Re: [PATCH 1/3 v2] mm: Batch unmapping of file mapped pages inshrink_page_list

[Tim Chen]

On Tue, 2012-09-11 at 12:05 +0100, Mel Gorman wrote:

> 
> One *massive* change here that is not called out in the changelog is that
> the reclaim path now holds the page lock on multiple pages at the same
> time waiting for them to be batch unlocked in __remove_mapping_batch.
> This is suspicious for two reasons.
> 
> The first suspicion is that it is expected that there are filesystems
> that lock multiple pages in page->index order and page reclaim tries to
> lock pages in a random order.  You are "ok" because you trylock the pages
> but there should be a comment explaining the situation and why you're
> ok.
> 
> My *far* greater concern is that the hold time for a locked page is
> now potentially much longer. You could lock a bunch of filesystem pages
> and then call pageout() on an swapcache page that takes a long time to
> write. This potentially causes a filesystem (or flusher threads etc)
> to stall on lock_page and that could cause all sorts of latency trouble.
> It will be hard to hit this bug and diagnose it but I believe it's
> there.
> 
> That second risk *really* must be commented upon and ideally reviewed by
> the filesystem people. However, I very strongly suspect that the outcome
> of such a review will be a suggestion to unlock the pages and reacquire
> the lock in __remove_mapping_batch(). Bear in mind that if you take this
> approach that you *must* use trylock when reacquiring the page lock and
> handle being unable to lock the page.
> 

Mel,

Thanks for your detailed comments and analysis.  If I unlock the pages,
will flusher threads be the only things that will touch them?  Or do I
have to worry about potentially other things done to the pages that will
make it invalid for me to unmap the pages later and put them on free
list?

Tim

--
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/