share_pool: Optimize compact procedure

ascend inclusion
category: perf
bugzilla: https://gitee.com/openeuler/kernel/issues/I4EUVI
CVE: NA

-------------------------------------------------

We did an compaction experiment on Hi1951. First creating external
fragmentation, then echo 1 > /proc/sys/vm/compact_memory.

Fragmentation size | 0G   | 2G   | 4G  | 8G
Compaction time(s) | 0.07 | 0.85 | 1.7 | 3.4

Obviously, sysctl_compaction_handler has a big performance impact.

We optimize share pool compact procedure as follows:
1. At most one compact daemon is allowed.
2. Creating a compact daemon is only allowed when the last one is
finished at least *sysctl_sp_compact_interval* seconds ago.
Signed-off-by: NTang Yizhou <tangyizhou@huawei.com>
Reviewed-by: NDing Tianhong <dingtianhong@huawei.com>
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>
Reviewed-by: NWeilong Chen <chenweilong@huawei.com>
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>

include/linux/share_pool.h

@@ -40,6 +40,10 @@ extern int enable_ascend_share_pool;
 
 extern int sysctl_share_pool_map_lock_enable;
 
+extern int sysctl_sp_compact_enable;
+extern unsigned long sysctl_sp_compact_interval;
+extern unsigned long sysctl_sp_compact_interval_max;
+
 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
 extern bool vmap_allow_huge;
 #endif

kernel/sysctl.c

@@ -1789,6 +1789,24 @@ static struct ctl_table vm_table[] = {
 		.extra1		= &zero,
 		.extra2		= &one,
 	},
+	{
+		.procname	= "sharepool_compact_enable",
+		.data		= &sysctl_sp_compact_enable,
+		.maxlen		= sizeof(sysctl_sp_compact_enable),
+		.mode		= 0600,
+		.proc_handler	= proc_dointvec_minmax,
+		.extra1		= &zero,
+		.extra2		= &one,
+	},
+	{
+		.procname	= "sharepool_compact_interval",
+		.data		= &sysctl_sp_compact_interval,
+		.maxlen		= sizeof(sysctl_sp_compact_interval),
+		.mode		= 0600,
+		.proc_handler	= proc_doulongvec_minmax,
+		.extra1		= &zero_ul,
+		.extra2		= &sysctl_sp_compact_interval_max,
+	},
 #endif
 	{ }
 };

mm/share_pool.c

@@ -1127,26 +1127,44 @@ void sp_area_drop(struct vm_area_struct *vma)
 	spin_unlock(&sp_area_lock);
 }
 
-static unsigned long last_jiffies;
+int sysctl_sp_compact_enable;
+unsigned long sysctl_sp_compact_interval = 30UL;
+unsigned long sysctl_sp_compact_interval_max = 1000UL;
+static unsigned long compact_last_jiffies;
+static unsigned long compact_daemon_status;
+#define COMPACT_START	1
+#define COMPACT_STOP	0
+
 static void sp_compact_nodes(struct work_struct *work)
 {
 	sysctl_compaction_handler(NULL, 1, NULL, NULL, NULL);
 
 	kfree(work);
+
+	compact_last_jiffies = jiffies;
+	cmpxchg(&compact_daemon_status, COMPACT_START, COMPACT_STOP);
 }
 
 static void sp_add_work_compact(void)
 {
 	struct work_struct *compact_work;
 
-	if (!time_after(jiffies, last_jiffies + 10 * HZ))
+	if (!sysctl_sp_compact_enable)
+		return;
+
+	/* experimental compaction time: 4GB->1.7s, 8GB->3.4s */
+	if (!time_after(jiffies,
+		compact_last_jiffies + sysctl_sp_compact_interval * HZ))
+		return;
+
+	if (cmpxchg(&compact_daemon_status, COMPACT_STOP, COMPACT_START) ==
+		    COMPACT_START)
 		return;
 
 	compact_work = kzalloc(sizeof(*compact_work), GFP_KERNEL);
 	if (!compact_work)
 		return;
 
-	last_jiffies = jiffies;
 	INIT_WORK(compact_work, sp_compact_nodes);
 	schedule_work(compact_work);
 }