Re: [PATCH]btrfs: speed up extent_io tree search

[Shaohua Li <shaohua.li@intel.com>]

On Tue, Apr 20, 2010 at 10:39:01PM +0800, Josef Bacik wrote:
> On Tue, Apr 20, 2010 at 05:21:58PM +0800, Shaohua Li wrote:
> > searching extent_io_tree is frequently used and tooks a lot of cpu time.
> > We could cache last found extent_state to skip some full search. In my
> > test, the hit rate is from 30% to 70% depending on different workload,
> > which can speed up the search.
> > 
> > Signed-off-by: Shaohua Li <shaohua.li@intel.com>
> > 
> > diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
> > index d2d0368..645f00c 100644
> > --- a/fs/btrfs/extent_io.c
> > +++ b/fs/btrfs/extent_io.c
> > @@ -110,6 +110,7 @@ void extent_io_tree_init(struct extent_io_tree *tree,
> >  	spin_lock_init(&tree->lock);
> >  	spin_lock_init(&tree->buffer_lock);
> >  	tree->mapping = mapping;
> > +	tree->cached_state = NULL;
> >  }
> >  
> >  static struct extent_state *alloc_extent_state(gfp_t mask)
> > @@ -135,6 +136,22 @@ static struct extent_state *alloc_extent_state(gfp_t mask)
> >  	return state;
> >  }
> >  
> > +static void remove_cached_extent(struct extent_io_tree *tree,
> > +	struct extent_state *state)
> > +{
> > +	if (!tree->cached_state)
> > +		return;
> > +	if (tree->cached_state == state)
> > +		tree->cached_state = NULL;
> > +}
> > +
> > +static void merge_cached_extent(struct extent_io_tree *tree,
> > +	struct extent_state *first, struct extent_state *last)
> > +{
> > +	if (tree->cached_state == first || tree->cached_state == last)
> > +		tree->cached_state = first;
> > +}
> > +
> >  static void free_extent_state(struct extent_state *state)
> >  {
> >  	if (!state)
> > @@ -188,6 +205,12 @@ static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
> >  	struct rb_node *orig_prev = NULL;
> >  	struct tree_entry *entry;
> >  	struct tree_entry *prev_entry = NULL;
> > +	struct tree_entry *cached_entry =
> > +				(struct tree_entry *)tree->cached_state;
> > +
> > +	if (likely(cached_entry && offset >= cached_entry->start &&
> > +		offset <= cached_entry->end))
> > +		return &cached_entry->rb_node;
> >  
> >  	while (n) {
> >  		entry = rb_entry(n, struct tree_entry, rb_node);
> > @@ -198,8 +221,10 @@ static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
> >  			n = n->rb_left;
> >  		else if (offset > entry->end)
> >  			n = n->rb_right;
> > -		else
> > +		else {
> > +			tree->cached_state = (struct extent_state *)entry;
> >  			return n;
> > +		}
> >  	}
> >  
> >  	if (prev_ret) {
> > @@ -313,6 +338,7 @@ static int merge_state(struct extent_io_tree *tree,
> >  			merge_cb(tree, state, other);
> >  			state->start = other->start;
> >  			other->tree = NULL;
> > +			merge_cached_extent(tree, state, other);
> >  			rb_erase(&other->rb_node, &tree->state);
> >  			free_extent_state(other);
> >  		}
> > @@ -325,6 +351,7 @@ static int merge_state(struct extent_io_tree *tree,
> >  			merge_cb(tree, state, other);
> >  			other->start = state->start;
> >  			state->tree = NULL;
> > +			merge_cached_extent(tree, other, state);
> >  			rb_erase(&state->rb_node, &tree->state);
> >  			free_extent_state(state);
> >  			state = NULL;
> > @@ -473,6 +500,7 @@ static int clear_state_bit(struct extent_io_tree *tree,
> >  		wake_up(&state->wq);
> >  	if (delete || state->state == 0) {
> >  		if (state->tree) {
> > +			remove_cached_extent(tree, state);
> >  			clear_state_cb(tree, state, state->state);
> >  			rb_erase(&state->rb_node, &tree->state);
> >  			state->tree = NULL;
> > diff --git a/fs/btrfs/extent_io.h b/fs/btrfs/extent_io.h
> > index bbab481..e60b367 100644
> > --- a/fs/btrfs/extent_io.h
> > +++ b/fs/btrfs/extent_io.h
> > @@ -89,6 +89,7 @@ struct extent_io_tree {
> >  	spinlock_t lock;
> >  	spinlock_t buffer_lock;
> >  	struct extent_io_ops *ops;
> > +	struct extent_state *cached_state;
> >  };
> >  
> >  struct extent_state {
> 
> Sorry I saw this earlier but then forgot about it.  So instead of doing a
> per-tree thing, which will end up with misses if somebody else tries to search
> the tree for a different offset, you will want to do something like this
> 
> http://git.kernel.org/?p=linux/kernel/git/mason/btrfs-unstable.git;a=commit;h=2ac55d41b5d6bf49e76bc85db5431240617e2f8f
> 
> So that way _anybody_ who does a search will have a cached state, and so all
> subsequent searches won't be needed, instead of only working for the first guy
> who gets their state cached.  Thanks,
Hmm, the patch you pointed out is already in upstream but I still saw the search
takes a lot of CPU.

Thanks,
Shaohua