Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Mel Gorman <mgorman@techsingularity.net>
To:     Andrew Morton <akpm@linux-foundation.org>
Cc:     Aaron Lu <aaron.lu@intel.com>,
        Dave Hansen <dave.hansen@linux.intel.com>,
        Vlastimil Babka <vbabka@suse.cz>,
        Michal Hocko <mhocko@kernel.org>,
        Jesper Dangaard Brouer <brouer@redhat.com>,
        LKML <linux-kernel@vger.kernel.org>,
        Linux-MM <linux-mm@kvack.org>,
        Mel Gorman <mgorman@techsingularity.net>
Subject: [PATCH 5/5] mm/page_alloc: Limit number of high-order pages on PCP during bulk free
Date:   Tue, 15 Feb 2022 14:51:11 +0000
Message-Id: <20220215145111.27082-6-mgorman@techsingularity.net>
In-Reply-To: <20220215145111.27082-1-mgorman@techsingularity.net>
References: <20220215145111.27082-1-mgorman@techsingularity.net>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

When a PCP is mostly used for frees then high-order pages can exist on PCP
lists for some time. This is problematic when the allocation pattern is all
allocations from one CPU and all frees from another resulting in colder
pages being used. When bulk freeing pages, limit the number of high-order
pages that are stored on the PCP lists.

Netperf running on localhost exhibits this pattern and while it does
not matter for some machines, it does matter for others with smaller
caches where cache misses cause problems due to reduced page reuse.
Pages freed directly to the buddy list may be reused quickly while still
cache hot where as storing on the PCP lists may be cold by the time
free_pcppages_bulk() is called.

Using perf kmem:mm_page_alloc, the 5 most used page frames were

5.17-rc3
  13041 pfn=0x111a30
  13081 pfn=0x5814d0
  13097 pfn=0x108258
  13121 pfn=0x689598
  13128 pfn=0x5814d8

5.17-revert-highpcp
 192009 pfn=0x54c140
 195426 pfn=0x1081d0
 200908 pfn=0x61c808
 243515 pfn=0xa9dc20
 402523 pfn=0x222bb8

5.17-full-series
 142693 pfn=0x346208
 162227 pfn=0x13bf08
 166413 pfn=0x2711e0
 166950 pfn=0x2702f8

The spread is wider as there is still time before pages freed to one
PCP get released with a tradeoff between fast reuse and reduced zone
lock acquisition.

From the machine used to gather the traces, the headline performance
was equivalent.

netperf-tcp
                            5.17.0-rc3             5.17.0-rc3             5.17.0-rc3
                               vanilla  mm-reverthighpcp-v1r1  mm-highpcplimit-v1r12
Hmean     64         839.93 (   0.00%)      840.77 (   0.10%)      835.34 *  -0.55%*
Hmean     128       1614.22 (   0.00%)     1622.07 *   0.49%*     1604.18 *  -0.62%*
Hmean     256       2952.00 (   0.00%)     2953.19 (   0.04%)     2959.46 (   0.25%)
Hmean     1024     10291.67 (   0.00%)    10239.17 (  -0.51%)    10287.05 (  -0.04%)
Hmean     2048     17335.08 (   0.00%)    17399.97 (   0.37%)    17125.73 *  -1.21%*
Hmean     3312     22628.15 (   0.00%)    22471.97 (  -0.69%)    22414.24 *  -0.95%*
Hmean     4096     25009.50 (   0.00%)    24752.83 *  -1.03%*    24620.03 *  -1.56%*
Hmean     8192     32745.01 (   0.00%)    31682.63 *  -3.24%*    32475.31 (  -0.82%)
Hmean     16384    39759.59 (   0.00%)    36805.78 *  -7.43%*    39291.42 (  -1.18%)

