Btrfs: do extent allocation and reference count updates in the background
Chris Mason [Fri, 13 Mar 2009 14:10:06 +0000 (10:10 -0400)]
The extent allocation tree maintains a reference count and full
back reference information for every extent allocated in the
filesystem.  For subvolume and snapshot trees, every time
a block goes through COW, the new copy of the block adds a reference
on every block it points to.

If a btree node points to 150 leaves, then the COW code needs to go
and add backrefs on 150 different extents, which might be spread all
over the extent allocation tree.

These updates currently happen during btrfs_cow_block, and most COWs
happen during btrfs_search_slot.  btrfs_search_slot has locks held
on both the parent and the node we are COWing, and so we really want
to avoid IO during the COW if we can.

This commit adds an rbtree of pending reference count updates and extent
allocations.  The tree is ordered by byte number of the extent and byte number
of the parent for the back reference.  The tree allows us to:

1) Modify back references in something close to disk order, reducing seeks
2) Significantly reduce the number of modifications made as block pointers
are balanced around
3) Do all of the extent insertion and back reference modifications outside
of the performance critical btrfs_search_slot code.

#3 has the added benefit of greatly reducing the btrfs stack footprint.
The extent allocation tree modifications are done without the deep
(and somewhat recursive) call chains used in the past.

These delayed back reference updates must be done before the transaction
commits, and so the rbtree is tied to the transaction.  Throttling is
implemented to help keep the queue of backrefs at a reasonable size.

Since there was a similar mechanism in place for the extent tree
extents, that is removed and replaced by the delayed reference tree.

Yan Zheng <yan.zheng@oracle.com> helped review and fixup this code.

Signed-off-by: Chris Mason <chris.mason@oracle.com>

fs/btrfs/Makefile
fs/btrfs/ctree.c
fs/btrfs/ctree.h
fs/btrfs/delayed-ref.c [new file with mode: 0644]
fs/btrfs/delayed-ref.h [new file with mode: 0644]
fs/btrfs/disk-io.c
fs/btrfs/extent-tree.c
fs/btrfs/file.c
fs/btrfs/transaction.c
fs/btrfs/transaction.h
fs/btrfs/tree-defrag.c

index d2cf5a5..9adf5e4 100644 (file)
@@ -8,7 +8,7 @@ btrfs-y := super.o ctree.o extent-tree.o print-tree.o root-tree.o dir-item.o \
           extent_map.o sysfs.o struct-funcs.o xattr.o ordered-data.o \
           extent_io.o volumes.o async-thread.o ioctl.o locking.o orphan.o \
           ref-cache.o export.o tree-log.o acl.o free-space-cache.o zlib.o \
-          compression.o
+          compression.o delayed-ref.o
 else
 
 # Normal Makefile
index 87c9038..bebc9fd 100644 (file)
@@ -922,6 +922,7 @@ static noinline int balance_level(struct btrfs_trans_handle *trans,
                spin_unlock(&root->node_lock);
 
                ret = btrfs_update_extent_ref(trans, root, child->start,
+                                             child->len,
                                              mid->start, child->start,
                                              root->root_key.objectid,
                                              trans->transid, level - 1);
@@ -2075,7 +2076,7 @@ static noinline int insert_new_root(struct btrfs_trans_handle *trans,
        spin_unlock(&root->node_lock);
 
        ret = btrfs_update_extent_ref(trans, root, lower->start,
-                                     lower->start, c->start,
+                                     lower->len, lower->start, c->start,
                                      root->root_key.objectid,
                                      trans->transid, level - 1);
        BUG_ON(ret);
index 3a37ba7..ced5fd8 100644 (file)
@@ -688,8 +688,6 @@ struct btrfs_fs_info {
        struct rb_root block_group_cache_tree;
 
        struct extent_io_tree pinned_extents;
-       struct extent_io_tree pending_del;
-       struct extent_io_tree extent_ins;
 
        /* logical->physical extent mapping */
        struct btrfs_mapping_tree mapping_tree;
@@ -717,7 +715,6 @@ struct btrfs_fs_info {
        struct mutex tree_log_mutex;
        struct mutex transaction_kthread_mutex;
        struct mutex cleaner_mutex;
-       struct mutex extent_ins_mutex;
        struct mutex pinned_mutex;
        struct mutex chunk_mutex;
        struct mutex drop_mutex;
@@ -1704,18 +1701,15 @@ static inline struct dentry *fdentry(struct file *file)
 }
 
