Where did the TI supplied kernel come from?

Kevin Hilman khilman at mvista.com
Wed Mar 28 11:16:55 CST 2007


Paul Jackson wrote:
> Kevin Hilman wrote:
>>
>> Do you have the DVEVM or the DVSDK?  Both are based on the MontaVista
>> Professional edition kernel, which is a 2.6.10-based kernel, plus lots
>> of patches, architectures and features as well as backported features
>> from newer kernels.   The DVEVM has an early-release of the MV kernel,
>> and the DVSDK has the official, supported product.
> 
> We have the DVEVM. It sounds like we ought to consider getting the DVSDK.
>>
>>> The reason I'm asking ...
>>>
>>> We really need to get the real time patches which have been applied to
>>> the main line since 2.6.10. The current kernel has them built in I
>>> believe.
>>
>> Which real-time patches are you speaking of?
>>
>> Actually, the real-time patches are still not part of mainline.  Several
>> sub-features of the -rt patch are in mainline (generic time-keeping,
>> hrtimers, dynamic tick, etc.) but the core realtime preemption feature
>> is still a separate patch maintained by Ingo Molnar[1].
>>
> Thanks for that - I had the impression that most of the real-time
> patches were in the main line. Once we move to 2.6.20 I suppose it would
> be fairly easy to apply the real time patches for preemption.

It would be easy to apply them, but there is still some DaVinci specific
work to do to enable support.  For example, DaVinci is not yet converted
over to use the generic timekeeping infrastructure which is a
prerequisite for support.   That is not very difficult, and my patches
in RMK's patch queue show how to do it for 3 other ARM platforms
(ixp4xx, Versatile and OMAP.)

>> The realtime work for ARM is not integrated an any one place yet either.
>>  I have submitted patches to the ARM kernel for all the core
>> functionality that is in mainline, but there would still be work to do
>> for validating this on DaVinci for 2.6.20+
>>
>> In other words, if you need working -rt support for the latest kernels,
>>  I can point you to the pieces you need, but you would still have a
>> decent amount of integration and validation work to do for DaVinci.
>>
>> Also note that some of the TI-specific software, particularily DSPLink,
>> is known not to work using full real-time preemption.  So there is not
>> only just kernel work to do.
> Do you know any more about the DSPLink problem with full real-time
> preemption ? Does it just need a rebuild or is it more serious than that?

First, it needs some changes to build against newer kernels (see
previous posts from myself and Alex Rier about running the demos on the
git kernels.)

But more importantly, it's internal locking and semaphore usage as well
as it's use of linux locking primitives needs a rethink to work cleanly
with (and without) the -rt patch.  It's been been a year or so since I
looked at this, so maybe the DSPLink folks have already worked through
the issues.

>>
>> The git tree is the work-in-progress of forward porting the 2.6.10
>> kernel to the latest kernel.  These drivers were not intentionally left
>> out, they just have not yet been forward ported.  Most of this work has
>> been done by MontaVista, but contributors are welcome!  Since you have
>> he source for 2.6.10, feel free to do the port and contribute.
>>
>> Remember that the git tree is an open-source project.  It is only as
>> good as its contributions (and contributors.)  It is not a TI or
>> MontaVista "supported" product, although both companies contribute.
>>
>> Hope that helps,
>>
>> Kevin
> 
> Thanks Kevin, that does help
> 
> Paul
> 
>>
>>
>> [1] http://people.redhat.com/mingo/realtime-preempt/
>> [2] RMK's patch system:
>> http://www.arm.linux.org.uk/developer/patches/section.php
>>
>>



More information about the Davinci-linux-open-source mailing list