lkml.org 
[lkml]   [2012]   [Aug]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v8 1/5] mm: introduce a common interface for balloon pages mobility
On Thu, Aug 23, 2012 at 09:13:39AM -0300, Rafael Aquini wrote:
> On Thu, Aug 23, 2012 at 01:01:07PM +0300, Michael S. Tsirkin wrote:
> > > So, when remove_common() calls leak_balloon() looping on
> > > vb->num_pages, that won't become a tight loop.
> > > The scheme was apparently working before this series, and it will remain working
> > > after it.
> >
> > It seems that before we would always leak all requested memory
> > in one go. I can't tell why we have a while loop there at all.
> > Rusty, could you clarify please?
> >
>
> It seems that your claim isn't right. leak_balloon() cannot do it all at once,
> as for each round it only releases 256 pages, at most; and the 'one go' would
> require a couple of loop rounds at remove_common().

You are right in this respect.

> So, nothing has changed here.

Yes, your patch does change things:
leak_balloon now might return without freeing any pages.
In that case we will not be making any progress, and just
spin, pinning CPU.

>
> > > Just as before, same thing here. If you leaked less than required, balloon()
> > > will keep calling leak_balloon() until the balloon target is reached. This
> > > scheme was working before, and it will keep working after this patch.
> > >
> >
> > IIUC we never hit this path before.
> >
> So, how does balloon() works then?
>

It gets a request to leak a given number of pages
and executes it, then tells host that it is done.
It never needs to spin busy-waiting on a CPU for this.

> > > > How about we signal config_change
> > > > event when pages are back to pages_list?
> > >
> > > I really don't know what to tell you here, but, to me, it seems like an
> > > overcomplication that isn't directly entangled with this patch purposes.
> > > Besides, you cannot expect compation / migration happening and racing against
> > > leak_balloon() all the time to make them signal events to the later, so we might
> > > just be creating a wait-forever condition for leak_balloon(), IMHO.
> >
> > So use wait_event or similar, check for existance of isolated pages.
> >
>
> The thing here is expecting compaction as being an external event to signal
> actions to the balloon driver won't work as you desire. Also, as far as the
> balloon driver is concerned, it's only a matter of time to accomplish a total,
> or partial, balloon leak, even when we have some pages isolated from balloon's
> page list.
>
> IMHO, you're attempting to complicate a simple thing that is already working
> well. As said before, there are no guarantees you'll have isolated pages
> by the time you're leaking the balloon, so you might leave it waiting forever
> on something that will not happen. And if there are isolated pages while balloon
> is leaking, they'll have their chance to get back to the list before the device
> finishes its leaking job.

Well busy wait pinning CPU is ugly. Instead we should block thread and
wake it up when done. I don't mind how we fix it specifically.

--
MST


\
 
 \ /
  Last update: 2012-08-23 15:21    [W:0.077 / U:0.204 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site