Re: [PATCH] nfsd: prevent integer overflow in decode_cb_compound4res()

[Chuck Lever <chuck.lever@oracle.com>]

On Mon, Sep 16, 2024 at 06:14:15PM +0300, Dan Carpenter wrote:
> If "length" is >= U32_MAX - 3 then the "length + 4" addition can result
> in an integer overflow.  The impact of this bug is not totally clear to
> me, but it's safer to not allow the integer overflow.
> 
> There is also some math in xdr_inline_decode() which could overflow, so
> really it's ">= U32_MAX - 7" which is problematic.  Let's just check
> against INT_MAX and make it easy.
> 
> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
> ---
>  fs/nfsd/nfs4callback.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/fs/nfsd/nfs4callback.c b/fs/nfsd/nfs4callback.c
> index 43b8320c8255..12b44c9246d1 100644
> --- a/fs/nfsd/nfs4callback.c
> +++ b/fs/nfsd/nfs4callback.c
> @@ -317,6 +317,8 @@ static int decode_cb_compound4res(struct xdr_stream *xdr,
>  	hdr->status = be32_to_cpup(p++);
>  	/* Ignore the tag */
>  	length = be32_to_cpup(p++);
> +	if (unlikely(length > INT_MAX))
> +		goto out_overflow;

I think we assume (wrongly) that xdr_inline_decode() will kick
back any length request that is larger than the xdr_stream's
xdr_buf. Here, the test could be the same: any @length value that is
larger than xdr->buf->len is bogus.

Another way to avoid the overflow would be to split the decode of
the tag and the following op count field. Untested:

        length = be32_to_cpup(p);
        p = xdr_inline_decode(xdr, length);
        if (unlikely(p == NULL))
                goto out_overflow;
	if (xdr_stream_decode_u32(xdr, &hdr->nops) < 0)
                goto out_overflow;

I have generally been using this approach in other NFSD XDR decoding
functions.

I wonder how many other xdr_inline_decode() call sites have a
similar issue.


>  	p = xdr_inline_decode(xdr, length + 4);
>  	if (unlikely(p == NULL))
>  		goto out_overflow;
> -- 
> 2.45.2
> 

-- 
Chuck Lever