Re: [patch] radix-tree: fix small lockless radix-tree bug

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 13 Jun 2008 05:03:45 +1000
Nick Piggin <nickpiggin@yahoo.com.au> wrote:

> Hi guys,
> 
> Although this doesn't seem like cause for alarm (as per the analysis),
> it may still be a good 2.6.26 candidate as we should have a few more
> weeks of testing left.
> 
> It should definitely go in -mm with the lockless pagecache patch.
> 
> When shrinking a radix-tree, we do it in a lockless manner by atomically
> switching the root pointer away from the redundant node (one that only
> has a single entry in the left most slot), and switching it over to its
> lone child.
> 
> Because a lockless lookup may have got a reference to the parent and be
> in the middle of deciding what to do with it while it is being swapped
> away for its child. For this reason, we also have to keep it around and
> in a valid state for the lookup to proceed and give a valid result, for
> at least an RCU grace period. So we need to keep the child in the left
> most slot there in case that is requested by the lookup.
> 
> This is all pretty standard RCU stuff. It is worth repeating because
> in my eagerness to obey the radix tree node constructor scheme, I had
> broken this by zeroing the radix tree node before the grace period.
> 
> Fix it by clearing those fields in the RCU callback. I would normally
> want to rip out the constructor entirely, but radix tree nodes are one
> of those places where they make sense (only few cachelines will be
> touched soon after allocation).
> 
> 
> This was never actually observed in any lockless pagecache testing or
> using the test harness, but as a rare problem testing my scalable vmap
> rewrite.
> 
> Fortunately, it is not a problem anywhere lockless pagecache is used in
> mainline kernels (pagecache probe is not a guarantee, and brd does not
> have concurrent lookups and deletes).
> 
> However, it would eventually pop up for someone using lockless pagecache :P
> 

OK, I give up.  A cannot spot what you actually changed amongst all the
code motion?