tree-wide: fix assorted typos all over the place
[safe/jmp/linux-2.6] / fs / btrfs / transaction.c
index 903edab..bca82a4 100644 (file)
@@ -25,7 +25,6 @@
 #include "disk-io.h"
 #include "transaction.h"
 #include "locking.h"
-#include "ref-cache.h"
 #include "tree-log.h"
 
 #define BTRFS_ROOT_TRANS_TAG 0
@@ -41,6 +40,12 @@ static noinline void put_transaction(struct btrfs_transaction *transaction)
        }
 }
 
+static noinline void switch_commit_root(struct btrfs_root *root)
+{
+       free_extent_buffer(root->commit_root);
+       root->commit_root = btrfs_root_node(root);
+}
+
 /*
  * either allocate a new transaction or hop into the existing one
  */
@@ -53,8 +58,6 @@ static noinline int join_transaction(struct btrfs_root *root)
                                             GFP_NOFS);
                BUG_ON(!cur_trans);
                root->fs_info->generation++;
-               root->fs_info->last_alloc = 0;
-               root->fs_info->last_data_alloc = 0;
                cur_trans->num_writers = 1;
                cur_trans->num_joined = 0;
                cur_trans->transid = root->fs_info->generation;
@@ -96,48 +99,39 @@ static noinline int join_transaction(struct btrfs_root *root)
  * to make sure the old root from before we joined the transaction is deleted
  * when the transaction commits
  */
-noinline int btrfs_record_root_in_trans(struct btrfs_root *root)
+static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
+                                        struct btrfs_root *root)
 {
-       struct btrfs_dirty_root *dirty;
-       u64 running_trans_id = root->fs_info->running_transaction->transid;
-       if (root->ref_cows && root->last_trans < running_trans_id) {
+       if (root->ref_cows && root->last_trans < trans->transid) {
                WARN_ON(root == root->fs_info->extent_root);
-               if (root->root_item.refs != 0) {
-                       radix_tree_tag_set(&root->fs_info->fs_roots_radix,
-                                  (unsigned long)root->root_key.objectid,
-                                  BTRFS_ROOT_TRANS_TAG);
-
-                       dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
-                       BUG_ON(!dirty);
-                       dirty->root = kmalloc(sizeof(*dirty->root), GFP_NOFS);
-                       BUG_ON(!dirty->root);
-                       dirty->latest_root = root;
-                       INIT_LIST_HEAD(&dirty->list);
-
-                       root->commit_root = btrfs_root_node(root);
-
-                       memcpy(dirty->root, root, sizeof(*root));
-                       spin_lock_init(&dirty->root->node_lock);
-                       spin_lock_init(&dirty->root->list_lock);
-                       mutex_init(&dirty->root->objectid_mutex);
-                       mutex_init(&dirty->root->log_mutex);
-                       INIT_LIST_HEAD(&dirty->root->dead_list);
-                       dirty->root->node = root->commit_root;
-                       dirty->root->commit_root = NULL;
-
-                       spin_lock(&root->list_lock);
-                       list_add(&dirty->root->dead_list, &root->dead_list);
-                       spin_unlock(&root->list_lock);
-
-                       root->dirty_root = dirty;
-               } else {
-                       WARN_ON(1);
-               }
-               root->last_trans = running_trans_id;
+               WARN_ON(root->commit_root != root->node);
+
+               radix_tree_tag_set(&root->fs_info->fs_roots_radix,
+                          (unsigned long)root->root_key.objectid,
+                          BTRFS_ROOT_TRANS_TAG);
+               root->last_trans = trans->transid;
+               btrfs_init_reloc_root(trans, root);
        }
        return 0;
 }
 
+int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
+                              struct btrfs_root *root)
+{
+       if (!root->ref_cows)
+               return 0;
+
+       mutex_lock(&root->fs_info->trans_mutex);
+       if (root->last_trans == trans->transid) {
+               mutex_unlock(&root->fs_info->trans_mutex);
+               return 0;
+       }
+
+       record_root_in_trans(trans, root);
+       mutex_unlock(&root->fs_info->trans_mutex);
+       return 0;
+}
+
 /* wait for commit against the current transaction to become unblocked
  * when this is done, it is safe to start a new transaction, but the current
  * transaction might not be fully on disk.
@@ -183,7 +177,6 @@ static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
        ret = join_transaction(root);
        BUG_ON(ret);
 