 /* extent-tree.c */
+int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
+                          struct btrfs_root *root, unsigned long count);
 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len);
-int btrfs_lookup_extent_ref(struct btrfs_trans_handle *trans,
-                           struct btrfs_root *root, u64 bytenr,
-                           u64 num_bytes, u32 *refs);
 int btrfs_update_pinned_extents(struct btrfs_root *root,
                                u64 bytenr, u64 num, int pin);
 int btrfs_drop_leaf_ref(struct btrfs_trans_handle *trans,
                        struct btrfs_root *root, struct extent_buffer *leaf);
 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
                          struct btrfs_root *root, u64 objectid, u64 bytenr);
-int btrfs_extent_post_op(struct btrfs_trans_handle *trans,
-                        struct btrfs_root *root);
 int btrfs_copy_pinned(struct btrfs_root *root, struct extent_io_tree *copy);
 struct btrfs_block_group_cache *btrfs_lookup_block_group(
                                                 struct btrfs_fs_info *info,
@@ -1777,7 +1771,7 @@ int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
                         u64 root_objectid, u64 ref_generation,
                         u64 owner_objectid);
 int btrfs_update_extent_ref(struct btrfs_trans_handle *trans,
-                           struct btrfs_root *root, u64 bytenr,
+                           struct btrfs_root *root, u64 bytenr, u64 num_bytes,
                            u64 orig_parent, u64 parent,
                            u64 root_objectid, u64 ref_generation,
                            u64 owner_objectid);
diff --git a/fs/btrfs/delayed-ref.c b/fs/btrfs/delayed-ref.c
new file mode 100644 (file)
index 0000000..874565a
--- /dev/null
@@ -0,0 +1,585 @@
+/*
+ * Copyright (C) 2009 Oracle.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public
+ * License along with this program; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 021110-1307, USA.
+ */
+
+#include <linux/sched.h>
+#include <linux/sort.h>
+#include <linux/ftrace.h>
+#include "ctree.h"
+#include "delayed-ref.h"
+#include "transaction.h"
+
+/*
+ * delayed back reference update tracking.  For subvolume trees
+ * we queue up extent allocations and backref maintenance for
+ * delayed processing.   This avoids deep call chains where we
+ * add extents in the middle of btrfs_search_slot, and it allows
+ * us to buffer up frequently modified backrefs in an rb tree instead
+ * of hammering updates on the extent allocation tree.
+ *
+ * Right now this code is only used for reference counted trees, but
+ * the long term goal is to get rid of the similar code for delayed
+ * extent tree modifications.
+ */
+
+/*
+ * entries in the rb tree are ordered by the byte number of the extent
+ * and by the byte number of the parent block.
+ */
+static int comp_entry(struct btrfs_delayed_ref_node *ref,
+                     u64 bytenr, u64 parent)
+{
+       if (bytenr < ref->bytenr)
+               return -1;
+       if (bytenr > ref->bytenr)
+               return 1;
+       if (parent < ref->parent)
+               return -1;
+       if (parent > ref->parent)
+               return 1;
+       return 0;
+}
+
+/*
+ * insert a new ref into the rbtree.  This returns any existing refs
+ * for the same (bytenr,parent) tuple, or NULL if the new node was properly
+ * inserted.
+ */
+static struct btrfs_delayed_ref_node *tree_insert(struct rb_root *root,
+                                                 u64 bytenr, u64 parent,
+                                                 struct rb_node *node)
+{
+       struct rb_node **p = &root->rb_node;
+       struct rb_node *parent_node = NULL;
+       struct btrfs_delayed_ref_node *entry;
+       int cmp;
+
+       while (*p) {
+               parent_node = *p;
+               entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
+                                rb_node);
+
+               cmp = comp_entry(entry, bytenr, parent);
+               if (cmp < 0)
+                       p = &(*p)->rb_left;
+               else if (cmp > 0)
+                       p = &(*p)->rb_right;
+               else
+                       return entry;
+       }
+
+       entry = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
+       rb_link_node(node, parent_node, p);
+       rb_insert_color(node, root);
+       return NULL;
+}
+
+/*
+ * find an entry based on (bytenr,parent).  This returns the delayed
+ * ref if it was able to find one, or NULL if nothing was in that spot
+ */
+static struct btrfs_delayed_ref_node *tree_search(struct rb_root *root,
+                                                 u64 bytenr, u64 parent)
+{
+       struct rb_node *n = root->rb_node;
+       struct btrfs_delayed_ref_node *entry;
+       int cmp;
+
+       while (n) {
+               entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
+               WARN_ON(!entry->in_tree);
+
+               cmp = comp_entry(entry, bytenr, parent);
+               if (cmp < 0)
+                       n = n->rb_left;
+               else if (cmp > 0)
+                       n = n->rb_right;
+               else
+                       return entry;
+       }
+       return NULL;
+}
+
+/*
+ * Locking on delayed refs is done by taking a lock on the head node,
+ * which has the (impossible) parent id of (u64)-1.  Once a lock is held
+ * on the head node, you're allowed (and required) to process all the
+ * delayed refs for a given byte number in the tree.
+ *
+ * This will walk forward in the rbtree until it finds a head node it
+ * is able to lock.  It might not lock the delayed ref you asked for,
+ * and so it will return the one it did lock in next_ret and return 0.
+ *
+ * If no locks are taken, next_ret is set to null and 1 is returned.  This
+ * means there are no more unlocked head nodes in the rbtree.
+ */
+int btrfs_lock_delayed_ref(struct btrfs_trans_handle *trans,
+                          struct btrfs_delayed_ref_node *ref,
+                          struct btrfs_delayed_ref_head **next_ret)
+{
+       struct rb_node *node;
+       struct btrfs_delayed_ref_head *head;
+       int ret = 0;
+
+       while (1) {
+               if (btrfs_delayed_ref_is_head(ref)) {
+                       head = btrfs_delayed_node_to_head(ref);
+                       if (mutex_trylock(&head->mutex)) {
+                               *next_ret = head;
+                               ret = 0;
+                               break;
+                       }
+               }
+               node = rb_next(&ref->rb_node);
+               if (!node) {
+                       ret = 1;
+                       *next_ret = NULL;
+                       break;
+               }
+               ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
+       }
+       return ret;
+}
+
+/*
+ * This checks to see if there are any delayed refs in the
+ * btree for a given bytenr.  It returns one if it finds any
+ * and zero otherwise.
+ *
+ * If it only finds a head node, it returns 0.
+ *
+ * The idea is to use this when deciding if you can safely delete an
+ * extent from the extent allocation tree.  There may be a pending
+ * ref in the rbtree that adds or removes references, so as long as this
+ * returns one you need to leave the BTRFS_EXTENT_ITEM in the extent
+ * allocation tree.
+ */
+int btrfs_delayed_ref_pending(struct btrfs_trans_handle *trans, u64 bytenr)
+{
+       struct btrfs_delayed_ref_node *ref;
+       struct btrfs_delayed_ref_root *delayed_refs;
+       struct rb_node *prev_node;
+       int ret = 0;
+
+       delayed_refs = &trans->transaction->delayed_refs;
+       spin_lock(&delayed_refs->lock);
+
+       ref = tree_search(&delayed_refs->root, bytenr, (u64)-1);
+       if (ref) {
+               prev_node = rb_prev(&ref->rb_node);
+               if (!prev_node)
+                       goto out;
+               ref = rb_entry(prev_node, struct btrfs_delayed_ref_node,
+                              rb_node);
+               if (ref->bytenr == bytenr)
+                       ret = 1;
+       }
+out:
+       spin_unlock(&delayed_refs->lock);
+       return ret;
+}
+
+/*
+ * helper function to lookup reference count
+ *
+ * the head node for delayed ref is used to store the sum of all the
+ * reference count modifications queued up in the rbtree.  This way you
+ * can check to see what the reference count would be if all of the
+ * delayed refs are processed.
+ */
+int btrfs_lookup_extent_ref(struct btrfs_trans_handle *trans,
+                           struct btrfs_root *root, u64 bytenr,
+                           u64 num_bytes, u32 *refs)
+{
+       struct btrfs_delayed_ref_node *ref;
+       struct btrfs_delayed_ref_head *head;
+       struct btrfs_delayed_ref_root *delayed_refs;
+       struct btrfs_path *path;
+       struct extent_buffer *leaf;
+       struct btrfs_extent_item *ei;
+       struct btrfs_key key;
+       u32 num_refs;
+       int ret;
+
+       path = btrfs_alloc_path();
+       if (!path)
+               return -ENOMEM;
+
+       key.objectid = bytenr;
+       key.type = BTRFS_EXTENT_ITEM_KEY;
+       key.offset = num_bytes;
+       delayed_refs = &trans->transaction->delayed_refs;
+again:
+       ret = btrfs_search_slot(trans, root->fs_info->extent_root,
+                               &key, path, 0, 0);
+       if (ret < 0)
+               goto out;
+
+       if (ret == 0) {
+               leaf = path->nodes[0];
+               ei = btrfs_item_ptr(leaf, path->slots[0],
+                                   struct btrfs_extent_item);
+               num_refs = btrfs_extent_refs(leaf, ei);
+       } else {
+               num_refs = 0;
+               ret = 0;
+       }
+
+       spin_lock(&delayed_refs->lock);
+       ref = tree_search(&delayed_refs->root, bytenr, (u64)-1);
+       if (ref) {
+               head = btrfs_delayed_node_to_head(ref);
+               if (mutex_trylock(&head->mutex)) {
+                       num_refs += ref->ref_mod;
+                       mutex_unlock(&head->mutex);
+                       *refs = num_refs;
+                       goto out;
+               }
+
+               atomic_inc(&ref->refs);
+               spin_unlock(&delayed_refs->lock);
+
+               btrfs_release_path(root->fs_info->extent_root, path);
+
+               mutex_lock(&head->mutex);
+               mutex_unlock(&head->mutex);
+               btrfs_put_delayed_ref(ref);
+               goto again;
+       } else {
+               *refs = num_refs;
+       }
+out:
+       spin_unlock(&delayed_refs->lock);
+       btrfs_free_path(path);
+       return ret;
+}
+
+/*
+ * helper function to update an extent delayed ref in the
+ * rbtree.  existing and update must both have the same
+ * bytenr and parent
+ *
+ * This may free existing if the update cancels out whatever
+ * operation it was doing.
+ */
+static noinline void
+update_existing_ref(struct btrfs_trans_handle *trans,
+                   struct btrfs_delayed_ref_root *delayed_refs,
+                   struct btrfs_delayed_ref_node *existing,
+                   struct btrfs_delayed_ref_node *update)
+{
+       struct btrfs_delayed_ref *existing_ref;
+       struct btrfs_delayed_ref *ref;
+
+       existing_ref = btrfs_delayed_node_to_ref(existing);
+       ref = btrfs_delayed_node_to_ref(update);
+
+       if (ref->pin)
+               existing_ref->pin = 1;
+
+       if (ref->action != existing_ref->action) {
+               /*
+                * this is effectively undoing either an add or a
+                * drop.  We decrement the ref_mod, and if it goes
+                * down to zero we just delete the entry without
+                * every changing the extent allocation tree.
+                */
+               existing->ref_mod--;
+               if (existing->ref_mod == 0) {
+                       rb_erase(&existing->rb_node,
+                                &delayed_refs->root);
+                       existing->in_tree = 0;
+                       btrfs_put_delayed_ref(existing);
+                       delayed_refs->num_entries--;
+                       if (trans->delayed_ref_updates)
+                               trans->delayed_ref_updates--;
+               }
+       } else {
+               if (existing_ref->action == BTRFS_ADD_DELAYED_REF) {
+                       /* if we're adding refs, make sure all the
+                        * details match up.  The extent could
+                        * have been totally freed and reallocated
+                        * by a different owner before the delayed
+                        * ref entries were removed.
+                        */
+                       existing_ref->owner_objectid = ref->owner_objectid;
+                       existing_ref->generation = ref->generation;
+                       existing_ref->root = ref->root;
+                       existing->num_bytes = update->num_bytes;
+               }
+               /*
+                * the action on the existing ref matches
+                * the action on the ref we're trying to add.
+                * Bump the ref_mod by one so the backref that
+                * is eventually added/removed has the correct
+                * reference count
+                */
+               existing->ref_mod += update->ref_mod;
+       }
+}
+
+/*
+ * helper function to update the accounting in the head ref
+ * existing and update must have the same bytenr
+ */
+static noinline void
+update_existing_head_ref(struct btrfs_delayed_ref_node *existing,
+                        struct btrfs_delayed_ref_node *update)
+{
+       struct btrfs_delayed_ref_head *existing_ref;
+       struct btrfs_delayed_ref_head *ref;
+
+       existing_ref = btrfs_delayed_node_to_head(existing);
+       ref = btrfs_delayed_node_to_head(update);
+
+       if (ref->must_insert_reserved) {
+               /* if the extent was freed and then
+                * reallocated before the delayed ref
+                * entries were processed, we can end up
+                * with an existing head ref without
+                * the must_insert_reserved flag set.
+                * Set it again here
+                */
+               existing_ref->must_insert_reserved = ref->must_insert_reserved;
+
+               /*
+                * update the num_bytes so we make sure the accounting
+                * is done correctly
+                */
+               existing->num_bytes = update->num_bytes;
+
+       }
+
+       /*
+        * update the reference mod on the head to reflect this new operation
+        */
+       existing->ref_mod += update->ref_mod;
+}
+
+/*
+ * helper function to actually insert a delayed ref into the rbtree.
+ * this does all the dirty work in terms of maintaining the correct
+ * overall modification count in the head node and properly dealing
+ * with updating existing nodes as new modifications are queued.
+ */
+static noinline int __btrfs_add_delayed_ref(struct btrfs_trans_handle *trans,
+                         struct btrfs_delayed_ref_node *ref,
+                         u64 bytenr, u64 num_bytes, u64 parent, u64 ref_root,
+                         u64 ref_generation, u64 owner_objectid, int action,
+                         int pin)
+{
+       struct btrfs_delayed_ref_node *existing;
+       struct btrfs_delayed_ref *full_ref;
+       struct btrfs_delayed_ref_head *head_ref;
+       struct btrfs_delayed_ref_root *delayed_refs;
+       int count_mod = 1;
+       int must_insert_reserved = 0;
+
+       /*
+        * the head node stores the sum of all the mods, so dropping a ref
+        * should drop the sum in the head node by one.
+        */
+       if (parent == (u64)-1 && action == BTRFS_DROP_DELAYED_REF)
+               count_mod = -1;
+
+       /*
+        * BTRFS_ADD_DELAYED_EXTENT means that we need to update
+        * the reserved accounting when the extent is finally added, or
+        * if a later modification deletes the delayed ref without ever
+        * inserting the extent into the extent allocation tree.
+        * ref->must_insert_reserved is the flag used to record
+        * that accounting mods are required.
+        *
+        * Once we record must_insert_reserved, switch the action to
+        * BTRFS_ADD_DELAYED_REF because other special casing is not required.
+        */
+       if (action == BTRFS_ADD_DELAYED_EXTENT) {
+               must_insert_reserved = 1;
+               action = BTRFS_ADD_DELAYED_REF;
+       } else {
+               must_insert_reserved = 0;
+       }
+
+
+       delayed_refs = &trans->transaction->delayed_refs;
+
+       /* first set the basic ref node struct up */
+       atomic_set(&ref->refs, 1);
+       ref->bytenr = bytenr;
+       ref->parent = parent;
+       ref->ref_mod = count_mod;
+       ref->in_tree = 1;
+       ref->num_bytes = num_bytes;
+
+       if (btrfs_delayed_ref_is_head(ref)) {
+               head_ref = btrfs_delayed_node_to_head(ref);
+               head_ref->must_insert_reserved = must_insert_reserved;
+               mutex_init(&head_ref->mutex);
+       } else {
+               full_ref = btrfs_delayed_node_to_ref(ref);
+               full_ref->root = ref_root;
+               full_ref->generation = ref_generation;
+               full_ref->owner_objectid = owner_objectid;
+               full_ref->pin = pin;
+               full_ref->action = action;
+       }
+
+       existing = tree_insert(&delayed_refs->root, bytenr,
+                              parent, &ref->rb_node);
+
+       if (existing) {
+               if (btrfs_delayed_ref_is_head(ref))
+                       update_existing_head_ref(existing, ref);
+               else
+                       update_existing_ref(trans, delayed_refs, existing, ref);
+
+               /*
+                * we've updated the existing ref, free the newly
+                * allocated ref
+                */
+               kfree(ref);
+       } else {
+               delayed_refs->num_entries++;
+               trans->delayed_ref_updates++;
+       }
+       return 0;
+}
+
+/*
+ * add a delayed ref to the tree.  This does all of the accounting required
+ * to make sure the delayed ref is eventually processed before this
+ * transaction commits.
+ */
+int btrfs_add_delayed_ref(struct btrfs_trans_handle *trans,
+                         u64 bytenr, u64 num_bytes, u64 parent, u64 ref_root,
+                         u64 ref_generation, u64 owner_objectid, int action,
+                         int pin)
+{
+       struct btrfs_delayed_ref *ref;
+       struct btrfs_delayed_ref_head *head_ref;
+       struct btrfs_delayed_ref_root *delayed_refs;
+       int ret;
+
+       ref = kmalloc(sizeof(*ref), GFP_NOFS);
+       if (!ref)
+               return -ENOMEM;
+
+       /*
+        * the parent = 0 case comes from cases where we don't actually
+        * know the parent yet.  It will get updated later via a add/drop
+        * pair.
+        */
+       if (parent == 0)
+               parent = bytenr;
+
+       head_ref = kmalloc(sizeof(*head_ref), GFP_NOFS);
+       if (!head_ref) {
+               kfree(ref);
+               return -ENOMEM;
+       }
+       delayed_refs = &trans->transaction->delayed_refs;
+       spin_lock(&delayed_refs->lock);
+
+       /*
+        * insert both the head node and the new ref without dropping
+        * the spin lock
+        */
+       ret = __btrfs_add_delayed_ref(trans, &head_ref->node, bytenr, num_bytes,
+                                     (u64)-1, 0, 0, 0, action, pin);
+       BUG_ON(ret);
+
+       ret = __btrfs_add_delayed_ref(trans, &ref->node, bytenr, num_bytes,
+                                     parent, ref_root, ref_generation,
+                                     owner_objectid, action, pin);
+       BUG_ON(ret);
+       spin_unlock(&delayed_refs->lock);
+       return 0;
+}
+
+/*
+ * add a delayed ref to the tree.  This does all of the accounting required
+ * to make sure the delayed ref is eventually processed before this
+ * transaction commits.
+ *
+ * The main point of this call is to add and remove a backreference in a single
+ * shot, taking the lock only once, and only searching for the head node once.
+ *
+ * It is the same as doing a ref add and delete in two separate calls.
+ */
+int btrfs_update_delayed_ref(struct btrfs_trans_handle *trans,
+                         u64 bytenr, u64 num_bytes, u64 orig_parent,
+                         u64 parent, u64 orig_ref_root, u64 ref_root,
+                         u64 orig_ref_generation, u64 ref_generation,
+                         u64 owner_objectid, int pin)
+{
+       struct btrfs_delayed_ref *ref;
+       struct btrfs_delayed_ref *old_ref;
+       struct btrfs_delayed_ref_head *head_ref;
+       struct btrfs_delayed_ref_root *delayed_refs;
+       int ret;
+
+       ref = kmalloc(sizeof(*ref), GFP_NOFS);
+       if (!ref)
+               return -ENOMEM;
+
+       old_ref = kmalloc(sizeof(*old_ref), GFP_NOFS);
+       if (!old_ref) {
+               kfree(ref);
+               return -ENOMEM;
+       }
+
+       /*
+        * the parent = 0 case comes from cases where we don't actually
+        * know the parent yet.  It will get updated later via a add/drop
+        * pair.
+        */
+       if (parent == 0)
+               parent = bytenr;
+       if (orig_parent == 0)
+               orig_parent = bytenr;
+
+       head_ref = kmalloc(sizeof(*head_ref), GFP_NOFS);
+       if (!head_ref) {
+               kfree(ref);
+               kfree(old_ref);
+               return -ENOMEM;
+       }
+       delayed_refs = &trans->transaction->delayed_refs;
+       spin_lock(&delayed_refs->lock);
+
+       /*
+        * insert both the head node and the new ref without dropping
+        * the spin lock
+        */
+       ret = __btrfs_add_delayed_ref(trans, &head_ref->node, bytenr, num_bytes,
+                                     (u64)-1, 0, 0, 0,
+                                     BTRFS_ADD_DELAYED_REF, 0);
+       BUG_ON(ret);
+
+       ret = __btrfs_add_delayed_ref(trans, &ref->node, bytenr, num_bytes,
+                                     parent, ref_root, ref_generation,
+                                     owner_objectid, BTRFS_ADD_DELAYED_REF, 0);
+       BUG_ON(ret);
+
+       ret = __btrfs_add_delayed_ref(trans, &old_ref->node, bytenr, num_bytes,
+                                     orig_parent, orig_ref_root,
+                                     orig_ref_generation, owner_objectid,
+                                     BTRFS_DROP_DELAYED_REF, pin);
+       BUG_ON(ret);
+       spin_unlock(&delayed_refs->lock);
+       return 0;
+}
diff --git a/fs/btrfs/delayed-ref.h b/fs/btrfs/delayed-ref.h
new file mode 100644 (file)
index 0000000..37919e5
--- /dev/null
@@ -0,0 +1,182 @@
+/*
+ * Copyright (C) 2008 Oracle.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public
+ * License along with this program; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 021110-1307, USA.
+ */
+#ifndef __DELAYED_REF__
+#define __DELAYED_REF__
+
+/* these are the possible values of struct btrfs_delayed_ref->action */
+#define BTRFS_ADD_DELAYED_REF    1 /* add one backref to the tree */
+#define BTRFS_DROP_DELAYED_REF   2 /* delete one backref from the tree */
+#define BTRFS_ADD_DELAYED_EXTENT 3 /* record a full extent allocation */
+
+struct btrfs_delayed_ref_node {
+       struct rb_node rb_node;
+
+       /* the starting bytenr of the extent */
+       u64 bytenr;
+
+       /* the parent our backref will point to */
+       u64 parent;
+
+       /* the size of the extent */
+       u64 num_bytes;
+
+       /* ref count on this data structure */
+       atomic_t refs;
+
+       /*
+        * how many refs is this entry adding or deleting.  For
+        * head refs, this may be a negative number because it is keeping
+        * track of the total mods done to the reference count.
+        * For individual refs, this will always be a positive number
+        *
+        * It may be more than one, since it is possible for a single
+        * parent to have more than one ref on an extent
+        */
+       int ref_mod;
+
+       /* is this node still in the rbtree? */
+       unsigned int in_tree:1;
+};
+
+/*
+ * the head refs are used to hold a lock on a given extent, which allows us
+ * to make sure that only one process is running the delayed refs
+ * at a time for a single extent.  They also store the sum of all the
+ * reference count modifications we've queued up.
+ */
+struct btrfs_delayed_ref_head {
+       struct btrfs_delayed_ref_node node;
+
+       /*
+        * the mutex is held while running the refs, and it is also
+        * held when checking the sum of reference modifications.
+        */
+       struct mutex mutex;
+
+       /*
+        * when a new extent is allocated, it is just reserved in memory
+        * The actual extent isn't inserted into the extent allocation tree
+        * until the delayed ref is processed.  must_insert_reserved is
+        * used to flag a delayed ref so the accounting can be updated
+        * when a full insert is done.
+        *
+        * It is possible the extent will be freed before it is ever
+        * inserted into the extent allocation tree.  In this case
+        * we need to update the in ram accounting to properly reflect
+        * the free has happened.
+        */
+       unsigned int must_insert_reserved:1;
+};
+
+struct btrfs_delayed_ref {
+       struct btrfs_delayed_ref_node node;
+
+       /* the root objectid our ref will point to */
+       u64 root;
+
+       /* the generation for the backref */
+       u64 generation;
+
+       /* owner_objectid of the backref  */
+       u64 owner_objectid;
+
+       /* operation done by this entry in the rbtree */
+       u8 action;
+
+       /* if pin == 1, when the extent is freed it will be pinned until
+        * transaction commit
+        */
+       unsigned int pin:1;
+};
+
+struct btrfs_delayed_ref_root {
+       struct rb_root root;
+
+       /* this spin lock protects the rbtree and the entries inside */
+       spinlock_t lock;
+
+       /* how many delayed ref updates we've queued, used by the
+        * throttling code
+        */
+       unsigned long num_entries;
+
+       /*
+        * set when the tree is flushing before a transaction commit,
+        * used by the throttling code to decide if new updates need
+        * to be run right away
+        */
+       int flushing;
+};
+
+static inline void btrfs_put_delayed_ref(struct btrfs_delayed_ref_node *ref)
+{
+       WARN_ON(atomic_read(&ref->refs) == 0);
+       if (atomic_dec_and_test(&ref->refs)) {
+               WARN_ON(ref->in_tree);
+               kfree(ref);
+       }
+}
+
+int btrfs_add_delayed_ref(struct btrfs_trans_handle *trans,
+                         u64 bytenr, u64 num_bytes, u64 parent, u64 ref_root,
+                         u64 ref_generation, u64 owner_objectid, int action,
+                         int pin);
+
+struct btrfs_delayed_ref *
+btrfs_find_delayed_ref(struct btrfs_trans_handle *trans, u64 bytenr,
+                      u64 parent);
+int btrfs_delayed_ref_pending(struct btrfs_trans_handle *trans, u64 bytenr);
+int btrfs_lock_delayed_ref(struct btrfs_trans_handle *trans,
+                          struct btrfs_delayed_ref_node *ref,
+                          struct btrfs_delayed_ref_head **next_ret);
+int btrfs_lookup_extent_ref(struct btrfs_trans_handle *trans,
+                           struct btrfs_root *root, u64 bytenr,
+                           u64 num_bytes, u32 *refs);
+int btrfs_update_delayed_ref(struct btrfs_trans_handle *trans,
+                         u64 bytenr, u64 num_bytes, u64 orig_parent,
+                         u64 parent, u64 orig_ref_root, u64 ref_root,
+                         u64 orig_ref_generation, u64 ref_generation,
+                         u64 owner_objectid, int pin);
+/*
+ * a node might live in a head or a regular ref, this lets you
+ * test for the proper type to use.
+ */
+static int btrfs_delayed_ref_is_head(struct btrfs_delayed_ref_node *node)
+{
+       return node->parent == (u64)-1;
+}
+
+/*
+ * helper functions to cast a node into its container
+ */
+static inline struct btrfs_delayed_ref *
+btrfs_delayed_node_to_ref(struct btrfs_delayed_ref_node *node)
+{
+       WARN_ON(btrfs_delayed_ref_is_head(node));
+       return container_of(node, struct btrfs_delayed_ref, node);
+
+}
+
+static inline struct btrfs_delayed_ref_head *
+btrfs_delayed_node_to_head(struct btrfs_delayed_ref_node *node)
+{
+       WARN_ON(!btrfs_delayed_ref_is_head(node));
+       return container_of(node, struct btrfs_delayed_ref_head, node);
+
+}
+#endif
index 3e18175..4f43e22 100644 (file)
@@ -1458,6 +1458,7 @@ static int transaction_kthread(void *arg)
        struct btrfs_root *root = arg;
        struct btrfs_trans_handle *trans;
        struct btrfs_transaction *cur;
+       struct btrfs_fs_info *info = root->fs_info;
        unsigned long now;
        unsigned long delay;
        int ret;
@@ -1471,12 +1472,6 @@ static int transaction_kthread(void *arg)
                vfs_check_frozen(root->fs_info->sb, SB_FREEZE_WRITE);
                mutex_lock(&root->fs_info->transaction_kthread_mutex);
 
