Re: [PATCH repost] sched: export sched_set/getaffinity to modules

["Michael S. Tsirkin" <mst@redhat.com>]

On Fri, Jul 30, 2010 at 04:19:01PM +0200, Oleg Nesterov wrote:
> Sorry for the delay, I can't be responsive these days...
> 
> On 07/27, Michael S. Tsirkin wrote:
> >
> > On Mon, Jul 26, 2010 at 08:08:34PM +0200, Oleg Nesterov wrote:
> > > On 07/26, Sridhar Samudrala wrote:
> > > >
> > > > I have been testing out a similar patch that uses kernel_thread() without CLONE_FILES
> > > > flag rather than create_kthread() and then closing the files.
> > >
> > > !CLONE_FILES can't help. copy_files() does dup_fd() in this case.
> > > The child still inherits the files.
> > >
> > > > Either version should be fine.
> > >
> > > I think neither version is fine ;)
> > >
> > > exit_files() is not enough too. How about the signals, reparenting?
> > >
> > >
> > > I already forgot all details, probably I missed somethig. But it
> > > seems to me that it is better to just export get/set affinity and
> > > forget about all complications.
> > >
> > > Oleg.
> >
> > Oleg, so can I attach your Ack to the patch in question, and merge
> > it all through net-next?
> 
> Well, I do not think you need my ack ;)
> 
> 
> But I must admit, I personally dislike this idea. A kernel thread which
> is the child of the user-space process, and in fact it is not the "real"
> kernel thread. I think this is against the common case. If you do not
> care the signals/reparenting, why can't you fork the user-space process
> which does all the work via ioctl's ? OK, I do not understand the problem
> domain, probably this can't work.
> 
> Anyway, the patch looks buggy to me. Starting from
> 
> 	create_kthread(&create);
> 	wait_for_completion(&create.done);
> 
> At least you should check create_kthread() suceeds, otherwise
> wait_for_completion() will hang forever. OTOH, if it suceeds then
> wait_for_completion() is not needed. But this is minor.
> 
> create_kthread()->kernel_thread() uses CLONE_VM, this means that the
> child will share ->mm. And this means that if the parent recieves
> the coredumping signal it will hang forever in kernel space waiting
> until this child exits.
> 
> This is just the immediate surprise I can see with this approach,
> I am afraid there is something else.
> 
> And once again. We are doing this hacks only because we lack a
> couples of exports (iiuk). This is, well, a bit strange ;)
> 
> Oleg.


Oleg, I mean Ack the exporting of get/set affinity.

Thanks!

-- 
MST