linux-kernel - Re: Bug in kernel 2.6.31, Slow wb

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:	Sat, 1 Aug 2009 10:02:28 +0800
From:	Wu Fengguang <fengguang.wu@...el.com>
To:	Martin Bligh <mbligh@...gle.com>
Cc:	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>,
	Michael Rubin <mrubin@...gle.com>,
	"sandeen@...hat.com" <sandeen@...hat.com>,
	Michael Davidson <md@...gle.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Peter Zijlstra <peterz@...radead.org>
Subject: Re: Bug in kernel 2.6.31, Slow wb_kupdate writeout

On Fri, Jul 31, 2009 at 03:55:44AM +0800, Martin Bligh wrote:
> > Note that this is a simple fix that may have suboptimal write performance.
> > Here is an old reasoning:
> >
> >        http://lkml.org/lkml/2009/3/28/235
> 
> The other thing I've been experimenting with is to disable the per-page
> check in write_cache_pages, ie:
> 
>                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
>                                 wb_stats_inc(WB_STATS_WCP_SECTION_CONG);
>                                 wbc->encountered_congestion = 1;
>                                 /* done = 1; */
> 
> This treats the congestion limits as soft, but encourages us to write
> back in larger, more efficient chunks. If that's not going to scare
> people unduly, I can submit that as well.

This risks hitting the hard limit (nr_requests), and block everyone,
including the ones with higher priority (ie. kswapd).

On the other hand, the simple fix in previous mails won't necessarily
act too sub-optimal. It's only a potential one. There is a window of
(1/16)*(nr_requests)*request_size (= 128*256KB/16 = 4MB) between
congestion-on and congestion-off states. So for the best we can inject
a big 4MB chunk into the async write queue once it becomes uncongested.

I have a writeback debug patch that can help find out how
that works out in your real world workloads (by monitoring
nr_to_write). You can also try doubling the ratio (1/16) in
blk_queue_congestion_threshold(), to see how an increased
congestion-on-off window may help.

Thanks,
Fengguang

View attachment "writeback-debug-2.6.31.patch" of type "text/x-diff" (2714 bytes)