Re: [Patch 04/12] GRU - collect per-context user statistics

[Jack Steiner <steiner@sgi.com>]

On Mon, Jun 08, 2009 at 04:07:40PM -0700, Andrew Morton wrote:
> On Mon, 08 Jun 2009 12:16:52 -0500
> steiner@sgi.com wrote:
> 
> >  /*
> > + * Fetch GSEG statisticss
> > + */
> > +long gru_get_gseg_statistics(unsigned long arg)
> > +{
> > +	struct gru_thread_state *gts;
> > +	struct gru_get_gseg_statistics_req req;
> > +
> > +	if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
> > +		return -EFAULT;
> > +
> > +	gts = gru_find_lock_gts(req.gseg);
> > +	if (gts) {
> > +		memcpy(&req.stats, &gts->ustats, sizeof(gts->ustats));
> > +		gru_unlock_gts(gts);
> > +	} else {
> > +		memset(&req.stats, 0, sizeof(gts->ustats));
> > +	}
> > +
> > +	if (copy_to_user((void __user *)arg, &req, sizeof(req)))
> > +		return -EFAULT;
> > +
> > +	return 0;
> > +}
> 
> So..  what's happening in the super-secret undocumented gts==NULL path?
> 
> It _looks_ like userspace passed into this ioctl a handle for something
> which the kernel doesn't know about.  If so, shouldn't we return
> -EINVAL or something?  

It makes sense but certainly needs a comment (will send later) to
explain the logic.

User space creates arrays of GRU contexts for threaded processes. The
library code that prints statistics scans the array & generates
statistic for each context. If an context was never referenced, there is
no GTS & all statistics are implicitly zero.

This could have been handled other ways but it is rare than an entry was
never referenced. A return of -EINVAL current is considered a bug, ie.
address is not a valid GRU address.


--- jack