Re: [Xen-devel] PAE PV guest kernel regression

["Jan Beulich" <jbeulich@xxxxxxxx>]

>>> Andrew Jones <drjones@xxxxxxxxxx> 06/07/12 12:47 PM >>>
>Ah, I didn't check to see what Fedora pulled in on top of 3.4. Had I
>done that I would have immediately suspected a different patch instead
>(mm-pmd_read_atomic-fix-32bit-PAE-pmd-walk-vs-pmd_populate-SMP-race-condition.patch,
>upstream commit 26c191788f18). We've already encountered one problem
>with this patch for RHEL6 and fixed it. The patch F17 has, however, is
>already the "fixed" version. Now the difference between RHEL6 and F17
>though is that F17 has CONFIG_TRANSPARENT_HUGEPAGE=y for 32b guests,
>but RHEL6 does not. So now with this patch F17 is calling
>atomic64_read() from pmd_none_or_trans_huge_or_clear_bad().

By itself the use of atomic64_read() on page table entries shouldn't be
a problem though, particularly not if what might get written is a not
present entry (while the way atoimc64_read_cx8() works would also
allow for values with the low bit set to be pseudo-written, but neither
did that appear to be the case in at least one of the two cases where
register dumps were provided in your bugzilla, nor should an attempt
to write back a value that was already there cause any problem, even
if the low bit was set).

In any case, the hypervisor log would be interesting to see. Plus - is
this perhaps a rather old hypervisor running there (i.e. potentially
lacking some fix)?

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel