L2 cache

Sebastian Sumpf Sebastian.Sumpf at ...1...
Wed Nov 28 12:09:46 CET 2012


Hi Vasily,

On 11/28/2012 09:56 AM, Vasily A. Sartakov wrote:
> 
> Inspired by article about Genode and Pandaboard, we have tried to turn on L2 cache on same board. Two patches are attached. One of them - new L2 features for foc kernel, extracted from linux and u-boot, the other one - small example for testing cache. This example show increasing performance on  memory operations with block sized more then L1. ( We made 4 tests: without L2, block size equal L1, block size equal quarter of L2; with L2 for same blocks)     Based on this I make assumption what L2 work correct. But USB host driver is stuck on boot. There were problems of opposition L2 and USB, something has changed?

Very good that you have a look at this tricky problem and cool thing
that you found out that one has to use 'smc' calls to program the PL310
controller on Panda board. However this issue is not as trivial as it
seems. As you noticed, problems arise when using DMA (like the USB
driver). Device drivers that allocate DMA memory need to map this memory
uncached. The problem here is that core maps all its RAM from sigma0 as
cached memory, there currently is no way to get uncached RAM. So what
core does if someone requests an uncached dataspace (see 'alloc' in
'Ram_session'), is to simply map the pages uncached to the client, which
is supported by Fiasco.OC. The other thing it does is to clean the L1
cache to make sure that all data is in RAM (see
'base-foc/src/core/ram_session_support.cc').
When the L2 cache is enabled it needs to be cleaned as well. In theory
it would suffice to call 'l4_cache_dma_coherent' instead of
'l4_cache_clean_data'. Unfortunately the PL310 on Panda must be
programed via these funny 'smc' calls and the implementation for that is
missing in Fiasco.
What one would have to do, is to implement the 'Outer_cache' interface
in Fiasco.OC (see: 'src/kern/arm/outer-cache.cpp'). The interface
compromises the 'clean', 'flush', and 'invalidate' functions. We already
have some code for rev38, but did not have any time to finish the issue.
Maybe you could use https://github.com/skalk/genode/tree/omap4_fb as a
starting point, if you want to further investigate this issue.
By the way, another possibility to enable the L2 cache is to let U-Boot
do it.

Greetings and thanks for your efforts,

Sebastian


-- 
Sebastian Sumpf
Genode Labs

http://www.genode-labs.com · http://genode.org

Genode Labs GmbH · Amtsgericht Dresden · HRB 28424 · Sitz Dresden
Geschäftsführer: Dr.-Ing. Norman Feske, Christian Helmuth







More information about the users mailing list