From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755614AbZEAX3g (ORCPT ); Fri, 1 May 2009 19:29:36 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1752262AbZEAX32 (ORCPT ); Fri, 1 May 2009 19:29:28 -0400 Received: from smtp1.linux-foundation.org ([140.211.169.13]:52624 "EHLO smtp1.linux-foundation.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751212AbZEAX31 (ORCPT ); Fri, 1 May 2009 19:29:27 -0400 Date: Fri, 1 May 2009 16:25:06 -0700 From: Andrew Morton To: Rik van Riel Cc: kosaki.motohiro@jp.fujitsu.com, peterz@infradead.org, elladan@eskimo.com, linux-kernel@vger.kernel.org, tytso@mit.edu, linux-mm@kvack.org Subject: Re: [PATCH] vmscan: evict use-once pages first (v3) Message-Id: <20090501162506.f9243dca.akpm@linux-foundation.org> In-Reply-To: <49FB8031.8000602@redhat.com> References: <20090428044426.GA5035@eskimo.com> <20090428192907.556f3a34@bree.surriel.com> <1240987349.4512.18.camel@laptop> <20090429114708.66114c03@cuia.bos.redhat.com> <2f11576a0904290907g48e94e74ye97aae593f6ac519@mail.gmail.com> <20090429131436.640f09ab@cuia.bos.redhat.com> <20090501153255.0f412420.akpm@linux-foundation.org> <49FB8031.8000602@redhat.com> X-Mailer: Sylpheed version 2.2.4 (GTK+ 2.8.20; i486-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 Fri, 01 May 2009 19:05:21 -0400 Rik van Riel wrote: > Andrew Morton wrote: > > On Wed, 29 Apr 2009 13:14:36 -0400 > > Rik van Riel wrote: > > > >> When the file LRU lists are dominated by streaming IO pages, > >> evict those pages first, before considering evicting other > >> pages. > >> > >> This should be safe from deadlocks or performance problems > >> because only three things can happen to an inactive file page: > >> 1) referenced twice and promoted to the active list > >> 2) evicted by the pageout code > >> 3) under IO, after which it will get evicted or promoted > >> > >> The pages freed in this way can either be reused for streaming > >> IO, or allocated for something else. If the pages are used for > >> streaming IO, this pageout pattern continues. Otherwise, we will > >> fall back to the normal pageout pattern. > >> > >> .. > >> > >> +int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg) > >> +{ > >> + unsigned long active; > >> + unsigned long inactive; > >> + > >> + inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE); > >> + active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE); > >> + > >> + return (active > inactive); > >> +} > > > > This function could trivially be made significantly more efficient by > > changing it to do a single pass over all the zones of all the nodes, > > rather than two passes. > > How would I do that in a clean way? copy-n-paste :( static unsigned long foo(struct mem_cgroup *mem, enum lru_list idx1, enum lru_list idx2) { int nid, zid; struct mem_cgroup_per_zone *mz; u64 total = 0; for_each_online_node(nid) for (zid = 0; zid < MAX_NR_ZONES; zid++) { mz = mem_cgroup_zoneinfo(mem, nid, zid); total += MEM_CGROUP_ZSTAT(mz, idx1); total += MEM_CGROUP_ZSTAT(mz, idx3); } return total; } dunno if that's justifiable. > The function mem_cgroup_inactive_anon_is_low and > the global versions all do the same. It would be > nice to make all four of them go fast :) > > If there is no standardized infrastructure for > getting multiple statistics yet, I can probably > whip something up. It depends how often it would be called for, I guess. One approach would be pass in a variable-length array of `enum lru_list's, get returned a same-lengthed array of totals. Or perhaps all we need to return is the sum of those totals. I'd let the memcg guys worry about this if I were you ;) > Optimizing them might make sense if it turns out to > use a significant amount of CPU. Yeah. By then it's often too late though. The sort of people for whom (num_online_nodes*MAX_NR_ZONES) is nuttily large tend not to run kernel.org kernels. From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail191.messagelabs.com (mail191.messagelabs.com [216.82.242.19]) by kanga.kvack.org (Postfix) with ESMTP id 1B4E86B003D for ; Fri, 1 May 2009 19:28:44 -0400 (EDT) Date: Fri, 1 May 2009 16:25:06 -0700 From: Andrew Morton Subject: Re: [PATCH] vmscan: evict use-once pages first (v3) Message-Id: <20090501162506.f9243dca.akpm@linux-foundation.org> In-Reply-To: <49FB8031.8000602@redhat.com> References: <20090428044426.GA5035@eskimo.com> <20090428192907.556f3a34@bree.surriel.com> <1240987349.4512.18.camel@laptop> <20090429114708.66114c03@cuia.bos.redhat.com> <2f11576a0904290907g48e94e74ye97aae593f6ac519@mail.gmail.com> <20090429131436.640f09ab@cuia.bos.redhat.com> <20090501153255.0f412420.akpm@linux-foundation.org> <49FB8031.8000602@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Sender: owner-linux-mm@kvack.org To: Rik van Riel Cc: kosaki.motohiro@jp.fujitsu.com, peterz@infradead.org, elladan@eskimo.com, linux-kernel@vger.kernel.org, tytso@mit.edu, linux-mm@kvack.org List-ID: On Fri, 01 May 2009 19:05:21 -0400 Rik van Riel wrote: > Andrew Morton wrote: > > On Wed, 29 Apr 2009 13:14:36 -0400 > > Rik van Riel wrote: > > > >> When the file LRU lists are dominated by streaming IO pages, > >> evict those pages first, before considering evicting other > >> pages. > >> > >> This should be safe from deadlocks or performance problems > >> because only three things can happen to an inactive file page: > >> 1) referenced twice and promoted to the active list > >> 2) evicted by the pageout code > >> 3) under IO, after which it will get evicted or promoted > >> > >> The pages freed in this way can either be reused for streaming > >> IO, or allocated for something else. If the pages are used for > >> streaming IO, this pageout pattern continues. Otherwise, we will > >> fall back to the normal pageout pattern. > >> > >> .. > >> > >> +int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg) > >> +{ > >> + unsigned long active; > >> + unsigned long inactive; > >> + > >> + inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE); > >> + active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE); > >> + > >> + return (active > inactive); > >> +} > > > > This function could trivially be made significantly more efficient by > > changing it to do a single pass over all the zones of all the nodes, > > rather than two passes. > > How would I do that in a clean way? copy-n-paste :( static unsigned long foo(struct mem_cgroup *mem, enum lru_list idx1, enum lru_list idx2) { int nid, zid; struct mem_cgroup_per_zone *mz; u64 total = 0; for_each_online_node(nid) for (zid = 0; zid < MAX_NR_ZONES; zid++) { mz = mem_cgroup_zoneinfo(mem, nid, zid); total += MEM_CGROUP_ZSTAT(mz, idx1); total += MEM_CGROUP_ZSTAT(mz, idx3); } return total; } dunno if that's justifiable. > The function mem_cgroup_inactive_anon_is_low and > the global versions all do the same. It would be > nice to make all four of them go fast :) > > If there is no standardized infrastructure for > getting multiple statistics yet, I can probably > whip something up. It depends how often it would be called for, I guess. One approach would be pass in a variable-length array of `enum lru_list's, get returned a same-lengthed array of totals. Or perhaps all we need to return is the sum of those totals. I'd let the memcg guys worry about this if I were you ;) > Optimizing them might make sense if it turns out to > use a significant amount of CPU. Yeah. By then it's often too late though. The sort of people for whom (num_online_nodes*MAX_NR_ZONES) is nuttily large tend not to run kernel.org kernels. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: email@kvack.org