Date:	Thu, 28 Jan 2016 09:36:37 +0800
From:	Chao Yu <chao2.yu@...sung.com>
To:	'Jaegeuk Kim' <jaegeuk@...nel.org>
Cc:	linux-f2fs-devel@...ts.sourceforge.net,
	linux-kernel@...r.kernel.org
Subject: RE: [PATCH] f2fs: remove vlist in extent node

Hi Jaegeuk,

> -----Original Message-----
> From: Jaegeuk Kim [mailto:jaegeuk@...nel.org]
> Sent: Thursday, January 28, 2016 5:38 AM
> To: Chao Yu
> Cc: linux-f2fs-devel@...ts.sourceforge.net; linux-kernel@...r.kernel.org
> Subject: Re: [PATCH] f2fs: remove vlist in extent node
> 
> Hi Chao,
> 
> Hmm. The original patch was just going under testing, and we couldn't post
> them since there is a kernel panic issue.
> This patch seems quite better approach, so I think we can integrate both of
> the patches together.
> 
> So, how about this patch?
> Hope you don't mind this.

No objection. :)

> 
> Thanks,
> 
> From a7844e0438db9ea9d12b7c0c40b655c3371bd6c9 Mon Sep 17 00:00:00 2001
> From: Hou Pengyang <houpengyang@...wei.com>
> Date: Tue, 26 Jan 2016 12:56:26 +0000
> Subject: [PATCH] f2fs: improve shrink performance of extent nodes
> 
> On the worst case, we need to scan the whole radix tree and every rb-tree to
> free the victimed extent_nodes when shrinking.
> 
> Pengyang initially introduced a victim_list to record the victimed extent_nodes,
> and free these extent_nodes by just scanning a list.
> 
> Later, Chao Yu enhances the original patch to improve memory footprint by
> removing victim list.
> 
> The policy of lru list shrinking becomes:
> 1) lock lru list's lock
> 2) trylock extent tree's lock
> 3) remove extent node from lru list
> 4) unlock lru list's lock
> 5) do shrink
> 6) repeat 1) to 5)
> 
> Signed-off-by: Hou Pengyang <houpengyang@...wei.com>
> Signed-off-by: Chao Yu <chao2.yu@...sung.com>
> Signed-off-by: Jaegeuk Kim <jaegeuk@...nel.org>
> ---
>  fs/f2fs/extent_cache.c | 87 +++++++++++++++++++++-----------------------------
>  fs/f2fs/f2fs.h         |  1 +
>  2 files changed, 37 insertions(+), 51 deletions(-)
> 
> diff --git a/fs/f2fs/extent_cache.c b/fs/f2fs/extent_cache.c
> index aae99f2..759b1b1 100644
> --- a/fs/f2fs/extent_cache.c
> +++ b/fs/f2fs/extent_cache.c
> @@ -33,6 +33,7 @@ static struct extent_node *__attach_extent_node(struct f2fs_sb_info *sbi,
> 
>  	en->ei = *ei;
>  	INIT_LIST_HEAD(&en->list);
> +	en->et = et;
> 
>  	rb_link_node(&en->rb_node, parent, p);
>  	rb_insert_color(&en->rb_node, &et->root);
> @@ -63,8 +64,8 @@ static void __release_extent_node(struct f2fs_sb_info *sbi,
>  			struct extent_tree *et, struct extent_node *en)
>  {
>  	spin_lock(&sbi->extent_lock);
> -	if (!list_empty(&en->list))
> -		list_del_init(&en->list);
> +	f2fs_bug_on(sbi, list_empty(&en->list));
> +	list_del_init(&en->list);
>  	spin_unlock(&sbi->extent_lock);
> 
>  	__detach_extent_node(sbi, et, en);
> @@ -147,7 +148,7 @@ static struct extent_node *__init_extent_tree(struct f2fs_sb_info *sbi,
>  }
> 
>  static unsigned int __free_extent_tree(struct f2fs_sb_info *sbi,
> -					struct extent_tree *et, bool free_all)
> +					struct extent_tree *et)
>  {
>  	struct rb_node *node, *next;
>  	struct extent_node *en;
> @@ -157,11 +158,7 @@ static unsigned int __free_extent_tree(struct f2fs_sb_info *sbi,
>  	while (node) {
>  		next = rb_next(node);
>  		en = rb_entry(node, struct extent_node, rb_node);
> -
> -		if (free_all)
> -			__release_extent_node(sbi, et, en);
> -		else if (list_empty(&en->list))
> -			__detach_extent_node(sbi, et, en);
> +		__release_extent_node(sbi, et, en);
>  		node = next;
>  	}
> 
> @@ -532,7 +529,7 @@ static unsigned int f2fs_update_extent_tree_range(struct inode *inode,
>  	}
> 
>  	if (is_inode_flag_set(F2FS_I(inode), FI_NO_EXTENT))
> -		__free_extent_tree(sbi, et, true);
> +		__free_extent_tree(sbi, et);
> 
>  	write_unlock(&et->lock);
> 
> @@ -541,14 +538,10 @@ static unsigned int f2fs_update_extent_tree_range(struct inode *inode,
> 
>  unsigned int f2fs_shrink_extent_tree(struct f2fs_sb_info *sbi, int nr_shrink)
>  {
> -	struct extent_tree *treevec[EXT_TREE_VEC_SIZE];
>  	struct extent_tree *et, *next;
> -	struct extent_node *en, *tmp;
> -	unsigned long ino = F2FS_ROOT_INO(sbi);
> -	unsigned int found;
> +	struct extent_node *en;
>  	unsigned int node_cnt = 0, tree_cnt = 0;
>  	int remained;
> -	bool do_free = false;
> 
>  	if (!test_opt(sbi, EXTENT_CACHE))
>  		return 0;
> @@ -561,11 +554,11 @@ unsigned int f2fs_shrink_extent_tree(struct f2fs_sb_info *sbi, int
> nr_shrink)
> 
>  	/* 1. remove unreferenced extent tree */
>  	list_for_each_entry_safe(et, next, &sbi->zombie_list, list) {
> -		if (atomic_read(&et->node_cnt)) {

