mm, thp: khugepaged should scan when sleep value is written
authorDavid Rientjes <rientjes@google.com>
Fri, 20 May 2016 23:58:56 +0000 (16:58 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 21 May 2016 00:58:30 +0000 (17:58 -0700)
If a large value is written to scan_sleep_millisecs, for example, that
period must lapse before khugepaged will wake up for periodic
collapsing.

If this value is tuned to 1 day, for example, and then re-tuned to its
default 10s, khugepaged will still wait for a day before scanning again.

This patch causes khugepaged to wakeup immediately when the value is
changed and then sleep until that value is rewritten or the new value
lapses.

Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1605181453200.4786@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/huge_memory.c

index 1764184c4774037dba0dcaf3e92050b20942305a..41ef7547e822bbcbf616ee929cf945a214c77444 100644 (file)
@@ -89,6 +89,7 @@ static unsigned int khugepaged_full_scans;
 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
 /* during fragmentation poll the hugepage allocator once every minute */
 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
+static unsigned long khugepaged_sleep_expire;
 static struct task_struct *khugepaged_thread __read_mostly;
 static DEFINE_MUTEX(khugepaged_mutex);
 static DEFINE_SPINLOCK(khugepaged_mm_lock);
@@ -467,6 +468,7 @@ static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
                return -EINVAL;
 
        khugepaged_scan_sleep_millisecs = msecs;
+       khugepaged_sleep_expire = 0;
        wake_up_interruptible(&khugepaged_wait);
 
        return count;
@@ -494,6 +496,7 @@ static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
                return -EINVAL;
 
        khugepaged_alloc_sleep_millisecs = msecs;
+       khugepaged_sleep_expire = 0;
        wake_up_interruptible(&khugepaged_wait);
 
        return count;
@@ -2791,15 +2794,25 @@ static void khugepaged_do_scan(void)
                put_page(hpage);
 }
 
+static bool khugepaged_should_wakeup(void)
+{
+       return kthread_should_stop() ||
+              time_after_eq(jiffies, khugepaged_sleep_expire);
+}
+
 static void khugepaged_wait_work(void)
 {
        if (khugepaged_has_work()) {
-               if (!khugepaged_scan_sleep_millisecs)
+               const unsigned long scan_sleep_jiffies =
+                       msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
+
+               if (!scan_sleep_jiffies)
                        return;
 
+               khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
                wait_event_freezable_timeout(khugepaged_wait,
-                                            kthread_should_stop(),
-                       msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
+                                            khugepaged_should_wakeup(),
+                                            scan_sleep_jiffies);
                return;
        }