Re: [Xen-devel] [PATCH v3] libxl: Wait for ballooning if free memory is increasing

[Wei Liu <wei.liu2@xxxxxxxxxx>]

On Tue, Feb 10, 2015 at 09:17:23PM -0700, Mike Latimer wrote:
> On Thursday, February 05, 2015 12:45:53 PM Ian Campbell wrote:
> > On Mon, 2015-02-02 at 08:17 -0700, Mike Latimer wrote:
> > > On Monday, February 02, 2015 02:35:39 PM Ian Campbell wrote:
> > > > On Fri, 2015-01-30 at 14:01 -0700, Mike Latimer wrote:
> > > > > During domain startup, all required memory ballooning must complete
> > > > > within a maximum window of 33 seconds (3 retries, 11 seconds of
> > > > > delay).
> > > > > If not, domain creation is aborted with a 'failed to free memory'
> > > > > error.
> ...
> > Thanks, applied.
> 
> Unfortunately, I just found an issue with this codepath...
> 
> In tools/libxl/xl_cmdimpl.c:freemem, the following call sets the memory 
> target 
> for dom0:
> 
>   rc = libxl_set_memory_target(ctx, 0, free_memkb - need_memkb, 1, 0);
> 
> This reduces the memory target of dom0 by the amount of memory still needed - 
> relative to the amount of memory currently free. In other words, during each 
> iteration of the loop, dom0's target memory is set lower, and lower, and 
> lower...
> 
> Prior to my changes, this issue would only be noticed when starting very 
> large 
> domains - due to the loop being limited to 3 iterations. (For example, when 
> ballooning 512G, dom0 memory could be reduced up to 1.5T.) With my changes, 
> this loop can take several more iterations (10 or many more in testing). With 
> each iteration lowering dom0's target, the end result can be dom0 ballooning 
> down 100's of gigabytes, just to satisfy a much smaller request. (On one 
> machine, a 64G guest can result in dom0 ballooning down 500G.)
> 

I'm curious why freemem doesn't not return when dom0 is ballooned down 64G?
I.e. I think libxl_wait_for_free_memory should return 0 in that case and
freeme should just return. Does it suggest there are other domains
competing for free memory and dom0 is doing the right thing?

Wei.

> It seems like a proper fix would be to use libxl_get_memory_target to first 
> check dom0's target and see if:
> 
>   (free_memory + (dom0_current_mem - dom0_target_mem)) >= needed_memory
> 
> If so, there will be sufficient memory when dom0 reaches the target, so don't 
> change it. Before going down that road, I'd like a sanity check on that 
> theory, and any advice on the overall picture (including my thread on freemem-
> slack).
> 
> I'll leave it up to you if my patch should be reverted until a new patch is 
> created.
> 
> Thanks!
> Mike

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel