Re: Possible memory ordering bug in page reclaim?

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

Benjamin Herrenschmidt wrote:
> On Sat, 2005-10-15 at 07:17 +0100, Hugh Dickins wrote:
> 
>>On Sat, 15 Oct 2005, Nick Piggin wrote:
>>

>>>1                                2
>>>find_get_page();
>>>write to page                    write_lock(tree_lock);
>>>SetPageDirty();                  if (page_count != 2
>>>put_page();                          || PageDirty())
>>>
>>>Now I'm worried that 2 might see PageDirty *before* SetPageDirty in
>>
>>                                  page->flags
>>
>>>1, and page_count *after* put_page in 1.
>>
>>I think you're right.  But I'm the last person to ask
>>barrier/ordering questions of.  CC'ed Ben and Andrea.
> 
> 
> yup, now the question is wether PG_Dirty will be visible to CPU 2 before
> the page count is decremented right ? That depends on put_page, I
> suppose. If it's doing a simple atomic, there is an issue. But atomics
> with return has been so often abused as locks that they may have been
> implemented with a barrier... (On ppc64, it will do an eieio, thus I
> think it should be ok).
> 

Well yes, that's on the store side (1, above). However can't a CPU
still speculatively (eg. guess the branch) load the page->flags
cacheline which might be satisfied from memory before the page->count
cacheline loads? Ie. you can still have the correct write ordering
but have incorrect read ordering?

Because neither PageDirty nor page_count is a barrier, and there is
no read barrier between them.

Nick

-- 
SUSE Labs, Novell Inc.

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