extent-tree.c

/*
 * Copyright (C) 2007 Oracle.  All rights reserved.
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public
 * License v2 as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License for more details.
 *
 * You should have received a copy of the GNU General Public
 * License along with this program; if not, write to the
 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 * Boston, MA 021110-1307, USA.
 */
#include <linux/sched.h>
#include <linux/pagemap.h>
#include <linux/writeback.h>
#include <linux/blkdev.h>
#include <linux/sort.h>
#include <linux/rcupdate.h>
#include <linux/kthread.h>
#include <linux/slab.h>
#include "compat.h"
#include "hash.h"
#include "ctree.h"
#include "disk-io.h"
#include "print-tree.h"
#include "transaction.h"
#include "volumes.h"
#include "locking.h"
#include "free-space-cache.h"

/* control flags for do_chunk_alloc's force field
 * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
 * if we really need one.
 *
 * CHUNK_ALLOC_FORCE means it must try to allocate one
 *
 * CHUNK_ALLOC_LIMITED means to only try and allocate one
 * if we have very few chunks already allocated.  This is
 * used as part of the clustering code to help make sure
 * we have a good pool of storage to cluster in, without
 * filling the FS with empty chunks
 *
 */
enum {
	CHUNK_ALLOC_NO_FORCE = 0,
	CHUNK_ALLOC_FORCE = 1,
	CHUNK_ALLOC_LIMITED = 2,
};

static int update_block_group(struct btrfs_trans_handle *trans,
			      struct btrfs_root *root,
			      u64 bytenr, u64 num_bytes, int alloc);
static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
				struct btrfs_root *root,
				u64 bytenr, u64 num_bytes, u64 parent,
				u64 root_objectid, u64 owner_objectid,
				u64 owner_offset, int refs_to_drop,
				struct btrfs_delayed_extent_op *extra_op);
static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
				    struct extent_buffer *leaf,
				    struct btrfs_extent_item *ei);
static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
				      struct btrfs_root *root,
				      u64 parent, u64 root_objectid,
				      u64 flags, u64 owner, u64 offset,
				      struct btrfs_key *ins, int ref_mod);
static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
				     struct btrfs_root *root,
				     u64 parent, u64 root_objectid,
				     u64 flags, struct btrfs_disk_key *key,
				     int level, struct btrfs_key *ins);
static int do_chunk_alloc(struct btrfs_trans_handle *trans,
			  struct btrfs_root *extent_root, u64 alloc_bytes,
			  u64 flags, int force);
static int find_next_key(struct btrfs_path *path, int level,
			 struct btrfs_key *key);
static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
			    int dump_block_groups);

static noinline int
block_group_cache_done(struct btrfs_block_group_cache *cache)
{
	smp_mb();
	return cache->cached == BTRFS_CACHE_FINISHED;
}

static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
{
	return (cache->flags & bits) == bits;
}

void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
{
	atomic_inc(&cache->count);
}

void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
{
	if (atomic_dec_and_test(&cache->count)) {
		WARN_ON(cache->pinned > 0);
		WARN_ON(cache->reserved > 0);
		WARN_ON(cache->reserved_pinned > 0);
		kfree(cache);
	}
}

/*
 * this adds the block group to the fs_info rb tree for the block group
 * cache
 */
static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
				struct btrfs_block_group_cache *block_group)
{
	struct rb_node **p;
	struct rb_node *parent = NULL;
	struct btrfs_block_group_cache *cache;

	spin_lock(&info->block_group_cache_lock);
	p = &info->block_group_cache_tree.rb_node;

	while (*p) {
		parent = *p;
		cache = rb_entry(parent, struct btrfs_block_group_cache,
				 cache_node);
		if (block_group->key.objectid < cache->key.objectid) {
			p = &(*p)->rb_left;
		} else if (block_group->key.objectid > cache->key.objectid) {
			p = &(*p)->rb_right;
		} else {
			spin_unlock(&info->block_group_cache_lock);
			return -EEXIST;
		}
	}

	rb_link_node(&block_group->cache_node, parent, p);
	rb_insert_color(&block_group->cache_node,
			&info->block_group_cache_tree);
	spin_unlock(&info->block_group_cache_lock);

