Re: [f2fs-dev][RFC PATCH 06/10] f2fs: add core functions for rb-tree extent cache

[Jaegeuk Kim <jaegeuk@kernel.org>]

On Fri, Jan 23, 2015 at 02:15:56PM +0800, Chao Yu wrote:
> Hi Jaegeuk,
> 
> > -----Original Message-----
> > From: Jaegeuk Kim [mailto:jaegeuk@kernel.org]
> > Sent: Friday, January 23, 2015 9:48 AM
> > To: Chao Yu
> > Cc: Changman Lee; linux-f2fs-devel@lists.sourceforge.net; linux-kernel@vger.kernel.org
> > Subject: Re: [f2fs-dev][RFC PATCH 06/10] f2fs: add core functions for rb-tree extent cache
> > 
> > On Mon, Jan 12, 2015 at 03:14:48PM +0800, Chao Yu wrote:
> > > This patch adds core functions including slab cache init function and
> > > init/lookup/update/shrink/destroy function for rb-tree based extent cache.
> > >
> > > Thank Jaegeuk Kim and Changman Lee as they gave much suggestion about detail
> > > design and implementation of extent cache.
> > >
> > > Todo:
> > >  * add a cached_ei into struct extent_tree for a quick recent cache.
> > >  * register rb-based extent cache shrink with mm shrink interface.
> > >  * disable dir inode's extent cache.
> > >

snip

> 
> > 
> > > +	}
> > 
> > How about adding __attach_extent_node()?
> 
> That's more readable.
> 
> How do you think of the following functions:
> 
> struct extent_node *__attach_extent_node(struct f2fs_sb_info *sbi,
> 				struct extent_tree *et, struct extent_info *ei,
> 				struct rb_node *parent, struct rb_node **p)
> {
> 	struct extent_node *en;
> 
> 	en = kmem_cache_alloc(extent_node_slab, GFP_ATOMIC);
> 	if (!en)
> 		return NULL;
> 
> 	en->ei = *ei;
> 	INIT_LIST_HEAD(&en->list);
> 
> 	rb_link_node(&en->rb_node, parent, p);
> 	rb_insert_color(&en->rb_node, &et->root);
> 	et->count++;
> 	atomic_inc(&sbi->total_ext_node);
> 	return en;
> }
> 
> static void __detach_extent_node(struct f2fs_sb_info *sbi,
> 				struct extent_tree *et, struct extent_node *en)
> {
> 	rb_erase(&en->rb_node, &et->root);
> 	et->count--;
> 	atomic_dec(&sbi->total_ext_node);
> }

Looks good to me.

> 
> > {
> > > +
> > > +	en = kmem_cache_alloc(extent_node_slab, GFP_ATOMIC);
> > 
> > 	en = f2fs_kmem_cache_alloc(.., GFP_ATOMIC);
> 
> We should avoid cond_resched() in f2fs_kmem_cache_alloc when we are holding write_lock.
> 
> IMO, it's better to return NULL if we fail to alloc extent_node here.
> Otherwise we'd better alloc extent_node before write_lock.

Oh, it's write_lock. Okay.

