Re: [PATCH v3 4/9] dm integrity: Remove VLA usage

From: Kees Cook
Date: Fri Jun 29 2018 - 17:56:46 EST


On Fri, Jun 29, 2018 at 1:43 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Fri, Jun 29, 2018 at 2:28 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
>> diff --git a/drivers/md/dm-integrity.c b/drivers/md/dm-integrity.c
>> index 86438b2f10dd..85e8ce1625a2 100644
>> --- a/drivers/md/dm-integrity.c
>> +++ b/drivers/md/dm-integrity.c
>> @@ -521,7 +521,12 @@ static void section_mac(struct dm_integrity_c *ic, unsigned section, __u8 result
>> }
>> memset(result + size, 0, JOURNAL_MAC_SIZE - size);
>> } else {
>> - __u8 digest[size];
>> + __u8 digest[SHASH_MAX_DIGESTSIZE];
>> +
>> + if (WARN_ON(size > sizeof(digest))) {
>> + dm_integrity_io_error(ic, "digest_size", -EINVAL);
>> + goto err;
>> + }
>
> I'm still slightly worried that some patches like this one could make
> things worse and lead to an actual stack overflow.

As in stack exhaustion? Yeah, this has been a concern of mine for the
crypto stuff because some combinations get BIG. My thinking has been
mainly that it means ALL cases will lead to a bad state instead of
only corner cases, which makes it easier to find and fix.

> You define SHASH_MAX_DIGESTSIZE
> as '512', which is still quite a lot to put on the kernel stack. The
> function also
> uses SHASH_DESC_ON_STACK(), so now you have two copies. Then you
> could call shash_final_unaligned(), which seems to put a third copy on
> the stack,
> so replacing each one with a fixed-size buffer adds quite a bit of bloat.
>
> Is there actually a digest that can be used in dm-integrity with more than 64
> byte output (matching JOURNAL_MAC_SIZE) here?

This conversion was following the existing check (PAGE_SIZE / 8), and
not via an analysis of alg.digestsize users. Let me double-check. For
predefined stuff, it looks like the largest is:

SKEIN1024_DIGEST_BIT_SIZE/8 == 128

I can drop this from 512 down to 128...

-Kees

--
Kees Cook
Pixel Security