Fixing page allocation failure

Fixing page allocation failure

[Morten W. Petersen]

[-- Attachment #1: Type: text/plain, Size: 685 bytes --]

Hi all,

I have a server that a couple of times each day squirts out messages 
about page allocation failures (python: page allocation failure. 
order:3, mode:0x20).  What's the reason for this, and could it affect 
the stability of the box?

The server that squirts these messages just crashed, for no apparent 
reason, so that's why I'm wondering.  It's a UML box.  Also, I'm 
wondering, are there any howto's for tweaking /proc settings so that the 
machine becomes more stable?  Are there any settings for increasing the 
verbosity of the kernel log so that the reason for a server crashing is 
easier to find?

Thanks in advance, and please CC me any replies :)

Regards,

Morten

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 256 bytes --]

Re: Fixing page allocation failure

[Nick Piggin]

Morten W. Petersen wrote:
> Hi all,
> 
> I have a server that a couple of times each day squirts out messages 
> about page allocation failures (python: page allocation failure. 
> order:3, mode:0x20).  What's the reason for this, and could it affect 
> the stability of the box?
> 
> The server that squirts these messages just crashed, for no apparent 
> reason, so that's why I'm wondering.  It's a UML box.  Also, I'm 
> wondering, are there any howto's for tweaking /proc settings so that the 
> machine becomes more stable?  Are there any settings for increasing the 
> verbosity of the kernel log so that the reason for a server crashing is 
> easier to find?
> 
> Thanks in advance, and please CC me any replies :)
> 
> Regards,
> 
> Morten

Yeah you can increase /proc/sys/vm/min_free_kbytes to help the problem.

There is a possibly better solution in the -mm kernels which should get
merged into 2.6 sooner or later - but there is still no way to guarantee
safety from allocation failures. If they cause crashes then that is a
bug, so report them here with traces and a description of the workload
and system hardware, etc.

Thanks
Nick

Re: Fixing page allocation failure

[Andrew Morton]

"Morten W. Petersen" <morten@nidelven-it.no> wrote:
>
> I have a server that a couple of times each day squirts out messages 
>  about page allocation failures (python: page allocation failure. 
>  order:3, mode:0x20).

It should also emit a stack backtrace.   If so, please send it.

Re: Fixing page allocation failure

[Morten W. Petersen]

[-- Attachment #1.1: Type: text/plain, Size: 556 bytes --]

Andrew Morton wrote:
> It should also emit a stack backtrace.   If so, please send it.

OK, I've attached the kern.log file, which should explain the setup and 
has some call traces.  The server has about 300-350 MBs of RAM, 500 MBs 
of swap, and typically uses about 2-300MBs of swap at any given time.

It's recorded load average is typically between 0.1 and 0.2, at times up 
to 1.0.

I've got some additional kern.log files if you're interested.

HTH,

Morten

PS:  I this is the second copy of this message, the first one I sent 
didn't come through.

[-- Attachment #1.2: kern.log.gz --]
[-- Type: application/gzip, Size: 5334 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 256 bytes --]

Re: Fixing page allocation failure

[Andrew Morton]

"Morten W. Petersen" <morten@nidelven-it.no> wrote:
>
> Andrew Morton wrote:
>  > It should also emit a stack backtrace.   If so, please send it.
> 
>  OK, I've attached the kern.log file

That's just a bunch of hex numbers.  Please enable CONFIG_KALLSYMS and
resend.