Skip to contentSkip to conversation

Btrfs: improve the delayed inode throttling

PATCH In review Chris Mason 3 participants 10 messages Mar 5, 2013 – Mar 7, 2013 Btrfs file system

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Btrfs: improve the delayed inode throttling

Hi everyone,

While running fs_mark against raid5/6 I noticed the delayed inode insertions were creating a lot of latencies. Reading through things, I think we need to move to a model where we fire off fewer work items and have the ones we do fire last longer.

With this commit our average file creation rates goes from 130K files/sec up to 160K files/sec for my big fs_mark run.

But before I toss it into git, I wanted to run it by everyone and see if there are workloads this doesn't fit well.

--

The delayed inode code batches up changes to the btree in hopes of doing them in bulk. As the changes build up, processes kick off worker threads and wait for them to make progress.

The current code kicks off an async work queue item for each delayed node, which creates a lot of churn. It also uses a fixed 1 HZ waiting period for the throttle, which allows us to build a lot of pending work and can slow down the commit.

This changes us to watch a sequence counter as it is bumped during the operations. We kick off fewer work items and have each work item do more work.

fs/btrfs/delayed-inode.c +97 -50
index 0b278b1..460d1a8 100644
@@ -22,8 +22,8 @@
 #include "disk-io.h"
 #include "transaction.h"
 
-#define BTRFS_DELAYED_WRITEBACK		400
-#define BTRFS_DELAYED_BACKGROUND	100
+#define BTRFS_DELAYED_WRITEBACK		512
+#define BTRFS_DELAYED_BACKGROUND	128
 
 static struct kmem_cache *delayed_node_cache;
 
@@ -494,6 +494,15 @@ static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
 					BTRFS_DELAYED_DELETION_ITEM);
 }
 
+static void finish_one_item(struct btrfs_delayed_root *delayed_root)
+{
+	int seq = atomic_inc_return(&delayed_root->items_seq);
+	if ((atomic_dec_return(&delayed_root->items) <
+	    BTRFS_DELAYED_BACKGROUND || seq % 16 == 0) &&
+	    waitqueue_active(&delayed_root->wait))
+		wake_up(&delayed_root->wait);
+}
+
 static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 {
 	struct rb_root *root;
@@ -512,10 +521,8 @@ static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 
 	rb_erase(&delayed_item->rb_node, root);
 	delayed_item->delayed_node->count--;
-	if (atomic_dec_return(&delayed_root->items) <
-	    BTRFS_DELAYED_BACKGROUND &&
-	    waitqueue_active(&delayed_root->wait))
-		wake_up(&delayed_root->wait);
+
+	finish_one_item(delayed_root);
 }
 
 static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
@@ -1056,10 +1063,7 @@ static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
 		delayed_node->count--;
 
 		delayed_root = delayed_node->root->fs_info->delayed_root;
-		if (atomic_dec_return(&delayed_root->items) <
-		    BTRFS_DELAYED_BACKGROUND &&
-		    waitqueue_active(&delayed_root->wait))
-			wake_up(&delayed_root->wait);
+		finish_one_item(delayed_root);
 	}
 }
 
@@ -1304,35 +1308,55 @@ void btrfs_remove_delayed_node(struct inode *inode)
 	btrfs_release_delayed_node(delayed_node);
 }
 
+#if 0
 struct btrfs_async_delayed_node {
 	struct btrfs_root *root;
 	struct btrfs_delayed_node *delayed_node;
 	struct btrfs_work work;
 };
+#endif
+
+struct btrfs_async_delayed_work {
+	struct btrfs_delayed_root *delayed_root;
+	int nr;
+	struct btrfs_work work;
+};
 
-static void btrfs_async_run_delayed_node_done(struct btrfs_work *work)
+static void btrfs_async_run_delayed_root(struct btrfs_work *work)
 {
-	struct btrfs_async_delayed_node *async_node;
+	struct btrfs_async_delayed_work *async_work;
+	struct btrfs_delayed_root *delayed_root;
 	struct btrfs_trans_handle *trans;
 	struct btrfs_path *path;
 	struct btrfs_delayed_node *delayed_node = NULL;
 	struct btrfs_root *root;
 	struct btrfs_block_rsv *block_rsv;
 	int need_requeue = 0;
+	int total_done = 0;
 
-	async_node = container_of(work, struct btrfs_async_delayed_node, work);
+	async_work = container_of(work, struct btrfs_async_delayed_work, work);
+	delayed_root = async_work->delayed_root;
 
 	path = btrfs_alloc_path();
 	if (!path)
 		goto out;
-	path->leave_spinning = 1;
 
-	delayed_node = async_node->delayed_node;
+again:
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND / 2)
+		goto free_path;
+
+	delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
+	if (!delayed_node)
+		goto free_path;
+
+requeue:
+	path->leave_spinning = 1;
+	need_requeue = 0;
 	root = delayed_node->root;
 
 	trans = btrfs_join_transaction(root);
 	if (IS_ERR(trans))
-		goto free_path;
+		goto release_path;
 
 	block_rsv = trans->block_rsv;
 	trans->block_rsv = &root->fs_info->delayed_block_rsv;
@@ -1373,47 +1397,48 @@ static void btrfs_async_run_delayed_node_done(struct btrfs_work *work)
 	trans->block_rsv = block_rsv;
 	btrfs_end_transaction_dmeta(trans, root);
 	btrfs_btree_balance_dirty_nodelay(root);
+
+release_path:
+	btrfs_release_path(path);
+	total_done++;
+
+	if (need_requeue) {
+		goto requeue;
+	} else {
+		btrfs_release_prepared_delayed_node(delayed_node);
+		if (async_work->nr == 0 || total_done < async_work->nr)
+			goto again;
+	}
+
 free_path:
 	btrfs_free_path(path);
 out:
-	if (need_requeue)
-		btrfs_requeue_work(&async_node->work);
-	else {
-		btrfs_release_prepared_delayed_node(delayed_node);
-		kfree(async_node);
-	}
+	wake_up(&delayed_root->wait);
+	kfree(async_work);
 }
 
+
 static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
 				     struct btrfs_root *root, int all)
 {
-	struct btrfs_async_delayed_node *async_node;
-	struct btrfs_delayed_node *curr;
-	int count = 0;
+	struct btrfs_async_delayed_work *async_work;
 
-again:
-	curr = btrfs_first_prepared_delayed_node(delayed_root);
-	if (!curr)
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return 0;
 
-	async_node = kmalloc(sizeof(*async_node), GFP_NOFS);
-	if (!async_node) {
-		btrfs_release_prepared_delayed_node(curr);
+	async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
+	if (!async_work)
 		return -ENOMEM;
-	}
-
-	async_node->root = root;
-	async_node->delayed_node = curr;
-
-	async_node->work.func = btrfs_async_run_delayed_node_done;
-	async_node->work.flags = 0;
-
-	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_node->work);
-	count++;
 
-	if (all || count < 4)
-		goto again;
+	async_work->delayed_root = delayed_root;
+	async_work->work.func = btrfs_async_run_delayed_root;
+	async_work->work.flags = 0;
+	if (all)
+		async_work->nr = 0;
+	else
+		async_work->nr = 16;
 
