On Wed, Feb 21, 2024 at 06:29:17PM -0500, Kent Overstreet wrote:
> On Wed, Feb 21, 2024 at 03:05:32PM -0800, Kees Cook wrote:
> > On Wed, Feb 21, 2024 at 11:40:27AM -0800, Suren Baghdasaryan wrote:
> > > [...]
> > > +struct alloc_tag {
> > > + struct codetag ct;
> > > + struct alloc_tag_counters __percpu *counters;
> > > +} __aligned(8);
> > > [...]
> > > +#define DEFINE_ALLOC_TAG(_alloc_tag)
> > > \
> > > + static DEFINE_PER_CPU(struct alloc_tag_counters, _alloc_tag_cntr);
> > > \
> > > + static struct alloc_tag _alloc_tag __used __aligned(8)
> > > \
> > > + __section("alloc_tags") = {
> > > \
> > > + .ct = CODE_TAG_INIT,
> > > \
> > > + .counters = &_alloc_tag_cntr };
> > > [...]
> > > +static inline struct alloc_tag *alloc_tag_save(struct alloc_tag *tag)
> > > +{
> > > + swap(current->alloc_tag, tag);
> > > + return tag;
> > > +}
> >
> > Future security hardening improvement idea based on this infrastructure:
> > it should be possible to implement per-allocation-site kmem caches. For
> > example, we could create:
> >
> > struct alloc_details {
> > u32 flags;
> > union {
> > u32 size; /* not valid after __init completes */
> > struct kmem_cache *cache;
> > };
> > };
> >
> > - add struct alloc_details to struct alloc_tag
> > - move the tags section into .ro_after_init
> > - extend alloc_hooks() to populate flags and size:
> > .flags = __builtin_constant_p(size) ? KMALLOC_ALLOCATE_FIXED
> > : KMALLOC_ALLOCATE_BUCKETS;
> > .size = __builtin_constant_p(size) ? size : SIZE_MAX;
> > - during kernel start or module init, walk the alloc_tag list
> > and create either a fixed-size kmem_cache or to allocate a
> > full set of kmalloc-buckets, and update the "cache" member.
> > - adjust kmalloc core routines to use current->alloc_tag->cache instead
> > of using the global buckets.
> >
> > This would get us fully separated allocations, producing better than
> > type-based levels of granularity, exceeding what we have currently with
> > CONFIG_RANDOM_KMALLOC_CACHES.
> >
> > Does this look possible, or am I misunderstanding something in the
> > infrastructure being created here?
>
> Definitely possible, but... would we want this?
Yes, very very much. One of the worst and mostly unaddressed weaknesses
with the kernel right now is use-after-free based type confusion[0], which
depends on merged caches (or cache reuse).
This doesn't solve cross-allocator (kmalloc/page_alloc) type confusion
(as terrifyingly demonstrated[1] by Jann Horn), but it does help with
what has been a very common case of "use msg_msg to impersonate your
target object"[2] exploitation.
> That would produce a _lot_ of kmem caches
Fewer than you'd expect, but yes, there is some overhead. However,
out-of-tree forks of Linux have successfully experimented with this
already and seen good results[3].
> and don't we already try to collapse those where possible to reduce
> internal fragmentation?
In the past, yes, but the desire for security has tended to have more
people building with SLAB_MERGE_DEFAULT=n and/or CONFIG_RANDOM_KMALLOC_CACHES=y
(or booting with "slab_nomerge").
Just doing the type safety isn't sufficient without the cross-allocator
safety, but we've also had solutions for that proposed[4].
-Kees
[0] https://github.com/KSPP/linux/issues/189
[1]
https://googleprojectzero.blogspot.com/2021/10/how-simple-linux-kernel-memory.html
[2] https://www.willsroot.io/2021/08/corctf-2021-fire-of-salvation-writeup.html
https://google.github.io/security-research/pocs/linux/cve-2021-22555/writeup.html#exploring-struct-msg_msg
[3] https://grsecurity.net/how_autoslab_changes_the_memory_unsafety_game
[4]
https://lore.kernel.org/linux-hardening/[email protected]/
--
Kees Cook
