From: ebiederm@xmission.com (Eric W. Biederman) Subject: Re: [PATCH] procfs: make /proc style symlinks behave like "normal" symlinks Date: Thu, 19 Nov 2009 10:57:08 -0800 Message-ID: References: <1258638251-20034-1-git-send-email-jlayton@redhat.com> <20091119132833.30bc93a4@barsoom.rdu.redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org, pavel@ucw.cz To: Jeff Layton Return-path: Received: from out01.mta.xmission.com ([166.70.13.231]:59488 "EHLO out01.mta.xmission.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751901AbZKSS5G (ORCPT ); Thu, 19 Nov 2009 13:57:06 -0500 In-Reply-To: <20091119132833.30bc93a4@barsoom.rdu.redhat.com> (Jeff Layton's message of "Thu\, 19 Nov 2009 13\:28\:33 -0500") Sender: linux-fsdevel-owner@vger.kernel.org List-ID: Jeff Layton writes: > On Thu, 19 Nov 2009 09:07:16 -0800 > ebiederm@xmission.com (Eric W. Biederman) wrote: > >> >> Nacked-by: "Eric W. Biederman" >> >> This is broken. If the referenced file is in a different mount namespace >> the path returned could point to a completely different path in your >> own mount namespace. Even in your own mount namespace this makes the >> proc symlinks racy and not guaranteed to return the file of interest. >> >> I don't see any hope of this approach ever working. >> >> Eric >> > > Then is proc_pid_readlink broken in the same way? proc_pid_readlink has the same deficiencies. The race is fundamental to all readlink operations, the difference is that for normal symlinks it is a don't care, and for proc it is incorrect behavior if you follow the symlink to the wrong file. If you are dealing with a file in a different namespace or a socket what you get back doesn't actually work as a file in your local namespace but that is the best we can do with a pathname, and if you know the context of what is going on readlink is still useful. Adding all of the short comings to followlink that readlink has is a problem, especially as followlink does much better now. At a practical level I think your changes are much easier to exploit than Pavels contrived example. I really don't have any problems with your first patch to proc to add the missing revalidate. Eric