Re: [PATCH] mm: add ___GFP_NOINIT flag which disables zeroing on alloc

[David Hildenbrand]

On 30.03.21 03:44, Hyunsoon Kim wrote:
> On Mon, Mar 29, 2021 at 12:53:48PM +0200, David Hildenbrand wrote:
> > On 29.03.21 07:29, Hyunsoon Kim wrote:
> > > This patch allows programmer to avoid zero initialization on page
> > > allocation even when the kernel config "CONFIG_INIT_ON_ALLOC_DEFAULT"
> > > is enabled. The configuration is made to prevent uninitialized
> > > heap memory flaws, and Android has applied this for security and
> > > deterministic execution times. Please refer to below.
> > >
> > > https://android-review.googlesource.com/c/kernel/common/+/1235132
> > >
> > > However, there is a case that the zeroing page memory is unnecessary
> > > when the page is used on specific purpose and will be zeroed
> > > automatically by hardware that accesses the memory through DMA.
> > > For instance, page allocation used for IP packet reception from Exynos
> > > modem is solely used for packet reception. Although the page will be
> > > freed eventually and reused for some other purpose, initialization at
> > > that moment of reuse will be sufficient to avoid uninitialized heap
> > > memory flaws. To support this kind of control, this patch creates new
> > > gfp type called ___GFP_NOINIT, that allows no zeroing at the moment
> > > of page allocation, called by many related APIs such as page_frag_alloc,
> > > alloc_pages, etc.
> > >
> > > Signed-off-by: Hyunsoon Kim <h10.kim@xxxxxxxxxxx>
> > > ---
> > >   include/linux/gfp.h | 2 ++
> > >   include/linux/mm.h  | 4 +++-
> > >   2 files changed, 5 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> > > index 8572a14..4ddd947 100644
> > > --- a/include/linux/gfp.h
> > > +++ b/include/linux/gfp.h
> > > @@ -58,6 +58,8 @@ struct vm_area_struct;
> > >   #else
> > >   #define ___GFP_NOLOCKDEP	0
> > >   #endif
> > > +#define ___GFP_NOINIT		0x1000000u
> > > +
> > >   /* If the above are modified, __GFP_BITS_SHIFT may need updating */
> > >   /*
> > > diff --git a/include/linux/mm.h b/include/linux/mm.h
> > > index 8ba4342..06a23bb 100644
> > > --- a/include/linux/mm.h
> > > +++ b/include/linux/mm.h
> > > @@ -2907,7 +2907,9 @@ static inline void kernel_unpoison_pages(struct page *page, int numpages) { }
> > >   DECLARE_STATIC_KEY_FALSE(init_on_alloc);
> > >   static inline bool want_init_on_alloc(gfp_t flags)
> > >   {
> > > -	if (static_branch_unlikely(&init_on_alloc))
> > > +	if (flags & ___GFP_NOINIT)
> > > +		return false;
> > > +	else if (static_branch_unlikely(&init_on_alloc))
> > >   		return true;
> > >   	return flags & __GFP_ZERO;
> > >   }
> >
> > We discussed that in the past - whatever leaves the buddy shall be
> > initialized. This is a security feature, not something random kernel modules
> > should be able to hack around.
> >
> > We also discussed back then to allow other allocators to eventually be able
> > to optimize in the future if we are sure it really makes sense. Then,
> > however, we need a new API that is not available to random modules, instead
> > of exposing ___GFP_NOINIT to anybody out there in the system.
> >
> > Nacked-by: David Hildenbrand <david@xxxxxxxxxx>
> >
> > --
> > Thanks,
> >
> > David / dhildenb
>
> If you don't mind, may i ask you exactly what security flaws you are expecting
> from uninitialized value allocation? I can think of below scenario:
>
> 1. Security related value is freed by security system.
> 2. Malicious module get allocation to the memory region that is freed by above.
> 3. Malicious module uses that uninitialized value, and breach the security.

I think one of the most important cases are "Content of process A is 
leaked via driver/mechanism X to process B". Or "Kernel content is 
leaked via driver/mechanism X to process Y".

> Could you please confirm that I am think in the right way? If so, isn't it
> possible to make the security system to zero on free? I am not talking about
> CONFIG_INIT_ON_FREE_DEFAULT_ON. I am just suggesting that isn't it better to
> make programs that generate important values to be forced to initialize on
> free, instead of making whole system to zeroing on alloc always, resulting
> in performance downgrade? I think this approach can make enhancement.

Well, it's not that easy. Then it's up to the freeing context to decide 
if a page should better be freed. Similarly, if you have a BUG (e.g., 
random put_page() from context X frees a user space page to the buddy) 
there, the whole security feature is - again - moot. That's why really 
only "init_on_free" vs "init_on_alloc" make sense - for anything that 
enters/leaves the buddy. As soon as you start poking holes you start 
opening the door for such security issues.

Enabling init_on_free has the downside that system boots gets slower, as 
everything that enters the buddy (== all memory) has to be zeroed out.

--
Thanks,

David / dhildenb