Re: [PATCH v2] hw/arm/virt enable support for virtio-mem

[David Hildenbrand]

On 25.11.20 09:38, Andrew Jones wrote:
> On Tue, Nov 24, 2020 at 08:17:35PM +0100, David Hildenbrand wrote:
>> On 24.11.20 19:11, Jonathan Cameron wrote:
>>> On Mon, 9 Nov 2020 20:47:09 +0100
>>> David Hildenbrand <david@redhat.com> wrote:
>>>
>>> +CC Eric based on similar query in other branch of the thread.
>>>
>>>> On 05.11.20 18:43, Jonathan Cameron wrote:
>>>>> Basically a cut and paste job from the x86 support with the exception of
>>>>> needing a larger block size as the Memory Block Size (MIN_SECTION_SIZE)
>>>>> on ARM64 in Linux is 1G.
>>>>>
>>>>> Tested:
>>>>> * In full emulation and with KVM on an arm64 server.
>>>>> * cold and hotplug for the virtio-mem-pci device.
>>>>> * Wide range of memory sizes, added at creation and later.
>>>>> * Fairly basic memory usage of memory added.  Seems to function as normal.
>>>>> * NUMA setup with virtio-mem-pci devices on each node.
>>>>> * Simple migration test.
>>>>>
>>>>> Related kernel patch just enables the Kconfig item for ARM64 as an
>>>>> alternative to x86 in drivers/virtio/Kconfig
>>>>>
>>>>> The original patches from David Hildenbrand stated that he thought it 
>>>>> should
>>>>> work for ARM64 but it wasn't enabled in the kernel [1]
>>>>> It appears he was correct and everything 'just works'.
>>>>>
>>>>> The build system related stuff is intended to ensure virtio-mem support is
>>>>> not built for arm32 (build will fail due no defined block size).
>>>>> If there is a more elegant way to do this, please point me in the right
>>>>> direction.  
>>>>
>>>> You might be aware of https://virtio-mem.gitlab.io/developer-guide.html 
>>>> and the "issue" with 64k base pages - 512MB granularity. Similar as the 
>>>> question from Auger, have you tried running arm64 with differing page 
>>>> sizes in host/guest?
>>>>
>>>
>>> Hi David,
>>>
>>>> With recent kernels, you can use "memhp_default_state=online_movable" on 
>>>> the kernel cmdline to make memory unplug more likely to succeed - 
>>>> especially with 64k base pages. You just have to be sure to not hotplug 
>>>> "too much memory" to a VM.
>>>
>>> Thanks for the pointer - that definitely simplifies testing.  Was getting a 
>>> bit
>>> tedious with out that.
>>>
>>> As ever other stuff got in the way, so I only just got back to looking at 
>>> this.
>>>
>>> I've not done a particularly comprehensive set of tests yet, but things seem
>>> to 'work' with mixed page sizes.
>>>
>>> With 64K pages in general, you run into a problem with the device 
>>> block_size being
>>> smaller than the subblock_size.  I've just added a check for that into the
>>
>> "device block size smaller than subblock size" - that's very common,
>> e.g.,  on x86-64.
>>
>> E.g., device_block_size is 2MiB, subblock size 4MiB - until we improve
>> that in the future in Linux guests.
>>
>> Or did you mean something else?
>>
>>> virtio-mem kernel driver and have it fail to probe if that happens.  I don't
>>> think such a setup makes any sense anyway so no loss there.  Should it make 
>>> sense
>>> to drop that restriction in the future we can deal with that then without 
>>> breaking
>>> backwards compatibility.
>>>
>>> So the question is whether it makes sense to bother with virtio-mem support
>>> at all on ARM64 with 64k pages given currently the minimum workable 
>>> block_size
>>> is 512MiB?  I guess there is an argument of virtio-mem being a possibly more
>>> convenient interface than full memory HP.  Curious to hear what people 
>>> think on
>>> this?
>>
>> IMHO we really want it. For example, RHEL is always 64k. This is a
>> current guest limitation, to be improved in the future - either by
>> moving away from 512MB huge pages with 64k or by improving
>> alloc_contig_range().
> 
> Even with 64k pages you may be able to have 2MB huge pages by setting
> default_hugepagesz=2M on the kernel command line.

Yes, but not for THP, right? Last time I checked that move was not
performed yet - resulting in MAX_ORDER/pageblock_order in Linux
corresponding to 512 MB.

> 
> Thanks,
> drew


-- 
Thanks,

David / dhildenb