Re: [PATCH 02/14] dt-bindings: soc: milbeaut: Add Milbeaut trampoline description

From: Sugaya, Taichi
Date: Wed Dec 05 2018 - 05:30:45 EST


Hi,

On 2018/12/04 22:32, Rob Herring wrote:
On Tue, Dec 4, 2018 at 5:30 AM Sugaya, Taichi
<sugaya.taichi@xxxxxxxxxxxxx> wrote:

Hi

On 2018/12/04 0:49, Rob Herring wrote:
On Mon, Dec 3, 2018 at 1:42 AM Sugaya, Taichi
<sugaya.taichi@xxxxxxxxxxxxx> wrote:

Hi,

On 2018/11/30 17:16, Stephen Boyd wrote:
Quoting Sugaya, Taichi (2018-11-29 04:24:51)
On 2018/11/28 11:01, Stephen Boyd wrote:
Quoting Sugaya Taichi (2018-11-18 17:01:07)
create mode 100644 Documentation/devicetree/bindings/soc/socionext/socionext,m10v.txt

diff --git a/Documentation/devicetree/bindings/soc/socionext/socionext,m10v.txt b/Documentation/devicetree/bindings/soc/socionext/socionext,m10v.txt
new file mode 100644
index 0000000..f5d906c
--- /dev/null
+++ b/Documentation/devicetree/bindings/soc/socionext/socionext,m10v.txt
@@ -0,0 +1,12 @@
+Socionext M10V SMP trampoline driver binding
+
+This is a driver to wait for sub-cores while boot process.
+
+- compatible: should be "socionext,smp-trampoline"
+- reg: should be <0x4C000100 0x100>
+
+EXAMPLE
+ trampoline: trampoline@0x4C000100 {
Drop the 0x part of unit addresses.

Okay.


+ compatible = "socionext,smp-trampoline";
+ reg = <0x4C000100 0x100>;
Looks like a software construct, which we wouldn't want to put into DT
this way. DT doesn't describe drivers.
We would like to use this node only getting the address of the
trampoline area
in which sub-cores wait. (They have finished to go to this area in previous
bootloader process.)

Is this area part of memory, or a special SRAM? If it's part of memory,
I would expect this node to be under the reserved-memory node and
pointed to by some other node that uses this region. Could even be the
CPU nodes.

Yes, 0x4C000100 is a part of memory under the reserved-memory node. So
we would like to use the SRAM ( allocated 0x00000000 ) area instead.
BTW, sorry, the trampoline address of this example is simply wrong. We
were going to use a part of the SRAM from the beginning.



So should we embed the constant value in source codes instead of getting
from
DT because the address is constant at the moment? Or is there other
approach?


If it's constant then that also works. Why does it need to come from DT
at all then?

We think it is not good to embed constant value in driver codes and do
not have another way...
Are there better ways?

If this is just memory, can you use the standard spin-table binding in
the DT spec? There are some requirements like 64-bit values even on
32-bit machines (though this gets violated).

The spin-table seems to be used on only 64-bit arch. Have it ever worked
on 32-bit machine?

Yes.

And I would like not to use it because avoid violation.

The issue now that I remember is cpu-release-addr is defined to always
be a 64-bit value while some platforms made it a 32-bit value.
'cpu-release-addr' is also used for some other enable-methods.

Thanks.
OK, try to use the spin-table.

Best Regards,
Sugaya Taichi


Rob