-               if (root->fs_info->total_ref_cache_size > 20 * 1024 * 1024) {
-                       printk(KERN_INFO "btrfs: total reference cache "
-                              "size %llu\n",
-                              root->fs_info->total_ref_cache_size);
-               }
-
                mutex_lock(&root->fs_info->trans_mutex);
                cur = root->fs_info->running_transaction;
                if (!cur) {
@@ -1486,13 +1481,30 @@ static int transaction_kthread(void *arg)
 
                now = get_seconds();
                if (now < cur->start_time || now - cur->start_time < 30) {
+                       unsigned long num_delayed;
+                       num_delayed = cur->delayed_refs.num_entries;
                        mutex_unlock(&root->fs_info->trans_mutex);
                        delay = HZ * 5;
+
+                       /*
+                        * we may have been woken up early to start
+                        * processing the delayed extent ref updates
+                        * If so, run some of them and then loop around again
+                        * to see if we need to force a commit
+                        */
+                       if (num_delayed > 64) {
+                               mutex_unlock(&info->transaction_kthread_mutex);
+                               trans = btrfs_start_transaction(root, 1);
+                               btrfs_run_delayed_refs(trans, root, 256);
+                               btrfs_end_transaction(trans, root);
+                               continue;
+                       }
                        goto sleep;
                }
                mutex_unlock(&root->fs_info->trans_mutex);
                trans = btrfs_start_transaction(root, 1);
                ret = btrfs_commit_transaction(trans, root);
+
 sleep:
                wake_up_process(root->fs_info->cleaner_kthread);
                mutex_unlock(&root->fs_info->transaction_kthread_mutex);
@@ -1611,10 +1623,6 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 
        extent_io_tree_init(&fs_info->pinned_extents,
                             fs_info->btree_inode->i_mapping, GFP_NOFS);
-       extent_io_tree_init(&fs_info->pending_del,
-                            fs_info->btree_inode->i_mapping, GFP_NOFS);
-       extent_io_tree_init(&fs_info->extent_ins,
-                            fs_info->btree_inode->i_mapping, GFP_NOFS);
        fs_info->do_barriers = 1;
 
        INIT_LIST_HEAD(&fs_info->dead_reloc_roots);
@@ -1629,7 +1637,6 @@ struct btrfs_root *open_ctree(struct super_block *sb,
        mutex_init(&fs_info->trans_mutex);
        mutex_init(&fs_info->tree_log_mutex);
        mutex_init(&fs_info->drop_mutex);
-       mutex_init(&fs_info->extent_ins_mutex);
        mutex_init(&fs_info->pinned_mutex);
        mutex_init(&fs_info->chunk_mutex);
        mutex_init(&fs_info->transaction_kthread_mutex);
index fefe83a..9b5da2b 100644 (file)
@@ -49,10 +49,13 @@ struct pending_extent_op {
        int del;
 };
 
-static int finish_current_insert(struct btrfs_trans_handle *trans,
-                                struct btrfs_root *extent_root, int all);
-static int del_pending_extents(struct btrfs_trans_handle *trans,
-                              struct btrfs_root *extent_root, int all);
+static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
+                                        struct btrfs_root *root, u64 parent,
+                                        u64 root_objectid, u64 ref_generation,
+                                        u64 owner, struct btrfs_key *ins,
+                                        int ref_mod);
+static int update_reserved_extents(struct btrfs_root *root,
+                                  u64 bytenr, u64 num, int reserve);
 static int pin_down_bytes(struct btrfs_trans_handle *trans,
                          struct btrfs_root *root,
                          u64 bytenr, u64 num_bytes, int is_data);
@@ -60,6 +63,12 @@ static int update_block_group(struct btrfs_trans_handle *trans,
                              struct btrfs_root *root,
                              u64 bytenr, u64 num_bytes, int alloc,
                              int mark_free);
+static noinline int __btrfs_free_extent(struct btrfs_trans_handle *trans,
+                                       struct btrfs_root *root,
+                                       u64 bytenr, u64 num_bytes, u64 parent,
+                                       u64 root_objectid, u64 ref_generation,
+                                       u64 owner_objectid, int pin,
+                                       int ref_to_drop);
 
 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
                          struct btrfs_root *extent_root, u64 alloc_bytes,
@@ -554,262 +563,13 @@ out:
        return ret;
 }
 
-/*
- * updates all the backrefs that are pending on update_list for the
- * extent_root
- */
-static noinline int update_backrefs(struct btrfs_trans_handle *trans,
-                                   struct btrfs_root *extent_root,
-                                   struct btrfs_path *path,
-                                   struct list_head *update_list)
-{
-       struct btrfs_key key;
-       struct btrfs_extent_ref *ref;
-       struct btrfs_fs_info *info = extent_root->fs_info;
-       struct pending_extent_op *op;
-       struct extent_buffer *leaf;
-       int ret = 0;
-       struct list_head *cur = update_list->next;
-       u64 ref_objectid;
-       u64 ref_root = extent_root->root_key.objectid;
-
-       op = list_entry(cur, struct pending_extent_op, list);
-
-search:
-       key.objectid = op->bytenr;
-       key.type = BTRFS_EXTENT_REF_KEY;
-       key.offset = op->orig_parent;
-
-       ret = btrfs_search_slot(trans, extent_root, &key, path, 0, 1);
-       BUG_ON(ret);
-
-       leaf = path->nodes[0];
-
-loop:
-       ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_ref);
-
-       ref_objectid = btrfs_ref_objectid(leaf, ref);
-
-       if (btrfs_ref_root(leaf, ref) != ref_root ||
-           btrfs_ref_generation(leaf, ref) != op->orig_generation ||
-           (ref_objectid != op->level &&
-            ref_objectid != BTRFS_MULTIPLE_OBJECTIDS)) {
-               printk(KERN_ERR "btrfs couldn't find %llu, parent %llu, "
-                      "root %llu, owner %u\n",
-                      (unsigned long long)op->bytenr,
-                      (unsigned long long)op->orig_parent,
-                      (unsigned long long)ref_root, op->level);
-               btrfs_print_leaf(extent_root, leaf);
-               BUG();
-       }
-
-       key.objectid = op->bytenr;
-       key.offset = op->parent;
-       key.type = BTRFS_EXTENT_REF_KEY;
-       ret = btrfs_set_item_key_safe(trans, extent_root, path, &key);
-       BUG_ON(ret);
-       ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_ref);
-       btrfs_set_ref_generation(leaf, ref, op->generation);
-
-       cur = cur->next;
-
-       list_del_init(&op->list);
-       unlock_extent(&info->extent_ins, op->bytenr,
-                     op->bytenr + op->num_bytes - 1, GFP_NOFS);
-       kfree(op);
-
-       if (cur == update_list) {
-               btrfs_mark_buffer_dirty(path->nodes[0]);
-               btrfs_release_path(extent_root, path);
-               goto out;
-       }
-
-       op = list_entry(cur, struct pending_extent_op, list);
-
-       path->slots[0]++;
-       while (path->slots[0] < btrfs_header_nritems(leaf)) {
-               btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
-               if (key.objectid == op->bytenr &&
-                   key.type == BTRFS_EXTENT_REF_KEY)
-                       goto loop;
-               path->slots[0]++;
-       }
-
-       btrfs_mark_buffer_dirty(path->nodes[0]);
-       btrfs_release_path(extent_root, path);
-       goto search;
-
-out:
-       return 0;
-}
-
-static noinline int insert_extents(struct btrfs_trans_handle *trans,
-                                  struct btrfs_root *extent_root,
-                                  struct btrfs_path *path,
-                                  struct list_head *insert_list, int nr)
-{
-       struct btrfs_key *keys;
-       u32 *data_size;
-       struct pending_extent_op *op;
-       struct extent_buffer *leaf;
-       struct list_head *cur = insert_list->next;
-       struct btrfs_fs_info *info = extent_root->fs_info;
-       u64 ref_root = extent_root->root_key.objectid;
-       int i = 0, last = 0, ret;
-       int total = nr * 2;
-
-       if (!nr)
-               return 0;
-
-       keys = kzalloc(total * sizeof(struct btrfs_key), GFP_NOFS);
-       if (!keys)
-               return -ENOMEM;
-
-       data_size = kzalloc(total * sizeof(u32), GFP_NOFS);
-       if (!data_size) {
-               kfree(keys);
-               return -ENOMEM;
-       }
-
-       list_for_each_entry(op, insert_list, list) {
-               keys[i].objectid = op->bytenr;
-               keys[i].offset = op->num_bytes;
-               keys[i].type = BTRFS_EXTENT_ITEM_KEY;
-               data_size[i] = sizeof(struct btrfs_extent_item);
-               i++;
-
-               keys[i].objectid = op->bytenr;
-               keys[i].offset = op->parent;
-               keys[i].type = BTRFS_EXTENT_REF_KEY;
-               data_size[i] = sizeof(struct btrfs_extent_ref);
-               i++;
-       }
-
-       op = list_entry(cur, struct pending_extent_op, list);
-       i = 0;
-       while (i < total) {
-               int c;
-               ret = btrfs_insert_some_items(trans, extent_root, path,
-                                             keys+i, data_size+i, total-i);
-               BUG_ON(ret < 0);
-
-               if (last && ret > 1)
-                       BUG();
-
-               leaf = path->nodes[0];
-               for (c = 0; c < ret; c++) {
-                       int ref_first = keys[i].type == BTRFS_EXTENT_REF_KEY;
-
-                       /*
-                        * if the first item we inserted was a backref, then
-                        * the EXTENT_ITEM will be the odd c's, else it will
-                        * be the even c's
-                        */
-                       if ((ref_first && (c % 2)) ||
-                           (!ref_first && !(c % 2))) {
-                               struct btrfs_extent_item *itm;
-
-                               itm = btrfs_item_ptr(leaf, path->slots[0] + c,
-                                                    struct btrfs_extent_item);
-                               btrfs_set_extent_refs(path->nodes[0], itm, 1);
-                               op->del++;
-                       } else {
-                               struct btrfs_extent_ref *ref;
-
-                               ref = btrfs_item_ptr(leaf, path->slots[0] + c,
-                                                    struct btrfs_extent_ref);
-                               btrfs_set_ref_root(leaf, ref, ref_root);
-                               btrfs_set_ref_generation(leaf, ref,
-                                                        op->generation);
-                               btrfs_set_ref_objectid(leaf, ref, op->level);
-                               btrfs_set_ref_num_refs(leaf, ref, 1);
-                               op->del++;
-                       }
-
-                       /*
-                        * using del to see when its ok to free up the
-                        * pending_extent_op.  In the case where we insert the
-                        * last item on the list in order to help do batching
-                        * we need to not free the extent op until we actually
-                        * insert the extent_item
-                        */
-                       if (op->del == 2) {
-                               unlock_extent(&info->extent_ins, op->bytenr,
-                                             op->bytenr + op->num_bytes - 1,
-                                             GFP_NOFS);
-                               cur = cur->next;
-                               list_del_init(&op->list);
-                               kfree(op);
-                               if (cur != insert_list)
-                                       op = list_entry(cur,
-                                               struct pending_extent_op,
-                                               list);
-                       }
-               }
-               btrfs_mark_buffer_dirty(leaf);
-               btrfs_release_path(extent_root, path);
-
-               /*
-                * Ok backref's and items usually go right next to eachother,
-                * but if we could only insert 1 item that means that we
-                * inserted on the end of a leaf, and we have no idea what may
-                * be on the next leaf so we just play it safe.  In order to
-                * try and help this case we insert the last thing on our
-                * insert list so hopefully it will end up being the last
-                * thing on the leaf and everything else will be before it,
-                * which will let us insert a whole bunch of items at the same
-                * time.
-                */
-               if (ret == 1 && !last && (i + ret < total)) {
-                       /*
-                        * last: where we will pick up the next time around
-                        * i: our current key to insert, will be total - 1
-                        * cur: the current op we are screwing with
-                        * op: duh
-                        */
-                       last = i + ret;
-                       i = total - 1;
-                       cur = insert_list->prev;
-                       op = list_entry(cur, struct pending_extent_op, list);
-               } else if (last) {
-                       /*
-                        * ok we successfully inserted the last item on the
-                        * list, lets reset everything
-                        *
-                        * i: our current key to insert, so where we left off
-                        *    last time
-                        * last: done with this
-                        * cur: the op we are messing with
-                        * op: duh
-                        * total: since we inserted the last key, we need to
-                        *        decrement total so we dont overflow
-                        */
-                       i = last;
-                       last = 0;
-                       total--;
-                       if (i < total) {
-                               cur = insert_list->next;
-                               op = list_entry(cur, struct pending_extent_op,
-                                               list);
-                       }
-               } else {
-                       i += ret;
-               }
-
-               cond_resched();
-       }
-       ret = 0;
-       kfree(keys);
-       kfree(data_size);
-       return ret;
-}
-
 static noinline int insert_extent_backref(struct btrfs_trans_handle *trans,
                                          struct btrfs_root *root,
                                          struct btrfs_path *path,
                                          u64 bytenr, u64 parent,
                                          u64 ref_root, u64 ref_generation,
-                                         u64 owner_objectid)
+                                         u64 owner_objectid,
+                                         int refs_to_add)
 {
        struct btrfs_key key;
        struct extent_buffer *leaf;
@@ -829,9 +589,10 @@ static noinline int insert_extent_backref(struct btrfs_trans_handle *trans,
                btrfs_set_ref_root(leaf, ref, ref_root);
                btrfs_set_ref_generation(leaf, ref, ref_generation);
                btrfs_set_ref_objectid(leaf, ref, owner_objectid);
-               btrfs_set_ref_num_refs(leaf, ref, 1);
+               btrfs_set_ref_num_refs(leaf, ref, refs_to_add);
        } else if (ret == -EEXIST) {
                u64 existing_owner;
+
                BUG_ON(owner_objectid < BTRFS_FIRST_FREE_OBJECTID);
                leaf = path->nodes[0];
                ref = btrfs_item_ptr(leaf, path->slots[0],
@@ -845,7 +606,7 @@ static noinline int insert_extent_backref(struct btrfs_trans_handle *trans,
 
