Btrfs: reduce stalls during transaction commit
[safe/jmp/linux-2.6] / fs / btrfs / transaction.c
index 903edab..01c9620 100644 (file)
@@ -192,6 +192,7 @@ 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;
+
        root->fs_info->running_transaction->use_count++;
        mutex_unlock(&root->fs_info->trans_mutex);
        return h;
@@ -281,7 +282,6 @@ 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);
 }
 
@@ -298,6 +298,13 @@ 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;
@@ -666,6 +673,31 @@ int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 }
 
 /*
+ * 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 wait_transaction_pre_flush(struct btrfs_fs_info *info)
+{
+       DEFINE_WAIT(wait);
+
+       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);
+               schedule();
+               mutex_lock(&info->trans_mutex);
+               finish_wait(&info->transaction_wait, &wait);
+       }
+       mutex_unlock(&info->trans_mutex);
+       return 0;
+}
+
+/*
  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
  * all of them
  */
@@ -692,7 +724,22 @@ static noinline int drop_dirty_roots(struct btrfs_root *tree_root,
                atomic_inc(&root->fs_info->throttles);
 
                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;
+                       }
+
                        mutex_lock(&root->fs_info->drop_mutex);
                        ret = btrfs_drop_snapshot(trans, dirty->root);
                        if (ret != -EAGAIN)
@@ -932,20 +979,20 @@ 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++;
+       INIT_LIST_HEAD(&dirty_fs_roots);
+       if (cur_trans->in_commit) {
+               cur_trans->use_count++;
                mutex_unlock(&root->fs_info->trans_mutex);
                btrfs_end_transaction(trans, root);
 
@@ -968,7 +1015,6 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 
        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);
@@ -1081,6 +1127,7 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
        btrfs_copy_pinned(root, pinned_copy);
 
        trans->transaction->blocked = 0;
+
        wake_up(&root->fs_info->transaction_throttle);
        wake_up(&root->fs_info->transaction_wait);
 
@@ -1107,6 +1154,7 @@ 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);