Btrfs: calc file extent num_bytes correctly in file clone
[linux-2.6.git] / fs / btrfs / free-space-cache.c
index d1e5f0e..41ac927 100644 (file)
  * Boston, MA 021110-1307, USA.
  */
 
+#include <linux/pagemap.h>
 #include <linux/sched.h>
+#include <linux/slab.h>
+#include <linux/math64.h>
 #include "ctree.h"
+#include "free-space-cache.h"
+#include "transaction.h"
+#include "disk-io.h"
+#include "extent_io.h"
+#include "inode-map.h"
 
-static int tree_insert_offset(struct rb_root *root, u64 offset,
-                             struct rb_node *node)
+#define BITS_PER_BITMAP                (PAGE_CACHE_SIZE * 8)
+#define MAX_CACHE_BYTES_PER_GIG        (32 * 1024)
+
+static int link_free_space(struct btrfs_free_space_ctl *ctl,
+                          struct btrfs_free_space *info);
+
+static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
+                                              struct btrfs_path *path,
+                                              u64 offset)
 {
-       struct rb_node **p = &root->rb_node;
-       struct rb_node *parent = NULL;
-       struct btrfs_free_space *info;
+       struct btrfs_key key;
+       struct btrfs_key location;
+       struct btrfs_disk_key disk_key;
+       struct btrfs_free_space_header *header;
+       struct extent_buffer *leaf;
+       struct inode *inode = NULL;
+       int ret;
 
-       while (*p) {
-               parent = *p;
-               info = rb_entry(parent, struct btrfs_free_space, offset_index);
+       key.objectid = BTRFS_FREE_SPACE_OBJECTID;
+       key.offset = offset;
+       key.type = 0;
 
-               if (offset < info->offset)
-                       p = &(*p)->rb_left;
-               else if (offset > info->offset)
-                       p = &(*p)->rb_right;
-               else
-                       return -EEXIST;
+       ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
+       if (ret < 0)
+               return ERR_PTR(ret);
+       if (ret > 0) {
+               btrfs_release_path(path);
+               return ERR_PTR(-ENOENT);
        }
 
-       rb_link_node(node, parent, p);
-       rb_insert_color(node, root);
+       leaf = path->nodes[0];
+       header = btrfs_item_ptr(leaf, path->slots[0],
+                               struct btrfs_free_space_header);
+       btrfs_free_space_key(leaf, header, &disk_key);
+       btrfs_disk_key_to_cpu(&location, &disk_key);
+       btrfs_release_path(path);
+
+       inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
+       if (!inode)
+               return ERR_PTR(-ENOENT);
+       if (IS_ERR(inode))
+               return inode;
+       if (is_bad_inode(inode)) {
+               iput(inode);
+               return ERR_PTR(-ENOENT);
+       }
+
+       inode->i_mapping->flags &= ~__GFP_FS;
+
+       return inode;
+}
+
+struct inode *lookup_free_space_inode(struct btrfs_root *root,
+                                     struct btrfs_block_group_cache
+                                     *block_group, struct btrfs_path *path)
+{
+       struct inode *inode = NULL;
+
+       spin_lock(&block_group->lock);
+       if (block_group->inode)
+               inode = igrab(block_group->inode);
+       spin_unlock(&block_group->lock);
+       if (inode)
+               return inode;
+
+       inode = __lookup_free_space_inode(root, path,
+                                         block_group->key.objectid);
+       if (IS_ERR(inode))
+               return inode;
+
+       spin_lock(&block_group->lock);
+       if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM) {
+               printk(KERN_INFO "Old style space inode found, converting.\n");
+               BTRFS_I(inode)->flags &= ~BTRFS_INODE_NODATASUM;
+               block_group->disk_cache_state = BTRFS_DC_CLEAR;
+       }
+
+       if (!btrfs_fs_closing(root->fs_info)) {
+               block_group->inode = igrab(inode);
+               block_group->iref = 1;
+       }
+       spin_unlock(&block_group->lock);
+
+       return inode;
+}
+
+int __create_free_space_inode(struct btrfs_root *root,
+                             struct btrfs_trans_handle *trans,
+                             struct btrfs_path *path, u64 ino, u64 offset)
+{
+       struct btrfs_key key;
+       struct btrfs_disk_key disk_key;
+       struct btrfs_free_space_header *header;
+       struct btrfs_inode_item *inode_item;
+       struct extent_buffer *leaf;
+       int ret;
+
+       ret = btrfs_insert_empty_inode(trans, root, path, ino);
+       if (ret)
+               return ret;
+
+       leaf = path->nodes[0];
+       inode_item = btrfs_item_ptr(leaf, path->slots[0],
+                                   struct btrfs_inode_item);
+       btrfs_item_key(leaf, &disk_key, path->slots[0]);
+       memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
+                            sizeof(*inode_item));
+       btrfs_set_inode_generation(leaf, inode_item, trans->transid);
+       btrfs_set_inode_size(leaf, inode_item, 0);
+       btrfs_set_inode_nbytes(leaf, inode_item, 0);
+       btrfs_set_inode_uid(leaf, inode_item, 0);
+       btrfs_set_inode_gid(leaf, inode_item, 0);
+       btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
+       btrfs_set_inode_flags(leaf, inode_item, BTRFS_INODE_NOCOMPRESS |
+                             BTRFS_INODE_PREALLOC);
+       btrfs_set_inode_nlink(leaf, inode_item, 1);
+       btrfs_set_inode_transid(leaf, inode_item, trans->transid);
+       btrfs_set_inode_block_group(leaf, inode_item, offset);
+       btrfs_mark_buffer_dirty(leaf);
+       btrfs_release_path(path);
+
+       key.objectid = BTRFS_FREE_SPACE_OBJECTID;
+       key.offset = offset;
+       key.type = 0;
+
+       ret = btrfs_insert_empty_item(trans, root, path, &key,
+                                     sizeof(struct btrfs_free_space_header));
+       if (ret < 0) {
+               btrfs_release_path(path);
+               return ret;
+       }
+       leaf = path->nodes[0];
+       header = btrfs_item_ptr(leaf, path->slots[0],
+                               struct btrfs_free_space_header);
+       memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
+       btrfs_set_free_space_key(leaf, header, &disk_key);
+       btrfs_mark_buffer_dirty(leaf);
+       btrfs_release_path(path);
+
+       return 0;
+}
+
+int create_free_space_inode(struct btrfs_root *root,
+                           struct btrfs_trans_handle *trans,
+                           struct btrfs_block_group_cache *block_group,
+                           struct btrfs_path *path)
+{
+       int ret;
+       u64 ino;
+
+       ret = btrfs_find_free_objectid(root, &ino);
+       if (ret < 0)
+               return ret;
+
+       return __create_free_space_inode(root, trans, path, ino,
+                                        block_group->key.objectid);
+}
+
+int btrfs_truncate_free_space_cache(struct btrfs_root *root,
+                                   struct btrfs_trans_handle *trans,
+                                   struct btrfs_path *path,
+                                   struct inode *inode)
+{
+       struct btrfs_block_rsv *rsv;
+       loff_t oldsize;
+       int ret = 0;
+
+       rsv = trans->block_rsv;
+       trans->block_rsv = root->orphan_block_rsv;
+       ret = btrfs_block_rsv_check(trans, root,
+                                   root->orphan_block_rsv,
+                                   0, 5);
+       if (ret)
+               return ret;
+
+       oldsize = i_size_read(inode);
+       btrfs_i_size_write(inode, 0);
+       truncate_pagecache(inode, oldsize, 0);
+
+       /*
+        * We don't need an orphan item because truncating the free space cache
+        * will never be split across transactions.
+        */
+       ret = btrfs_truncate_inode_items(trans, root, inode,
+                                        0, BTRFS_EXTENT_DATA_KEY);
+
+       trans->block_rsv = rsv;
+       if (ret) {
+               WARN_ON(1);
+               return ret;
+       }
+
+       ret = btrfs_update_inode(trans, root, inode);
+       return ret;
+}
+
+static int readahead_cache(struct inode *inode)
+{
+       struct file_ra_state *ra;
+       unsigned long last_index;
+
+       ra = kzalloc(sizeof(*ra), GFP_NOFS);
+       if (!ra)
+               return -ENOMEM;
+
+       file_ra_state_init(ra, inode->i_mapping);
+       last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
+
+       page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
+
+       kfree(ra);
 
