Re: [PATCH nft v2 11/18] netlink_linearize: round binop bitmask length up.

[Jeremy Sowden <jeremy@azazel.net>]

[-- Attachment #1: Type: text/plain, Size: 1171 bytes --]

On 2020-03-02, at 22:19:09 +0000, Jeremy Sowden wrote:
> In this example:
>
> nft --debug=netlink add rule ip t c ip dscp set ip dscp
> ip t c
>   [ payload load 2b @ network header + 0 => reg 1 ]
>   [ bitwise reg 1 = (reg=1 & 0x000003ff ) ^ 0x00000000 ]
>   [ payload load 1b @ network header + 1 => reg 2 ]
>   [ bitwise reg 2 = (reg=2 & 0x0000003c ) ^ 0x00000000 ]
>   [ bitwise reg 2 = ( reg 2 >> 0x00000002 ) ]
>   [ bitwise reg 2 = ( reg 2 << 0x00000002 ) ]
>   [ bitwise reg 1 = (reg=1 & 0x0000ffff ) ^ reg 2 ]
>   [ payload write reg 1 => 2b @ network header + 0 csum_type 1 csum_off 10 csum_flags 0x0 ]
>
> The mask at line 4 should be 0xfc, not 0x3c.
>
> Evaluation of the payload expression munges it from `ip dscp` to
> `(ip dscp & 0xfc) >> 2`.  When this AND expression is evaluated, its
> length is set to 6, the length of `ip dscp`.  When the bitwise netlink
> expression is generated, the length of the AND is used to generate the
> mask, 0x3f, used in combining the binop's.  The upshot of this is that
> the original mask gets mangled to 0x3c.  We can fix this by rounding
> the length of the mask to the nearest byte.

This is the wrong solution.

J.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 659 bytes --]