From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id D73AAC433EF for ; Fri, 18 Feb 2022 20:24:13 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S239427AbiBRUY2 (ORCPT ); Fri, 18 Feb 2022 15:24:28 -0500 Received: from mxb-00190b01.gslb.pphosted.com ([23.128.96.19]:59476 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S236590AbiBRUY1 (ORCPT ); Fri, 18 Feb 2022 15:24:27 -0500 Received: from zeniv-ca.linux.org.uk (zeniv-ca.linux.org.uk [IPv6:2607:5300:60:148a::1]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id F35C75BD16; Fri, 18 Feb 2022 12:24:10 -0800 (PST) Received: from viro by zeniv-ca.linux.org.uk with local (Exim 4.94.2 #2 (Red Hat Linux)) id 1nL9nV-002omL-JE; Fri, 18 Feb 2022 20:24:09 +0000 Date: Fri, 18 Feb 2022 20:24:09 +0000 From: Al Viro To: Rik van Riel Cc: linux-kernel@vger.kernel.org, kernel-team@fb.com, linux-fsdevel@vger.kernel.org, paulmck@kernel.org, gscrivan@redhat.com, Eric Biederman , Chris Mason Subject: Re: [PATCH 1/2] vfs: free vfsmount through rcu work from kern_unmount Message-ID: References: <20220218183114.2867528-1-riel@surriel.com> <20220218183114.2867528-2-riel@surriel.com> <5f442a7770fe4ac06b2837e4f937d559f5d17b8b.camel@surriel.com> MIME-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Disposition: inline Content-Transfer-Encoding: 8bit In-Reply-To: Sender: Al Viro Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Fri, Feb 18, 2022 at 07:43:43PM +0000, Al Viro wrote: > On Fri, Feb 18, 2022 at 02:33:31PM -0500, Rik van Riel wrote: > > On Fri, 2022-02-18 at 19:26 +0000, Al Viro wrote: > > > On Fri, Feb 18, 2022 at 01:31:13PM -0500, Rik van Riel wrote: > > > > After kern_unmount returns, callers can no longer access the > > > > vfsmount structure. However, the vfsmount structure does need > > > > to be kept around until the end of the RCU grace period, to > > > > make sure other accesses have all gone away too. > > > > > > > > This can be accomplished by either gating each kern_unmount > > > > on synchronize_rcu (the comment in the code says it all), or > > > > by deferring the freeing until the next grace period, where > > > > it needs to be handled in a workqueue due to the locking in > > > > mntput_no_expire(). > > > > > > NAK.  There's code that relies upon kern_unmount() being > > > synchronous.  That's precisely the reason why MNT_INTERNAL > > > is treated that way in mntput_no_expire(). > > > > Fair enough. Should I make a kern_unmount_rcu() version > > that gets called just from mq_put_mnt()? > > Umm... I'm not sure you can afford having struct ipc_namespace > freed and reused before the mqueue superblock gets at least to > deactivate_locked_super(). BTW, that's a good demonstration of the problems with making those beasts async. struct mount is *not* accessed past kern_unmount(), but the objects used by the superblock might very well be - in this case they (struct ipc_namespace, pointed to by s->s_fs_data) are freed by the caller after kern_unmount() returns. And possibly reused. Now note that they are used as search keys by mqueue_get_tree() and it becomes very fishy. If you want to go that way, make it something like void put_ipc_ns(struct ipc_namespace *ns) { if (refcount_dec_and_lock(&ns->ns.count, &mq_lock)) { mq_clear_sbinfo(ns); spin_unlock(&mq_lock); kern_unmount_rcu(ns->mq_mnt); } } and give mqueue this for ->kill_sb(): static void mqueue_kill_super(struct super_block *sb) { struct ipc_namespace *ns = sb->s_fs_info; kill_litter_super(sb); do the rest of free_ipc_ns(); } One thing: kern_unmount_rcu() needs a very big warning about the caution needed from its callers. It's really not safe for general use, and it will be a temptation for folks with scalability problems like this one to just use it instead of kern_unmount() and declare the problem solved.