writeback: switch to per-bdi threads for flushing data

This gets rid of pdflush for bdi writeout and kupdated style cleaning.
This is an experiment to see if we get better writeout behaviour with
per-bdi flushing. Some initial tests look pretty encouraging. A sample
ffsb workload that does random writes to files is about 8% faster here
on a simple SATA drive during the benchmark phase. File layout also seems
a LOT more smooth in vmstat:

 r  b   swpd   free   buff  cache   si   so    bi    bo   in    cs us sy id wa
 0  1      0 608848   2652 375372    0    0     0 71024  604    24  1 10 48 42
 0  1      0 549644   2712 433736    0    0     0 60692  505    27  1  8 48 44
 1  0      0 476928   2784 505192    0    0     4 29540  553    24  0  9 53 37
 0  1      0 457972   2808 524008    0    0     0 54876  331    16  0  4 38 58
 0  1      0 366128   2928 614284    0    0     4 92168  710    58  0 13 53 34
 0  1      0 295092   3000 684140    0    0     0 62924  572    23  0  9 53 37
 0  1      0 236592   3064 741704    0    0     4 58256  523    17  0  8 48 44
 0  1      0 165608   3132 811464    0    0     0 57460  560    21  0  8 54 38
 0  1      0 102952   3200 873164    0    0     4 74748  540    29  1 10 48 41
 0  1      0  48604   3252 926472    0    0     0 53248  469    29  0  7 47 45

where vanilla tends to fluctuate a lot in the creation phase:

 r  b   swpd   free   buff  cache   si   so    bi    bo   in    cs us sy id wa
 1  1      0 678716   5792 303380    0    0     0 74064  565    50  1 11 52 36
 1  0      0 662488   5864 319396    0    0     4   352  302   329  0  2 47 51
 0  1      0 599312   5924 381468    0    0     0 78164  516    55  0  9 51 40
 0  1      0 519952   6008 459516    0    0     4 78156  622    56  1 11 52 37
 1  1      0 436640   6092 541632    0    0     0 82244  622    54  0 11 48 41
 0  1      0 436640   6092 541660    0    0     0     8  152    39  0  0 51 49
 0  1      0 332224   6200 644252    0    0     4 102800  728    46  1 13 49 36
 1  0      0 274492   6260 701056    0    0     4 12328  459    49  0  7 50 43
 0  1      0 211220   6324 763356    0    0     0 106940  515    37  1 10 51 39
 1  0      0 160412   6376 813468    0    0     0  8224  415    43  0  6 49 45
 1  1      0  85980   6452 886556    0    0     4 113516  575    39  1 11 54 34
 0  2      0  85968   6452 886620    0    0     0  1640  158   211  0  0 46 54

So apart from seemingly behaving better for buffered writeout, this also
allows us to potentially have more than one bdi thread flushing out data.
This may be useful for NUMA type setups.

A 10 disk test with btrfs performs 26% faster with per-bdi flushing. Other
tests pending. mmap heavy writing also improves considerably.

A separate thread is added to sync the super blocks. In the long term,
adding sync_supers_bdi() functionality could get rid of this thread again.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

3 files changed, 270 insertions, 146 deletions

diff --git a/mm/backing-dev.c b/mm/backing-dev.c
index de0bbfe08371..9e17baea0ca9 100644
--- a/mm/backing-dev.c
+++ b/mm/backing-dev.c
@@ -1,8 +1,11 @@
 
 #include <linux/wait.h>
 #include <linux/backing-dev.h>
+#include <linux/kthread.h>
+#include <linux/freezer.h>
 #include <linux/fs.h>
 #include <linux/pagemap.h>
+#include <linux/mm.h>
 #include <linux/sched.h>
 #include <linux/module.h>
 #include <linux/writeback.h>
@@ -22,8 +25,18 @@ struct backing_dev_info default_backing_dev_info = {
 EXPORT_SYMBOL_GPL(default_backing_dev_info);
 
 static struct class *bdi_class;
-DEFINE_MUTEX(bdi_lock);
+DEFINE_SPINLOCK(bdi_lock);
 LIST_HEAD(bdi_list);
+LIST_HEAD(bdi_pending_list);
+
+static struct task_struct *sync_supers_tsk;
+static struct timer_list sync_supers_timer;
+
+static int bdi_sync_supers(void *);
+static void sync_supers_timer_fn(unsigned long);
+static void arm_supers_timer(void);
+
+static void bdi_add_default_flusher_task(struct backing_dev_info *bdi);
 