                num_refs = btrfs_ref_num_refs(leaf, ref);
                BUG_ON(num_refs == 0);
-               btrfs_set_ref_num_refs(leaf, ref, num_refs + 1);
+               btrfs_set_ref_num_refs(leaf, ref, num_refs + refs_to_add);
 
                existing_owner = btrfs_ref_objectid(leaf, ref);
                if (existing_owner != owner_objectid &&
@@ -865,7 +626,8 @@ out:
 
 static noinline int remove_extent_backref(struct btrfs_trans_handle *trans,
                                          struct btrfs_root *root,
-                                         struct btrfs_path *path)
+                                         struct btrfs_path *path,
+                                         int refs_to_drop)
 {
        struct extent_buffer *leaf;
        struct btrfs_extent_ref *ref;
@@ -875,8 +637,8 @@ static noinline int remove_extent_backref(struct btrfs_trans_handle *trans,
        leaf = path->nodes[0];
        ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_ref);
        num_refs = btrfs_ref_num_refs(leaf, ref);
-       BUG_ON(num_refs == 0);
-       num_refs -= 1;
+       BUG_ON(num_refs < refs_to_drop);
+       num_refs -= refs_to_drop;
        if (num_refs == 0) {
                ret = btrfs_del_item(trans, root, path);
        } else {
@@ -927,332 +689,28 @@ static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
 #endif
 }
 
-static noinline int free_extents(struct btrfs_trans_handle *trans,
-                                struct btrfs_root *extent_root,
-                                struct list_head *del_list)
-{
-       struct btrfs_fs_info *info = extent_root->fs_info;
-       struct btrfs_path *path;
-       struct btrfs_key key, found_key;
-       struct extent_buffer *leaf;
-       struct list_head *cur;
-       struct pending_extent_op *op;
-       struct btrfs_extent_item *ei;
-       int ret, num_to_del, extent_slot = 0, found_extent = 0;
-       u32 refs;
-       u64 bytes_freed = 0;
-
-       path = btrfs_alloc_path();
-       if (!path)
-               return -ENOMEM;
-       path->reada = 1;
-
-search:
-       /* search for the backref for the current ref we want to delete */
-       cur = del_list->next;
-       op = list_entry(cur, struct pending_extent_op, list);
-       ret = lookup_extent_backref(trans, extent_root, path, op->bytenr,
-                                   op->orig_parent,
-                                   extent_root->root_key.objectid,
-                                   op->orig_generation, op->level, 1);
-       if (ret) {
-               printk(KERN_ERR "btrfs unable to find backref byte nr %llu "
-                      "root %llu gen %llu owner %u\n",
-                      (unsigned long long)op->bytenr,
-                      (unsigned long long)extent_root->root_key.objectid,
-                      (unsigned long long)op->orig_generation, op->level);
-               btrfs_print_leaf(extent_root, path->nodes[0]);
-               WARN_ON(1);
-               goto out;
-       }
-
-       extent_slot = path->slots[0];
-       num_to_del = 1;
-       found_extent = 0;
-
-       /*
-        * if we aren't the first item on the leaf we can move back one and see
-        * if our ref is right next to our extent item
-        */
-       if (likely(extent_slot)) {
-               extent_slot--;
-               btrfs_item_key_to_cpu(path->nodes[0], &found_key,
-                                     extent_slot);
-               if (found_key.objectid == op->bytenr &&
-                   found_key.type == BTRFS_EXTENT_ITEM_KEY &&
-                   found_key.offset == op->num_bytes) {
-                       num_to_del++;
-                       found_extent = 1;
-               }
-       }
-
-       /*
-        * if we didn't find the extent we need to delete the backref and then
-        * search for the extent item key so we can update its ref count
-        */
-       if (!found_extent) {
-               key.objectid = op->bytenr;
-               key.type = BTRFS_EXTENT_ITEM_KEY;
-               key.offset = op->num_bytes;
-
-               ret = remove_extent_backref(trans, extent_root, path);
-               BUG_ON(ret);
-               btrfs_release_path(extent_root, path);
-               ret = btrfs_search_slot(trans, extent_root, &key, path, -1, 1);
-               BUG_ON(ret);
-               extent_slot = path->slots[0];
-       }
-
-       /* this is where we update the ref count for the extent */
-       leaf = path->nodes[0];
-       ei = btrfs_item_ptr(leaf, extent_slot, struct btrfs_extent_item);
-       refs = btrfs_extent_refs(leaf, ei);
-       BUG_ON(refs == 0);
-       refs--;
-       btrfs_set_extent_refs(leaf, ei, refs);
-
-       btrfs_mark_buffer_dirty(leaf);
-
-       /*
-        * This extent needs deleting.  The reason cur_slot is extent_slot +
-        * num_to_del is because extent_slot points to the slot where the extent
-        * is, and if the backref was not right next to the extent we will be
-        * deleting at least 1 item, and will want to start searching at the
-        * slot directly next to extent_slot.  However if we did find the
-        * backref next to the extent item them we will be deleting at least 2
-        * items and will want to start searching directly after the ref slot
-        */
-       if (!refs) {
-               struct list_head *pos, *n, *end;
-               int cur_slot = extent_slot+num_to_del;
-               u64 super_used;
-               u64 root_used;
-
-               path->slots[0] = extent_slot;
-               bytes_freed = op->num_bytes;
-
-               mutex_lock(&info->pinned_mutex);
-               ret = pin_down_bytes(trans, extent_root, op->bytenr,
-                                    op->num_bytes, op->level >=
-                                    BTRFS_FIRST_FREE_OBJECTID);
-               mutex_unlock(&info->pinned_mutex);
-               BUG_ON(ret < 0);
-               op->del = ret;
-
-               /*
-                * we need to see if we can delete multiple things at once, so
-                * start looping through the list of extents we are wanting to
-                * delete and see if their extent/backref's are right next to
-                * eachother and the extents only have 1 ref
-                */
-               for (pos = cur->next; pos != del_list; pos = pos->next) {
-                       struct pending_extent_op *tmp;
-
-                       tmp = list_entry(pos, struct pending_extent_op, list);
-
-                       /* we only want to delete extent+ref at this stage */
-                       if (cur_slot >= btrfs_header_nritems(leaf) - 1)
-                               break;
-
-                       btrfs_item_key_to_cpu(leaf, &found_key, cur_slot);
-                       if (found_key.objectid != tmp->bytenr ||
-                           found_key.type != BTRFS_EXTENT_ITEM_KEY ||
-                           found_key.offset != tmp->num_bytes)
-                               break;
-
-                       /* check to make sure this extent only has one ref */
-                       ei = btrfs_item_ptr(leaf, cur_slot,
-                                           struct btrfs_extent_item);
-                       if (btrfs_extent_refs(leaf, ei) != 1)
-                               break;
-
-                       btrfs_item_key_to_cpu(leaf, &found_key, cur_slot+1);
-                       if (found_key.objectid != tmp->bytenr ||
-                           found_key.type != BTRFS_EXTENT_REF_KEY ||
-                           found_key.offset != tmp->orig_parent)
-                               break;
-
-                       /*
-                        * the ref is right next to the extent, we can set the
-                        * ref count to 0 since we will delete them both now
-                        */
-                       btrfs_set_extent_refs(leaf, ei, 0);
-
-                       /* pin down the bytes for this extent */
-                       mutex_lock(&info->pinned_mutex);
-                       ret = pin_down_bytes(trans, extent_root, tmp->bytenr,
-                                            tmp->num_bytes, tmp->level >=
-                                            BTRFS_FIRST_FREE_OBJECTID);
-                       mutex_unlock(&info->pinned_mutex);
-                       BUG_ON(ret < 0);
-
-                       /*
-                        * use the del field to tell if we need to go ahead and
-                        * free up the extent when we delete the item or not.
-                        */
-                       tmp->del = ret;
-                       bytes_freed += tmp->num_bytes;
-
-                       num_to_del += 2;
-                       cur_slot += 2;
-               }
-               end = pos;
-
-               /* update the free space counters */
-               spin_lock(&info->delalloc_lock);
-               super_used = btrfs_super_bytes_used(&info->super_copy);
-               btrfs_set_super_bytes_used(&info->super_copy,
-                                          super_used - bytes_freed);
-
-               root_used = btrfs_root_used(&extent_root->root_item);
-               btrfs_set_root_used(&extent_root->root_item,
-                                   root_used - bytes_freed);
-               spin_unlock(&info->delalloc_lock);
-
-               /* delete the items */
-               ret = btrfs_del_items(trans, extent_root, path,
-                                     path->slots[0], num_to_del);
-               BUG_ON(ret);
-
-               /*
-                * loop through the extents we deleted and do the cleanup work
-                * on them
-                */
-               for (pos = cur, n = pos->next; pos != end;
-                    pos = n, n = pos->next) {
-                       struct pending_extent_op *tmp;
-                       tmp = list_entry(pos, struct pending_extent_op, list);
-
-                       /*
-                        * remember tmp->del tells us wether or not we pinned
-                        * down the extent
-                        */
-                       ret = update_block_group(trans, extent_root,
-                                                tmp->bytenr, tmp->num_bytes, 0,
-                                                tmp->del);
-                       BUG_ON(ret);
-
-                       list_del_init(&tmp->list);
-                       unlock_extent(&info->extent_ins, tmp->bytenr,
-                                     tmp->bytenr + tmp->num_bytes - 1,
-                                     GFP_NOFS);
-                       kfree(tmp);
-               }
-       } else if (refs && found_extent) {
-               /*
-                * the ref and extent were right next to eachother, but the
-                * extent still has a ref, so just free the backref and keep
-                * going
-                */
-               ret = remove_extent_backref(trans, extent_root, path);
-               BUG_ON(ret);
-
-               list_del_init(&op->list);
-               unlock_extent(&info->extent_ins, op->bytenr,
-                             op->bytenr + op->num_bytes - 1, GFP_NOFS);
-               kfree(op);
-       } else {
-               /*
-                * the extent has multiple refs and the backref we were looking
-                * for was not right next to it, so just unlock and go next,
-                * we're good to go
-                */
-               list_del_init(&op->list);
-               unlock_extent(&info->extent_ins, op->bytenr,
-                             op->bytenr + op->num_bytes - 1, GFP_NOFS);
-               kfree(op);
-       }
-
-       btrfs_release_path(extent_root, path);
-       if (!list_empty(del_list))
-               goto search;
-
-out:
-       btrfs_free_path(path);
-       return ret;
-}
-
 static int __btrfs_update_extent_ref(struct btrfs_trans_handle *trans,
                                     struct btrfs_root *root, u64 bytenr,
+                                    u64 num_bytes,
                                     u64 orig_parent, u64 parent,
                                     u64 orig_root, u64 ref_root,
                                     u64 orig_generation, u64 ref_generation,
                                     u64 owner_objectid)
 {
        int ret;
-       struct btrfs_root *extent_root = root->fs_info->extent_root;
-       struct btrfs_path *path;
-
-       if (root == root->fs_info->extent_root) {
-               struct pending_extent_op *extent_op;
-               u64 num_bytes;
-
-               BUG_ON(owner_objectid >= BTRFS_MAX_LEVEL);
-               num_bytes = btrfs_level_size(root, (int)owner_objectid);
-               mutex_lock(&root->fs_info->extent_ins_mutex);
-               if (test_range_bit(&root->fs_info->extent_ins, bytenr,
-                               bytenr + num_bytes - 1, EXTENT_WRITEBACK, 0)) {
-                       u64 priv;
-                       ret = get_state_private(&root->fs_info->extent_ins,
-                                               bytenr, &priv);
-                       BUG_ON(ret);
-                       extent_op = (struct pending_extent_op *)
-                                                       (unsigned long)priv;
-                       BUG_ON(extent_op->parent != orig_parent);
-                       BUG_ON(extent_op->generation != orig_generation);
-
-                       extent_op->parent = parent;
-                       extent_op->generation = ref_generation;
-               } else {
-                       extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
-                       BUG_ON(!extent_op);
-
-                       extent_op->type = PENDING_BACKREF_UPDATE;
-                       extent_op->bytenr = bytenr;
-                       extent_op->num_bytes = num_bytes;
-                       extent_op->parent = parent;
-                       extent_op->orig_parent = orig_parent;
-                       extent_op->generation = ref_generation;
-                       extent_op->orig_generation = orig_generation;
-                       extent_op->level = (int)owner_objectid;
-                       INIT_LIST_HEAD(&extent_op->list);
-                       extent_op->del = 0;
-
-                       set_extent_bits(&root->fs_info->extent_ins,
-                                       bytenr, bytenr + num_bytes - 1,
-                                       EXTENT_WRITEBACK, GFP_NOFS);
-                       set_state_private(&root->fs_info->extent_ins,
-                                         bytenr, (unsigned long)extent_op);
-               }
-               mutex_unlock(&root->fs_info->extent_ins_mutex);
-               return 0;
-       }
+       int pin = owner_objectid < BTRFS_FIRST_FREE_OBJECTID;
 
