Re: [PATCH] lib/idr.c: Zero memory properly in idr_remove_all

[Stefan Richter <stefanr@s5r6.in-berlin.de>]

Kristian Høgsberg wrote:
> On Wed, 2009-01-14 at 17:21 +0100, Stefan Richter wrote:
>> Kristian Høgsberg wrote:
>>> On Tue, 2009-01-13 at 14:48 -0800, Andrew Morton wrote:
>>>> What do we think of just removing the constructor and using
>>>> kmem_cache_zalloc()?
>>> We still need to zero out the idr_layer before returning it to the idr's
>>> internal free list.
>> I think so too.  But a more robust solution would IMO be to initialize
>> an idr_layer /before/ use, not /after/ use.  Will send a patch later.
> 
> The reason it was done this way is that normally, when the layers are
> returned to the free list they're already zero'ed out (since their
> elements have been removed one by one), so no need to do this on later
> re-initialization or when freeing.  It's a silly little
> sup-optimization.

Ah, right.

> idr_remove_all() is different in that it doesn't
> incrementally zero out the layer, and so it has to do it using a memset.
> 
> If you want rework how this works, I'd suggest just not using the free
> list except for in the idr_pre_get()+idr_get() sequence.  When a layer
> is no longer used, just free it, don't put it back on the free list.
> And use kmem_cache_zalloc() in idr_pre_get() as Andrew suggests.

Wait a moment:  Nadia's change (which introduced the bug) also already
implements your suggestion.  idr_remove_all feeds to the kmem cache now,
not to the free list anymore.

So, Andrew, I take back my assertion that the sequence

	idr_remove_all(idr);

	if (idr_pre_get(idr, GFP_KERNEL))
		id = idr_get_new(idr, p, h);

would be unsafe.  Your fix has this covered as well.
-- 
Stefan Richter
-=====-==--= ---= -===-
http://arcgraph.de/sr/