	return 0;
}

/*
 * This will return the block group at or after bytenr if contains is 0, else
 * it will return the block group that contains the bytenr
 */
static struct btrfs_block_group_cache *
block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
			      int contains)
{
	struct btrfs_block_group_cache *cache, *ret = NULL;
	struct rb_node *n;
	u64 end, start;

	spin_lock(&info->block_group_cache_lock);
	n = info->block_group_cache_tree.rb_node;

	while (n) {
		cache = rb_entry(n, struct btrfs_block_group_cache,
				 cache_node);
		end = cache->key.objectid + cache->key.offset - 1;
		start = cache->key.objectid;

		if (bytenr < start) {
			if (!contains && (!ret || start < ret->key.objectid))
				ret = cache;
			n = n->rb_left;
		} else if (bytenr > start) {
			if (contains && bytenr <= end) {
				ret = cache;
				break;
			}
			n = n->rb_right;
		} else {
			ret = cache;
			break;
		}
	}
	if (ret)
		btrfs_get_block_group(ret);
	spin_unlock(&info->block_group_cache_lock);

	return ret;
}

static int add_excluded_extent(struct btrfs_root *root,
			       u64 start, u64 num_bytes)
{
	u64 end = start + num_bytes - 1;
	set_extent_bits(&root->fs_info->freed_extents[0],
			start, end, EXTENT_UPTODATE, GFP_NOFS);
	set_extent_bits(&root->fs_info->freed_extents[1],
			start, end, EXTENT_UPTODATE, GFP_NOFS);
	return 0;
}

static void free_excluded_extents(struct btrfs_root *root,
				  struct btrfs_block_group_cache *cache)
{
	u64 start, end;

	start = cache->key.objectid;
	end = start + cache->key.offset - 1;

	clear_extent_bits(&root->fs_info->freed_extents[0],
			  start, end, EXTENT_UPTODATE, GFP_NOFS);
	clear_extent_bits(&root->fs_info->freed_extents[1],
			  start, end, EXTENT_UPTODATE, GFP_NOFS);
}

static int exclude_super_stripes(struct btrfs_root *root,
				 struct btrfs_block_group_cache *cache)
{
	u64 bytenr;
	u64 *logical;
	int stripe_len;
	int i, nr, ret;

	if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
		stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
		cache->bytes_super += stripe_len;
		ret = add_excluded_extent(root, cache->key.objectid,
					  stripe_len);
		BUG_ON(ret);
	}

	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
		bytenr = btrfs_sb_offset(i);
		ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
				       cache->key.objectid, bytenr,
				       0, &logical, &nr, &stripe_len);
		BUG_ON(ret);

		while (nr--) {
			cache->bytes_super += stripe_len;
			ret = add_excluded_extent(root, logical[nr],
						  stripe_len);
			BUG_ON(ret);
		}

		kfree(logical);
	}
	return 0;
}

static struct btrfs_caching_control *
get_caching_control(struct btrfs_block_group_cache *cache)
{
	struct btrfs_caching_control *ctl;

	spin_lock(&cache->lock);
	if (cache->cached != BTRFS_CACHE_STARTED) {
		spin_unlock(&cache->lock);
		return NULL;
	}

	/* We're loading it the fast way, so we don't have a caching_ctl. */
	if (!cache->caching_ctl) {
		spin_unlock(&cache->lock);
		return NULL;
	}

	ctl = cache->caching_ctl;
	atomic_inc(&ctl->count);
	spin_unlock(&cache->lock);
	return ctl;
}

static void put_caching_control(struct btrfs_caching_control *ctl)
{
	if (atomic_dec_and_test(&ctl->count))
		kfree(ctl);
}