-       btrfs_record_root_in_trans(root);
        h->transid = root->fs_info->running_transaction->transid;
        h->transaction = root->fs_info->running_transaction;
        h->blocks_reserved = num_blocks;
@@ -192,7 +185,12 @@ static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
        h->alloc_exclude_nr = 0;
        h->alloc_exclude_start = 0;
        h->delayed_ref_updates = 0;
+
+       if (!current->journal_info)
+               current->journal_info = h;
+
        root->fs_info->running_transaction->use_count++;
+       record_root_in_trans(h, root);
        mutex_unlock(&root->fs_info->trans_mutex);
        return h;
 }
@@ -234,6 +232,7 @@ static noinline int wait_for_commit(struct btrfs_root *root,
        return 0;
 }
 
+#if 0
 /*
  * rate limit against the drop_snapshot code.  This helps to slow down new
  * operations if the drop_snapshot code isn't able to keep up.
@@ -274,6 +273,7 @@ harder:
                        goto harder;
        }
 }
+#endif
 
 void btrfs_throttle(struct btrfs_root *root)
 {
@@ -281,8 +281,6 @@ void btrfs_throttle(struct btrfs_root *root)
        if (!root->fs_info->open_ioctl_trans)
                wait_current_trans(root);
        mutex_unlock(&root->fs_info->trans_mutex);
-
-       throttle_on_drops(root);
 }
 
 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
@@ -298,6 +296,13 @@ static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
                if (cur &&
                    trans->transaction->delayed_refs.num_heads_ready > 64) {
                        trans->delayed_ref_updates = 0;
+
+                       /*
+                        * do a full flush if the transaction is trying
+                        * to close
+                        */
+                       if (trans->transaction->delayed_refs.flushing)
+                               cur = 0;
                        btrfs_run_delayed_refs(trans, root, cur);
                } else {
                        break;
@@ -315,12 +320,12 @@ static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
                wake_up(&cur_trans->writer_wait);
        put_transaction(cur_trans);
        mutex_unlock(&info->trans_mutex);
+
+       if (current->journal_info == trans)
+               current->journal_info = NULL;
        memset(trans, 0, sizeof(*trans));
        kmem_cache_free(btrfs_trans_handle_cachep, trans);
 
-       if (throttle)
-               throttle_on_drops(root);
-
        return 0;
 }
 
@@ -339,10 +344,10 @@ int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 /*
  * when btree blocks are allocated, they have some corresponding bits set for
  * them in one of two extent_io trees.  This is used to make sure all of
- * those extents are on disk for transaction or log commit
+ * those extents are sent to disk but does not wait on them
  */
-int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
-                                       struct extent_io_tree *dirty_pages)
+int btrfs_write_marked_extents(struct btrfs_root *root,
+                              struct extent_io_tree *dirty_pages)
 {
        int ret;
        int err = 0;
@@ -389,6 +394,29 @@ int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
                        page_cache_release(page);
                }
        }
+       if (err)
+               werr = err;
+       return werr;
+}
+
+/*
+ * when btree blocks are allocated, they have some corresponding bits set for
+ * them in one of two extent_io trees.  This is used to make sure all of
+ * those extents are on disk for transaction or log commit.  We wait
+ * on all the pages and clear them from the dirty pages state tree
+ */
+int btrfs_wait_marked_extents(struct btrfs_root *root,
+                             struct extent_io_tree *dirty_pages)
+{
+       int ret;
+       int err = 0;
+       int werr = 0;
+       struct page *page;
+       struct inode *btree_inode = root->fs_info->btree_inode;
+       u64 start = 0;
+       u64 end;
+       unsigned long index;
+
        while (1) {
                ret = find_first_extent_bit(dirty_pages, 0, &start, &end,
                                            EXTENT_DIRTY);
@@ -419,6 +447,22 @@ int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
        return werr;
 }
 
