optimize-ext4_should_retry_alloc-to-improve-ENOSPC-performance

   1 ext4: optimize ext4_should_retry_alloc() to improve ENOSPC performance
   2
   3 If there are no pending blocks to be released after a commit, forcing
   4 a journal commit has no hope of helping.  It's possible that a commit
   5 had just completed, so if there are now free blocks available for
   6 allocation, it's worth retrying the commit.
   7
   8 Reported-by: Chao Yu <yuchao0@huawei.com>
   9 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
  10 ---
  11  fs/ext4/balloc.c    |  4 +++-
  12  fs/ext4/ext4.h      |  1 +
  13  fs/ext4/ext4_jbd2.h | 10 +++++++++-
  14  fs/ext4/mballoc.c   | 12 ++++++++++--
  15  4 files changed, 23 insertions(+), 4 deletions(-)
  16
  17 diff --git a/fs/ext4/balloc.c b/fs/ext4/balloc.c
  18 index 3020fd7..0b8105b 100644
  19 --- a/fs/ext4/balloc.c
  20 +++ b/fs/ext4/balloc.c
  21 @@ -610,7 +610,9 @@ int ext4_should_retry_alloc(struct super_block *sb, int *retries)
  22
  23         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
  24
  25 -       jbd2_journal_force_commit_nested(EXT4_SB(sb)->s_journal);
  26 +       smp_mb();
  27 +       if (EXT4_SB(sb)->s_mb_free_pending)
  28 +               jbd2_journal_force_commit_nested(EXT4_SB(sb)->s_journal);
  29         return 1;
  30  }
  31
  32 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
  33 index b84aa1c..96c73e6 100644
  34 --- a/fs/ext4/ext4.h
  35 +++ b/fs/ext4/ext4.h
  36 @@ -1430,6 +1430,7 @@ struct ext4_sb_info {
  37         unsigned short *s_mb_offsets;
  38         unsigned int *s_mb_maxs;
  39         unsigned int s_group_info_size;
  40 +       unsigned int s_mb_free_pending;
  41
  42         /* tunables */
  43         unsigned long s_stripe;
  44 diff --git a/fs/ext4/ext4_jbd2.h b/fs/ext4/ext4_jbd2.h
  45 index 09c1ef3..b1d52c1 100644
  46 --- a/fs/ext4/ext4_jbd2.h
  47 +++ b/fs/ext4/ext4_jbd2.h
  48 @@ -175,6 +175,13 @@ struct ext4_journal_cb_entry {
  49   * There is no guaranteed calling order of multiple registered callbacks on
  50   * the same transaction.
  51   */
  52 +static inline void _ext4_journal_callback_add(handle_t *handle,
  53 +                       struct ext4_journal_cb_entry *jce)
  54 +{
  55 +       /* Add the jce to transaction's private list */
  56 +       list_add_tail(&jce->jce_list, &handle->h_transaction->t_private_list);
  57 +}
  58 +
  59  static inline void ext4_journal_callback_add(handle_t *handle,
  60                         void (*func)(struct super_block *sb,
  61                                      struct ext4_journal_cb_entry *jce,
  62 @@ -187,10 +194,11 @@ static inline void ext4_journal_callback_add(handle_t *handle,
  63         /* Add the jce to transaction's private list */
  64         jce->jce_func = func;
  65         spin_lock(&sbi->s_md_lock);
  66 -       list_add_tail(&jce->jce_list, &handle->h_transaction->t_private_list);
  67 +       _ext4_journal_callback_add(handle, jce);
  68         spin_unlock(&sbi->s_md_lock);
  69  }
  70
  71 +
  72  /**
  73   * ext4_journal_callback_del: delete a registered callback
  74   * @handle: active journal transaction handle on which callback was registered
  75 diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
  76 index c1ab3ec..77249e1 100644
  77 --- a/fs/ext4/mballoc.c
  78 +++ b/fs/ext4/mballoc.c
  79 @@ -2627,6 +2627,7 @@ int ext4_mb_init(struct super_block *sb)
  80
  81         spin_lock_init(&sbi->s_md_lock);
  82         spin_lock_init(&sbi->s_bal_lock);
  83 +       sbi->s_mb_free_pending = 0;
  84
  85         sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
  86         sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
  87 @@ -2814,6 +2815,9 @@ static void ext4_free_data_callback(struct super_block *sb,
  88         /* we expect to find existing buddy because it's pinned */
  89         BUG_ON(err != 0);
  90
  91 +       spin_lock(&EXT4_SB(sb)->s_md_lock);
  92 +       EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
  93 +       spin_unlock(&EXT4_SB(sb)->s_md_lock);
  94
  95         db = e4b.bd_info;
  96         /* there are blocks to put in buddy to make them really free */
  97 @@ -4583,6 +4587,7 @@ ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
  98  {
  99         ext4_group_t group = e4b->bd_group;
 100         ext4_grpblk_t cluster;
 101 +       ext4_grpblk_t clusters = new_entry->efd_count;
 102         struct ext4_free_data *entry;
 103         struct ext4_group_info *db = e4b->bd_info;
 104         struct super_block *sb = e4b->bd_sb;
 105 @@ -4649,8 +4654,11 @@ ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
 106                 }
 107         }
 108         /* Add the extent to transaction's private list */
 109 -       ext4_journal_callback_add(handle, ext4_free_data_callback,
 110 -                                 &new_entry->efd_jce);
 111 +       new_entry->efd_jce.jce_func = ext4_free_data_callback;
 112 +       spin_lock(&sbi->s_md_lock);
 113 +       _ext4_journal_callback_add(handle, &new_entry->efd_jce);
 114 +       sbi->s_mb_free_pending += clusters;
 115 +       spin_unlock(&sbi->s_md_lock);
 116         return 0;
 117  }
 118