On Wed, May 08, 2024 at 01:52:21PM +0300, Vitaliy Makkoveev wrote:
> The simplest case. Nothing to change in sockets layer, only set
> SB_MTXLOCK on buffers. This diff doesn't conflict with "Turn sblock()
> to `sb_lock' ..." [1] diff.
> 
> 1. https://marc.info/?t=171478067200001&r=1&w=2
> 
> Not related to this diff, but pkpcb uses solock() only to protect
> `kcb_flags' and `kcb_reg'. Protected paths are pretty small and have no
> context switch, so per-pkpcb mutex(9) could be used instead of solock().
> This removes context switch from `kcb_list' walkthroughs and this SRP
> could be converted to SMR list.

OK bluhm@

> @@ -595,14 +594,14 @@ pfkeyv2_sendmessage(void **headers, int 
>  
>  		/* Send to all registered promiscuous listeners */
>  		SRPL_FOREACH(kp, &sr, &pkptable.pkp_list, kcb_list) {
> +			int flags = READ_ONCE(kp->kcb_flags);
> +
>  			if (kp->kcb_rdomain != rdomain)
>  				continue;
>  

You can move the READ_ONCE after continue.