+	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_work->work);
 	return 0;
 }
 
@@ -1424,30 +1449,52 @@ void btrfs_assert_delayed_root_empty(struct btrfs_root *root)
 	WARN_ON(btrfs_first_delayed_node(delayed_root));
 }
 
+static int refs_newer(struct btrfs_delayed_root *delayed_root,
+		      int seq, int count)
+{
+	int val = atomic_read(&delayed_root->items_seq);
+
+	if (val < seq || val >= seq + count)
+		return 1;
+	return 0;
+}
+
 void btrfs_balance_delayed_items(struct btrfs_root *root)
 {
 	struct btrfs_delayed_root *delayed_root;
+	int seq;
 
 	delayed_root = btrfs_get_delayed_root(root);
 
 	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return;
 
+	seq = atomic_read(&delayed_root->items_seq);
+
 	if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
 		int ret;
+		DEFINE_WAIT(__wait);
+
 		ret = btrfs_wq_run_delayed_node(delayed_root, root, 1);
 		if (ret)
 			return;
 
-		wait_event_interruptible_timeout(
-				delayed_root->wait,
-				(atomic_read(&delayed_root->items) <
-				 BTRFS_DELAYED_BACKGROUND),
-				HZ);
-		return;
+		while (1) {
+			prepare_to_wait(&delayed_root->wait, &__wait,
+					TASK_INTERRUPTIBLE);
+
+			if (refs_newer(delayed_root, seq, 16) ||
+			    atomic_read(&delayed_root->items) <
+			    BTRFS_DELAYED_BACKGROUND) {
+				break;
+			}
+			if (!signal_pending(current))
+				schedule();
+		}
+		finish_wait(&delayed_root->wait, &__wait);
 	}
 
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, 16);
 }
 
 /* Will return 0 or -ENOMEM */
fs/btrfs/delayed-inode.h +2 -0
index 78b6ad0..1d5c5f7 100644
@@ -43,6 +43,7 @@ struct btrfs_delayed_root {
 	 */
 	struct list_head prepare_list;
 	atomic_t items;		/* for delayed items */
+	atomic_t items_seq;	/* for delayed items */
 	int nodes;		/* for delayed nodes */
 	wait_queue_head_t wait;
 };
@@ -86,6 +87,7 @@ static inline void btrfs_init_delayed_root(
 				struct btrfs_delayed_root *delayed_root)
 {
 	atomic_set(&delayed_root->items, 0);
+	atomic_set(&delayed_root->items_seq, 0);
 	delayed_root->nodes = 0;
 	spin_lock_init(&delayed_root->lock);
 	init_waitqueue_head(&delayed_root->wait);
#1 lore raw
Quoted 35 lines · Chris Mason

Hi everyone,

While running fs_mark against raid5/6 I noticed the delayed inode insertions were creating a lot of latencies. Reading through things, I think we need to move to a model where we fire off fewer work items and have the ones we do fire last longer.

With this commit our average file creation rates goes from 130K files/sec up to 160K files/sec for my big fs_mark run.

But before I toss it into git, I wanted to run it by everyone and see if there are workloads this doesn't fit well.

--

The delayed inode code batches up changes to the btree in hopes of doing them in bulk. As the changes build up, processes kick off worker threads and wait for them to make progress.

The current code kicks off an async work queue item for each delayed node, which creates a lot of churn. It also uses a fixed 1 HZ waiting period for the throttle, which allows us to build a lot of pending work and can slow down the commit.

This changes us to watch a sequence counter as it is bumped during the operations. We kick off fewer work items and have each work item do more work.

fs/btrfs/delayed-inode.c +0 -0
index 0b278b1..460d1a8 100644
@@ -22,8 +22,8 @@

[...]

Quoted 41 lines
fs/btrfs/delayed-inode.c +16 -6
@@ -1424,30 +1449,52 @@ void btrfs_assert_delayed_root_empty(struct btrfs_root *root)
 	WARN_ON(btrfs_first_delayed_node(delayed_root));
 }
 
+static int refs_newer(struct btrfs_delayed_root *delayed_root,
+		      int seq, int count)
+{
+	int val = atomic_read(&delayed_root->items_seq);
+
+	if (val < seq || val >= seq + count)
+		return 1;
+	return 0;
+}
+
 void btrfs_balance_delayed_items(struct btrfs_root *root)
 {
 	struct btrfs_delayed_root *delayed_root;
+	int seq;
 
 	delayed_root = btrfs_get_delayed_root(root);
 
 	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return;
 
+	seq = atomic_read(&delayed_root->items_seq);
+
 	if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
 		int ret;
+		DEFINE_WAIT(__wait);
+
 		ret = btrfs_wq_run_delayed_node(delayed_root, root, 1);
 		if (ret)
 			return;
 
-		wait_event_interruptible_timeout(
-				delayed_root->wait,
-				(atomic_read(&delayed_root->items) <
-				 BTRFS_DELAYED_BACKGROUND),
-				HZ);
-		return;
+		while (1) {

fs/btrfs/delayed-inode.c void btrfs_assert_delayed_root_empty(struct btrfs_root *root) · line 1492

Quoted 10 lines
fs/btrfs/delayed-inode.c +10 -0
+			prepare_to_wait(&delayed_root->wait, &__wait,
+					TASK_INTERRUPTIBLE);
+
+			if (refs_newer(delayed_root, seq, 16) ||
+			    atomic_read(&delayed_root->items) <
+			    BTRFS_DELAYED_BACKGROUND) {
+				break;
+			}
+			if (!signal_pending(current))
+				schedule();

Do we need a 'break' here?

Like

if (!signal_pending(current))

	schedule();
else
	break;

thanks,
liubo

Quoted 33 lines
fs/btrfs/delayed-inode.c +3 -1
+		}
+		finish_wait(&delayed_root->wait, &__wait);
 	}
 
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, 16);
 }
 
 /* Will return 0 or -ENOMEM */
fs/btrfs/delayed-inode.h +2 -1
index 78b6ad0..1d5c5f7 100644
@@ -43,6 +43,7 @@ struct btrfs_delayed_root {
 	 */
 	struct list_head prepare_list;
 	atomic_t items;		/* for delayed items */
+	atomic_t items_seq;	/* for delayed items */
 	int nodes;		/* for delayed nodes */
 	wait_queue_head_t wait;
 };
@@ -86,6 +87,7 @@ static inline void btrfs_init_delayed_root(
 				struct btrfs_delayed_root *delayed_root)
 {
 	atomic_set(&delayed_root->items, 0);
+	atomic_set(&delayed_root->items_seq, 0);
 	delayed_root->nodes = 0;
 	spin_lock_init(&delayed_root->lock);
 	init_waitqueue_head(&delayed_root->wait);
--

To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#2 lore raw
Quoted 10 lines · Liu Bo

fs/btrfs/delayed-inode.c void btrfs_assert_delayed_root_empty(struct btrfs_root *root) · line 1492

Quoted 6 lines · Chris Mason
fs/btrfs/delayed-inode.c +2 -0
index 0b278b1..460d1a8 100644
+			if (!signal_pending(current))
+				schedule();

Do we need a 'break' here?