 #ifdef CONFIG_DEBUG_FS
 #include <linux/debugfs.h>
@@ -187,6 +200,13 @@ static int __init default_bdi_init(void)
 {
 	int err;
 
+	sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
+	BUG_ON(IS_ERR(sync_supers_tsk));
+
+	init_timer(&sync_supers_timer);
+	setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
+	arm_supers_timer();
+
 	err = bdi_init(&default_backing_dev_info);
 	if (!err)
 		bdi_register(&default_backing_dev_info, NULL, "default");
@@ -195,6 +215,184 @@ static int __init default_bdi_init(void)
 }
 subsys_initcall(default_bdi_init);
 
+static int bdi_start_fn(void *ptr)
+{
+	struct backing_dev_info *bdi = ptr;
+	struct task_struct *tsk = current;
+
+	/*
+	 * Add us to the active bdi_list
+	 */
+	spin_lock(&bdi_lock);
+	list_add(&bdi->bdi_list, &bdi_list);
+	spin_unlock(&bdi_lock);
+
+	tsk->flags |= PF_FLUSHER | PF_SWAPWRITE;
+	set_freezable();
+
+	/*
+	 * Our parent may run at a different priority, just set us to normal
+	 */
+	set_user_nice(tsk, 0);
+
+	/*
+	 * Clear pending bit and wakeup anybody waiting to tear us down
+	 */
+	clear_bit(BDI_pending, &bdi->state);
+	smp_mb__after_clear_bit();
+	wake_up_bit(&bdi->state, BDI_pending);
+
+	return bdi_writeback_task(bdi);
+}
+
+static void bdi_flush_io(struct backing_dev_info *bdi)
+{
+	struct writeback_control wbc = {
+		.bdi			= bdi,
+		.sync_mode		= WB_SYNC_NONE,
+		.older_than_this	= NULL,
+		.range_cyclic		= 1,
+		.nr_to_write		= 1024,
+	};
+
+	generic_sync_bdi_inodes(NULL, &wbc);
+}
+
+/*
+ * kupdated() used to do this. We cannot do it from the bdi_forker_task()
+ * or we risk deadlocking on ->s_umount. The longer term solution would be
+ * to implement sync_supers_bdi() or similar and simply do it from the
+ * bdi writeback tasks individually.
+ */
+static int bdi_sync_supers(void *unused)
+{
+	set_user_nice(current, 0);
+
+	while (!kthread_should_stop()) {
+		set_current_state(TASK_INTERRUPTIBLE);
+		schedule();
+
+		/*
+		 * Do this periodically, like kupdated() did before.
+		 */
+		sync_supers();
+	}
+
+	return 0;
+}
+
+static void arm_supers_timer(void)
+{
+	unsigned long next;
+
+	next = msecs_to_jiffies(dirty_writeback_interval * 10) + jiffies;
+	mod_timer(&sync_supers_timer, round_jiffies_up(next));
+}
+
+static void sync_supers_timer_fn(unsigned long unused)
+{
+	wake_up_process(sync_supers_tsk);
+	arm_supers_timer();
+}
+
+static int bdi_forker_task(void *ptr)
+{
+	struct backing_dev_info *me = ptr;
+
+	for (;;) {
+		struct backing_dev_info *bdi, *tmp;
+
+		/*
+		 * Temporary measure, we want to make sure we don't see
+		 * dirty data on the default backing_dev_info
+		 */
+		if (bdi_has_dirty_io(me))
+			bdi_flush_io(me);
+
+		spin_lock(&bdi_lock);
+
+		/*
+		 * Check if any existing bdi's have dirty data without
+		 * a thread registered. If so, set that up.
+		 */
+		list_for_each_entry_safe(bdi, tmp, &bdi_list, bdi_list) {
+			if (bdi->task || !bdi_has_dirty_io(bdi))
+				continue;
+
+			bdi_add_default_flusher_task(bdi);
+		}
+
+		set_current_state(TASK_INTERRUPTIBLE);
+
+		if (list_empty(&bdi_pending_list)) {
+			unsigned long wait;
+
+			spin_unlock(&bdi_lock);
+			wait = msecs_to_jiffies(dirty_writeback_interval * 10);
+			schedule_timeout(wait);
+			try_to_freeze();
+			continue;
+		}
+
+		__set_current_state(TASK_RUNNING);
+
+		/*
+		 * This is our real job - check for pending entries in
+		 * bdi_pending_list, and create the tasks that got added
+		 */
+		bdi = list_entry(bdi_pending_list.next, struct backing_dev_info,
+				 bdi_list);
+		list_del_init(&bdi->bdi_list);
+		spin_unlock(&bdi_lock);
+
+		BUG_ON(bdi->task);
+
+		bdi->task = kthread_run(bdi_start_fn, bdi, "flush-%s",
+					dev_name(bdi->dev));
+		/*
+		 * If task creation fails, then readd the bdi to
+		 * the pending list and force writeout of the bdi
+		 * from this forker thread. That will free some memory
+		 * and we can try again.
+		 */
+		if (IS_ERR(bdi->task)) {
+			bdi->task = NULL;
+
+			/*
+			 * Add this 'bdi' to the back, so we get
+			 * a chance to flush other bdi's to free
+			 * memory.
+			 */
+			spin_lock(&bdi_lock);
+			list_add_tail(&bdi->bdi_list, &bdi_pending_list);
+			spin_unlock(&bdi_lock);
+
+			bdi_flush_io(bdi);
+		}
+	}
+
+	return 0;
+}
+
+/*
+ * Add a new flusher task that gets created for any bdi
+ * that has dirty data pending writeout
+ */
+static void bdi_add_default_flusher_task(struct backing_dev_info *bdi)
+{
+	/*
+	 * Someone already marked this pending for task creation
+	 */
+	if (test_and_set_bit(BDI_pending, &bdi->state))
+		return;
+
+	spin_lock(&bdi_lock);
+	list_move_tail(&bdi->bdi_list, &bdi_pending_list);
+	spin_unlock(&bdi_lock);
+
+	wake_up_process(default_backing_dev_info.task);
+}
+
 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 		const char *fmt, ...)
 {
@@ -213,13 +411,31 @@ int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 		goto exit;
 	}
 
