Btrfs: fix the missing error information in create_pending_snapshot()
[firefly-linux-kernel-4.4.55.git] / fs / btrfs / transaction.c
index 27c26004e050a33211674363cfd80c02c98d1063..910ff8051ba93c99b6f655baf354e2e091a2913c 100644 (file)
@@ -272,6 +272,7 @@ enum btrfs_trans_type {
        TRANS_JOIN,
        TRANS_USERSPACE,
        TRANS_JOIN_NOLOCK,
+       TRANS_JOIN_FREEZE,
 };
 
 static int may_wait_transaction(struct btrfs_root *root, int type)
@@ -290,7 +291,8 @@ static int may_wait_transaction(struct btrfs_root *root, int type)
 }
 
 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
-                                                   u64 num_items, int type)
+                                                   u64 num_items, int type,
+                                                   int noflush)
 {
        struct btrfs_trans_handle *h;
        struct btrfs_transaction *cur_trans;
@@ -324,9 +326,14 @@ static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
                }
 
                num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
-               ret = btrfs_block_rsv_add(root,
-                                         &root->fs_info->trans_block_rsv,
-                                         num_bytes);
+               if (noflush)
+                       ret = btrfs_block_rsv_add_noflush(root,
+                                               &root->fs_info->trans_block_rsv,
+                                               num_bytes);
+               else
+                       ret = btrfs_block_rsv_add(root,
+                                               &root->fs_info->trans_block_rsv,
+                                               num_bytes);
                if (ret)
                        return ERR_PTR(ret);
        }
@@ -335,7 +342,19 @@ again:
        if (!h)
                return ERR_PTR(-ENOMEM);
 
-       sb_start_intwrite(root->fs_info->sb);
+       /*
+        * If we are JOIN_NOLOCK we're already committing a transaction and
+        * waiting on this guy, so we don't need to do the sb_start_intwrite
+        * because we're already holding a ref.  We need this because we could
+        * have raced in and did an fsync() on a file which can kick a commit
+        * and then we deadlock with somebody doing a freeze.
+        */
+       if (type != TRANS_JOIN_NOLOCK &&
+           !__sb_start_write(root->fs_info->sb, SB_FREEZE_FS, false)) {
+               if (type == TRANS_JOIN_FREEZE)
+                       return ERR_PTR(-EPERM);
+               sb_start_intwrite(root->fs_info->sb);
+       }
 
        if (may_wait_transaction(root, type))
                wait_current_trans(root);
@@ -368,6 +387,7 @@ again:
        h->qgroup_reserved = qgroup_reserved;
        h->delayed_ref_elem.seq = 0;
        INIT_LIST_HEAD(&h->qgroup_ref_list);
+       INIT_LIST_HEAD(&h->new_bgs);
 
        smp_mb();
        if (cur_trans->blocked && may_wait_transaction(root, type)) {
@@ -393,21 +413,33 @@ got_it:
 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
                                                   int num_items)
 {
-       return start_transaction(root, num_items, TRANS_START);
+       return start_transaction(root, num_items, TRANS_START, 0);
 }
+
+struct btrfs_trans_handle *btrfs_start_transaction_noflush(
+                                       struct btrfs_root *root, int num_items)
+{
+       return start_transaction(root, num_items, TRANS_START, 1);
+}
+
 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 {
-       return start_transaction(root, 0, TRANS_JOIN);
+       return start_transaction(root, 0, TRANS_JOIN, 0);
 }
 
 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 {
-       return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
+       return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 }
 
 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 {
-       return start_transaction(root, 0, TRANS_USERSPACE);
+       return start_transaction(root, 0, TRANS_USERSPACE, 0);
+}
+
+struct btrfs_trans_handle *btrfs_join_transaction_freeze(struct btrfs_root *root)
+{
+       return start_transaction(root, 0, TRANS_JOIN_FREEZE, 0);
 }
 
 /* wait for a transaction commit to be fully complete */
@@ -536,6 +568,9 @@ static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
                trans->qgroup_reserved = 0;
        }
 
+       if (!list_empty(&trans->new_bgs))
+               btrfs_create_pending_block_groups(trans, root);
+
        while (count < 2) {
                unsigned long cur = trans->delayed_ref_updates;
                trans->delayed_ref_updates = 0;
@@ -551,7 +586,8 @@ static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
        btrfs_trans_release_metadata(trans, root);
        trans->block_rsv = NULL;
 
-       sb_end_intwrite(root->fs_info->sb);
+       if (!list_empty(&trans->new_bgs))
+               btrfs_create_pending_block_groups(trans, root);
 
        if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
            should_end_transaction(trans, root)) {
@@ -573,6 +609,9 @@ static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
                }
        }
 
