Re: Possible memory ordering bug in page reclaim?

[Nick Piggin <nickpiggin@yahoo.com.au>]

Linus Torvalds wrote:

> I agree, however, that it looks like PG_dirty is racy. Probably not in 
> practice, but still.
> 
> So I'd suggest adding a smp_wmb() into set_page_dirty, and the rmb where 
> Nick suggested.
> 
> So I'd suggest a patch something more like this.. Marking the dirty/count 
> cases unlikely too in mm/page-writeback.c, since we should have tested for 
> these conditions optimistically outside the lock.
> 

As Dave suggested, I think there is too much other code that depends on
these atomics to be barriers for us to change it (at least not in this
patch! :)).


> Comments? Nick, did you have some test-case that you think might actually 
> have been impacted by this?
> 

I guess your vmscan.c hunks are slightly nicer, though I might put
'cannot_free' right at the end, because it will be a very uncommon case.

And no, I don't have a test case. In fact, I wouldn't be surprised if
nobody anywhere has ever hit it :) I was just browsing code...

Thanks,
Nick

-- 
SUSE Labs, Novell Inc.

Send instant messages to your online friends http://au.messenger.yahoo.com