writeback: optimize periodic bdi thread wakeups
Artem Bityutskiy [Sun, 25 Jul 2010 11:29:22 +0000 (14:29 +0300)]
Whe the first inode for a bdi is marked dirty, we wake up the bdi thread which
should take care of the periodic background write-out. However, the write-out
will actually start only 'dirty_writeback_interval' centisecs later, so we can
delay the wake-up.

This change was requested by Nick Piggin who pointed out that if we delay the
wake-up, we weed out 2 unnecessary contex switches, which matters because
'__mark_inode_dirty()' is a hot-path function.

This patch introduces a new function - 'bdi_wakeup_thread_delayed()', which
sets up a timer to wake-up the bdi thread and returns. So the wake-up is
delayed.

We also delete the timer in bdi threads just before writing-back. And
synchronously delete it when unregistering bdi. At the unregister point the bdi
does not have any users, so no one can arm it again.

Since now we take 'bdi->wb_lock' in the timer, which can execute in softirq
context, we have to use 'spin_lock_bh()' for 'bdi->wb_lock'. This patch makes
this change as well.

This patch also moves the 'bdi_wb_init()' function down in the file to avoid
forward-declaration of 'bdi_wakeup_thread_delayed()'.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

fs/fs-writeback.c
include/linux/backing-dev.h
mm/backing-dev.c

index 55f6e46..bfa2df2 100644 (file)
@@ -76,7 +76,7 @@ static void bdi_queue_work(struct backing_dev_info *bdi,
 {
        trace_writeback_queue(bdi, work);
 
-       spin_lock(&bdi->wb_lock);
+       spin_lock_bh(&bdi->wb_lock);
        list_add_tail(&work->list, &bdi->work_list);
        if (bdi->wb.task) {
                wake_up_process(bdi->wb.task);
@@ -88,7 +88,7 @@ static void bdi_queue_work(struct backing_dev_info *bdi,
                trace_writeback_nothread(bdi, work);
                wake_up_process(default_backing_dev_info.wb.task);
        }
-       spin_unlock(&bdi->wb_lock);
+       spin_unlock_bh(&bdi->wb_lock);
 }
 
 static void
@@ -704,13 +704,13 @@ get_next_work_item(struct backing_dev_info *bdi)
 {
        struct wb_writeback_work *work = NULL;
 
-       spin_lock(&bdi->wb_lock);
+       spin_lock_bh(&bdi->wb_lock);
        if (!list_empty(&bdi->work_list)) {
                work = list_entry(bdi->work_list.next,
                                  struct wb_writeback_work, list);
                list_del_init(&work->list);
        }
-       spin_unlock(&bdi->wb_lock);
+       spin_unlock_bh(&bdi->wb_lock);
        return work;
 }
 
@@ -810,6 +810,12 @@ int bdi_writeback_thread(void *data)
        trace_writeback_thread_start(bdi);
 
        while (!kthread_should_stop()) {
+               /*
+                * Remove own delayed wake-up timer, since we are already awake
+                * and we'll take care of the preriodic write-back.
+                */
+               del_timer(&wb->wakeup_timer);
+
                pages_written = wb_do_writeback(wb, 0);
 
                trace_writeback_pages_written(pages_written);
@@ -868,26 +874,6 @@ void wakeup_flusher_threads(long nr_pages)
        rcu_read_unlock();
 }
 
-/*
- * This function is used when the first inode for this bdi is marked dirty. It
- * wakes-up the corresponding bdi thread which should then take care of the
- * periodic background write-out of dirty inodes.
- */
-static void wakeup_bdi_thread(struct backing_dev_info *bdi)
-{
-       spin_lock(&bdi->wb_lock);
-       if (bdi->wb.task)
-               wake_up_process(bdi->wb.task);
-       else
-               /*
-                * When bdi tasks are inactive for long time, they are killed.
-                * In this case we have to wake-up the forker thread which
-                * should create and run the bdi thread.
-                */
-               wake_up_process(default_backing_dev_info.wb.task);
-       spin_unlock(&bdi->wb_lock);
-}
-
 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
 {
        if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
@@ -1019,7 +1005,7 @@ out:
        spin_unlock(&inode_lock);
 
        if (wakeup_bdi)
-               wakeup_bdi_thread(bdi);
+               bdi_wakeup_thread_delayed(bdi);
 }
 EXPORT_SYMBOL(__mark_inode_dirty);
 
index 71b6223..7628219 100644 (file)
@@ -52,6 +52,7 @@ struct bdi_writeback {
        unsigned long last_active;      /* last time bdi thread was active */
 
        struct task_struct *task;       /* writeback thread */
+       struct timer_list wakeup_timer; /* used for delayed bdi thread wakeup */
        struct list_head b_dirty;       /* dirty inodes */
        struct list_head b_io;          /* parked for writeback */
        struct list_head b_more_io;     /* parked for more writeback */
@@ -105,6 +106,7 @@ void bdi_start_background_writeback(struct backing_dev_info *bdi);
 int bdi_writeback_thread(void *data);
 int bdi_has_dirty_io(struct backing_dev_info *bdi);
 void bdi_arm_supers_timer(void);
+void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi);
 
 extern spinlock_t bdi_lock;
 extern struct list_head bdi_list;
index a9a08d8..cfff722 100644 (file)
@@ -248,17 +248,6 @@ static int __init default_bdi_init(void)
 }
 subsys_initcall(default_bdi_init);
 