/*
 * this is only called by cache_block_group, since we could have freed extents
 * we need to check the pinned_extents for any extents that can't be used yet
 * since their free space will be released as soon as the transaction commits.
 */
static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
			      struct btrfs_fs_info *info, u64 start, u64 end)
{
	u64 extent_start, extent_end, size, total_added = 0;
	int ret;

	while (start < end) {
		ret = find_first_extent_bit(info->pinned_extents, start,
					    &extent_start, &extent_end,
					    EXTENT_DIRTY | EXTENT_UPTODATE);
		if (ret)
			break;

		if (extent_start <= start) {
			start = extent_end + 1;
		} else if (extent_start > start && extent_start < end) {
			size = extent_start - start;
			total_added += size;
			ret = btrfs_add_free_space(block_group, start,
						   size);
			BUG_ON(ret);
			start = extent_end + 1;
		} else {
			break;
		}
	}

	if (start < end) {
		size = end - start;
		total_added += size;
		ret = btrfs_add_free_space(block_group, start, size);
		BUG_ON(ret);
	}

	return total_added;
}

static int caching_kthread(void *data)
{
	struct btrfs_block_group_cache *block_group = data;
	struct btrfs_fs_info *fs_info = block_group->fs_info;
	struct btrfs_caching_control *caching_ctl = block_group->caching_ctl;
	struct btrfs_root *extent_root = fs_info->extent_root;
	struct btrfs_path *path;
	struct extent_buffer *leaf;
	struct btrfs_key key;
	u64 total_found = 0;
	u64 last = 0;
	u32 nritems;
	int ret = 0;

	path = btrfs_alloc_path();
	if (!path)
		return -ENOMEM;

	last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);

	/*
	 * We don't want to deadlock with somebody trying to allocate a new
	 * extent for the extent root while also trying to search the extent
	 * root to add free space.  So we skip locking and search the commit
	 * root, since its read-only
	 */
	path->skip_locking = 1;
	path->search_commit_root = 1;
	path->reada = 2;

	key.objectid = last;
	key.offset = 0;
	key.type = BTRFS_EXTENT_ITEM_KEY;
again:
	mutex_lock(&caching_ctl->mutex);
	/* need to make sure the commit_root doesn't disappear */
	down_read(&fs_info->extent_commit_sem);

	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
	if (ret < 0)
		goto err;

	leaf = path->nodes[0];
	nritems = btrfs_header_nritems(leaf);

	while (1) {
		smp_mb();
		if (fs_info->closing > 1) {
			last = (u64)-1;
			break;
		}

		if (path->slots[0] < nritems) {
			btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
		} else {
			ret = find_next_key(path, 0, &key);
			if (ret)
				break;

			if (need_resched() ||
			    btrfs_next_leaf(extent_root, path)) {
				caching_ctl->progress = last;
				btrfs_release_path(extent_root, path);
				up_read(&fs_info->extent_commit_sem);
				mutex_unlock(&caching_ctl->mutex);
				cond_resched();
				goto again;
			}
			leaf = path->nodes[0];
			nritems = btrfs_header_nritems(leaf);
			continue;
		}

		if (key.objectid < block_group->key.objectid) {
			path->slots[0]++;
			continue;
		}

		if (key.objectid >= block_group->key.objectid +
		    block_group->key.offset)
			break;

		if (key.type == BTRFS_EXTENT_ITEM_KEY) {
			total_found += add_new_free_space(block_group,
							  fs_info, last,
							  key.objectid);
			last = key.objectid + key.offset;

			if (total_found > (1024 * 1024 * 2)) {
				total_found = 0;
				wake_up(&caching_ctl->wait);
			}
		}
		path->slots[0]++;
	}
	ret = 0;

	total_found += add_new_free_space(block_group, fs_info, last,
					  block_group->key.objectid +
					  block_group->key.offset);
	caching_ctl->progress = (u64)-1;

	spin_lock(&block_group->lock);
	block_group->caching_ctl = NULL;
	block_group->cached = BTRFS_CACHE_FINISHED;
	spin_unlock(&block_group->lock);

err:
	btrfs_free_path(path);
	up_read(&fs_info->extent_commit_sem);

	free_excluded_extents(extent_root, block_group);

	mutex_unlock(&caching_ctl->mutex);
	wake_up(&caching_ctl->wait);

	put_caching_control(caching_ctl);
	atomic_dec(&block_group->space_info->caching_threads);
	btrfs_put_block_group(block_group);

	return 0;
}

static int cache_block_group(struct btrfs_block_group_cache *cache,
			     struct btrfs_trans_handle *trans,
			     struct btrfs_root *root,
			     int load_cache_only)
{
	struct btrfs_fs_info *fs_info = cache->fs_info;
	struct btrfs_caching_control *caching_ctl;
	struct task_struct *tsk;
	int ret = 0;

	smp_mb();
	if (cache->cached != BTRFS_CACHE_NO)
		return 0;

