stop-using-VFS-for-dirty-superblock-management

   1 ext4: stop using VFS for dirty superblock management
   2
   3 From: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
   4
   5 Do not use VFS for managing dirty superblock but do it inside
   6 ext4. Remove the 'ext4_write_super()' VFS callback and instead,
   7 schedule a delayed work when the superblock is marked as dirty. Use
   8 the 'dio_unwritten_wq' which we already have for these purposes.
   9
  10 We add memory barriers to make sure the 's_dirt' flag changes are
  11 visible to other CPUs as soon as possible to avoid queuing unnecessary
  12 works.
  13
  14 The big changes comparing to the previous behavior:
  15 1. We use 'dio_unwritten' queue, so the superblock will be written by
  16    per-filesystem 'ext4-dio-unwrit' thread, instead of 'sync_supers'
  17    thread.
  18
  19 2. We allocate memory in 'ext4_mark_super_dirty()'.
  20
  21 Note: the final goal is to get rid of the 'sync_supers()' kernel
  22 thread which wakes up every 5 seconds and even if there is nothing to
  23 do. Thus, we are pushing superblock management from VFS down to
  24 file-systems.
  25
  26 Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
  27 Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
  28 ---
  29  fs/ext4/super.c |   73 +++++++++++++++++++++++++++++++++++++++++++++++-------
  30  1 files changed, 63 insertions(+), 10 deletions(-)
  31
  32 diff --git a/fs/ext4/super.c b/fs/ext4/super.c
  33 index d9543f3..7d453f7 100644
  34 --- a/fs/ext4/super.c
  35 +++ b/fs/ext4/super.c
  36 @@ -73,7 +73,6 @@ static const char *ext4_decode_error(struct super_block *sb, int errno,
  37  static int ext4_remount(struct super_block *sb, int *flags, char *data);
  38  static int ext4_statfs(struct dentry *dentry, struct kstatfs *buf);
  39  static int ext4_unfreeze(struct super_block *sb);
  40 -static void ext4_write_super(struct super_block *sb);
  41  static int ext4_freeze(struct super_block *sb);
  42  static struct dentry *ext4_mount(struct file_system_type *fs_type, int flags,
  43                        const char *dev_name, void *data);
  44 @@ -1294,7 +1293,6 @@ static const struct super_operations ext4_nojournal_sops = {
  45         .dirty_inode    = ext4_dirty_inode,
  46         .drop_inode     = ext4_drop_inode,
  47         .evict_inode    = ext4_evict_inode,
  48 -       .write_super    = ext4_write_super,
  49         .put_super      = ext4_put_super,
  50         .statfs         = ext4_statfs,
  51         .remount_fs     = ext4_remount,
  52 @@ -4140,6 +4138,14 @@ static int ext4_commit_super(struct super_block *sb, int sync)
  53
  54         if (!sbh || block_device_ejected(sb))
  55                 return error;
  56 +
  57 +       sb->s_dirt = 0;
  58 +       /*
  59 +        * Make sure we first mark the superblock as clean and then start
  60 +        * writing it out.
  61 +        */
  62 +       smp_wmb();
  63 +
  64         if (buffer_write_io_error(sbh)) {
  65                 /*
  66                  * Oh, dear.  A previous attempt to write the
  67 @@ -4180,7 +4186,6 @@ static int ext4_commit_super(struct super_block *sb, int sync)
  68         es->s_free_inodes_count =
  69                 cpu_to_le32(percpu_counter_sum_positive(
  70                                 &EXT4_SB(sb)->s_freeinodes_counter));
  71 -       sb->s_dirt = 0;
  72         BUFFER_TRACE(sbh, "marking dirty");
  73         mark_buffer_dirty(sbh);
  74         if (sync) {
  75 @@ -4199,9 +4204,64 @@ static int ext4_commit_super(struct super_block *sb, int sync)
  76         return error;
  77  }
  78
  79 +struct sb_delayed_work {
  80 +       struct delayed_work dwork;
  81 +       struct super_block *sb;
  82 +};
  83 +
  84 +static struct sb_delayed_work *work_to_sbwork(struct work_struct *work)
  85 +{
  86 +       struct delayed_work *dwork;
  87 +
  88 +       dwork = container_of(work, struct delayed_work, work);
  89 +       return container_of(dwork, struct sb_delayed_work, dwork);
  90 +}
  91 +
  92 +static void write_super(struct work_struct *work)
  93 +{
  94 +       struct sb_delayed_work *sbwork = work_to_sbwork(work);
  95 +       struct super_block *sb = sbwork->sb;
  96 +
  97 +       kfree(sbwork);
  98 +
  99 +       smp_rmb();
 100 +       if (!sb->s_dirt)
 101 +               return;
 102 +
 103 +       lock_super(sb);
 104 +       ext4_commit_super(sb, 1);
 105 +       unlock_super(sb);
 106 +}
 107 +
 108  void __ext4_mark_super_dirty(struct super_block *sb)
 109  {
 110 +       struct ext4_sb_info *sbi = EXT4_SB(sb);
 111 +       struct sb_delayed_work *sbwork;
 112 +       unsigned long delay;
 113 +
 114 +       /* Make sure we see 's_dirt' changes ASAP */
 115 +       smp_rmb();
 116 +       if (sb->s_dirt == 1)
 117 +               return;
 118         sb->s_dirt = 1;
 119 +       /* Make other CPUs see the 's_dirt' change as soon as possible */
 120 +       smp_wmb();
 121 +
 122 +       sbwork = kmalloc(sizeof(struct sb_delayed_work), GFP_NOFS);
 123 +       if (!sbwork) {
 124 +               /*
 125 +                * Well, should not be a big deal - the system must be in
 126 +                * trouble anyway, and the SB will be written out on unmount or
 127 +                * we may be luckier next time it is marked as dirty.
 128 +                */
 129 +               sb->s_dirt = 2;
 130 +               return;
 131 +       }
 132 +
 133 +       INIT_DELAYED_WORK(&sbwork->dwork, write_super);
 134 +       sbwork->sb = sb;
 135 +       delay = msecs_to_jiffies(dirty_writeback_interval * 10);
 136 +       queue_delayed_work(sbi->dio_unwritten_wq, &sbwork->dwork, delay);
 137  }
 138
 139  /*
 140 @@ -4291,13 +4351,6 @@ int ext4_force_commit(struct super_block *sb)
 141         return ret;
 142  }
 143
 144 -static void ext4_write_super(struct super_block *sb)
 145 -{
 146 -       lock_super(sb);
 147 -       ext4_commit_super(sb, 1);
 148 -       unlock_super(sb);
 149 -}
 150 -
 151  static int ext4_sync_fs(struct super_block *sb, int wait)
 152  {
 153         int ret = 0;
 154 --
 155 1.7.7.6
 156
 157