From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from dkim1.fusionio.com ([66.114.96.53]:44783 "EHLO dkim1.fusionio.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754489Ab3CEPkU (ORCPT ); Tue, 5 Mar 2013 10:40:20 -0500 Received: from mx2.fusionio.com (unknown [10.101.1.160]) by dkim1.fusionio.com (Postfix) with ESMTP id 284267C039C for ; Tue, 5 Mar 2013 08:40:20 -0700 (MST) Date: Tue, 5 Mar 2013 10:40:17 -0500 From: Chris Mason To: Linux Btrfs List , Miao Xie Subject: [PATCH] Btrfs: improve the delayed inode throttling Message-ID: <20130305154017.GF30680@shiny.masoncoding.com> MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Sender: linux-btrfs-owner@vger.kernel.org List-ID: 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. Signed-off-by: Chris Mason diff --git a/fs/btrfs/delayed-inode.c b/fs/btrfs/delayed-inode.c index 0b278b1..460d1a8 100644 --- a/fs/btrfs/delayed-inode.c +++ b/fs/btrfs/delayed-inode.c @@ -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 */ diff --git a/fs/btrfs/delayed-inode.h b/fs/btrfs/delayed-inode.h index 78b6ad0..1d5c5f7 100644 --- a/fs/btrfs/delayed-inode.h +++ b/fs/btrfs/delayed-inode.h @@ -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);