From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752201Ab2BDHmx (ORCPT ); Sat, 4 Feb 2012 02:42:53 -0500 Received: from idcmail-mo1so.shaw.ca ([24.71.223.10]:38661 "EHLO idcmail-mo1so.shaw.ca" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751657Ab2BDHmu convert rfc822-to-8bit (ORCPT ); Sat, 4 Feb 2012 02:42:50 -0500 X-Cloudmark-SP-Filtered: true X-Cloudmark-SP-Result: v=1.1 cv=W+ZrOcpwHTeQLvZQ9vbIxXRdvbK5p6H2/T6SuQCN924= c=1 sm=1 a=-dEdcLQ1GN0A:10 a=BLceEmwcHowA:10 a=kj9zAlcOel0A:10 a=k8hqzbHzjweV03hO1KKrVA==:17 a=NvX0e53ABAVZLJygwU0A:9 a=CjuIK1q_8ugA:10 a=CtS22CzEzCZcoyER:21 a=xcl7UIqVqkBu_JbL:21 a=HpAAvcLHHh0Zw7uRqdWCyQ==:117 Subject: Re: [RFC] killing boilerplate checks in ->link/->mkdir/->rename Mime-Version: 1.0 (Apple Message framework v1084) Content-Type: text/plain; charset=us-ascii From: Andreas Dilger In-Reply-To: <20120203170300.GW23916@ZenIV.linux.org.uk> Date: Sat, 4 Feb 2012 00:42:48 -0700 Cc: linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org, Linus Torvalds Content-Transfer-Encoding: 8BIT Message-Id: References: <4F27C6EB.2070305@suse.cz> <20120202012258.GQ23916@ZenIV.linux.org.uk> <20120202212400.GR23916@ZenIV.linux.org.uk> <20120203170300.GW23916@ZenIV.linux.org.uk> To: Al Viro X-Mailer: Apple Mail (2.1084) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 2012-02-03, at 10:03 AM, Al Viro wrote: > On Fri, Feb 03, 2012 at 01:25:26AM -0700, Andreas Dilger wrote: >> On 2012-02-02, at 2:24 PM, Al Viro wrote: >>> FWIW, there's something we really should've done a long time ago: putting >>> that limit into sb->s_max_links. With 0 meaning "leave all checks to >>> ->link/->mkdir/->rename". Something like the following would make a >>> reasonable start - just the conversion of obvious cases. As the next >>> step I'd probably initialize it as ~0U instead of 0 and let the filesystems >>> that want something trickier (reiserfs, ext4, gfs2, ocfs2) explicitly set >>> it to 0 in their foo_fill_super(). That would take care of a bunch of cases >>> where we forgot to do those checks (ubifs, hfsplus, jffs2, ramfs, etc.) and >>> it's probably a saner default anyway. >> >> This would also give userspace some hope of pathconf(path, _PC_LINK_MAX) >> returning the actual value from the filesystem, instead of hard-coding >> this into glibc itself based on the statfs-returned f_type magic value. > > *snort* > > Even skipping the standard flame about pathconf() as an API, this will > not work. > * we have filesystems that do not allow link creation at all and > do keep track of subdirectories count in i_nlink of directories. What > would you have them store? As it is, ~0U works just fine, but pathconf() > users won't be happy with it. > * we have filesystems that allow unlimited subdirectories, while > limiting the number of links to non-directories; ->s_max_links == 0 will > work just fine, but won't make pathconf() happy. > * we have filesystems that have more complex rules re links to > non-directory (see mail from Chris in this thread). What would you have > pathconf() do? No comment on how good an API pathconf() is, but getting a per-filesystem value from the kernel has to be better than a hard-coded value coded in a library in userspace. Cheers, Andreas