From mboxrd@z Thu Jan 1 00:00:00 1970 From: Hugh Dickins Subject: Re: fs_stack/eCryptfs: remove 3rd arg of copy_attr_all, add locking to copy_inode_size Date: Thu, 3 Apr 2008 20:26:54 +0100 (BST) Message-ID: References: <200804031907.m33J70fQ025437@agora.fsl.cs.sunysb.edu> Mime-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Cc: "Josef 'Jeff' Sipek" , akpm@linux-foundation.org, linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org, Al Viro , hch@infradead.org To: Erez Zadok Return-path: Received: from extu-mxob-1.symantec.com ([216.10.194.28]:33008 "EHLO extu-mxob-1.symantec.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755583AbYDCT07 (ORCPT ); Thu, 3 Apr 2008 15:26:59 -0400 In-Reply-To: <200804031907.m33J70fQ025437@agora.fsl.cs.sunysb.edu> Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Thu, 3 Apr 2008, Erez Zadok wrote: > In message <20080403182001.GB30189@josefsipek.net>, "Josef 'Jeff' Sipek" writes: > > I think you need to check CONFIG_PREEMPT as well. > > I'm not sure if it's needed in case of CONFIG_PREEMPT. Anyone? The code > for i_size_write (below), and the comment at the top of the function, > suggest that the spinlock is needed only to prevent the lots seqcount. Correct. > BTW, some time ago I reviewed all callers of i_size_write. I did so again > just now, and the results were the same: > > - a LOT of callers of i_size_write don't take any lock They mostly know that i_mutex is already held (as i_size_write comment mentions); but I believe that's up to the individual filesystem. > - some take another spinlock in a different data structure > - those that do take the spinlock, do so unconditionally > - only unionfs and fs/stack.c wrap the spinlock in > > #if BITS_PER_LONG == 32 && defined(CONFIG_SMP) I chose to follow the #ifdeffery of i_size_write(), but you could do it unconditionally if you prefer: just a little more overhead when it's not needed. As I've said elsewhere, I don't think the result can be entirely safe against concurrent changes in the lower filesystem, using different locking; but I don't know how resilient unionfs is expected to be against messing directly with lower at the same time as upper level. Hugh