+/*
+ * when btree blocks are allocated, they have some corresponding bits set for
+ * them in one of two extent_io trees.  This is used to make sure all of
+ * those extents are on disk for transaction or log commit
+ */
+int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
+                                       struct extent_io_tree *dirty_pages)
+{
+       int ret;
+       int ret2;
+
+       ret = btrfs_write_marked_extents(root, dirty_pages);
+       ret2 = btrfs_wait_marked_extents(root, dirty_pages);
+       return ret || ret2;
+}
+
 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
                                     struct btrfs_root *root)
 {
@@ -450,36 +494,32 @@ static int update_cowonly_root(struct btrfs_trans_handle *trans,
 
        btrfs_write_dirty_block_groups(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);
                if (old_root_bytenr == root->node->start)
                        break;
-               btrfs_set_root_bytenr(&root->root_item,
-                                      root->node->start);
-               btrfs_set_root_level(&root->root_item,
-                                    btrfs_header_level(root->node));
-               btrfs_set_root_generation(&root->root_item, trans->transid);
 
+               btrfs_set_root_node(&root->root_item, root->node);
                ret = btrfs_update_root(trans, tree_root,
                                        &root->root_key,
                                        &root->root_item);
                BUG_ON(ret);
-               btrfs_write_dirty_block_groups(trans, root);
 
-               ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
+               ret = btrfs_write_dirty_block_groups(trans, root);
                BUG_ON(ret);
        }
+
+       if (root != root->fs_info->extent_root)
+               switch_commit_root(root);
+
        return 0;
 }
 
 /*
  * update all the cowonly tree roots on disk
  */
-int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
-                           struct btrfs_root *root)
+static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
+                                        struct btrfs_root *root)
 {
        struct btrfs_fs_info *fs_info = root->fs_info;
        struct list_head *next;
@@ -503,10 +543,12 @@ 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);
        }
+
+       down_write(&fs_info->extent_commit_sem);
+       switch_commit_root(fs_info->extent_root);
+       up_write(&fs_info->extent_commit_sem);
+
        return 0;
 }
 
@@ -515,118 +557,53 @@ int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
  * a dirty root struct and adds it into the list of dead roots that need to
  * be deleted
  */
-int btrfs_add_dead_root(struct btrfs_root *root, struct btrfs_root *latest)
+int btrfs_add_dead_root(struct btrfs_root *root)
 {
-       struct btrfs_dirty_root *dirty;
-
-       dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
-       if (!dirty)
-               return -ENOMEM;
-       dirty->root = root;
-       dirty->latest_root = latest;
-
        mutex_lock(&root->fs_info->trans_mutex);
-       list_add(&dirty->list, &latest->fs_info->dead_roots);
+       list_add(&root->root_list, &root->fs_info->dead_roots);
        mutex_unlock(&root->fs_info->trans_mutex);
        return 0;
 }
 
 /*
- * at transaction commit time we need to schedule the old roots for
- * deletion via btrfs_drop_snapshot.  This runs through all the
- * reference counted roots that were modified in the current
- * transaction and puts them into the drop list
+ * update all the cowonly tree roots on disk
  */
-static noinline int add_dirty_roots(struct btrfs_trans_handle *trans,
-                                   struct radix_tree_root *radix,
-                                   struct list_head *list)
+static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
+                                   struct btrfs_root *root)
 {
-       struct btrfs_dirty_root *dirty;
        struct btrfs_root *gang[8];
-       struct btrfs_root *root;
+       struct btrfs_fs_info *fs_info = root->fs_info;
        int i;
        int ret;
        int err = 0;
-       u32 refs;
 
        while (1) {
-               ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
+               ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
+                                                (void **)gang, 0,
                                                 ARRAY_SIZE(gang),
                                                 BTRFS_ROOT_TRANS_TAG);
                if (ret == 0)
                        break;
                for (i = 0; i < ret; i++) {
                        root = gang[i];
-                       radix_tree_tag_clear(radix,
-                                    (unsigned long)root->root_key.objectid,
-                                    BTRFS_ROOT_TRANS_TAG);
-
-                       BUG_ON(!root->ref_tree);
-                       dirty = root->dirty_root;
+                       radix_tree_tag_clear(&fs_info->fs_roots_radix,
+                                       (unsigned long)root->root_key.objectid,
+                                       BTRFS_ROOT_TRANS_TAG);
 
