On Tue, Feb 20, 2024 at 10:14:10 +0100, Michal Privoznik wrote:
As of v9.8.0-rc1~7 we check whether two <memory/> devices
don't
overlap (since we allow setting where a <memory/> device should
be mapped to). We do this pretty straightforward, by comparing
start and end address of each <memory/> device combination.
But since only the start address is given (an exposed in the
XML), the end address is computed trivially as:
start + mem->size * 1024
And for majority of memory device types this works. Except for
NVDIMMs. For them the <memory/> device consists of two separate
regions: 1) actual memory device, and 2) label.
Label is where NVDIMM stores some additional information like
namespaces partition and so on. But it's not mapped into the
guest the same way as actual memory device. In fact, mem->size is
a sum of both actual memory device and label sizes. And to make
things a bit worse, both sizes are subject to alignment (either
the alignsize value specified in XML, or system page size if not
specified in XML).
Therefore, to get the size of actual memory device we need to
take mem->size and substract label size rounded up to alignment.
If we don't do this we report there's an overlap between two
NVDIMMs even when in reality there's none.
Resolves:
https://issues.redhat.com/browse/RHEL-4452?focusedId=23805174#comment-238...
Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com>
---
Please add a 'Fixes: ' with the commit that introduced the broken
check.
I'd also suggest you add negative test case which will specifically try
the scenario.
It should be fairly straightforward in this case to setup the nvdimm so
that it has a label and add another device right after it.