From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-fsdevel-owner@vger.kernel.org>
Received: from bombadil.infradead.org ([198.137.202.9]:32884 "EHLO
        bombadil.infradead.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
        with ESMTP id S1755873AbcJQTyl (ORCPT
        <rfc822;linux-fsdevel@vger.kernel.org>);
        Mon, 17 Oct 2016 15:54:41 -0400
Date: Mon, 17 Oct 2016 21:54:28 +0200
From: Peter Zijlstra <peterz@infradead.org>
To: Christoph Hellwig <hch@lst.de>
Cc: Jeff Moyer <jmoyer@redhat.com>, viro@zeniv.linux.org.uk,
        jack@suse.cz, dmonakhov@openvz.org, linux-fsdevel@vger.kernel.org,
        linux-aio@kvack.org, linux-kernel@vger.kernel.org,
        Oleg Nesterov <oleg@redhat.com>,
        Mikulas Patocka <mpatocka@redhat.com>
Subject: Re: [PATCH] aio: fix a use after free (and fix freeze protection of
 aio writes)
Message-ID: <20161017195428.GZ3568@worktop.programming.kicks-ass.net>
References: <1476597082-15317-1-git-send-email-hch@lst.de>
 <x49twcabzy4.fsf@segfault.boston.devel.redhat.com>
 <20161017185552.GA24653@lst.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20161017185552.GA24653@lst.de>
Sender: linux-fsdevel-owner@vger.kernel.org
List-ID: <linux-fsdevel.vger.kernel.org>

On Mon, Oct 17, 2016 at 08:55:52PM +0200, Christoph Hellwig wrote:
> On Mon, Oct 17, 2016 at 02:19:47PM -0400, Jeff Moyer wrote:
> > This ends up being a call to __sb_end_write:
> > 
> > void __sb_end_write(struct super_block *sb, int level)
> > {
> >         percpu_up_read(sb->s_writers.rw_sem + level-1);
> > }
> > 
> > Nothing guarantees that submission and completion happen on the same
> > CPU.  Is this safe?
> 
> Good point.  From my reading of the percpu_rwsem implementation it
> is not safe to release it from a different CPU.  Which makes me
> wonder how we can protect aio writes properly here..

percpu-rwsem has the same semantics as regular rwsems, so preemptible
and 'owner' stuff.

Therefore we must support doing up from a different cpu than we did down
on; the owner could've been migrated while we held it.

And while there's a metric ton of tricky in the implementation, this
part is actually fairly straight forward. We only care about the direct
sum of the per-cpu counter, see readers_active_check() -> per_cpu_sum().

So one cpu doing an inc and another doing a dec summed is still 0.