On Mon, 4 Jun 2012, Minchan Kim wrote:

On 06/02/2012 01:40 PM, Hugh Dickins wrote:

quoted

On Fri, 1 Jun 2012, Linus Torvalds wrote:

quoted

On Fri, Jun 1, 2012 at 3:17 PM, Hugh Dickins [off-list ref] wrote:

quoted

+       spin_lock_irqsave(&zone->lock, flags);
       for (page = start_page, pfn = start_pfn; page < end_page; pfn++,
                                                                 page++) {

So holding the spinlock (and disabling irqs!) over the whole loop
sounds horrible.

There looks to be a pretty similar loop inside move_freepages_block(),
which is the part which I believe really needs the lock - it's moving
free pages from one lru to another.

quoted

At the same time, the iterators don't seem to require the spinlock, so
it should be possible to just move the lock into the loop, no?

Move the lock after the loop, I think you meant.

I put the lock before the loop because it's deciding whether it can
usefully proceed, and then proceeding: I was thinking that the lock
would stabilize the conditions that it bases that decision on.

We do it with two phase.
In first phase, we don't need lock because we don't need to be exact.
In second phase where move pages really, we need a lock so we already hold it.

No, see Linus's point elsewhere in this thread.

To spell it out further, page_order(page) uses page_private(page),
and you've no idea what someone might put into page_private(page)
once it's no longer PageBuddy but perhaps allocated to a user.

So the unlocked advancment by page_order(page) may even take you
way out of this or any pageblock.

Linus was suggesting to take and drop the lock around that little
block each time.  Maybe.  I'm wary, I don't pretend to have thought
it through (nor shall further).

ret = suitable_migration_target(page, cc);
..
..
spin_lock_irqsave(&zone->lock, flags);
ret = suitable_migration_target(page, cc); 

So you shouldn't put the lock in loop.

quoted

But it certainly does not stabilize all of them (most obviously not
PageLRU), so I'm guesssing that this is a best-effort decision which

quoted

can safely go wrong some of the time.

Right.

quoted

In which case, yes, much better to follow your suggestion, and hold
the lock (with irqs disabled) for only half the time.

Similarly untested patch below.

But I'm entirely unfamiliar with this code: best Cc people more familiar
with it.  Does this addition of locking to rescue_unmovable_pageblock()
look correct to you, and do you think it has a good chance of fixing the

No.I think we need to use start_page instead of page and

I thought so, but Linus points out why not (pfn_valid_within).

we need a last page of page block to check cross-over zones,
not first page in next page block.

Yes, that's the off-by-one I was alluding to.

Yes.

Yes.

No.  I guess we can assume the incoming page was valid (fair?),
so should still use that, but something else for the loop iterator.

And you seem to have missed out all the locking needed.

        return true;
 }

So Nack to that on several grounds.

And I'd like to hear evidence that this really is useful code,
justifying the locking and interrupt-disabling which would have to
be added.  My 0 out of 25000 was not reassuring.  Nor the original
test results, when it was doing completely the wrong thing unnoticed.

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>