Quoted 6 lines

Like

if (!signal_pending(current))

	schedule();
else
	break;

Seems like a good idea ;)

Thanks

-chris

#3 lore raw
In reply to Chris Mason
Quoted 119 lines · Chris Mason
fs/btrfs/delayed-inode.c +40 -16
index 0b278b1..460d1a8 100644
@@ -22,8 +22,8 @@
 #include "disk-io.h"
 #include "transaction.h"
 
-#define BTRFS_DELAYED_WRITEBACK		400
-#define BTRFS_DELAYED_BACKGROUND	100
+#define BTRFS_DELAYED_WRITEBACK		512
+#define BTRFS_DELAYED_BACKGROUND	128
 
 static struct kmem_cache *delayed_node_cache;
 
@@ -494,6 +494,15 @@ static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
 					BTRFS_DELAYED_DELETION_ITEM);
 }
 
+static void finish_one_item(struct btrfs_delayed_root *delayed_root)
+{
+	int seq = atomic_inc_return(&delayed_root->items_seq);
+	if ((atomic_dec_return(&delayed_root->items) <
+	    BTRFS_DELAYED_BACKGROUND || seq % 16 == 0) &&
+	    waitqueue_active(&delayed_root->wait))
+		wake_up(&delayed_root->wait);
+}
+
 static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 {
 	struct rb_root *root;
@@ -512,10 +521,8 @@ static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 
 	rb_erase(&delayed_item->rb_node, root);
 	delayed_item->delayed_node->count--;
-	if (atomic_dec_return(&delayed_root->items) <
-	    BTRFS_DELAYED_BACKGROUND &&
-	    waitqueue_active(&delayed_root->wait))
-		wake_up(&delayed_root->wait);
+
+	finish_one_item(delayed_root);
 }
 
 static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
@@ -1056,10 +1063,7 @@ static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
 		delayed_node->count--;
 
 		delayed_root = delayed_node->root->fs_info->delayed_root;
-		if (atomic_dec_return(&delayed_root->items) <
-		    BTRFS_DELAYED_BACKGROUND &&
-		    waitqueue_active(&delayed_root->wait))
-			wake_up(&delayed_root->wait);
+		finish_one_item(delayed_root);
 	}
 }
 
@@ -1304,35 +1308,55 @@ void btrfs_remove_delayed_node(struct inode *inode)
 	btrfs_release_delayed_node(delayed_node);
 }
 
+#if 0
 struct btrfs_async_delayed_node {
 	struct btrfs_root *root;
 	struct btrfs_delayed_node *delayed_node;
 	struct btrfs_work work;
 };
+#endif
+
+struct btrfs_async_delayed_work {
+	struct btrfs_delayed_root *delayed_root;
+	int nr;
+	struct btrfs_work work;
+};
 
-static void btrfs_async_run_delayed_node_done(struct btrfs_work *work)
+static void btrfs_async_run_delayed_root(struct btrfs_work *work)
 {
-	struct btrfs_async_delayed_node *async_node;
+	struct btrfs_async_delayed_work *async_work;
+	struct btrfs_delayed_root *delayed_root;
 	struct btrfs_trans_handle *trans;
 	struct btrfs_path *path;
 	struct btrfs_delayed_node *delayed_node = NULL;
 	struct btrfs_root *root;
 	struct btrfs_block_rsv *block_rsv;
 	int need_requeue = 0;
+	int total_done = 0;
 
-	async_node = container_of(work, struct btrfs_async_delayed_node, work);
+	async_work = container_of(work, struct btrfs_async_delayed_work, work);
+	delayed_root = async_work->delayed_root;
 
 	path = btrfs_alloc_path();
 	if (!path)
 		goto out;
-	path->leave_spinning = 1;
 
-	delayed_node = async_node->delayed_node;
+again:
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND / 2)
+		goto free_path;
+
+	delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
+	if (!delayed_node)
+		goto free_path;
+
+requeue:
+	path->leave_spinning = 1;
+	need_requeue = 0;
 	root = delayed_node->root;
 
 	trans = btrfs_join_transaction(root);
 	if (IS_ERR(trans))
-		goto free_path;
+		goto release_path;
 
 	block_rsv = trans->block_rsv;
 	trans->block_rsv = &root->fs_info->delayed_block_rsv;

fs/btrfs/delayed-inode.c static void btrfs_async_run_delayed_node_done(struct btrfs_work *work) · line 1406

Quoted 10 lines
fs/btrfs/delayed-inode.c +7 -0
@@ -1373,47 +1397,48 @@ static void btrfs_async_run_delayed_node_done(struct btrfs_work *work)
 	trans->block_rsv = block_rsv;
 	btrfs_end_transaction_dmeta(trans, root);
 	btrfs_btree_balance_dirty_nodelay(root);
+
+release_path:
+	btrfs_release_path(path);
+	total_done++;
+
+	if (need_requeue) {
+		goto requeue;

We re-queue the node just when there are some delayed items in the current node. But if the node still has delayed items after we deal with it, that is to say someone is accessing the node. So it is better to release it and deal with it later. In this way, we can amass more items and deal with them in batches.

fs/btrfs/delayed-inode.c static void btrfs_async_run_delayed_node_done(struct btrfs_work *work) · line 1410

Quoted 4 lines
fs/btrfs/delayed-inode.c +4 -0
+	} else {
+		btrfs_release_prepared_delayed_node(delayed_node);
+		if (async_work->nr == 0 || total_done < async_work->nr)
+			goto again;

If joining transaction fails, we should end the async handle. And for case ->nr == 0 (it means there are too many items, we need flush all), we can set ->blocked of the current transaction, in this way, the users can not insert any delayed item for a while, and will wait until the current transation is committed

Quoted 108 lines
fs/btrfs/delayed-inode.c +39 -33
+	}
+
 free_path:
 	btrfs_free_path(path);
 out:
-	if (need_requeue)
-		btrfs_requeue_work(&async_node->work);
-	else {
-		btrfs_release_prepared_delayed_node(delayed_node);
-		kfree(async_node);
-	}
+	wake_up(&delayed_root->wait);
+	kfree(async_work);
 }
 
+
 static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
 				     struct btrfs_root *root, int all)
 {
-	struct btrfs_async_delayed_node *async_node;
-	struct btrfs_delayed_node *curr;
-	int count = 0;
+	struct btrfs_async_delayed_work *async_work;
 
-again:
-	curr = btrfs_first_prepared_delayed_node(delayed_root);
-	if (!curr)
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return 0;
 
-	async_node = kmalloc(sizeof(*async_node), GFP_NOFS);
-	if (!async_node) {
-		btrfs_release_prepared_delayed_node(curr);
+	async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
+	if (!async_work)
 		return -ENOMEM;
-	}
-
-	async_node->root = root;
-	async_node->delayed_node = curr;
-
-	async_node->work.func = btrfs_async_run_delayed_node_done;
-	async_node->work.flags = 0;
-
-	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_node->work);
-	count++;
 
