Date: Mon, 3 Apr 2017 08:26:43 -0700
From: Davidlohr Bueso <dave@stgolabs.net>
To: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: mingo@kernel.org, peterz@infradead.org, akpm@linux-foundation.org,
        jack@suse.cz, kirill.shutemov@linux.intel.com, mhocko@suse.com,
        mgorman@techsingularity.net, linux-kernel@vger.kernel.org,
        Davidlohr Bueso <dbueso@suse.de>
Subject: Re: [PATCH 1/5] locking: Introduce range reader/writer lock
Message-ID: <20170403152643.GB2270@linux-80c1.suse>
References: <1488863010-13028-1-git-send-email-dave@stgolabs.net>
 <1488863010-13028-2-git-send-email-dave@stgolabs.net>
 <2f7628f4-58e1-22c4-ccbe-3106c15cb405@linux.vnet.ibm.com>
 <20170328163918.GA27446@linux-80c1.suse>
 <20170403161917.4ef46f17@nimbus>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1; format=flowed
Content-Disposition: inline
Content-Transfer-Encoding: 8BIT
In-Reply-To: <20170403161917.4ef46f17@nimbus>
User-Agent: Mutt/1.5.24 (2015-08-30)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1706
Lines: 45

On Mon, 03 Apr 2017, Laurent Dufour wrote:

>Le Tue, 28 Mar 2017 09:39:18 -0700,
>Davidlohr Bueso <dave@stgolabs.net> a ?crit :
>> I'll wait to see if there are any more concerns and send a v2 with
>> your corrections.
>
>Hi Bavidlohr, I think there is a major issue regarding the task
>catching a signal in wait_for_range().
>I can see it when a thread is catching a signal, the process deadlock
>in exit path.
>
>Let's imagine all these tasks waiting for the complete range lock, so
>range doesn't matter:
>
>A get the lock in write
>B want the read lock => B->blocking_range=1 (because of A)
>C want the write lock => C->blocking_range=2 (A,B)
>D want the read lock => D->blocking_range=3 (A,B,C)
>=> C catch a signal and exit wait_for_ranges()
>A release the lock
>        => B->blocking_range=0
>        => D->blocking_range=2 (D has not seen C removal)
>=> B get the lock
>B release the lock
>        => D->blocking_range=1
>
>D remains blocked while no one has the lock !
>
>The issue is when removing a task from the interval tree, we
>should decrement all the blocking_ranges of the task added to that
>range after the one leaving... I can't see an easy fix for that :(
>
>Am I right ?

Yes. Peter had also mentioned the issue too. One way I though of fixing
the problem was to track the jiffies timestamp in a per range_rwlock
basis for when it was added, and in the signal_pending() case, along with
removing the lock from the tree, we iterate the tree again and decrement
the blocking_ranges for those with a higher timestamp. It would add some
overhead, but again this is the unlikely() case. It also adds an extra 8
bytes of footprint, but this is usually stack allocated.

Thanks,
Davidlohr