        return 0;
 }
 
-static int tree_insert_bytes(struct rb_root *root, u64 bytes,
-                            struct rb_node *node)
+int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
+                           struct btrfs_free_space_ctl *ctl,
+                           struct btrfs_path *path, u64 offset)
+{
+       struct btrfs_free_space_header *header;
+       struct extent_buffer *leaf;
+       struct page *page;
+       struct btrfs_key key;
+       struct list_head bitmaps;
+       u64 num_entries;
+       u64 num_bitmaps;
+       u64 generation;
+       pgoff_t index = 0;
+       int ret = 0;
+
+       INIT_LIST_HEAD(&bitmaps);
+
+       /* Nothing in the space cache, goodbye */
+       if (!i_size_read(inode))
+               goto out;
+
+       key.objectid = BTRFS_FREE_SPACE_OBJECTID;
+       key.offset = offset;
+       key.type = 0;
+
+       ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
+       if (ret < 0)
+               goto out;
+       else if (ret > 0) {
+               btrfs_release_path(path);
+               ret = 0;
+               goto out;
+       }
+
+       ret = -1;
+
+       leaf = path->nodes[0];
+       header = btrfs_item_ptr(leaf, path->slots[0],
+                               struct btrfs_free_space_header);
+       num_entries = btrfs_free_space_entries(leaf, header);
+       num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
+       generation = btrfs_free_space_generation(leaf, header);
+       btrfs_release_path(path);
+
+       if (BTRFS_I(inode)->generation != generation) {
+               printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
+                      " not match free space cache generation (%llu)\n",
+                      (unsigned long long)BTRFS_I(inode)->generation,
+                      (unsigned long long)generation);
+               goto out;
+       }
+
+       if (!num_entries)
+               goto out;
+
+       ret = readahead_cache(inode);
+       if (ret)
+               goto out;
+
+       while (1) {
+               struct btrfs_free_space_entry *entry;
+               struct btrfs_free_space *e;
+               void *addr;
+               unsigned long offset = 0;
+               int need_loop = 0;
+
+               if (!num_entries && !num_bitmaps)
+                       break;
+
+               page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
+               if (!page)
+                       goto free_cache;
+
+               if (!PageUptodate(page)) {
+                       btrfs_readpage(NULL, page);
+                       lock_page(page);
+                       if (!PageUptodate(page)) {
+                               unlock_page(page);
+                               page_cache_release(page);
+                               printk(KERN_ERR "btrfs: error reading free "
+                                      "space cache\n");
+                               goto free_cache;
+                       }
+               }
+               addr = kmap(page);
+
+               if (index == 0) {
+                       u64 *gen;
+
+                       /*
+                        * We put a bogus crc in the front of the first page in
+                        * case old kernels try to mount a fs with the new
+                        * format to make sure they discard the cache.
+                        */
+                       addr += sizeof(u64);
+                       offset += sizeof(u64);
+
+                       gen = addr;
+                       if (*gen != BTRFS_I(inode)->generation) {
+                               printk(KERN_ERR "btrfs: space cache generation"
+                                      " (%llu) does not match inode (%llu)\n",
+                                      (unsigned long long)*gen,
+                                      (unsigned long long)
+                                      BTRFS_I(inode)->generation);
+                               kunmap(page);
+                               unlock_page(page);
+                               page_cache_release(page);
+                               goto free_cache;
+                       }
+                       addr += sizeof(u64);
+                       offset += sizeof(u64);
+               }
+               entry = addr;
+
+               while (1) {
+                       if (!num_entries)
+                               break;
+
+                       need_loop = 1;
+                       e = kmem_cache_zalloc(btrfs_free_space_cachep,
+                                             GFP_NOFS);
+                       if (!e) {
+                               kunmap(page);
+                               unlock_page(page);
+                               page_cache_release(page);
+                               goto free_cache;
+                       }
+
+                       e->offset = le64_to_cpu(entry->offset);
+                       e->bytes = le64_to_cpu(entry->bytes);
+                       if (!e->bytes) {
+                               kunmap(page);
+                               kmem_cache_free(btrfs_free_space_cachep, e);
+                               unlock_page(page);
+                               page_cache_release(page);
+                               goto free_cache;
+                       }
+
+                       if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
+                               spin_lock(&ctl->tree_lock);
+                               ret = link_free_space(ctl, e);
+                               spin_unlock(&ctl->tree_lock);
+                               if (ret) {
+                                       printk(KERN_ERR "Duplicate entries in "
+                                              "free space cache, dumping\n");
+                                       kunmap(page);
+                                       unlock_page(page);
+                                       page_cache_release(page);
+                                       goto free_cache;
+                               }
+                       } else {
+                               e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
+                               if (!e->bitmap) {
+                                       kunmap(page);
+                                       kmem_cache_free(
+                                               btrfs_free_space_cachep, e);
+                                       unlock_page(page);
+                                       page_cache_release(page);
+                                       goto free_cache;
+                               }
+                               spin_lock(&ctl->tree_lock);
+                               ret = link_free_space(ctl, e);
+                               ctl->total_bitmaps++;
+                               ctl->op->recalc_thresholds(ctl);
+                               spin_unlock(&ctl->tree_lock);
+                               if (ret) {
+                                       printk(KERN_ERR "Duplicate entries in "
+                                              "free space cache, dumping\n");
+                                       kunmap(page);
+                                       unlock_page(page);
+                                       page_cache_release(page);
+                                       goto free_cache;
+                               }
+                               list_add_tail(&e->list, &bitmaps);
+                       }
+
+                       num_entries--;
+                       offset += sizeof(struct btrfs_free_space_entry);
+                       if (offset + sizeof(struct btrfs_free_space_entry) >=
+                           PAGE_CACHE_SIZE)
+                               break;
+                       entry++;
+               }
+
+               /*
+                * We read an entry out of this page, we need to move on to the
+                * next page.
+                */
+               if (need_loop) {
+                       kunmap(page);
+                       goto next;
+               }
+
+               /*
+                * We add the bitmaps at the end of the entries in order that
+                * the bitmap entries are added to the cache.
+                */
+               e = list_entry(bitmaps.next, struct btrfs_free_space, list);
+               list_del_init(&e->list);
+               memcpy(e->bitmap, addr, PAGE_CACHE_SIZE);
+               kunmap(page);
+               num_bitmaps--;
+next:
+               unlock_page(page);
+               page_cache_release(page);
+               index++;
+       }
+
+       ret = 1;
+out:
+       return ret;
+free_cache:
+       __btrfs_remove_free_space_cache(ctl);
+       goto out;
+}
+
+int load_free_space_cache(struct btrfs_fs_info *fs_info,
+                         struct btrfs_block_group_cache *block_group)
+{
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_root *root = fs_info->tree_root;
+       struct inode *inode;
+       struct btrfs_path *path;
+       int ret;
+       bool matched;
+       u64 used = btrfs_block_group_used(&block_group->item);
+
+       /*
+        * If we're unmounting then just return, since this does a search on the
+        * normal root and not the commit root and we could deadlock.
+        */
+       if (btrfs_fs_closing(fs_info))
+               return 0;
+
+       /*
+        * If this block group has been marked to be cleared for one reason or
+        * another then we can't trust the on disk cache, so just return.
+        */
+       spin_lock(&block_group->lock);
+       if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
+               spin_unlock(&block_group->lock);
+               return 0;
+       }
+       spin_unlock(&block_group->lock);
+
+       path = btrfs_alloc_path();
+       if (!path)
+               return 0;
+
+       inode = lookup_free_space_inode(root, block_group, path);
+       if (IS_ERR(inode)) {
+               btrfs_free_path(path);
+               return 0;
+       }
+
+       ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
+                                     path, block_group->key.objectid);
+       btrfs_free_path(path);
+       if (ret <= 0)
+               goto out;
+
+       spin_lock(&ctl->tree_lock);
+       matched = (ctl->free_space == (block_group->key.offset - used -
+                                      block_group->bytes_super));
+       spin_unlock(&ctl->tree_lock);
+
+       if (!matched) {
+               __btrfs_remove_free_space_cache(ctl);
+               printk(KERN_ERR "block group %llu has an wrong amount of free "
+                      "space\n", block_group->key.objectid);
+               ret = -1;
+       }
+out:
+       if (ret < 0) {
+               /* This cache is bogus, make sure it gets cleared */
+               spin_lock(&block_group->lock);
+               block_group->disk_cache_state = BTRFS_DC_CLEAR;
+               spin_unlock(&block_group->lock);
+               ret = 0;
+
+               printk(KERN_ERR "btrfs: failed to load free space cache "
+                      "for block group %llu\n", block_group->key.objectid);
+       }
+
+       iput(inode);
+       return ret;
+}
+
+int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
+                           struct btrfs_free_space_ctl *ctl,
+                           struct btrfs_block_group_cache *block_group,
+                           struct btrfs_trans_handle *trans,
+                           struct btrfs_path *path, u64 offset)
+{
+       struct btrfs_free_space_header *header;
+       struct extent_buffer *leaf;
+       struct rb_node *node;
+       struct list_head *pos, *n;
+       struct page **pages;
+       struct page *page;
+       struct extent_state *cached_state = NULL;
+       struct btrfs_free_cluster *cluster = NULL;
+       struct extent_io_tree *unpin = NULL;
+       struct list_head bitmap_list;
+       struct btrfs_key key;
+       u64 start, end, len;
+       u64 bytes = 0;
+       u32 crc = ~(u32)0;
+       int index = 0, num_pages = 0;
+       int entries = 0;
+       int bitmaps = 0;
+       int ret = -1;
+       bool next_page = false;
+       bool out_of_space = false;
+
+       INIT_LIST_HEAD(&bitmap_list);
+
+       node = rb_first(&ctl->free_space_offset);
+       if (!node)
+               return 0;
+
+       if (!i_size_read(inode))
+               return -1;
+
+       num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
+               PAGE_CACHE_SHIFT;
+
+       filemap_write_and_wait(inode->i_mapping);
+       btrfs_wait_ordered_range(inode, inode->i_size &
+                                ~(root->sectorsize - 1), (u64)-1);
+
+       pages = kzalloc(sizeof(struct page *) * num_pages, GFP_NOFS);
+       if (!pages)
+               return -1;
+
+       /* Get the cluster for this block_group if it exists */
+       if (block_group && !list_empty(&block_group->cluster_list))
+               cluster = list_entry(block_group->cluster_list.next,
+                                    struct btrfs_free_cluster,
+                                    block_group_list);
+
+       /*
+        * We shouldn't have switched the pinned extents yet so this is the
+        * right one
+        */
+       unpin = root->fs_info->pinned_extents;
+
+       /*
+        * Lock all pages first so we can lock the extent safely.
+        *
+        * NOTE: Because we hold the ref the entire time we're going to write to
+        * the page find_get_page should never fail, so we don't do a check
+        * after find_get_page at this point.  Just putting this here so people
+        * know and don't freak out.
+        */
+       while (index < num_pages) {
+               page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
+               if (!page) {
+                       int i;
+
+                       for (i = 0; i < num_pages; i++) {
+                               unlock_page(pages[i]);
+                               page_cache_release(pages[i]);
+                       }
+                       goto out;
+               }
+               pages[index] = page;
+               index++;
+       }
+
+       index = 0;
+       lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
+                        0, &cached_state, GFP_NOFS);
+
+       /*
+        * When searching for pinned extents, we need to start at our start
+        * offset.
+        */
+       if (block_group)
+               start = block_group->key.objectid;
+
+       /* Write out the extent entries */
+       do {
+               struct btrfs_free_space_entry *entry;
+               void *addr, *orig;
+               unsigned long offset = 0;
+
+               next_page = false;
+
+               if (index >= num_pages) {
+                       out_of_space = true;
+                       break;
+               }
+
+               page = pages[index];
+
+               orig = addr = kmap(page);
+               if (index == 0) {
+                       u64 *gen;
+
+                       /*
+                        * We're going to put in a bogus crc for this page to
+                        * make sure that old kernels who aren't aware of this
+                        * format will be sure to discard the cache.
+                        */
+                       addr += sizeof(u64);
+                       offset += sizeof(u64);
+
+                       gen = addr;
+                       *gen = trans->transid;
+                       addr += sizeof(u64);
+                       offset += sizeof(u64);
+               }
+               entry = addr;
+
+               memset(addr, 0, PAGE_CACHE_SIZE - offset);
+               while (node && !next_page) {
+                       struct btrfs_free_space *e;
+
+                       e = rb_entry(node, struct btrfs_free_space, offset_index);
+                       entries++;
+
+                       entry->offset = cpu_to_le64(e->offset);
+                       entry->bytes = cpu_to_le64(e->bytes);
+                       if (e->bitmap) {
+                               entry->type = BTRFS_FREE_SPACE_BITMAP;
+                               list_add_tail(&e->list, &bitmap_list);
+                               bitmaps++;
+                       } else {
+                               entry->type = BTRFS_FREE_SPACE_EXTENT;
+                       }
+                       node = rb_next(node);
+                       if (!node && cluster) {
+                               node = rb_first(&cluster->root);
+                               cluster = NULL;
+                       }
+                       offset += sizeof(struct btrfs_free_space_entry);
+                       if (offset + sizeof(struct btrfs_free_space_entry) >=
+                           PAGE_CACHE_SIZE)
+                               next_page = true;
+                       entry++;
+               }
+
+               /*
+                * We want to add any pinned extents to our free space cache
+                * so we don't leak the space
+                */
+               while (block_group && !next_page &&
+                      (start < block_group->key.objectid +
+                       block_group->key.offset)) {
+                       ret = find_first_extent_bit(unpin, start, &start, &end,
+                                                   EXTENT_DIRTY);
+                       if (ret) {
+                               ret = 0;
+                               break;
+                       }
+
+                       /* This pinned extent is out of our range */
+                       if (start >= block_group->key.objectid +
+                           block_group->key.offset)
+                               break;
+
+                       len = block_group->key.objectid +
+                               block_group->key.offset - start;
+                       len = min(len, end + 1 - start);
+
+                       entries++;
+                       entry->offset = cpu_to_le64(start);
+                       entry->bytes = cpu_to_le64(len);
+                       entry->type = BTRFS_FREE_SPACE_EXTENT;
+
+                       start = end + 1;
+                       offset += sizeof(struct btrfs_free_space_entry);
+                       if (offset + sizeof(struct btrfs_free_space_entry) >=
+                           PAGE_CACHE_SIZE)
+                               next_page = true;
+                       entry++;
+               }
+
+               /* Generate bogus crc value */
+               if (index == 0) {
+                       u32 *tmp;
+                       crc = btrfs_csum_data(root, orig + sizeof(u64), crc,
+                                             PAGE_CACHE_SIZE - sizeof(u64));
+                       btrfs_csum_final(crc, (char *)&crc);
+                       crc++;
+                       tmp = orig;
+                       *tmp = crc;
+               }
+
+               kunmap(page);
+
+               bytes += PAGE_CACHE_SIZE;
+
+               index++;
+       } while (node || next_page);
+
+       /* Write out the bitmaps */
+       list_for_each_safe(pos, n, &bitmap_list) {
+               void *addr;
+               struct btrfs_free_space *entry =
+                       list_entry(pos, struct btrfs_free_space, list);
+
+               if (index >= num_pages) {
+                       out_of_space = true;
+                       break;
+               }
+               page = pages[index];
+
+               addr = kmap(page);
+               memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
+               kunmap(page);
+               bytes += PAGE_CACHE_SIZE;
+
+               list_del_init(&entry->list);
+               index++;
+       }
+
+       if (out_of_space) {
+               btrfs_drop_pages(pages, num_pages);
+               unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
+                                    i_size_read(inode) - 1, &cached_state,
+                                    GFP_NOFS);
+               ret = 0;
+               goto out;
+       }
+
+       /* Zero out the rest of the pages just to make sure */
+       while (index < num_pages) {
+               void *addr;
+
+               page = pages[index];
+               addr = kmap(page);
+               memset(addr, 0, PAGE_CACHE_SIZE);
+               kunmap(page);
+               bytes += PAGE_CACHE_SIZE;
+               index++;
+       }
+
+       ret = btrfs_dirty_pages(root, inode, pages, num_pages, 0,
+                                           bytes, &cached_state);
+       btrfs_drop_pages(pages, num_pages);
+       unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
+                            i_size_read(inode) - 1, &cached_state, GFP_NOFS);
+
+       if (ret) {
+               ret = 0;
+               goto out;
+       }
+
+       BTRFS_I(inode)->generation = trans->transid;
+
+       filemap_write_and_wait(inode->i_mapping);
+
+       key.objectid = BTRFS_FREE_SPACE_OBJECTID;
+       key.offset = offset;
+       key.type = 0;
+
+       ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
+       if (ret < 0) {
+               ret = -1;
+               clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
+                                EXTENT_DIRTY | EXTENT_DELALLOC |
+                                EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
+               goto out;
+       }
+       leaf = path->nodes[0];
+       if (ret > 0) {
+               struct btrfs_key found_key;
+               BUG_ON(!path->slots[0]);
+               path->slots[0]--;
+               btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
+               if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
+                   found_key.offset != offset) {
+                       ret = -1;
+                       clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
+                                        EXTENT_DIRTY | EXTENT_DELALLOC |
+                                        EXTENT_DO_ACCOUNTING, 0, 0, NULL,
+                                        GFP_NOFS);
+                       btrfs_release_path(path);
+                       goto out;
+               }
+       }
+       header = btrfs_item_ptr(leaf, path->slots[0],
+                               struct btrfs_free_space_header);
+       btrfs_set_free_space_entries(leaf, header, entries);
+       btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
+       btrfs_set_free_space_generation(leaf, header, trans->transid);
+       btrfs_mark_buffer_dirty(leaf);
+       btrfs_release_path(path);
+
+       ret = 1;
+
+out:
+       kfree(pages);
+       if (ret != 1) {
+               invalidate_inode_pages2_range(inode->i_mapping, 0, index);
+               BTRFS_I(inode)->generation = 0;
+       }
+       btrfs_update_inode(trans, root, inode);
+       return ret;
+}
+
+int btrfs_write_out_cache(struct btrfs_root *root,
+                         struct btrfs_trans_handle *trans,
+                         struct btrfs_block_group_cache *block_group,
+                         struct btrfs_path *path)
+{
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct inode *inode;
+       int ret = 0;
+
+       root = root->fs_info->tree_root;
+
+       spin_lock(&block_group->lock);
+       if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
+               spin_unlock(&block_group->lock);
+               return 0;
+       }
+       spin_unlock(&block_group->lock);
+
+       inode = lookup_free_space_inode(root, block_group, path);
+       if (IS_ERR(inode))
+               return 0;
+
+       ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
+                                     path, block_group->key.objectid);
+       if (ret < 0) {
+               spin_lock(&block_group->lock);
+               block_group->disk_cache_state = BTRFS_DC_ERROR;
+               spin_unlock(&block_group->lock);
+               ret = 0;
+
+               printk(KERN_ERR "btrfs: failed to write free space cace "
+                      "for block group %llu\n", block_group->key.objectid);
+       }
+
+       iput(inode);
+       return ret;
+}
+
+static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
+                                         u64 offset)
+{
+       BUG_ON(offset < bitmap_start);
+       offset -= bitmap_start;
+       return (unsigned long)(div_u64(offset, unit));
+}
+
+static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
+{
+       return (unsigned long)(div_u64(bytes, unit));
+}
+
+static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
+                                  u64 offset)
+{
+       u64 bitmap_start;
+       u64 bytes_per_bitmap;
+
+       bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
+       bitmap_start = offset - ctl->start;
+       bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
+       bitmap_start *= bytes_per_bitmap;
+       bitmap_start += ctl->start;
+
+       return bitmap_start;
+}
+
+static int tree_insert_offset(struct rb_root *root, u64 offset,
+                             struct rb_node *node, int bitmap)
 {
        struct rb_node **p = &root->rb_node;
        struct rb_node *parent = NULL;
@@ -53,12 +920,40 @@ static int tree_insert_bytes(struct rb_root *root, u64 bytes,
 