> 
> > 
> > > +	en->ei = *ei;
> > > +	INIT_LIST_HEAD(&en->list);
> > > +
> > > +	rb_link_node(&en->rb_node, parent, p);
> > > +	rb_insert_color(&en->rb_node, &et->root);
> > > +	atomic_inc(&sbi->total_ext_node);
> > > +	et->count++;
> > 
> > }
> > 
> > > +
> > > +	return en;
> > > +}
> > > +
> > > +static struct extent_node *__remove_extent_tree(struct extent_tree *et,
> > > +							unsigned int fofs)
> > 
> > This is __detach_extent_node()?
> > 
> > > +{
> > > +	struct rb_node *p = et->root.rb_node;
> > > +	struct extent_node *en;
> > > +
> > > +	while (p) {
> > > +		en = rb_entry(p, struct extent_node, rb_node);
> > > +
> > > +		if (fofs < en->ei.fofs)
> > > +			p = p->rb_left;
> > 
> > Coding style.
> > 
> > if () {
> > } else {
> > }
> 
> will fix.
> 
> > 
> > > +		else if (fofs >= en->ei.fofs + en->ei.len)
> > > +			p = p->rb_right;
> > > +		else {
> > > +			rb_erase(&en->rb_node, &et->root);
> > > +			et->count--;
> > 
> > Add here?
> > 			atomic_dec(&sbi->total_ext_node);
> 
> Agree, see __detach_extent_node() implementation above.
> 
> > 
> > > +			return en;
> > > +		}
> > > +	}
> > > +	return NULL;
> > > +}
> > > +
> > > +static unsigned int __free_extent_tree(struct f2fs_sb_info *sbi,
> > > +					struct extent_tree *et, bool free_all)
> > > +{
> > > +	struct rb_node *node, *next;
> > > +	struct extent_node *en;
> > > +	unsigned int count = et->count;
> > > +
> > > +	node = rb_first(&et->root);
> > > +	while (node) {
> > > +		next = rb_next(node);
> > > +		en = rb_entry(node, struct extent_node, rb_node);
> > > +
> > > +		if (free_all) {
> > > +			spin_lock(&sbi->extent_lock);
> > > +			if (!list_empty(&en->list))
> > > +				list_del_init(&en->list);
> > > +			spin_unlock(&sbi->extent_lock);
> > > +		}
> > > +
> > > +		if (free_all || list_empty(&en->list)) {
> > > +			rb_erase(node, &et->root);
> > > +			kmem_cache_free(extent_node_slab, en);
> > > +			atomic_dec(&sbi->total_ext_node);
> > > +			et->count--;
> > > +		}
> > > +		node = next;
> > > +	}
> > > +
> > > +	return count - et->count;
> > > +}
> > > +
> > > +static bool f2fs_lookup_extent_tree(struct inode *inode, pgoff_t pgofs,
> > > +							struct extent_info *ei)
> > > +{
> > > +	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> > > +	struct extent_tree *et;
> > > +	struct extent_node *en;
> > > +
> > > +	if (is_inode_flag_set(F2FS_I(inode), FI_NO_EXTENT))
> > > +		return false;
> > > +
> > > +	down_read(&sbi->extent_tree_lock);
> > > +	et = radix_tree_lookup(&sbi->extent_tree_root, inode->i_ino);
> > > +	if (!et) {
> > > +		up_read(&sbi->extent_tree_lock);
> > > +		return false;
> > > +	}
> > > +	atomic_inc(&et->refcount);
> > > +	up_read(&sbi->extent_tree_lock);
> > > +
> > > +	read_lock(&et->lock);
> > > +	en = __lookup_extent_tree(et, pgofs);
> > > +	if (en) {
> > > +		*ei = en->ei;
> > > +		spin_lock(&sbi->extent_lock);
> > > +		if (!list_empty(&en->list))
> > > +			list_move_tail(&en->list, &sbi->extent_list);
> > > +		spin_unlock(&sbi->extent_lock);
> > > +		stat_inc_read_hit(sbi->sb);
> > > +	}
> > > +	stat_inc_total_hit(sbi->sb);
> > > +	read_unlock(&et->lock);
> > > +
> > > +	atomic_dec(&et->refcount);
> > > +	return en ? true : false;
> > > +}
> > > +
> > > +static void f2fs_update_extent_tree(struct inode *inode, pgoff_t fofs,
> > > +							block_t blkaddr)
> > > +{
> > > +	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> > > +	nid_t ino = inode->i_ino;
> > > +	struct extent_tree *et;
> > > +	struct extent_node *en = NULL, *en1 = NULL, *en2 = NULL, *en3 = NULL;
> > > +	struct extent_node *den = NULL;
> > > +	struct extent_info *pei;
> > > +	struct extent_info ei;
> > > +	unsigned int endofs;
> > > +
> > > +	if (is_inode_flag_set(F2FS_I(inode), FI_NO_EXTENT))
> > > +		return;
> > > +
> > > +retry:
> > > +	down_write(&sbi->extent_tree_lock);
> > > +	et = radix_tree_lookup(&sbi->extent_tree_root, ino);
> > > +	if (!et) {
> > > +		et = kmem_cache_alloc(extent_tree_slab, GFP_ATOMIC);
> > 
> > et = f2fs_kmem_cache_alloc(.., GFP_ATOMIC);
> 
> How about modifying as below to avoid holding extent_tree_lock for long time?

