Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Tue, 27 Jun 2023 11:28:29 +0200
From:   Jan Kara <jack@suse.cz>
To:     Baokun Li <libaokun1@huawei.com>
Cc:     Jan Kara <jack@suse.cz>, linux-fsdevel@vger.kernel.org,
        linux-ext4@vger.kernel.org, linux-kernel@vger.kernel.org,
        yi.zhang@huawei.com, yangerkun@huawei.com, chengzhihao1@huawei.com,
        yukuai3@huawei.com
Subject: Re: [PATCH] quota: fix race condition between dqput() and
 dquot_mark_dquot_dirty()
Message-ID: <20230627092829.d3s3x4nkprux7jmo@quack3>
References: <20230616085608.42435-1-libaokun1@huawei.com>
 <20230616152824.ndpgvkegvvip2ahh@quack3>
 <c8daf4a0-769f-f769-50f6-8b7063542499@huawei.com>
 <20230622145620.hk3bdjxtlr64gtzl@quack3>
 <b73894fc-0c7a-0503-25ad-ab5a9dfbd852@huawei.com>
 <20230626130957.kvfli23djxc2opkq@quack3>
 <2486ec73-55e0-00cb-fc76-97b9b285a9ce@huawei.com>
 <20230627083406.hhjf55e2tqnwqaf6@quack3>
 <fe7d3b03-4d08-34ac-a695-a5c57c751aeb@huawei.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <fe7d3b03-4d08-34ac-a695-a5c57c751aeb@huawei.com>
Precedence: bulk

On Tue 27-06-23 17:08:27, Baokun Li wrote:
> Hello!
> 
> On 2023/6/27 16:34, Jan Kara wrote:
> > Hello!
> > 
> > On Mon 26-06-23 21:55:49, Baokun Li wrote:
> > > On 2023/6/26 21:09, Jan Kara wrote:
> > > > On Sun 25-06-23 15:56:10, Baokun Li wrote:
> > > > > > > I think we can simply focus on the race between the DQ_ACTIVE_B flag and
> > > > > > > the DQ_MOD_B flag, which is the core problem, because the same quota
> > > > > > > should not have both flags. These two flags are protected by dq_list_lock
> > > > > > > and dquot->dq_lock respectively, so it makes sense to add a
> > > > > > > wait_on_dquot() to ensure the accuracy of DQ_ACTIVE_B.
> > > > > > But the fundamental problem is not only the race with DQ_MOD_B setting. The
> > > > > > dquot structure can be completely freed by the time
> > > > > > dquot_claim_space_nodirty() calls dquot_mark_dquot_dirty() on it. That's
> > > > > > why I think making __dquot_transfer() obey dquot_srcu rules is the right
> > > > > > solution.
> > > > > Yes, now I also think that making __dquot_transfer() obey dquot_srcu
> > > > > rules is a better solution. But with inode->i_lock protection, why would
> > > > > the dquot structure be completely freed?
> > > > Well, when dquot_claim_space_nodirty() calls mark_all_dquot_dirty() it does
> > > > not hold any locks (only dquot_srcu). So nothing prevents dquot_transfer()
> > > > to go, swap dquot structure pointers and drop dquot references and after
> > > > that mark_all_dquot_dirty() can use a stale pointer to call
> > > > mark_dquot_dirty() on already freed memory.
> > > > 
> > > No, this doesn't look like it's going to happen.  The
> > > mark_all_dquot_dirty() uses a pointer array pointer, the dquot in the
> > > array is dynamically changing, so after swap dquot structure pointers,
> > > mark_all_dquot_dirty() uses the new pointer, and the stale pointer is
> > > always destroyed after swap, so there is no case of using the stale
> > > pointer here.
> > There is a case - CPU0 can prefetch the values from dquots[] array into its
> > local cache, then CPU1 can update the dquots[] array (these writes can
> > happily stay in CPU1 store cache invisible to other CPUs) and free the
> > dquots via dqput(). Then CPU0 can pass the prefetched dquot pointers to
> > mark_dquot_dirty(). There are no locks or memory barries preventing CPUs
> > from ordering instructions and memory operations like this in the code...
> > You can read Documentation/memory-barriers.txt about all the perils current
> > CPU architecture brings wrt coordination of memory accesses among CPUs ;)
> > 
> > 								Honza
> 
> Got it!
> 
> Sorry for misunderstanding you (I thought "completely freed" meant
> dquot_destroy(), but you should have meant dquot_release()).

Well, the dquot can even get to dquot_destroy(). There's nothing really
preventing CPU2 going into memory reclaim and free the dquot in
dqcache_shrink_scan() still before CPU0 even calls mark_dquot_dirty() on
it. Sure such timing on real hardware is very unlikely but in a VM where a
virtual CPU can get starved for a significant amount of time this could
happen.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR