Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932084Ab1BWQhS (ORCPT ); Wed, 23 Feb 2011 11:37:18 -0500 Received: from mx1.redhat.com ([209.132.183.28]:14022 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754886Ab1BWQhQ (ORCPT ); Wed, 23 Feb 2011 11:37:16 -0500 Date: Wed, 23 Feb 2011 17:36:39 +0100 From: Andrea Arcangeli To: Arthur Marsh Cc: Clemens Ladisch , alsa-user@lists.sourceforge.net, linux-kernel@vger.kernel.org, Mel Gorman Subject: Re: [Alsa-user] new source of MIDI playback slow-down identified - 5a03b051ed87e72b959f32a86054e1142ac4cf55 thp: use compaction in kswapd for GFP_ATOMIC order > 0 Message-ID: <20110223163639.GM31195@random.random> References: <4D6367B3.9050306@googlemail.com> <20110222134047.GT13092@random.random> <20110222161513.GC13092@random.random> <4D63F6C0.7060204@internode.on.net> <20110223162432.GL31195@random.random> MIME-Version: 1.0 Content-Type: multipart/mixed; boundary="IiVenqGWf+H9Y6IX" Content-Disposition: inline In-Reply-To: <20110223162432.GL31195@random.random> Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 7910 Lines: 210 --IiVenqGWf+H9Y6IX Content-Type: text/plain; charset=us-ascii Content-Disposition: inline On Wed, Feb 23, 2011 at 05:24:32PM +0100, Andrea Arcangeli wrote: > On Wed, Feb 23, 2011 at 04:17:44AM +1030, Arthur Marsh wrote: > > OK, these patches applied together against upstream didn't cause a crash > > but I did observe: > > > > significant slowdowns of MIDI playback (moreso than in previous cases, > > and with less than 20 Meg of swap file in use); > > > > kswapd0 sharing equal top place in CPU usage at times (e.g. 20 percent). > > > > If I should try only one of the patches or something else entirely, > > please let me know. > > Yes, with irq off, schedule won't run and need_resched won't get set. > > So let's try this. > > In case this doesn't fix I definitely give it up with compaction in > kswapd as GFP_ATOMIC will likely not get an huge benefit out of > compaction anyway because 1) the allocations from GFP_ATOMIC are > likely short lived, 2) the cost of the compaction scan loop + > migration may be higher than the benefit we get from jumbo frames > > So if this doesn't fix it, I'll already post a definitive fix that > removes compaction from kswapd (but leaves it enabled for direct > reclaim for all order sizes including kernel stack). I already > verified that this solves not just the midi latency issue but the > other server benchmark that I'm dealing with. Then we can think at > compaction+kswapd later for 2.6.39 but I think we need to close this > kswapd issue in time for 2.6.38. So if the below isn't enough the next > patch should be applied. I'll get those two patches tested in the > server load too, and I'll let you know how the results are when it > finishes. > > Please apply also the attached "kswapd-high-wmark" before the below > one. If the previous patch please test the below after the attached patch (as usual). If the previous patch (last attempt for 2.6.38 to add compaction in kswapd) fails this is the way to go for 2.6.38. === Subject: compaction: fix high compaction latencies and remove compaction-kswapd From: Andrea Arcangeli We need to proper spin_unlock_irq/cond_resched in the compaction loop to avoid hurting latencies. We must also stop calling compaction from kswapd as that creates too high load during memory pressure that can't be offseted by the improved performance of hugepage allocations. NOTE: this is not related to THP as all THP allocations uses __GFP_NO_KSWAPD, this is only related to usually small order allocations like the kernel stack that make kswapd go wild with compaction. Signed-off-by: Andrea Arcangeli --- mm/compaction.c | 40 +++++++++++++++++++++------------------- 1 file changed, 21 insertions(+), 19 deletions(-) --- a/mm/compaction.c +++ b/mm/compaction.c @@ -271,9 +271,27 @@ static unsigned long isolate_migratepage } /* Time to isolate some pages for migration */ + cond_resched(); spin_lock_irq(&zone->lru_lock); for (; low_pfn < end_pfn; low_pfn++) { struct page *page; + int unlocked = 0; + + /* give a chance to irqs before checking need_resched() */ + if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) { + spin_unlock_irq(&zone->lru_lock); + unlocked = 1; + } + if (need_resched() || spin_is_contended(&zone->lru_lock)) { + if (fatal_signal_pending(current)) + break; + if (!unlocked) + spin_unlock_irq(&zone->lru_lock); + cond_resched(); + spin_lock_irq(&zone->lru_lock); + } else if (unlocked) + spin_lock_irq(&zone->lru_lock); + if (!pfn_valid_within(low_pfn)) continue; nr_scanned++; @@ -397,10 +415,7 @@ static int compact_finished(struct zone return COMPACT_COMPLETE; /* Compaction run is not finished if the watermark is not met */ - if (cc->compact_mode != COMPACT_MODE_KSWAPD) - watermark = low_wmark_pages(zone); - else - watermark = high_wmark_pages(zone); + watermark = low_wmark_pages(zone); watermark += (1 << cc->order); if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0)) @@ -413,15 +428,6 @@ static int compact_finished(struct zone if (cc->order == -1) return COMPACT_CONTINUE; - /* - * Generating only one page of the right order is not enough - * for kswapd, we must continue until we're above the high - * watermark as a pool for high order GFP_ATOMIC allocations - * too. - */ - if (cc->compact_mode == COMPACT_MODE_KSWAPD) - return COMPACT_CONTINUE; - /* Direct compactor: Is a suitable page free? */ for (order = cc->order; order < MAX_ORDER; order++) { /* Job done if page is free of the right migratetype */ @@ -543,8 +549,7 @@ static int compact_zone(struct zone *zon unsigned long compact_zone_order(struct zone *zone, int order, gfp_t gfp_mask, - bool sync, - int compact_mode) + bool sync) { struct compact_control cc = { .nr_freepages = 0, @@ -553,7 +558,6 @@ unsigned long compact_zone_order(struct .migratetype = allocflags_to_migratetype(gfp_mask), .zone = zone, .sync = sync, - .compact_mode = compact_mode, }; INIT_LIST_HEAD(&cc.freepages); INIT_LIST_HEAD(&cc.migratepages); @@ -599,8 +603,7 @@ unsigned long try_to_compact_pages(struc nodemask) { int status; - status = compact_zone_order(zone, order, gfp_mask, sync, - COMPACT_MODE_DIRECT_RECLAIM); + status = compact_zone_order(zone, order, gfp_mask, sync); rc = max(status, rc); /* If a normal allocation would succeed, stop compacting */ @@ -631,7 +634,6 @@ static int compact_node(int nid) .nr_freepages = 0, .nr_migratepages = 0, .order = -1, - .compact_mode = COMPACT_MODE_DIRECT_RECLAIM, }; zone = &pgdat->node_zones[zoneid]; --IiVenqGWf+H9Y6IX Content-Type: text/plain; charset=us-ascii Content-Disposition: attachment; filename=kswapd-high_wmark Subject: vmscan: kswapd must not free more than high_wmark pages From: Andrea Arcangeli When the min_free_kbytes is set with `hugeadm --set-recommended-min_free_kbytes" or with THP enabled (which runs the equivalent of "hugeadm --set-recommended-min_free_kbytes" to activate anti-frag at full effectiveness automatically at boot) the high wmark of some zone is as high as ~88M. 88M free on a 4G system isn't horrible, but 88M*8 = 704M free on a 4G system is definitely unbearable. This only tends to be visible on 4G systems with tiny over-4g zone where kswapd insists to reach the high wmark on the over-4g zone but doing so it shrunk up to 704M from the normal zone by mistake. For the trivial case where kswapd isn't waken until all zones hit the low wmark and there is no concurrency between allocator and kswapd freeing, rotating more the tiny above4g lru than "high-low" despite we only allocated "high-low" cache into it doesn't sound obviously right either. Bigger gap to me looks like will do more harm than good and if we need a real guarantee of balancing we should rotate the allocations across the zones (bigger lru in a zone will require it to be hit more frequently because it'll rotate slower than the other zones, the bias should not even dependent on the zone size but on the lru size). Signed-off-by: Andrea Arcangeli --- mm/vmscan.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) --- a/mm/vmscan.c +++ b/mm/vmscan.c @@ -2408,7 +2408,7 @@ loop_again: * zone has way too many pages free already. */ if (!zone_watermark_ok_safe(zone, order, - 8*high_wmark_pages(zone), end_zone, 0)) + high_wmark_pages(zone), end_zone, 0)) shrink_zone(priority, zone, &sc); reclaim_state->reclaimed_slab = 0; nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, --IiVenqGWf+H9Y6IX-- -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/