Re: ima: use of radix tree cache indexing == massive waste of memory?

[Eric Paris <eparis@redhat.com>]

On Mon, 2010-10-18 at 20:19 +0200, Ingo Molnar wrote:
> * Eric Paris <eparis@redhat.com> wrote:
> 
> > On Mon, 2010-10-18 at 10:56 -0700, Linus Torvalds wrote:
> > > On Mon, Oct 18, 2010 at 9:48 AM, Eric Paris <eparis@redhat.com> wrote:
> > > >
> > > > 1) IMA uses radix trees which end up wasting 500 bytes per inode because
> > > > the key is too sparse.  I've got a patch which uses an rbtree instead
> > > > I'm testing and will send along shortly.  I found it funny working on
> > > > the patch to see that Documentation/rbtree.txt says "This differs from
> > > > radix trees (which are used to efficiently store sparse arrays and thus
> > > > use long integer indexes to insert/access/delete nodes)"  Which flys in
> > > > the face of this report.
> > > 
> > > Please. Look at the report more carefully.
> > > 
> > > The radix tree memory use is disgusting. Yes. But it is absolutely NOT
> > > sufficient to try to just fix that part. Go back, look at the original
> > > report email, and this line in particular:
> > > 
> > >    2235648 2069791  92%    0.12K  69864       32    279456K iint_cache
> > > 
> > > There's 2.2 million iint_cache allocations too, each 128 bytes in
> > > size. That's still a quarter _gigabyte_ of crap that adds zero value
> > > at all.
> > 
> > That was #2 in my list of things to fix:
> >
> > 2) IMA creates an entire integrity structure for every inode even when most or all 
> > of this structure will not be needed.
> > 
> > I'm stating with #1 since that was 2G of wasted space (thus far my switch to 
> > rbtree seems to be surviving an xfstest) so I expect to send the patch this 
> > afternoon.  #2 should attack the size of the iint_cache entries.  #3 should attack 
> > the scalability.  I'm certainly hoping I didn't miss part of the report....
> 
> I think it would be fair to argue that #2 is the thing that should be fixed first 
> and foremost - before touching any data structure details.
> 
> Because if you fix #2 then all the other items will become no-op to 99.9% of the 
> people who are affected by this bug today.
> 
> It's also probably a much simpler fix for -stable, so should be done first, etc.
> 
> If you do the data structure changes first then #2 will likely not be backportable 
> standalone and #1 will be risky to backport - creating nasty dependencies.

Good point.  I'll keep that in mind and possibly reorder.

-Eric