Re: xen: oops at atomic64_read_cx8+0x4

[Andrea Arcangeli <aarcange@redhat.com>]

Hi,

On Thu, Jun 07, 2012 at 11:56:47AM -0400, Konrad Rzeszutek Wilk wrote:
> then the high part. Why that would trip the hypervisor
> is not clear to me. Perhaps in the past it only read the

That is the CONFIG_TRANSPARENT_HUGEPAGE=n case and in fact it doesn't
trip the hypervisor. That was tested too, it should work fine.

The problem is with the atomic64_read version, that one uses cmpxchg8b
to read the contents of the pmdp.

> Ah, by just skipping the thing if the low bits are zero.

Yep.

> didn't like somebody reading the high and low bit, but isn't the
> pmdval_t still 64-bit ? So I would have thought this would

The pmd format is unchanged, that's hardware.

> The thing is that I did install a 32-bit PAE guest (a Fedora) on a Fedora
> 17 dom0. So it looks like this is reading high part is fixed on the newer
> hypervisors, but now with the older ones. And the older one is Amazon EC2
> so some .. hack to workaround older hypervisors could be added.

The insn oopsing is cmpxchg8b and it's not reading the low/high part
in two separate insn but reading it in a single insn, which means the
kernel oopsing was built with CONFIG_TRANSPARENT_HUGEPAGE=y.