Re: smp race fix between invalidate_inode_pages* and do_no_page

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

Andrea Arcangeli wrote:
> On Fri, Dec 16, 2005 at 02:51:47PM +0100, Andrea Arcangeli wrote:
> 
>>There was a minor buglet in the previous patch an update is here:
>>
>>	http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.15-rc5/seqschedlock-2
> 
> 
> JFYI: I got a few hours ago positive confirmation of the fix from the
> customer that was capable of reproducing this. I guess this is good
> enough for production use (it's at the very least certainly better than
> the previous code and it's guaranteed not to hurt the scalability of the
> fast path in smp, so it's the least intrusive fix I could imagine).
> 
> So we can start to think if we should using this new primitive I
> created, and if to replace the yield() with a proper waitqueue (and
> how). Or if to take the risk of hitting a bit of scalability in the
> nopage page faults of processes, by rewriting the fix with a
> find_lock_page in the do_no_page handler, that would avoid the need of
> my new locking primitive.
> 
> Comments welcome thanks!

I'd be inclined to think a lock_page is not a big SMP scalability
problem because the struct page's cacheline(s) will be written to
several times in the process of refcounting anyway. Such a workload
would also be running into tree_lock as well.

Not that I have done any measurements.

-- 
SUSE Labs, Novell Inc.

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