-static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
-{
-       memset(wb, 0, sizeof(*wb));
-
-       wb->bdi = bdi;
-       wb->last_old_flush = jiffies;
-       INIT_LIST_HEAD(&wb->b_dirty);
-       INIT_LIST_HEAD(&wb->b_io);
-       INIT_LIST_HEAD(&wb->b_more_io);
-}
-
 int bdi_has_dirty_io(struct backing_dev_info *bdi)
 {
        return wb_has_dirty_io(&bdi->wb);
@@ -316,6 +305,43 @@ static void sync_supers_timer_fn(unsigned long unused)
        bdi_arm_supers_timer();
 }
 
+static void wakeup_timer_fn(unsigned long data)
+{
+       struct backing_dev_info *bdi = (struct backing_dev_info *)data;
+
+       spin_lock_bh(&bdi->wb_lock);
+       if (bdi->wb.task) {
+               wake_up_process(bdi->wb.task);
+       } else {
+               /*
+                * When bdi tasks are inactive for long time, they are killed.
+                * In this case we have to wake-up the forker thread which
+                * should create and run the bdi thread.
+                */
+               wake_up_process(default_backing_dev_info.wb.task);
+       }
+       spin_unlock_bh(&bdi->wb_lock);
+}
+
+/*
+ * This function is used when the first inode for this bdi is marked dirty. It
+ * wakes-up the corresponding bdi thread which should then take care of the
+ * periodic background write-out of dirty inodes. Since the write-out would
+ * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
+ * set up a timer which wakes the bdi thread up later.
+ *
+ * Note, we wouldn't bother setting up the timer, but this function is on the
+ * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
+ * by delaying the wake-up.
+ */
+void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
+{
+       unsigned long timeout;
+
+       timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
+       mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
+}
+
 /*
  * Calculate the longest interval (jiffies) bdi threads are allowed to be
  * inactive.
@@ -353,8 +379,10 @@ static int bdi_forker_thread(void *ptr)
                 * Temporary measure, we want to make sure we don't see
                 * dirty data on the default backing_dev_info
                 */
-               if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list))
+               if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
+                       del_timer(&me->wakeup_timer);
                        wb_do_writeback(me, 0);
+               }
 
                spin_lock_bh(&bdi_lock);
                set_current_state(TASK_INTERRUPTIBLE);
@@ -386,7 +414,7 @@ static int bdi_forker_thread(void *ptr)
                                break;
                        }
 
-                       spin_lock(&bdi->wb_lock);
+                       spin_lock_bh(&bdi->wb_lock);
                        /*
                         * If there is no work to do and the bdi thread was
                         * inactive long enough - kill it. The wb_lock is taken
@@ -403,7 +431,7 @@ static int bdi_forker_thread(void *ptr)
                                action = KILL_THREAD;
                                break;
                        }
-                       spin_unlock(&bdi->wb_lock);
+                       spin_unlock_bh(&bdi->wb_lock);
                }
                spin_unlock_bh(&bdi_lock);
 
@@ -427,9 +455,9 @@ static int bdi_forker_thread(void *ptr)
                                 * The spinlock makes sure we do not lose
                                 * wake-ups when racing with 'bdi_queue_work()'.
                                 */
-                               spin_lock(&bdi->wb_lock);
+                               spin_lock_bh(&bdi->wb_lock);
                                bdi->wb.task = task;
-                               spin_unlock(&bdi->wb_lock);
+                               spin_unlock_bh(&bdi->wb_lock);
                        }
                        break;
 
@@ -586,6 +614,7 @@ void bdi_unregister(struct backing_dev_info *bdi)
        if (bdi->dev) {
                trace_writeback_bdi_unregister(bdi);
                bdi_prune_sb(bdi);
+               del_timer_sync(&bdi->wb.wakeup_timer);
 
                if (!bdi_cap_flush_forker(bdi))
                        bdi_wb_shutdown(bdi);
@@ -596,6 +625,18 @@ void bdi_unregister(struct backing_dev_info *bdi)
 }
 EXPORT_SYMBOL(bdi_unregister);
 
+static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
+{
+       memset(wb, 0, sizeof(*wb));
+
+       wb->bdi = bdi;
+       wb->last_old_flush = jiffies;
+       INIT_LIST_HEAD(&wb->b_dirty);
+       INIT_LIST_HEAD(&wb->b_io);
+       INIT_LIST_HEAD(&wb->b_more_io);
+       setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
+}
+
 int bdi_init(struct backing_dev_info *bdi)
 {
        int i, err;