Recalculate per-cpu page allocator batch and high limits after deferred meminit

Both the percpu_pagelist_fraction sysctl handler and memory hotplug have a common requirement of updating the pcpu page allocation batch and high values. Split the relevant helper to share common code.

No functional change.

Signed-off-by Mel Gorman <[email protected]>

---
 mm/page_alloc.c | 23 ++++++++++++-----------
 1 file changed, 12 insertions(+), 11 deletions(-)

Both the percpu_pagelist_fraction sysctl handler and memory hotplug have a common requirement of updating the pcpu page allocation batch and high values. Split the relevant helper to share common code.

No functional change.

Signed-off-by Mel Gorman <[email protected]>

---
 mm/page_alloc.c | 23 ++++++++++++-----------
 1 file changed, 12 insertions(+), 11 deletions(-)

mm/page_alloc.c +12 -11

index c0b2e0306720..cafe568d36f6 100644
@@ -7983,6 +7983,15 @@ int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
 	return 0;
 }
 
+static void __zone_pcp_update(struct zone *zone)
+{
+	unsigned int cpu;
+
+	for_each_possible_cpu(cpu)
+		pageset_set_high_and_batch(zone,
+				per_cpu_ptr(zone->pageset, cpu));
+}
+
 /*
  * percpu_pagelist_fraction - changes the pcp->high for each zone on each
  * cpu.  It is the fraction of total pages in each zone that a hot per cpu

@@ -8014,13 +8023,8 @@ int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
 	if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
 		goto out;
 
-	for_each_populated_zone(zone) {
-		unsigned int cpu;
-
-		for_each_possible_cpu(cpu)
-			pageset_set_high_and_batch(zone,
-					per_cpu_ptr(zone->pageset, cpu));
-	}
+	for_each_populated_zone(zone)
+		__zone_pcp_update(zone);
 out:
 	mutex_unlock(&pcp_batch_high_lock);
 	return ret;

@@ -8519,11 +8523,8 @@ void free_contig_range(unsigned long pfn, unsigned int nr_pages)
  */
 void __meminit zone_pcp_update(struct zone *zone)
 {
-	unsigned cpu;
 	mutex_lock(&pcp_batch_high_lock);
-	for_each_possible_cpu(cpu)
-		pageset_set_high_and_batch(zone,
-				per_cpu_ptr(zone->pageset, cpu));
+	__zone_pcp_update(zone);
 	mutex_unlock(&pcp_batch_high_lock);
 }
 #endif

Signature · 1 line

2.16.4

Deferred memory initialisation updates zone->managed_pages during the initialisation phase but before that finishes, the per-cpu page allocator (pcpu) calculates the number of pages allocated/freed in batches as well as the maximum number of pages allowed on a per-cpu list. As zone->managed_pages is not up to date yet, the pcpu initialisation calculates inappropriately low batch and high values.

This increases zone lock contention quite severely in some cases with the degree of severity depending on how many CPUs share a local zone and the size of the zone. A private report indicated that kernel build times were excessive with extremely high system CPU usage. A perf profile indicated that a large chunk of time was lost on zone->lock contention.

This patch recalculates the pcpu batch and high values after deferred initialisation completes on each node. It was tested on a 2-socket AMD EPYC 2 machine using a kernel compilation workload -- allmodconfig and all available CPUs.

mmtests configuration: config-workload-kernbench-max Configuration was modified to build on a fresh XFS partition.

kernbench

                                5.4.0-rc3              5.4.0-rc3
                                  vanilla         resetpcpu-v1r1
Amean     user-256    13249.50 (   0.00%)    15928.40 * -20.22%*
Amean     syst-256    14760.30 (   0.00%)     4551.77 *  69.16%*
Amean     elsp-256      162.42 (   0.00%)      118.46 *  27.06%*
Stddev    user-256       42.97 (   0.00%)       50.83 ( -18.30%)
Stddev    syst-256      336.87 (   0.00%)       33.70 (  90.00%)
Stddev    elsp-256        2.46 (   0.00%)        0.81 (  67.01%)

                   5.4.0-rc3   5.4.0-rc3
                     vanillaresetpcpu-v1r1
Duration User       39766.24    47802.92
Duration System     44298.10    13671.93

Duration Elapsed 519.11 387.65

The patch reduces system CPU usage by 69.16% and total build time by 27.06%. The variance of system CPU usage is also much reduced.

