Return-Path: linux-nfs-owner@vger.kernel.org
Received: from mail-wg0-f43.google.com ([74.125.82.43]:54068 "EHLO
	mail-wg0-f43.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
	with ESMTP id S1755051AbaDGWCT (ORCPT
	<rfc822;linux-nfs@vger.kernel.org>); Mon, 7 Apr 2014 18:02:19 -0400
Received: by mail-wg0-f43.google.com with SMTP id x13so69121wgg.14
        for <linux-nfs@vger.kernel.org>; Mon, 07 Apr 2014 15:02:18 -0700 (PDT)
Message-ID: <1396908136.5563.9.camel@leira.trondhjem.org>
Subject: Re: NFS deadlock between 'sync' and commit after unmount....
From: Trond Myklebust <trond.myklebust@primarydata.com>
To: Jan Kara <jack@suse.cz>
Cc: Brown Neil <neilb@suse.de>, Viro Alexander <viro@zeniv.linux.org.uk>,
        NFS <linux-nfs@vger.kernel.org>
Date: Mon, 07 Apr 2014 18:02:16 -0400
In-Reply-To: <20140407202750.GE23670@quack.suse.cz>
References: <20140407135001.56ef9f36@notabene.brown>
	 <F59673DE-8047-4FE0-AFF9-6A3DC37348A5@primarydata.com>
	 <20140407202750.GE23670@quack.suse.cz>
Content-Type: text/plain; charset="UTF-8"
Mime-Version: 1.0
Sender: linux-nfs-owner@vger.kernel.org
List-ID: <linux-nfs.vger.kernel.org>

On Mon, 2014-04-07 at 22:27 +0200, Jan Kara wrote:
>   Hello,
> 
> On Mon 07-04-14 10:10:27, Trond Myklebust wrote:
> > On Apr 6, 2014, at 23:50, NeilBrown <neilb@suse.de> wrote:
> > > I've just hit a deadlock in NFS that seems very strange.
> > > The kernel is 3.14-rc8 which some local changes which shouldn't affect the
> > > deadlocking code.
> > > 
> > > Shortly after umounting the NFS filesystem with "umount -f" (though I don't
> > > think the -f is important), I ran "sync".
> > > 
> > > The sync is now stuck in
> > > 
> > > [<ffffffff81197fc1>] sync_inodes_sb+0xa1/0x1c0
> > > [<ffffffff8119cd99>] sync_inodes_one_sb+0x19/0x20
> > > [<ffffffff81173372>] iterate_supers+0xb2/0x110
> > > [<ffffffff8119cfd0>] sys_sync+0x30/0x90
> > > [<ffffffff81aa4622>] system_call_fastpath+0x16/0x1b
> > > [<ffffffffffffffff>] 0xffffffffffffffff
> > > 
> > > while kworker/u16:1 is stuck:
> > > 
> > > [<ffffffff815420b3>] call_rwsem_down_write_failed+0x13/0x20
> > > [<ffffffff81172889>] deactivate_super+0x39/0x60
> > > [<ffffffff812d56f1>] nfs_sb_deactive+0x21/0x30
> > > [<ffffffff812d2ef9>] __put_nfs_open_context+0xc9/0x100
> > > [<ffffffff812d2f3b>] put_nfs_open_context+0xb/0x10
> > > [<ffffffff812ddd14>] nfs_commitdata_release+0x14/0x30
> > > [<ffffffff812ddd4a>] nfs_commit_release+0x1a/0x20
> > > [<ffffffff81a45a05>] rpc_free_task+0x25/0x70
> > > [<ffffffff81a45fd8>] rpc_do_put_task+0x78/0x80
> > > [<ffffffff81a45feb>] rpc_put_task+0xb/0x10
> > > [<ffffffff812de3fe>] nfs_initiate_commit+0xce/0x110
> > > [<ffffffff812df112>] nfs_commit_list+0x62/0x90
> > > [<ffffffff812dfd26>] nfs_commit_inode+0xa6/0x170
> > > [<ffffffff812dfe4d>] nfs_write_inode+0x5d/0xa0
> > > [<ffffffff81300d69>] nfs4_write_inode+0x9/0x10
> > > [<ffffffff811978ec>] __writeback_single_inode+0x10c/0x2c0
> > > [<ffffffff811987ea>] writeback_sb_inodes+0x2ca/0x450
> > > [<ffffffff81198b2c>] wb_writeback+0xec/0x320
> > > [<ffffffff81199365>] bdi_writeback_workfn+0x115/0x4c0
> > > [<ffffffff810a595b>] process_one_work+0x16b/0x430
> > > [<ffffffff810a6619>] worker_thread+0x119/0x3a0
> > > [<ffffffff810ac2bd>] kthread+0xcd/0xf0
> > > [<ffffffff81aa457c>] ret_from_fork+0x7c/0xb0
> > > [<ffffffffffffffff>] 0xffffffffffffffff
> > > 
> > > 
> > > So sync is holding sb->s_umount, queued some bdi work on the filesystem
> > > and is waiting for it to complete.  Mean while, that work has (I think)
> > > submitted a 'commit' (via ->write_inode) and that commit wants to
> > > deactivate_super and so needs to get ->s_umount.
> > > 
> > > I suspect this could happen even more easily with a lazy unmount.
> > > 
> > > It seems that this commit request is that last thing that is keeping
> > > ->s_active elevated and it deadlocks trying to drop the last s_active.
> > > 
> > > I have no idea how to fix it....  help?
> > > 
> > 
> > The problem seems to be the use of iterate_supers(), which grabs a
> > passive reference, and conflicts with our use of an active reference in
> > the open context.
>   Yeah, we cannot really do otherwise in iterate_supers() - we have to grab
> some superblock reference and we don't really want to get an active one
> since that would result in spurious EBUSY returns from umount.
> 
> Cannot we just punt the deactivate_super() call to a workqueue to avoid
> this deadlock? It's a bit ugly but it should do the trick. Or is
> nfs_sb_deactive() called too often and we'd see some adverse effects for
> that? We could also offload it to workqueue only in the special case where
> sb->s_active == 1. That should be really rare then but it's a bit ugly
> poking in VFS internals.

The activate/deactivate super is basically there to save our bacon when
NFS file state extends beyond the usual VFS path walk, open() and
close(). Examples include sillyrename and NFSv4 delegations. Even
ordinary read and write state can extend beyond close() if the user
decides to 'kill -9' in the wrong places.
In most of these situations, we need to keep a dentry around until we're
finished, which means that we want to keep the super block alive too.

>  BTW why is nfs_sb_active() safe doing just
> atomic_inc() on s_active? grab_super() is more careful...

We're never grabbing a reference without already holding a reference via
a struct path or a struct file passed down to us by the VFS. However we
may want to keep the struct dentry around for longer than the lifetime
of that struct path/file.

-- 
Trond Myklebust
Linux NFS client maintainer, PrimaryData
trond.myklebust@primarydata.com