-	if (all || count < 4)
-		goto again;
+	async_work->delayed_root = delayed_root;
+	async_work->work.func = btrfs_async_run_delayed_root;
+	async_work->work.flags = 0;
+	if (all)
+		async_work->nr = 0;
+	else
+		async_work->nr = 16;
 
+	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_work->work);
 	return 0;
 }
 
@@ -1424,30 +1449,52 @@ void btrfs_assert_delayed_root_empty(struct btrfs_root *root)
 	WARN_ON(btrfs_first_delayed_node(delayed_root));
 }
 
+static int refs_newer(struct btrfs_delayed_root *delayed_root,
+		      int seq, int count)
+{
+	int val = atomic_read(&delayed_root->items_seq);
+
+	if (val < seq || val >= seq + count)
+		return 1;
+	return 0;
+}
+
 void btrfs_balance_delayed_items(struct btrfs_root *root)
 {
 	struct btrfs_delayed_root *delayed_root;
+	int seq;
 
 	delayed_root = btrfs_get_delayed_root(root);
 
 	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return;
 
+	seq = atomic_read(&delayed_root->items_seq);
+
 	if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
 		int ret;
+		DEFINE_WAIT(__wait);
+
 		ret = btrfs_wq_run_delayed_node(delayed_root, root, 1);
 		if (ret)
 			return;
 
-		wait_event_interruptible_timeout(
-				delayed_root->wait,
-				(atomic_read(&delayed_root->items) <
-				 BTRFS_DELAYED_BACKGROUND),
-				HZ);
-		return;
+		while (1) {
+			prepare_to_wait(&delayed_root->wait, &__wait,
+					TASK_INTERRUPTIBLE);
+
+			if (refs_newer(delayed_root, seq, 16) ||
+			    atomic_read(&delayed_root->items) <
+			    BTRFS_DELAYED_BACKGROUND) {

fs/btrfs/delayed-inode.c void btrfs_assert_delayed_root_empty(struct btrfs_root *root) · line 1497

Quoted 10 lines
fs/btrfs/delayed-inode.c +7 -1
+				break;
+			}
+			if (!signal_pending(current))
+				schedule();
+		}
+		finish_wait(&delayed_root->wait, &__wait);
 	}
 
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, 16);

The last argument is a bool variant, it shoud be 0 here.

Thanks
Miao

Quoted 28 lines
fs/btrfs/delayed-inode.c +0 -0
 }
 
 /* Will return 0 or -ENOMEM */
fs/btrfs/delayed-inode.h +2 -1
index 78b6ad0..1d5c5f7 100644
@@ -43,6 +43,7 @@ struct btrfs_delayed_root {
 	 */
 	struct list_head prepare_list;
 	atomic_t items;		/* for delayed items */
+	atomic_t items_seq;	/* for delayed items */
 	int nodes;		/* for delayed nodes */
 	wait_queue_head_t wait;
 };
@@ -86,6 +87,7 @@ static inline void btrfs_init_delayed_root(
 				struct btrfs_delayed_root *delayed_root)
 {
 	atomic_set(&delayed_root->items, 0);
+	atomic_set(&delayed_root->items_seq, 0);
 	delayed_root->nodes = 0;
 	spin_lock_init(&delayed_root->lock);
 	init_waitqueue_head(&delayed_root->wait);
--

To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#4 lore raw
Quoted 5 lines · Miao Xie

We re-queue the node just when there are some delayed items in the current node. But if the node still has delayed items after we deal with it, that is to say someone is accessing the node. So it is better to release it and deal with it later. In this way, we can amass more items and deal with them in batches.

Thanks, I've made this change.

Quoted 11 lines

Review

Quoted 5 lines
(quoted diff) +4 -0
+	} else {
+		btrfs_release_prepared_delayed_node(delayed_node);
+		if (async_work->nr == 0 || total_done < async_work->nr)
+			goto again;

If joining transaction fails, we should end the async handle. And for case ->nr == 0 (it means there are too many items, we need flush all), we can set ->blocked of the current transaction, in this way, the users can not insert any delayed item for a while, and will wait until the current transation is committed

This one I've left out for now, the old code didn't block and I'd prefer that we test that change independently.

V2 below, it also has the break Liu Bo mentioned.

From: Chris Mason <[email protected]>
Date: Mon, 4 Mar 2013 17:13:31 -0500
Subject: [PATCH] Btrfs: improve the delayed inode throttling

The delayed inode code batches up changes to the btree in hopes of doing them in bulk. As the changes build up, processes kick off worker threads and wait for them to make progress.

The current code kicks off an async work queue item for each delayed node, which creates a lot of churn. It also uses a fixed 1 HZ waiting period for the throttle, which allows us to build a lot of pending work and can slow down the commit.

This changes us to watch a sequence counter as it is bumped during the operations. We kick off fewer work items and have each work item do more work.

---
 fs/btrfs/delayed-inode.c | 152 ++++++++++++++++++++++++++++-------------------
 fs/btrfs/delayed-inode.h |   2 +
 2 files changed, 94 insertions(+), 60 deletions(-)
fs/btrfs/delayed-inode.c +92 -60
index 0b278b1..46f354a 100644
@@ -22,8 +22,9 @@
 #include "disk-io.h"
 #include "transaction.h"
 
-#define BTRFS_DELAYED_WRITEBACK		400
-#define BTRFS_DELAYED_BACKGROUND	100
+#define BTRFS_DELAYED_WRITEBACK		512
+#define BTRFS_DELAYED_BACKGROUND	128
+#define BTRFS_DELAYED_BATCH		16
 
 static struct kmem_cache *delayed_node_cache;
 
@@ -494,6 +495,15 @@ static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
 					BTRFS_DELAYED_DELETION_ITEM);
 }
 
+static void finish_one_item(struct btrfs_delayed_root *delayed_root)
+{
+	int seq = atomic_inc_return(&delayed_root->items_seq);
+	if ((atomic_dec_return(&delayed_root->items) <
+	    BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0) &&
+	    waitqueue_active(&delayed_root->wait))
+		wake_up(&delayed_root->wait);
+}
+
 static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 {
 	struct rb_root *root;
@@ -512,10 +522,8 @@ static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 
 	rb_erase(&delayed_item->rb_node, root);
 	delayed_item->delayed_node->count--;
-	if (atomic_dec_return(&delayed_root->items) <
-	    BTRFS_DELAYED_BACKGROUND &&
-	    waitqueue_active(&delayed_root->wait))
-		wake_up(&delayed_root->wait);
+
+	finish_one_item(delayed_root);
 }
 
 static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
@@ -1056,10 +1064,7 @@ static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
 		delayed_node->count--;
 
 		delayed_root = delayed_node->root->fs_info->delayed_root;
-		if (atomic_dec_return(&delayed_root->items) <
-		    BTRFS_DELAYED_BACKGROUND &&
-		    waitqueue_active(&delayed_root->wait))
-			wake_up(&delayed_root->wait);
+		finish_one_item(delayed_root);
 	}
 }
 