Hmm. I don't think it doesn't take such the long time.
It's GFP_ATOMIC.

Moreover, for radix_tree, I prefer to use f2fs_radix_tree_insert with GFP_NOIO.
Since, we actually don't need to call kmem_cache_free.

> 
> et = kmem_cache_alloc(extent_tree_slab, GFP_ATOMIC);
> if (!et) {
> 	up_write(&sbi->extent_tree_lock);
> 	cond_resched();
> 	goto retry;
> }
> 
> > 
> > > +		if (!et) {
> > > +			up_write(&sbi->extent_tree_lock);
> > > +			goto retry;
> > > +		}
> > > +		if (radix_tree_insert(&sbi->extent_tree_root, ino, et)) {
> > > +			up_write(&sbi->extent_tree_lock);
> > > +			kmem_cache_free(extent_tree_slab, et);
> 
> cond_resched()?

I'm not sure why this should be needed.
There is rw_semaphore, so we have already a rescheduling point.

> 
> > > +			goto retry;
> > > +		}
> > > +		memset(et, 0, sizeof(struct extent_tree));
> > > +		et->ino = ino;
> > > +		et->root = RB_ROOT;
> > > +		rwlock_init(&et->lock);
> > > +		atomic_set(&et->refcount, 0);
> > > +		et->count = 0;
> > > +		sbi->total_ext_tree++;
> > > +	}
> > > +	atomic_inc(&et->refcount);
> > > +	up_write(&sbi->extent_tree_lock);
> > > +
> > > +	write_lock(&et->lock);
> > > +
> > > +	/* 1. lookup and remove exist extent info in cache */
> > 
> >                                 existing
> 
> got it.
> 
> > 
> > > +	en = __remove_extent_tree(et, fofs);
> > 
> > en = __detach_extent_node();?
> 
> OK, it seems that most codes of __remove_extent_tree and __lookup_extent_tree 
> are the same, so here I want to remove extent node like this:
> 
> 	en = __lookup_extent_tree(et, fofs);
> 	if (!en)
> 		goto update_extent;
> 
> 	dei = en->ei;
> 	__detach_extent_node(sbi, et, en);
> 
> How do you think?

Looks good to me.

> 
> > 
> > > +	if (!en)
> > > +		goto update_extent;
> > > +
> > > +	pei = &en->ei;
> > > +	/* 2. if extent can be split more, split and insert the left part */
> > > +	if (pei->len > 1) {
> > > +		/*  insert left part of split extent into cache */
> > > +		if (pei->fofs < fofs) {
> > > +			set_extent_info(&ei, pei->fofs, pei->blk,
> > > +							fofs - pei->fofs);
> > > +			en1 = __insert_extent_tree(sbi, et, &ei, NULL);
> > > +		}
> > > +
> > > +		/* insert right part of split extent into cache */
> > > +		endofs = pei->fofs + pei->len - 1;
> > > +		if (endofs > fofs) {
> > > +			set_extent_info(&ei, fofs + 1,
> > > +				fofs - pei->fofs + pei->blk, endofs - fofs);
> > > +			en2 = __insert_extent_tree(sbi, et, &ei, NULL);
> > > +		}
> > > +	}
> > > +
> > > +update_extent:
> > > +	/* 3. update extent in extent cache */
> > > +	if (blkaddr) {
> > > +		set_extent_info(&ei, fofs, blkaddr, 1);
> > > +		en3 = __insert_extent_tree(sbi, et, &ei, &den);
> > > +	}
> > > +
> > > +	/* 4. update in global extent list */
> > > +	spin_lock(&sbi->extent_lock);
> > > +	if (en && !list_empty(&en->list))
> > > +		list_del_init(&en->list);
> > > +	/*
> > > +	 * en1 and en2 split from en, they will become more and more smaller
> > > +	 * fragments after splitting several times. So if the length is smaller
> > > +	 * than F2FS_MIN_EXTENT_LEN, we will not add them into extent_list,
> > > +	 * but just waiting shrinker to free them for reclaiming when OOM.
> > > +	 */
> > 
> > Can we just remove en1 and en2 in __insert_extent_tree?
> 
> en1 and en2 is newly added, in __attach_extent_node we do not add en1 and en2 into
> lru list, so if you do not want to keep split part in lru list, let's just remove
> the below codes.

