Re: [PATCH] net: openvswitch: Use for_each_cpu_from() where appropriate

[Yury Norov <yury.norov@gmail.com>]

On Thu, Aug 14, 2025 at 10:49:30PM +0200, Ilya Maximets wrote:
> On 8/14/25 9:58 PM, Yury Norov wrote:
> > From: Yury Norov (NVIDIA) <yury.norov@gmail.com>
> > 
> > Openvswitch opencodes for_each_cpu_from(). Fix it and drop some
> > housekeeping code.
> > 
> > Signed-off-by: Yury Norov (NVIDIA) <yury.norov@gmail.com>
> > ---
> >  net/openvswitch/flow.c       | 14 ++++++--------
> >  net/openvswitch/flow_table.c |  8 ++++----
> >  2 files changed, 10 insertions(+), 12 deletions(-)
> > 
> > diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c
> > index b80bd3a90773..b464ab120731 100644
> > --- a/net/openvswitch/flow.c
> > +++ b/net/openvswitch/flow.c
> > @@ -129,15 +129,14 @@ void ovs_flow_stats_get(const struct sw_flow *flow,
> >  			struct ovs_flow_stats *ovs_stats,
> >  			unsigned long *used, __be16 *tcp_flags)
> >  {
> > -	int cpu;
> > +	/* CPU 0 is always considered */
> > +	unsigned int cpu = 1;
> 
> Hmm.  I'm a bit confused here.  Where is CPU 0 considered if we start
> iteration from 1?

I didn't touch this part of the original comment, as you see, and I'm
not a domain expert, so don't know what does this wording mean.

Most likely 'always considered' means that CPU0 is not accounted in this
statistics.
  
> >  	*used = 0;
> >  	*tcp_flags = 0;
> >  	memset(ovs_stats, 0, sizeof(*ovs_stats));
> >  
> > -	/* We open code this to make sure cpu 0 is always considered */
> > -	for (cpu = 0; cpu < nr_cpu_ids;
> > -	     cpu = cpumask_next(cpu, flow->cpu_used_mask)) {
> > +	for_each_cpu_from(cpu, flow->cpu_used_mask) {
> 
> And why it needs to be a for_each_cpu_from() and not just for_each_cpu() ?

The original code explicitly ignores CPU0. If we use for_each_cpu(),
it would ignore initial value in 'cpu'. Contrary, for_each_cpu_from()
does respect it.

> Note: the original logic here came from using for_each_node() back when
> stats were collected per numa, and it was important to check node 0 when
> the system didn't have it, so the loop was open-coded, see commit:
>   40773966ccf1 ("openvswitch: fix flow stats accounting when node 0 is not possible")
> 
> Later the stats collection was changed to be per-CPU instead of per-NUMA,
> th eloop was adjusted to CPUs, but remained open-coded, even though it
> was probbaly safe to use for_each_cpu() macro here, as it accepts the
> mask and doesn't limit it to available CPUs, unlike the for_each_node()
> macro that only iterates over possible NUMA node numbers and will skip
> the zero.  The zero is importnat, because it is used as long as only one
> core updates the stats, regardless of the number of that core, AFAIU.
> 
> So, the comments in the code do not really make a lot of sense, especially
> in this patch.

I can include CPU0 and iterate over it, but it would break the existing
logic. The intention of my work is to minimize direct cpumask_next()
usage over the kernel, and as I said I'm not a domain expert here.

Let's wait for more comments. If it's indeed a bug in current logic,
I'll happily send v2.

Thanks,
Yury