        while (*p) {
                parent = *p;
-               info = rb_entry(parent, struct btrfs_free_space, bytes_index);
+               info = rb_entry(parent, struct btrfs_free_space, offset_index);
 
-               if (bytes < info->bytes)
+               if (offset < info->offset) {
                        p = &(*p)->rb_left;
-               else
+               } else if (offset > info->offset) {
                        p = &(*p)->rb_right;
+               } else {
+                       /*
+                        * we could have a bitmap entry and an extent entry
+                        * share the same offset.  If this is the case, we want
+                        * the extent entry to always be found first if we do a
+                        * linear search through the tree, since we want to have
+                        * the quickest allocation time, and allocating from an
+                        * extent is faster than allocating from a bitmap.  So
+                        * if we're inserting a bitmap and we find an entry at
+                        * this offset, we want to go right, or after this entry
+                        * logically.  If we are inserting an extent and we've
+                        * found a bitmap, we want to go left, or before
+                        * logically.
+                        */
+                       if (bitmap) {
+                               if (info->bitmap) {
+                                       WARN_ON_ONCE(1);
+                                       return -EEXIST;
+                               }
+                               p = &(*p)->rb_right;
+                       } else {
+                               if (!info->bitmap) {
+                                       WARN_ON_ONCE(1);
+                                       return -EEXIST;
+                               }
+                               p = &(*p)->rb_left;
+                       }
+               }
        }
 
        rb_link_node(node, parent, p);
@@ -68,249 +963,757 @@ static int tree_insert_bytes(struct rb_root *root, u64 bytes,
 }
 
 /*
- * searches the tree for the given offset.  If contains is set we will return
- * the free space that contains the given offset.  If contains is not set we
- * will return the free space that starts at or after the given offset and is
- * at least bytes long.
+ * searches the tree for the given offset.
+ *
+ * fuzzy - If this is set, then we are trying to make an allocation, and we just
+ * want a section that has at least bytes size and comes at or after the given
+ * offset.
  */
-static struct btrfs_free_space *tree_search_offset(struct rb_root *root,
-                                                  u64 offset, u64 bytes,
-                                                  int contains)
+static struct btrfs_free_space *
+tree_search_offset(struct btrfs_free_space_ctl *ctl,
+                  u64 offset, int bitmap_only, int fuzzy)
 {
-       struct rb_node *n = root->rb_node;
-       struct btrfs_free_space *entry, *ret = NULL;
+       struct rb_node *n = ctl->free_space_offset.rb_node;
+       struct btrfs_free_space *entry, *prev = NULL;
+
+       /* find entry that is closest to the 'offset' */
+       while (1) {
+               if (!n) {
+                       entry = NULL;
+                       break;
+               }
 
-       while (n) {
                entry = rb_entry(n, struct btrfs_free_space, offset_index);
+               prev = entry;
 
-               if (offset < entry->offset) {
-                       if (!contains &&
-                           (!ret || entry->offset < ret->offset) &&
-                           (bytes <= entry->bytes))
-                               ret = entry;
+               if (offset < entry->offset)
                        n = n->rb_left;
-               } else if (offset > entry->offset) {
-                       if ((entry->offset + entry->bytes - 1) >= offset &&
-                           bytes <= entry->bytes) {
-                               ret = entry;
-                               break;
-                       }
+               else if (offset > entry->offset)
                        n = n->rb_right;
-               } else {
-                       if (bytes > entry->bytes) {
-                               n = n->rb_right;
-                               continue;
-                       }
-                       ret = entry;
+               else
                        break;
-               }
        }
 
-       return ret;
-}
-
-/*
- * return a chunk at least bytes size, as close to offset that we can get.
- */
-static struct btrfs_free_space *tree_search_bytes(struct rb_root *root,
-                                                 u64 offset, u64 bytes)
-{
-       struct rb_node *n = root->rb_node;
-       struct btrfs_free_space *entry, *ret = NULL;
+       if (bitmap_only) {
+               if (!entry)
+                       return NULL;
+               if (entry->bitmap)
+                       return entry;
 
-       while (n) {
-               entry = rb_entry(n, struct btrfs_free_space, bytes_index);
+               /*
+                * bitmap entry and extent entry may share same offset,
+                * in that case, bitmap entry comes after extent entry.
+                */
+               n = rb_next(n);
+               if (!n)
+                       return NULL;
+               entry = rb_entry(n, struct btrfs_free_space, offset_index);
+               if (entry->offset != offset)
+                       return NULL;
 
