Re: [PATCH 3/3] mm: Extend memory hotplug API to allow memory hotplug in virtual machines

From: Daniel Kiper
Date: Tue Mar 29 2011 - 17:54:56 EST


On Tue, Mar 29, 2011 at 12:15:41PM -0700, Andrew Morton wrote:
> On Tue, 29 Mar 2011 20:59:13 +0200
> Daniel Kiper <dkiper@xxxxxxxxxxxx> wrote:
>
> > > This is a bit strange. Normally we'll use a notifier chain to tell
> > > listeners "hey, X just happened". But this code is different - it
> > > instead uses a notifier chain to tell handlers "hey, do X". Where in
> > > this case, X is "free a page".
> > >
> > > And this (ab)use of notifiers is not a good fit! Because we have the
> > > obvious problem that if there are three registered noftifiers, we don't
> > > want to be freeing the page three times. Hence the tricks with
> > > notifier callout return values.
> > >
> > > If there are multiple independent notifier handlers, how do we manage
> > > their priorities? And what are the effects of the ordering of the
> > > registration calls?
> > >
> > > And when one callback overrides an existing one, is there any point in
> > > leaving the original one installed at all?
> > >
> > > I dunno, it's all a bit confusing and strange. Perhaps it would help
> > > if you were to explain exactly what behaviour you want here, and we can
> > > look to see if there is a more idiomatic way of doing it.
> >
> > OK. I am looking for simple generic mechanism which allow runtime
> > registration/unregistration of generic or module specific (in that
> > case Xen) page onlining function. Dave Hansen sugested compile time
> > solution (https://lkml.org/lkml/2011/2/8/235), however, it does not
> > fit well in my new project on which I am working on (I am going post
> > details at the end of April).
>
> Well, without a complete description of what you're trying to do and
> without any indication of what "does not fit well" means, I'm at a bit
> of a loss to suggest anything.

The most important thing for me is runtime registration/unregistration.
It will be good if it is possible to register more than one callback
at a time (e.g. for counting), however, it is not required now. It
appears that your proposal fits quite well my requirements. I will check
that. Thank you.

> If we are assured that only one callback will ever be registered at a
> time then a simple
>
> typdef void (*callback_t)(struct page *);
>
> static callback_t g_callback;
>
> int register_callback(callback_t callback)
> {
> int ret = -EINVAL;
>
> lock(some_lock);
> if (g_callback == NULL) {
> g_callback = callback;
> ret = 0;
> }
> unlock(some_lock)
> return ret;
> }
>
> would suffice. That's rather nasty because calls to (*g_callback)
> require some_lock. Use RCU.

I think that in this case lock_memory_hotplug()/unlock_memory_hotplug()
is much better because it is used for locking during memory hotplug
operation. That means they protect against callback changes during
memory hotplug. It appears sufficient here.

> > > Also... I don't think we need (the undocumented)
> > > OP_DO_NOT_INCREMENT_TOTAL_COUNTERS and OP_INCREMENT_TOTAL_COUNTERS.
> > > Just do
> > >
> > > void __online_page_increment_counters(struct page *page,
> > > bool inc_total_counters);
> > >
> > > and pass it "true" or false".
> >
> > What do you think about __online_page_increment_counters()
> > (totalram_pages and totalhigh_pages) and
> > __online_page_set_limits() (num_physpages and max_mapnr) ???
>
> I don't understand the proposal.

void __online_page_increment_counters(struct page *page)
{
totalram_pages++;

#ifdef CONFIG_HIGHMEM
if (PageHighMem(page))
totalhigh_pages++;
#endif
}

void __online_page_set_limits(struct page *page)
{
unsigned long pfn = page_to_pfn(page);

if (pfn >= num_physpages)
num_physpages = pfn + 1;

#ifdef CONFIG_FLATMEM
max_mapnr = max(pfn, max_mapnr);
#endif
}

Daniel
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/