@@ -1304,35 +1309,44 @@ void btrfs_remove_delayed_node(struct inode *inode)
 	btrfs_release_delayed_node(delayed_node);
 }
 
-struct btrfs_async_delayed_node {
-	struct btrfs_root *root;
-	struct btrfs_delayed_node *delayed_node;
+struct btrfs_async_delayed_work {
+	struct btrfs_delayed_root *delayed_root;
+	int nr;
 	struct btrfs_work work;
 };
 
-static void btrfs_async_run_delayed_node_done(struct btrfs_work *work)
+static void btrfs_async_run_delayed_root(struct btrfs_work *work)
 {
-	struct btrfs_async_delayed_node *async_node;
+	struct btrfs_async_delayed_work *async_work;
+	struct btrfs_delayed_root *delayed_root;
 	struct btrfs_trans_handle *trans;
 	struct btrfs_path *path;
 	struct btrfs_delayed_node *delayed_node = NULL;
 	struct btrfs_root *root;
 	struct btrfs_block_rsv *block_rsv;
-	int need_requeue = 0;
+	int total_done = 0;
 
-	async_node = container_of(work, struct btrfs_async_delayed_node, work);
+	async_work = container_of(work, struct btrfs_async_delayed_work, work);
+	delayed_root = async_work->delayed_root;
 
 	path = btrfs_alloc_path();
 	if (!path)
 		goto out;
-	path->leave_spinning = 1;
 
-	delayed_node = async_node->delayed_node;
+again:
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND / 2)
+		goto free_path;
+
+	delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
+	if (!delayed_node)
+		goto free_path;
+
+	path->leave_spinning = 1;
 	root = delayed_node->root;
 
 	trans = btrfs_join_transaction(root);
 	if (IS_ERR(trans))
-		goto free_path;
+		goto release_path;
 
 	block_rsv = trans->block_rsv;
 	trans->block_rsv = &root->fs_info->delayed_block_rsv;
@@ -1363,57 +1377,50 @@ static void btrfs_async_run_delayed_node_done(struct btrfs_work *work)
 	 * Task1 will sleep until the transaction is commited.
 	 */
 	mutex_lock(&delayed_node->mutex);
-	if (delayed_node->count)
-		need_requeue = 1;
-	else
-		btrfs_dequeue_delayed_node(root->fs_info->delayed_root,
-					   delayed_node);
+	btrfs_dequeue_delayed_node(root->fs_info->delayed_root, delayed_node);
 	mutex_unlock(&delayed_node->mutex);
 
 	trans->block_rsv = block_rsv;
 	btrfs_end_transaction_dmeta(trans, root);
 	btrfs_btree_balance_dirty_nodelay(root);
+
+release_path:
+	btrfs_release_path(path);
+	total_done++;
+
+	btrfs_release_prepared_delayed_node(delayed_node);
+	if (async_work->nr == 0 || total_done < async_work->nr)
+		goto again;
+
 free_path:
 	btrfs_free_path(path);
 out:
-	if (need_requeue)
-		btrfs_requeue_work(&async_node->work);
-	else {
-		btrfs_release_prepared_delayed_node(delayed_node);
-		kfree(async_node);
-	}
+	wake_up(&delayed_root->wait);
+	kfree(async_work);
 }
 
+
 static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
-				     struct btrfs_root *root, int all)
+				     struct btrfs_root *root, int nr)
 {
-	struct btrfs_async_delayed_node *async_node;
-	struct btrfs_delayed_node *curr;
-	int count = 0;
+	struct btrfs_async_delayed_work *async_work;
 
-again:
-	curr = btrfs_first_prepared_delayed_node(delayed_root);
-	if (!curr)
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return 0;
 
-	async_node = kmalloc(sizeof(*async_node), GFP_NOFS);
-	if (!async_node) {
-		btrfs_release_prepared_delayed_node(curr);
+	async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
+	if (!async_work)
 		return -ENOMEM;
-	}
-
-	async_node->root = root;
-	async_node->delayed_node = curr;
-
-	async_node->work.func = btrfs_async_run_delayed_node_done;
-	async_node->work.flags = 0;
-
-	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_node->work);
-	count++;
 
-	if (all || count < 4)
-		goto again;
+	async_work->delayed_root = delayed_root;
+	async_work->work.func = btrfs_async_run_delayed_root;
+	async_work->work.flags = 0;
+	if (nr)
+		async_work->nr = 0;
+	else
+		async_work->nr = nr;
 
+	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_work->work);
 	return 0;
 }
 
@@ -1424,30 +1431,55 @@ void btrfs_assert_delayed_root_empty(struct btrfs_root *root)
 	WARN_ON(btrfs_first_delayed_node(delayed_root));
 }
 
+static int refs_newer(struct btrfs_delayed_root *delayed_root,
+		      int seq, int count)
+{
+	int val = atomic_read(&delayed_root->items_seq);
+
+	if (val < seq || val >= seq + count)
+		return 1;
+	return 0;
+}
+
 void btrfs_balance_delayed_items(struct btrfs_root *root)
 {
 	struct btrfs_delayed_root *delayed_root;
+	int seq;
 
 	delayed_root = btrfs_get_delayed_root(root);
 
 	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return;
 
+	seq = atomic_read(&delayed_root->items_seq);
+
 	if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
 		int ret;
+		DEFINE_WAIT(__wait);
+
 		ret = btrfs_wq_run_delayed_node(delayed_root, root, 1);
 		if (ret)
 			return;
 
-		wait_event_interruptible_timeout(
-				delayed_root->wait,
-				(atomic_read(&delayed_root->items) <
-				 BTRFS_DELAYED_BACKGROUND),
-				HZ);
-		return;
+		while (1) {
+			prepare_to_wait(&delayed_root->wait, &__wait,
+					TASK_INTERRUPTIBLE);
+
+			if (refs_newer(delayed_root, seq,
+				       BTRFS_DELAYED_BATCH) ||
+			    atomic_read(&delayed_root->items) <
+			    BTRFS_DELAYED_BACKGROUND) {
+				break;
+			}
+			if (!signal_pending(current))
+				schedule();
+			else
+				break;
+		}
+		finish_wait(&delayed_root->wait, &__wait);
 	}
 
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, BTRFS_DELAYED_BATCH);
 }
 
 /* Will return 0 or -ENOMEM */
fs/btrfs/delayed-inode.h +2 -0
index 78b6ad0..1d5c5f7 100644
@@ -43,6 +43,7 @@ struct btrfs_delayed_root {
 	 */
 	struct list_head prepare_list;
 	atomic_t items;		/* for delayed items */
+	atomic_t items_seq;	/* for delayed items */
 	int nodes;		/* for delayed nodes */
 	wait_queue_head_t wait;
 };
@@ -86,6 +87,7 @@ static inline void btrfs_init_delayed_root(
 				struct btrfs_delayed_root *delayed_root)
 {
 	atomic_set(&delayed_root->items, 0);
+	atomic_set(&delayed_root->items_seq, 0);
 	delayed_root->nodes = 0;
 	spin_lock_init(&delayed_root->lock);
 	init_waitqueue_head(&delayed_root->wait);
Signature · 1 line
1.8.1.3
#5 lore raw
Quoted 162 lines · Chris Mason
Quoted 5 lines · Miao Xie

We re-queue the node just when there are some delayed items in the current node. But if the node still has delayed items after we deal with it, that is to say someone is accessing the node. So it is better to release it and deal with it later. In this way, we can amass more items and deal with them in batches.