                        btrfs_free_log(trans, root);
-                       btrfs_free_reloc_root(trans, root);
-
-                       if (root->commit_root == root->node) {
-                               WARN_ON(root->node->start !=
-                                       btrfs_root_bytenr(&root->root_item));
-
-                               free_extent_buffer(root->commit_root);
-                               root->commit_root = NULL;
-                               root->dirty_root = NULL;
+                       btrfs_update_reloc_root(trans, root);
 
-                               spin_lock(&root->list_lock);
-                               list_del_init(&dirty->root->dead_list);
-                               spin_unlock(&root->list_lock);
-
-                               kfree(dirty->root);
-                               kfree(dirty);
-
-                               /* make sure to update the root on disk
-                                * so we get any updates to the block used
-                                * counts
-                                */
-                               err = btrfs_update_root(trans,
-                                               root->fs_info->tree_root,
-                                               &root->root_key,
-                                               &root->root_item);
-                               continue;
+                       if (root->commit_root != root->node) {
+                               switch_commit_root(root);
+                               btrfs_set_root_node(&root->root_item,
+                                                   root->node);
                        }
 
-                       memset(&root->root_item.drop_progress, 0,
-                              sizeof(struct btrfs_disk_key));
-                       root->root_item.drop_level = 0;
-                       root->commit_root = NULL;
-                       root->dirty_root = NULL;
-                       root->root_key.offset = root->fs_info->generation;
-                       btrfs_set_root_bytenr(&root->root_item,
-                                             root->node->start);
-                       btrfs_set_root_level(&root->root_item,
-                                            btrfs_header_level(root->node));
-                       btrfs_set_root_generation(&root->root_item,
-                                                 root->root_key.offset);
-
-                       err = btrfs_insert_root(trans, root->fs_info->tree_root,
+                       err = btrfs_update_root(trans, fs_info->tree_root,
                                                &root->root_key,
                                                &root->root_item);
                        if (err)
                                break;
-
-                       refs = btrfs_root_refs(&dirty->root->root_item);
-                       btrfs_set_root_refs(&dirty->root->root_item, refs - 1);
-                       err = btrfs_update_root(trans, root->fs_info->tree_root,
-                                               &dirty->root->root_key,
-                                               &dirty->root->root_item);
-
-                       BUG_ON(err);
-                       if (refs == 1) {
-                               list_add(&dirty->list, list);
-                       } else {
-                               WARN_ON(1);
-                               free_extent_buffer(dirty->root->node);
-                               kfree(dirty->root);
-                               kfree(dirty);
-                       }
                }
        }
        return err;
@@ -665,102 +642,96 @@ int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
        return 0;
 }
 
+#if 0
 /*
- * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
- * all of them
+ * when dropping snapshots, we generate a ton of delayed refs, and it makes
+ * sense not to join the transaction while it is trying to flush the current
+ * queue of delayed refs out.
+ *
+ * This is used by the drop snapshot code only
  */
-static noinline int drop_dirty_roots(struct btrfs_root *tree_root,
-                                    struct list_head *list)
+static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 {
-       struct btrfs_dirty_root *dirty;
-       struct btrfs_trans_handle *trans;
-       unsigned long nr;
-       u64 num_bytes;
-       u64 bytes_used;
-       u64 max_useless;
-       int ret = 0;
-       int err;
-
-       while (!list_empty(list)) {
-               struct btrfs_root *root;
+       DEFINE_WAIT(wait);
 
-               dirty = list_entry(list->prev, struct btrfs_dirty_root, list);
-               list_del_init(&dirty->list);
+       mutex_lock(&info->trans_mutex);
+       while (info->running_transaction &&
+              info->running_transaction->delayed_refs.flushing) {
+               prepare_to_wait(&info->transaction_wait, &wait,
+                               TASK_UNINTERRUPTIBLE);
+               mutex_unlock(&info->trans_mutex);
 
-               num_bytes = btrfs_root_used(&dirty->root->root_item);
-               root = dirty->latest_root;
-               atomic_inc(&root->fs_info->throttles);
+               schedule();
 
