Re: [PATCH resend] cpumask: don't perform while loop in cpumask_next_and()

[Sergey Senozhatsky <sergey.senozhatsky@gmail.com>]

Hello,

On (06/15/15 15:12), Peter Zijlstra wrote:
> > +++ b/lib/cpumask.c
> > @@ -37,10 +37,11 @@ EXPORT_SYMBOL(__next_cpu_nr);
> >  int cpumask_next_and(int n, const struct cpumask *src1p,
> >  		     const struct cpumask *src2p)
> >  {
> > +	struct cpumask tmp;
> > +
> > +	if (cpumask_and(&tmp, src1p, src2p))
> > +		return cpumask_next(n, &tmp);
> > +	return nr_cpu_ids;
> >  }
> >  EXPORT_SYMBOL(cpumask_next_and);
> 
> Just ran into this; I though we were not supposed to put cpumasks on the
> stack because $BIG. ?!

Gosh, I didn't think $BIG enough. So, on a _big_ 4096 x86_64 it's like...
64 bytes on stack. That's bad. alloc_cpumask_var()/free_cpumask_var()
version just doesn't look like a win (inlined below) so I guess I'll
ask to revert. It makes sense on smaller systems, but loses on huge
ones.

---

 lib/cpumask.c | 14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)

diff --git a/lib/cpumask.c b/lib/cpumask.c
index 5f62708..95ce89a 100644
--- a/lib/cpumask.c
+++ b/lib/cpumask.c
@@ -16,11 +16,17 @@
 int cpumask_next_and(int n, const struct cpumask *src1p,
 		     const struct cpumask *src2p)
 {
-	struct cpumask tmp;
+	int ret = nr_cpu_ids;
+	cpumask_var_t tmp;
 
-	if (cpumask_and(&tmp, src1p, src2p))
-		return cpumask_next(n, &tmp);
-	return nr_cpu_ids;
+	if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
+		return ret;
+
+	if (cpumask_and(tmp, src1p, src2p))
+		ret = cpumask_next(n, tmp);
+
+	free_cpumask_var(tmp);
+	return ret;
 }
 EXPORT_SYMBOL(cpumask_next_and);