-       path = btrfs_alloc_path();
-       if (!path)
-               return -ENOMEM;
-       ret = lookup_extent_backref(trans, extent_root, path,
-                                   bytenr, orig_parent, orig_root,
-                                   orig_generation, owner_objectid, 1);
-       if (ret)
-               goto out;
-       ret = remove_extent_backref(trans, extent_root, path);
-       if (ret)
-               goto out;
-       ret = insert_extent_backref(trans, extent_root, path, bytenr,
-                                   parent, ref_root, ref_generation,
-                                   owner_objectid);
+       ret = btrfs_update_delayed_ref(trans, bytenr, num_bytes,
+                                      orig_parent, parent, orig_root,
+                                      ref_root, orig_generation,
+                                      ref_generation, owner_objectid, pin);
        BUG_ON(ret);
-       finish_current_insert(trans, extent_root, 0);
-       del_pending_extents(trans, extent_root, 0);
-out:
-       btrfs_free_path(path);
        return ret;
 }
 
 int btrfs_update_extent_ref(struct btrfs_trans_handle *trans,
                            struct btrfs_root *root, u64 bytenr,
-                           u64 orig_parent, u64 parent,
+                           u64 num_bytes, u64 orig_parent, u64 parent,
                            u64 ref_root, u64 ref_generation,
                            u64 owner_objectid)
 {
@@ -1260,20 +718,36 @@ int btrfs_update_extent_ref(struct btrfs_trans_handle *trans,
        if (ref_root == BTRFS_TREE_LOG_OBJECTID &&
            owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
                return 0;
-       ret = __btrfs_update_extent_ref(trans, root, bytenr, orig_parent,
-                                       parent, ref_root, ref_root,
-                                       ref_generation, ref_generation,
-                                       owner_objectid);
+
+       ret = __btrfs_update_extent_ref(trans, root, bytenr, num_bytes,
+                                       orig_parent, parent, ref_root,
+                                       ref_root, ref_generation,
+                                       ref_generation, owner_objectid);
        return ret;
 }
-
 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
                                  struct btrfs_root *root, u64 bytenr,
+                                 u64 num_bytes,
                                  u64 orig_parent, u64 parent,
                                  u64 orig_root, u64 ref_root,
                                  u64 orig_generation, u64 ref_generation,
                                  u64 owner_objectid)
 {
+       int ret;
+
+       ret = btrfs_add_delayed_ref(trans, bytenr, num_bytes, parent, ref_root,
+                                   ref_generation, owner_objectid,
+                                   BTRFS_ADD_DELAYED_REF, 0);
+       BUG_ON(ret);
+       return ret;
+}
+
+static noinline_for_stack int add_extent_ref(struct btrfs_trans_handle *trans,
+                         struct btrfs_root *root, u64 bytenr,
+                         u64 num_bytes, u64 parent, u64 ref_root,
+                         u64 ref_generation, u64 owner_objectid,
+                         int refs_to_add)
+{
        struct btrfs_path *path;
        int ret;
        struct btrfs_key key;
@@ -1288,15 +762,19 @@ static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
        path->reada = 1;
        key.objectid = bytenr;
        key.type = BTRFS_EXTENT_ITEM_KEY;
-       key.offset = (u64)-1;
+       key.offset = num_bytes;
 
-       ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key, path,
-                               0, 1);
+       /* first find the extent item and update its reference count */
+       ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
+                               path, 0, 1);
        if (ret < 0)
                return ret;
-       BUG_ON(ret == 0 || path->slots[0] == 0);
 
-       path->slots[0]--;
+       if (ret > 0) {
+               WARN_ON(1);
+               btrfs_free_path(path);
+               return -EIO;
+       }
        l = path->nodes[0];
 
        btrfs_item_key_to_cpu(l, &key, path->slots[0]);
@@ -1310,21 +788,20 @@ static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
        BUG_ON(key.type != BTRFS_EXTENT_ITEM_KEY);
 
        item = btrfs_item_ptr(l, path->slots[0], struct btrfs_extent_item);
+
        refs = btrfs_extent_refs(l, item);
-       btrfs_set_extent_refs(l, item, refs + 1);
+       btrfs_set_extent_refs(l, item, refs + refs_to_add);
        btrfs_mark_buffer_dirty(path->nodes[0]);
 
        btrfs_release_path(root->fs_info->extent_root, path);
 
        path->reada = 1;
+       /* now insert the actual backref */
        ret = insert_extent_backref(trans, root->fs_info->extent_root,
                                    path, bytenr, parent,
                                    ref_root, ref_generation,
-                                   owner_objectid);
+                                   owner_objectid, refs_to_add);
        BUG_ON(ret);
-       finish_current_insert(trans, root->fs_info->extent_root, 0);
-       del_pending_extents(trans, root->fs_info->extent_root, 0);
-
        btrfs_free_path(path);
        return 0;
 }
@@ -1339,68 +816,245 @@ int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
        if (ref_root == BTRFS_TREE_LOG_OBJECTID &&
            owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
                return 0;
-       ret = __btrfs_inc_extent_ref(trans, root, bytenr, 0, parent,
+
+       ret = __btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0, parent,
                                     0, ref_root, 0, ref_generation,
                                     owner_objectid);
        return ret;
 }
 
-int btrfs_extent_post_op(struct btrfs_trans_handle *trans,
-                        struct btrfs_root *root)
+static int drop_delayed_ref(struct btrfs_trans_handle *trans,
+                                       struct btrfs_root *root,
+                                       struct btrfs_delayed_ref_node *node)
+{
+       int ret = 0;
+       struct btrfs_delayed_ref *ref = btrfs_delayed_node_to_ref(node);
+
+       BUG_ON(node->ref_mod == 0);
+       ret = __btrfs_free_extent(trans, root, node->bytenr, node->num_bytes,
+                                 node->parent, ref->root, ref->generation,
+                                 ref->owner_objectid, ref->pin, node->ref_mod);
+
+       return ret;
+}
+
+/* helper function to actually process a single delayed ref entry */
+static noinline int run_one_delayed_ref(struct btrfs_trans_handle *trans,
+                                       struct btrfs_root *root,
+                                       struct btrfs_delayed_ref_node *node,
+                                       int insert_reserved)
 {
-       u64 start;
-       u64 end;
        int ret;
+       struct btrfs_delayed_ref *ref;
 
-       while(1) {
-               finish_current_insert(trans, root->fs_info->extent_root, 1);
-               del_pending_extents(trans, root->fs_info->extent_root, 1);
+       if (node->parent == (u64)-1) {
+               struct btrfs_delayed_ref_head *head;
+               /*
+                * we've hit the end of the chain and we were supposed
+                * to insert this extent into the tree.  But, it got
+                * deleted before we ever needed to insert it, so all
+                * we have to do is clean up the accounting
+                */
+               if (insert_reserved) {
+                       update_reserved_extents(root, node->bytenr,
+                                               node->num_bytes, 0);
+               }
+               head = btrfs_delayed_node_to_head(node);
+               mutex_unlock(&head->mutex);
+               return 0;
+       }
 
-               /* is there more work to do? */
-               ret = find_first_extent_bit(&root->fs_info->pending_del,
-                                           0, &start, &end, EXTENT_WRITEBACK);
-               if (!ret)
-                       continue;
-               ret = find_first_extent_bit(&root->fs_info->extent_ins,
-                                           0, &start, &end, EXTENT_WRITEBACK);
-               if (!ret)
-                       continue;
-               break;
+       ref = btrfs_delayed_node_to_ref(node);
+       if (ref->action == BTRFS_ADD_DELAYED_REF) {
+               if (insert_reserved) {
+                       struct btrfs_key ins;
+
+                       ins.objectid = node->bytenr;
+                       ins.offset = node->num_bytes;
+                       ins.type = BTRFS_EXTENT_ITEM_KEY;
+
+                       /* record the full extent allocation */
+                       ret = __btrfs_alloc_reserved_extent(trans, root,
+                                       node->parent, ref->root,
+                                       ref->generation, ref->owner_objectid,
+                                       &ins, node->ref_mod);
+                       update_reserved_extents(root, node->bytenr,
+                                               node->num_bytes, 0);
+               } else {
+                       /* just add one backref */
+                       ret = add_extent_ref(trans, root, node->bytenr,
+                                    node->num_bytes,
+                                    node->parent, ref->root, ref->generation,
+                                    ref->owner_objectid, node->ref_mod);
+               }
+               BUG_ON(ret);
+       } else if (ref->action == BTRFS_DROP_DELAYED_REF) {
+               WARN_ON(insert_reserved);
+               ret = drop_delayed_ref(trans, root, node);
        }
        return 0;
 }
 
-int btrfs_lookup_extent_ref(struct btrfs_trans_handle *trans,
-                           struct btrfs_root *root, u64 bytenr,
-                           u64 num_bytes, u32 *refs)
+static noinline struct btrfs_delayed_ref_node *
+select_delayed_ref(struct btrfs_delayed_ref_head *head)
 {
-       struct btrfs_path *path;
+       struct rb_node *node;
+       struct btrfs_delayed_ref_node *ref;
+       int action = BTRFS_ADD_DELAYED_REF;
+again:
+       /*
+        * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
+        * this prevents ref count from going down to zero when
+        * there still are pending delayed ref.
+        */
+       node = rb_prev(&head->node.rb_node);
+       while (1) {
+               if (!node)
+                       break;
+               ref = rb_entry(node, struct btrfs_delayed_ref_node,
+                               rb_node);
+               if (ref->bytenr != head->node.bytenr)
+                       break;
+               if (btrfs_delayed_node_to_ref(ref)->action == action)
+                       return ref;
+               node = rb_prev(node);
+       }
+       if (action == BTRFS_ADD_DELAYED_REF) {
+               action = BTRFS_DROP_DELAYED_REF;
+               goto again;
+       }
+       return NULL;
+}
+
+/*
+ * this starts processing the delayed reference count updates and
+ * extent insertions we have queued up so far.  count can be
+ * 0, which means to process everything in the tree at the start
+ * of the run (but not newly added entries), or it can be some target
+ * number you'd like to process.
+ */
+int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
+                          struct btrfs_root *root, unsigned long count)
+{
+       struct rb_node *node;
+       struct btrfs_delayed_ref_root *delayed_refs;
+       struct btrfs_delayed_ref_node *ref;
+       struct btrfs_delayed_ref_head *locked_ref = NULL;
        int ret;
-       struct btrfs_key key;
-       struct extent_buffer *l;
-       struct btrfs_extent_item *item;
+       int must_insert_reserved = 0;
+       int run_all = count == (unsigned long)-1;
 
-       WARN_ON(num_bytes < root->sectorsize);
-       path = btrfs_alloc_path();
-       path->reada = 1;
-       key.objectid = bytenr;
-       key.offset = num_bytes;
-       btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
-       ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key, path,
-                               0, 0);
-       if (ret < 0)
-               goto out;
-       if (ret != 0) {
-               btrfs_print_leaf(root, path->nodes[0]);
-               printk(KERN_INFO "btrfs failed to find block number %llu\n",
-                      (unsigned long long)bytenr);
-               BUG();
+       if (root == root->fs_info->extent_root)
+               root = root->fs_info->tree_root;
+
+       delayed_refs = &trans->transaction->delayed_refs;
+again:
+       spin_lock(&delayed_refs->lock);
+       if (count == 0)
+               count = delayed_refs->num_entries;
+       while (1) {
+               if (!locked_ref) {
+                       /*
+                        * no locked ref, go find something we can
+                        * process in the rbtree.  We start at
+                        * the beginning of the tree, there may be less
+                        * lock contention if we do something smarter here.
+                        */
+                       node = rb_first(&delayed_refs->root);
+                       if (!node) {
+                               spin_unlock(&delayed_refs->lock);
+                               break;
+                       }
+
+                       ref = rb_entry(node, struct btrfs_delayed_ref_node,
+                                      rb_node);
+                       ret = btrfs_lock_delayed_ref(trans, ref, &locked_ref);
+                       if (ret) {
+                               spin_unlock(&delayed_refs->lock);
+                               break;
+                       }
+               }
+
+               /*
+                * record the must insert reserved flag before we
+                * drop the spin lock.
+                */
+               must_insert_reserved = locked_ref->must_insert_reserved;
+               locked_ref->must_insert_reserved = 0;
+
+               /*
+                * locked_ref is the head node, so we have to go one
+                * node back for any delayed ref updates
+                */
+
+               ref = select_delayed_ref(locked_ref);
+               if (!ref) {
+                       /* All delayed refs have been processed, Go ahead
+                        * and send the head node to run_one_delayed_ref,
+                        * so that any accounting fixes can happen
+                        */
+                       ref = &locked_ref->node;
+                       locked_ref = NULL;
+               }
+
+               ref->in_tree = 0;
+               rb_erase(&ref->rb_node, &delayed_refs->root);
+               delayed_refs->num_entries--;
+               spin_unlock(&delayed_refs->lock);
+
+               ret = run_one_delayed_ref(trans, root, ref,
+                                         must_insert_reserved);
+               BUG_ON(ret);
+               btrfs_put_delayed_ref(ref);
+
+               /* once we lock the head ref, we have to process all the
+                * entries for it.  So, we might end up doing more entries
+                * that count was asking us to do.
+                */
+               if (count > 0)
+                       count--;
+
+               /*
+                * we set locked_ref to null above if we're all done
+                * with this bytenr
+                */
+               if (!locked_ref && count == 0)
+                       break;
+
+               spin_lock(&delayed_refs->lock);
+       }
+       if (run_all) {
+               spin_lock(&delayed_refs->lock);
+               node = rb_first(&delayed_refs->root);
+               if (!node) {
+                       spin_unlock(&delayed_refs->lock);
+                       goto out;
+               }
+
+               while (node) {
+                       ref = rb_entry(node, struct btrfs_delayed_ref_node,
+                                      rb_node);
+                       if (btrfs_delayed_ref_is_head(ref)) {
+                               struct btrfs_delayed_ref_head *head;
+
+                               head = btrfs_delayed_node_to_head(ref);
+                               atomic_inc(&ref->refs);
+
+                               spin_unlock(&delayed_refs->lock);
+                               mutex_lock(&head->mutex);
+                               mutex_unlock(&head->mutex);
+
+                               btrfs_put_delayed_ref(ref);
+                               goto again;
+                       }
+                       node = rb_next(node);
+               }
+               spin_unlock(&delayed_refs->lock);
+               count = (unsigned long)-1;
+               schedule_timeout(1);
+               goto again;
        }
-       l = path->nodes[0];
-       item = btrfs_item_ptr(l, path->slots[0], struct btrfs_extent_item);
-       *refs = btrfs_extent_refs(l, item);
 out:
-       btrfs_free_path(path);
        return 0;
 }
 