-               while (1) {
-                       trans = btrfs_start_transaction(tree_root, 1);
-                       mutex_lock(&root->fs_info->drop_mutex);
-                       ret = btrfs_drop_snapshot(trans, dirty->root);
-                       if (ret != -EAGAIN)
-                               break;
-                       mutex_unlock(&root->fs_info->drop_mutex);
+               mutex_lock(&info->trans_mutex);
+               finish_wait(&info->transaction_wait, &wait);
+       }
+       mutex_unlock(&info->trans_mutex);
+       return 0;
+}
 
-                       err = btrfs_update_root(trans,
-                                       tree_root,
-                                       &dirty->root->root_key,
-                                       &dirty->root->root_item);
-                       if (err)
-                               ret = err;
-                       nr = trans->blocks_used;
-                       ret = btrfs_end_transaction(trans, tree_root);
-                       BUG_ON(ret);
+/*
+ * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
+ * all of them
+ */
+int btrfs_drop_dead_root(struct btrfs_root *root)
+{
+       struct btrfs_trans_handle *trans;
+       struct btrfs_root *tree_root = root->fs_info->tree_root;
+       unsigned long nr;
+       int ret;
 
-                       btrfs_btree_balance_dirty(tree_root, nr);
-                       cond_resched();
+       while (1) {
+               /*
+                * we don't want to jump in and create a bunch of
+                * delayed refs if the transaction is starting to close
+                */
+               wait_transaction_pre_flush(tree_root->fs_info);
+               trans = btrfs_start_transaction(tree_root, 1);
+
+               /*
+                * we've joined a transaction, make sure it isn't
+                * closing right now
+                */
+               if (trans->transaction->delayed_refs.flushing) {
+                       btrfs_end_transaction(trans, tree_root);
+                       continue;
                }
-               BUG_ON(ret);
-               atomic_dec(&root->fs_info->throttles);
-               wake_up(&root->fs_info->transaction_throttle);
 
-               num_bytes -= btrfs_root_used(&dirty->root->root_item);
-               bytes_used = btrfs_root_used(&root->root_item);
-               if (num_bytes) {
-                       mutex_lock(&root->fs_info->trans_mutex);
-                       btrfs_record_root_in_trans(root);
-                       mutex_unlock(&root->fs_info->trans_mutex);
-                       btrfs_set_root_used(&root->root_item,
-                                           bytes_used - num_bytes);
-               }
+               ret = btrfs_drop_snapshot(trans, root);
+               if (ret != -EAGAIN)
+                       break;
 
-               ret = btrfs_del_root(trans, tree_root, &dirty->root->root_key);
-               if (ret) {
-                       BUG();
+               ret = btrfs_update_root(trans, tree_root,
+                                       &root->root_key,
+                                       &root->root_item);
+               if (ret)
                        break;
-               }
-               mutex_unlock(&root->fs_info->drop_mutex);
-
-               spin_lock(&root->list_lock);
-               list_del_init(&dirty->root->dead_list);
-               if (!list_empty(&root->dead_list)) {
-                       struct btrfs_root *oldest;
-                       oldest = list_entry(root->dead_list.prev,
-                                           struct btrfs_root, dead_list);
-                       max_useless = oldest->root_key.offset - 1;
-               } else {
-                       max_useless = root->root_key.offset - 1;
-               }
-               spin_unlock(&root->list_lock);
 
                nr = trans->blocks_used;
                ret = btrfs_end_transaction(trans, tree_root);
                BUG_ON(ret);
 
-               ret = btrfs_remove_leaf_refs(root, max_useless, 0);
-               BUG_ON(ret);
-
-               free_extent_buffer(dirty->root->node);
-               kfree(dirty->root);
-               kfree(dirty);
-
                btrfs_btree_balance_dirty(tree_root, nr);
                cond_resched();
        }
+       BUG_ON(ret);
+
+       ret = btrfs_del_root(trans, tree_root, &root->root_key);
+       BUG_ON(ret);
+
+       nr = trans->blocks_used;
+       ret = btrfs_end_transaction(trans, tree_root);
+       BUG_ON(ret);
+
+       free_extent_buffer(root->node);
+       free_extent_buffer(root->commit_root);
+       kfree(root);
+
+       btrfs_btree_balance_dirty(tree_root, nr);
        return ret;
 }
+#endif
 
 /*
  * new snapshots need to be created at a very specific time in the
@@ -788,24 +759,24 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
        if (ret)
                goto fail;
 