	/*
	 * We can't do the read from on-disk cache during a commit since we need
	 * to have the normal tree locking.  Also if we are currently trying to
	 * allocate blocks for the tree root we can't do the fast caching since
	 * we likely hold important locks.
	 */
	if (trans && (!trans->transaction->in_commit) &&
	    (root && root != root->fs_info->tree_root)) {
		spin_lock(&cache->lock);
		if (cache->cached != BTRFS_CACHE_NO) {
			spin_unlock(&cache->lock);
			return 0;
		}
		cache->cached = BTRFS_CACHE_STARTED;
		spin_unlock(&cache->lock);

		ret = load_free_space_cache(fs_info, cache);

		spin_lock(&cache->lock);
		if (ret == 1) {
			cache->cached = BTRFS_CACHE_FINISHED;
			cache->last_byte_to_unpin = (u64)-1;
		} else {
			cache->cached = BTRFS_CACHE_NO;
		}
		spin_unlock(&cache->lock);
		if (ret == 1) {
			free_excluded_extents(fs_info->extent_root, cache);
			return 0;
		}
	}

	if (load_cache_only)
		return 0;

	caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
	BUG_ON(!caching_ctl);

	INIT_LIST_HEAD(&caching_ctl->list);
	mutex_init(&caching_ctl->mutex);
	init_waitqueue_head(&caching_ctl->wait);
	caching_ctl->block_group = cache;
	caching_ctl->progress = cache->key.objectid;
	/* one for caching kthread, one for caching block group list */
	atomic_set(&caching_ctl->count, 2);

	spin_lock(&cache->lock);
	if (cache->cached != BTRFS_CACHE_NO) {
		spin_unlock(&cache->lock);
		kfree(caching_ctl);
		return 0;
	}
	cache->caching_ctl = caching_ctl;
	cache->cached = BTRFS_CACHE_STARTED;
	spin_unlock(&cache->lock);

	down_write(&fs_info->extent_commit_sem);
	list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
	up_write(&fs_info->extent_commit_sem);

	atomic_inc(&cache->space_info->caching_threads);
	btrfs_get_block_group(cache);

	tsk = kthread_run(caching_kthread, cache, "btrfs-cache-%llu\n",
			  cache->key.objectid);
	if (IS_ERR(tsk)) {
		ret = PTR_ERR(tsk);
		printk(KERN_ERR "error running thread %d\n", ret);
		BUG();
	}

	return ret;
}

/*
 * return the block group that starts at or after bytenr
 */
static struct btrfs_block_group_cache *
btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
{
	struct btrfs_block_group_cache *cache;

	cache = block_group_cache_tree_search(info, bytenr, 0);

	return cache;
}

/*
 * return the block group that contains the given bytenr
 */
struct btrfs_block_group_cache *btrfs_lookup_block_group(
						 struct btrfs_fs_info *info,
						 u64 bytenr)
{
	struct btrfs_block_group_cache *cache;

	cache = block_group_cache_tree_search(info, bytenr, 1);

	return cache;
}

static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
						  u64 flags)
{
	struct list_head *head = &info->space_info;
	struct btrfs_space_info *found;

	flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
		 BTRFS_BLOCK_GROUP_METADATA;

	rcu_read_lock();
	list_for_each_entry_rcu(found, head, list) {
		if (found->flags & flags) {
			rcu_read_unlock();
			return found;
		}
	}
	rcu_read_unlock();
	return NULL;
}

