Re: [PATCH v2] rtw_security: fix cast to restricted __le32

[Jhih Ming Huang]

On Tue, Jun 15, 2021 at 1:03 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> On Mon, Jun 14, 2021 at 11:27:03PM +0800, Jhih Ming Huang wrote:
>
> > Thanks for your explanation.
> >
> > To clarify, even though it might be false positives in some senses,
> > following "hold the variable native-endian and check the conversion
> > done correctly"
> > is much easier than the other way. And it's exactly the current implementation.
> >
> > So it's better to keep the current implementation and ignore the
> > warnings, right?
>
> Umm...  If that's the case, the warnings should go away if you use
> cpu_to_le32() for conversions from native to l-e and le32_to_cpu()
> for conversions from l-e to native.
>
> IOW, the choice between those should annotate what's going on.
>
> In your case doing
>         *((u32 *)crc) = le32_to_cpu((__force __le32)~crc32_le(~0, payload, length - 4));
> is wrong - you have
> crc32_le(...) native-endian
> ~crc32_le(...) - ditto
> le32_to_cpu(~crc32_le(...)) - byteswapped native-endian on b-e, unchanged on
> l-e.  So result will be little-endian representation of ~crc32(...) in all
> cases.  IOW, it's cpu_to_le32(~crc32_le(...)), misannotated as native-endian
> instead of little-endian it actually is.
>
> Then you store that value (actually __le32) into *(u32 *)crc.  Seeing that
> crc is u8[4] there, that *(u32 *) is misleading - you are actually storing
> __le32 there (and, AFAICS, doing noting with the result).  The same story
> in rtw_tkip_decrypt(), only there you do use the result later.
>
> So just make it __le32 crc and
>         crc = cpu_to_le32(~crc32_le(~0, payload, length - 4));
> with
>                         if (crc[3] != payload[length - 1] || crc[2] != payload[length - 2] ||
>                             crc[1] != payload[length - 3] || crc[0] != payload[length - 4])
> turned into
>                         if (memcmp(&crc, payload + length - 4, 4) != 0)
> (or (crc != get_unaligned((__le32 *)(payload + length - 4))),
> for that matter, to document what's going on and let the damn thing
> pick the optimal implementation for given architecture).
>
> Incidentally, your secmicgetuint32() is simply get_unaligned_le32()
> and secmicputuint32() - put_unaligned_le32().  No need to reinvent
> that wheel...
>

Thanks for your comprehensive explanation.

I just sent the v3 PATCH, but I replied to this thread.
Should I create the other thread?

For the secmicgetuint32(), I am not the author of this function,
but you are right we should not reinvent the wheel.

Let's focus on sparse warning fixing in this commit.

thanks.

--jmhuang