From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753506AbaAUEqr (ORCPT ); Mon, 20 Jan 2014 23:46:47 -0500 Received: from ipmail07.adl2.internode.on.net ([150.101.137.131]:6208 "EHLO ipmail07.adl2.internode.on.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751380AbaAUEqh (ORCPT ); Mon, 20 Jan 2014 23:46:37 -0500 X-IronPort-Anti-Spam-Filtered: true X-IronPort-Anti-Spam-Result: ApkHAOX63VJ5LKVw/2dsb2JhbABZgwu3FYVSgQwXdIIlAQEBAwE6HCMFCwgDDgoJJQ8FJQMhE4d9B8QGFxaOIEkHgySBFASYIZIZg0EogS0 Date: Tue, 21 Jan 2014 15:46:32 +1100 From: Dave Chinner To: Andy Lutomirski Cc: "linux-kernel@vger.kernel.org" , "linux-mm@kvack.org" , Linux FS Devel Subject: Re: Dirty deleted files cause pointless I/O storms (unless truncated first) Message-ID: <20140121044632.GA25923@dastard> References: MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Mon, Jan 20, 2014 at 04:59:23PM -0800, Andy Lutomirski wrote: > The code below runs quickly for a few iterations, and then it slows > down and the whole system becomes laggy for far too long. > > Removing the sync_file_range call results in no I/O being performed at > all (which means that the kernel isn't totally screwing this up), and > changing "4096" to SIZE causes lots of I/O but without > the going-out-to-lunch bit (unsurprisingly). More details please. hardware, storage, kernel version, etc. I can't reproduce any slowdown with the code as posted on a VM running 3.31-rc5 with 16GB RAM and an SSD w/ ext4 or XFS. The workload is only generating about 80 IOPS on ext4 so even a slow spindle should be able handle this without problems... > Surprisingly, uncommenting the ftruncate call seems to fix the > problem. This suggests that all the necessary infrastructure to avoid > wasting time writing to deleted files is there but that it's not > getting used. Not surprising at all - if it's stuck in a writeback loop somewhere, truncating the file will terminate writeback because it end up being past EOF and so stops immediately... Cheers, Dave. -- Dave Chinner david@fromorbit.com