Re: [PATCH 2/2] Drivers: hv: Add Hyper-V balloon driver

[Rusty Russell <rusty@ozlabs.org>]

Andrew Morton <akpm@linux-foundation.org> writes:
> On Wed, 10 Oct 2012 16:34:37 -0700
> Jeremy Fitzhardinge <jeremy@goop.org> wrote:
>
>> On 10/09/2012 06:14 PM, Andrew Morton wrote:
>> > On Wed, 10 Oct 2012 00:09:12 +0000 KY Srinivasan <kys@microsoft.com> wrote:
>> >
>> >>>> +		if (!pg) {
>> >>>> +			*alloc_error = true;
>> >>>> +			return i * alloc_unit;
>> >>>> +		}
>> >>>> +
>> >>>> +		totalram_pages -= alloc_unit;
>> >>> Well, I'd consider totalram_pages to be an mm-private thing which drivers
>> >>> shouldn't muck with.  Why is this done?
>> >> By modifying the totalram_pages, the information presented in /proc/meminfo
>> >> correctly reflects what is currently assigned to the guest (MemTotal).
>> > eh?  /proc/meminfo:MemTotal tells you the total memory in the machine. 
>> > The only thing which should change it after boot is memory hotplug. 
>> [...]
>> > Why on earth do balloon drivers do this?  If the amount of memory which
>> > is consumed by balloons is interesting then it should be exported via a
>> > standalone metric, not by mucking with totalram_pages.
>> 
>> Balloon drivers are trying to fake a form of page-by-page memory
>> hotplug.  When they allocate memory from the kernel, they're actually
>> giving the pages back to the hypervisor to redistribute to other
>> guests.  They reduce totalram_pages to try and reflect that the memory
>> is no longer the kernel (in Xen, at least, the pfns will no longer have
>> any physical page underlying them).
>> 
>> I agree this is pretty ugly; it would be nice to have some better
>> interface to indicate what's going on.  At one point I tried to use the
>> memory hotplug interfaces for larger-scale dynamic transfers of memory
>> between a domain and the host, but when I last looked at it, it was too
>> coarse grained and heavyweight to replace the balloon mechanism.
>> 
>
> urgh.
>
> I suppose the least we can do here would be to stop directly dinking
> with totalram_pages and create some sort of interface for this
> operation.  That interface would run the memory hotplug notifier so
> that code which cares about changes in the amount of physical memory
> can take appropriate steps.  The implications would be that the balloon
> drivers would need to call this interface at low frequency (ie: batch
> the pages) and in some reasonably lock-free context.
>
> I guess that's solving a non-problem at this stage.

Yep.  drivers/virtio/virtio_balloon manipulates it too.  This, it's best
practice!

Cheers,
Rusty.