Re: *alloc API changes

[Matthew Wilcox <willy@infradead.org>]

On Mon, May 07, 2018 at 01:27:38PM -0700, Kees Cook wrote:
> On Mon, May 7, 2018 at 1:19 PM, Matthew Wilcox <willy@infradead.org> wrote:
> > Yes.  And today with kvmalloc.  However, I proposed to Linus that
> > kvmalloc() shouldn't allow it -- we should have kvmalloc_large() which
> > would, but kvmalloc wouldn't.  He liked that idea, so I'm going with it.
> 
> How would we handle size calculations for _large?

I'm not sure we should, at least initially.  The very few places which
need a large kvmalloc really are special and can do their own careful
checking.  Because, as Linus pointed out, we shouldn't be letting the
user ask us to allocate a terabyte of RAM.  We should just fail that.

let's see how those users pan out, and then see what we can offer in
terms of safety.

> > There are very, very few places which should need kvmalloc_large.
> > That's one million 8-byte pointers.  If you need more than that inside
> > the kernel, you're doing something really damn weird and should do
> > something that looks obviously different.
> 
> I'm CCing John since I remember long ago running into problems loading
> the AppArmor DFA with kmalloc and switching it to kvmalloc. John, how
> large can the DFAs for AppArmor get? Would an 8MB limit be a problem?

Great!  Opinions from people who'll use this interface are exceptionally
useful.

> And do we have any large IO or network buffers >8MB?

Not that get allocated with kvmalloc ... because you can't DMA map vmalloc
(without doing some unusual contortions).

> > but I thought of another problem with array_size.  We already have
> > ARRAY_SIZE and it means "the number of elements in the array".
> >
> > so ... struct_bytes(), array_bytes(), array3_bytes()?
> 
> Maybe "calc"? struct_calc(), array_calc(), array3_calc()? This has the
> benefit of actually saying more about what it is doing, rather than
> its return value... In the end, I don't care. :)

I don't have a strong feeling on this either.

> > Keeping our focus on allocations ... do we have plain additions (as
> > opposed to multiply-and-add?)  And subtraction?
> 
> All I've seen are just rare "weird" cases of lots of mult/add. Some
> are way worse than others:
> http://www.ozlabs.org/~akpm/mmotm/broken-out/exofs-avoid-vla-in-structures.patch
> 
> Just having the mult/add saturation would be lovely.

Ow.  My brain just oozed out of my ears.