/*
 * after adding space to the filesystem, we need to clear the full flags
 * on all the space infos.
 */
void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
{
	struct list_head *head = &info->space_info;
	struct btrfs_space_info *found;

	rcu_read_lock();
	list_for_each_entry_rcu(found, head, list)
		found->full = 0;
	rcu_read_unlock();
}

static u64 div_factor(u64 num, int factor)
{
	if (factor == 10)
		return num;
	num *= factor;
	do_div(num, 10);
	return num;
}

static u64 div_factor_fine(u64 num, int factor)
{
	if (factor == 100)
		return num;
	num *= factor;
	do_div(num, 100);
	return num;
}

u64 btrfs_find_block_group(struct btrfs_root *root,
			   u64 search_start, u64 search_hint, int owner)
{
	struct btrfs_block_group_cache *cache;
	u64 used;
	u64 last = max(search_hint, search_start);
	u64 group_start = 0;
	int full_search = 0;
	int factor = 9;
	int wrapped = 0;
again:
	while (1) {
		cache = btrfs_lookup_first_block_group(root->fs_info, last);
		if (!cache)
			break;

		spin_lock(&cache->lock);
		last = cache->key.objectid + cache->key.offset;
		used = btrfs_block_group_used(&cache->item);

		if ((full_search || !cache->ro) &&
		    block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
			if (used + cache->pinned + cache->reserved <
			    div_factor(cache->key.offset, factor)) {
				group_start = cache->key.objectid;
				spin_unlock(&cache->lock);
				btrfs_put_block_group(cache);
				goto found;
			}
		}
		spin_unlock(&cache->lock);
		btrfs_put_block_group(cache);
		cond_resched();
	}
	if (!wrapped) {
		last = search_start;
		wrapped = 1;
		goto again;
	}
	if (!full_search && factor < 10) {
		last = search_start;
		full_search = 1;
		factor = 10;
		goto again;
	}
found:
	return group_start;
}

/* simple helper to search for an existing extent at a given offset */
int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
{
	int ret;
	struct btrfs_key key;
	struct btrfs_path *path;

	path = btrfs_alloc_path();
	BUG_ON(!path);
	key.objectid = start;
	key.offset = len;
	btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
	ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
				0, 0);
	btrfs_free_path(path);
	return ret;
}

/*
 * helper function to lookup reference count and flags of extent.
 *
 * the head node for delayed ref is used to store the sum of all the
 * reference count modifications queued up in the rbtree. the head
 * node may also store the extent flags to set. This way you can check
 * to see what the reference count and extent flags would be if all of
 * the delayed refs are not processed.
 */
int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
			     struct btrfs_root *root, u64 bytenr,
			     u64 num_bytes, u64 *refs, u64 *flags)
{
	struct btrfs_delayed_ref_head *head;
	struct btrfs_delayed_ref_root *delayed_refs;
	struct btrfs_path *path;
	struct btrfs_extent_item *ei;
	struct extent_buffer *leaf;
	struct btrfs_key key;
	u32 item_size;
	u64 num_refs;
	u64 extent_flags;
	int ret;

	path = btrfs_alloc_path();
	if (!path)
		return -ENOMEM;

	key.objectid = bytenr;
	key.type = BTRFS_EXTENT_ITEM_KEY;
	key.offset = num_bytes;
	if (!trans) {
		path->skip_locking = 1;
		path->search_commit_root = 1;
	}
again:
	ret = btrfs_search_slot(trans, root->fs_info->extent_root,
				&key, path, 0, 0);
	if (ret < 0)
		goto out_free;

	if (ret == 0) {
		leaf = path->nodes[0];
		item_size = btrfs_item_size_nr(leaf, path->slots[0]);
		if (item_size >= sizeof(*ei)) {
			ei = btrfs_item_ptr(leaf, path->slots[0],
					    struct btrfs_extent_item);
			num_refs = btrfs_extent_refs(leaf, ei);
			extent_flags = btrfs_extent_flags(leaf, ei);
		} else {
#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
			struct btrfs_extent_item_v0 *ei0;
			BUG_ON(item_size != sizeof(*ei0));
			ei0 = btrfs_item_ptr(leaf, path->slots[0],
					     struct btrfs_extent_item_v0);
			num_refs = btrfs_extent_refs_v0(leaf, ei0);
			/* FIXME: this isn't correct for data */
			extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
#else
			BUG();
#endif
		}
		BUG_ON(num_refs == 0);
	} else {
		num_refs = 0;
		extent_flags = 0;
		ret = 0;
	}

	if (!trans)
		goto out;

