Fiasco.OC performance issues

[Sebastian Sumpf]

On 01/29/2013 12:09 AM, Adam Lackorzynski wrote:
> On Fri Jan 25, 2013 at 00:00:26 +0100, Sebastian Sumpf wrote:
>> On 01/17/2013 11:31 PM, Adam Lackorzynski wrote:
>>> On Thu Jan 17, 2013 at 17:03:36 +0100, Sebastian Sumpf wrote:
>>>> I recently upgraded Fiasco.OC to SVN revision 42 and experience some
>>>> pretty severe performance degradation compared to revision 40 on the
>>>> Pandaboard (SMP). It seems that 'simga0' and the root task stall for 5
>>>> to 10 seconds during boot up. I tracked the issue down to be caused by
>>>> the initial mapping operations, especially our root task maps all the
>>>> available memory during bootstrap. Within the kernel the
>>>> 'Context::xcpu_tlb_flush' is called for each mapping. The function sends
>>>> an IPI (to CPU1 which is idle) and then waits for an IPI in order to
>>>> signal the end of the operation. The whole operation seems to have
>>>> gotten slower compared to revision 40, but I could not find many
>>>> differences in the IPI-handling code. Do you have any ideas or
>>>> suggestions what could cause the delay (maybe scheduling changes) and
>>>> how to fix it?
>>>
>>> I noticed a similar/same thing but hadn't time to investigate yet.
>>
>> Okay, I just wanted to make sure that the problem is not at our side nor
>> at our usage pattern.
>> Another thing I wonder is: Since you now have second level cache support
>> for the PandaBoard, how do I map DMA memory to a client? The problem
>> seems to be that sigma0 maps all memory as cached. So what we have been
>> trying to do is this: When someone requests DMA memory we map the page
>> as uncached and then call 'l4_cache_dma_coherent' afterwards. This
>> doesn't seem to work out well for our drivers. The thing I think I could
>> gather is that memory that is mapped cached (sigma0, roottask) and
>> uncached (client) at the same time has an undefined behavior (I might be
>> wrong here) on ARM. So, what is the protocol to implement this on
>> Fiasco.OC/L4RE setups?
> 
> Indeed, having memory with different attributes must be avoided.
> But it's also about accessing that memory. So for example for sigma0
> this isn't a problem because sigma0 does not touch the memory itself.
> Is your roottask accessing the memory, i.e. pulling it into caches?

Yes this is the problem we're trying to solve. We don't have a notice of
normal RAM and DMA pools within our roottask (if we had this, the
question of how to dimension DMA pools would arise, also this seems to
be an ARM only issue). So here is what we did with L1 caches enabled
only: Acquire the memory in roottask, map it to the client as
non-cacheable, zero out the memory (as it might have been previously
used as normal RAM and for security reasons), clean and invalidate the
data-cache. With the L2-cache enabled on PandaBoard, I tried to use the
'l4_cache_dma_coherent' function to accomplish the same behavior, which
worked out well for the L1-case ... but it didn't. So, what am I doing
wrong here, or isn't this supported anyways?

Thanks a lot,

Sebstian