-               if (bytes < entry->bytes) {
+               WARN_ON(!entry->bitmap);
+               return entry;
+       } else if (entry) {
+               if (entry->bitmap) {
                        /*
-                        * We prefer to get a hole size as close to the size we
-                        * are asking for so we don't take small slivers out of
-                        * huge holes, but we also want to get as close to the
-                        * offset as possible so we don't have a whole lot of
-                        * fragmentation.
+                        * if previous extent entry covers the offset,
+                        * we should return it instead of the bitmap entry
                         */
-                       if (offset <= entry->offset) {
-                               if (!ret)
-                                       ret = entry;
-                               else if (entry->bytes < ret->bytes)
-                                       ret = entry;
-                               else if (entry->offset < ret->offset)
-                                       ret = entry;
+                       n = &entry->offset_index;
+                       while (1) {
+                               n = rb_prev(n);
+                               if (!n)
+                                       break;
+                               prev = rb_entry(n, struct btrfs_free_space,
+                                               offset_index);
+                               if (!prev->bitmap) {
+                                       if (prev->offset + prev->bytes > offset)
+                                               entry = prev;
+                                       break;
+                               }
                        }
-                       n = n->rb_left;
-               } else if (bytes > entry->bytes) {
-                       n = n->rb_right;
+               }
+               return entry;
+       }
+
+       if (!prev)
+               return NULL;
+
+       /* find last entry before the 'offset' */
+       entry = prev;
+       if (entry->offset > offset) {
+               n = rb_prev(&entry->offset_index);
+               if (n) {
+                       entry = rb_entry(n, struct btrfs_free_space,
+                                       offset_index);
+                       BUG_ON(entry->offset > offset);
                } else {
-                       /*
-                        * Ok we may have multiple chunks of the wanted size,
-                        * so we don't want to take the first one we find, we
-                        * want to take the one closest to our given offset, so
-                        * keep searching just in case theres a better match.
-                        */
-                       n = n->rb_right;
-                       if (offset > entry->offset)
-                               continue;
-                       else if (!ret || entry->offset < ret->offset)
-                               ret = entry;
+                       if (fuzzy)
+                               return entry;
+                       else
+                               return NULL;
                }
        }
 
-       return ret;
+       if (entry->bitmap) {
+               n = &entry->offset_index;
+               while (1) {
+                       n = rb_prev(n);
+                       if (!n)
+                               break;
+                       prev = rb_entry(n, struct btrfs_free_space,
+                                       offset_index);
+                       if (!prev->bitmap) {
+                               if (prev->offset + prev->bytes > offset)
+                                       return prev;
+                               break;
+                       }
+               }
+               if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
+                       return entry;
+       } else if (entry->offset + entry->bytes > offset)
+               return entry;
+
+       if (!fuzzy)
+               return NULL;
+
+       while (1) {
+               if (entry->bitmap) {
+                       if (entry->offset + BITS_PER_BITMAP *
+                           ctl->unit > offset)
+                               break;
+               } else {
+                       if (entry->offset + entry->bytes > offset)
+                               break;
+               }
+
+               n = rb_next(&entry->offset_index);
+               if (!n)
+                       return NULL;
+               entry = rb_entry(n, struct btrfs_free_space, offset_index);
+       }
+       return entry;
+}
+
+static inline void
+__unlink_free_space(struct btrfs_free_space_ctl *ctl,
+                   struct btrfs_free_space *info)
+{
+       rb_erase(&info->offset_index, &ctl->free_space_offset);
+       ctl->free_extents--;
 }
 
-static void unlink_free_space(struct btrfs_block_group_cache *block_group,
+static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
                              struct btrfs_free_space *info)
 {
-       rb_erase(&info->offset_index, &block_group->free_space_offset);
-       rb_erase(&info->bytes_index, &block_group->free_space_bytes);
+       __unlink_free_space(ctl, info);
+       ctl->free_space -= info->bytes;
 }
 
-static int link_free_space(struct btrfs_block_group_cache *block_group,
+static int link_free_space(struct btrfs_free_space_ctl *ctl,
                           struct btrfs_free_space *info)
 {
        int ret = 0;
 
-
-       ret = tree_insert_offset(&block_group->free_space_offset, info->offset,
-                                &info->offset_index);
+       BUG_ON(!info->bitmap && !info->bytes);
+       ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
+                                &info->offset_index, (info->bitmap != NULL));
        if (ret)
                return ret;
 
-       ret = tree_insert_bytes(&block_group->free_space_bytes, info->bytes,
-                               &info->bytes_index);
-       if (ret)
-               return ret;
+       ctl->free_space += info->bytes;
+       ctl->free_extents++;
+       return ret;
+}
+
+static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
+{
+       struct btrfs_block_group_cache *block_group = ctl->private;
+       u64 max_bytes;
+       u64 bitmap_bytes;
+       u64 extent_bytes;
+       u64 size = block_group->key.offset;
+       u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
+       int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
+
+       BUG_ON(ctl->total_bitmaps > max_bitmaps);
+
+       /*
+        * The goal is to keep the total amount of memory used per 1gb of space
+        * at or below 32k, so we need to adjust how much memory we allow to be
+        * used by extent based free space tracking
+        */
+       if (size < 1024 * 1024 * 1024)
+               max_bytes = MAX_CACHE_BYTES_PER_GIG;
+       else
+               max_bytes = MAX_CACHE_BYTES_PER_GIG *
+                       div64_u64(size, 1024 * 1024 * 1024);
+
+       /*
+        * we want to account for 1 more bitmap than what we have so we can make
+        * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
+        * we add more bitmaps.
+        */
+       bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
+
+       if (bitmap_bytes >= max_bytes) {
+               ctl->extents_thresh = 0;
+               return;
+       }
+
+       /*
+        * we want the extent entry threshold to always be at most 1/2 the maxw
+        * bytes we can have, or whatever is less than that.
+        */
+       extent_bytes = max_bytes - bitmap_bytes;
+       extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
+
+       ctl->extents_thresh =
+               div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
+}
+
+static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
+                                      struct btrfs_free_space *info,
+                                      u64 offset, u64 bytes)
+{
+       unsigned long start, count;
+
+       start = offset_to_bit(info->offset, ctl->unit, offset);
+       count = bytes_to_bits(bytes, ctl->unit);
+       BUG_ON(start + count > BITS_PER_BITMAP);
+
+       bitmap_clear(info->bitmap, start, count);
+
+       info->bytes -= bytes;
+}
+
+static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
+                             struct btrfs_free_space *info, u64 offset,
+                             u64 bytes)
+{
+       __bitmap_clear_bits(ctl, info, offset, bytes);
+       ctl->free_space -= bytes;
+}
+
+static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
+                           struct btrfs_free_space *info, u64 offset,
+                           u64 bytes)
+{
+       unsigned long start, count;
+
+       start = offset_to_bit(info->offset, ctl->unit, offset);
+       count = bytes_to_bits(bytes, ctl->unit);
+       BUG_ON(start + count > BITS_PER_BITMAP);
+
+       bitmap_set(info->bitmap, start, count);
+
+       info->bytes += bytes;
+       ctl->free_space += bytes;
+}
+
+static int search_bitmap(struct btrfs_free_space_ctl *ctl,
+                        struct btrfs_free_space *bitmap_info, u64 *offset,
+                        u64 *bytes)
+{
+       unsigned long found_bits = 0;
+       unsigned long bits, i;
+       unsigned long next_zero;
+
+       i = offset_to_bit(bitmap_info->offset, ctl->unit,
+                         max_t(u64, *offset, bitmap_info->offset));
+       bits = bytes_to_bits(*bytes, ctl->unit);
+
+       for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
+            i < BITS_PER_BITMAP;
+            i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
+               next_zero = find_next_zero_bit(bitmap_info->bitmap,
+                                              BITS_PER_BITMAP, i);
+               if ((next_zero - i) >= bits) {
+                       found_bits = next_zero - i;
+                       break;
+               }
+               i = next_zero;
+       }
+
+       if (found_bits) {
+               *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
+               *bytes = (u64)(found_bits) * ctl->unit;
+               return 0;
+       }
+
+       return -1;
+}
+
+static struct btrfs_free_space *
+find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes)
+{
+       struct btrfs_free_space *entry;
+       struct rb_node *node;
+       int ret;
+
+       if (!ctl->free_space_offset.rb_node)
+               return NULL;
+
+       entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
+       if (!entry)
+               return NULL;
+
+       for (node = &entry->offset_index; node; node = rb_next(node)) {
+               entry = rb_entry(node, struct btrfs_free_space, offset_index);
+               if (entry->bytes < *bytes)
+                       continue;
+
+               if (entry->bitmap) {
+                       ret = search_bitmap(ctl, entry, offset, bytes);
+                       if (!ret)
+                               return entry;
+                       continue;
+               }
+
+               *offset = entry->offset;
+               *bytes = entry->bytes;
+               return entry;
+       }
+
+       return NULL;
+}
+
+static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
+                          struct btrfs_free_space *info, u64 offset)
+{
+       info->offset = offset_to_bitmap(ctl, offset);
+       info->bytes = 0;
+       link_free_space(ctl, info);
+       ctl->total_bitmaps++;
+
+       ctl->op->recalc_thresholds(ctl);
+}
+
+static void free_bitmap(struct btrfs_free_space_ctl *ctl,
+                       struct btrfs_free_space *bitmap_info)
+{
+       unlink_free_space(ctl, bitmap_info);
+       kfree(bitmap_info->bitmap);
+       kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
+       ctl->total_bitmaps--;
+       ctl->op->recalc_thresholds(ctl);
+}
+
+static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
+                             struct btrfs_free_space *bitmap_info,
+                             u64 *offset, u64 *bytes)
+{
+       u64 end;
+       u64 search_start, search_bytes;
+       int ret;
+
+again:
+       end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
+
+       /*
+        * XXX - this can go away after a few releases.
+        *
+        * since the only user of btrfs_remove_free_space is the tree logging
+        * stuff, and the only way to test that is under crash conditions, we
+        * want to have this debug stuff here just in case somethings not
+        * working.  Search the bitmap for the space we are trying to use to
+        * make sure its actually there.  If its not there then we need to stop
+        * because something has gone wrong.
+        */
+       search_start = *offset;
+       search_bytes = *bytes;
+       search_bytes = min(search_bytes, end - search_start + 1);
+       ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
+       BUG_ON(ret < 0 || search_start != *offset);
+
+       if (*offset > bitmap_info->offset && *offset + *bytes > end) {
+               bitmap_clear_bits(ctl, bitmap_info, *offset, end - *offset + 1);
+               *bytes -= end - *offset + 1;
+               *offset = end + 1;
+       } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
+               bitmap_clear_bits(ctl, bitmap_info, *offset, *bytes);
+               *bytes = 0;
+       }
+
+       if (*bytes) {
+               struct rb_node *next = rb_next(&bitmap_info->offset_index);
+               if (!bitmap_info->bytes)
+                       free_bitmap(ctl, bitmap_info);
+
+               /*
+                * no entry after this bitmap, but we still have bytes to
+                * remove, so something has gone wrong.
+                */
+               if (!next)
+                       return -EINVAL;
+
+               bitmap_info = rb_entry(next, struct btrfs_free_space,
+                                      offset_index);
+
+               /*
+                * if the next entry isn't a bitmap we need to return to let the
+                * extent stuff do its work.
+                */
+               if (!bitmap_info->bitmap)
+                       return -EAGAIN;
+
+               /*
+                * Ok the next item is a bitmap, but it may not actually hold
+                * the information for the rest of this free space stuff, so
+                * look for it, and if we don't find it return so we can try
+                * everything over again.
+                */
+               search_start = *offset;
+               search_bytes = *bytes;
+               ret = search_bitmap(ctl, bitmap_info, &search_start,
+                                   &search_bytes);
+               if (ret < 0 || search_start != *offset)
+                       return -EAGAIN;
+
+               goto again;
+       } else if (!bitmap_info->bytes)
+               free_bitmap(ctl, bitmap_info);
+
+       return 0;
+}
+
+static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
+                              struct btrfs_free_space *info, u64 offset,
+                              u64 bytes)
+{
+       u64 bytes_to_set = 0;
+       u64 end;
+
+       end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
+
+       bytes_to_set = min(end - offset, bytes);
+
+       bitmap_set_bits(ctl, info, offset, bytes_to_set);
+
+       return bytes_to_set;
+
+}
+
+static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
+                     struct btrfs_free_space *info)
+{
+       struct btrfs_block_group_cache *block_group = ctl->private;
+
+       /*
+        * If we are below the extents threshold then we can add this as an
+        * extent, and don't have to deal with the bitmap
+        */
+       if (ctl->free_extents < ctl->extents_thresh) {
+               /*
+                * If this block group has some small extents we don't want to
+                * use up all of our free slots in the cache with them, we want
+                * to reserve them to larger extents, however if we have plent
+                * of cache left then go ahead an dadd them, no sense in adding
+                * the overhead of a bitmap if we don't have to.
+                */
+               if (info->bytes <= block_group->sectorsize * 4) {
+                       if (ctl->free_extents * 2 <= ctl->extents_thresh)
+                               return false;
+               } else {
+                       return false;
+               }
+       }
+
+       /*
+        * some block groups are so tiny they can't be enveloped by a bitmap, so
+        * don't even bother to create a bitmap for this
+        */
+       if (BITS_PER_BITMAP * block_group->sectorsize >
+           block_group->key.offset)
+               return false;
+
+       return true;
+}
+
+static struct btrfs_free_space_op free_space_op = {
+       .recalc_thresholds      = recalculate_thresholds,
+       .use_bitmap             = use_bitmap,
+};
+
+static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
+                             struct btrfs_free_space *info)
+{
+       struct btrfs_free_space *bitmap_info;
+       struct btrfs_block_group_cache *block_group = NULL;
+       int added = 0;
+       u64 bytes, offset, bytes_added;
+       int ret;
+
+       bytes = info->bytes;
+       offset = info->offset;
+
+       if (!ctl->op->use_bitmap(ctl, info))
+               return 0;
+
+       if (ctl->op == &free_space_op)
+               block_group = ctl->private;
+again:
+       /*
+        * Since we link bitmaps right into the cluster we need to see if we
+        * have a cluster here, and if so and it has our bitmap we need to add
+        * the free space to that bitmap.
+        */
+       if (block_group && !list_empty(&block_group->cluster_list)) {
+               struct btrfs_free_cluster *cluster;
+               struct rb_node *node;
+               struct btrfs_free_space *entry;
+
+               cluster = list_entry(block_group->cluster_list.next,
+                                    struct btrfs_free_cluster,
+                                    block_group_list);
+               spin_lock(&cluster->lock);
+               node = rb_first(&cluster->root);
+               if (!node) {
+                       spin_unlock(&cluster->lock);
+                       goto no_cluster_bitmap;
+               }
+
+               entry = rb_entry(node, struct btrfs_free_space, offset_index);
+               if (!entry->bitmap) {
+                       spin_unlock(&cluster->lock);
+                       goto no_cluster_bitmap;
+               }
+
+               if (entry->offset == offset_to_bitmap(ctl, offset)) {
+                       bytes_added = add_bytes_to_bitmap(ctl, entry,
+                                                         offset, bytes);
+                       bytes -= bytes_added;
+                       offset += bytes_added;
+               }
+               spin_unlock(&cluster->lock);
+               if (!bytes) {
+                       ret = 1;
+                       goto out;
+               }
+       }
+
+no_cluster_bitmap:
+       bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
+                                        1, 0);
+       if (!bitmap_info) {
+               BUG_ON(added);
+               goto new_bitmap;
+       }
+
+       bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
+       bytes -= bytes_added;
+       offset += bytes_added;
+       added = 0;
+
+       if (!bytes) {
+               ret = 1;
+               goto out;
+       } else
+               goto again;
+
+new_bitmap:
+       if (info && info->bitmap) {
+               add_new_bitmap(ctl, info, offset);
+               added = 1;
+               info = NULL;
+               goto again;
+       } else {
+               spin_unlock(&ctl->tree_lock);
+
+               /* no pre-allocated info, allocate a new one */
+               if (!info) {
+                       info = kmem_cache_zalloc(btrfs_free_space_cachep,
+                                                GFP_NOFS);
+                       if (!info) {
+                               spin_lock(&ctl->tree_lock);
+                               ret = -ENOMEM;
+                               goto out;
+                       }
+               }
+
+               /* allocate the bitmap */
+               info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
+               spin_lock(&ctl->tree_lock);
+               if (!info->bitmap) {
+                       ret = -ENOMEM;
+                       goto out;
+               }
+               goto again;
+       }
+
+out:
+       if (info) {
+               if (info->bitmap)
+                       kfree(info->bitmap);
+               kmem_cache_free(btrfs_free_space_cachep, info);
+       }
 