Cc [email protected] # v4.15+

Signed-off-by Mel Gorman <[email protected]>

---
 mm/page_alloc.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

Deferred memory initialisation updates zone->managed_pages during the initialisation phase but before that finishes, the per-cpu page allocator (pcpu) calculates the number of pages allocated/freed in batches as well as the maximum number of pages allowed on a per-cpu list. As zone->managed_pages is not up to date yet, the pcpu initialisation calculates inappropriately low batch and high values.

This increases zone lock contention quite severely in some cases with the degree of severity depending on how many CPUs share a local zone and the size of the zone. A private report indicated that kernel build times were excessive with extremely high system CPU usage. A perf profile indicated that a large chunk of time was lost on zone->lock contention.

This patch recalculates the pcpu batch and high values after deferred initialisation completes on each node. It was tested on a 2-socket AMD EPYC 2 machine using a kernel compilation workload -- allmodconfig and all available CPUs.

mmtests configuration: config-workload-kernbench-max Configuration was modified to build on a fresh XFS partition.

kernbench

                                5.4.0-rc3              5.4.0-rc3
                                  vanilla         resetpcpu-v1r1
Amean     user-256    13249.50 (   0.00%)    15928.40 * -20.22%*
Amean     syst-256    14760.30 (   0.00%)     4551.77 *  69.16%*
Amean     elsp-256      162.42 (   0.00%)      118.46 *  27.06%*
Stddev    user-256       42.97 (   0.00%)       50.83 ( -18.30%)
Stddev    syst-256      336.87 (   0.00%)       33.70 (  90.00%)
Stddev    elsp-256        2.46 (   0.00%)        0.81 (  67.01%)

                   5.4.0-rc3   5.4.0-rc3
                     vanillaresetpcpu-v1r1
Duration User       39766.24    47802.92
Duration System     44298.10    13671.93

Duration Elapsed 519.11 387.65

The patch reduces system CPU usage by 69.16% and total build time by 27.06%. The variance of system CPU usage is also much reduced.

Cc [email protected] # v4.15+

Signed-off-by Mel Gorman <[email protected]>

---
 mm/page_alloc.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

mm/page_alloc.c +8 -2

index cafe568d36f6..0a0dd74edc83 100644
@@ -1818,6 +1818,14 @@ static int __init deferred_init_memmap(void *data)
 	 */
 	while (spfn < epfn)
 		nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
+
+	/*
+	 * The number of managed pages has changed due to the initialisation
+	 * so the pcpu batch and high limits needs to be updated or the limits
+	 * will be artificially small.
+	 */
+	zone_pcp_update(zone);
+
 zone_empty:
 	pgdat_resize_unlock(pgdat, &flags);
 

@@ -8516,7 +8524,6 @@ void free_contig_range(unsigned long pfn, unsigned int nr_pages)
 	WARN(count != 0, "%d pages are still in use!\n", count);
 }
 
-#ifdef CONFIG_MEMORY_HOTPLUG
 /*
  * The zone indicated has a new number of managed_pages; batch sizes and percpu
  * page high values need to be recalulated.

@@ -8527,7 +8534,6 @@ void __meminit zone_pcp_update(struct zone *zone)
 	__zone_pcp_update(zone);
 	mutex_unlock(&pcp_batch_high_lock);
 }
-#endif
 
 void zone_pcp_reset(struct zone *zone)
 {

Signature · 1 line

2.16.4

Quoted 37 lines · Mel Gorman

Deferred memory initialisation updates zone->managed_pages during the initialisation phase but before that finishes, the per-cpu page allocator (pcpu) calculates the number of pages allocated/freed in batches as well as the maximum number of pages allowed on a per-cpu list. As zone->managed_pages is not up to date yet, the pcpu initialisation calculates inappropriately low batch and high values.

