Re: [RFC][PATCH] Add prctl to set sibling thread names

[KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>]

> On Wed, 21 Oct 2009 17:52:24 -0700 john stultz <johnstul@us.ibm.com> wrote:
> 
> > On Wed, 2009-10-21 at 17:48 -0700, Arjan van de Ven wrote:
> > > On Wed, 21 Oct 2009 16:21:37 -0700
> > > john stultz <johnstul@us.ibm.com> wrote:
> > > 
> > > > 
> > > > Taking a very raw attempt at this, I scratched out the following
> > > > simple implementation. I'd appreciate any review or suggestions for
> > > > improvements. I'm not at all certain the passing of the thread pid_t
> > > > through the unsigned long is valid, for instance, or if
> > > > same_thread_group() is the right check to make sure we only change
> > > > siblings and not tid from other processes. So any advice on better
> > > > approaches would be great.
> > > > 
> > > > +				return -EINVAL;
> > > > +
> > > > +			set_task_comm(tsk, comm);
> > > 
> > > 
> > > you're pretty much the first now who touches ->comm from
> > > not-the-thread-itself.... are you sure that is safe?
> > 
> > No, I'm not sure at all :)
> > 
> > Thanks for pointing this out. I'll see whats needed in set_task_comm().
> > 
> 
> set_task_comm() is OK.  The problem will be the unwritten rule that
> processes can read *their own* ->comm without task_lock(), because nobody
> ever alters ->comm apart from tack which owns it.
> 
> You've changed that, so all the open-coded accesses to current->comm are
> now racy.
> 
> Also, you appear to be running set_task_comm() against a task_struct
> without holding a reference on that task.  Will a well-timed exit() cause a
> modify-after-free?

John, I'd prefer to suggested another design.
How about this?

1. remove pid argument from prctl
2. cancel pthread_setname_np()
3. instead, create pthread_attr_setname_np()
4. pthread_create() change own thread name by pthread_attr.


It avoid many racy problem automatically.