[okl4-developer] Fast address-space switching on GUMSTIX?

[xi yang]

On Sat, May 24, 2008 at 5:42 AM, Abi Nourai <anourai at ok-labs.com> wrote:
> Hi Xi
>
> On 23/05/2008, at 6:58 PM, xi yang wrote:
>
>> Hi folks:
>>
>> I write a test, running two thread in two different address space,
>> each thread execute
>> while(1){
>>  print the timestamp of performance counter of xscale
>>  set counter of xscale to 0
>>  L4_Yield()
>> }
>> It shows that the cost of switching  to other address space is about
>> 4000 cycles and 7 TLB miss.
>> I find that after each time of address space switching, when new
>> thread begin execute in user level, a exception occurs, which is a
>> domain exception and then kernel flush the TLB and cache.
>>
>> Dose Fast address space switching works on gumstix(Xscale) ?
>
>
> Do the two address spaces you're switching between reside in different 1MB
> aligned regions of virtual memory ?  This is a constraint of ARMv5 hardware
> (that's derived from top level page-table entries mapping 1MB regions) and
> is required to avoid domain faults which is where you're taking the hit.
1)But it seems that iguana adds 0x8000,0000 - 0x8006,0000(In the
kdebuger, user print ptab command) to every PD, what 's the function
of this mapping, how to tell the iguana do not mapping these spaces,
since this will cause domain fault.

2) Vtimer and other server, they will create a memsection and map it
to the hardware address, it seems that iguana will map this memsection
to 0x800X,XXXX, which means after address space switching, there will
be a domain fault when you access this area.

How to configure iguana to eliminate the overlap mapping? I am not
familiar with the internal of iguana.
>
> Regards
> Abi
>