From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755521Ab3JCVKL (ORCPT ); Thu, 3 Oct 2013 17:10:11 -0400 Received: from mail.linuxfoundation.org ([140.211.169.12]:57948 "EHLO mail.linuxfoundation.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754917Ab3JCVKK (ORCPT ); Thu, 3 Oct 2013 17:10:10 -0400 Date: Thu, 3 Oct 2013 14:10:08 -0700 From: Andrew Morton To: David Rientjes Cc: Bob Liu , linux-mm@kvack.org, linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org, viro@zeniv.linux.org.uk, mgorman@suse.de, hannes@cmpxchg.org, riel@redhat.com, minchan@kernel.org, Bob Liu Subject: Re: [PATCH] mm: pagevec: cleanup: drop pvec->cold argument in all places Message-Id: <20131003141008.0bf310445041a0919fe84fb9@linux-foundation.org> In-Reply-To: References: <1380357239-30102-1-git-send-email-bob.liu@oracle.com> <20130930150207.3661b5c146b6ecea84194547@linux-foundation.org> X-Mailer: Sylpheed 3.2.0beta5 (GTK+ 2.24.10; x86_64-pc-linux-gnu) Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, 1 Oct 2013 22:47:36 -0700 (PDT) David Rientjes wrote: > On Mon, 30 Sep 2013, Andrew Morton wrote: > > > > Nobody uses the pvec->cold argument of pagevec and it's also unreasonable for > > > pages in pagevec released as cold page, so drop the cold argument from pagevec. > > > > Is it unreasonable? I'd say it's unreasonable to assume that all pages > > in all cases are likely to be cache-hot. Example: what if the pages > > are being truncated and were found to be on the inactive LRU, > > unreferenced? > > > > A useful exercise would be to go through all those pagevec_init() sites > > and convince ourselves that the decision at each place was the correct > > one. > > > > Agreed, and the "cold" argument to release_pages() becomes a no-op if this > patch is merged meaning that anything released through it will > automatically go to the start of the pcp lists. If the pages aren't hot > then this is exactly the opposite of what we wanted to do; the fact that > the pvec length doesn't take into account the size of cpu cache can almost > guarantee that everything isn't cache hot. The hot/cold pages code was very marginal when we first merged it and I suspect it has rotted since. It would be a useful exercise for someone to disable it then run some benchmarks with a view to removing it all. But the problem I have with this approach is perhaps the code *could* become effective if some careful maintenance work was done on it - we should at least get the hot/cold decisions optimised before making a decision about the overall desirability of keeping it.