Re: [PATCH] netfilter: replace horrible hack with ksize()

[Pekka Enberg <penberg@cs.helsinki.fi>]

Hi Patrick,

Patrick McHardy wrote:
> > I think you are misunderstanding ksize() (see mm/slub.c::ksize() for 
> > example).
> 
> The ksize() description in mm/slab.c matches exactly what netfilter
> wants to do:

Agreed.

Patrick McHardy wrote:
> The initial allocation size is calculated as max(size, min slab size)
> and is stored as ext->alloc_size. When adding the first extension,

Yes, this part is correct, however...

> it allocates ext->alloc_size of memory and stores both the real amount
> of space used (ext->len) and the actual size (ext->real_len).
> When adding further extensions, it calculates the new total amount of
> space needed (newlen). If that is larger than the real amount of
> memory allocated (real_len), it reallocates.

...looking at nf_ct_ext_create() you do:

         *ext = kzalloc(real_len, gfp);
                        ^^^^^^^^
         if (!*ext)
                 return NULL;

         (*ext)->offset[id] = off;
         (*ext)->len = len;
         (*ext)->real_len = real_len;
                            ^^^^^^^^

You are storing the _object size_ (total amount of memory requested) and 
not the _buffer size_ (total amount of memory allocated). Keep in mind 
that object size < buffer size and that ksize() returns the latter.

Now continuing in __nf_ct_ext_add() you do:

        if (newlen >= ct->ext->real_len) {
                               ^^^^^^^^
                 new = kmalloc(newlen, gfp);
                 if (!new)
                         return NULL;

So you're comparing newlen to the object size and not the buffer size 
which is what you want and what ksize() and consequently my patch does.

Take a look at mm/util.c::krealloc(). It does exactly what you want 
modulo the RCU bits. My patch converts the netfilter code to follow the 
exact same semantics.

			Pekka