[RFC PATCH] mm/workingset : judge file page activity via timestamp

[Zhaoyang Huang]

From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>

This patch introduce timestamp into workingset's entry and judge if the page
is active or inactive via active_file/refault_ratio instead of refault distance.

The original thought is coming from the logs we got from trace_printk in this
patch, we can find about 1/5 of the file pages' refault are under the
scenario[1],which will be counted as inactive as they have a long refault distance
in between access. However, we can also know from the time information that the
page refault quickly as comparing to the average refault time which is calculated
by the number of active file and refault ratio. We want to save these kinds of
pages from evicted earlier as it used to be. The refault ratio is the value
which can reflect lru's average file access frequency and also can be deemed as a
prediction of future.

The patch is tested on an android system and reduce 30% of page faults, while
60% of the pages remain the original status as (refault_distance < active_file)
indicates. Pages status got from ftrace during the test can refer to [2].

[1]
system_server workingset_refault: WKST_ACT[0]:rft_dis 265976, act_file 34268 rft_ratio 3047 rft_time 0 avg_rft_time 11 refault 295592 eviction 29616 secs 97 pre_secs 97
HwBinder:922  workingset_refault: WKST_ACT[0]:rft_dis 264478, act_file 35037 rft_ratio 3070 rft_time 2 avg_rft_time 11 refault 310078 eviction 45600 secs 101 pre_secs 99

[2]
WKST_ACT[0]:   original--INACTIVE  commit--ACTIVE
WKST_ACT[1]:   original--ACTIVE    commit--ACTIVE
WKST_INACT[0]: original--INACTIVE  commit--INACTIVE
WKST_INACT[1]: original--ACTIVE    commit--INACTIVE

Signed-off-by: Zhaoyang Huang <huangzhaoyang@xxxxxxxxx>
---
 include/linux/mmzone.h |   1 +
 mm/workingset.c        | 120 +++++++++++++++++++++++++++++++++++++++++++++----
 2 files changed, 112 insertions(+), 9 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 32699b2..6f30673 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -240,6 +240,7 @@ struct lruvec {
 	atomic_long_t			inactive_age;
 	/* Refaults at the time of last reclaim cycle */
 	unsigned long			refaults;
+	atomic_long_t			refaults_ratio;
 #ifdef CONFIG_MEMCG
 	struct pglist_data *pgdat;
 #endif
diff --git a/mm/workingset.c b/mm/workingset.c
index 40ee02c..66c177b 100644
--- a/mm/workingset.c
+++ b/mm/workingset.c
@@ -160,6 +160,21 @@
 			 MEM_CGROUP_ID_SHIFT)
 #define EVICTION_MASK	(~0UL >> EVICTION_SHIFT)
 
+#ifdef CONFIG_64BIT
+#define EVICTION_SECS_POS_SHIFT 20
+#define EVICTION_SECS_SHRINK_SHIFT 4
+#define EVICTION_SECS_POS_MASK  ((1UL << EVICTION_SECS_POS_SHIFT) - 1)
+#else
+#ifndef CONFIG_MEMCG
+#define EVICTION_SECS_POS_SHIFT 12
+#define EVICTION_SECS_SHRINK_SHIFT 4
+#define EVICTION_SECS_POS_MASK  ((1UL << EVICTION_SECS_POS_SHIFT) - 1)
+#else
+#define EVICTION_SECS_POS_SHIFT 0
+#define EVICTION_SECS_SHRINK_SHIFT 0
+#define NO_SECS_IN_WORKINGSET
+#endif
+#endif
 /*
  * Eviction timestamps need to be able to cover the full range of
  * actionable refaults. However, bits are tight in the radix tree
@@ -169,10 +184,54 @@
  * evictions into coarser buckets by shaving off lower timestamp bits.
  */
 static unsigned int bucket_order __read_mostly;
-
+#ifdef NO_SECS_IN_WORKINGSET
+static void pack_secs(unsigned long *peviction) { }
+static unsigned int unpack_secs(unsigned long entry) {return 0; }
+#else
+/*
+ * Shrink the timestamp according to its value and store it together
+ * with the shrink size in the entry.
+ */
+static void pack_secs(unsigned long *peviction)
+{
+	unsigned int secs;
+	unsigned long eviction;
+	int order;
+	int secs_shrink_size;
+	struct timespec ts;
+
+	get_monotonic_boottime(&ts);
+	secs = (unsigned int)ts.tv_sec ? (unsigned int)ts.tv_sec : 1;
+	order = get_count_order(secs);
+	secs_shrink_size = (order <= EVICTION_SECS_POS_SHIFT)
+			? 0 : (order - EVICTION_SECS_POS_SHIFT);
+
+	eviction = *peviction;
+	eviction = (eviction << EVICTION_SECS_POS_SHIFT)
+			| ((secs >> secs_shrink_size) & EVICTION_SECS_POS_MASK);
+	eviction = (eviction << EVICTION_SECS_SHRINK_SHIFT) | (secs_shrink_size & 0xf);
+	*peviction = eviction;
+}
+/*
+ * Unpack the second from the entry and restore the value according to the
+ * shrink size.
+ */
+static unsigned int unpack_secs(unsigned long entry)
+{
+	unsigned int secs;
+	int secs_shrink_size;
+
+	secs_shrink_size = entry & ((1 << EVICTION_SECS_SHRINK_SHIFT) - 1);
+	entry >>= EVICTION_SECS_SHRINK_SHIFT;
+	secs = entry & EVICTION_SECS_POS_MASK;
+	secs = secs << secs_shrink_size;
+	return secs;
+}
+#endif
 static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction)
 {
 	eviction >>= bucket_order;
+	pack_secs(&eviction);
 	eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
 	eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
 	eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
@@ -181,20 +240,24 @@ static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction)
 }
 
 static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
-			  unsigned long *evictionp)
+			  unsigned long *evictionp, unsigned int *prev_secs)
 {
 	unsigned long entry = (unsigned long)shadow;
 	int memcgid, nid;
+	unsigned int secs;
 
 	entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
 	nid = entry & ((1UL << NODES_SHIFT) - 1);
 	entry >>= NODES_SHIFT;
 	memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
 	entry >>= MEM_CGROUP_ID_SHIFT;
+	secs = unpack_secs(entry);
+	entry >>= (EVICTION_SECS_POS_SHIFT + EVICTION_SECS_SHRINK_SHIFT);
 
 	*memcgidp = memcgid;
 	*pgdat = NODE_DATA(nid);
 	*evictionp = entry << bucket_order;
+	*prev_secs = secs;
 }
 
 /**
@@ -242,9 +305,22 @@ bool workingset_refault(void *shadow)
 	unsigned long refault;
 	struct pglist_data *pgdat;
 	int memcgid;
+#ifndef NO_SECS_IN_WORKINGSET
+	unsigned long avg_refault_time;
+	unsigned long refault_time;
+	int tradition;
+	unsigned int prev_secs;
+	unsigned int secs;
+	unsigned long refaults_ratio;
+#endif
+	struct timespec ts;
+	/*
+	convert jiffies to second
+	*/
+	get_monotonic_boottime(&ts);
+	secs = (unsigned int)ts.tv_sec ? (unsigned int)ts.tv_sec : 1;
 
-	unpack_shadow(shadow, &memcgid, &pgdat, &eviction);
-
+	unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &prev_secs);
 	rcu_read_lock();
 	/*
 	 * Look up the memcg associated with the stored ID. It might
@@ -288,14 +364,37 @@ bool workingset_refault(void *shadow)
 	 * list is not a problem.
 	 */
 	refault_distance = (refault - eviction) & EVICTION_MASK;
-
 	inc_lruvec_state(lruvec, WORKINGSET_REFAULT);
-
-	if (refault_distance <= active_file) {
+#ifndef NO_SECS_IN_WORKINGSET
+	refaults_ratio = (atomic_long_read(&lruvec->inactive_age) + 1) / secs;
+	atomic_long_set(&lruvec->refaults_ratio, refaults_ratio);
+	refault_time = secs - prev_secs;
+	avg_refault_time = active_file / refaults_ratio;
+	tradition = !!(refault_distance < active_file);
+	if (refault_time <= avg_refault_time) {
+#else
+	if (refault_distance < active_file) {
+#endif
 		inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE);
+#ifndef NO_SECS_IN_WORKINGSET
+		trace_printk("WKST_ACT[%d]:rft_dis %ld, act_file %ld \
+				rft_ratio %ld rft_time %ld avg_rft_time %ld \
+				refault %ld eviction %ld secs %d pre_secs %d\n",
+				tradition, refault_distance, active_file,
+				refaults_ratio, refault_time, avg_refault_time,
+				refault, eviction, secs, prev_secs);
+#endif
 		rcu_read_unlock();
 		return true;
 	}
+#ifndef NO_SECS_IN_WORKINGSET
+	trace_printk("WKST_INACT[%d]:rft_dis %ld, act_file %ld \
+			rft_ratio %ld rft_time %ld avg_rft_time %ld \
+			refault %ld eviction %ld secs %d pre_secs %d\n",
+			tradition, refault_distance, active_file,
+			refaults_ratio, refault_time, avg_refault_time,
+			refault, eviction, secs, prev_secs);
+#endif
 	rcu_read_unlock();
 	return false;
 }
@@ -513,7 +612,9 @@ static int __init workingset_init(void)
 	unsigned int max_order;
 	int ret;
 
-	BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
+	BUILD_BUG_ON(BITS_PER_LONG < (EVICTION_SHIFT
+				+ EVICTION_SECS_POS_SHIFT
+				+ EVICTION_SECS_SHRINK_SHIFT));
 	/*
 	 * Calculate the eviction bucket size to cover the longest
 	 * actionable refault distance, which is currently half of
@@ -521,7 +622,8 @@ static int __init workingset_init(void)
 	 * some more pages at runtime, so keep working with up to
 	 * double the initial memory by using totalram_pages as-is.
 	 */
-	timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
+	timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT
+			- EVICTION_SECS_POS_SHIFT - EVICTION_SECS_SHRINK_SHIFT;
 	max_order = fls_long(totalram_pages - 1);
 	if (max_order > timestamp_bits)
 		bucket_order = max_order - timestamp_bits;
-- 
1.9.1