On 06.04.22 15:32, Juergen Gross wrote:
When onlining a new memory page in a guest the Xen balloon driver is adding it to the ballooned pages instead making it available to be used immediately. This is meant to enable to add a new upper memory limit to a guest via hotplugging memory, without having to assign the new memory in one go.
In case the upper memory limit will be raised above 4G, the new memory will populate the ZONE_NORMAL memory zone, which wasn't populated before. The newly populated zone won't be added to the list of zones looked at by the page allocator though, as only zones with available memory are being added, and the memory isn't yet available as it is ballooned out.
I think we just recently discussed these corner cases on the -mm list. The issue is having effectively populated zones without manages pages because everything is inflated in a balloon.
That can theoretically also happen when managing to fully inflate the balloon in one zone and then, somehow, the zones get rebuilt.
build_zonerefs_node() documents "Add all populated zones of a node to the zonelist" but checks for managed zones, which is wrong.
See https://lkml.kernel.org/r/20220201070044.zbm3obsoimhz3xd3@master
This will result in the new memory being assigned to the guest, but without the allocator being able to use it.
When running as a PV guest the situation is even worse: when having been started with less memory than allowed, and the upper limit being lower than 4G, ballooning up will have the same effect as hotplugging new memory. This is due to the usage of the zone device functionality since commit 9e2369c06c8a ("xen: add helpers to allocate unpopulated memory") for creating mappings of other guest's pages, which as a side effect is being used for PV guest ballooning, too.
Fix this by checking in xen_online_page() whether the new memory page will be the first in a new zone. If this is the case, add another page to the balloon and use the first memory page of the new chunk as a replacement for this now ballooned out page. This will result in the newly populated zone containing one page being available for the page allocator, which in turn will lead to the zone being added to the allocator.
This somehow feels like a hack for something that should be handled in the core instead :/