        return ret;
 }
 
-static int __btrfs_add_free_space(struct btrfs_block_group_cache *block_group,
-                                 u64 offset, u64 bytes)
+static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
+                         struct btrfs_free_space *info, bool update_stat)
 {
-       struct btrfs_free_space *right_info;
        struct btrfs_free_space *left_info;
-       struct btrfs_free_space *info = NULL;
-       struct btrfs_free_space *alloc_info;
-       int ret = 0;
-
-       alloc_info = kzalloc(sizeof(struct btrfs_free_space), GFP_NOFS);
-       if (!alloc_info)
-               return -ENOMEM;
+       struct btrfs_free_space *right_info;
+       bool merged = false;
+       u64 offset = info->offset;
+       u64 bytes = info->bytes;
 
        /*
         * first we want to see if there is free space adjacent to the range we
         * are adding, if there is remove that struct and add a new one to
         * cover the entire range
         */
-       right_info = tree_search_offset(&block_group->free_space_offset,
-                                       offset+bytes, 0, 1);
-       left_info = tree_search_offset(&block_group->free_space_offset,
-                                      offset-1, 0, 1);
-
-       if (right_info && right_info->offset == offset+bytes) {
-               unlink_free_space(block_group, right_info);
-               info = right_info;
-               info->offset = offset;
-               info->bytes += bytes;
-       } else if (right_info && right_info->offset != offset+bytes) {
-               printk(KERN_ERR "btrfs adding space in the middle of an "
-                      "existing free space area. existing: "
-                      "offset=%llu, bytes=%llu. new: offset=%llu, "
-                      "bytes=%llu\n", (unsigned long long)right_info->offset,
-                      (unsigned long long)right_info->bytes,
-                      (unsigned long long)offset,
-                      (unsigned long long)bytes);
-               BUG();
-       }
-
-       if (left_info) {
-               unlink_free_space(block_group, left_info);
-
-               if (unlikely((left_info->offset + left_info->bytes) !=
-                            offset)) {
-                       printk(KERN_ERR "btrfs free space to the left "
-                              "of new free space isn't "
-                              "quite right. existing: offset=%llu, "
-                              "bytes=%llu. new: offset=%llu, bytes=%llu\n",
-                              (unsigned long long)left_info->offset,
-                              (unsigned long long)left_info->bytes,
-                              (unsigned long long)offset,
-                              (unsigned long long)bytes);
-                       BUG();
-               }
-
-               if (info) {
-                       info->offset = left_info->offset;
-                       info->bytes += left_info->bytes;
-                       kfree(left_info);
-               } else {
-                       info = left_info;
-                       info->bytes += bytes;
-               }
+       right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
+       if (right_info && rb_prev(&right_info->offset_index))
+               left_info = rb_entry(rb_prev(&right_info->offset_index),
+                                    struct btrfs_free_space, offset_index);
+       else
+               left_info = tree_search_offset(ctl, offset - 1, 0, 0);
+
+       if (right_info && !right_info->bitmap) {
+               if (update_stat)
+                       unlink_free_space(ctl, right_info);
+               else
+                       __unlink_free_space(ctl, right_info);
+               info->bytes += right_info->bytes;
+               kmem_cache_free(btrfs_free_space_cachep, right_info);
+               merged = true;
        }
 
-       if (info) {
-               ret = link_free_space(block_group, info);
-               if (!ret)
-                       info = NULL;
-               goto out;
+       if (left_info && !left_info->bitmap &&
+           left_info->offset + left_info->bytes == offset) {
+               if (update_stat)
+                       unlink_free_space(ctl, left_info);
+               else
+                       __unlink_free_space(ctl, left_info);
+               info->offset = left_info->offset;
+               info->bytes += left_info->bytes;
+               kmem_cache_free(btrfs_free_space_cachep, left_info);
+               merged = true;
        }
 
-       info = alloc_info;
-       alloc_info = NULL;
+       return merged;
+}
+
+int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
+                          u64 offset, u64 bytes)
+{
+       struct btrfs_free_space *info;
+       int ret = 0;
+
+       info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
+       if (!info)
+               return -ENOMEM;
+
        info->offset = offset;
        info->bytes = bytes;
 
-       ret = link_free_space(block_group, info);
+       spin_lock(&ctl->tree_lock);
+
+       if (try_merge_free_space(ctl, info, true))
+               goto link;
+
+       /*
+        * There was no extent directly to the left or right of this new
+        * extent then we know we're going to have to allocate a new extent, so
+        * before we do that see if we need to drop this into a bitmap
+        */
+       ret = insert_into_bitmap(ctl, info);
+       if (ret < 0) {
+               goto out;
+       } else if (ret) {
+               ret = 0;
+               goto out;
+       }
+link:
+       ret = link_free_space(ctl, info);
        if (ret)
-               kfree(info);
+               kmem_cache_free(btrfs_free_space_cachep, info);
 out:
+       spin_unlock(&ctl->tree_lock);
+
        if (ret) {
-               printk(KERN_ERR "btrfs: unable to add free space :%d\n", ret);
-               if (ret == -EEXIST)
-                       BUG();
+               printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
+               BUG_ON(ret == -EEXIST);
        }
 
-       kfree(alloc_info);
-
        return ret;
 }
 
-static int
-__btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
-                         u64 offset, u64 bytes)
+int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
+                           u64 offset, u64 bytes)
 {
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
        struct btrfs_free_space *info;
+       struct btrfs_free_space *next_info = NULL;
        int ret = 0;
 
-       info = tree_search_offset(&block_group->free_space_offset, offset, 0,
-                                 1);
+       spin_lock(&ctl->tree_lock);
+
+again:
+       info = tree_search_offset(ctl, offset, 0, 0);
+       if (!info) {
+               /*
+                * oops didn't find an extent that matched the space we wanted
+                * to remove, look for a bitmap instead
+                */
+               info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
+                                         1, 0);
+               if (!info) {
+                       WARN_ON(1);
+                       goto out_lock;
+               }
+       }
+
+       if (info->bytes < bytes && rb_next(&info->offset_index)) {
+               u64 end;
+               next_info = rb_entry(rb_next(&info->offset_index),
+                                            struct btrfs_free_space,
+                                            offset_index);
 