-       btrfs_record_root_in_trans(root);
+       record_root_in_trans(trans, root);
        btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
        memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 
        key.objectid = objectid;
+       /* record when the snapshot was created in key.offset */
        key.offset = trans->transid;
        btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 
        old = btrfs_lock_root_node(root);
        btrfs_cow_block(trans, root, old, NULL, 0, &old);
+       btrfs_set_lock_blocking(old);
 
        btrfs_copy_root(trans, root, old, &tmp, objectid);
        btrfs_tree_unlock(old);
        free_extent_buffer(old);
 
-       btrfs_set_root_bytenr(new_root_item, tmp->start);
-       btrfs_set_root_level(new_root_item, btrfs_header_level(tmp));
-       btrfs_set_root_generation(new_root_item, trans->transid);
+       btrfs_set_root_node(new_root_item, tmp);
        ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
                                new_root_item);
        btrfs_tree_unlock(tmp);
@@ -817,6 +788,7 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
        memcpy(&pending->root_key, &key, sizeof(key));
 fail:
        kfree(new_root_item);
+       btrfs_unreserve_metadata_space(root, 6);
        return ret;
 }
 
@@ -852,24 +824,14 @@ static noinline int finish_pending_snapshot(struct btrfs_fs_info *fs_info,
        ret = btrfs_update_inode(trans, parent_root, parent_inode);
        BUG_ON(ret);
 
-       /* add the backref first */
        ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
                                 pending->root_key.objectid,
-                                BTRFS_ROOT_BACKREF_KEY,
                                 parent_root->root_key.objectid,
                                 parent_inode->i_ino, index, pending->name,
                                 namelen);
 
        BUG_ON(ret);
 
-       /* now add the forward ref */
-       ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
-                                parent_root->root_key.objectid,
-                                BTRFS_ROOT_REF_KEY,
-                                pending->root_key.objectid,
-                                parent_inode->i_ino, index, pending->name,
-                                namelen);
-
        inode = btrfs_lookup_dentry(parent_inode, pending->dentry);
        d_instantiate(pending->dentry, inode);
 fail:
@@ -913,6 +875,34 @@ static noinline int finish_pending_snapshots(struct btrfs_trans_handle *trans,
        return 0;
 }
 
+static void update_super_roots(struct btrfs_root *root)
+{
+       struct btrfs_root_item *root_item;
+       struct btrfs_super_block *super;
+
+       super = &root->fs_info->super_copy;
+
+       root_item = &root->fs_info->chunk_root->root_item;
+       super->chunk_root = root_item->bytenr;
+       super->chunk_root_generation = root_item->generation;
+       super->chunk_root_level = root_item->level;
+
+       root_item = &root->fs_info->tree_root->root_item;
+       super->root = root_item->bytenr;
+       super->generation = root_item->generation;
+       super->root_level = root_item->level;
+}
+
+int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
+{
+       int ret = 0;
+       spin_lock(&info->new_trans_lock);
+       if (info->running_transaction)
+               ret = info->running_transaction->in_commit;
+       spin_unlock(&info->new_trans_lock);
+       return ret;
+}
+
 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
                             struct btrfs_root *root)
 {
@@ -920,11 +910,13 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
        unsigned long timeout = 1;
        struct btrfs_transaction *cur_trans;
        struct btrfs_transaction *prev_trans = NULL;
-       struct btrfs_root *chunk_root = root->fs_info->chunk_root;
-       struct list_head dirty_fs_roots;
-       struct extent_io_tree *pinned_copy;
        DEFINE_WAIT(wait);
        int ret;
+       int should_grow = 0;
+       unsigned long now = get_seconds();
+       int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
+
+       btrfs_run_ordered_operations(root, 0);
 
        /* make a pass through all the delayed refs we have so far
         * any runnings procs may add more while we are here
@@ -932,20 +924,19 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
        ret = btrfs_run_delayed_refs(trans, root, 0);
        BUG_ON(ret);
 
+       cur_trans = trans->transaction;
        /*
         * set the flushing flag so procs in this transaction have to
         * start sending their work down.
         */
-       trans->transaction->delayed_refs.flushing = 1;
+       cur_trans->delayed_refs.flushing = 1;
 
        ret = btrfs_run_delayed_refs(trans, root, 0);
        BUG_ON(ret);
 
