Re: [PATCH R3 7/7] xen/balloon: Memory hotplug support for Xen balloon driver

[Daniel Kiper]

On Tue, Feb 08, 2011 at 09:22:26AM -0800, Dave Hansen wrote:
> On Mon, 2011-02-07 at 15:12 +0100, Daniel Kiper wrote:
> > On Thu, Feb 03, 2011 at 10:12:24AM -0800, Dave Hansen wrote:
> > > On Thu, 2011-02-03 at 17:30 +0100, Daniel Kiper wrote:
> > > > +static struct resource *allocate_memory_resource(unsigned long nr_pages)
> > > > +{
> > > > +	resource_size_t r_min, r_size;
> > > > +	struct resource *r;
> > > > +
> > > > +	/*
> > > > +	 * Look for first unused memory region starting at page
> > > > +	 * boundary. Skip last memory section created at boot time
> > > > +	 * because it may contains unused memory pages with PG_reserved
> > > > +	 * bit not set (online_pages() require PG_reserved bit set).
> > > > +	 */
> > >
> > > Could you elaborate on this comment a bit?  I think it's covering both
> > > the "PAGE_SIZE" argument to allocate_resource() and something else, but
> > > I'm not quite sure.
> >
> > Yes, you are right. Aligment to PAGE_SIZE is done by allocate_resource().
> > Additionally, r_min (calculated below) sets lower limit at which hoplugged
> > memory could be installed (due to PG_reserved bit requirment set up by
> > online_pages()). Later r_min is put as an argument to allocate_resource() call.
>
> OK, and you'll update the comment on that?

OK.

> > > > +	r = kzalloc(sizeof(struct resource), GFP_KERNEL);
> > > > +
> > > > +	if (!r)
> > > > +		return NULL;
> > > > +
> > > > +	r->name = "System RAM";
> > > > +	r->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
> > > > +	r_min = PFN_PHYS(section_nr_to_pfn(pfn_to_section_nr(balloon_stats.boot_max_pfn) + 1));
> > >
> > > Did you do this for alignment reasons?  It might be a better idea to
> > > just make a nice sparsemem function to do alignment.
> >
> > Please look above.
>
> You spoke about page alignment up there.  Why is this section-aligned?
> Should we make an "align to section" function in generic sparsemem code?

It is done because all pages in relevant section starting from max_pfn
to the end of that section do not have PG_reserved bit set. It was tested
on Linux Kernel Ver. 2.6.32.x, however, I am going to do some tests on
current Linus tree. Currently, I do not expect that "align to section"
function is required by others.

> > > > +	r_size = nr_pages << PAGE_SHIFT;
> > > > +
> > > > +	if (allocate_resource(&iomem_resource, r, r_size, r_min,
> > > > +					ULONG_MAX, PAGE_SIZE, NULL, NULL) < 0) {
> > > > +		kfree(r);
> > > > +		return NULL;
> > > > +	}
> > > > +
> > > > +	return r;
> > > > +}
> > >
> > > This function should probably be made generic.  I bet some more
> > > hypervisors come around and want to use this.  They generally won't care
> > > where the memory goes, and the kernel can allocate a spot for them.
> >
> > Yes, you are right. I think about new project in which
> > this function will be generic and then I would move it to
> > some more generic place. Now, I think it should stay here.
>
> Please move it to generic code.  It doesn't belong in Xen code.

OK.

> > > In other words, I think you can take everything from and including
> > > online_pages() down in the function and take it out.  Use a udev hotplug
> > > rule to online it immediately if that's what you want.
> >
> > I agree. I discussed a bit about this problem with Jeremy, too. However,
> > there are some problems to implement that solution now. First of all it is
> > possible to online hotplugged memory using sysfs interface only in chunks
> > called sections. It means that it is not possible online once again section
> > which was onlined ealier partialy populated and now it contains new pages
> > to online. In this situation sysfs interface emits Invalid argument error.
> > In theory it should be possible to offline and then online whole section
> > once again, however, if memory from this section was used is not possible
> > to do that. It means that those properties does not allow hotplug memory
> > in guest in finer granulity than section and sysfs interface is too inflexible
> > to be used in that solution. That is why I decided to online hoplugged memory
> > using API which does not have those limitations.
>
> Sure, you have to _online_ the whole thing at once, but you don't have
> to actually make the pages available.  You also don't need to hook in to
> the memory resource code like you're doing.  It's sufficient to just try
> and add the memory.  If you get -EEXIST, then you can't add it there, so
> move up and try again.
>
> int xen_balloon_add_memory(u64 size)
> {
> 	unsigned long top_of_mem = max_pfn;
> 	top_of_mem = section_align_up(top_of_mem);
>
> 	while (1) {
> 		int ret = add_memory(nid, top_of_mem, size);
> 		if (ret == -EEXIST)
> 			continue;
> 		// error handling...
> 		break;
> 	}
> 	return...;
> }
>
> As for telling the hypervisor where you've mapped things, that should be
> done in arch_add_memory().
>
> When it comes down to online_page(), you don't want your pages freed
> back in to the buddy allocator, you want them put in to the balloon.
> So, take the __free_page() in online_page(), and put a Xen hook in
> there.
>
> +void __attribute__((weak)) arch_free_hotplug_page(struct page *page)
> +{
> +	__free_page(page);
> +}

I think that this function should be registered dynamically at init
stage by specific balloon driver (in this case Xen balloon driver).

> void online_page(struct page *page)
> {
>         unsigned long pfn = page_to_pfn(page);
> ...
> -       __free_page(page);
> +	arch_free_hotplug_page(page);
> }
>
> Then, have Xen override it:
>
> void arch_free_hotplug_page(struct page *page)
> {
> 	if (xen_need_to_inflate_balloon())
> 		put_page_in_balloon(page);
> 	else
> 		__free_page(page);
> }

Whole idea looks quiet interesting. I will try to migrate
to that solution.

> > Additionally, IIRC, add_memory() requires
> > that underlying memory is available before its call.
>
> No, that's not correct.  s390's memory isn't available until after it
> calls vmem_add_mapping().  See arch/s390/mm/init.c

I was right to some extent. First versions of memory hotplug code were
written on the base of Linux Kernel Ver. 2.6.32.x. Tests done on that
versions showed that add_memory() required that underlying memory should
be available before its call. However, after short investigation it came
out that there are some issues with some Xen calls. Those issues does
not exists in current Linus tree.

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/