Thanks, I've made this change.

Quoted 11 lines

fs/btrfs/delayed-inode.c static void btrfs_async_run_delayed_node_done(struct btrfs_work *work) · line 1393

Quoted 5 lines
fs/btrfs/delayed-inode.c +4 -0
+	} else {
+		btrfs_release_prepared_delayed_node(delayed_node);
+		if (async_work->nr == 0 || total_done < async_work->nr)
+			goto again;

If joining transaction fails, we should end the async handle. And for case ->nr == 0 (it means there are too many items, we need flush all), we can set ->blocked of the current transaction, in this way, the users can not insert any delayed item for a while, and will wait until the current transation is committed

This one I've left out for now, the old code didn't block and I'd prefer that we test that change independently.

V2 below, it also has the break Liu Bo mentioned.

From: Chris Mason <[email protected]>
Date: Mon, 4 Mar 2013 17:13:31 -0500
Subject: [PATCH] Btrfs: improve the delayed inode throttling

The delayed inode code batches up changes to the btree in hopes of doing them in bulk. As the changes build up, processes kick off worker threads and wait for them to make progress.

The current code kicks off an async work queue item for each delayed node, which creates a lot of churn. It also uses a fixed 1 HZ waiting period for the throttle, which allows us to build a lot of pending work and can slow down the commit.

This changes us to watch a sequence counter as it is bumped during the operations. We kick off fewer work items and have each work item do more work.

---
 fs/btrfs/delayed-inode.c | 152 ++++++++++++++++++++++++++++-------------------
 fs/btrfs/delayed-inode.h |   2 +
 2 files changed, 94 insertions(+), 60 deletions(-)
fs/btrfs/delayed-inode.c +34 -20
index 0b278b1..46f354a 100644
@@ -22,8 +22,9 @@
 #include "disk-io.h"
 #include "transaction.h"
 
-#define BTRFS_DELAYED_WRITEBACK		400
-#define BTRFS_DELAYED_BACKGROUND	100
+#define BTRFS_DELAYED_WRITEBACK		512
+#define BTRFS_DELAYED_BACKGROUND	128
+#define BTRFS_DELAYED_BATCH		16
 
 static struct kmem_cache *delayed_node_cache;
 
@@ -494,6 +495,15 @@ static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
 					BTRFS_DELAYED_DELETION_ITEM);
 }
 
+static void finish_one_item(struct btrfs_delayed_root *delayed_root)
+{
+	int seq = atomic_inc_return(&delayed_root->items_seq);
+	if ((atomic_dec_return(&delayed_root->items) <
+	    BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0) &&
+	    waitqueue_active(&delayed_root->wait))
+		wake_up(&delayed_root->wait);
+}
+
 static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 {
 	struct rb_root *root;
@@ -512,10 +522,8 @@ static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 
 	rb_erase(&delayed_item->rb_node, root);
 	delayed_item->delayed_node->count--;
-	if (atomic_dec_return(&delayed_root->items) <
-	    BTRFS_DELAYED_BACKGROUND &&
-	    waitqueue_active(&delayed_root->wait))
-		wake_up(&delayed_root->wait);
+
+	finish_one_item(delayed_root);
 }
 
 static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
@@ -1056,10 +1064,7 @@ static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
 		delayed_node->count--;
 
 		delayed_root = delayed_node->root->fs_info->delayed_root;
-		if (atomic_dec_return(&delayed_root->items) <
-		    BTRFS_DELAYED_BACKGROUND &&
-		    waitqueue_active(&delayed_root->wait))
-			wake_up(&delayed_root->wait);
+		finish_one_item(delayed_root);
 	}
 }
 
@@ -1304,35 +1309,44 @@ void btrfs_remove_delayed_node(struct inode *inode)
 	btrfs_release_delayed_node(delayed_node);
 }
 
-struct btrfs_async_delayed_node {
-	struct btrfs_root *root;
-	struct btrfs_delayed_node *delayed_node;
+struct btrfs_async_delayed_work {
+	struct btrfs_delayed_root *delayed_root;
+	int nr;
 	struct btrfs_work work;
 };
 
-static void btrfs_async_run_delayed_node_done(struct btrfs_work *work)
+static void btrfs_async_run_delayed_root(struct btrfs_work *work)
 {
-	struct btrfs_async_delayed_node *async_node;
+	struct btrfs_async_delayed_work *async_work;
+	struct btrfs_delayed_root *delayed_root;
 	struct btrfs_trans_handle *trans;
 	struct btrfs_path *path;
 	struct btrfs_delayed_node *delayed_node = NULL;
 	struct btrfs_root *root;
 	struct btrfs_block_rsv *block_rsv;
-	int need_requeue = 0;
+	int total_done = 0;
 
-	async_node = container_of(work, struct btrfs_async_delayed_node, work);
+	async_work = container_of(work, struct btrfs_async_delayed_work, work);
+	delayed_root = async_work->delayed_root;
 
 	path = btrfs_alloc_path();
 	if (!path)
 		goto out;
-	path->leave_spinning = 1;
 
-	delayed_node = async_node->delayed_node;
+again:
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND / 2)
+		goto free_path;
+
+	delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
+	if (!delayed_node)
+		goto free_path;
+
+	path->leave_spinning = 1;
 	root = delayed_node->root;
 
 	trans = btrfs_join_transaction(root);
 	if (IS_ERR(trans))
-		goto free_path;
+		goto release_path;
 
 	block_rsv = trans->block_rsv;
 	trans->block_rsv = &root->fs_info->delayed_block_rsv;

fs/btrfs/delayed-inode.c static void btrfs_async_run_delayed_node_done(struct btrfs_work *work) · line 1381

Quoted 10 lines
fs/btrfs/delayed-inode.c +1 -5
@@ -1363,57 +1377,50 @@ static void btrfs_async_run_delayed_node_done(struct btrfs_work *work)
 	 * Task1 will sleep until the transaction is commited.
 	 */
 	mutex_lock(&delayed_node->mutex);
-	if (delayed_node->count)
-		need_requeue = 1;
-	else
-		btrfs_dequeue_delayed_node(root->fs_info->delayed_root,
-					   delayed_node);
+	btrfs_dequeue_delayed_node(root->fs_info->delayed_root, delayed_node);
 	mutex_unlock(&delayed_node->mutex);

The code in this mutex can be removed, this implementation won't trigger the problem that the comment mentions.