From a 1-socket skylake machine with a small CPU cache that suffers
more if cache misses are too high

netperf-tcp
                            5.17.0-rc3             5.17.0-rc3             5.17.0-rc3
                               vanilla    mm-reverthighpcp-v1     mm-highpcplimit-v1
Min       64         935.38 (   0.00%)      939.40 (   0.43%)      940.11 (   0.51%)
Min       128       1831.69 (   0.00%)     1856.15 (   1.34%)     1849.30 (   0.96%)
Min       256       3560.61 (   0.00%)     3659.25 (   2.77%)     3654.12 (   2.63%)
Min       1024     13165.24 (   0.00%)    13444.74 (   2.12%)    13281.71 (   0.88%)
Min       2048     22706.44 (   0.00%)    23219.67 (   2.26%)    23027.31 (   1.41%)
Min       3312     30960.26 (   0.00%)    31985.01 (   3.31%)    31484.40 (   1.69%)
Min       4096     35149.03 (   0.00%)    35997.44 (   2.41%)    35891.92 (   2.11%)
Min       8192     48064.73 (   0.00%)    49574.05 (   3.14%)    48928.89 (   1.80%)
Min       16384    58017.25 (   0.00%)    60352.93 (   4.03%)    60691.14 (   4.61%)
Hmean     64         938.95 (   0.00%)      941.50 *   0.27%*      940.47 (   0.16%)
Hmean     128       1843.10 (   0.00%)     1857.58 *   0.79%*     1855.83 *   0.69%*
Hmean     256       3573.07 (   0.00%)     3667.45 *   2.64%*     3662.08 *   2.49%*
Hmean     1024     13206.52 (   0.00%)    13487.80 *   2.13%*    13351.11 *   1.09%*
Hmean     2048     22870.23 (   0.00%)    23337.96 *   2.05%*    23149.68 *   1.22%*
Hmean     3312     31001.99 (   0.00%)    32206.50 *   3.89%*    31849.40 *   2.73%*
Hmean     4096     35364.59 (   0.00%)    36490.96 *   3.19%*    36112.91 *   2.12%*
Hmean     8192     48497.71 (   0.00%)    49954.05 *   3.00%*    49384.50 *   1.83%*
Hmean     16384    58410.86 (   0.00%)    60839.80 *   4.16%*    61362.12 *   5.05%*

Note that this was a machine that did not benefit from caching high-order
pages and performance is almost restored with the series applied. It's not
fully restored as cache misses are still higher. This is a trade-off
between optimising for a workload that does all allocs on one CPU and frees
on another or more general workloads that need high-order pages for SLUB
and benefit from avoiding zone->lock for every SLUB refill/drain.

Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
---
 mm/page_alloc.c | 26 +++++++++++++++++++++-----
 1 file changed, 21 insertions(+), 5 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 6881175b27df..cfb3cbad152c 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -3314,10 +3314,15 @@ static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
 	return true;
 }
 
-static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
+static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
+		       bool free_high)
 {
 	int min_nr_free, max_nr_free;
 
+	/* Free everything if batch freeing high-order pages. */
+	if (unlikely(free_high))
+		return pcp->count;
+
 	/* Check for PCP disabled or boot pageset */
 	if (unlikely(high < batch))
 		return 1;
@@ -3338,11 +3343,12 @@ static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
 	return batch;
 }
 
-static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
+static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
+		       bool free_high)
 {
 	int high = READ_ONCE(pcp->high);
 
-	if (unlikely(!high))
+	if (unlikely(!high || free_high))
 		return 0;
 
 	if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
@@ -3362,17 +3368,27 @@ static void free_unref_page_commit(struct page *page, unsigned long pfn,
 	struct per_cpu_pages *pcp;
 	int high;
 	int pindex;
+	bool free_high;
 
 	__count_vm_event(PGFREE);
 	pcp = this_cpu_ptr(zone->per_cpu_pageset);
 	pindex = order_to_pindex(migratetype, order);
 	list_add(&page->lru, &pcp->lists[pindex]);
 	pcp->count += 1 << order;
-	high = nr_pcp_high(pcp, zone);
+
+	/*
+	 * As high-order pages other than THP's stored on PCP can contribute
+	 * to fragmentation, limit the number stored when PCP is heavily
+	 * freeing without allocation. The remainder after bulk freeing
+	 * stops will be drained from vmstat refresh context.
+	 */
+	free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
+
+	high = nr_pcp_high(pcp, zone, free_high);
 	if (pcp->count >= high) {
 		int batch = READ_ONCE(pcp->batch);
 
-		free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp, pindex);
+		free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
 	}
 }
 
-- 
2.31.1