[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] Re: [PATCH v2] Use ballooned pages for gntdev



On 03/10/2011 04:18 AM, Ian Campbell wrote:
> On Wed, 2011-03-09 at 23:07 +0000, Daniel De Graaf wrote:
>> Split up the balloon module, similar to how grants and events are
>> handled. This allow gntdev to use ballooned pages without adding a
>> dependency on the balloon module.
>>
>> The interface is slightly different from that exposed in 2.6.32; the
>> page vector is allocated by the caller, not by the balloon driver. This
>> allows more freedom in how the returned pages are used, since they do
>> not all have to be returned to the balloon driver at the same time. It
>> also tries to reuse already ballooned pages rather than always
>> ballooning new pages to ensure they are in low memory as the 2.6.32
>> version did.
>>
>> Changes since v1:
>>      * Rebased on konrad/devel/{balloon-hotplug,next-2.6.38} merge
>>      * Better function names
>>      * Adjust balloon stats for requested pages
>>
>> [PATCH 1/3] xen-balloon: Move core balloon functionality out of module
>> [PATCH 2/3] xen-balloon: Add interface to retrieve ballooned pages
>> [PATCH 3/3] xen-gntdev: Use ballooned pages for grant mappings
> 
> Does this introduces a new potential failure case? When a normal balloon
> operation takes place is it is now possible for the ballooned_pages list
> to be empty (because gntdev has stolen stuff from it) where before we
> knew the list was non-empty at that point?
> 
> e.g. increase_reservation includes:
>               page = balloon_retrieve();
>               BUG_ON(page == NULL);
> as well as:
>       page = balloon_first_page();
>       for (i = 0; i < nr_pages; i++) {
>               BUG_ON(page == NULL);
>               frame_list[i] = page_to_pfn(page);
>               page = balloon_next_page(page);
>       }
> 
> Are we sure we aren't about to exercise those BUG_ONs?

This is safe because increase_reservation is protected by the balloon
mutex. There is a loop just before the hypercall that verifies that
there are enough pages in the list.

> As a secondary concern, assuming we don't crash, does the balloon
> process correctly sleep until explicitly kicked when ballooned_pages
> becomes empty? Or does it keep needlessly waking up on the jiffy timer?
> IOW does credit in balloon_process become 0 when this situation occurs?
> 
> Or does adjusting the balloon stats ensure this all works out alright?

Adjusting the statistics should make this work out properly. Since only
pages actually in the list will be counted in balloon_low+balloon_high,
which is the lower bound for credit, we will always be able to satisfy
the reservation changes requested by balloon_process. The jiffy timer is
used only when Linux or Xen cannot satisfy the memory allocations required
to adjust the balloon size.

> An interesting experiment would probably be to allocate a bunch of
> address space via gntdev and then manually balloon the guest above it's
> current allocation, back below the lower limit due to the gntdev
> allocation, back up to starting point etc etc.
> 
> Ian.
> 

That does sound like a useful test case.

-- 
Daniel De Graaf
National Security Agency

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.