by Kees Cook

[permalink] [raw]

Subject: Re: crypto: Kernel memory overwrite attempt detected to spans multiple pages

On Wed, Apr 10, 2019 at 4:12 PM Eric Biggers <[email protected]> wrote:
> You've explained *what* it does again, but not *why*. *Why* do you want
> hardened usercopy to detect copies across page boundaries, when there is no
> actual buffer overflow?

But that *is* how it determines it was a buffer overflow: "if you
cross page boundaries (of a non-compound allocation), it *is* a buffer
overflow". This assertion, however, is flawed because many contiguous
allocations are not marked as being grouped together when it reality
they were. It was an attempt to get allocation size information out of
the page allocator, similar to how slab can be queries about
allocation size. I'm open to improvements here, since it's obviously
broken in its current state. :)

--
Kees Cook

2019-04-11 01:40:40

On 17/04/2019 09:09, Russell King - ARM Linux admin wrote:
> On Tue, Apr 16, 2019 at 09:08:22PM -0700, Matthew Wilcox wrote:
>> On Mon, Apr 15, 2019 at 10:14:51PM -0500, Kees Cook wrote:
>>> On Mon, Apr 15, 2019 at 9:18 PM Matthew Wilcox <[email protected]> wrote:
>>>> I agree; if the crypto code is never going to try to go from the address of
>>>> a byte in the allocation back to the head page, then there's no need to
>>>> specify GFP_COMP.
>>>>
>>>> But that leaves us in the awkward situation where
>>>> HARDENED_USERCOPY_PAGESPAN does need to be able to figure out whether
>>>> 'ptr + n - 1' lies within the same allocation as ptr. Without using
>>>> a compound page, there's no indication in the VM structures that these
>>>> two pages were allocated as part of the same allocation.
>>>>
>>>> We could force all multi-page allocations to be compound pages if
>>>> HARDENED_USERCOPY_PAGESPAN is enabled, but I worry that could break
>>>> something. We could make it catch fewer problems by succeeding if the
>>>> page is not compound. I don't know, these all seem like bad choices
>>>> to me.
>>>
>>> If GFP_COMP is _not_ the correct signal about adjacent pages being
>>> part of the same allocation, then I agree: we need to drop this check
>>> entirely from PAGESPAN. Is there anything else that indicates this
>>> property? (Or where might we be able to store that info?)
>>
>> As far as I know, the page allocator does not store size information
>> anywhere, unless you use GFP_COMP. That's why you have to pass
>> the 'order' to free_pages() and __free_pages(). It's also why
>> alloc_pages_exact() works (follow all the way into split_page()).
>>
>>> There are other pagespan checks, though, so those could stay. But I'd
>>> really love to gain page allocator allocation size checking ...
>>
>> I think that's a great idea, but I'm not sure how you'll be able to
>> do that.
>
> However, we have had code (maybe historically now) that has allocated
> a higher order page and then handed back pages that it doesn't need -
> for example, when the code requires multiple contiguous pages but does
> not require a power-of-2 size of contiguous pages.

'git grep alloc_pages_exact' suggests it's not historical yet...

Robin.