apr 20, 2013 at 10:52:35PM +0800, Ganbold Tsagaankhuu wrote:
> > > > Hi,
> > > >
> > > > I'm trying to boot HEAD after updating, but unfortunately it panics
> with
> > > > following message:
> > > >
> > > > panic: kmem_suballoc: bad status return of 3.
> > > >
> > > > I was only able to get image of the panic.
> > > >
> > > > http://www.mnbsd.org/ganbold/IMG_20130420_222353-2.jpg
> > > >
> > > > Does anybody see same panic booting on i386 lately?
> > > > Tried clang, and it panics also.
> > >
> > > How much memory do you have ?
> > > Do you have any tunables in the loader.conf ?
> > >
> >
> > Following settings caused the panic:
> >
> > vm.kmem_size="999M"
> > vm.kmem_size_max="999M"
> I cannot imagine how this could work earlier as well, except by chance.
>

Right, probably by chance it didn't cause panic in previous version.


>
> >
> > Whether it is regression or not in previous version of kernel (r244046)
> it
> > didn't panic on boot.
> > Seems no information on UPDATING either.
> Do you propose to enumerate all non-working or panic-provoking settings in
> loader.conf ?
>

Not really.


>
> Why did you set this value at all ?
>

These are the settings that I did for zfs, since my machine has zfs and
probably it is not even worth to run zfs on 2GB system.
It is my fault.


>
> KVA on i386 is limited to slightly less then 1GB, where all the kernel
> maps must be instantiated. Kernel uses the value of the tunable
> vm.kmem_size literally, except it makes a mild attempt to prevent
> foot-shooting by capping kmem_size to 2 * physical memory size.
>
> You neglected to answer how much memory is installed on your machine,
>


As mentioned in one of my previous email I have 2GB machine which is
probably not worth to have zfs on such system.
Sorry for the noise.

thanks,

Ganbold




> but I suspect you have enough so that overblown kmem_map cannot coexists
> with other kernel VA consumers.
>