]> git.proxmox.com Git - mirror_ubuntu-disco-kernel.git/commitdiff
nilfs2: fix deadlock of segment constructor over I_SYNC flag
authorRyusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Thu, 5 Feb 2015 20:25:20 +0000 (12:25 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 5 Feb 2015 21:35:29 +0000 (13:35 -0800)
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>
fs/nilfs2/nilfs.h
fs/nilfs2/segment.c
fs/nilfs2/segment.h

index 91093cd74f0da15e7f972d9c695f221e5afb454d..385704027575525474673b78d385d7ead580ef6d 100644 (file)
@@ -141,7 +141,6 @@ enum {
  * @ti_save: Backup of journal_info field of task_struct
  * @ti_flags: Flags
  * @ti_count: Nest level
- * @ti_garbage:        List of inode to be put when releasing semaphore
  */
 struct nilfs_transaction_info {
        u32                     ti_magic;
@@ -150,7 +149,6 @@ struct nilfs_transaction_info {
                                   one of other filesystems has a bug. */
        unsigned short          ti_flags;
        unsigned short          ti_count;
-       struct list_head        ti_garbage;
 };
 
 /* ti_magic */
index 7ef18fc656c28568047c30c24adb0bc029bada5c..469086b9f99bc8e20053e492237d48a3373bdb37 100644 (file)
@@ -305,7 +305,6 @@ static void nilfs_transaction_lock(struct super_block *sb,
        ti->ti_count = 0;
        ti->ti_save = cur_ti;
        ti->ti_magic = NILFS_TI_MAGIC;
-       INIT_LIST_HEAD(&ti->ti_garbage);
        current->journal_info = ti;
 
        for (;;) {
@@ -332,8 +331,6 @@ static void nilfs_transaction_unlock(struct super_block *sb)
 
        up_write(&nilfs->ns_segctor_sem);
        current->journal_info = ti->ti_save;
-       if (!list_empty(&ti->ti_garbage))
-               nilfs_dispose_list(nilfs, &ti->ti_garbage, 0);
 }
 
 static void *nilfs_segctor_map_segsum_entry(struct nilfs_sc_info *sci,
@@ -746,6 +743,15 @@ static void nilfs_dispose_list(struct the_nilfs *nilfs,
        }
 }
 
+static void nilfs_iput_work_func(struct work_struct *work)
+{
+       struct nilfs_sc_info *sci = container_of(work, struct nilfs_sc_info,
+                                                sc_iput_work);
+       struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
+
+       nilfs_dispose_list(nilfs, &sci->sc_iput_queue, 0);
+}
+
 static int nilfs_test_metadata_dirty(struct the_nilfs *nilfs,
                                     struct nilfs_root *root)
 {
@@ -1900,8 +1906,8 @@ static int nilfs_segctor_collect_dirty_files(struct nilfs_sc_info *sci,
 static void nilfs_segctor_drop_written_files(struct nilfs_sc_info *sci,
                                             struct the_nilfs *nilfs)
 {
-       struct nilfs_transaction_info *ti = current->journal_info;
        struct nilfs_inode_info *ii, *n;
+       int defer_iput = false;
 
        spin_lock(&nilfs->ns_inode_lock);
        list_for_each_entry_safe(ii, n, &sci->sc_dirty_files, i_dirty) {
@@ -1912,9 +1918,24 @@ static void nilfs_segctor_drop_written_files(struct nilfs_sc_info *sci,
                clear_bit(NILFS_I_BUSY, &ii->i_state);
                brelse(ii->i_bh);
                ii->i_bh = NULL;
-               list_move_tail(&ii->i_dirty, &ti->ti_garbage);
+               list_del_init(&ii->i_dirty);
+               if (!ii->vfs_inode.i_nlink) {
+                       /*
+                        * Defer calling iput() to avoid a deadlock
+                        * over I_SYNC flag for inodes with i_nlink == 0
+                        */
+                       list_add_tail(&ii->i_dirty, &sci->sc_iput_queue);
+                       defer_iput = true;
+               } else {
+                       spin_unlock(&nilfs->ns_inode_lock);
+                       iput(&ii->vfs_inode);
+                       spin_lock(&nilfs->ns_inode_lock);
+               }
        }
        spin_unlock(&nilfs->ns_inode_lock);
+
+       if (defer_iput)
+               schedule_work(&sci->sc_iput_work);
 }
 
 /*
@@ -2583,6 +2604,8 @@ static struct nilfs_sc_info *nilfs_segctor_new(struct super_block *sb,
        INIT_LIST_HEAD(&sci->sc_segbufs);
        INIT_LIST_HEAD(&sci->sc_write_logs);
        INIT_LIST_HEAD(&sci->sc_gc_inodes);
+       INIT_LIST_HEAD(&sci->sc_iput_queue);
+       INIT_WORK(&sci->sc_iput_work, nilfs_iput_work_func);
        init_timer(&sci->sc_timer);
 
        sci->sc_interval = HZ * NILFS_SC_DEFAULT_TIMEOUT;
@@ -2609,6 +2632,8 @@ static void nilfs_segctor_write_out(struct nilfs_sc_info *sci)
                ret = nilfs_segctor_construct(sci, SC_LSEG_SR);
                nilfs_transaction_unlock(sci->sc_super);
 
+               flush_work(&sci->sc_iput_work);
+
        } while (ret && retrycount-- > 0);
 }
 
@@ -2633,6 +2658,9 @@ static void nilfs_segctor_destroy(struct nilfs_sc_info *sci)
                || sci->sc_seq_request != sci->sc_seq_done);
        spin_unlock(&sci->sc_state_lock);
 
+       if (flush_work(&sci->sc_iput_work))
+               flag = true;
+
        if (flag || !nilfs_segctor_confirm(sci))
                nilfs_segctor_write_out(sci);
 
@@ -2642,6 +2670,12 @@ static void nilfs_segctor_destroy(struct nilfs_sc_info *sci)
                nilfs_dispose_list(nilfs, &sci->sc_dirty_files, 1);
        }
 
+       if (!list_empty(&sci->sc_iput_queue)) {
+               nilfs_warning(sci->sc_super, __func__,
+                             "iput queue is not empty\n");
+               nilfs_dispose_list(nilfs, &sci->sc_iput_queue, 1);
+       }
+
        WARN_ON(!list_empty(&sci->sc_segbufs));
        WARN_ON(!list_empty(&sci->sc_write_logs));
 
index 38a1d0013314395938ceb78fd1f1e906705206f1..a48d6de1e02cc276019fb150fee8e0bc6c41bcbe 100644 (file)
@@ -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;