[PATCH 00/10] HYP mode boot and cache manager updates for ARM and RaspberryPi specially.

Alan Cudmore alan.cudmore at gmail.com
Mon Jul 4 22:29:09 UTC 2016


I updated my RTEMS git and rebuilt the Pi and Pi2 BSPs.
Pi 1 works, but Pi 2 does not boot.

Pi 1 has a speed up on the dhrystone command:
I get 90000 dhrystones per second.

When I run your github pi branch, the Pi2 works with both the new and old firmware. 
The new firmware Pi 2 dhrystone command is 26315 dhrystones per second
The old firmware Pi 2 dhrystone command is 45454 dhrystones per second.
So something in the firmware and caches is different, I can double check the configuration.

But first, I will have to try to figure out the differences between the RTEMS git and your github and see why the Pi 2 is not booting for me.

Thanks,
Alan


> On Jul 4, 2016, at 4:17 PM, Pavel Pisa <pisa at cmp.felk.cvut.cz> wrote:
> 
> Hello Alan and Rohini Kulkarni,
> 
> On Monday 04 of July 2016 18:47:09 Alan Cudmore wrote:
>> Hi Pavel,
>> I can confirm that your rtems-rpi-devel branch from your github account
>> works on the Raspberry Pi A+, Raspberry Pi Zero, and Raspberry Pi 2 with
>> the latest firmware.
>> 
>> I did notice that the Dhrystone benchmark runs 3 times faster on the Pi 1
>> than the Pi 2, but that could be the cache configuration.
> 
> Is this problem of the new firmware only?
> If answer is yes, then it can be problem with providing
> permission to control cache from SVC mode. Then some
> register setup before switching from HYP mode could be required.
> 
> If it is not related to new hardware then I think that
> there is no change in caching attributes introduced
> by all my changes and experience of Rohini Kulkarni working
> on RPi2 SMP project in frame of GSoC 2015 could help.
> Code is on GitHub, 
> 
>  https://github.com/krohini1593/rtems
> 
> which is not merged to RTEMS mainline yet.
> 
> There are some changes to change cache configuration.
> 
> I have tried some changes in bsp_start_hook_1
> 
> +#include <bsp/arm-cp15-start.h>
> 
> 
> -  bsp_memory_management_initialize();
> +  uint32_t ctrl;
> 
> +  /* turn mmu off first in case it's on */
> +  ctrl = arm_cp15_start_setup_mmu_and_cache(
> +    ARM_CP15_CTRL_M | ARM_CP15_CTRL_A,  /* clear - mmu off */
> +    ARM_CP15_CTRL_AFE
> +  );
> +
> +  ctrl |= ARM_CP15_CTRL_Z;
> +
> +  arm_cp15_start_setup_translation_table_and_enable_mmu_and_cache(
> +    ctrl,
> +    (uint32_t *) bsp_translation_table_base,
> +    ARM_MMU_DEFAULT_CLIENT_DOMAIN,
> +    &arm_cp15_start_mmu_config_table[0],
> +    arm_cp15_start_mmu_config_table_size
> +  );
> 
> to ensure that ARM_CP15_CTRL_Z is set for branch predictor.
> But I have tried some modifications in BSP and I have observed
> no changes. RPi1 seems to run about 4 times faster in my case.
> 
> So it is probable that at cache is not enabled. May be only L2???
> 
> There can be mitchmatch between configured MMU table format/bit use
> and the way bits are filled. I have not time to recheck that
> yet but RPi2 code matches Cyclone Cortex-A9 BSP in CPU setup
> so it should be OK if there is not some Broadcom specific control
> register.
> 
> Best wishes,
> 
>              Pavel




More information about the devel mailing list