Re: [PATCH V2 1/2] mm: hotplug: implement non-movable version ofget_user_pages() called get_user_pages_non_movable()

[Tang Chen]

Hi Benjamin, Mel,

Please see below.

On 05/14/2013 09:58 PM, Benjamin LaHaise wrote:
> On Tue, May 14, 2013 at 09:24:58AM +0800, Tang Chen wrote:
> > Hi Mel, Benjamin, Jeff,
> >
> > On 05/13/2013 11:01 PM, Benjamin LaHaise wrote:
> > > On Mon, May 13, 2013 at 10:54:03AM -0400, Jeff Moyer wrote:
> > > > How do you propose to move the ring pages?
> > >
> > > It's the same problem as doing a TLB shootdown: flush the old pages from
> > > userspace's mapping, copy any existing data to the new pages, then
> > > repopulate the page tables.  It will likely require the addition of
> > > address_space_operations for the mapping, but that's not too hard to do.
> >
> > I think we add migrate_unpin() callback to decrease page->count if
> > necessary,
> > and migrate the page to a new page, and add migrate_pin() callback to pin
> > the new page again.
>
> You can't just decrease the page count for this to work.  The pages are
> pinned because aio_complete() can occur at any time and needs to have a
> place to write the completion events.  When changing pages, aio has to
> take the appropriate lock when changing one page for another.

In aio_complete(),

aio_complete() {
	......
	spin_lock_irqsave(&ctx->completion_lock, flags);
	//write the completion event.
	spin_unlock_irqrestore(&ctx->completion_lock, flags);
	......
}

So for this problem, I think we can hold ctx->completion_lock in the aio
callbacks to prevent aio subsystem accessing pages who are being migrated.

> > The migrate procedure will work just as before. We use callbacks to
> > decrease
> > the page->count before migration starts, and increase it when the migration
> > is done.
> >
> > And migrate_pin() and migrate_unpin() callbacks will be added to
> > struct address_space_operations.
>
> I think the existing migratepage operation in address_space_operations can
> be used.  Does it get called when hot unplug occurs?  That is: is testing
> with the migrate_pages syscall similar enough to the memory removal case?

But as I said, for anonymous pages such as aio ring buffer, they don't have
address_space_operations. So where should we put the callbacks' pointers ?

Add something like address_space_operations to struct anon_vma ?

Thanks. :)






--
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/