+       if (lock)
+               sb_end_intwrite(root->fs_info->sb);
+
        WARN_ON(cur_trans != info->running_transaction);
        WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
        atomic_dec(&cur_trans->num_writers);
@@ -955,6 +994,8 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
        struct btrfs_root *parent_root;
        struct btrfs_block_rsv *rsv;
        struct inode *parent_inode;
+       struct btrfs_path *path;
+       struct btrfs_dir_item *dir_item;
        struct dentry *parent;
        struct dentry *dentry;
        struct extent_buffer *tmp;
@@ -967,18 +1008,22 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
        u64 root_flags;
        uuid_le new_uuid;
 
-       rsv = trans->block_rsv;
+       path = btrfs_alloc_path();
+       if (!path) {
+               ret = pending->error = -ENOMEM;
+               goto path_alloc_fail;
+       }
 
        new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
        if (!new_root_item) {
                ret = pending->error = -ENOMEM;
-               goto fail;
+               goto root_item_alloc_fail;
        }
 
        ret = btrfs_find_free_objectid(tree_root, &objectid);
        if (ret) {
                pending->error = ret;
-               goto fail;
+               goto no_free_objectid;
        }
 
        btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
@@ -988,22 +1033,22 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
                                                  to_reserve);
                if (ret) {
                        pending->error = ret;
-                       goto fail;
+                       goto no_free_objectid;
                }
        }
 
        ret = btrfs_qgroup_inherit(trans, fs_info, root->root_key.objectid,
                                   objectid, pending->inherit);
-       kfree(pending->inherit);
        if (ret) {
                pending->error = ret;
-               goto fail;
+               goto no_free_objectid;
        }
 
        key.objectid = objectid;
        key.offset = (u64)-1;
        key.type = BTRFS_ROOT_ITEM_KEY;
 
+       rsv = trans->block_rsv;
        trans->block_rsv = &pending->block_rsv;
 
        dentry = pending->dentry;
@@ -1017,24 +1062,21 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
         */
        ret = btrfs_set_inode_index(parent_inode, &index);
        BUG_ON(ret); /* -ENOMEM */
-       ret = btrfs_insert_dir_item(trans, parent_root,
-                               dentry->d_name.name, dentry->d_name.len,
-                               parent_inode, &key,
-                               BTRFS_FT_DIR, index);
-       if (ret == -EEXIST) {
+
+       /* check if there is a file/dir which has the same name. */
+       dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
+                                        btrfs_ino(parent_inode),
+                                        dentry->d_name.name,
+                                        dentry->d_name.len, 0);
+       if (dir_item != NULL && !IS_ERR(dir_item)) {
                pending->error = -EEXIST;
-               dput(parent);
                goto fail;
-       } else if (ret) {
-               goto abort_trans_dput;
+       } else if (IS_ERR(dir_item)) {
+               ret = PTR_ERR(dir_item);
+               btrfs_abort_transaction(trans, root, ret);
+               goto fail;
        }
-
-       btrfs_i_size_write(parent_inode, parent_inode->i_size +
-                                        dentry->d_name.len * 2);
-       parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
-       ret = btrfs_update_inode(trans, parent_root, parent_inode);
-       if (ret)
-               goto abort_trans_dput;
+       btrfs_release_path(path);
 
        /*
         * pull in the delayed directory update
@@ -1043,8 +1085,8 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
         * snapshot
         */
        ret = btrfs_run_delayed_items(trans, root);
-       if (ret) { /* Transaction aborted */
-               dput(parent);
+       if (ret) {      /* Transaction aborted */
+               btrfs_abort_transaction(trans, root, ret);
                goto fail;
        }
 
@@ -1079,7 +1121,8 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
        if (ret) {
                btrfs_tree_unlock(old);
                free_extent_buffer(old);
-               goto abort_trans_dput;
+               btrfs_abort_transaction(trans, root, ret);
+               goto fail;
        }
 
        btrfs_set_lock_blocking(old);
@@ -1088,8 +1131,10 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
        /* clean up in any case */
        btrfs_tree_unlock(old);
        free_extent_buffer(old);
-       if (ret)
-               goto abort_trans_dput;
+       if (ret) {
+               btrfs_abort_transaction(trans, root, ret);
+               goto fail;
+       }
 
        /* see comments in should_cow_block() */
        root->force_cow = 1;
@@ -1101,8 +1146,10 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
        ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
        btrfs_tree_unlock(tmp);
        free_extent_buffer(tmp);
-       if (ret)
-               goto abort_trans_dput;
+       if (ret) {
+               btrfs_abort_transaction(trans, root, ret);
+               goto fail;
+       }
 
