f2fs: cover global locks for reserve_new_block

The fill_zero() from fallocate() calls get_new_data_page() in which calls
reserve_new_block().
The reserve_new_block() should be covered by *DATA_NEW*, one of global locks.
And also, before getting the lock, we should check free sections by calling
f2fs_balance_fs().

If we break this rule, f2fs is able to face with out-of-control free space
management and fall into infinite loop like the following scenario as well.

[f2fs_sync_fs()]             [fallocate()]
 - write_checkpoint()        - fill_zero()
  - block_operations()        - get_new_data_page()
   : grab NODE_NEW             - get_dnode_of_data()
                                : get locked dirty node page
    - sync_node_pages()
                                : try to grab NODE_NEW for data allocation
     : trylock and skip the dirty node page
   : call sync_node_pages() repeatedly in order to flush all the dirty node
     pages!

In order to avoid this, we should grab another global lock such as DATA_NEW
before calling get_new_data_page() in fill_zero().

Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>

diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
index 3191b52aafb04cdce0ead1a5dbdb4a1ed6c7b082..6cdab2c64fc6eec7801127512fb0ddc27662a3fd 100644 (file)
--- a/fs/f2fs/file.c
+++ b/fs/f2fs/file.c
@@ -387,12 +387,17 @@ const struct inode_operations f2fs_file_inode_operations = {
 static void fill_zero(struct inode *inode, pgoff_t index,
                                        loff_t start, loff_t len)
 {
+       struct f2fs_sb_info *sbi = F2FS_SB(inode->i_sb);
        struct page *page;
 
        if (!len)
                return;
 
+       f2fs_balance_fs(sbi);
+
+       mutex_lock_op(sbi, DATA_NEW);
        page = get_new_data_page(inode, index, false);
+       mutex_unlock_op(sbi, DATA_NEW);
 
        if (!IS_ERR(page)) {
                wait_on_page_writeback(page);