Re: MDIO Debug Interface

[Vladimir Oltean <olteanv@gmail.com>]

Hi Tobias,

On Thu, Jul 09, 2020 at 10:47:54PM +0200, Tobias Waldekranz wrote:
> Hi netdev,
> 
> TL;DR: Is something like https://github.com/wkz/mdio-tools a good
> idea?
> 
> The kernel does not, as far as I know, have a low-level debug
> interface to MDIO devices. I.e. something equivalent to i2c-dev or
> spi-dev for example. The closest thing I've found are the
> SIOCG/SMIIREG ioctls, but they have several drawbacks:
> 
> 1. "Write register" is not always exactly that. The kernel will try to
>    be extra helpful and do things behind the scenes if it detects a
>    write to the reset bit of a PHY for example.
> 
> 2. Only one op per syscall. This means that is impossible to implement
>    many operations in a safe manner. Something as simple as a
>    read/mask/write cycle can race against an in-kernel driver.
> 
> 3. Addressing is awkward since you don't address the MDIO bus
>    directly, rather "the MDIO bus to which this netdev's PHY is
>    connected". This makes it hard to talk to devices on buses to which
>    no PHYs are connected, the typical case being Ethernet switches.
> 
> The kernel side of mdio-tools, mdio-netlink, tries to address these
> problems by adding a GENL interface with which a user can interact
> with an MDIO bus directly.
> 
> The user sends a program that the mdio-netlink module executes,
> possibly emitting data back to the user. I.e. it implements a very
> simple VM. Read/mask/write operations could be implemented by
> dedicated commands, but when you start looking at more advanced things
> like reading out the VLAN database of a switch you need to state and
> branching.
> 
> FAQ:
> 
> - A VM just for MDIO, that seems ridiculous, no?
> 
>   It does. But if you want to support the complex kinds of operations
>   that I'm looking for, without the kernel module having to be aware
>   of every kind of MDIO device in the world, I haven't found an easier
>   way.
> 
> - Why not use BPF?
> 
>   That could absolutely be one way forward, but the GENL approach was
>   easy to build out-of-tree to prove the idea. Its not obvious how it
>   would work though as BPF programs typically run async on some event
>   (probe hit, packet received etc.) whereas this is a single execution
>   on behalf of a user. So to what would the program be attached? The
>   output path is also not straight forward, but it could be done with
>   perf events i suppose.
> 
> My question is thus; do you think mdio-netlink, or something like it,
> is a candidate for mainline?
> 
> Thank you

I will let the PHY library maintainers comment about implementation
design choices made by mdio-netlink. However, I want to add a big "+1"
from my side for identifying the correct issues in the existing PHY
ioctls and doing something about it. I think the mainline kernel needs
this.
Please be aware that, if your mdio-netlink module, or something
equivalent to it, lands in mainline, QEMU/KVM is going to be one of its
users (for virtualizing an MDIO bus). So this is going to be more than
just for debugging.
And, while we're at it: context switches from a VM to a host are
expensive. And the PHY library polls around 5 MDIO registers per PHY
every second. It would be nice if your mdio-netlink module had some sort
of concept of "poll offload": just do the polling in the kernel side and
notify the user space only of a change.

Thanks,
-Vladimir