	delayed_refs = &trans->transaction->delayed_refs;
	spin_lock(&delayed_refs->lock);
	head = btrfs_find_delayed_ref_head(trans, bytenr);
	if (head) {
		if (!mutex_trylock(&head->mutex)) {
			atomic_inc(&head->node.refs);
			spin_unlock(&delayed_refs->lock);

			btrfs_release_path(root->fs_info->extent_root, path);

			mutex_lock(&head->mutex);
			mutex_unlock(&head->mutex);
			btrfs_put_delayed_ref(&head->node);
			goto again;
		}
		if (head->extent_op && head->extent_op->update_flags)
			extent_flags |= head->extent_op->flags_to_set;
		else
			BUG_ON(num_refs == 0);

		num_refs += head->node.ref_mod;
		mutex_unlock(&head->mutex);
	}
	spin_unlock(&delayed_refs->lock);
out:
	WARN_ON(num_refs == 0);
	if (refs)
		*refs = num_refs;
	if (flags)
		*flags = extent_flags;
out_free:
	btrfs_free_path(path);
	return ret;
}

/*
 * Back reference rules.  Back refs have three main goals:
 *
 * 1) differentiate between all holders of references to an extent so that
 *    when a reference is dropped we can make sure it was a valid reference
 *    before freeing the extent.
 *
 * 2) Provide enough information to quickly find the holders of an extent
 *    if we notice a given block is corrupted or bad.
 *
 * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 *    maintenance.  This is actually the same as #2, but with a slightly
 *    different use case.
 *
 * There are two kinds of back refs. The implicit back refs is optimized
 * for pointers in non-shared tree blocks. For a given pointer in a block,
 * back refs of this kind provide information about the block's owner tree
 * and the pointer's key. These information allow us to find the block by
 * b-tree searching. The full back refs is for pointers in tree blocks not
 * referenced by their owner trees. The location of tree block is recorded
 * in the back refs. Actually the full back refs is generic, and can be
 * used in all cases the implicit back refs is used. The major shortcoming
 * of the full back refs is its overhead. Every time a tree block gets
 * COWed, we have to update back refs entry for all pointers in it.
 *
 * For a newly allocated tree block, we use implicit back refs for
 * pointers in it. This means most tree related operations only involve
 * implicit back refs. For a tree block created in old transaction, the
 * only way to drop a reference to it is COW it. So we can detect the
 * event that tree block loses its owner tree's reference and do the
 * back refs conversion.
 *
 * When a tree block is COW'd through a tree, there are four cases:
 *
 * The reference count of the block is one and the tree is the block's
 * owner tree. Nothing to do in this case.
 *
 * The reference count of the block is one and the tree is not the
 * block's owner tree. In this case, full back refs is used for pointers
 * in the block. Remove these full back refs, add implicit back refs for
 * every pointers in the new block.
 *
 * The reference count of the block is greater than one and the tree is
 * the block's owner tree. In this case, implicit back refs is used for
 * pointers in the block. Add full back refs for every pointers in the
 * block, increase lower level extents' reference counts. The original
 * implicit back refs are entailed to the new block.
 *
 * The reference count of the block is greater than one and the tree is
 * not the block's owner tree. Add implicit back refs for every pointer in
 * the new block, increase lower level extents' reference count.
 *
 * Back Reference Key composing:
 *
 * The key objectid corresponds to the first byte in the extent,
 * The key type is used to differentiate between types of back refs.
 * There are different meanings of the key offset for different types
 * of back refs.
 *
 * File extents can be referenced by:
 *
 * - multiple snapshots, subvolumes, or different generations in one subvol
 * - different files inside a single subvolume
 * - different offsets inside a file (bookend extents in file.c)
 *
 * The extent ref structure for the implicit back refs has fields for:
 *
 * - Objectid of the subvolume root
 * - objectid of the file holding the reference
 * - original offset in the file
 * - how many bookend extents
 *
 * The key offset for the implicit back refs is hash of the first
 * three fields.
 *
 * The extent ref structure for the full back refs has field for:
 *
 * - number of pointers in the tree leaf
 *
 * The key offset for the implicit back refs is the first byte of
 * the tree leaf
 *
 * When a file extent is allocated, The implicit back refs is used.
 * the fields are filled in:
 *
 *     (root_key.objectid, inode objectid, offset in file, 1)
 *
 * When a file extent is removed file truncation, we find the
 * corresponding implicit back refs and check the following fields:
 *
 *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 *
 * Btree extents can be referenced by:
 *
 * - Different subvolumes
 *
 * Both the implicit back refs and the full back refs for tree blocks
 * only consist of key. The key offset for the implicit back refs is
 * objectid of block's owner tree. The key offset for the full back refs
 * is the first byte of parent block.
 *
 * When implicit back refs is used, information about the lowest key and
 * level of the tree block are required. These information are stored in
 * tree block info structure.
 */

