Re: cpusets and kthreads, inconsistent behaviour

[Max Krasnyansky <maxk@qualcomm.com>]

David Rientjes wrote:
> On Tue, 10 Jun 2008, Max Krasnyansky wrote:
> 
>> Basically the issue is that current behaviour of the cpusets is inconsistent
>> with regards to kthreads. Kthreads inherit cpuset from a parent properly but
>> they simply ignore cpuset.cpus when their cpu affinity is set/updated.
>> I think the behaviour must be consistent across the board. cpuset.cpus must
>> apply to _all_ the tasks in the set, not just some of the tasks. If kthread
>> must run on the cpus other than current_cpuset.cpus then it should detach from
>> the cpuset.
>>
> 
> I disagree that a cpuset's set of allowable cpus should apply to all tasks 
> attached to that cpuset.  It's certainly beneficial to be able to further 
> constrict the set of allowed cpus for a task using sched_setaffinity().
> 
> It makes more sense to argue that for each task p, p->cpus_allowed is a 
> subset of task_cs(p)->cpus_allowed.
Yes that's exactly what I meant :). Sorry for not being clear. I did not mean
that each task in a cpuset must have the same affinity mask. So we're on the
same page here.

>> To give you an example kthreads like scsi_eh, kswapd, kacpid, pdflush,
>> kseriod, etc are all started with cpus_allows=ALL_CPUS even though they
>> inherit a cpuset from kthreadd. Yes they can moved manually (with
>> sched_setaffinity) but the behaviour is not consistent, and for no good
>> reason. kthreads can be stopped/started at any time (module load for example)
>> which means that the user will have to keep moving them.
>>
> 
> This doesn't seem to be purely a kthread issue.  Tasks can be moved to a 
> disjoint set of cpus by any caller to set_cpus_allowed_ptr() in the 
> kernel.
Hmm, technically you are correct of course. But we do not have any other
kernel tasks besides kthreads. All the kernel tasks I have on my machines have
kthreadd as their parent.
And I'm not aware of any kernel code that changes affinity mask of a
user-space task without paying attention to the cpuset the task belongs to. If
you know of any we should fix it because it'd clearly be a bug.

Thanx
Max