-	mutex_lock(&bdi_lock);
+	spin_lock(&bdi_lock);
 	list_add_tail(&bdi->bdi_list, &bdi_list);
-	mutex_unlock(&bdi_lock);
+	spin_unlock(&bdi_lock);
 
 	bdi->dev = dev;
-	bdi_debug_register(bdi, dev_name(dev));
 
+	/*
+	 * Just start the forker thread for our default backing_dev_info,
+	 * and add other bdi's to the list. They will get a thread created
+	 * on-demand when they need it.
+	 */
+	if (bdi_cap_flush_forker(bdi)) {
+		bdi->task = kthread_run(bdi_forker_task, bdi, "bdi-%s",
+						dev_name(dev));
+		if (IS_ERR(bdi->task)) {
+			bdi->task = NULL;
+			spin_lock(&bdi_lock);
+			list_del(&bdi->bdi_list);
+			spin_unlock(&bdi_lock);
+			ret = -ENOMEM;
+			goto exit;
+		}
+	}
+
+	bdi_debug_register(bdi, dev_name(dev));
 exit:
 	return ret;
 }
@@ -231,17 +447,40 @@ int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 }
 EXPORT_SYMBOL(bdi_register_dev);
 
-static void bdi_remove_from_list(struct backing_dev_info *bdi)
+static int sched_wait(void *word)
 {
-	mutex_lock(&bdi_lock);
+	schedule();
+	return 0;
+}
+
+/*
+ * Remove bdi from the global list
+ */
+static void bdi_wb_shutdown(struct backing_dev_info *bdi)
+{
+	/*
+	 * If setup is pending, wait for that to complete first
+	 */
+	wait_on_bit(&bdi->state, BDI_pending, sched_wait, TASK_UNINTERRUPTIBLE);
+
+	/*
+	 * Make sure nobody finds us on the bdi_list anymore
+	 */
+	spin_lock(&bdi_lock);
 	list_del(&bdi->bdi_list);
-	mutex_unlock(&bdi_lock);
+	spin_unlock(&bdi_lock);
 }
 
 void bdi_unregister(struct backing_dev_info *bdi)
 {
 	if (bdi->dev) {
-		bdi_remove_from_list(bdi);
+		if (!bdi_cap_flush_forker(bdi)) {
+			bdi_wb_shutdown(bdi);
+			if (bdi->task) {
+				kthread_stop(bdi->task);
+				bdi->task = NULL;
+			}
+		}
 		bdi_debug_unregister(bdi);
 		device_unregister(bdi->dev);
 		bdi->dev = NULL;
@@ -251,8 +490,7 @@ EXPORT_SYMBOL(bdi_unregister);
 
 int bdi_init(struct backing_dev_info *bdi)
 {
-	int i;
-	int err;
+	int i, err;
 
 	bdi->dev = NULL;
 
