Device tree bindings for linux ramoops use

[Mitch Bradley]

On 1/6/2012 12:55 PM, Olof Johansson wrote:
> On Fri, Jan 6, 2012 at 2:38 PM, Rob Herring<robherring2 at gmail.com>  wrote:
>>
>>
>> On 01/06/2012 04:05 PM, Grant Likely wrote:
>>> On Fri, Jan 06, 2012 at 05:53:57PM +0000, Jamie Iles wrote:
>>>> On Fri, Jan 06, 2012 at 09:47:22AM -0800, Olof Johansson wrote:
>>>>> On Fri, Jan 6, 2012 at 8:58 AM, Jamie Iles<jamie at jamieiles.com>  wrote:
>>>>>> On Fri, Jan 06, 2012 at 08:28:51AM -0800, Olof Johansson wrote:
>>>>>>> On Thu, Jan 5, 2012 at 11:22 PM, Mitch Bradley<wmb at firmworks.com>  wrote:
>>>>>>>>
>>>>>>>> On 1/5/2012 6:39 PM, Olof Johansson wrote:
>>>>>>>>>
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> I'm considering how to best describe the data that ramoops needs in
>>>>>>>>> the device tree.
>>>>>>>>>
>>>>>>>>> The idea is really about describing a memory area that is (likely to>  >>  >>  be) nonvolatile across reboots. Said area is not to be included in the
>>>>>>>>> regular memory map of the system (i.e. not covered by /memory).
>>>>>>>>>
>>>>>>>>> I have a few options on where to do it. It's not really a hardware
>>>>>>>>> device per se, so it's a gray area for the device tree alltogether.
>>>>>>>>>
>>>>>>>>> How about something like?
>>>>>>>>>
>>>>>>>>> compatible = "linux,ramoops"
>>>>>>>>> linux,ramoops-start =<start address of preserved ram>
>>>>>>>>> linux,ramoops-size = ...
>>>>>>>>> linux,ramoops-record-size = ...
>>>>>>>>> linux,ramoops-include-oopses = ... (this one is a bit of a corner
>>>>>>>>> case, it's truly a software setting -- probably leave it out)
>>>>>>>>>
>>>>>>>>> Anybody have a better idea?
>>>>>>>>
>>>>>>>>
>>>>>>>> If it is addressable, it should appear as a device node underneath the node
>>>>>>>> that creates the address space in which it appears, and the start and size
>>>>>>>> should be described by a "reg" property.
>>>>>>>
>>>>>>> A yes, of course.
>>>>>>>
>>>>>>> I got on the wrong track due to the lack of use of resources in the
>>>>>>> linux platform_driver.
>>>>>>
>>>>>> But you still need some ramoops specific configuration though right?
>>>>>> Could this be represented with a generic binding for the onchip RAM as
>>>>>> has already proposed then inside the chosen node something like:
>>>>>>
>>>>>>         chosen {
>>>>>>                 ramoops {
>>>>>>                         linux,ramoops-record-size =<12>;
>>>>>>                         linux,ramoops-include-oopses =<1>;
>>>>>>                         /* phandle to ram, offset, size */
>>>>>>                         linux,ramoops-ram =<&iram 0x1000 0x200>;
>>>>>>                 };
>>>>>>         };
>>>>>>
>>>>>> to decouple the runtime configuration from the hardware binding?
>>>>>
>>>>> Only the ramoops-include-oopses is really the runtime configuration,
>>>>> so that alone in /chosen could be a good idea. But I would rather have
>>>>> the "partition" described as a device with a compatible field that the
>>>>> driver can bind against.
>>>>
>>>> I can see why that would be nice too, but to me this feels different to
>>>> say MTD partitions as it really is Linux specific and it doesn't seem
>>>> unreasonable that someone may want to include ramoops support when
>>>> debugging something, but for another application, use the whole of the
>>>> onchip RAM as a buffer.  Requiring modifications for the DT on identical
>>>> hardware platforms but different applications doesn't feel quite right
>>>> to me.  Seeing as chosen is special that doesn't feel too bad.
>>>
>>> I agree.  This is pretty low level core stuff.  The actual ramoops
>>> region could be anywhere as Olof says; in regular ram, in sram,
>>> somewhere else, but the common case is really just memory mapped ram
>>> that Linux needs to be told "hands off".  I would be fine with
>>> properties in /chosen and using /memreserve/ sections to inhibit the
>>> kernel from using them if they are in main memory.  There can still be
>>> nodes for the actual device, but that doesn't need to be explicitly
>>> connected to the ramoops properties in /chosen.  Implicitly by base
>>> address is fine by me.
>>>
>>
>> Using actual system RAM will not work. Ramoops uses ioremap which is
>> forbidden on RAM (at least for ARM). /memreserve/ doesn't unmap the RAM,
>> but just doesn't allocate it. So you would have to lie to Linux and save
>> some RAM. There's also the issue that most likely main memory is not
>> preserved across resets.
>
> The way to handle this is to have u-boot or whatever firmware you use,
> completely remove that range from the memory node. Or leave it out of
> the memory map.
>
> In practice, we have had excellent results with regular RAM contents
> being preserved across reboots. I was skeptical at first as well, but
> it's not a problem in reality. Of course, firmware needs to be
> adjusted such that it does not clear that memory range on reboot.
> Chrome OS uses this both on x86 and ARM.


DRAM tends to hold its data for a good fraction of a second, so if the 
memory controller is quickly re-initialized to the point of running 
refresh cycles, you win.

You also have to worry about memory change due to auto-probing for the 
memory size.

But, yeah, it can usually be made to work if you control the firmware.


>
>> So practically speaking, we are talking about an auxiliary RAM device.
>
> Sure, it's a separate region of RAM that is not used by the rest of the system.
>
>> I personally like the partition scheme and then device nodes like video
>> accelerators or audio h/w can be given portions of the RAM.
>
> Me too.
>
>
> -Olof
> _______________________________________________
> devicetree-discuss mailing list
> devicetree-discuss at lists.ozlabs.org
> https://lists.ozlabs.org/listinfo/devicetree-discuss
>
>