#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
				  struct btrfs_root *root,
				  struct btrfs_path *path,
				  u64 owner, u32 extra_size)
{
	struct btrfs_extent_item *item;
	struct btrfs_extent_item_v0 *ei0;
	struct btrfs_extent_ref_v0 *ref0;
	struct btrfs_tree_block_info *bi;
	struct extent_buffer *leaf;
	struct btrfs_key key;
	struct btrfs_key found_key;
	u32 new_size = sizeof(*item);
	u64 refs;
	int ret;

	leaf = path->nodes[0];
	BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));

	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
	ei0 = btrfs_item_ptr(leaf, path->slots[0],
			     struct btrfs_extent_item_v0);
	refs = btrfs_extent_refs_v0(leaf, ei0);

	if (owner == (u64)-1) {
		while (1) {
			if (path->slots[0] >= btrfs_header_nritems(leaf)) {
				ret = btrfs_next_leaf(root, path);
				if (ret < 0)
					return ret;
				BUG_ON(ret > 0);
				leaf = path->nodes[0];
			}
			btrfs_item_key_to_cpu(leaf, &found_key,
					      path->slots[0]);
			BUG_ON(key.objectid != found_key.objectid);
			if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
				path->slots[0]++;
				continue;
			}
			ref0 = btrfs_item_ptr(leaf, path->slots[0],
					      struct btrfs_extent_ref_v0);
			owner = btrfs_ref_objectid_v0(leaf, ref0);
			break;
		}
	}
	btrfs_release_path(root, path);

	if (owner < BTRFS_FIRST_FREE_OBJECTID)
		new_size += sizeof(*bi);

	new_size -= sizeof(*ei0);
	ret = btrfs_search_slot(trans, root, &key, path,
				new_size + extra_size, 1);
	if (ret < 0)
		return ret;
	BUG_ON(ret);

	ret = btrfs_extend_item(trans, root, path, new_size);
	BUG_ON(ret);

	leaf = path->nodes[0];
	item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
	btrfs_set_extent_refs(leaf, item, refs);
	/* FIXME: get real generation */
	btrfs_set_extent_generation(leaf, item, 0);
	if (owner < BTRFS_FIRST_FREE_OBJECTID) {
		btrfs_set_extent_flags(leaf, item,
				       BTRFS_EXTENT_FLAG_TREE_BLOCK |
				       BTRFS_BLOCK_FLAG_FULL_BACKREF);
		bi = (struct btrfs_tree_block_info *)(item + 1);
		/* FIXME: get first key of the block */
		memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
		btrfs_set_tree_block_level(leaf, bi, (int)owner);
	} else {
		btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
	}
	btrfs_mark_buffer_dirty(leaf);
	return 0;
}
#endif

static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
{
	u32 high_crc = ~(u32)0;
	u32 low_crc = ~(u32)0;
	__le64 lenum;

	lenum = cpu_to_le64(root_objectid);
	high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
	lenum = cpu_to_le64(owner);
	low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
	lenum = cpu_to_le64(offset);
	low_crc = crc32c(low_crc, &lenum, sizeof(lenum));

	return ((u64)high_crc << 31) ^ (u64)low_crc;
}

static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
				     struct btrfs_extent_data_ref *ref)
{
	return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
				    btrfs_extent_data_ref_objectid(leaf, ref),
				    btrfs_extent_data_ref_offset(leaf, ref));
}

static int match_extent_data_ref(struct extent_buffer *leaf,