On Sat, Aug 13, 2016 at 06:41:45PM +0200, Jean-Francois Moine wrote:
Some MMC devices as mmc2 in the A83T or mmc1 and mmc2 in the H3 have
a 'New Timing' mode.
When aware about this capability, and when this is possible (clock
rate great enough), the clock driver switches the MMC clock to the
'new mode', meaning that the phase delays are defined in the MMC
registers instead of in the clock registers.
To alert the MMC driver about this switch, the clock driver returns
the error code EPERM on calling clk_set_phase().
This patch makes the MMC driver to handle this returned code and to
activate or not the 'New Timing' mode on the MMC side.
Signed-off-by: Jean-Francois Moine <redacted>
---
Some explanations:
In the old timing, the phase delays are set in the clock only
(that's why there is a function clk_set_phase() which is called from
the MMC side).
In the new timing, the delays are in the MMC register SDXC_REG_NTSR
only.
The new timing works only when the clock rate is greater or equal
to 50MHz.
There are 2 flags saying that the new timing is used:
- the bit 'mode select' in the clock register, and
- the bit 'new timing' in the MMC register.
Both bits must be set/reset at the same time, otherwise the device
does not work (tested with wifi and eMMC in H3 and A83T boards).
So, some synchronization must exist.
The previous versions was using a DT property for the MMC and a flag
in the clock driver. This did work with a correct configuration
on both sides, but experiment showed that it was easy to do an error.
I still believe that we will need a property, at least to identify on
which we can try the new mode, and on which clocks it's irrelevant (at
least for the A33 and A83T).
However, I also believe we should make that mode switching explicit
through a function call, instead of relying on some side effect (of
some non-upstream code).
Maxime
--
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20160830/b503a2fa/attachment.sig>
On Tue, 30 Aug 2016 18:26:13 +0200
Maxime Ripard [off-list ref] wrote:
quoted
There are 2 flags saying that the new timing is used:
- the bit 'mode select' in the clock register, and
- the bit 'new timing' in the MMC register.
Both bits must be set/reset at the same time, otherwise the device
does not work (tested with wifi and eMMC in H3 and A83T boards).
So, some synchronization must exist.
The previous versions was using a DT property for the MMC and a flag
in the clock driver. This did work with a correct configuration
on both sides, but experiment showed that it was easy to do an error.
I still believe that we will need a property, at least to identify on
which we can try the new mode, and on which clocks it's irrelevant (at
least for the A33 and A83T).
As told above, setting the new mode on side (clock or MMC) and not on
the other one prevents the devices to work. Then, it is safer to have
the new mode capability flag only once for both sides.
Now, as the clocks are defined by memory tables and not by the DT, it
seems natural to have the flag on the clock side.
However, I also believe we should make that mode switching explicit
through a function call, instead of relying on some side effect (of
some non-upstream code).
Do you mean a direct call from the MMC driver to the clock driver?
--
Ken ar c'henta? | ** Breizh ha Linux atav! **
Jef | http://moinejf.free.fr/
On Tue, Aug 30, 2016 at 07:32:02PM +0200, Jean-Francois Moine wrote:
On Tue, 30 Aug 2016 18:26:13 +0200
Maxime Ripard [off-list ref] wrote:
quoted
quoted
There are 2 flags saying that the new timing is used:
- the bit 'mode select' in the clock register, and
- the bit 'new timing' in the MMC register.
Both bits must be set/reset at the same time, otherwise the device
does not work (tested with wifi and eMMC in H3 and A83T boards).
So, some synchronization must exist.
The previous versions was using a DT property for the MMC and a flag
in the clock driver. This did work with a correct configuration
on both sides, but experiment showed that it was easy to do an error.
I still believe that we will need a property, at least to identify on
which we can try the new mode, and on which clocks it's irrelevant (at
least for the A33 and A83T).
As told above, setting the new mode on side (clock or MMC) and not on
the other one prevents the devices to work. Then, it is safer to have
the new mode capability flag only once for both sides.
Now, as the clocks are defined by memory tables and not by the DT, it
seems natural to have the flag on the clock side.
You can look at it from two sides.
Either you try to switch to the new mode all the time, or you try to
do it only for MMC controllers (and their associated clocks) that
support it.
In the former case, you'll need to ignore a failure in the switch
(mostly if the clock doesn't support it) only for the MMC controllers
that do not have that mode. For the controllers that support it, you
cannot ignore that error anymore. So you have to have some way to
identify in which case you are.
And you need to have that in the former case too, so there's really no
way you can go without such a flag.
quoted
However, I also believe we should make that mode switching explicit
through a function call, instead of relying on some side effect (of
some non-upstream code).
Do you mean a direct call from the MMC driver to the clock driver?
Yes.
Maxime
--
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20161011/69218957/attachment.sig>