What I meant was, how about avoiding attaching en1 and en2, if they are splits whose
lens are less than F2FS_MIN_EXTENT_LEN in advance?

> 
> > 
> > > +	if (en1 && en1->ei.len >= F2FS_MIN_EXTENT_LEN)
> > > +		list_add_tail(&en1->list, &sbi->extent_list);
> > > +	if (en2 && en2->ei.len >= F2FS_MIN_EXTENT_LEN)
> > > +		list_add_tail(&en2->list, &sbi->extent_list);
> > > +	if (en3) {
> > > +		if (list_empty(&en3->list))
> > > +			list_add_tail(&en3->list, &sbi->extent_list);
> > > +		else
> > > +			list_move_tail(&en3->list, &sbi->extent_list);
> > > +	}
> > > +	if (den && !list_empty(&den->list))
> > > +		list_del_init(&den->list);
> > > +	spin_unlock(&sbi->extent_lock);
> > > +
> > > +	if (en) {
> > > +		kmem_cache_free(extent_node_slab, en);
> > > +		atomic_dec(&sbi->total_ext_node);
> > 
> > 		--> move into __detach_extent_node().
> 
> will do.
> 
> > 
> > > +	}
> > > +	if (den) {
> > > +		kmem_cache_free(extent_node_slab, den);
> > > +		atomic_dec(&sbi->total_ext_node);
> > > +	}
> > > +
> > > +	write_unlock(&et->lock);
> > > +	atomic_dec(&et->refcount);
> > > +}
> > > +
> > > +void f2fs_shrink_extent_tree(struct f2fs_sb_info *sbi, int nr_shrink)
> > > +{
> > > +	struct extent_tree *treevec[EXT_TREE_VEC_SIZE];
> > > +	struct extent_node *en, *tmp;
> > > +	unsigned long ino = F2FS_ROOT_INO(sbi);
> > > +	struct radix_tree_iter iter;
> > > +	void **slot;
> > > +	unsigned int found;
> > > +	unsigned int node_cnt = 0, tree_cnt = 0;
> > > +
> > > +	if (available_free_memory(sbi, EXTENT_CACHE))
> > > +		return;
> > > +
> > > +	spin_lock(&sbi->extent_lock);
> > > +	list_for_each_entry_safe(en, tmp, &sbi->extent_list, list) {
> > > +		if (!nr_shrink--)
> > > +			break;
> > > +		list_del_init(&en->list);
> > > +	}
> > > +	spin_unlock(&sbi->extent_lock);
> > > +
> > > +	down_read(&sbi->extent_tree_lock);
> > > +	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];
> > > +
> > > +			atomic_inc(&et->refcount);
> > > +			write_lock(&et->lock);
> > > +			node_cnt += __free_extent_tree(sbi, et, false);
> > > +			write_unlock(&et->lock);
> > > +			atomic_dec(&et->refcount);
> > > +		}
> > > +	}
> > > +	up_read(&sbi->extent_tree_lock);
> > > +
> > > +	down_write(&sbi->extent_tree_lock);
> > > +	radix_tree_for_each_slot(slot, &sbi->extent_tree_root, &iter,
> > > +							F2FS_ROOT_INO(sbi)) {
> > > +		struct extent_tree *et = (struct extent_tree *)*slot;
> > > +
> > > +		if (!atomic_read(&et->refcount) && !et->count) {
> > > +			radix_tree_delete(&sbi->extent_tree_root, et->ino);
> > > +			kmem_cache_free(extent_tree_slab, et);
> > > +			sbi->total_ext_tree--;
> > > +			tree_cnt++;
> > 
> > No use of tree_cnt.
> 
> This is used by trace function in RFC PATCH 10/10.

Well, then, it needs to add tree_cnt in Patch #10. :)

Thanks,

> 
> Thanks for your review! :)
> 
> Regards,
> Yu
> 
> > 
> > Thanks,
> 
> [snip]