Bug 1576780

+++ This bug was initially created as a clone of Bug #1571078 +++

Description of problem:

The way the POWER HPT MMU is virtualized means that guests are only able to use page sizes equal to or smaller than the host pages used for the guest's RAM.

Currently we handle that by advertising different available pagesizes to the guest depending on the pagesize backing guest RAM.  This is messy, because it means the guest gets a visibly different environment depending on should be host side configuration details.  This can cause considerable confusion on migration, particularly from POWER8 to POWER9, which have different available hugepage sizes on the host.

Bug 1571076 is tracking a fix for that in qemu, creating an explicit option to control what page sizes should be available to an HPT guest.

To handle this gracefully in libvirt we need it to:

1) If a guest XML specifies hugepages, determine the available hugepage size on the host, and pass that value to the guest parameter to control the pagesizes that are available to it.

2) Once that's done, update the XML to explicitly list the required pagesize, so that future migrations won't change the available pagesizes.

--- Additional comment from Andrea Bolognani on 2018-05-07 09:34:23 EDT ---

(In reply to David Gibson from comment #0)
> Bug 1571076 is tracking a fix for that in qemu, creating an explicit option
> to control what page sizes should be available to an HPT guest.
> 
> To handle this gracefully in libvirt we need it to:
> 
> 1) If a guest XML specifies hugepages, determine the available hugepage size
> on the host, and pass that value to the guest parameter to control the
> pagesizes that are available to it.
> 
> 2) Once that's done, update the XML to explicitly list the required
> pagesize, so that future migrations won't change the available pagesizes.

As discussed on qemu-devel[1], libvirt can't sanely handle this
automatically without learning to differentiate between TCG, KVM
PR and KVM HV guests, as well as between HPT and RPT guests, all
of which have long been considered non-goals.

So libvirt will provide a configuration knob that the user can
take advantage of to control what page sizes are available to the
guest, but will not override QEMU's default out of the box: it
will be up to higher-level management tools to use the mechanism
to implement a suitable policy.


[1] http://lists.nongnu.org/archive/html/qemu-devel/2018-04/msg04595.html