Re: [PATCH] bonding:avoid repeated display of same link status change

[Michal Kubecek <mkubecek@suse.cz>]

On Fri, Nov 02, 2018 at 11:31:38PM -0700, David Miller wrote:
> From: mk.singh@oracle.com
> Date: Wed, 31 Oct 2018 16:27:28 +0530
> 
> > -			if (slave->delay) {
> > +			if (slave->delay &&
> > +			    !atomic64_read(&bond->rtnl_needed)) {
>  ...
> > +			    !atomic64_read(&bond->rtnl_needed)) {
>  ...
> > +			atomic64_set(&bond->rtnl_needed, 1);
>  ...
> > +		atomic64_set(&bond->rtnl_needed, 0);
>  ...
> > @@ -229,6 +229,7 @@ struct bonding {
> >  	struct	 dentry *debug_dir;
> >  #endif /* CONFIG_DEBUG_FS */
> >  	struct rtnl_link_stats64 bond_stats;
> > +	atomic64_t rtnl_needed;
> 
> There is nothing "atomic" about a value that is only set and read.
> 
> And using a full 64-bit value for something taking on only '0' and
> '1' is unnecessary as well.

Part of the misunderstanding is caused by the fact that this is actually
a v4 but not marked as such:

  v1: https://patchwork.ozlabs.org/patch/955789/
  v2: https://patchwork.ozlabs.org/patch/970421/
  v3: https://patchwork.ozlabs.org/patch/988241/

When commenting v3, I didn't know about the v2 discussion where Eric
Dumazet NACKed the patch because of potential conflict issues:

  https://patchwork.ozlabs.org/patch/970421/#1992397
  https://patchwork.ozlabs.org/patch/988241/#2017317

On the other hand, there is no need for atomic64_t. Simple atomic_t
(with explaining comment) would suffice. On architectures allowing
atomic read/write for 32-bit integers, there would be no performance
penalty. On architectures not allowing it, atomic_read() and
atomic_set() are implemented to be safe.

Michal Kubecek