[fedora-arm] ARMv5 and atomic operations

[Andrew Haley]

On 04/24/2012 04:49 PM, Nicolas Pitre wrote:
> On Tue, 24 Apr 2012, Andrew Haley wrote:
> 
>> On 04/24/2012 12:39 PM, Nicolas Pitre wrote:
>>> On Tue, 24 Apr 2012, Andrew Haley wrote:
>>>
>>>> On 04/23/2012 09:31 PM, Nicolas Pitre wrote:
>>>>> On Mon, 23 Apr 2012, Andrew Haley wrote:
>>>>>
>>>>>> On 04/23/2012 06:36 PM, Thomas Meyer wrote:
>>>>>>> I'm running the Ubuntun 2.6.38 Tegra2 kernel (because of their fbdev
>>>>>>> support) on top of Fedora 17 armv5el on an Toshiba AC100 Laptop. The
>>>>>>> rsyslog package crashed everytime because of the missing kernel support
>>>>>>> of cmpxchg64. So when relying on the kernel helpers make sure that the
>>>>>>> resp. kernel support exists.
>>>>>>
>>>>>> Indeed.  I had to write a workaround in IcedTea (i.e. java) on ARM for
>>>>>> just this reason.  If you can't depend on a kernel helper being there I
>>>>>> can't see it's of any use.
>>>>>
>>>>> Kernel helpers don't disappear with time.  You therefore can probe for 
>>>>> their availability (see the documentation) in case the kernel support 
>>>>> could be backported, or just refuse to run if the kernel version isn't 
>>>>> recent enough.  This is not much different from relying on a new 
>>>>> syscall.
>>>>
>>>> Indeed it is.  What would I gain from adding such a test?  All I can
>>>> see is extra complication, untested code paths, and larger programs.
>>>
>>> What alternative do you have, other than not using any atomic 
>>> operations?
>>
>> What I've done already: I have my own routines.
> 
> You may get some atomicity with the SWP instruction, but that is not 
> sufficient to implement atomic increments and the like. 
>
> You may use the SWP to implement some locking and perform the "atomic" 
> operations under the lock protection, but that doesn't work if atomic 
> operations are also performed in signal handlers, and that scales badly 
> if concurency is high.
> 
> So that really depends what your atomic needs are.  Implementing the 
> latest library standards without any kernel help on ARMv5 or earlier is 
> simply impossible.  But custom code can get away with less.

Oh I see what you mean now.  I'm inside the Java VM, so I can fall
back to locking if I'm on ARMv5; ldrexd is an optimization for doing
long-word updates rather than needed for correctness.

So, if I'm on v7 I have to have my own cmpxchg64 routines for older
kernels.  I have to fall back to locks on older CPUs, because I still
want to work on old kernels too.  I can't see any circumstances where
the kernel helper will be of any use.

Andrew.