Re: New 2.6.6 pagg and job patches available

Peter Williams wrote:
> Erik Jacobson wrote:
> 
>> New pagg and job patches are available from the pagg web site for the
>> 2.6.6 kernel.
>>
>> oss.sgi.com/projects/pagg
>>
>> Click 'download' on the left.
>>
>> The patches I made available are:
>> linux-2.6.6-pagg.patch
>> and
>> linux-2.6.6-job.patch
>>
>> Check out the README for some installation tips.
>>
>> Note: we haven't finished processing all the community feedback yet.  So
>> there will likely be at least one more 2.6.6 pagg patch to come.
>> I also haven't implemented the patches from Peter Williams yet but 
>> hope to
>> soon.
> 
> 
> Attached is a patch (against your latest 2.6.6 patch) to do the 
> initialisation of tasks at registration and add real uid/gid and CPU 
> affinity hooks.
> 
> The "initialisation of tasks" code still needs some work so that it 
> cleans up properly when faced with memory allocation failures.  Probably 
> best left to someone more familiar with PAGG than me.
> 
> On a related note, shouldn't the deregistration code do something 
> similar? i.e. call detach() on all of the deregistering clients paggs 
> and remove form any task lists that they are on rather than just 
> refusing to unload if there are any tasks with paggs belonging to the 
> deregistering client.  Otherwise every client will have to reinvent the 
> wheel in order to do this when it is unloaded from the kernel.

I've just been perusing pagg_hook_unregister() with a view to 
implementing the above idea and noticed that there is a possible race 
condition in the code that checks for the presence of paggs from the 
client being deregistered in tasks.  Because the tasklist_lock is 
released within the for loop it is possible for new tasks to be created 
and added in such a way that they miss being processed by this for loop.

-- 
Dr Peter Williams                                [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
  -- Ambrose Bierce