        /*
         * insert root back/forward references
@@ -1111,32 +1158,58 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
                                 parent_root->root_key.objectid,
                                 btrfs_ino(parent_inode), index,
                                 dentry->d_name.name, dentry->d_name.len);
-       dput(parent);
-       if (ret)
+       if (ret) {
+               btrfs_abort_transaction(trans, root, ret);
                goto fail;
+       }
 
        key.offset = (u64)-1;
        pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
        if (IS_ERR(pending->snap)) {
                ret = PTR_ERR(pending->snap);
-               goto abort_trans;
+               btrfs_abort_transaction(trans, root, ret);
+               goto fail;
        }
 
        ret = btrfs_reloc_post_snapshot(trans, pending);
+       if (ret) {
+               btrfs_abort_transaction(trans, root, ret);
+               goto fail;
+       }
+
+       ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
+       if (ret) {
+               btrfs_abort_transaction(trans, root, ret);
+               goto fail;
+       }
+
+       ret = btrfs_insert_dir_item(trans, parent_root,
+                                   dentry->d_name.name, dentry->d_name.len,
+                                   parent_inode, &key,
+                                   BTRFS_FT_DIR, index);
+       /* We have check then name at the beginning, so it is impossible. */
+       BUG_ON(ret == -EEXIST);
+       if (ret) {
+               btrfs_abort_transaction(trans, root, ret);
+               goto fail;
+       }
+
+       btrfs_i_size_write(parent_inode, parent_inode->i_size +
+                                        dentry->d_name.len * 2);
+       parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
+       ret = btrfs_update_inode(trans, parent_root, parent_inode);
        if (ret)
-               goto abort_trans;
-       ret = 0;
+               btrfs_abort_transaction(trans, root, ret);
 fail:
-       kfree(new_root_item);
+       dput(parent);
        trans->block_rsv = rsv;
+no_free_objectid:
+       kfree(new_root_item);
+root_item_alloc_fail:
+       btrfs_free_path(path);
+path_alloc_fail:
        btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
        return ret;
-
-abort_trans_dput:
-       dput(parent);
-abort_trans:
-       btrfs_abort_transaction(trans, root, ret);
-       goto fail;
 }
 
 /*
@@ -1229,6 +1302,16 @@ static void do_async_commit(struct work_struct *work)
        struct btrfs_async_commit *ac =
                container_of(work, struct btrfs_async_commit, work.work);
 
+       /*
+        * We've got freeze protection passed with the transaction.
+        * Tell lockdep about it.
+        */
+       rwsem_acquire_read(
+               &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
+               0, 1, _THIS_IP_);
+
+       current->journal_info = ac->newtrans;
+
        btrfs_commit_transaction(ac->newtrans, ac->root);
        kfree(ac);
 }
@@ -1258,6 +1341,14 @@ int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
        atomic_inc(&cur_trans->use_count);
 
        btrfs_end_transaction(trans, root);
+
+       /*
+        * Tell lockdep we've released the freeze rwsem, since the
+        * async commit thread will be the one to unlock it.
+        */
+       rwsem_release(&root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
+                     1, _THIS_IP_);
+
        schedule_delayed_work(&ac->work, 0);
 
        /* wait for transaction to start and unblock */
@@ -1348,6 +1439,9 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
         */
        cur_trans->delayed_refs.flushing = 1;
 
+       if (!list_empty(&trans->new_bgs))
+               btrfs_create_pending_block_groups(trans, root);
+
        ret = btrfs_run_delayed_refs(trans, root, 0);
        if (ret)
                goto cleanup_transaction;
@@ -1403,7 +1497,7 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 
                if (flush_on_commit || snap_pending) {
                        btrfs_start_delalloc_inodes(root, 1);
-                       btrfs_wait_ordered_extents(root, 0, 1);
+                       btrfs_wait_ordered_extents(root, 1);
                }
 
                ret = btrfs_run_delayed_items(trans, root);
@@ -1456,13 +1550,28 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
         */
        mutex_lock(&root->fs_info->reloc_mutex);
 
-       ret = btrfs_run_delayed_items(trans, root);
+       /*
+        * We needn't worry about the delayed items because we will
+        * deal with them in create_pending_snapshot(), which is the
+        * core function of the snapshot creation.
+        */
+       ret = create_pending_snapshots(trans, root->fs_info);
        if (ret) {
                mutex_unlock(&root->fs_info->reloc_mutex);
                goto cleanup_transaction;
        }
 
-       ret = create_pending_snapshots(trans, root->fs_info);
+       /*
+        * We insert the dir indexes of the snapshots and update the inode
+        * of the snapshots' parents after the snapshot creation, so there
+        * are some delayed items which are not dealt with. Now deal with
+        * them.
+        *
+        * We needn't worry that this operation will corrupt the snapshots,
+        * because all the tree which are snapshoted will be forced to COW
+        * the nodes and leaves.
+        */
+       ret = btrfs_run_delayed_items(trans, root);
        if (ret) {
                mutex_unlock(&root->fs_info->reloc_mutex);
                goto cleanup_transaction;