running into OOM killer with kexec loading large ramdisk

[Philipp Rudo <prudo@redhat.com>]

Hi Tobias,

On Wed, 23 Mar 2022 12:14:59 +0100
Tobias Powalowski <tobias.powalowski@googlemail.com> wrote:

> Hi,
> again me,
> I try to load a 900MB ramdisk on a ramfs rootfs  with kexec and a 10MB kernel.
> With 2800MB RAM assigned to qemu.
> Memory free by /proc/memstat: 2.2GB
> It keeps on OOM killed while executing:
> kexec -l kernel --initrd=initrd.img
> I can safely unpack the initrd in the ramfs without getting OOM killed.
> What is kexec doing wrong here?

I don't think that kexec is doing anything wrong here.

The kexec_load syscall is designed in a way that user space prepares
everything in a huge buffer and passes a pointer to it to the system
call. The systemcall then needs to copy everything from the user buffer
to a kernel buffer. So there are three copies of the initrd in memory
(including the one in ramfs). Together they take up 3 * 900MB = 2.7GB.
So basically all the memory of your guest.

You can try using the kexec_file_load systemcall. That at least
eliminates the copy in the user buffer so there are only two copies
left. 400MB of free memory is still quite scarce, though.

Thanks
Philipp