From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753436Ab2GIMDr (ORCPT ); Mon, 9 Jul 2012 08:03:47 -0400 Received: from mail.digidescorp.com ([50.73.98.161]:2404 "EHLO mail.digidescorp.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753090Ab2GIMDo (ORCPT ); Mon, 9 Jul 2012 08:03:44 -0400 DomainKey-Signature: a=rsa-sha1; s=MDaemon; d=digidescorp.com; c=simple; q=dns; h=message-id:from; b=bWQi6HPZQmKS2+LS+ap+/45fFCpLSe/XpqEpr7kFGVO4CFuEWXAFS1RBWX18 wXdkEVpsj4yJjjdjBXd9DimUzHZuJYNo8YoLix37n80uTIDb1T3/L22Pj mZ0QM54LnWAd9PdEyhXvvOivzQC9JP2S4Zp23IXS3X2LwtkZ/ueC2Q=; X-Spam-Processed: mail.digidescorp.com, Mon, 09 Jul 2012 07:03:42 -0500 (not processed: message from trusted or authenticated source) X-Authenticated-Sender: steve@digidescorp.com X-Return-Path: prvs=1537d3712a=steve@digidescorp.com X-Envelope-From: steve@digidescorp.com X-MDaemon-Deliver-To: linux-kernel@vger.kernel.org Message-ID: <1341835422.2215.6.camel@iscandar.digidescorp.com> Subject: Re: [PATCH 2/2] fat (exportfs): reconnect file handles to evicted inodes/dentries From: "Steven J. Magnani" To: OGAWA Hirofumi Cc: linux-kernel@vger.kernel.org Date: Mon, 09 Jul 2012 07:03:42 -0500 In-Reply-To: <87zk7bo6aq.fsf@devron.myhome.or.jp> References: <1341342576-15394-1-git-send-email-steve@digidescorp.com> <1341342576-15394-3-git-send-email-steve@digidescorp.com> <87pq8bokcp.fsf@devron.myhome.or.jp> <1341606824.2214.14.camel@iscandar.digidescorp.com> <871ukobntl.fsf@devron.myhome.or.jp> <1341623782.2021.9.camel@iscandar> <87wr2g9kh3.fsf@devron.myhome.or.jp> <1341679286.2435.12.camel@iscandar> <87zk7bo6aq.fsf@devron.myhome.or.jp> Content-Type: text/plain; charset="UTF-8" X-Mailer: Evolution 3.4.3 (3.4.3-1.fc17) Mime-Version: 1.0 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Sun, 2012-07-08 at 02:00 +0900, OGAWA Hirofumi wrote: > "Steven J. Magnani" writes: > > >> You mean the unhashed inode is created by ->get_parent()? If so, the > >> root cause sounds like ->get_parent() itself. If not, I'm not > >> understanding the meaning of the temporary/unofficial inode here. > > > > Maybe "private" is a better word than "unofficial". Private inodes are > > created anywhere fat_new_inode (nee fat_build_unhashed_inode) is called > > directly, instead of through fat_build_inode. So yes, this is on the > > get_parent paths (via fat_lookup_dir), and also on the fh_to_dentry path > > when inode reconstruction is necessary. > > > > With private inodes, I don't see how anyone but the code that created > > them could find them to lock them. The reason they're private is that > > they're temporary aliases; at the time they're created, we don't have > > enough information to register them in a way that others could find > > them. A lookup, etc. operation will look for the inode of the "drivers" > > directory, not the ".." of the "usb" directory. We do need these private > > inodes in order to walk directory entries. I don't think they're a > > problem that needs solving; if we didn't use private inodes, we'd still > > need a way to walk directory entries in the context of these NFS > > operations, and there would still be potential races between that and > > other operations on the filesystem. > > How do you prevent to modify or free the those inode/blocks from other > path? Yeah, it is racy. And if races is not solved, that's simply wrong > and not solution. > > Although I'm not thinking deeply about NFS support on FAT. Just a idea, > the one of possible solutions would be register it to hash, and find it > on all path. So, all path will use same inode and lock. > > We need the key, possible key is - if it is only directory, FAT may be > able to use i_start as additional search key. Interesting idea. I think this, and reformulating the FAT NFS file handle to include the parent's i_ino, will greatly simplify (and speed up) the code. I am having a hard time seeing how inclusion of i_pos in the file handle is useful. The only scenarios I can conceive where an i_ino lookup fails and an i_pos lookup succeeds are the object referenced by the file handle has: (A) been replaced by something else, or (2) had its inode evicted, but later re-instantiated Perhaps the design was intended to support case (2), but since the file handle of the re-instantiated inode will differ from the original (since i_ino and i_generation will have changed), NFS clients will cry ESTALE anyway. Steve