Re: [PATCH] Export symbol ksize()

[Andrew Morton]

On Sun, 15 Feb 2009 17:49:41 -0600 Matt Mackall <mpm@xxxxxxxxxxx> wrote:

> On Sun, 2009-02-15 at 13:55 -0800, Andrew Morton wrote:
> > On Sun, 15 Feb 2009 15:43:14 -0600 Matt Mackall <mpm@xxxxxxxxxxx> wrote:
> > 
> > > On Sun, 2009-02-15 at 13:36 -0800, Andrew Morton wrote:
> > > > On Thu, 12 Feb 2009 17:55:04 +0200 Pekka Enberg <penberg@xxxxxxxxxxxxxx> wrote:
> > > > 
> > > > > On Thu, Feb 12, 2009 at 12:45:21PM +0200, Pekka Enberg wrote:
> > > > > > > 
> > > > > > > Because the API was being widely abused in the nommu code, for example.
> > > > > > > I'd rather not add it back for this special case which can be handled
> > > > > > > otherwise.
> > > > > 
> > > > > On Thu, 2009-02-12 at 18:50 +0800, Herbert Xu wrote:
> > > > > > I'm sorry but that's like banning the use of heaters just because
> > > > > > they can abused and cause fires.
> > > > > > 
> > > > > > I think I've said this to you before but in networking we very much
> > > > > > want to use ksize because the standard case of a 1500-byte packet
> > > > > > has loads of extra room given by kmalloc which all goes to waste
> > > > > > right now.
> > > > > > 
> > > > > > If we could use ksize then we can stuff loads of metadata in that
> > > > > > space.
> > > > > 
> > > > > OK, fair enough, I applied Kirill's patch. Thanks.
> > > > > 
> > > > 
> > > > Could we please have more details regarding this:
> > > > 
> > > > > The ksize() function is not exported to modules because it has non-standard
> > > > > behavour across different slab allocators. 
> > > > 
> > > > How does the behaviour differ?  It this documented?  Can we fix it?
> > > 
> > > SLAB and SLUB support calling ksize() on objects returned by
> > > kmem_cache_alloc.
> > > 
> > > SLOB only supports it on objects from kmalloc. This is because it does
> > > not store any size or type information in kmem_cache_alloc'ed objects.
> > > Instead, it infers them from the cache argument.
> > 
> > OK.  This is really bad, isn't it?
> 
> No. There are very few ksize callers and very few of those are making
> this particular category error.
> 
> And it -is- a category error. The fact that kmalloc is implemented on
> top of kmem_cache_alloc is an implementation detail that callers should
> not assume. They shouldn't call kfree() on kmem_cache_alloc objects
> (even though it might just happen to work), nor should they call
> ksize().

But they could call a new kmem_cache_size(cachep, obj)?

> > > Ideally SLAB and SLUB would complain about using ksize inappropriately
> > > when debugging was enabled.
> > > 
> > 
> > OK, thanks.
> > 
> > Ideally we would support ksize() for both kmalloc() and
> > kmem_cache_alloc() memory across all implementations.
> 
> There's never a good reason to call ksize on a kmem_cache_alloced
> object. You -must- statically know what type of object you have already
> to be able to free it later with kmem_cache_free, ergo, you can
> statically know how big it is too.

But kmem_cache_size() would tell you how much extra secret memory there
is available after the object?

How that gets along with redzoning is a bit of a mystery though.

The whole concept is quite hacky and nasty, isn't it?.  Does
networking/crypto actually show any gain from pulling this stunt?

> Another alternative to the above is to throw sparse at it, and have it
> track what allocators a pointer might have come through. 
> 
> But as far as I'm aware, there's only been one actual bug in this area:
> nommu was calling ksize on pointers of all kinds, including stuff
> allocated at compile time.
> 
> > Gee this sucks.  Biggest mistake I ever made.  Are we working hard
> > enough to remove some of these sl?b implementations?  Would it help if
> > I randomly deleted a couple?
> 
> Again, I think there's a strong argument for having two. We can't
> reasonably expect one allocator to work well on supercomputers and
> phones.

We can't reasonably expect an OS to work well on supercomputers and
phones ;)  It's a matter of how much person-power gets tossed at it.

> One will likely value performance significantly higher than
> memory usage and vice-versa.
> 
> I think most of the pain here is actually peripheral. SLUB in particular
> has churned a lot of interfaces. But we would have had that had we
> instead decided to throw a lot of effort into making SLAB better.

hm.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/