From owner-freebsd-current  Tue Feb 20 16:57:03 1996
Return-Path: owner-current
Received: (from root@localhost)
          by freefall.freebsd.org (8.7.3/8.7.3) id QAA20620
          for current-outgoing; Tue, 20 Feb 1996 16:57:03 -0800 (PST)
Received: from research.gate.nec.co.jp (research.gate.nec.co.jp [202.32.8.49])
          by freefall.freebsd.org (8.7.3/8.7.3) with ESMTP id QAA20614
          Tue, 20 Feb 1996 16:56:59 -0800 (PST)
Received: from sbl-gw.sbl.cl.nec.co.jp by research.gate.nec.co.jp (8.7.3+2.6Wbeta5/950912) with ESMTP id JAA12918; Wed, 21 Feb 1996 09:56:40 +0900 (JST)
Received: from sirius.sbl.cl.nec.co.jp by sbl-gw.sbl.cl.nec.co.jp (8.7.3/3.3W6) with SMTP
	id JAA14149; Wed, 21 Feb 1996 09:56:39 +0900 (JST)
Received: by sirius.sbl.cl.nec.co.jp (8.6.9/3.3W6) with UUCP
	id JAA17442; Wed, 21 Feb 1996 09:58:33 +0900
Date: Wed, 21 Feb 1996 09:58:33 +0900
From: Naoki Hamada <nao@sbl.cl.nec.co.jp>
Message-Id: <199602210058.JAA17442@sirius.sbl.cl.nec.co.jp>
References: <199602202331.PAA04282@Root.COM>
To: davidg@Root.COM
CC: andreas@knobel.gun.de, hackers@FreeBSD.org, current@FreeBSD.org
In-reply-to: David Greenman's message of "Tue, 20 Feb 1996 15:31:50 -0800"
	     <199602202331.PAA04282@Root.COM>
Subject: Re: mbuf enhancement patch
Sender: owner-current@FreeBSD.org
Precedence: bulk

>   No. The performance improvemment is actually quite small and the effect of
>this is that any buffers malloced up to the high water mark won't be available
>to other parts of the system after they are freed. We generally try to avoid
>private pools of buffers unless it's absolutely necessary - which is case for
>mbuf clusters, for example, which has a mechanism for maintaining reference
>counts that requires them to be allocated out of a private pool.

Agreed. Sounds quite reasonable.

>   We once had changes similar to the ones you've provided, except we had it
>so that the buffers over a certain threshold were returned back to malloc. The
>problem with this was that the malloc type was lost in the process and this
>messed up the malloc-type accounting (which eventually leads to malloc
>failures).

I found the ep driver always keeps some mbuf's in its pool. Is this
because mbuf allocation is too expensive for boards which equip small
receive buffer? If this is the case, some improvement (not mine :-) is
desirable.

-nao