From owner-freebsd-arch@FreeBSD.ORG Thu May 15 12:07:52 2008 Return-Path: Delivered-To: freebsd-arch@freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2001:4f8:fff6::34]) by hub.freebsd.org (Postfix) with ESMTP id 1CF79106566B for ; Thu, 15 May 2008 12:07:52 +0000 (UTC) (envelope-from jhb@freebsd.org) Received: from elvis.mu.org (elvis.mu.org [192.203.228.196]) by mx1.freebsd.org (Postfix) with ESMTP id F18D98FC17 for ; Thu, 15 May 2008 12:07:51 +0000 (UTC) (envelope-from jhb@freebsd.org) Received: from server.baldwin.cx (unknown [208.65.91.234]) by elvis.mu.org (Postfix) with ESMTP id 46FE41A4D8C; Thu, 15 May 2008 05:07:51 -0700 (PDT) Received: from localhost.corp.yahoo.com (john@localhost [127.0.0.1]) (authenticated bits=0) by server.baldwin.cx (8.14.2/8.14.2) with ESMTP id m4FC7fd2083425; Thu, 15 May 2008 08:07:42 -0400 (EDT) (envelope-from jhb@freebsd.org) From: John Baldwin To: freebsd-arch@freebsd.org Date: Thu, 15 May 2008 07:30:07 -0400 User-Agent: KMail/1.9.7 References: <20080504171002.GN18958@deviant.kiev.zoral.com.ua> <20080513093734.GF18958@deviant.kiev.zoral.com.ua> <20080514093223.GN18958@deviant.kiev.zoral.com.ua> In-Reply-To: <20080514093223.GN18958@deviant.kiev.zoral.com.ua> MIME-Version: 1.0 Content-Disposition: inline Message-Id: <200805150730.08418.jhb@freebsd.org> Content-Type: text/plain; charset="iso-8859-15" Content-Transfer-Encoding: 7bit X-Greylist: Sender succeeded SMTP AUTH authentication, not delayed by milter-greylist-2.0.2 (server.baldwin.cx [127.0.0.1]); Thu, 15 May 2008 08:07:42 -0400 (EDT) X-Virus-Scanned: ClamAV 0.91.2/7127/Thu May 15 05:59:12 2008 on server.baldwin.cx X-Virus-Status: Clean X-Spam-Status: No, score=-4.4 required=4.2 tests=ALL_TRUSTED,AWL,BAYES_00 autolearn=ham version=3.1.3 X-Spam-Checker-Version: SpamAssassin 3.1.3 (2006-06-01) on server.baldwin.cx Cc: Kostik Belousov Subject: Re: Per-open file private data for the cdevs X-BeenThere: freebsd-arch@freebsd.org X-Mailman-Version: 2.1.5 Precedence: list List-Id: Discussion related to FreeBSD architecture List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Thu, 15 May 2008 12:07:52 -0000 On Wednesday 14 May 2008 05:32:23 am Kostik Belousov wrote: > On Tue, May 13, 2008 at 12:37:34PM +0300, Kostik Belousov wrote: > > On Mon, May 12, 2008 at 10:03:14AM -1000, Jeff Roberson wrote: > > > > > > On Mon, 12 May 2008, Kostik Belousov wrote: > > > > > > >On Sun, May 11, 2008 at 03:40:14PM -1000, Jeff Roberson wrote: > > > >> > > > >>On Sun, 11 May 2008, Kostik Belousov wrote: > > > >> > > > >>>On Sun, May 11, 2008 at 02:50:30PM +0300, Kostik Belousov wrote: > > > >>>>On Sat, May 10, 2008 at 09:53:12PM -1000, Jeff Roberson wrote: > > > >>>>>On Sun, 4 May 2008, Kostik Belousov wrote: > > > >>>>> > > > >>>>>>Since the review for the clone-at-open patch (fdclone) posted some > > > >>>>>>time > > > >>>>>>ago > > > >>>>>>mostly says that it would be better to implement per-file private data > > > >>>>>>instead, I produced the patch along this line, > > > >>>>>> > > > >>>>>>The patch does not change the cdevsw ABI, instead, three new functions > > > >>>>>>nt devfs_get_cdevpriv(void **datap); > > > >>>>>>int devfs_set_cdevpriv(void *priv, cdevpriv_dtr_t dtr); > > > >>>>>>void devfs_clear_cdevpriv(void); > > > >>>>>>are provided for manipulation of the per-file private data. > > > >>>>>> > > > >>>>>>devfs_set_cdevpriv assigns the priv as private data for the file > > > >>>>>>descriptor > > > >>>>>>which is used to initiate currently performed driver operation. dtr > > > >>>>>>is the function that will be called when either the last refernce to > > > >>>>>>the file goes away or devfs_clear_cdevpriv is called. > > > >>>>>> > > > >>>>>>devfs_get_cdevpriv is the obvious accessor. > > > >>>>>> > > > >>>>>>devfs_clear_cdevpriv allows to clear the private data for the still > > > >>>>>>open file. > > > >>>>>> > > > >>>>>>The synchronization of the cdev data and file private data is left > > > >>>>>>to the driver code, I did not found any generic helper mechanism that > > > >>>>>>could be useful there. > > > >>>>>> > > > >>>>>>Patch: > > > >>>>>>http://people.freebsd.org/~kib/misc/fdpriv.1.patch > > > >>>>>> > > > >>>>>>Dumb driver that shows the basic usage of the proposed KPI: > > > >>>>>>http://people.freebsd.org/~kib/misc/fpclone.c > > > >>>>>> > > > >>>>>>Previous version of the patch was tested by Peter Holm. > > > >>>>>> > > > >>>>> > > > >>>>>Hi Kostik, > > > >>>>> > > > >>>>>Are these per-instances structures intended to be used by anything > > > >>>>>other > > > >>>>>than devices? If not can we make them a union with the DTYPE_VNODE > > > >>>>>fields to save space? > > > >>>>> > > > >>>>>Thanks, > > > >>>>>Jeff > > > >>>> > > > >>>>The current version of the patch is at > > > >>>>http://people.freebsd.org/~kib/misc/fdpriv.3.patch > > > >>>> > > > >>>>Per insistence of John Baldwin and request of Eric Anholt, the > > > >>>>destructors > > > >>>>are called now when either file is last closed, or the device is > > > >>>>destroyed. > > > >>>>This versions adds only one pointer to the struct file. > > > >>>> > > > >>>>Jeff, would you, please, explicitely specify what field you propose to > > > >>>>union with the f_cdevpriv ? > > > >> > > > >>f_nextoff and f_seqcount are only used if vn_read() and vn_write() are > > > >>used. They do not apply to any other descriptors. > > > >I use the f_cdevpriv != NULL as an indicator for the necessity to enter > > > >the cdevpriv code, in particular, locking the cdevpriv_mtx, that would > > > >otherwise needed to be entered at each last close. I think that one > > > >pointer for the struct file is not too big cost, do you agree ? > > > > > > No, it's not a big cost, however if it is possible to avoid that is best. > > > > > > Can you not check the type before checking f_cdevpriv? Should we not only > > > be checking cdevpriv in contexts where we know that it is not a vnode? > > > > I am sorry, my english may be not enough, so I may interpret your > > proposal mistakenly. I read it as a suggestion to check the file type > > before accessing the f_cdevpriv. > > > > The problem with the f_cdevpriv exists only at the _fdrop(). There, we > > have a file of f_type == DTYPE_VNODE both for devfs and normal files. > > I cannot check the f_vnode since the vnode may be reclaimed. The only > > differentiator is the f_ops, that is devfs_ops_f for devfs file, and > > vnops for the normal file during the file lifetime. Unfortunately, > > f_ops is reset to the badfileops by vn_closefile before the _fdrop() is > > getting called. > > > > Reserving the flag in the f_flag looks not good due to interaction with > > the userspace. > > > > I do not want the callback to be called before the d_close() driver method > > gets a chance to clean the data. > > So, I cannot implement overlay of the cdevpriv over the seqaccess data > without some additional flag. On the other hand, use of 2 bytes for the > f_type is overkill when the range of the types is [1,8]. I divided it > to two one-byte fields, and new field is used for filetype-specific > flags. > > http://people.freebsd.org/~kib/misc/fdpriv.4.patch > shall give the same size of the struct file while not causing unconditional > acquire of the cdevpriv_mtx on the last file close. > > Jeff, do you agree with the proposed vivisection of the f_type ? I will > ask Peter to retest the patch then. I would rather add the new pointer to struct file and avoid breaking the ABI of fstat. That is, I would like this patch to be MFCable, but you can't MFC this w/o breaking the ABI of struct file since fstat(1) will be reading the full short to get the DTYPE. -- John Baldwin