From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andy Lutomirski Subject: Re: [RFC PATCH] fs: Add user_file_or_path_at and use it for truncate Date: Tue, 27 Aug 2013 13:28:27 -0700 Message-ID: References: <7d1419dda1da70a8ad915f85b093a58b86bcaf3b.1377630856.git.luto@amacapital.net> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Cc: Al Viro , Willy Tarreau , Ingo Molnar , "security@kernel.org" , Linux Kernel Mailing List , Oleg Nesterov , Linux FS Devel , Brad Spengler To: Linus Torvalds Return-path: In-Reply-To: Sender: linux-kernel-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org On Tue, Aug 27, 2013 at 12:32 PM, Linus Torvalds wrote: > On Tue, Aug 27, 2013 at 12:16 PM, Andy Lutomirski wrote: >> This is an experiment to see if we can get nice semantics for all syscalls >> that either follow symlinks or allow AT_EMPTY_PATH without jumping through >> enormous hoops. This converts truncate (although you can't tell using >> truncate from coreutils, because it actually uses open + ftruncate). > > So this seems *way* too complex. I'd much rather see "nd->flags" get a > LOOKUP_READONLY flag, for example, that gets set by > proc_pid_follow_link() when it hits a read-only file descriptor (and > gets cleared by other lookups). > > Wouldn't that be *much* more straightforward? It would if it works. It certainly would for truncate, setxattr, etc. There are funny cases, though. For example, execing an O_WRONLY fd should probably fail, as should opening an O_WRONLY fd as O_RDWR. O_APPEND is also funny. flink will (I suspect) always want to be a bit special. There are also O_PATH fds, and I'm not sure what the semantics of O_PATH fds are or should be when they refer to something other than a directory. The benefit of my approach is that it's really obvious that truncate("/proc/self/fd/N") and ftruncate(N) do exactly the same thing. The downside is that the namei code is a bit gross and there was more rearranging of ftruncate than I would have liked. (I also lost the benefit of fget_light, but I could fix that by passing a struct fd around instead of a struct file *.) --Andy