-       if (info && info->offset == offset) {
-               if (info->bytes < bytes) {
-                       printk(KERN_ERR "Found free space at %llu, size %llu,"
-                              "trying to use %llu\n",
-                              (unsigned long long)info->offset,
-                              (unsigned long long)info->bytes,
-                              (unsigned long long)bytes);
+               if (next_info->bitmap)
+                       end = next_info->offset +
+                             BITS_PER_BITMAP * ctl->unit - 1;
+               else
+                       end = next_info->offset + next_info->bytes;
+
+               if (next_info->bytes < bytes ||
+                   next_info->offset > offset || offset > end) {
+                       printk(KERN_CRIT "Found free space at %llu, size %llu,"
+                             " trying to use %llu\n",
+                             (unsigned long long)info->offset,
+                             (unsigned long long)info->bytes,
+                             (unsigned long long)bytes);
                        WARN_ON(1);
                        ret = -EINVAL;
-                       goto out;
+                       goto out_lock;
                }
-               unlink_free_space(block_group, info);
 
-               if (info->bytes == bytes) {
-                       kfree(info);
-                       goto out;
+               info = next_info;
+       }
+
+       if (info->bytes == bytes) {
+               unlink_free_space(ctl, info);
+               if (info->bitmap) {
+                       kfree(info->bitmap);
+                       ctl->total_bitmaps--;
                }
+               kmem_cache_free(btrfs_free_space_cachep, info);
+               goto out_lock;
+       }
 
+       if (!info->bitmap && info->offset == offset) {
+               unlink_free_space(ctl, info);
                info->offset += bytes;
                info->bytes -= bytes;
+               link_free_space(ctl, info);
+               goto out_lock;
+       }
 
-               ret = link_free_space(block_group, info);
-               BUG_ON(ret);
-       } else if (info && info->offset < offset &&
-                  info->offset + info->bytes >= offset + bytes) {
+       if (!info->bitmap && info->offset <= offset &&
+           info->offset + info->bytes >= offset + bytes) {
                u64 old_start = info->offset;
                /*
                 * we're freeing space in the middle of the info,
@@ -319,177 +1722,931 @@ __btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
                 * first unlink the old info and then
                 * insert it again after the hole we're creating
                 */
-               unlink_free_space(block_group, info);
+               unlink_free_space(ctl, info);
                if (offset + bytes < info->offset + info->bytes) {
                        u64 old_end = info->offset + info->bytes;
 
                        info->offset = offset + bytes;
                        info->bytes = old_end - info->offset;
-                       ret = link_free_space(block_group, info);
-                       BUG_ON(ret);
+                       ret = link_free_space(ctl, info);
+                       WARN_ON(ret);
+                       if (ret)
+                               goto out_lock;
                } else {
                        /* the hole we're creating ends at the end
                         * of the info struct, just free the info
                         */
-                       kfree(info);
+                       kmem_cache_free(btrfs_free_space_cachep, info);
                }
+               spin_unlock(&ctl->tree_lock);
 
-               /* step two, insert a new info struct to cover anything
-                * before the hole
+               /* step two, insert a new info struct to cover
+                * anything before the hole
                 */
-               ret = __btrfs_add_free_space(block_group, old_start,
-                                            offset - old_start);
-               BUG_ON(ret);
+               ret = btrfs_add_free_space(block_group, old_start,
+                                          offset - old_start);
+               WARN_ON(ret);
+               goto out;
+       }
+
+       ret = remove_from_bitmap(ctl, info, &offset, &bytes);
+       if (ret == -EAGAIN)
+               goto again;
+       BUG_ON(ret);
+out_lock:
+       spin_unlock(&ctl->tree_lock);
+out:
+       return ret;
+}
+
+void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
+                          u64 bytes)
+{
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_free_space *info;
+       struct rb_node *n;
+       int count = 0;
+
+       for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
+               info = rb_entry(n, struct btrfs_free_space, offset_index);
+               if (info->bytes >= bytes)
+                       count++;
+               printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
+                      (unsigned long long)info->offset,
+                      (unsigned long long)info->bytes,
+                      (info->bitmap) ? "yes" : "no");
+       }
+       printk(KERN_INFO "block group has cluster?: %s\n",
+              list_empty(&block_group->cluster_list) ? "no" : "yes");
+       printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
+              "\n", count);
+}
+
+void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
+{
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+
+       spin_lock_init(&ctl->tree_lock);
+       ctl->unit = block_group->sectorsize;
+       ctl->start = block_group->key.objectid;
+       ctl->private = block_group;
+       ctl->op = &free_space_op;
+
+       /*
+        * we only want to have 32k of ram per block group for keeping
+        * track of free space, and if we pass 1/2 of that we want to
+        * start converting things over to using bitmaps
+        */
+       ctl->extents_thresh = ((1024 * 32) / 2) /
+                               sizeof(struct btrfs_free_space);
+}
+
+/*
+ * for a given cluster, put all of its extents back into the free
+ * space cache.  If the block group passed doesn't match the block group
+ * pointed to by the cluster, someone else raced in and freed the
+ * cluster already.  In that case, we just return without changing anything
+ */
+static int
+__btrfs_return_cluster_to_free_space(
+                            struct btrfs_block_group_cache *block_group,
+                            struct btrfs_free_cluster *cluster)
+{
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_free_space *entry;
+       struct rb_node *node;
+
+       spin_lock(&cluster->lock);
+       if (cluster->block_group != block_group)
+               goto out;
+
+       cluster->block_group = NULL;
+       cluster->window_start = 0;
+       list_del_init(&cluster->block_group_list);
+
+       node = rb_first(&cluster->root);
+       while (node) {
+               bool bitmap;
+
+               entry = rb_entry(node, struct btrfs_free_space, offset_index);
+               node = rb_next(&entry->offset_index);
+               rb_erase(&entry->offset_index, &cluster->root);
+
+               bitmap = (entry->bitmap != NULL);
+               if (!bitmap)
+                       try_merge_free_space(ctl, entry, false);
+               tree_insert_offset(&ctl->free_space_offset,
+                                  entry->offset, &entry->offset_index, bitmap);
+       }
+       cluster->root = RB_ROOT;
+
+out:
+       spin_unlock(&cluster->lock);
+       btrfs_put_block_group(block_group);
+       return 0;
+}
+
+void __btrfs_remove_free_space_cache_locked(struct btrfs_free_space_ctl *ctl)
+{
+       struct btrfs_free_space *info;
+       struct rb_node *node;
+
+       while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
+               info = rb_entry(node, struct btrfs_free_space, offset_index);
+               if (!info->bitmap) {
+                       unlink_free_space(ctl, info);
+                       kmem_cache_free(btrfs_free_space_cachep, info);
+               } else {
+                       free_bitmap(ctl, info);
+               }
+               if (need_resched()) {
+                       spin_unlock(&ctl->tree_lock);
+                       cond_resched();
+                       spin_lock(&ctl->tree_lock);
+               }
+       }
+}
+
+void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
+{
+       spin_lock(&ctl->tree_lock);
+       __btrfs_remove_free_space_cache_locked(ctl);
+       spin_unlock(&ctl->tree_lock);
+}
+
+void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
+{
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_free_cluster *cluster;
+       struct list_head *head;
+
+       spin_lock(&ctl->tree_lock);
+       while ((head = block_group->cluster_list.next) !=
+              &block_group->cluster_list) {
+               cluster = list_entry(head, struct btrfs_free_cluster,
+                                    block_group_list);
+
+               WARN_ON(cluster->block_group != block_group);
+               __btrfs_return_cluster_to_free_space(block_group, cluster);
+               if (need_resched()) {
+                       spin_unlock(&ctl->tree_lock);
+                       cond_resched();
+                       spin_lock(&ctl->tree_lock);
+               }
+       }
+       __btrfs_remove_free_space_cache_locked(ctl);
+       spin_unlock(&ctl->tree_lock);
+
+}
+
+u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
+                              u64 offset, u64 bytes, u64 empty_size)
+{
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_free_space *entry = NULL;
+       u64 bytes_search = bytes + empty_size;
+       u64 ret = 0;
+
+       spin_lock(&ctl->tree_lock);
+       entry = find_free_space(ctl, &offset, &bytes_search);
+       if (!entry)
+               goto out;
+
+       ret = offset;
+       if (entry->bitmap) {
+               bitmap_clear_bits(ctl, entry, offset, bytes);
+               if (!entry->bytes)
+                       free_bitmap(ctl, entry);
        } else {
-               WARN_ON(1);
+               unlink_free_space(ctl, entry);
+               entry->offset += bytes;
+               entry->bytes -= bytes;
+               if (!entry->bytes)
+                       kmem_cache_free(btrfs_free_space_cachep, entry);
+               else
+                       link_free_space(ctl, entry);
        }
+
 out:
+       spin_unlock(&ctl->tree_lock);
+
        return ret;
 }
 
-int btrfs_add_free_space(struct btrfs_block_group_cache *block_group,
-                        u64 offset, u64 bytes)
+/*
+ * given a cluster, put all of its extents back into the free space
+ * cache.  If a block group is passed, this function will only free
+ * a cluster that belongs to the passed block group.
+ *
+ * Otherwise, it'll get a reference on the block group pointed to by the
+ * cluster and remove the cluster from it.
+ */
+int btrfs_return_cluster_to_free_space(
+                              struct btrfs_block_group_cache *block_group,
+                              struct btrfs_free_cluster *cluster)
 {
+       struct btrfs_free_space_ctl *ctl;
        int ret;
-       struct btrfs_free_space *sp;
 
-       mutex_lock(&block_group->alloc_mutex);
-       ret = __btrfs_add_free_space(block_group, offset, bytes);
-       sp = tree_search_offset(&block_group->free_space_offset, offset, 0, 1);
-       BUG_ON(!sp);
-       mutex_unlock(&block_group->alloc_mutex);
+       /* first, get a safe pointer to the block group */
+       spin_lock(&cluster->lock);
+       if (!block_group) {
+               block_group = cluster->block_group;
+               if (!block_group) {
+                       spin_unlock(&cluster->lock);
+                       return 0;
+               }
+       } else if (cluster->block_group != block_group) {
+               /* someone else has already freed it don't redo their work */
+               spin_unlock(&cluster->lock);
+               return 0;
+       }
+       atomic_inc(&block_group->count);
+       spin_unlock(&cluster->lock);
+
+       ctl = block_group->free_space_ctl;
+
+       /* now return any extents the cluster had on it */
+       spin_lock(&ctl->tree_lock);
+       ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
+       spin_unlock(&ctl->tree_lock);
 
+       /* finally drop our ref */
+       btrfs_put_block_group(block_group);
        return ret;
 }
 
-int btrfs_add_free_space_lock(struct btrfs_block_group_cache *block_group,
-                             u64 offset, u64 bytes)
+static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
+                                  struct btrfs_free_cluster *cluster,
+                                  struct btrfs_free_space *entry,
+                                  u64 bytes, u64 min_start)
 {
-       int ret;
-       struct btrfs_free_space *sp;
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       int err;
+       u64 search_start = cluster->window_start;
+       u64 search_bytes = bytes;
+       u64 ret = 0;
+
+       search_start = min_start;
+       search_bytes = bytes;
 