@@ -1624,7 +1278,7 @@ noinline int btrfs_inc_ref(struct btrfs_trans_handle *trans,
        int refi = 0;
        int slot;
        int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
-                           u64, u64, u64, u64, u64, u64, u64, u64);
+                           u64, u64, u64, u64, u64, u64, u64, u64, u64);
 
        ref_root = btrfs_header_owner(buf);
        ref_generation = btrfs_header_generation(buf);
@@ -1696,12 +1350,19 @@ noinline int btrfs_inc_ref(struct btrfs_trans_handle *trans,
 
                if (level == 0) {
                        btrfs_item_key_to_cpu(buf, &key, slot);
+                       fi = btrfs_item_ptr(buf, slot,
+                                           struct btrfs_file_extent_item);
+
+                       bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
+                       if (bytenr == 0)
+                               continue;
 
                        ret = process_func(trans, root, bytenr,
-                                          orig_buf->start, buf->start,
-                                          orig_root, ref_root,
-                                          orig_generation, ref_generation,
-                                          key.objectid);
+                                  btrfs_file_extent_disk_num_bytes(buf, fi),
+                                  orig_buf->start, buf->start,
+                                  orig_root, ref_root,
+                                  orig_generation, ref_generation,
+                                  key.objectid);
 
                        if (ret) {
                                faili = slot;
@@ -1709,7 +1370,7 @@ noinline int btrfs_inc_ref(struct btrfs_trans_handle *trans,
                                goto fail;
                        }
                } else {
-                       ret = process_func(trans, root, bytenr,
+                       ret = process_func(trans, root, bytenr, buf->len,
                                           orig_buf->start, buf->start,
                                           orig_root, ref_root,
                                           orig_generation, ref_generation,
@@ -1786,17 +1447,17 @@ int btrfs_update_ref(struct btrfs_trans_handle *trans,
                        if (bytenr == 0)
                                continue;
                        ret = __btrfs_update_extent_ref(trans, root, bytenr,
-                                           orig_buf->start, buf->start,
-                                           orig_root, ref_root,
-                                           orig_generation, ref_generation,
-                                           key.objectid);
+                                   btrfs_file_extent_disk_num_bytes(buf, fi),
+                                   orig_buf->start, buf->start,
+                                   orig_root, ref_root, orig_generation,
+                                   ref_generation, key.objectid);
                        if (ret)
                                goto fail;
                } else {
                        bytenr = btrfs_node_blockptr(buf, slot);
                        ret = __btrfs_update_extent_ref(trans, root, bytenr,
-                                           orig_buf->start, buf->start,
-                                           orig_root, ref_root,
+                                           buf->len, orig_buf->start,
+                                           buf->start, orig_root, ref_root,
                                            orig_generation, ref_generation,
                                            level - 1);
                        if (ret)
@@ -1815,7 +1476,6 @@ static int write_one_cache_group(struct btrfs_trans_handle *trans,
                                 struct btrfs_block_group_cache *cache)
 {
        int ret;
-       int pending_ret;
        struct btrfs_root *extent_root = root->fs_info->extent_root;
        unsigned long bi;
        struct extent_buffer *leaf;
@@ -1831,12 +1491,8 @@ static int write_one_cache_group(struct btrfs_trans_handle *trans,
        btrfs_mark_buffer_dirty(leaf);
        btrfs_release_path(extent_root, path);
 fail:
-       finish_current_insert(trans, extent_root, 0);
-       pending_ret = del_pending_extents(trans, extent_root, 0);
        if (ret)
                return ret;
-       if (pending_ret)
-               return pending_ret;
        return 0;
 
 }
@@ -2474,193 +2130,6 @@ int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
        return ret;
 }
 
-static int finish_current_insert(struct btrfs_trans_handle *trans,
-                                struct btrfs_root *extent_root, int all)
-{
-       u64 start;
-       u64 end;
-       u64 priv;
-       u64 search = 0;
-       struct btrfs_fs_info *info = extent_root->fs_info;
-       struct btrfs_path *path;
-       struct pending_extent_op *extent_op, *tmp;
-       struct list_head insert_list, update_list;
-       int ret;
-       int num_inserts = 0, max_inserts, restart = 0;
-
-       path = btrfs_alloc_path();
-       INIT_LIST_HEAD(&insert_list);
-       INIT_LIST_HEAD(&update_list);
-
-       max_inserts = extent_root->leafsize /
-               (2 * sizeof(struct btrfs_key) + 2 * sizeof(struct btrfs_item) +
-                sizeof(struct btrfs_extent_ref) +
-                sizeof(struct btrfs_extent_item));
-again:
-       mutex_lock(&info->extent_ins_mutex);
-       while (1) {
-               ret = find_first_extent_bit(&info->extent_ins, search, &start,
-                                           &end, EXTENT_WRITEBACK);
-               if (ret) {
-                       if (restart && !num_inserts &&
-                           list_empty(&update_list)) {
-                               restart = 0;
-                               search = 0;
-                               continue;
-                       }
-                       break;
-               }
-
-               ret = try_lock_extent(&info->extent_ins, start, end, GFP_NOFS);
-               if (!ret) {
-                       if (all)
-                               restart = 1;
-                       search = end + 1;
-                       if (need_resched()) {
-                               mutex_unlock(&info->extent_ins_mutex);
-                               cond_resched();
-                               mutex_lock(&info->extent_ins_mutex);
-                       }
-                       continue;
-               }
-
-               ret = get_state_private(&info->extent_ins, start, &priv);
-               BUG_ON(ret);
-               extent_op = (struct pending_extent_op *)(unsigned long) priv;
-
-               if (extent_op->type == PENDING_EXTENT_INSERT) {
-                       num_inserts++;
-                       list_add_tail(&extent_op->list, &insert_list);
-                       search = end + 1;
-                       if (num_inserts == max_inserts) {
-                               restart = 1;
-                               break;
-                       }
-               } else if (extent_op->type == PENDING_BACKREF_UPDATE) {
-                       list_add_tail(&extent_op->list, &update_list);
-                       search = end + 1;
-               } else {
-                       BUG();
-               }
-       }
-
-       /*
-        * process the update list, clear the writeback bit for it, and if
-        * somebody marked this thing for deletion then just unlock it and be
-        * done, the free_extents will handle it
-        */
-       list_for_each_entry_safe(extent_op, tmp, &update_list, list) {
-               clear_extent_bits(&info->extent_ins, extent_op->bytenr,
-                                 extent_op->bytenr + extent_op->num_bytes - 1,
-                                 EXTENT_WRITEBACK, GFP_NOFS);
-               if (extent_op->del) {
-                       list_del_init(&extent_op->list);
-                       unlock_extent(&info->extent_ins, extent_op->bytenr,
-                                     extent_op->bytenr + extent_op->num_bytes
-                                     - 1, GFP_NOFS);
-                       kfree(extent_op);
-               }
-       }
-       mutex_unlock(&info->extent_ins_mutex);
-
-       /*
-        * still have things left on the update list, go ahead an update
-        * everything
-        */
-       if (!list_empty(&update_list)) {
-               ret = update_backrefs(trans, extent_root, path, &update_list);
-               BUG_ON(ret);
-
-               /* we may have COW'ed new blocks, so lets start over */
-               if (all)
-                       restart = 1;
-       }
-
-       /*
-        * if no inserts need to be done, but we skipped some extents and we
-        * need to make sure everything is cleaned then reset everything and
-        * go back to the beginning
-        */
-       if (!num_inserts && restart) {
-               search = 0;
-               restart = 0;
-               INIT_LIST_HEAD(&update_list);
-               INIT_LIST_HEAD(&insert_list);
-               goto again;
-       } else if (!num_inserts) {
-               goto out;
-       }
-
-       /*
-        * process the insert extents list.  Again if we are deleting this
-        * extent, then just unlock it, pin down the bytes if need be, and be
-        * done with it.  Saves us from having to actually insert the extent
-        * into the tree and then subsequently come along and delete it
-        */
-       mutex_lock(&info->extent_ins_mutex);
-       list_for_each_entry_safe(extent_op, tmp, &insert_list, list) {
-               clear_extent_bits(&info->extent_ins, extent_op->bytenr,
-                                 extent_op->bytenr + extent_op->num_bytes - 1,
-                                 EXTENT_WRITEBACK, GFP_NOFS);
-               if (extent_op->del) {
-                       u64 used;
-                       list_del_init(&extent_op->list);
-                       unlock_extent(&info->extent_ins, extent_op->bytenr,
-                                     extent_op->bytenr + extent_op->num_bytes
-                                     - 1, GFP_NOFS);
-
-                       mutex_lock(&extent_root->fs_info->pinned_mutex);
-                       ret = pin_down_bytes(trans, extent_root,
-                                            extent_op->bytenr,
-                                            extent_op->num_bytes, 0);
-                       mutex_unlock(&extent_root->fs_info->pinned_mutex);
-
-                       spin_lock(&info->delalloc_lock);
-                       used = btrfs_super_bytes_used(&info->super_copy);
-                       btrfs_set_super_bytes_used(&info->super_copy,
-                                       used - extent_op->num_bytes);
-                       used = btrfs_root_used(&extent_root->root_item);
-                       btrfs_set_root_used(&extent_root->root_item,
-                                       used - extent_op->num_bytes);
-                       spin_unlock(&info->delalloc_lock);
-
-                       ret = update_block_group(trans, extent_root,
-                                                extent_op->bytenr,
-                                                extent_op->num_bytes,
-                                                0, ret > 0);
-                       BUG_ON(ret);
-                       kfree(extent_op);
-                       num_inserts--;
-               }
-       }
-       mutex_unlock(&info->extent_ins_mutex);
-
-       ret = insert_extents(trans, extent_root, path, &insert_list,
-                            num_inserts);
-       BUG_ON(ret);
-
-       /*
-        * if restart is set for whatever reason we need to go back and start
-        * searching through the pending list again.
-        *
-        * We just inserted some extents, which could have resulted in new
-        * blocks being allocated, which would result in new blocks needing
-        * updates, so if all is set we _must_ restart to get the updated
-        * blocks.
-        */
-       if (restart || all) {
-               INIT_LIST_HEAD(&insert_list);
-               INIT_LIST_HEAD(&update_list);
-               search = 0;
-               restart = 0;
-               num_inserts = 0;
-               goto again;
-       }
-out:
-       btrfs_free_path(path);
-       return 0;
-}
-
 static int pin_down_bytes(struct btrfs_trans_handle *trans,
                          struct btrfs_root *root,
                          u64 bytenr, u64 num_bytes, int is_data)
@@ -2686,6 +2155,7 @@ static int pin_down_bytes(struct btrfs_trans_handle *trans,
                u64 header_transid = btrfs_header_generation(buf);
                if (header_owner != BTRFS_TREE_LOG_OBJECTID &&
                    header_owner != BTRFS_TREE_RELOC_OBJECTID &&
+                   header_owner != BTRFS_DATA_RELOC_TREE_OBJECTID &&
                    header_transid == trans->transid &&
                    !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
                        clean_tree_block(NULL, root, buf);
@@ -2710,7 +2180,8 @@ static int __free_extent(struct btrfs_trans_handle *trans,
                         struct btrfs_root *root,
                         u64 bytenr, u64 num_bytes, u64 parent,
                         u64 root_objectid, u64 ref_generation,
-                        u64 owner_objectid, int pin, int mark_free)
+                        u64 owner_objectid, int pin, int mark_free,
+                        int refs_to_drop)
 {
        struct btrfs_path *path;
        struct btrfs_key key;
@@ -2753,7 +2224,8 @@ static int __free_extent(struct btrfs_trans_handle *trans,
                                break;
                }
                if (!found_extent) {
-                       ret = remove_extent_backref(trans, extent_root, path);
+                       ret = remove_extent_backref(trans, extent_root, path,
+                                                   refs_to_drop);
                        BUG_ON(ret);
                        btrfs_release_path(extent_root, path);
                        ret = btrfs_search_slot(trans, extent_root,
@@ -2771,8 +2243,9 @@ static int __free_extent(struct btrfs_trans_handle *trans,
                btrfs_print_leaf(extent_root, path->nodes[0]);
                WARN_ON(1);
                printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
-                      "root %llu gen %llu owner %llu\n",
+                      "parent %llu root %llu gen %llu owner %llu\n",
                       (unsigned long long)bytenr,
+                      (unsigned long long)parent,
                       (unsigned long long)root_objectid,
                       (unsigned long long)ref_generation,
                       (unsigned long long)owner_objectid);
@@ -2782,17 +2255,23 @@ static int __free_extent(struct btrfs_trans_handle *trans,
        ei = btrfs_item_ptr(leaf, extent_slot,
                            struct btrfs_extent_item);
        refs = btrfs_extent_refs(leaf, ei);
-       BUG_ON(refs == 0);
-       refs -= 1;
-       btrfs_set_extent_refs(leaf, ei, refs);
 
