From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-fsdevel-owner@vger.kernel.org>
Received: from mfb02-md.ns.itscom.net ([175.177.155.110]:53440 "EHLO
	mfb02-md.ns.itscom.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
	with ESMTP id S1750774AbcFQUz4 (ORCPT
	<rfc822;linux-fsdevel@vger.kernel.org>);
	Fri, 17 Jun 2016 16:55:56 -0400
Received: from mail04-md.ns.itscom.net (mail04-md.ns.itscom.net [175.177.155.114])
	by mfb02-md.ns.itscom.net (Postfix) with ESMTP id 35283170BB94
	for <linux-fsdevel@vger.kernel.org>; Sat, 18 Jun 2016 05:50:33 +0900 (JST)
From: "J. R. Okajima" <hooanon05g@gmail.com>
To: viro@ZenIV.linux.org.uk, linux-fsdevel@vger.kernel.org
Subject: Q. hlist_bl_add_head_rcu() in d_alloc_parallel()
Date: Sat, 18 Jun 2016 05:50:30 +0900
Message-ID: <13136.1466196630@jrobl>
Sender: linux-fsdevel-owner@vger.kernel.org
List-ID: <linux-fsdevel.vger.kernel.org>


I am afraid there may exist another violation of "no lookups on the same
name in parallel" rule, but I am not sure.

Roughly d_alloc_parallel() behaves like this.

struct dentry *d_alloc_parallel()
{
	new = d_alloc(parent, name);

	rcu_read_lock();
	hlist_bl_lock(b);
	rcu_read_unlock();
	hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
		if (!matched_dentry_found)
			continue;
		dget(dentry);
		hlist_bl_unlock(b);
		return dentry;
	}
	hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
	hlist_bl_unlock(b);
	return new;
}

When two processes try opening a single existing file and enters
d_alloc_parallel() at the same time, only one process wins and should
succeeds hlist_bl_add_head_rcu(). The other process should find the
dentry in d_u.d_in_lookup_hash and return 'dentry' (instead of
'new'). Am I right?

My question is when will 'new' be added into d_u.d_in_lookup_hash?
It should be between these two lines, I guess.
	rcu_read_unlock();
	hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
But can it surely happen?
If 'new' is not added here because someone else is in rcu_read_lock
region or other reason, then both processes will add the same named but
different dentry?

Is it better to change the lock/unlock-order like this?

	rcu_read_unlock();
	rcu_barrier();
	hlist_bl_lock(b);
	hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {


J. R. Okajima