I have been trying to move an 8360 based system from a 3.0 kernel to a
3.12 (on the way to 3.14 with ipipe/xenomai) kernel and encountered an
oops in the ucc_geth driver when using RTBI mode on one of the ucc
ports. I haven't managed to find any commits to of_mdio or ucc_geth or
fsl_pq_mdio that would appear to address this problem, so I believe it
is still present in the latest kernel, but have not confirmed that with
testing yet.
Commit 058112c7efc9ef43bb511c137293dddbe6e42908 appears to have broken
ucc support for tbi phy detection.
With the patch in place, I am unable to get the mdio bus to create phy
devices for the tbi phy in the ucc on an 8360e, and the ucc_geth driver
causes a kernel oops, while with the patch reverted, it does create them
and the driver comes up and works.
The tbi phy is needed when using a ucc in RTBI, TBI or SGMII mode.
I am not convinced that the tbi phy really behaves quite like a real phy,
which may be why get_phy_device does not work with it. Perhaps there
is a better way to deal with the tbi phy on the ucc for this purpose.
Certainly as it is, this patch has caused a regression though, although
probably not very many systems with ucc ports actually use one of the
affected modes so the damage isn't that great.
--
Len Sorensen
2014-12-18 19:49 GMT-08:00 Lennart Sorensen <[email protected]>:
> I have been trying to move an 8360 based system from a 3.0 kernel to a
> 3.12 (on the way to 3.14 with ipipe/xenomai) kernel and encountered an
> oops in the ucc_geth driver when using RTBI mode on one of the ucc
> ports. I haven't managed to find any commits to of_mdio or ucc_geth or
> fsl_pq_mdio that would appear to address this problem, so I believe it
> is still present in the latest kernel, but have not confirmed that with
> testing yet.
>
> Commit 058112c7efc9ef43bb511c137293dddbe6e42908 appears to have broken
> ucc support for tbi phy detection.
>
> With the patch in place, I am unable to get the mdio bus to create phy
> devices for the tbi phy in the ucc on an 8360e, and the ucc_geth driver
> causes a kernel oops, while with the patch reverted, it does create them
> and the driver comes up and works.
>
> The tbi phy is needed when using a ucc in RTBI, TBI or SGMII mode.
>
> I am not convinced that the tbi phy really behaves quite like a real phy,
> which may be why get_phy_device does not work with it. Perhaps there
> is a better way to deal with the tbi phy on the ucc for this purpose.
There are some comments in ucc_geth that also lead me to believe this
is a just a hack instead of a real Ethernet PHY device. Part of what I
think got broken is because of this comment:
/* Initialize TBI PHY interface for communicating with the
* SERDES lynx PHY on the chip. We communicate with this PHY
* through the MDIO bus on each controller, treating it as a
* "normal" PHY at the address found in the UTBIPA register. We assume
* that the UTBIPA register is valid. Either the MDIO bus code will set
* it to a value that doesn't conflict with other PHYs on the bus, or the
* value doesn't matter, as there are no other PHYs on the bus.
*/
In particular this one:
"Either the MDIO bus code will set
* it to a value that doesn't conflict with other PHYs on the bus, or the
* value doesn't matter, as there are no other PHYs on the bus."
and what Sebastian removed did exactly that, we used the special MDIO
broadcast address 0 to provide this "whatever". If this is such a
requirement from the ucc_geth driver and TBI PHYs, maybe we should
have this hack somewhere in the actual MDIO driver used by the
ucc_geth driver instead, or set a flag/read the PHY connection mode
and do this in drivers/of/of_mdio.c
>
> Certainly as it is, this patch has caused a regression though, although
> probably not very many systems with ucc ports actually use one of the
> affected modes so the damage isn't that great.
>
> --
> Len Sorensen
--
Florian
On Sat, Dec 20, 2014 at 09:08:51AM -0800, Florian Fainelli wrote:
> There are some comments in ucc_geth that also lead me to believe this
> is a just a hack instead of a real Ethernet PHY device. Part of what I
> think got broken is because of this comment:
>
> /* Initialize TBI PHY interface for communicating with the
> * SERDES lynx PHY on the chip. We communicate with this PHY
> * through the MDIO bus on each controller, treating it as a
> * "normal" PHY at the address found in the UTBIPA register. We assume
> * that the UTBIPA register is valid. Either the MDIO bus code will set
> * it to a value that doesn't conflict with other PHYs on the bus, or the
> * value doesn't matter, as there are no other PHYs on the bus.
> */
>
> In particular this one:
>
> "Either the MDIO bus code will set
> * it to a value that doesn't conflict with other PHYs on the bus, or the
> * value doesn't matter, as there are no other PHYs on the bus."
>
> and what Sebastian removed did exactly that, we used the special MDIO
> broadcast address 0 to provide this "whatever". If this is such a
> requirement from the ucc_geth driver and TBI PHYs, maybe we should
> have this hack somewhere in the actual MDIO driver used by the
> ucc_geth driver instead, or set a flag/read the PHY connection mode
> and do this in drivers/of/of_mdio.c
Well it used to be that it would look for an unused address and assign
that, but that was changed to just use 0 unless the dtb specified an
address (essentially making specifying the address in the dtb mandetory).
Unfortunately after this patch, specifying it in the dtb isn't enough,
and the ucc_geth actually hits a null pointer because the tbi phy no
longer exists.
Before commit 28d8ea2d568534026ccda3e8936f5ea1e04a86a1, the tbi address
was in fact _not_ 0. So yes it used to set it to a non conflicting
address, but no longer does. It used to pick the highest unused address
for the tbi. Now it uses 0 unless the dtb specifies the address.
Unfortunately no one ever fixed that comment. It appears to be entirely
inaccurate.
In the case of the board I am dealing with, setting the address to 0
when it isn't used (port is not in SGMII or RTBI mode) actually breaks
things because we have a switch chip at address 0 on the MDIO bus that we
now can't reach. Adding explicit addresses for the tbi phy on each ucc
solves that though so that is no big deal. The fact that the ucc that
needs to actually use the tbi phy for SGMII or RTBI though can't find it
anymore because it is no longer created does seem like a problem, and it
isn't being created no matter what the address (it is not 0 in this case).
So right now it is broken with ucc_geth segfaulting if you use SGMII or
RTBI mode. I would love a clean solution to fixing it, although for
now reverting this patch has solved the problem.
--
Len Sorensen