+       /*
+        * we're not allowed to delete the extent item if there
+        * are other delayed ref updates pending
+        */
+
+       BUG_ON(refs < refs_to_drop);
+       refs -= refs_to_drop;
+       btrfs_set_extent_refs(leaf, ei, refs);
        btrfs_mark_buffer_dirty(leaf);
 
-       if (refs == 0 && found_extent && path->slots[0] == extent_slot + 1) {
+       if (refs == 0 && found_extent &&
+           path->slots[0] == extent_slot + 1) {
                struct btrfs_extent_ref *ref;
                ref = btrfs_item_ptr(leaf, path->slots[0],
                                     struct btrfs_extent_ref);
-               BUG_ON(btrfs_ref_num_refs(leaf, ref) != 1);
+               BUG_ON(btrfs_ref_num_refs(leaf, ref) != refs_to_drop);
                /* if the back ref and the extent are next to each other
                 * they get deleted below in one shot
                 */
@@ -2800,7 +2279,8 @@ static int __free_extent(struct btrfs_trans_handle *trans,
                num_to_del = 2;
        } else if (found_extent) {
                /* otherwise delete the extent back ref */
-               ret = remove_extent_backref(trans, extent_root, path);
+               ret = remove_extent_backref(trans, extent_root, path,
+                                           refs_to_drop);
                BUG_ON(ret);
                /* if refs are 0, we need to setup the path for deletion */
                if (refs == 0) {
@@ -2850,218 +2330,35 @@ static int __free_extent(struct btrfs_trans_handle *trans,
                BUG_ON(ret);
        }
        btrfs_free_path(path);
-       finish_current_insert(trans, extent_root, 0);
        return ret;
 }
 
 /*
- * find all the blocks marked as pending in the radix tree and remove
- * them from the extent map
- */
-static int del_pending_extents(struct btrfs_trans_handle *trans,
-                              struct btrfs_root *extent_root, int all)
-{
-       int ret;
-       int err = 0;
-       u64 start;
-       u64 end;
-       u64 priv;
-       u64 search = 0;
-       int nr = 0, skipped = 0;
-       struct extent_io_tree *pending_del;
-       struct extent_io_tree *extent_ins;
-       struct pending_extent_op *extent_op;
-       struct btrfs_fs_info *info = extent_root->fs_info;
-       struct list_head delete_list;
-
-       INIT_LIST_HEAD(&delete_list);
-       extent_ins = &extent_root->fs_info->extent_ins;
-       pending_del = &extent_root->fs_info->pending_del;
-
-again:
-       mutex_lock(&info->extent_ins_mutex);
-       while (1) {
-               ret = find_first_extent_bit(pending_del, search, &start, &end,
-                                           EXTENT_WRITEBACK);
-               if (ret) {
-                       if (all && skipped && !nr) {
-                               search = 0;
-                               skipped = 0;
-                               continue;
-                       }
-                       mutex_unlock(&info->extent_ins_mutex);
-                       break;
-               }
-
-               ret = try_lock_extent(extent_ins, start, end, GFP_NOFS);
-               if (!ret) {
-                       search = end+1;
-                       skipped = 1;
-
-                       if (need_resched()) {
-                               mutex_unlock(&info->extent_ins_mutex);
-                               cond_resched();
-                               mutex_lock(&info->extent_ins_mutex);
-                       }
-
-                       continue;
-               }
-               BUG_ON(ret < 0);
-
-               ret = get_state_private(pending_del, start, &priv);
-               BUG_ON(ret);
-               extent_op = (struct pending_extent_op *)(unsigned long)priv;
-
-               clear_extent_bits(pending_del, start, end, EXTENT_WRITEBACK,
-                                 GFP_NOFS);
-               if (!test_range_bit(extent_ins, start, end,
-                                   EXTENT_WRITEBACK, 0)) {
-                       list_add_tail(&extent_op->list, &delete_list);
-                       nr++;
-               } else {
-                       kfree(extent_op);
-
-                       ret = get_state_private(&info->extent_ins, start,
-                                               &priv);
-                       BUG_ON(ret);
-                       extent_op = (struct pending_extent_op *)
-                                               (unsigned long)priv;
-
-                       clear_extent_bits(&info->extent_ins, start, end,
-                                         EXTENT_WRITEBACK, GFP_NOFS);
-
-                       if (extent_op->type == PENDING_BACKREF_UPDATE) {
-                               list_add_tail(&extent_op->list, &delete_list);
-                               search = end + 1;
-                               nr++;
-                               continue;
-                       }
-
-                       mutex_lock(&extent_root->fs_info->pinned_mutex);
-                       ret = pin_down_bytes(trans, extent_root, start,
-                                            end + 1 - start, 0);
-                       mutex_unlock(&extent_root->fs_info->pinned_mutex);
-
-                       ret = update_block_group(trans, extent_root, start,
-                                               end + 1 - start, 0, ret > 0);
-
-                       unlock_extent(extent_ins, start, end, GFP_NOFS);
-                       BUG_ON(ret);
-                       kfree(extent_op);
-               }
-               if (ret)
-                       err = ret;
-
-               search = end + 1;
-
-               if (need_resched()) {
-                       mutex_unlock(&info->extent_ins_mutex);
-                       cond_resched();
-                       mutex_lock(&info->extent_ins_mutex);
-               }
-       }
-
-       if (nr) {
-               ret = free_extents(trans, extent_root, &delete_list);
-               BUG_ON(ret);
-       }
-
-       if (all && skipped) {
-               INIT_LIST_HEAD(&delete_list);
-               search = 0;
-               nr = 0;
-               goto again;
-       }
-
-       if (!err)
-               finish_current_insert(trans, extent_root, 0);
-       return err;
-}
-
-/*
  * remove an extent from the root, returns 0 on success
  */
 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
-                              struct btrfs_root *root,
-                              u64 bytenr, u64 num_bytes, u64 parent,
-                              u64 root_objectid, u64 ref_generation,
-                              u64 owner_objectid, int pin)
+                                       struct btrfs_root *root,
+                                       u64 bytenr, u64 num_bytes, u64 parent,
+                                       u64 root_objectid, u64 ref_generation,
+                                       u64 owner_objectid, int pin,
+                                       int refs_to_drop)
 {
-       struct btrfs_root *extent_root = root->fs_info->extent_root;
-       int pending_ret;
-       int ret;
-
        WARN_ON(num_bytes < root->sectorsize);
-       if (root == extent_root) {
-               struct pending_extent_op *extent_op = NULL;
-
-               mutex_lock(&root->fs_info->extent_ins_mutex);
-               if (test_range_bit(&root->fs_info->extent_ins, bytenr,
-                               bytenr + num_bytes - 1, EXTENT_WRITEBACK, 0)) {
-                       u64 priv;
-                       ret = get_state_private(&root->fs_info->extent_ins,
-                                               bytenr, &priv);
-                       BUG_ON(ret);
-                       extent_op = (struct pending_extent_op *)
-                                               (unsigned long)priv;
-
-                       extent_op->del = 1;
-                       if (extent_op->type == PENDING_EXTENT_INSERT) {
-                               mutex_unlock(&root->fs_info->extent_ins_mutex);
-                               return 0;
-                       }
-               }
-
-               if (extent_op) {
-                       ref_generation = extent_op->orig_generation;
-                       parent = extent_op->orig_parent;
-               }
 
-               extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
-               BUG_ON(!extent_op);
-
-               extent_op->type = PENDING_EXTENT_DELETE;
-               extent_op->bytenr = bytenr;
-               extent_op->num_bytes = num_bytes;
-               extent_op->parent = parent;
-               extent_op->orig_parent = parent;
-               extent_op->generation = ref_generation;
-               extent_op->orig_generation = ref_generation;
-               extent_op->level = (int)owner_objectid;
-               INIT_LIST_HEAD(&extent_op->list);
-               extent_op->del = 0;
-
-               set_extent_bits(&root->fs_info->pending_del,
-                               bytenr, bytenr + num_bytes - 1,
-                               EXTENT_WRITEBACK, GFP_NOFS);
-               set_state_private(&root->fs_info->pending_del,
-                                 bytenr, (unsigned long)extent_op);
-               mutex_unlock(&root->fs_info->extent_ins_mutex);
-               return 0;
-       }
-       /* if metadata always pin */
-       if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
-               if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
-                       mutex_lock(&root->fs_info->pinned_mutex);
-                       btrfs_update_pinned_extents(root, bytenr, num_bytes, 1);
-                       mutex_unlock(&root->fs_info->pinned_mutex);
-                       update_reserved_extents(root, bytenr, num_bytes, 0);
-                       return 0;
-               }
+       /*
+        * if metadata always pin
+        * if data pin when any transaction has committed this
+        */
+       if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID ||
+           ref_generation != trans->transid)
                pin = 1;
-       }
 
-       /* if data pin when any transaction has committed this */
        if (ref_generation != trans->transid)
                pin = 1;
 
-       ret = __free_extent(trans, root, bytenr, num_bytes, parent,
+       return __free_extent(trans, root, bytenr, num_bytes, parent,
                            root_objectid, ref_generation,
-                           owner_objectid, pin, pin == 0);
-
-       finish_current_insert(trans, root->fs_info->extent_root, 0);
-       pending_ret = del_pending_extents(trans, root->fs_info->extent_root, 0);
-       return ret ? ret : pending_ret;
+                           owner_objectid, pin, pin == 0, refs_to_drop);
 }
 
 int btrfs_free_extent(struct btrfs_trans_handle *trans,
@@ -3072,9 +2369,26 @@ int btrfs_free_extent(struct btrfs_trans_handle *trans,
 {
        int ret;
 
-       ret = __btrfs_free_extent(trans, root, bytenr, num_bytes, parent,
-                                 root_objectid, ref_generation,
-                                 owner_objectid, pin);
+       /*
+        * tree log blocks never actually go into the extent allocation
+        * tree, just update pinning info and exit early.
+        *
+        * data extents referenced by the tree log do need to have
+        * their reference counts bumped.
+        */
+       if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID &&
+           owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
+               mutex_lock(&root->fs_info->pinned_mutex);
+               btrfs_update_pinned_extents(root, bytenr, num_bytes, 1);
+               mutex_unlock(&root->fs_info->pinned_mutex);
+               update_reserved_extents(root, bytenr, num_bytes, 0);
+               ret = 0;
+       } else {
+               ret = btrfs_add_delayed_ref(trans, bytenr, num_bytes, parent,
+                                      root_objectid, ref_generation,
+                                      owner_objectid,
+                                      BTRFS_DROP_DELAYED_REF, 1);
+       }
        return ret;
 }
 
@@ -3475,10 +2789,10 @@ int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
 static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
                                         struct btrfs_root *root, u64 parent,
                                         u64 root_objectid, u64 ref_generation,
-                                        u64 owner, struct btrfs_key *ins)
+                                        u64 owner, struct btrfs_key *ins,
+                                        int ref_mod)
 {
        int ret;
-       int pending_ret;
        u64 super_used;
        u64 root_used;
        u64 num_bytes = ins->offset;
@@ -3503,33 +2817,6 @@ static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
        btrfs_set_root_used(&root->root_item, root_used + num_bytes);
        spin_unlock(&info->delalloc_lock);
 
-       if (root == extent_root) {
-               struct pending_extent_op *extent_op;
-
-               extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
-               BUG_ON(!extent_op);
-
-               extent_op->type = PENDING_EXTENT_INSERT;
-               extent_op->bytenr = ins->objectid;
-               extent_op->num_bytes = ins->offset;
-               extent_op->parent = parent;
-               extent_op->orig_parent = 0;
-               extent_op->generation = ref_generation;
-               extent_op->orig_generation = 0;
-               extent_op->level = (int)owner;
-               INIT_LIST_HEAD(&extent_op->list);
-               extent_op->del = 0;
-
-               mutex_lock(&root->fs_info->extent_ins_mutex);
-               set_extent_bits(&root->fs_info->extent_ins, ins->objectid,
-                               ins->objectid + ins->offset - 1,
-                               EXTENT_WRITEBACK, GFP_NOFS);
-               set_state_private(&root->fs_info->extent_ins,
-                                 ins->objectid, (unsigned long)extent_op);
-               mutex_unlock(&root->fs_info->extent_ins_mutex);
-               goto update_block;
-       }
-
        memcpy(&keys[0], ins, sizeof(*ins));
        keys[1].objectid = ins->objectid;
        keys[1].type = BTRFS_EXTENT_REF_KEY;
@@ -3546,31 +2833,24 @@ static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
 
        extent_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
                                     struct btrfs_extent_item);
-       btrfs_set_extent_refs(path->nodes[0], extent_item, 1);
+       btrfs_set_extent_refs(path->nodes[0], extent_item, ref_mod);
        ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
                             struct btrfs_extent_ref);
 
        btrfs_set_ref_root(path->nodes[0], ref, root_objectid);
        btrfs_set_ref_generation(path->nodes[0], ref, ref_generation);
        btrfs_set_ref_objectid(path->nodes[0], ref, owner);
-       btrfs_set_ref_num_refs(path->nodes[0], ref, 1);
+       btrfs_set_ref_num_refs(path->nodes[0], ref, ref_mod);
 
        btrfs_mark_buffer_dirty(path->nodes[0]);
 
        trans->alloc_exclude_start = 0;
        trans->alloc_exclude_nr = 0;
        btrfs_free_path(path);
-       finish_current_insert(trans, extent_root, 0);
-       pending_ret = del_pending_extents(trans, extent_root, 0);
 
        if (ret)
                goto out;
-       if (pending_ret) {
-               ret = pending_ret;
-               goto out;
-       }
 