@@ -277,8 +515,6 @@ int bdi_init(struct backing_dev_info *bdi)
 err:
 		while (i--)
 			percpu_counter_destroy(&bdi->bdi_stat[i]);
-
-		bdi_remove_from_list(bdi);
 	}
 
 	return err;
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index d6a1b33d081c..7b87d108602f 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -36,15 +36,6 @@
 #include <linux/pagevec.h>
 
 /*
- * The maximum number of pages to writeout in a single bdflush/kupdate
- * operation.  We do this so we don't hold I_SYNC against an inode for
- * enormous amounts of time, which would block a userspace task which has
- * been forced to throttle against that inode.  Also, the code reevaluates
- * the dirty each time it has written this many pages.
- */
-#define MAX_WRITEBACK_PAGES	1024
-
-/*
  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  * will look to see if it needs to force writeback or throttling.
  */
@@ -117,8 +108,6 @@ EXPORT_SYMBOL(laptop_mode);
 /* End of sysctl-exported parameters */
 
 
-static void background_writeout(unsigned long _min_pages);
-
 /*
  * Scale the writeback cache size proportional to the relative writeout speeds.
  *
@@ -326,7 +315,7 @@ int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 {
 	int ret = 0;
 
-	mutex_lock(&bdi_lock);
+	spin_lock(&bdi_lock);
 	if (min_ratio > bdi->max_ratio) {
 		ret = -EINVAL;
 	} else {
@@ -338,7 +327,7 @@ int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 			ret = -EINVAL;
 		}
 	}
-	mutex_unlock(&bdi_lock);
+	spin_unlock(&bdi_lock);
 
 	return ret;
 }
@@ -350,14 +339,14 @@ int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
 	if (max_ratio > 100)
 		return -EINVAL;
 
-	mutex_lock(&bdi_lock);
+	spin_lock(&bdi_lock);
 	if (bdi->min_ratio > max_ratio) {
 		ret = -EINVAL;
 	} else {
 		bdi->max_ratio = max_ratio;
 		bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
 	}
-	mutex_unlock(&bdi_lock);
+	spin_unlock(&bdi_lock);
 
 	return ret;
 }
@@ -540,7 +529,7 @@ static void balance_dirty_pages(struct address_space *mapping)
 		 * been flushed to permanent storage.
 		 */
 		if (bdi_nr_reclaimable) {
-			writeback_inodes(&wbc);
+			generic_sync_bdi_inodes(NULL, &wbc);
 			pages_written += write_chunk - wbc.nr_to_write;
 			get_dirty_limits(&background_thresh, &dirty_thresh,
 				       &bdi_thresh, bdi);
@@ -591,7 +580,7 @@ static void balance_dirty_pages(struct address_space *mapping)
 			(!laptop_mode && (global_page_state(NR_FILE_DIRTY)
 					  + global_page_state(NR_UNSTABLE_NFS)
 					  > background_thresh)))
-		pdflush_operation(background_writeout, 0);
+		bdi_start_writeback(bdi, NULL, 0, WB_SYNC_NONE);
 }
 
 void set_page_dirty_balance(struct page *page, int page_mkwrite)
@@ -676,152 +665,53 @@ void throttle_vm_writeout(gfp_t gfp_mask)
 }
 
 /*
- * writeback at least _min_pages, and keep writing until the amount of dirty
- * memory is less than the background threshold, or until we're all clean.
+ * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
+ * the whole world.
  */
-static void background_writeout(unsigned long _min_pages)
+void wakeup_flusher_threads(long nr_pages)
 {
-	long min_pages = _min_pages;
 	struct writeback_control wbc = {
-		.bdi		= NULL,
 		.sync_mode	= WB_SYNC_NONE,
 		.older_than_this = NULL,
-		.nr_to_write	= 0,
-		.nonblocking	= 1,
 		.range_cyclic	= 1,
 	};
 
-	for ( ; ; ) {
-		unsigned long background_thresh;
-		unsigned long dirty_thresh;
-
-		get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
-		if (global_page_state(NR_FILE_DIRTY) +
-			global_page_state(NR_UNSTABLE_NFS) < background_thresh
-				&& min_pages <= 0)
-			break;
-		wbc.more_io = 0;
-		wbc.encountered_congestion = 0;
-		wbc.nr_to_write = MAX_WRITEBACK_PAGES;
-		wbc.pages_skipped = 0;
-		writeback_inodes(&wbc);
-		min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
-		if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
-			/* Wrote less than expected */
-			if (wbc.encountered_congestion || wbc.more_io)
-				congestion_wait(WRITE, HZ/10);
-			else
-				break;
-		}
-	}
-}
-
-/*
- * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
- * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
- * -1 if all pdflush threads were busy.
- */
-int wakeup_pdflush(long nr_pages)
-{
 	if (nr_pages == 0)
 		nr_pages = global_page_state(NR_FILE_DIRTY) +
 				global_page_state(NR_UNSTABLE_NFS);
-	return pdflush_operation(background_writeout, nr_pages);
+	wbc.nr_to_write = nr_pages;
+	bdi_writeback_all(NULL, &wbc);
 }
 
