[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: 2.6.18 - Xen coredump problem



Am Montag, 11. Dezember 2006 12:32 schrieb Bastian Blank:
> Hi folks
>
> Our 2.6.18 i386 xen images sometimes dies on coredumps[1]. Fedora, which
> uses the same patch, does not have this problem. The main different is
> PAE, Fedora don't longer ships kernels without. I asked the Fedora Xen
> maintainer but did not get a response on this problem yet. I did a quick
> check with our kernel, and the problem did not show up with PAE enabled.
>
> We have the following possibilities:
> - Ignore it. Each i386 xen machine can be crashed by a simple program.
> - Switch the i386 image(s) to PAE.
>   - With ABI bump.
>   - Without ABI bump but new package names.

Hi Bastian & "folks",

what about a thrid possibility (but I don't know if it's possible at all):

 - Have the kernel decice at boot time if pae is needed or not. If the system 
has >3.5 gb of ram, then use pae, else use non-pae.

I am not sure if pae in general is really a good idea, because I got some 
problems with pae and xen on quite usual serverhardware in the past (xen 
3.0.0 & 3.0.1). For example the 3ware 9xxx controllers seems to have some 
ugly problems with 36bit wide adressspace (pae) and until I recieved a patch 
from Ian Pratt for the xen hypervisor, xen was not really useable.

Ian's words in this context: "PAE is a damn crook. Just a cheap hack... I hate 
it! Don't use it, if you don't need it."

If I remember it correctly, using pae also costs a bit of performance, so it 
would be not wise to use it if not really needed.

Bastian, have you asked the xen-devel ml if they have a clou why #399113 
happens? Maybe they can help.

> Bastian
>
> [1]: #399113



Reply to: