Re: [dm-devel] [PATCH v3 9/9] crypto: shash: Remove VLA usage in unaligned hashing
From: Kees Cook
Date: Mon Jul 02 2018 - 13:34:56 EST
On Sun, Jul 1, 2018 at 10:20 AM, Eric Biggers <ebiggers3@xxxxxxxxx> wrote:
> On Sun, Jul 01, 2018 at 10:04:59AM -0700, Kees Cook wrote:
>> On Sat, Jun 30, 2018 at 12:03 AM, Eric Biggers <ebiggers3@xxxxxxxxx> wrote:
>> > On Thu, Jun 28, 2018 at 05:28:43PM -0700, Kees Cook wrote:
>> >> @@ -88,11 +81,13 @@ static int shash_update_unaligned(struct shash_desc *desc, const u8 *data,
>> >> unsigned long alignmask = crypto_shash_alignmask(tfm);
>> >> unsigned int unaligned_len = alignmask + 1 -
>> >> ((unsigned long)data & alignmask);
>> >> - u8 ubuf[shash_align_buffer_size(unaligned_len, alignmask)]
>> >> - __aligned_largest;
>> >> + u8 ubuf[MAX_ALGAPI_ALIGNMASK + 1];
>> >> u8 *buf = PTR_ALIGN(&ubuf, alignmask + 1);
>> >> int err;
>> >> + if (WARN_ON(buf + unaligned_len > ubuf + sizeof(ubuf)))
>> >> + return -EINVAL;
>> >> +
>> > How is 'ubuf' guaranteed to be large enough? You removed the __aligned
>> > attribute, so 'ubuf' can have any alignment. So the aligned pointer 'buf' may
>> > be as high as '&ubuf[alignmask]'. Then, up to 'alignmask' bytes of data will be
>> > copied into 'buf'... resulting in up to '2 * alignmask' bytes needed in 'ubuf'.
>> > But you've only guaranteed 'alignmask + 1' bytes.
>> Hm, good point. Adding __aligned(MAX_ALGAPI_ALIGNMASK + 1) looks to
>> fix this, yes?
>> Also, if __aligned() is used here, can't PTR_ALIGN() be dropped? (I
>> think you pointed this out earlier.)
> Sure, I'm just not sure whether __aligned() with such a large alignment is
> guaranteed to work on stack variables on all architectures. See e.g.
That's terrible. :( That seems like a compiler bug, but okay.
>> Also, is "unaligned_len" being calculated correctly? Let's say
>> alignmask is 63. If data is binary ...111111, then unaligned_len will
>> be 64 - 63 == 1, which is fine: we copy 1 byte out, bump the address
>> by 1, and we're happily aligned to ...000000. If data is ...000000,
>> then unaligned_len will be 64. But it should be 0. Shouldn't this be:
>> unsigned int unaligned_len;
>> unaligned_len = (unsigned long)data & alignmask;
>> if (unaligned_len)
>> unaligned_len = alignmask + 1 - unaligned_len;
>> And then ubuf only needs to be MAX_ALGAPI_ALIGNMASK, without the +1?
> shash_update_unaligned() is only called when 'data & alignmask'.
> Similarly with shash_final_unaligned().
Ah! I see that now.
> Though, calculating 'unaligned_len' could be simplified to
> unsigned int unaligned_len = -(unsigned long)data & alignmask;
> which works either way.
So, since we can't depend on __aligned() working, I'll just keep the
PTR_ALIGN and add MAX_ALGAPI_ALIGNMASK to each array. That'll be less
memory-efficient, but it'll actually get aligned correctly.