From mboxrd@z Thu Jan 1 00:00:00 1970 From: Bernd Schubert Subject: Re: [PATCH 3/3] nfsd: vfs_llseek() with O_32BITHASH or O_64BITHASH Date: Thu, 28 Jul 2011 11:19:07 +0200 Message-ID: <4E31298B.7050800@itwm.fraunhofer.de> References: <20110727110148.204979.49551.stgit@localhost.localdomain> <20110727110259.204979.56782.stgit@localhost.localdomain> <20110727210344.GC9066@infradead.org> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Cc: linux-nfs@vger.kernel.org, linux-ext4@vger.kernel.org, linux-fsdevel@vger.kernel.org, yong.fan@whamcloud.com, adilger@whamcloud.com, tytso@mit.edu To: Christoph Hellwig Return-path: In-Reply-To: <20110727210344.GC9066@infradead.org> Sender: linux-ext4-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org On 07/27/2011 11:03 PM, Christoph Hellwig wrote: > On Wed, Jul 27, 2011 at 01:02:59PM +0200, Bernd Schubert wrote: >> Use 32-bit or 64-bit llseek() hashes for directory offsets depending on >> the NFS version. NFSv2 gets 32-bit hashes only. > > Independent of the O_ vs FMODE thing make sure you pass the correct > flag at open time, instead of racy runtime modifications. > Christoph, before I'm going to work further on the patch sets, I have few questions first. Could you please help me with that? file->f_mode is set in __dentry_open() based on O_ flags. So if f_mode is supposed to be set directly during the NFS open call we would need O_ *and* FMODE flags, Now I still do not understand why we cannot add a flag *after* the open call and only in nfsd_readdir()? I do not see any races there. nfsd_readdir() gets its 'struct file' from get_empty_filp() and __dentry_open(). Now as 'struct file' is allocated by get_empty_filp() it also cannot be used by any other thread. nfsd_readdir() just reads the directory and closes it immedeatily after the readdir(). So where is there supposed to be a race? And lastly, if we are going to set f_mode directly at open time, we have the choice to 1) Specify those new O_ flags for all files. While setting a flag is a cheap operation, it still wastes CPU cycles for file opens, as files do not need that flag. 2) Duplicate nfsd_open() to implement a new function for directories only. I think not a good idea either. 3) Rewrite the nfsd_open() function to allow to set flags from calling functions. That would mean to update the nfsd code at at least 8 places. Do we really want to go that way? So altogether, updating the patches to replace O_ by FMODE flags is easy, but setting that flag in nfsd at open time, would mean a large overhead. Thanks, Bernd