Thanks
Miao

Quoted 157 lines
fs/btrfs/delayed-inode.c +57 -35
 	trans->block_rsv = block_rsv;
 	btrfs_end_transaction_dmeta(trans, root);
 	btrfs_btree_balance_dirty_nodelay(root);
+
+release_path:
+	btrfs_release_path(path);
+	total_done++;
+
+	btrfs_release_prepared_delayed_node(delayed_node);
+	if (async_work->nr == 0 || total_done < async_work->nr)
+		goto again;
+
 free_path:
 	btrfs_free_path(path);
 out:
-	if (need_requeue)
-		btrfs_requeue_work(&async_node->work);
-	else {
-		btrfs_release_prepared_delayed_node(delayed_node);
-		kfree(async_node);
-	}
+	wake_up(&delayed_root->wait);
+	kfree(async_work);
 }
 
+
 static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
-				     struct btrfs_root *root, int all)
+				     struct btrfs_root *root, int nr)
 {
-	struct btrfs_async_delayed_node *async_node;
-	struct btrfs_delayed_node *curr;
-	int count = 0;
+	struct btrfs_async_delayed_work *async_work;
 
-again:
-	curr = btrfs_first_prepared_delayed_node(delayed_root);
-	if (!curr)
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return 0;
 
-	async_node = kmalloc(sizeof(*async_node), GFP_NOFS);
-	if (!async_node) {
-		btrfs_release_prepared_delayed_node(curr);
+	async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
+	if (!async_work)
 		return -ENOMEM;
-	}
-
-	async_node->root = root;
-	async_node->delayed_node = curr;
-
-	async_node->work.func = btrfs_async_run_delayed_node_done;
-	async_node->work.flags = 0;
-
-	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_node->work);
-	count++;
 
-	if (all || count < 4)
-		goto again;
+	async_work->delayed_root = delayed_root;
+	async_work->work.func = btrfs_async_run_delayed_root;
+	async_work->work.flags = 0;
+	if (nr)
+		async_work->nr = 0;
+	else
+		async_work->nr = nr;
 
+	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_work->work);
 	return 0;
 }
 
@@ -1424,30 +1431,55 @@ void btrfs_assert_delayed_root_empty(struct btrfs_root *root)
 	WARN_ON(btrfs_first_delayed_node(delayed_root));
 }
 
+static int refs_newer(struct btrfs_delayed_root *delayed_root,
+		      int seq, int count)
+{
+	int val = atomic_read(&delayed_root->items_seq);
+
+	if (val < seq || val >= seq + count)
+		return 1;
+	return 0;
+}
+
 void btrfs_balance_delayed_items(struct btrfs_root *root)
 {
 	struct btrfs_delayed_root *delayed_root;
+	int seq;
 
 	delayed_root = btrfs_get_delayed_root(root);
 
 	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return;
 
+	seq = atomic_read(&delayed_root->items_seq);
+
 	if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
 		int ret;
+		DEFINE_WAIT(__wait);
+
 		ret = btrfs_wq_run_delayed_node(delayed_root, root, 1);
 		if (ret)
 			return;
 
-		wait_event_interruptible_timeout(
-				delayed_root->wait,
-				(atomic_read(&delayed_root->items) <
-				 BTRFS_DELAYED_BACKGROUND),
-				HZ);
-		return;
+		while (1) {
+			prepare_to_wait(&delayed_root->wait, &__wait,
+					TASK_INTERRUPTIBLE);
+
+			if (refs_newer(delayed_root, seq,
+				       BTRFS_DELAYED_BATCH) ||
+			    atomic_read(&delayed_root->items) <
+			    BTRFS_DELAYED_BACKGROUND) {
+				break;
+			}
+			if (!signal_pending(current))
+				schedule();
+			else
+				break;
+		}
+		finish_wait(&delayed_root->wait, &__wait);
 	}
 
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, BTRFS_DELAYED_BATCH);
 }
 
 /* Will return 0 or -ENOMEM */
fs/btrfs/delayed-inode.h +2 -0
index 78b6ad0..1d5c5f7 100644
@@ -43,6 +43,7 @@ struct btrfs_delayed_root {
 	 */
 	struct list_head prepare_list;
 	atomic_t items;		/* for delayed items */
+	atomic_t items_seq;	/* for delayed items */
 	int nodes;		/* for delayed nodes */
 	wait_queue_head_t wait;
 };
@@ -86,6 +87,7 @@ static inline void btrfs_init_delayed_root(
 				struct btrfs_delayed_root *delayed_root)
 {
 	atomic_set(&delayed_root->items, 0);
+	atomic_set(&delayed_root->items_seq, 0);
 	delayed_root->nodes = 0;
 	spin_lock_init(&delayed_root->lock);
 	init_waitqueue_head(&delayed_root->wait);
#6 lore raw
In reply to Chris Mason

On wed, 6 Mar 2013 09:53:28 -0500, Chris Mason wrote:
[SNIP]

Quoted 31 lines

static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,

fs/btrfs/delayed-inode.c +5 -20
-				     struct btrfs_root *root, int all)
+				     struct btrfs_root *root, int nr)
 {
-	struct btrfs_async_delayed_node *async_node;
-	struct btrfs_delayed_node *curr;
-	int count = 0;
+	struct btrfs_async_delayed_work *async_work;
 
-again:
-	curr = btrfs_first_prepared_delayed_node(delayed_root);
-	if (!curr)
+	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return 0;
 
-	async_node = kmalloc(sizeof(*async_node), GFP_NOFS);
-	if (!async_node) {
-		btrfs_release_prepared_delayed_node(curr);
+	async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
+	if (!async_work)
 		return -ENOMEM;
-	}
-
-	async_node->root = root;
-	async_node->delayed_node = curr;
-
-	async_node->work.func = btrfs_async_run_delayed_node_done;
-	async_node->work.flags = 0;
-
-	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_node->work);
-	count++;

fs/btrfs/delayed-inode.c static void btrfs_async_run_delayed_node_done(struct btrfs_work *work) · line 1421

Quoted 10 lines
fs/btrfs/delayed-inode.c +7 -2
 
-	if (all || count < 4)
-		goto again;
+	async_work->delayed_root = delayed_root;
+	async_work->work.func = btrfs_async_run_delayed_root;
+	async_work->work.flags = 0;
+	if (nr)
+		async_work->nr = 0;
+	else
+		async_work->nr = nr;

the code here is wrong.
the argument nr is the number we want to deal with, if it is 0, we will deal with all. so
- if (nr)
- async_work->nr = 0;
- else
- async_work->nr = nr;
+ async_work->nr = nr;

Quoted 26 lines
fs/btrfs/delayed-inode.c +12 -0
+	btrfs_queue_worker(&root->fs_info->delayed_workers, &async_work->work);
 	return 0;
 }
@@ -1424,30 +1431,55 @@ void btrfs_assert_delayed_root_empty(struct btrfs_root *root)
 	WARN_ON(btrfs_first_delayed_node(delayed_root));
 }
 
+static int refs_newer(struct btrfs_delayed_root *delayed_root,
+		      int seq, int count)
+{
+	int val = atomic_read(&delayed_root->items_seq);
+
+	if (val < seq || val >= seq + count)
+		return 1;
+	return 0;
+}
+
 void btrfs_balance_delayed_items(struct btrfs_root *root)
 {
 	struct btrfs_delayed_root *delayed_root;
+	int seq;
 
 	delayed_root = btrfs_get_delayed_root(root);
 

fs/btrfs/delayed-inode.c void btrfs_assert_delayed_root_empty(struct btrfs_root *root) · line 1460