-update_block:
        ret = update_block_group(trans, root, ins->objectid,
                                 ins->offset, 1, 0);
        if (ret) {
@@ -3592,9 +2872,12 @@ int btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
 
        if (root_objectid == BTRFS_TREE_LOG_OBJECTID)
                return 0;
-       ret = __btrfs_alloc_reserved_extent(trans, root, parent, root_objectid,
-                                           ref_generation, owner, ins);
-       update_reserved_extents(root, ins->objectid, ins->offset, 0);
+
+       ret = btrfs_add_delayed_ref(trans, ins->objectid,
+                                   ins->offset, parent, root_objectid,
+                                   ref_generation, owner,
+                                   BTRFS_ADD_DELAYED_EXTENT, 0);
+       BUG_ON(ret);
        return ret;
 }
 
@@ -3621,7 +2904,7 @@ int btrfs_alloc_logged_extent(struct btrfs_trans_handle *trans,
        BUG_ON(ret);
        put_block_group(block_group);
        ret = __btrfs_alloc_reserved_extent(trans, root, parent, root_objectid,
-                                           ref_generation, owner, ins);
+                                           ref_generation, owner, ins, 1);
        return ret;
 }
 
@@ -3640,20 +2923,18 @@ int btrfs_alloc_extent(struct btrfs_trans_handle *trans,
                       u64 search_end, struct btrfs_key *ins, u64 data)
 {
        int ret;
-
        ret = __btrfs_reserve_extent(trans, root, num_bytes,
                                     min_alloc_size, empty_size, hint_byte,
                                     search_end, ins, data);
        BUG_ON(ret);
        if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
-               ret = __btrfs_alloc_reserved_extent(trans, root, parent,
-                                       root_objectid, ref_generation,
-                                       owner_objectid, ins);
+               ret = btrfs_add_delayed_ref(trans, ins->objectid,
+                                           ins->offset, parent, root_objectid,
+                                           ref_generation, owner_objectid,
+                                           BTRFS_ADD_DELAYED_EXTENT, 0);
                BUG_ON(ret);
-
-       } else {
-               update_reserved_extents(root, ins->objectid, ins->offset, 1);
        }
+       update_reserved_extents(root, ins->objectid, ins->offset, 1);
        return ret;
 }
 
@@ -3789,7 +3070,7 @@ int btrfs_drop_leaf_ref(struct btrfs_trans_handle *trans,
 
                fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 
-               ret = __btrfs_free_extent(trans, root, disk_bytenr,
+               ret = btrfs_free_extent(trans, root, disk_bytenr,
                                btrfs_file_extent_disk_num_bytes(leaf, fi),
                                leaf->start, leaf_owner, leaf_generation,
                                key.objectid, 0);
@@ -3829,7 +3110,7 @@ static noinline int cache_drop_leaf_ref(struct btrfs_trans_handle *trans,
         */
        for (i = 0; i < ref->nritems; i++) {
                info = ref->extents + sorted[i].slot;
-               ret = __btrfs_free_extent(trans, root, info->bytenr,
+               ret = btrfs_free_extent(trans, root, info->bytenr,
                                          info->num_bytes, ref->bytenr,
                                          ref->owner, ref->generation,
                                          info->objectid, 0);
@@ -3846,12 +3127,13 @@ static noinline int cache_drop_leaf_ref(struct btrfs_trans_handle *trans,
        return 0;
 }
 
-static int drop_snap_lookup_refcount(struct btrfs_root *root, u64 start,
+static int drop_snap_lookup_refcount(struct btrfs_trans_handle *trans,
+                                    struct btrfs_root *root, u64 start,
                                     u64 len, u32 *refs)
 {
        int ret;
 
-       ret = btrfs_lookup_extent_ref(NULL, root, start, len, refs);
+       ret = btrfs_lookup_extent_ref(trans, root, start, len, refs);
        BUG_ON(ret);
 
 #if 0 /* some debugging code in case we see problems here */
@@ -3959,7 +3241,8 @@ static noinline int drop_level_one_refs(struct btrfs_trans_handle *trans,
                 * we just decrement it below and don't update any
                 * of the refs the leaf points to.
                 */
-               ret = drop_snap_lookup_refcount(root, bytenr, blocksize, &refs);
+               ret = drop_snap_lookup_refcount(trans, root, bytenr,
+                                               blocksize, &refs);
                BUG_ON(ret);
                if (refs != 1)
                        continue;
@@ -4010,7 +3293,7 @@ static noinline int drop_level_one_refs(struct btrfs_trans_handle *trans,
         */
        for (i = 0; i < refi; i++) {
                bytenr = sorted[i].bytenr;
-               ret = __btrfs_free_extent(trans, root, bytenr,
+               ret = btrfs_free_extent(trans, root, bytenr,
                                        blocksize, eb->start,
                                        root_owner, root_gen, 0, 1);
                BUG_ON(ret);
@@ -4053,7 +3336,7 @@ static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
 
        WARN_ON(*level < 0);
        WARN_ON(*level >= BTRFS_MAX_LEVEL);
-       ret = drop_snap_lookup_refcount(root, path->nodes[*level]->start,
+       ret = drop_snap_lookup_refcount(trans, root, path->nodes[*level]->start,
                                path->nodes[*level]->len, &refs);
        BUG_ON(ret);
        if (refs > 1)
@@ -4104,7 +3387,8 @@ static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
                ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
                blocksize = btrfs_level_size(root, *level - 1);
 
-               ret = drop_snap_lookup_refcount(root, bytenr, blocksize, &refs);
+               ret = drop_snap_lookup_refcount(trans, root, bytenr,
+                                               blocksize, &refs);
                BUG_ON(ret);
 
                /*
@@ -4119,7 +3403,7 @@ static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
                        root_gen = btrfs_header_generation(parent);
                        path->slots[*level]++;
 
-                       ret = __btrfs_free_extent(trans, root, bytenr,
+                       ret = btrfs_free_extent(trans, root, bytenr,
                                                blocksize, parent->start,
                                                root_owner, root_gen,
                                                *level - 1, 1);
@@ -4165,7 +3449,7 @@ out:
         * cleanup and free the reference on the last node
         * we processed
         */
-       ret = __btrfs_free_extent(trans, root, bytenr, blocksize,
+       ret = btrfs_free_extent(trans, root, bytenr, blocksize,
                                  parent->start, root_owner, root_gen,
                                  *level, 1);
        free_extent_buffer(path->nodes[*level]);
@@ -5457,6 +4741,7 @@ static noinline int replace_extents_in_leaf(struct btrfs_trans_handle *trans,
                                        root->root_key.objectid,
                                        trans->transid, key.objectid);
                BUG_ON(ret);
+
                ret = btrfs_free_extent(trans, root,
                                        bytenr, num_bytes, leaf->start,
                                        btrfs_header_owner(leaf),
@@ -5768,9 +5053,6 @@ static noinline int relocate_tree_block(struct btrfs_trans_handle *trans,
                                ref_path, NULL, NULL);
        BUG_ON(ret);
 
-       if (root == root->fs_info->extent_root)
-               btrfs_extent_post_op(trans, root);
-
        return 0;
 }
 
@@ -6208,6 +5490,9 @@ again:
        btrfs_remove_leaf_refs(info->tree_root, (u64)-1, 1);
        mutex_unlock(&root->fs_info->cleaner_mutex);
 
+       trans = btrfs_start_transaction(info->tree_root, 1);
+       btrfs_commit_transaction(trans, info->tree_root);
+
        while (1) {
                ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
                if (ret < 0)
@@ -6500,9 +5785,6 @@ int btrfs_make_block_group(struct btrfs_trans_handle *trans,
                                sizeof(cache->item));
        BUG_ON(ret);
 
-       finish_current_insert(trans, extent_root, 0);
-       ret = del_pending_extents(trans, extent_root, 0);
-       BUG_ON(ret);
        set_avail_alloc_bits(extent_root->fs_info, type);
 
        return 0;
index dc78954..c800754 100644 (file)
@@ -643,7 +643,9 @@ next_slot:
 
                        if (disk_bytenr != 0) {
                                ret = btrfs_update_extent_ref(trans, root,
-                                               disk_bytenr, orig_parent,
+                                               disk_bytenr,
+                                               le64_to_cpu(old.disk_num_bytes),
+                                               orig_parent,
                                                leaf->start,
                                                root->root_key.objectid,
                                                trans->transid, ins.objectid);
@@ -912,7 +914,7 @@ again:
        btrfs_set_file_extent_other_encoding(leaf, fi, 0);
 
        if (orig_parent != leaf->start) {
-               ret = btrfs_update_extent_ref(trans, root, bytenr,
+               ret = btrfs_update_extent_ref(trans, root, bytenr, num_bytes,
                                              orig_parent, leaf->start,
                                              root->root_key.objectid,
                                              trans->transid, inode->i_ino);
index d638c54..f94c2ad 100644 (file)
@@ -65,6 +65,12 @@ static noinline int join_transaction(struct btrfs_root *root)
                cur_trans->use_count = 1;
                cur_trans->commit_done = 0;
                cur_trans->start_time = get_seconds();
+
+               cur_trans->delayed_refs.root.rb_node = NULL;
+               cur_trans->delayed_refs.num_entries = 0;
+               cur_trans->delayed_refs.flushing = 0;
+               spin_lock_init(&cur_trans->delayed_refs.lock);
+
                INIT_LIST_HEAD(&cur_trans->pending_snapshots);
                list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
                extent_io_tree_init(&cur_trans->dirty_pages,
@@ -182,6 +188,7 @@ static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
        h->block_group = 0;
        h->alloc_exclude_nr = 0;
        h->alloc_exclude_start = 0;
+       h->delayed_ref_updates = 0;
        root->fs_info->running_transaction->use_count++;
        mutex_unlock(&root->fs_info->trans_mutex);
        return h;
@@ -281,6 +288,14 @@ static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
        struct btrfs_transaction *cur_trans;
        struct btrfs_fs_info *info = root->fs_info;
 
+       if (trans->delayed_ref_updates &&
+           (trans->transaction->delayed_refs.flushing ||
+           trans->transaction->delayed_refs.num_entries > 16384)) {
+               btrfs_run_delayed_refs(trans, root, trans->delayed_ref_updates);
+       } else if (trans->transaction->delayed_refs.num_entries > 64) {
+               wake_up_process(root->fs_info->transaction_kthread);
+       }
+
        mutex_lock(&info->trans_mutex);
        cur_trans = info->running_transaction;
        WARN_ON(cur_trans != trans->transaction);
@@ -424,9 +439,10 @@ static int update_cowonly_root(struct btrfs_trans_handle *trans,
        u64 old_root_bytenr;
        struct btrfs_root *tree_root = root->fs_info->tree_root;
 
-       btrfs_extent_post_op(trans, root);
        btrfs_write_dirty_block_groups(trans, root);
-       btrfs_extent_post_op(trans, root);
+
+       ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
+       BUG_ON(ret);
 
        while (1) {
                old_root_bytenr = btrfs_root_bytenr(&root->root_item);
@@ -438,14 +454,14 @@ static int update_cowonly_root(struct btrfs_trans_handle *trans,
                                     btrfs_header_level(root->node));
                btrfs_set_root_generation(&root->root_item, trans->transid);
 
-               btrfs_extent_post_op(trans, root);
-
                ret = btrfs_update_root(trans, tree_root,
                                        &root->root_key,
                                        &root->root_item);
                BUG_ON(ret);
                btrfs_write_dirty_block_groups(trans, root);
-               btrfs_extent_post_op(trans, root);
+
+               ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
+               BUG_ON(ret);
        }
        return 0;
 }
@@ -459,15 +475,18 @@ int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
        struct btrfs_fs_info *fs_info = root->fs_info;
        struct list_head *next;
        struct extent_buffer *eb;
+       int ret;
 
-       btrfs_extent_post_op(trans, fs_info->tree_root);
+       ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
+       BUG_ON(ret);
 
        eb = btrfs_lock_root_node(fs_info->tree_root);
        btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
        btrfs_tree_unlock(eb);
        free_extent_buffer(eb);
 
-       btrfs_extent_post_op(trans, fs_info->tree_root);
+       ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
+       BUG_ON(ret);
 
        while (!list_empty(&fs_info->dirty_cowonly_roots)) {
                next = fs_info->dirty_cowonly_roots.next;
@@ -475,6 +494,9 @@ int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
                root = list_entry(next, struct btrfs_root, dirty_list);
 
                update_cowonly_root(trans, root);
+
+               ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
+               BUG_ON(ret);
        }
        return 0;
 }
@@ -895,6 +917,21 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
        DEFINE_WAIT(wait);
        int ret;
 
+       /* make a pass through all the delayed refs we have so far
+        * any runnings procs may add more while we are here
+        */
+       ret = btrfs_run_delayed_refs(trans, root, 0);
+       BUG_ON(ret);
+
+       /*
+        * set the flushing flag so procs in this transaction have to
+        * start sending their work down.
+        */
+       trans->transaction->delayed_refs.flushing = 1;
+
+       ret = btrfs_run_delayed_refs(trans, root, (u64)-1);
+       BUG_ON(ret);
+
        INIT_LIST_HEAD(&dirty_fs_roots);
        mutex_lock(&root->fs_info->trans_mutex);
        if (trans->transaction->in_commit) {
@@ -969,6 +1006,9 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
        ret = create_pending_snapshots(trans, root->fs_info);
        BUG_ON(ret);
 
+       ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
+       BUG_ON(ret);
+
        WARN_ON(cur_trans != trans->transaction);
 
        /* btrfs_commit_tree_roots is responsible for getting the
index ea29211..9487670 100644 (file)
@@ -19,6 +19,7 @@
 #ifndef __BTRFS_TRANSACTION__
 #define __BTRFS_TRANSACTION__
 #include "btrfs_inode.h"
+#include "delayed-ref.h"
 
 struct btrfs_transaction {
        u64 transid;
@@ -34,6 +35,7 @@ struct btrfs_transaction {
        wait_queue_head_t writer_wait;
        wait_queue_head_t commit_wait;
        struct list_head pending_snapshots;
+       struct btrfs_delayed_ref_root delayed_refs;
 };
 
 struct btrfs_trans_handle {
@@ -44,6 +46,7 @@ struct btrfs_trans_handle {
        u64 block_group;
        u64 alloc_exclude_start;
        u64 alloc_exclude_nr;
+       unsigned long delayed_ref_updates;
 };
 
 struct btrfs_pending_snapshot {
index 98d25fa..b10eacd 100644 (file)
@@ -124,8 +124,6 @@ int btrfs_defrag_leaves(struct btrfs_trans_handle *trans,
        }
 
        btrfs_release_path(root, path);
-       if (is_extent)
-               btrfs_extent_post_op(trans, root);
 out:
        if (path)
                btrfs_free_path(path);