From owner-cvs-sys  Sun Feb 16 06:06:33 1997
Return-Path: <owner-cvs-sys>
Received: (from root@localhost)
          by freefall.freebsd.org (8.8.5/8.8.5) id GAA05813
          for cvs-sys-outgoing; Sun, 16 Feb 1997 06:06:33 -0800 (PST)
Received: from godzilla.zeta.org.au (godzilla.zeta.org.au [203.2.228.19])
          by freefall.freebsd.org (8.8.5/8.8.5) with ESMTP id GAA05807;
          Sun, 16 Feb 1997 06:06:26 -0800 (PST)
Received: (from bde@localhost) by godzilla.zeta.org.au (8.8.3/8.6.9) id BAA20251; Mon, 17 Feb 1997 01:04:24 +1100
Date: Mon, 17 Feb 1997 01:04:24 +1100
From: Bruce Evans <bde@zeta.org.au>
Message-Id: <199702161404.BAA20251@godzilla.zeta.org.au>
To: ache@nagual.ru, peter@spinner.dialix.com
Subject: Re: cvs commit: src/sys/sys types.h
Cc: bde@zeta.org.au, cvs-all@freefall.freebsd.org,
        CVS-committers@freefall.freebsd.org, cvs-sys@freefall.freebsd.org,
        mpp@freefall.freebsd.org
Sender: owner-cvs-sys@FreeBSD.ORG
X-Loop: FreeBSD.org
Precedence: bulk

>I don't understand why rounding is even needed here (maybe to make less

32-bit rounding is better for bzero, copyout, etc, and doesn't cost more
storage.  256-bit rounding avoids slow startup but isn't space-efficient
since we're allocating 6 times that.  Anyway, I think it would be
better to use an auto buffer for small allocations even p_selbits has
been allocated.  This would have better locality.  Most applications
wouldn't need a malloced buffer.  Half of implementation:

	fd_mask selbits[6 * OLD_FD_SETSIZE / NFDBITS];
	...
	ni = bits_to_copyin();		/* multiple of NFDBITS */
	if (ni <= OLD_FD_SETSIZE)
		use_selbits();
	else
		use_malloced_bits();

>*** sys_generic.c.orig	Tue Jan 14 23:11:28 1997
>--- sys_generic.c	Sun Feb 16 16:00:48 1997
>***************
>*** 508,513 ****
>--- 508,515 ----
>  	return (error);
>  }
>  
>+ #define FD_ROUNDSIZE 256
>+ 

Change to something like (2048 / 6) = 341.  Now that's an odd size :-).

>--- 562,580 ----
>  			p->p_selbits_size);
>  	}
>  
>  	/* The amount of space we need to copyin/copyout */
>  	ni = howmany(uap->nd, NFDBITS) * sizeof(fd_mask);
>+ 
>+ 	if (ni > 8) {   /* 64 bits - arbitary guess at tradeoff */
>+ 		/*
>+ 		 * Clear the entire return-to-user buffer space
>+ 		 */
>+ 		bzero (p->p_selbits + 3 * p->p_selbits_size,
>+ 			p->p_selbits_size * 3);
>+ 	} else {
>+ 		for (i = 0; i < 3; i++)
>+ 			bzero(obits[i], ni);
>+ 	}

Don't use this except possibly for very large ni together with unusual
combinations of fds.  Put 3 contiguous copyout buffers first (readfds,
writefds, exexptfds), then 3 copyin buffers, and bzero 1*ni, 2*ni or
3*ni bits as necessary (exceptfds is usually NULL so the bzero doesn't
need to cover it).  The old code got this wrong mainly by spacing the
buffers with separation p->p_selbits_size instead of ni.

Bruce