This increases zone lock contention quite severely in some cases with the degree of severity depending on how many CPUs share a local zone and the size of the zone. A private report indicated that kernel build times were excessive with extremely high system CPU usage. A perf profile indicated that a large chunk of time was lost on zone->lock contention.

This patch recalculates the pcpu batch and high values after deferred initialisation completes on each node. It was tested on a 2-socket AMD EPYC 2 machine using a kernel compilation workload -- allmodconfig and all available CPUs.

mmtests configuration: config-workload-kernbench-max Configuration was modified to build on a fresh XFS partition.

kernbench

                                5.4.0-rc3              5.4.0-rc3
                                  vanilla         resetpcpu-v1r1
Amean     user-256    13249.50 (   0.00%)    15928.40 * -20.22%*
Amean     syst-256    14760.30 (   0.00%)     4551.77 *  69.16%*
Amean     elsp-256      162.42 (   0.00%)      118.46 *  27.06%*
Stddev    user-256       42.97 (   0.00%)       50.83 ( -18.30%)
Stddev    syst-256      336.87 (   0.00%)       33.70 (  90.00%)
Stddev    elsp-256        2.46 (   0.00%)        0.81 (  67.01%)

                   5.4.0-rc3   5.4.0-rc3
                     vanillaresetpcpu-v1r1
Duration User       39766.24    47802.92
Duration System     44298.10    13671.93

Duration Elapsed 519.11 387.65

Quoted 2 lines

The patch reduces system CPU usage by 69.16% and total build time by 27.06%. The variance of system CPU usage is also much reduced.

The fix makes sense. It would be nice to see the difference in the batch sizes from the initial setup compared to the one after the deferred intialization is done

Quoted 1 line

Cc [email protected] # v4.15+

Hmm, are you sure about 4.15? Doesn't this go all the way down to deferred initialization? I do not see any recent changes on when setup_per_cpu_pageset is called.

Quoted 6 lines

Quoted 1 line

Cc [email protected] # v4.15+

Hmm, are you sure about 4.15? Doesn't this go all the way down to deferred initialization? I do not see any recent changes on when setup_per_cpu_pageset is called.

No, I'm not 100% sure. It looks like this was always an issue from the code but did not happen on at least one 4.12-based distribution kernel for reasons that are non-obvious. Either way, the tag should have been "v4.1+"

Quoted 10 lines · On Fri, 18 Oct 2019 15:09:59 +0100 Mel Gorman <[email protected]>…

Quoted 6 lines

Quoted 1 line

Cc [email protected] # v4.15+

Hmm, are you sure about 4.15? Doesn't this go all the way down to deferred initialization? I do not see any recent changes on when setup_per_cpu_pageset is called.

No, I'm not 100% sure. It looks like this was always an issue from the code but did not happen on at least one 4.12-based distribution kernel for reasons that are non-obvious. Either way, the tag should have been "v4.1+"

I could mark

mm-pcp-share-common-code-between-memory-hotplug-and-percpu-sysctl-handler.patch mm-meminit-recalculate-pcpu-batch-and-high-limits-after-init-completes.patch

as Cc: <[email protected]> [4.1+]

But for backporting purposes it's a bit cumbersome that [2/3] is the important patch. I think I'll switch the ordering so that mm-meminit-recalculate-pcpu-batch-and-high-limits-after-init-completes.patch is the first patch and the other two can be queued for 5.5-rc1, OK?

Also, is a Reported-by:Matt appropriate here?