-       ret = __btrfs_add_free_space(block_group, offset, bytes);
-       sp = tree_search_offset(&block_group->free_space_offset, offset, 0, 1);
-       BUG_ON(!sp);
+       err = search_bitmap(ctl, entry, &search_start, &search_bytes);
+       if (err)
+               return 0;
+
+       ret = search_start;
+       __bitmap_clear_bits(ctl, entry, ret, bytes);
 
        return ret;
 }
 
-int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
-                           u64 offset, u64 bytes)
+/*
+ * given a cluster, try to allocate 'bytes' from it, returns 0
+ * if it couldn't find anything suitably large, or a logical disk offset
+ * if things worked out
+ */
+u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
+                            struct btrfs_free_cluster *cluster, u64 bytes,
+                            u64 min_start)
 {
-       int ret = 0;
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_free_space *entry = NULL;
+       struct rb_node *node;
+       u64 ret = 0;
+
+       spin_lock(&cluster->lock);
+       if (bytes > cluster->max_size)
+               goto out;
+
+       if (cluster->block_group != block_group)
+               goto out;
+
+       node = rb_first(&cluster->root);
+       if (!node)
+               goto out;
+
+       entry = rb_entry(node, struct btrfs_free_space, offset_index);
+       while(1) {
+               if (entry->bytes < bytes ||
+                   (!entry->bitmap && entry->offset < min_start)) {
+                       node = rb_next(&entry->offset_index);
+                       if (!node)
+                               break;
+                       entry = rb_entry(node, struct btrfs_free_space,
+                                        offset_index);
+                       continue;
+               }
+
+               if (entry->bitmap) {
+                       ret = btrfs_alloc_from_bitmap(block_group,
+                                                     cluster, entry, bytes,
+                                                     min_start);
+                       if (ret == 0) {
+                               node = rb_next(&entry->offset_index);
+                               if (!node)
+                                       break;
+                               entry = rb_entry(node, struct btrfs_free_space,
+                                                offset_index);
+                               continue;
+                       }
+               } else {
+                       ret = entry->offset;
+
+                       entry->offset += bytes;
+                       entry->bytes -= bytes;
+               }
+
+               if (entry->bytes == 0)
+                       rb_erase(&entry->offset_index, &cluster->root);
+               break;
+       }
+out:
+       spin_unlock(&cluster->lock);
+
+       if (!ret)
+               return 0;
 
-       mutex_lock(&block_group->alloc_mutex);
-       ret = __btrfs_remove_free_space(block_group, offset, bytes);
-       mutex_unlock(&block_group->alloc_mutex);
+       spin_lock(&ctl->tree_lock);
+
+       ctl->free_space -= bytes;
+       if (entry->bytes == 0) {
+               ctl->free_extents--;
+               if (entry->bitmap) {
+                       kfree(entry->bitmap);
+                       ctl->total_bitmaps--;
+                       ctl->op->recalc_thresholds(ctl);
+               }
+               kmem_cache_free(btrfs_free_space_cachep, entry);
+       }
+
+       spin_unlock(&ctl->tree_lock);
 
        return ret;
 }
 
-int btrfs_remove_free_space_lock(struct btrfs_block_group_cache *block_group,
-                                u64 offset, u64 bytes)
+static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
+                               struct btrfs_free_space *entry,
+                               struct btrfs_free_cluster *cluster,
+                               u64 offset, u64 bytes, u64 min_bytes)
 {
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       unsigned long next_zero;
+       unsigned long i;
+       unsigned long search_bits;
+       unsigned long total_bits;
+       unsigned long found_bits;
+       unsigned long start = 0;
+       unsigned long total_found = 0;
        int ret;
+       bool found = false;
+
+       i = offset_to_bit(entry->offset, block_group->sectorsize,
+                         max_t(u64, offset, entry->offset));
+       search_bits = bytes_to_bits(bytes, block_group->sectorsize);
+       total_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
+
+again:
+       found_bits = 0;
+       for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
+            i < BITS_PER_BITMAP;
+            i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
+               next_zero = find_next_zero_bit(entry->bitmap,
+                                              BITS_PER_BITMAP, i);
+               if (next_zero - i >= search_bits) {
+                       found_bits = next_zero - i;
+                       break;
+               }
+               i = next_zero;
+       }
 
-       ret = __btrfs_remove_free_space(block_group, offset, bytes);
+       if (!found_bits)
+               return -ENOSPC;
 
-       return ret;
+       if (!found) {
+               start = i;
+               found = true;
+       }
+
+       total_found += found_bits;
+
+       if (cluster->max_size < found_bits * block_group->sectorsize)
+               cluster->max_size = found_bits * block_group->sectorsize;
+
+       if (total_found < total_bits) {
+               i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
+               if (i - start > total_bits * 2) {
+                       total_found = 0;
+                       cluster->max_size = 0;
+                       found = false;
+               }
+               goto again;
+       }
+
+       cluster->window_start = start * block_group->sectorsize +
+               entry->offset;
+       rb_erase(&entry->offset_index, &ctl->free_space_offset);
+       ret = tree_insert_offset(&cluster->root, entry->offset,
+                                &entry->offset_index, 1);
+       BUG_ON(ret);
+
+       return 0;
 }
 
-void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
-                          u64 bytes)
+/*
+ * This searches the block group for just extents to fill the cluster with.
+ */
+static noinline int
+setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
+                       struct btrfs_free_cluster *cluster,
+                       struct list_head *bitmaps, u64 offset, u64 bytes,
+                       u64 min_bytes)
 {
-       struct btrfs_free_space *info;
-       struct rb_node *n;
-       int count = 0;
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_free_space *first = NULL;
+       struct btrfs_free_space *entry = NULL;
+       struct btrfs_free_space *prev = NULL;
+       struct btrfs_free_space *last;
+       struct rb_node *node;
+       u64 window_start;
+       u64 window_free;
+       u64 max_extent;
+       u64 max_gap = 128 * 1024;
 
-       for (n = rb_first(&block_group->free_space_offset); n; n = rb_next(n)) {
-               info = rb_entry(n, struct btrfs_free_space, offset_index);
-               if (info->bytes >= bytes)
-                       count++;
+       entry = tree_search_offset(ctl, offset, 0, 1);
+       if (!entry)
+               return -ENOSPC;
+
+       /*
+        * We don't want bitmaps, so just move along until we find a normal
+        * extent entry.
+        */
+       while (entry->bitmap) {
+               if (list_empty(&entry->list))
+                       list_add_tail(&entry->list, bitmaps);
+               node = rb_next(&entry->offset_index);
+               if (!node)
+                       return -ENOSPC;
+               entry = rb_entry(node, struct btrfs_free_space, offset_index);
        }
-       printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
-              "\n", count);
+
+       window_start = entry->offset;
+       window_free = entry->bytes;
+       max_extent = entry->bytes;
+       first = entry;
+       last = entry;
+       prev = entry;
+
+       while (window_free <= min_bytes) {
+               node = rb_next(&entry->offset_index);
+               if (!node)
+                       return -ENOSPC;
+               entry = rb_entry(node, struct btrfs_free_space, offset_index);
+
+               if (entry->bitmap) {
+                       if (list_empty(&entry->list))
+                               list_add_tail(&entry->list, bitmaps);
+                       continue;
+               }
+
+               /*
+                * we haven't filled the empty size and the window is
+                * very large.  reset and try again
+                */
+               if (entry->offset - (prev->offset + prev->bytes) > max_gap ||
+                   entry->offset - window_start > (min_bytes * 2)) {
+                       first = entry;
+                       window_start = entry->offset;
+                       window_free = entry->bytes;
+                       last = entry;
+                       max_extent = entry->bytes;
+               } else {
+                       last = entry;
+                       window_free += entry->bytes;
+                       if (entry->bytes > max_extent)
+                               max_extent = entry->bytes;
+               }
+               prev = entry;
+       }
+
+       cluster->window_start = first->offset;
+
+       node = &first->offset_index;
+
+       /*
+        * now we've found our entries, pull them out of the free space
+        * cache and put them into the cluster rbtree
+        */
+       do {
+               int ret;
+
+               entry = rb_entry(node, struct btrfs_free_space, offset_index);
+               node = rb_next(&entry->offset_index);
+               if (entry->bitmap)
+                       continue;
+
+               rb_erase(&entry->offset_index, &ctl->free_space_offset);
+               ret = tree_insert_offset(&cluster->root, entry->offset,
+                                        &entry->offset_index, 0);
+               BUG_ON(ret);
+       } while (node && entry != last);
+
+       cluster->max_size = max_extent;
+
+       return 0;
 }
 
-u64 btrfs_block_group_free_space(struct btrfs_block_group_cache *block_group)
+/*
+ * This specifically looks for bitmaps that may work in the cluster, we assume
+ * that we have already failed to find extents that will work.
+ */
+static noinline int
+setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
+                    struct btrfs_free_cluster *cluster,
+                    struct list_head *bitmaps, u64 offset, u64 bytes,
+                    u64 min_bytes)
 {
-       struct btrfs_free_space *info;
-       struct rb_node *n;
-       u64 ret = 0;
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_free_space *entry;
+       struct rb_node *node;
+       int ret = -ENOSPC;
 
-       for (n = rb_first(&block_group->free_space_offset); n;
-            n = rb_next(n)) {
-               info = rb_entry(n, struct btrfs_free_space, offset_index);
-               ret += info->bytes;
+       if (ctl->total_bitmaps == 0)
+               return -ENOSPC;
+
+       /*
+        * First check our cached list of bitmaps and see if there is an entry
+        * here that will work.
+        */
+       list_for_each_entry(entry, bitmaps, list) {
+               if (entry->bytes < min_bytes)
+                       continue;
+               ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
+                                          bytes, min_bytes);
+               if (!ret)
+                       return 0;
+       }
+
+       /*
+        * If we do have entries on our list and we are here then we didn't find
+        * anything, so go ahead and get the next entry after the last entry in
+        * this list and start the search from there.
+        */
+       if (!list_empty(bitmaps)) {
+               entry = list_entry(bitmaps->prev, struct btrfs_free_space,
+                                  list);
+               node = rb_next(&entry->offset_index);
+               if (!node)
+                       return -ENOSPC;
+               entry = rb_entry(node, struct btrfs_free_space, offset_index);
+               goto search;
        }
 
+       entry = tree_search_offset(ctl, offset_to_bitmap(ctl, offset), 0, 1);
+       if (!entry)
+               return -ENOSPC;
+
+search:
+       node = &entry->offset_index;
+       do {
+               entry = rb_entry(node, struct btrfs_free_space, offset_index);
+               node = rb_next(&entry->offset_index);
+               if (!entry->bitmap)
+                       continue;
+               if (entry->bytes < min_bytes)
+                       continue;
+               ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
+                                          bytes, min_bytes);
+       } while (ret && node);
+
        return ret;
 }
 
-void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
+/*
+ * here we try to find a cluster of blocks in a block group.  The goal
+ * is to find at least bytes free and up to empty_size + bytes free.
+ * We might not find them all in one contiguous area.
+ *
+ * returns zero and sets up cluster if things worked out, otherwise
+ * it returns -enospc
+ */
+int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
+                            struct btrfs_root *root,
+                            struct btrfs_block_group_cache *block_group,
+                            struct btrfs_free_cluster *cluster,
+                            u64 offset, u64 bytes, u64 empty_size)
 {
-       struct btrfs_free_space *info;
-       struct rb_node *node;
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct list_head bitmaps;
+       struct btrfs_free_space *entry, *tmp;
+       u64 min_bytes;
+       int ret;
 
