From owner-freebsd-current  Tue Oct 15 13:26:25 1996
Return-Path: owner-current
Received: (from root@localhost)
          by freefall.freebsd.org (8.7.5/8.7.3) id NAA05997
          for current-outgoing; Tue, 15 Oct 1996 13:26:25 -0700 (PDT)
Received: from godzilla.zeta.org.au (godzilla.zeta.org.au [203.2.228.19])
          by freefall.freebsd.org (8.7.5/8.7.3) with ESMTP id NAA05990
          for <current@freebsd.org>; Tue, 15 Oct 1996 13:26:19 -0700 (PDT)
Received: (from bde@localhost) by godzilla.zeta.org.au (8.7.6/8.6.9) id GAA21446; Wed, 16 Oct 1996 06:22:23 +1000
Date: Wed, 16 Oct 1996 06:22:23 +1000
From: Bruce Evans <bde@zeta.org.au>
Message-Id: <199610152022.GAA21446@godzilla.zeta.org.au>
To: current@freebsd.org, phk@critter.tfs.com
Subject: Re: device driver open/close inconsistency
Sender: owner-current@freebsd.org
X-Loop: FreeBSD.org
Precedence: bulk

>Why do we call a device-driver->open() on all opens, but only ->close()
>once on the last close ?

Because it's easier to count references in vfs than in hundreds of
drivers.  The device close function would also need to be called on all
dup()s and implicit dup()s for fork()s to give it a chance of counting
references.  There is also an extra reference for controlling terminals.
There are also complications for /dev/tty.

Unfortunately, it's hard to count references properly in vfs too:

Bug 1: the reference count is incremented before calling the device
open function, so processes sleeping in open are effectively counted
as successful opens.   Suppose there is a process sleeping in open and
another process opens the device, does some ioctls to mess up the device
state, and closes the device.  Then it is necessary to call the device
close function to clean up and kick the process sleeping in open, but
the device close function is never called because the reference count is
always >= 1.  I believe this is (mis)handled in some versions of **ix
by not allowing processes to sleep in open when another open of the
same device to complete.  This causes other problems.  O_NONBLOCK is
per-open, so I think it is wrong to let a !O_NOBLOCK open succeed in
the presences of the blocking condition just because another process
has done a successful open using O_NOBLOCK.

Bug 2: It is difficult to get rid of controlling terminals.  In 4.4Lite,
exiting was the only way.  Thus device close functions for controlliing
terminals were never called until exit.  In 4.4Lite2 and most versions
of FreeBSD-2.x, the controlling terminal can be changed using TIOCSCTTY.

Bruce