Re: [PATCH] Export symbol ksize()

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



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().

> > 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.

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. 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.

-- 
http://selenic.com : development and support for Mercurial and Linux


--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[Index of Archives]     [Kernel]     [Gnu Classpath]     [Gnu Crypto]     [DM Crypt]     [Netfilter]     [Bugtraq]

  Powered by Linux