-       INIT_LIST_HEAD(&dirty_fs_roots);
        mutex_lock(&root->fs_info->trans_mutex);
-       if (trans->transaction->in_commit) {
-               cur_trans = trans->transaction;
-               trans->transaction->use_count++;
+       if (cur_trans->in_commit) {
+               cur_trans->use_count++;
                mutex_unlock(&root->fs_info->trans_mutex);
                btrfs_end_transaction(trans, root);
 
@@ -959,16 +950,8 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
                return 0;
        }
 
-       pinned_copy = kmalloc(sizeof(*pinned_copy), GFP_NOFS);
-       if (!pinned_copy)
-               return -ENOMEM;
-
-       extent_io_tree_init(pinned_copy,
-                            root->fs_info->btree_inode->i_mapping, GFP_NOFS);
-
        trans->transaction->in_commit = 1;
        trans->transaction->blocked = 1;
-       cur_trans = trans->transaction;
        if (cur_trans->list.prev != &root->fs_info->trans_list) {
                prev_trans = list_entry(cur_trans->list.prev,
                                        struct btrfs_transaction, list);
@@ -983,6 +966,9 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
                }
        }
 
+       if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
+               should_grow = 1;
+
        do {
                int snap_pending = 0;
                joined = cur_trans->num_joined;
@@ -995,22 +981,37 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 
                if (cur_trans->num_writers > 1)
                        timeout = MAX_SCHEDULE_TIMEOUT;
-               else
+               else if (should_grow)
                        timeout = 1;
 
                mutex_unlock(&root->fs_info->trans_mutex);
 
-               if (snap_pending) {
+               if (flush_on_commit) {
+                       btrfs_start_delalloc_inodes(root);
+                       ret = btrfs_wait_ordered_extents(root, 0);
+                       BUG_ON(ret);
+               } else if (snap_pending) {
                        ret = btrfs_wait_ordered_extents(root, 1);
                        BUG_ON(ret);
                }
 
-               schedule_timeout(timeout);
+               /*
+                * rename don't use btrfs_join_transaction, so, once we
+                * set the transaction to blocked above, we aren't going
+                * to get any new ordered operations.  We can safely run
+                * it here and no for sure that nothing new will be added
+                * to the list
+                */
+               btrfs_run_ordered_operations(root, 1);
+
+               smp_mb();
+               if (cur_trans->num_writers > 1 || should_grow)
+                       schedule_timeout(timeout);
 
                mutex_lock(&root->fs_info->trans_mutex);
                finish_wait(&cur_trans->writer_wait, &wait);
        } while (cur_trans->num_writers > 1 ||
-                (cur_trans->num_joined != joined));
+                (should_grow && cur_trans->num_joined != joined));
 
        ret = create_pending_snapshots(trans, root->fs_info);
        BUG_ON(ret);
@@ -1034,41 +1035,34 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
         * with the tree-log code.
         */
        mutex_lock(&root->fs_info->tree_log_mutex);
-       /*
-        * keep tree reloc code from adding new reloc trees
-        */
-       mutex_lock(&root->fs_info->tree_reloc_mutex);
-
 
-       ret = add_dirty_roots(trans, &root->fs_info->fs_roots_radix,
-                             &dirty_fs_roots);
+       ret = commit_fs_roots(trans, root);
        BUG_ON(ret);
 
-       /* add_dirty_roots gets rid of all the tree log roots, it is now
+       /* commit_fs_roots gets rid of all the tree log roots, it is now
         * safe to free the root of tree log roots
         */
        btrfs_free_log_root_tree(trans, root->fs_info);
 
-       ret = btrfs_commit_tree_roots(trans, root);
+       ret = commit_cowonly_roots(trans, root);
        BUG_ON(ret);
 
