Re: [RFC] Changing COW detection to be memory hotplug friendly

[Andrea Arcangeli <andrea@suse.de>]

On Thu, Feb 10, 2005 at 11:16:44AM -0800, Dave Hansen wrote:
> We actually do that, in addition to the more active methods of capturing
> the memory that we're about to remove.

This sounds the best way to go. btw, is this a different codebase from
the one that Toshihiro is testing?

> You're right, I don't really see a problem with ignoring those pages, at
> least in the active migration code.  We would, of course, like to keep
> the number of things that we depend on good faith to get migrated to a
> minimum, but things under I/O are the least of our problems.

Indeed. It's very similar to locked pages. All pages can be pinned for a
transient amount of time, either in the pte or with
PG_pinned/PG_writeback (now perhaps Hugh found a way to drop the pin on
the pte [I'm still unconvinced about that], but sure you're left with
transient pinning with PG_locked or PG_writeback).

> The only thing we might want to do is put something in the rawio code to
> look for the PG_capture pages to ensure that it gives the migration code
> a shot at them every once in a while (when I/O is not in flight,
> obviously).

If there are persistent usages PG_capture sounds a good idea. Perhaps
the whole point that Toshihiro has problem with is that there are really
persistent users that require PG_capture? He mentioned direct IO, that's
not a long time, if something core dump could be a long time.