This is used to avoid lock overhead if there are no nodes in the tree.
Why should we change this?

> -			write_lock(&et->lock);
> -			node_cnt += __free_extent_tree(sbi, et, true);
> -			write_unlock(&et->lock);
> -		}
> +		write_lock(&et->lock);
> +		node_cnt += __free_extent_tree(sbi, et);
> +		write_unlock(&et->lock);
> +		if (atomic_read(&et->node_cnt) > 0)
> +			goto unlock_out;
> 
>  		list_del_init(&et->list);
>  		radix_tree_delete(&sbi->extent_tree_root, et->ino);
> @@ -587,42 +580,29 @@ free_node:
>  	remained = nr_shrink - (node_cnt + tree_cnt);
> 
>  	spin_lock(&sbi->extent_lock);
> -	list_for_each_entry_safe(en, tmp, &sbi->extent_list, list) {
> -		if (!remained--)
> +	for (; remained > 0; remained--) {
> +		if (list_empty(&sbi->extent_list))
>  			break;
> -		list_del_init(&en->list);
> -		do_free = true;
> -	}
> -	spin_unlock(&sbi->extent_lock);
> -
> -	if (do_free == false)
> -		goto unlock_out;
> -
> -	/*
> -	 * reset ino for searching victims from beginning of global extent tree.
> -	 */
> -	ino = F2FS_ROOT_INO(sbi);
> -
> -	while ((found = radix_tree_gang_lookup(&sbi->extent_tree_root,
> -				(void **)treevec, ino, EXT_TREE_VEC_SIZE))) {
> -		unsigned i;
> -
> -		ino = treevec[found - 1]->ino + 1;
> -		for (i = 0; i < found; i++) {
> -			struct extent_tree *et = treevec[i];
> +		en = list_first_entry(&sbi->extent_list,
> +					struct extent_node, list);
> +		et = en->et;
> +		if (!write_trylock(&et->lock)) {
> +			/* refresh this extent node's position in extent list */
> +			list_move_tail(&en->list, &sbi->extent_list);
> +			continue;
> +		}
> 
> -			if (!atomic_read(&et->node_cnt))
> -				continue;
> +		list_del_init(&en->list);
> +		spin_unlock(&sbi->extent_lock);
> 
> -			if (write_trylock(&et->lock)) {
> -				node_cnt += __free_extent_tree(sbi, et, false);
> -				write_unlock(&et->lock);
> -			}
> +		__detach_extent_node(sbi, et, en);
> 
> -			if (node_cnt + tree_cnt >= nr_shrink)
> -				goto unlock_out;
> -		}
> +		write_unlock(&et->lock);
> +		node_cnt++;
> +		spin_lock(&sbi->extent_lock);
>  	}
> +	spin_unlock(&sbi->extent_lock);
> +
>  unlock_out:
>  	up_write(&sbi->extent_tree_lock);
>  out:
> @@ -641,7 +621,7 @@ unsigned int f2fs_destroy_extent_node(struct inode *inode)
>  		return 0;
> 
>  	write_lock(&et->lock);
> -	node_cnt = __free_extent_tree(sbi, et, true);
> +	node_cnt = __free_extent_tree(sbi, et);
>  	write_unlock(&et->lock);
> 
>  	return node_cnt;
> @@ -666,10 +646,15 @@ void f2fs_destroy_extent_tree(struct inode *inode)
>  	}
> 
>  	/* free all extent info belong to this extent tree */
> +free_more:
>  	node_cnt = f2fs_destroy_extent_node(inode);
> 
>  	/* delete extent tree entry in radix tree */
>  	down_write(&sbi->extent_tree_lock);
> +	if (atomic_read(&et->node_cnt) > 0) {
> +		up_write(&sbi->extent_tree_lock);
> +		goto free_more;
> +	}

If I understand correctly here, there is no race condition between shrinker
and destroyer, so it would be safe to usef2fs_bug_on(, et->node_cnt)?

Thanks,

>  	f2fs_bug_on(sbi, atomic_read(&et->node_cnt));
>  	radix_tree_delete(&sbi->extent_tree_root, inode->i_ino);
>  	kmem_cache_free(extent_tree_slab, et);
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index c4e723b..4e7eab9 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -354,6 +354,7 @@ struct extent_node {
>  	struct rb_node rb_node;		/* rb node located in rb-tree */
>  	struct list_head list;		/* node in global extent list of sbi */
>  	struct extent_info ei;		/* extent info */
> +	struct extent_tree *et;		/* extent tree pointer */
>  };
> 
>  struct extent_tree {
> --
> 2.6.3

Powered by blists - more mailing lists