RHEV should determine if the guest supports HPT resizing, then increase the default max size of memory.
+++ This bug was initially created as a clone of Bug #1308743 +++
In case libvirt requires changes to support HPT resizing. If no changes required, set this BZ to TestOnly.
+++ This bug was initially created as a clone of Bug #1305398 +++
Description of problem:
Allow the hash page table (HPT) of PAPR guests to be resized at runtime.
This is important for practical memory hotplug. Without this the HPT needs to be sized for the guest's maximum possible memory - since RHEV wants to set that to 4T, this can result in a much bigger than necessary HPT which wastes host resources and can cause allocation failures. With HV KVM the HPT is unswappable, contiguous host memory.
This BZ covers the qemu parts of this including TCG and PR KVM implementation of the necessary hypercalls, feature negotation with the guest and enabling the necessary KVM host pieces.
--- Additional comment from David Gibson on 2016-02-07 19:52:00 EST ---
An RFC has been posted upstream:
*** Bug 1308746 has been marked as a duplicate of this bug. ***
The resize-hpt=required machine option in qemu might be useful for RHEV to determine if the guest supports HPT resizing. With that option qemu will refuse to boot a guest which does not support it (exiting with an error during boot).
So, RHEV could boot with that option, then if the boot fails adjust the max memory size down and restart with without the option to run the non-HPT-resize aware guest.
Obviously RHEV might then want to cache the value, and/or pre-populate it when it knows the distro / version of the guest.
Note that the necessary qemu and kvm parts for this are now merged downstream.
That least only libvirt work as a prerequisite (which should be relatively simple) to use this in RHV.
It should be possible to unblock this. The qemu and libvirt changes are merged and released downstream, and seem to be working well (after some initial bugs were found and fixed).
There shouldn't be a lot that's needed on the oVirt / RHV side. Basically qemu and the guest should negotiate HPT resizing automatically, and it will be triggered automatically when memory is hotplugged or unplugged. libvirt should correctly handle locked memory allocation for it.
The only real impact for RHV is that for guests which do support HPT resizing, it can (again) freely use a large maximum memory size without that causing allocation of excessively sized HPTs any more.
I guess we do not really need to change anything on RHV side. Since the bug was opened we added a notion of max memory (upper hotplug limit) for all platforms and use a conservative default of 4x the configured memory. That helped "enough" with the original problem, for sane VMs we're not allocating 4TB(2TB on ppc) blindly, and user have a way how to change that when they do not plan to use hotplug at all
With the HPT resizing implemented we will just see more VMs to fit into the host as the default overhead of counting with 4x as much memory is gone. But that's entirely transparent to the end user, it will just "work better" now.
Great, sounds like we're in agreement.
Verified on PPC environment.
Reference to verification:
We cloned this bug to 4.2.6, I wasn't sure if the status for the 4.3 should stay verified or move back to ON_QA, please update if QE plans to verify it also on 4.3
(In reply to Eyal Edri from comment #12)
> We cloned this bug to 4.2.6, I wasn't sure if the status for the 4.3 should
> stay verified or move back to ON_QA, please update if QE plans to verify it
> also on 4.3
You can keep it on modified as it is a medium and TestOnly
moving back to Modified so QE can test it for 4.3
As comment #9, the tests passed on 4.3 environment.
Power9 hosts, RHEL7.6
Moving to verified.
Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.
For information on the advisory, and where to find the updated
files, follow the link below.
If the solution does not work for you, open a new bug report.