nilfs2: fix deadlock of segment constructor over I_SYNC flag

Nilfs2 eventually hangs in a stress test with fsstress program.  This
issue was caused by the following deadlock over I_SYNC flag between
nilfs_segctor_thread() and writeback_sb_inodes():

  nilfs_segctor_thread()
    nilfs_segctor_thread_construct()
      nilfs_segctor_unlock()
        nilfs_dispose_list()
          iput()
            iput_final()
              evict()
                inode_wait_for_writeback()  * wait for I_SYNC flag

  writeback_sb_inodes()
     * set I_SYNC flag on inode->i_state
    __writeback_single_inode()
      do_writepages()
        nilfs_writepages()
          nilfs_construct_dsync_segment()
            nilfs_segctor_sync()
               * wait for completion of segment constructor
    inode_sync_complete()
       * clear I_SYNC flag after __writeback_single_inode() completed

writeback_sb_inodes() calls do_writepages() for dirty inodes after
setting I_SYNC flag on inode->i_state.  do_writepages() in turn calls
nilfs_writepages(), which can run segment constructor and wait for its
completion.  On the other hand, segment constructor calls iput(), which
can call evict() and wait for the I_SYNC flag on
inode_wait_for_writeback().

Since segment constructor doesn't know when I_SYNC will be set, it
cannot know whether iput() will block or not unless inode->i_nlink has a
non-zero count.  We can prevent evict() from being called in iput() by
implementing sop->drop_inode(), but it's not preferable to leave inodes
with i_nlink == 0 for long periods because it even defers file
truncation and inode deallocation.  So, this instead resolves the
deadlock by calling iput() asynchronously with a workqueue for inodes
with i_nlink == 0.

Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Tested-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 5 insertions, 0 deletions

diff --git a/fs/nilfs2/segment.h b/fs/nilfs2/segment.h
index 38a1d0013314..a48d6de1e02c 100644
--- a/fs/nilfs2/segment.h
+++ b/fs/nilfs2/segment.h
@@ -26,6 +26,7 @@
 #include <linux/types.h>
 #include <linux/fs.h>
 #include <linux/buffer_head.h>
+#include <linux/workqueue.h>
 #include <linux/nilfs2_fs.h>
 #include "nilfs.h"
 
@@ -92,6 +93,8 @@ struct nilfs_segsum_pointer {
  * @sc_nblk_inc: Block count of current generation
  * @sc_dirty_files: List of files to be written
  * @sc_gc_inodes: List of GC inodes having blocks to be written
+ * @sc_iput_queue: list of inodes for which iput should be done
+ * @sc_iput_work: work struct to defer iput call
  * @sc_freesegs: array of segment numbers to be freed
  * @sc_nfreesegs: number of segments on @sc_freesegs
  * @sc_dsync_inode: inode whose data pages are written for a sync operation
@@ -135,6 +138,8 @@ struct nilfs_sc_info {
 
 	struct list_head	sc_dirty_files;
 	struct list_head	sc_gc_inodes;
+	struct list_head	sc_iput_queue;
+	struct work_struct	sc_iput_work;
 
 	__u64		       *sc_freesegs;
 	size_t			sc_nfreesegs;