Re: [RFC PATCH] net: phy/mdio: enable mmd indirect access through phy_mii_ioctl()

[Tobias Waldekranz]

On Thu, Nov 04, 2021 at 12:35, "Russell King (Oracle)" <linux@xxxxxxxxxxxxxxx> wrote:
> On Thu, Nov 04, 2021 at 12:17:47PM +0100, Tobias Waldekranz wrote:
>> On Wed, Nov 03, 2021 at 20:36, Andrew Lunn <andrew@xxxxxxx> wrote:
>> > On Wed, Nov 03, 2021 at 08:42:07PM +0200, Grygorii Strashko wrote:
>> >> 
>> >> 
>> >> On 03/11/2021 02:27, Andrew Lunn wrote:
>> >> > > > What i find interesting is that you and the other resent requester are
>> >> > > > using the same user space tool. If you implement C45 over C22 in that
>> >> > > > tool, you get your solution, and it will work for older kernels as
>> >> > > > well. Also, given the diverse implementations of this IOTCL, it
>> >> > > > probably works for more drivers than just those using phy_mii_ioctl().
>> >> > > 
>> >> > > Do you mean change uapi, like
>> >> > >   add mdio_phy_id_is_c45_over_c22() and
>> >> > >   flag #define MDIO_PHY_ID_C45_OVER_C22 0x4000?
>> >> > 
>> >> > No, i mean user space implements C45 over C22. Make phytool write
>> >> > MII_MMD_CTRL and MII_MMD_DATA to perform a C45 over C22.
>> >> 
>> >> Now I give up - as mentioned there is now way to sync User space vs Kernel
>> >> MMD transactions and so no way to get trusted results.
>> 
>> Except that there is a way: https://github.com/wkz/mdio-tools
>
> I'm guessing that this hasn't had much in the way of review, as it has
> a nice exploitable bug - you really want "pc" to be unsigned in
> mdio_nl_eval(), otherwise one can write a branch instruction that makes
> "pc" negative.

You are quite right, it never got that far as it was NAKed on principle
before that. I welcome the review, this is one of the reasons why I
would love to have it in mainline. Alternatively, if someone has a
better idea, I wouldn't mind adapting mdio-tools to whatever that
interface would be.

I agree that there should be much more rigorous checks around the
modification of the PC. I will get on that.

> Also it looks like one can easily exploit this to trigger any of your
> BUG_ON()/BUG() statements, thereby crashing while holding the MDIO bus
> lock causing a denial of service attack.

The idea is that this is pre-validated in mdio_nl_validate_insn. Each
instruction lists their acceptable argument types in mdio_nl_op_protos.

> I also see nothing that protects against any user on a system being
> able to use this interface, so the exploits above can be triggered by
> any user. Moreover, this lack of protection means any user on the
> system can use this interface to write to a PHY.

I was under the impression that specifying GENL_ADMIN_PERM in the
`struct genl_ops` would require the caller to hold CAP_NET_ADMIN?

> Given that some PHYs today contain firmware, this gives anyone access
> to reprogram the PHY firmware, possibly introducing malicious firmware.
>
> I hope no one is using this module in a production environment.

Thanks for your review.