Memory hotplug needs to be able to reset and reinit the pcpu allocator batch and high limits but this action is internal to the VM. Move the declaration to internal.h

Signed-off-by Mel Gorman <[email protected]>

---
 include/linux/mm.h | 3 ---
 mm/internal.h      | 3 +++
 2 files changed, 3 insertions(+), 3 deletions(-)

Memory hotplug needs to be able to reset and reinit the pcpu allocator batch and high limits but this action is internal to the VM. Move the declaration to internal.h

Signed-off-by Mel Gorman <[email protected]>

---
 include/linux/mm.h | 3 ---
 mm/internal.h      | 3 +++
 2 files changed, 3 insertions(+), 3 deletions(-)

include/linux/mm.h +0 -3

index cc292273e6ba..22d6104f2341 100644
@@ -2219,9 +2219,6 @@ void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...);
 
 extern void setup_per_cpu_pageset(void);
 
-extern void zone_pcp_update(struct zone *zone);
-extern void zone_pcp_reset(struct zone *zone);
-
 /* page_alloc.c */
 extern int min_free_kbytes;
 extern int watermark_boost_factor;

mm/internal.h +3 -0

index 0d5f720c75ab..0a3d41c7b3c5 100644
@@ -165,6 +165,9 @@ extern void post_alloc_hook(struct page *page, unsigned int order,
 					gfp_t gfp_flags);
 extern int user_min_free_kbytes;
 
+extern void zone_pcp_update(struct zone *zone);
+extern void zone_pcp_reset(struct zone *zone);
+
 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
 
 /*

Signature · 1 line

2.16.4

A private report stated that system CPU usage was excessive on an AMD EPYC 2 machine while building kernels with much longer build times than expected. The issue is partially explained by high zone lock contention due to the per-cpu page allocator batch and high limits being calculated incorrectly. This series addresses a large chunk of the problem. Patch 1 is mostly cosmetic but prepares for patch 2 which is the real fix. Patch 3 is definiely cosmetic but was noticed while implementing the fix. Proper details are in the changelog for patch 2.

 include/linux/mm.h |  3 ---
 mm/internal.h      |  3 +++
 mm/page_alloc.c    | 33 ++++++++++++++++++++-------------
 3 files changed, 23 insertions(+), 16 deletions(-)

Quoted 9 lines · Mel Gorman

A private report stated that system CPU usage was excessive on an AMD EPYC 2 machine while building kernels with much longer build times than expected. The issue is partially explained by high zone lock contention due to the per-cpu page allocator batch and high limits being calculated incorrectly. This series addresses a large chunk of the problem. Patch 1 is mostly cosmetic but prepares for patch 2 which is the real fix. Patch 3 is definiely cosmetic but was noticed while implementing the fix. Proper details are in the changelog for patch 2.

Quoted 4 lines

 include/linux/mm.h |  3 ---
 mm/internal.h      |  3 +++
 mm/page_alloc.c    | 33 ++++++++++++++++++++-------------
 3 files changed, 23 insertions(+), 16 deletions(-)

Just to confirm, these patches don't fix the issue we're seeing on the EPYC 2 machines, but they do return the batch sizes to sensible values.

Quoted 15 lines · Matt Fleming

Quoted 13 lines · Mel Gorman

A private report stated that system CPU usage was excessive on an AMD EPYC 2 machine while building kernels with much longer build times than expected. The issue is partially explained by high zone lock contention due to the per-cpu page allocator batch and high limits being calculated incorrectly. This series addresses a large chunk of the problem. Patch 1 is mostly cosmetic but prepares for patch 2 which is the real fix. Patch 3 is definiely cosmetic but was noticed while implementing the fix. Proper details are in the changelog for patch 2.

 include/linux/mm.h |  3 ---
 mm/internal.h      |  3 +++
 mm/page_alloc.c    | 33 ++++++++++++++++++++-------------
 3 files changed, 23 insertions(+), 16 deletions(-)

Quoted 2 lines

Just to confirm, these patches don't fix the issue we're seeing on the EPYC 2 machines, but they do return the batch sizes to sensible values.

To be clear, does the patch a) fix *some* of the issue and there is something else also going on that needs to be chased down or b) has no impact on build time or system CPU usage on your machine?