Re: [RFC net-next PATCH 00/13] Add PCS core support

[Sean Anderson <sean.anderson@linux.dev>]

On 4/7/25 13:21, Christian Marangi (Ansuel) wrote:
> Il giorno lun 7 apr 2025 alle ore 19:00 Sean Anderson
> <sean.anderson@linux.dev> ha scritto:
>>
>> On 4/7/25 12:46, Christian Marangi (Ansuel) wrote:
>> > Il giorno lun 7 apr 2025 alle ore 18:33 Sean Anderson
>> > <sean.anderson@linux.dev> ha scritto:
>> >>
>> >> On 4/7/25 12:27, Kory Maincent wrote:
>> >> > On Thu,  3 Apr 2025 14:18:54 -0400
>> >> > Sean Anderson <sean.anderson@linux.dev> wrote:
>> >> >
>> >> >> This series adds support for creating PCSs as devices on a bus with a
>> >> >> driver (patch 3). As initial users,
>> >> >>
>> >> >> - The Lynx PCS (and all of its users) is converted to this system (patch 5)
>> >> >> - The Xilinx PCS is broken out from the AXI Ethernet driver (patches 6-8)
>> >> >> - The Cadence MACB driver is converted to support external PCSs (namely
>> >> >>   the Xilinx PCS) (patches 9-10).
>> >> >>
>> >> >> The last few patches add device links for pcs-handle to improve boot times,
>> >> >> and add compatibles for all Lynx PCSs.
>> >> >>
>> >> >> Care has been taken to ensure backwards-compatibility. The main source
>> >> >> of this is that many PCS devices lack compatibles and get detected as
>> >> >> PHYs. To address this, pcs_get_by_fwnode_compat allows drivers to edit
>> >> >> the devicetree to add appropriate compatibles.
>> >> >
>> >> > I don't dive into your patch series and I don't know if you have heard about it
>> >> > but Christian Marangi is currently working on fwnode for PCS:
>> >> > https://lore.kernel.org/netdev/20250406221423.9723-1-ansuelsmth@gmail.com
>> >> >
>> >> > Maybe you should sync with him!
>> >>
>> >> I saw that series and made some comments. He is CC'd on this one.
>> >>
>> >> I think this approach has two advantages:
>> >>
>> >> - It completely solves the problem of the PCS being unregistered while the netdev
>> >>   (or whatever) is up
>> >> - I have designed the interface to make it easy to convert existing
>> >>   drivers that may not be able to use the "standard" probing process
>> >>   (because they have to support other devicetree structures for
>> >>   backwards-compatibility).
>> >>
>> >
>> > I notice this and it's my fault for taking too long to post v2 of the PCS patch.
>> > There was also this idea of entering the wrapper hell but I scrapped that early
>> > as I really feel it's a workaround to the current problem present for
>> > PCS handling.
>>
>> It's no workaround. The fundamental problem is that drivers can become
>> unbound at any time, and we cannot make consumers drop their references.
>> Every subsystem must deal with this reality, or suffer from
>> user-after-free bugs. See [1-3] for discussion of this problem in
>> relation to PCSs and PHYs, and [4] for more discussion of my approach.
>>
>> [1] https://lore.kernel.org/netdev/YV7Kp2k8VvN7J0fY@shell.armlinux.org.uk/
>> [2] https://lore.kernel.org/netdev/20220816163701.1578850-1-sean.anderson@seco.com/
>> [3] https://lore.kernel.org/netdev/9747f8ef-66b3-0870-cbc0-c1783896b30d@seco.com/
>> [3] https://lpc.events/event/17/contributions/1627/
>>
>> > And the real problem IMHO is that currently PCS handling is fragile and with too
>> > many assumptions. With Daniel we also discussed backwards-compatibility.
>> > (mainly needed for mt7621 and mt7986 (for mediatek side those are the 2
>> > that slipped in before it was correctly complained that things were
>> > taking a bad path)
>> >
>> > We feel v2 permits correct support of old implementations.
>> > The ""legacy"" implementation pose the assumption that PCS is never removed
>> > (unless the MAC driver is removed)
>> > That fits v2 where a MAC has to initially provide a list of PCS to
>> > phylink instance.
>>
>> And what happens when the driver is unbound from the device and suddenly
>> a PCS on that list is free'd memory but is in active use by a netdev?
>>
> 
> driver bug for not correctly implementing the removal task.
> 
> The approach is remove as provider and call phylink removal phase
> under rtnl lock.
> We tested this with unbind, that is actually the main problem we are
> trying to address
> and works correctly.

OK, so this is a different approach since your last submission. Please
CC me on your series.

- Fundamentally this is going to make backwards compatibility very
  difficult, since your approach cannot work with mac_select_pcs. How
  are you going to handle the case of MAC-internal PCSs? Are you going
  to make them create a swnode and bind to it just to create a PCS for
  e.g. MMIO registers? And how is the MAC supposed to know how to select
  the PCS? From what I can tell you don't even notify the MAC about
  which PCS it's using.

  I considered an approach like this, where the phylink would be in the
  driver's seat (so to speak), but I decided not to persue it due to
  the problems listed above. A lot of PCSs are tightly-integrated with
  their MACs, so it does not make sense to introduce this little
  coupling. I think it is better to let the MAC select the PCS e.g.
  based on the phy interface. This tends to be a few lines of code for
  the MAC and saves so much complexity in phylink.

  I think you should try doing the macb and lynx conversions for your
  approach. It will make the above problems obvious.

- Your approach is very intrusive. There are lots of changes all over
  phylink across several patches and it is hard to verify all the
  assumptions. Whereas a wrapper keeps everything contained to one file,
  and most of the functions can be evaluated independently.

--Sean