Btrfs: improve the delayed inode throttling
authorChris Mason <chris.mason@fusionio.com>
Mon, 4 Mar 2013 22:13:31 +0000 (17:13 -0500)
committerChris Mason <chris.mason@fusionio.com>
Thu, 7 Mar 2013 12:52:40 +0000 (07:52 -0500)
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.

Signed-off-by: Chris Mason <chris.mason@fusionio.com>
fs/btrfs/delayed-inode.c
fs/btrfs/delayed-inode.h

index 0b278b117cbe609611764cf45127f89a133eed5f..14fce27b4780803c53d0a04b5b8557c1096e61fb 100644 (file)
@@ -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 +1377,47 @@ 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;
+       async_work->nr = nr;
 
+       btrfs_queue_worker(&root->fs_info->delayed_workers, &async_work->work);
        return 0;
 }
 
@@ -1424,30 +1428,55 @@ 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;
-               ret = btrfs_wq_run_delayed_node(delayed_root, root, 1);
+               DEFINE_WAIT(__wait);
+
+               ret = btrfs_wq_run_delayed_node(delayed_root, root, 0);
                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 */
index 78b6ad0fc6699c5c59d5f9ccddef641aa80c5d09..1d5c5f7abe3e01bf872913d123f5c43b2797fd2b 100644 (file)
@@ -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);