Quoted 10 lines
fs/btrfs/delayed-inode.c +4 -0
 	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
 		return;
 
+	seq = atomic_read(&delayed_root->items_seq);
+
 	if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
 		int ret;
+		DEFINE_WAIT(__wait);
+
 		ret = btrfs_wq_run_delayed_node(delayed_root, root, 1);

here
- ret = btrfs_wq_run_delayed_node(delayed_root, root, 1);
+ ret = btrfs_wq_run_delayed_node(delayed_root, root, 0);

Quoted 20 lines
fs/btrfs/delayed-inode.c +11 -6
 		if (ret)
 			return;
 
-		wait_event_interruptible_timeout(
-				delayed_root->wait,
-				(atomic_read(&delayed_root->items) <
-				 BTRFS_DELAYED_BACKGROUND),
-				HZ);
-		return;
+		while (1) {
+			prepare_to_wait(&delayed_root->wait, &__wait,
+					TASK_INTERRUPTIBLE);
+
+			if (refs_newer(delayed_root, seq,
+				       BTRFS_DELAYED_BATCH) ||
+			    atomic_read(&delayed_root->items) <
+			    BTRFS_DELAYED_BACKGROUND) {
+				break;
+			}
+			if (!signal_pending(current))

fs/btrfs/delayed-inode.c void btrfs_assert_delayed_root_empty(struct btrfs_root *root) · line 1483

Quoted 10 lines
fs/btrfs/delayed-inode.c +6 -1
+				schedule();
+			else
+				break;
+		}
+		finish_wait(&delayed_root->wait, &__wait);
 	}
 
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, BTRFS_DELAYED_BATCH);
 }

There is a problem that we may introduce lots of btrfs_works, we need avoid it.

Thanks
Miao

#7 lore raw
Quoted 12 lines · Miao Xie

On wed, 6 Mar 2013 09:53:28 -0500, Chris Mason wrote:
[SNIP]

fs/btrfs/delayed-inode.c static void btrfs_async_run_delayed_node_done(struct btrfs_work *work) · line 1421

Quoted 7 lines
fs/btrfs/delayed-inode.c +7 -0
+	async_work->delayed_root = delayed_root;
+	async_work->work.func = btrfs_async_run_delayed_root;
+	async_work->work.flags = 0;
+	if (nr)
+		async_work->nr = 0;
+	else
+		async_work->nr = nr;

the code here is wrong.
the argument nr is the number we want to deal with, if it is 0, we will deal with all.

Whoops, thanks. I missed that when I was cleaning things up.

Quoted 7 lines

fs/btrfs/delayed-inode.c void btrfs_assert_delayed_root_empty(struct btrfs_root *root) · line 1483

Quoted 4 lines
fs/btrfs/delayed-inode.c +1 -1
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, BTRFS_DELAYED_BATCH);
 }

There is a problem that we may introduce lots of btrfs_works, we need avoid it.

It is possible, but we won't make more than we used to. The real solution is to limit the workers per root, but the code isn't currently structured for that. Right now the workers will exit out if the number of pending items is below the delayed limit, which isn't perfect but I think it's the best I can do right now.

Do you see better ways to improve it?

#8 lore raw
Quoted 24 lines · Chris Mason
Quoted 12 lines · Miao Xie

On wed, 6 Mar 2013 09:53:28 -0500, Chris Mason wrote:
[SNIP]

fs/btrfs/delayed-inode.c static void btrfs_async_run_delayed_node_done(struct btrfs_work *work) · line 1421

Quoted 7 lines
fs/btrfs/delayed-inode.c +7 -0
+	async_work->delayed_root = delayed_root;
+	async_work->work.func = btrfs_async_run_delayed_root;
+	async_work->work.flags = 0;
+	if (nr)
+		async_work->nr = 0;
+	else
+		async_work->nr = nr;

the code here is wrong.
the argument nr is the number we want to deal with, if it is 0, we will deal with all.

Whoops, thanks. I missed that when I was cleaning things up.

Quoted 7 lines

fs/btrfs/delayed-inode.c void btrfs_assert_delayed_root_empty(struct btrfs_root *root) · line 1483

Quoted 4 lines
fs/btrfs/delayed-inode.c +1 -1
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, BTRFS_DELAYED_BATCH);
 }

There is a problem that we may introduce lots of btrfs_works, we need avoid it.

Quoted 7 lines

It is possible, but we won't make more than we used to. The real solution is to limit the workers per root, but the code isn't currently structured for that. Right now the workers will exit out if the number of pending items is below the delayed limit, which isn't perfect but I think it's the best I can do right now.

Do you see better ways to improve it?

How do you think about per-cpu btrfs_work? If btrfs_work on the current cpu is dealt with, we don't queue it, just update ->nr if need and tell the workers that we need do flush again.

(This way is a bit ugly because btrfs_work might not be handled on its cpu)

Thanks
Miao

Quoted 5 lines

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#9 lore raw
Quoted 31 lines · Miao Xie
Quoted 30 lines · Chris Mason
Quoted 12 lines · Miao Xie

On wed, 6 Mar 2013 09:53:28 -0500, Chris Mason wrote:
[SNIP]

fs/btrfs/delayed-inode.c static void btrfs_async_run_delayed_node_done(struct btrfs_work *work) · line 1421

Quoted 7 lines
fs/btrfs/delayed-inode.c +7 -0
+	async_work->delayed_root = delayed_root;
+	async_work->work.func = btrfs_async_run_delayed_root;
+	async_work->work.flags = 0;
+	if (nr)
+		async_work->nr = 0;
+	else
+		async_work->nr = nr;

the code here is wrong.
the argument nr is the number we want to deal with, if it is 0, we will deal with all.

Whoops, thanks. I missed that when I was cleaning things up.

Quoted 7 lines

fs/btrfs/delayed-inode.c void btrfs_assert_delayed_root_empty(struct btrfs_root *root) · line 1483

Quoted 4 lines
fs/btrfs/delayed-inode.c +1 -1
-	btrfs_wq_run_delayed_node(delayed_root, root, 0);
+	btrfs_wq_run_delayed_node(delayed_root, root, BTRFS_DELAYED_BATCH);
 }

There is a problem that we may introduce lots of btrfs_works, we need avoid it.

It is possible, but we won't make more than we used to. The real solution is to limit the workers per root, but the code isn't currently structured for that. Right now the workers will exit out if the number of pending items is below the delayed limit, which isn't perfect but I think it's the best I can do right now.

Quoted 7 lines
Quoted 1 line

Do you see better ways to improve it?

How do you think about per-cpu btrfs_work? If btrfs_work on the current cpu is dealt with, we don't queue it, just update ->nr if need and tell the workers that we need do flush again.

(This way is a bit ugly because btrfs_work might not be handled on its cpu)

Yeah, I'd prefer that we add a per-root counter of some kind.

-chris