-       mutex_lock(&block_group->alloc_mutex);
-       while ((node = rb_last(&block_group->free_space_bytes)) != NULL) {
-               info = rb_entry(node, struct btrfs_free_space, bytes_index);
-               unlink_free_space(block_group, info);
-               kfree(info);
-               if (need_resched()) {
-                       mutex_unlock(&block_group->alloc_mutex);
-                       cond_resched();
-                       mutex_lock(&block_group->alloc_mutex);
-               }
+       /* for metadata, allow allocates with more holes */
+       if (btrfs_test_opt(root, SSD_SPREAD)) {
+               min_bytes = bytes + empty_size;
+       } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
+               /*
+                * we want to do larger allocations when we are
+                * flushing out the delayed refs, it helps prevent
+                * making more work as we go along.
+                */
+               if (trans->transaction->delayed_refs.flushing)
+                       min_bytes = max(bytes, (bytes + empty_size) >> 1);
+               else
+                       min_bytes = max(bytes, (bytes + empty_size) >> 4);
+       } else
+               min_bytes = max(bytes, (bytes + empty_size) >> 2);
+
+       spin_lock(&ctl->tree_lock);
+
+       /*
+        * If we know we don't have enough space to make a cluster don't even
+        * bother doing all the work to try and find one.
+        */
+       if (ctl->free_space < min_bytes) {
+               spin_unlock(&ctl->tree_lock);
+               return -ENOSPC;
        }
-       mutex_unlock(&block_group->alloc_mutex);
+
+       spin_lock(&cluster->lock);
+
+       /* someone already found a cluster, hooray */
+       if (cluster->block_group) {
+               ret = 0;
+               goto out;
+       }
+
+       INIT_LIST_HEAD(&bitmaps);
+       ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
+                                     bytes, min_bytes);
+       if (ret)
+               ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
+                                          offset, bytes, min_bytes);
+
+       /* Clear our temporary list */
+       list_for_each_entry_safe(entry, tmp, &bitmaps, list)
+               list_del_init(&entry->list);
+
+       if (!ret) {
+               atomic_inc(&block_group->count);
+               list_add_tail(&cluster->block_group_list,
+                             &block_group->cluster_list);
+               cluster->block_group = block_group;
+       }
+out:
+       spin_unlock(&cluster->lock);
+       spin_unlock(&ctl->tree_lock);
+
+       return ret;
 }
 
-#if 0
-static struct btrfs_free_space *btrfs_find_free_space_offset(struct
-                                                     btrfs_block_group_cache
-                                                     *block_group, u64 offset,
-                                                     u64 bytes)
+/*
+ * simple code to zero out a cluster
+ */
+void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
 {
-       struct btrfs_free_space *ret;
+       spin_lock_init(&cluster->lock);
+       spin_lock_init(&cluster->refill_lock);
+       cluster->root = RB_ROOT;
+       cluster->max_size = 0;
+       INIT_LIST_HEAD(&cluster->block_group_list);
+       cluster->block_group = NULL;
+}
+
+int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
+                          u64 *trimmed, u64 start, u64 end, u64 minlen)
+{
+       struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
+       struct btrfs_free_space *entry = NULL;
+       struct btrfs_fs_info *fs_info = block_group->fs_info;
+       u64 bytes = 0;
+       u64 actually_trimmed;
+       int ret = 0;
+
+       *trimmed = 0;
+
+       while (start < end) {
+               spin_lock(&ctl->tree_lock);
+
+               if (ctl->free_space < minlen) {
+                       spin_unlock(&ctl->tree_lock);
+                       break;
+               }
+
+               entry = tree_search_offset(ctl, start, 0, 1);
+               if (!entry)
+                       entry = tree_search_offset(ctl,
+                                                  offset_to_bitmap(ctl, start),
+                                                  1, 1);
+
+               if (!entry || entry->offset >= end) {
+                       spin_unlock(&ctl->tree_lock);
+                       break;
+               }
+
+               if (entry->bitmap) {
+                       ret = search_bitmap(ctl, entry, &start, &bytes);
+                       if (!ret) {
+                               if (start >= end) {
+                                       spin_unlock(&ctl->tree_lock);
+                                       break;
+                               }
+                               bytes = min(bytes, end - start);
+                               bitmap_clear_bits(ctl, entry, start, bytes);
+                               if (entry->bytes == 0)
+                                       free_bitmap(ctl, entry);
+                       } else {
+                               start = entry->offset + BITS_PER_BITMAP *
+                                       block_group->sectorsize;
+                               spin_unlock(&ctl->tree_lock);
+                               ret = 0;
+                               continue;
+                       }
+               } else {
+                       start = entry->offset;
+                       bytes = min(entry->bytes, end - start);
+                       unlink_free_space(ctl, entry);
+                       kmem_cache_free(btrfs_free_space_cachep, entry);
+               }
 
-       mutex_lock(&block_group->alloc_mutex);
-       ret = tree_search_offset(&block_group->free_space_offset, offset,
-                                bytes, 0);
-       mutex_unlock(&block_group->alloc_mutex);
+               spin_unlock(&ctl->tree_lock);
+
+               if (bytes >= minlen) {
+                       int update_ret;
+                       update_ret = btrfs_update_reserved_bytes(block_group,
+                                                                bytes, 1, 1);
+
+                       ret = btrfs_error_discard_extent(fs_info->extent_root,
+                                                        start,
+                                                        bytes,
+                                                        &actually_trimmed);
+
+                       btrfs_add_free_space(block_group, start, bytes);
+                       if (!update_ret)
+                               btrfs_update_reserved_bytes(block_group,
+                                                           bytes, 0, 1);
+
+                       if (ret)
+                               break;
+                       *trimmed += actually_trimmed;
+               }
+               start += bytes;
+               bytes = 0;
+
+               if (fatal_signal_pending(current)) {
+                       ret = -ERESTARTSYS;
+                       break;
+               }
+
+               cond_resched();
+       }
 
        return ret;
 }
 
-static struct btrfs_free_space *btrfs_find_free_space_bytes(struct
-                                                    btrfs_block_group_cache
-                                                    *block_group, u64 offset,
-                                                    u64 bytes)
+/*
+ * Find the left-most item in the cache tree, and then return the
+ * smallest inode number in the item.
+ *
+ * Note: the returned inode number may not be the smallest one in
+ * the tree, if the left-most item is a bitmap.
+ */
+u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
+{
+       struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
+       struct btrfs_free_space *entry = NULL;
+       u64 ino = 0;
+
+       spin_lock(&ctl->tree_lock);
+
+       if (RB_EMPTY_ROOT(&ctl->free_space_offset))
+               goto out;
+
+       entry = rb_entry(rb_first(&ctl->free_space_offset),
+                        struct btrfs_free_space, offset_index);
+
+       if (!entry->bitmap) {
+               ino = entry->offset;
+
+               unlink_free_space(ctl, entry);
+               entry->offset++;
+               entry->bytes--;
+               if (!entry->bytes)
+                       kmem_cache_free(btrfs_free_space_cachep, entry);
+               else
+                       link_free_space(ctl, entry);
+       } else {
+               u64 offset = 0;
+               u64 count = 1;
+               int ret;
+
+               ret = search_bitmap(ctl, entry, &offset, &count);
+               BUG_ON(ret);
+
+               ino = offset;
+               bitmap_clear_bits(ctl, entry, offset, 1);
+               if (entry->bytes == 0)
+                       free_bitmap(ctl, entry);
+       }
+out:
+       spin_unlock(&ctl->tree_lock);
+
+       return ino;
+}
+
+struct inode *lookup_free_ino_inode(struct btrfs_root *root,
+                                   struct btrfs_path *path)
+{
+       struct inode *inode = NULL;
+
+       spin_lock(&root->cache_lock);
+       if (root->cache_inode)
+               inode = igrab(root->cache_inode);
+       spin_unlock(&root->cache_lock);
+       if (inode)
+               return inode;
+
+       inode = __lookup_free_space_inode(root, path, 0);
+       if (IS_ERR(inode))
+               return inode;
+
+       spin_lock(&root->cache_lock);
+       if (!btrfs_fs_closing(root->fs_info))
+               root->cache_inode = igrab(inode);
+       spin_unlock(&root->cache_lock);
+
+       return inode;
+}
+
+int create_free_ino_inode(struct btrfs_root *root,
+                         struct btrfs_trans_handle *trans,
+                         struct btrfs_path *path)
+{
+       return __create_free_space_inode(root, trans, path,
+                                        BTRFS_FREE_INO_OBJECTID, 0);
+}
+
+int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
 {
-       struct btrfs_free_space *ret;
+       struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
+       struct btrfs_path *path;
+       struct inode *inode;
+       int ret = 0;
+       u64 root_gen = btrfs_root_generation(&root->root_item);
 
-       mutex_lock(&block_group->alloc_mutex);
+       if (!btrfs_test_opt(root, INODE_MAP_CACHE))
+               return 0;
 
-       ret = tree_search_bytes(&block_group->free_space_bytes, offset, bytes);
-       mutex_unlock(&block_group->alloc_mutex);
+       /*
+        * If we're unmounting then just return, since this does a search on the
+        * normal root and not the commit root and we could deadlock.
+        */
+       if (btrfs_fs_closing(fs_info))
+               return 0;
+
+       path = btrfs_alloc_path();
+       if (!path)
+               return 0;
+
+       inode = lookup_free_ino_inode(root, path);
+       if (IS_ERR(inode))
+               goto out;
+
+       if (root_gen != BTRFS_I(inode)->generation)
+               goto out_put;
 
+       ret = __load_free_space_cache(root, inode, ctl, path, 0);
+
+       if (ret < 0)
+               printk(KERN_ERR "btrfs: failed to load free ino cache for "
+                      "root %llu\n", root->root_key.objectid);
+out_put:
+       iput(inode);
+out:
+       btrfs_free_path(path);
        return ret;
 }
-#endif
 
-struct btrfs_free_space *btrfs_find_free_space(struct btrfs_block_group_cache
-                                              *block_group, u64 offset,
-                                              u64 bytes)
+int btrfs_write_out_ino_cache(struct btrfs_root *root,
+                             struct btrfs_trans_handle *trans,
+                             struct btrfs_path *path)
 {
-       struct btrfs_free_space *ret = NULL;
+       struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
+       struct inode *inode;
+       int ret;
 
-       ret = tree_search_offset(&block_group->free_space_offset, offset,
-                                bytes, 0);
-       if (!ret)
-               ret = tree_search_bytes(&block_group->free_space_bytes,
-                                       offset, bytes);
+       if (!btrfs_test_opt(root, INODE_MAP_CACHE))
+               return 0;
+
+       inode = lookup_free_ino_inode(root, path);
+       if (IS_ERR(inode))
+               return 0;
+
+       ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
+       if (ret < 0)
+               printk(KERN_ERR "btrfs: failed to write free ino cache "
+                      "for root %llu\n", root->root_key.objectid);
 
+       iput(inode);
        return ret;
 }