-static void wb_timer_fn(unsigned long unused);
 static void laptop_timer_fn(unsigned long unused);
 
-static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
 
 /*
- * Periodic writeback of "old" data.
- *
- * Define "old": the first time one of an inode's pages is dirtied, we mark the
- * dirtying-time in the inode's address_space.  So this periodic writeback code
- * just walks the superblock inode list, writing back any inodes which are
- * older than a specific point in time.
- *
- * Try to run once per dirty_writeback_interval.  But if a writeback event
- * takes longer than a dirty_writeback_interval interval, then leave a
- * one-second gap.
- *
- * older_than_this takes precedence over nr_to_write.  So we'll only write back
- * all dirty pages if they are all attached to "old" mappings.
- */
-static void wb_kupdate(unsigned long arg)
-{
-	unsigned long oldest_jif;
-	unsigned long start_jif;
-	unsigned long next_jif;
-	long nr_to_write;
-	struct writeback_control wbc = {
-		.bdi		= NULL,
-		.sync_mode	= WB_SYNC_NONE,
-		.older_than_this = &oldest_jif,
-		.nr_to_write	= 0,
-		.nonblocking	= 1,
-		.for_kupdate	= 1,
-		.range_cyclic	= 1,
-	};
-
-	sync_supers();
-
-	oldest_jif = jiffies - msecs_to_jiffies(dirty_expire_interval * 10);
-	start_jif = jiffies;
-	next_jif = start_jif + msecs_to_jiffies(dirty_writeback_interval * 10);
-	nr_to_write = global_page_state(NR_FILE_DIRTY) +
-			global_page_state(NR_UNSTABLE_NFS) +
-			(inodes_stat.nr_inodes - inodes_stat.nr_unused);
-	while (nr_to_write > 0) {
-		wbc.more_io = 0;
-		wbc.encountered_congestion = 0;
-		wbc.nr_to_write = MAX_WRITEBACK_PAGES;
-		writeback_inodes(&wbc);
-		if (wbc.nr_to_write > 0) {
-			if (wbc.encountered_congestion || wbc.more_io)
-				congestion_wait(WRITE, HZ/10);
-			else
-				break;	/* All the old data is written */
-		}
-		nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
-	}
-	if (time_before(next_jif, jiffies + HZ))
-		next_jif = jiffies + HZ;
-	if (dirty_writeback_interval)
-		mod_timer(&wb_timer, next_jif);
-}
-
-/*
  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
  */
 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
 	struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
 {
 	proc_dointvec(table, write, file, buffer, length, ppos);
-	if (dirty_writeback_interval)
-		mod_timer(&wb_timer, jiffies +
-			msecs_to_jiffies(dirty_writeback_interval * 10));
-	else
-		del_timer(&wb_timer);
 	return 0;
 }
 
-static void wb_timer_fn(unsigned long unused)
+static void do_laptop_sync(struct work_struct *work)
 {
-	if (pdflush_operation(wb_kupdate, 0) < 0)
-		mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
-}
-
-static void laptop_flush(unsigned long unused)
-{
-	sys_sync();
+	wakeup_flusher_threads(0);
+	kfree(work);
 }
 
 static void laptop_timer_fn(unsigned long unused)
 {
-	pdflush_operation(laptop_flush, 0);
+	struct work_struct *work;
+
+	work = kmalloc(sizeof(*work), GFP_ATOMIC);
+	if (work) {
+		INIT_WORK(work, do_laptop_sync);
+		schedule_work(work);
+	}
 }
 
 /*
@@ -904,8 +794,6 @@ void __init page_writeback_init(void)
 {
 	int shift;
 
-	mod_timer(&wb_timer,
-		  jiffies + msecs_to_jiffies(dirty_writeback_interval * 10));
 	writeback_set_ratelimit();
 	register_cpu_notifier(&ratelimit_nb);
 
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 54155268dfca..27558aa0b10a 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1715,7 +1715,7 @@ static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
 		 */
 		if (total_scanned > sc->swap_cluster_max +
 					sc->swap_cluster_max / 2) {
-			wakeup_pdflush(laptop_mode ? 0 : total_scanned);
+			wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
 			sc->may_writepage = 1;
 		}