+       btrfs_prepare_extent_commit(trans, root);
+
        cur_trans = root->fs_info->running_transaction;
        spin_lock(&root->fs_info->new_trans_lock);
        root->fs_info->running_transaction = NULL;
        spin_unlock(&root->fs_info->new_trans_lock);
-       btrfs_set_super_generation(&root->fs_info->super_copy,
-                                  cur_trans->transid);
-       btrfs_set_super_root(&root->fs_info->super_copy,
-                            root->fs_info->tree_root->node->start);
-       btrfs_set_super_root_level(&root->fs_info->super_copy,
-                          btrfs_header_level(root->fs_info->tree_root->node));
-
-       btrfs_set_super_chunk_root(&root->fs_info->super_copy,
-                                  chunk_root->node->start);
-       btrfs_set_super_chunk_root_level(&root->fs_info->super_copy,
-                                        btrfs_header_level(chunk_root->node));
-       btrfs_set_super_chunk_root_generation(&root->fs_info->super_copy,
-                               btrfs_header_generation(chunk_root->node));
+
+       btrfs_set_root_node(&root->fs_info->tree_root->root_item,
+                           root->fs_info->tree_root->node);
+       switch_commit_root(root->fs_info->tree_root);
+
+       btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
+                           root->fs_info->chunk_root->node);
+       switch_commit_root(root->fs_info->chunk_root);
+
+       update_super_roots(root);
 
        if (!root->fs_info->log_root_recovering) {
                btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
@@ -1078,10 +1072,8 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
        memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
               sizeof(root->fs_info->super_copy));
 
-       btrfs_copy_pinned(root, pinned_copy);
-
        trans->transaction->blocked = 0;
-       wake_up(&root->fs_info->transaction_throttle);
+
        wake_up(&root->fs_info->transaction_wait);
 
        mutex_unlock(&root->fs_info->trans_mutex);
@@ -1095,11 +1087,7 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
         */
        mutex_unlock(&root->fs_info->tree_log_mutex);
 
-       btrfs_finish_extent_commit(trans, root, pinned_copy);
-       kfree(pinned_copy);
-
-       btrfs_drop_dead_reloc_roots(root);
-       mutex_unlock(&root->fs_info->tree_reloc_mutex);
+       btrfs_finish_extent_commit(trans, root);
 
        /* do the directory inserts of any pending snapshot creations */
        finish_pending_snapshots(trans, root->fs_info);
@@ -1107,22 +1095,20 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
        mutex_lock(&root->fs_info->trans_mutex);
 
        cur_trans->commit_done = 1;
+
        root->fs_info->last_trans_committed = cur_trans->transid;
+
        wake_up(&cur_trans->commit_wait);
 
        put_transaction(cur_trans);
        put_transaction(cur_trans);
 
-       list_splice_init(&dirty_fs_roots, &root->fs_info->dead_roots);
-       if (root->fs_info->closing)
-               list_splice_init(&root->fs_info->dead_roots, &dirty_fs_roots);
-
        mutex_unlock(&root->fs_info->trans_mutex);
 
-       kmem_cache_free(btrfs_trans_handle_cachep, trans);
+       if (current->journal_info == trans)
+               current->journal_info = NULL;
 
-       if (root->fs_info->closing)
-               drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
+       kmem_cache_free(btrfs_trans_handle_cachep, trans);
        return ret;
 }
 
@@ -1131,16 +1117,22 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
  */
 int btrfs_clean_old_snapshots(struct btrfs_root *root)
 {
-       struct list_head dirty_roots;
-       INIT_LIST_HEAD(&dirty_roots);
-again:
-       mutex_lock(&root->fs_info->trans_mutex);
-       list_splice_init(&root->fs_info->dead_roots, &dirty_roots);
-       mutex_unlock(&root->fs_info->trans_mutex);
+       LIST_HEAD(list);
+       struct btrfs_fs_info *fs_info = root->fs_info;
+
+       mutex_lock(&fs_info->trans_mutex);
+       list_splice_init(&fs_info->dead_roots, &list);
+       mutex_unlock(&fs_info->trans_mutex);
 
-       if (!list_empty(&dirty_roots)) {
-               drop_dirty_roots(root, &dirty_roots);
-               goto again;
+       while (!list_empty(&list)) {
+               root = list_entry(list.next, struct btrfs_root, root_list);
+               list_del(&root->root_list);
+
+               if (btrfs_header_backref_rev(root->node) <
+                   BTRFS_MIXED_BACKREF_REV)
+                       btrfs_drop_snapshot(root, 0);
+               else
+                       btrfs_drop_snapshot(root, 1);
        }
        return 0;
 }