From: "Ted Unangst" <tedu@tedunangst.com>
Subject: Re: sysctl: unlock KERN_CONSBUF and KERN_MSGBUF
To: "Alexander Bluhm" <bluhm@openbsd.org>
Cc: "Vitaliy Makkoveev" <mvs@openbsd.org>, tech@openbsd.org
Date: Thu, 08 May 2025 12:24:57 -0400

On 2025-05-08, Alexander Bluhm wrote:
> I was considering memset() and mutex again.  Although clang does
> not optimize away the memset stores, the processor still could to
> this.  But putting a mutex barrier between clearing and filling the
> memory of struct ump prevents all optimization.
> 
> Of course sysctl(2) is not the hot path, it does not matter for
> speed, and I have no benchmarks.  It is more about MP programming
> in general.

I like this, if only because it makes the mtx_enter a little easier to see.


> 
> What do you think?
> 
> bluhm
> 
> Index: kern/kern_sysctl.c
> ===================================================================
> RCS file: /data/mirror/openbsd/cvs/src/sys/kern/kern_sysctl.c,v
> diff -u -p -r1.467 kern_sysctl.c
> --- kern/kern_sysctl.c	6 May 2025 18:34:26 -0000	1.467
> +++ kern/kern_sysctl.c	8 May 2025 12:58:42 -0000
> @@ -598,8 +598,8 @@ kern_sysctl(int *name, u_int namelen, vo
>  		} else 
>  			return (0);
>  
> -		memset(&ump, 0, sizeof(ump));
>  		mtx_enter(&log_mtx);
> +		memset(&ump, 0, sizeof(ump));
>  		ump.msg_magic = mp->msg_magic;
>  		ump.msg_bufs = mp->msg_bufs;
>  		ump.msg_bufx = mp->msg_bufx;
> 
>