Re: [RFC PATCH v5 4/7] sched: bias task wakeups to preferred semi-idle packages

[Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>]

* Balbir Singh <balbir@linux.vnet.ibm.com> [2008-12-15 23:32:36]:

> >> > > Sure its racy, but so what?
> >> > >
> >> > > The worst I can see it that we exclude a dying task from this logic,
> >> > > which isn't a problem at all, since its dying anyway.
> >> >
> >> > At which point I seriously doubt it'd still be on the rq anyway.
> >> >
> >>
> >> I forgot to mention that, the check should be (p->mm && !(p->flags & PF_KTHREAD))
> >
> > I can check for PF_KTHREAD for now.  However, I should reduce the
> > number of checks since this may slow down wake_idle for sched_mc=2.
> >
> > We can tolerate p->mm check on a dying process as Peter has suggested,
> > hence we don't need to protect it.  We are not going to access any
> > contents of the mm struct.
> >
> > If PF_KTHREAD is only being used by AIO, then I feel we can drop the
> > check since the threads will not have affinity and they can be moved
> > to other cpus anyway.
> >
> > The main reason for skipping kthread is that they may be using per-cpu
> > variables and sleep/preempted.  I did not want the wake_idle() logic
> > to move them around forcefully.  This is not the general case and this
> > situation should not happen.
> >
> > Second reason is to optimise on the affinity check since most of the
> > kthreads have affinity and cannot be moved.
> >
> > This condition check needs optimisation after getting the framework
> > functionally correct and useful.
> 
> 
> Vaidy, you (or your mailer) seem to have dropped me off of the to/cc
> list while replying and this seems to be the case for all replies.

Hi Balbir,

Thanks for pointing that out.  I will review my mutt setup.  This is
strange... since a group-reply to this message puts you in the to list
as expected, but not the ones that I sent earlier in reply to your
messages.  The header seems to be correct for others except for your
message :(

--Vaidy