>>> On Thu, Aug 21, 2008 at 01:02:48PM +0200, Peter Zijlstra wrote:
>>>> On Thu, 2008-08-21 at 12:57 +0200, Stefan Richter wrote:
>>>>> Peter Zijlstra wrote:
>>>>>> On Wed, 2008-08-20 at 20:50 -0600,
jmerkey@wolfmountaingroup.com
>>>>>> wrote:
>>>>>>
>>>>>>> volatiles left in the code due to the previously stated
>>>>>>> (and still present) severe breakage of the GNU compiler with SMP
>>>>>>> shared data. most of the barrier() functions are just plain
>>>>>>> broken
>>>>>>> and do not result in proper compiler behavior in this tree.
>>>>>> Can you provide explicit detail?
>>>>>>
>>>>>> By using barrier() the compiler should clobber all its memory and
>>>>>> registers therefore forcing a write/reload of the variable.
>>>>> I hope Jeff didn't try mere barrier()s only. smp_wmb() and smp_rmb()
>>>>> are the more relevant barrier variants for mdb, from what I remember
>>>>> when I last looked at it.
>>>> Sure, but volatile isn't a replacement for memory barriers.
>>>
>>> Let's face it, the C standard does not support concurrency, so we are
>>> all in a state of sin in any case, forced to rely on combinations of
>>> gcc-specific non-standard language extensions and assembly language.
>>>
>>> Could be worse!!!
>>
>> Nevertheless, an analysis of which particular parts of code generation
>> are insufficient if one particular volatile qualification is removed is
>> IMO likely to turn up places in mdb where a clearer or/and more
>> efficient implementation is possible. (Based on what I saw a few
>> revisions ago; I haven't looked at the current one yet.)
>> --
>> Stefan Richter
>> -=====-==--- =--- =-=-=
>>
http://arcgraph.de/sr/
>>
>
> I used the smp_wmb() functions. I noted a couple of things. a) some of
> these macros just emit __asm__ __volatile__ into the code so why not just
> say "volatile" to begin with b) smp_wmb() in some cases worked and in
> other cases jut optimized away the global reference. c) I can go back and
> break the code again by inserting them and building broken assembler d) I
> ave been doing hardware and software design since the early 1980;s, I
> invented SMP affinity scheduling, and yes, I understand barriers and this
> concept of instruction score-boarding and optimization very well -- its
> not an excuse for a busted C compiler.
>
> It did not break all the places in the code, but broke enough for SMP to
> lock up and fail, It turned global variables into local variables. If
> you want me to reproduce this I can but it will have to wait til this
> evening
> because I have some product releases to get out the door at Omega 8 today.
>
> It's simple to reproduce. Take away the volatile declaration for the
> rlock_t structure in mdb-ia32.c (rlock_t debug_lock) in all code
> references and watch the thing lock up in SMP with multiple processors in
> the debugger each stuck with their own local copy of debug_lock.
