avoid-pointless-scanning-of-checkpoint-lists

   1 jbd2: avoid pointless scanning of checkpoint lists
   2
   3 From: Jan Kara <jack@suse.cz>
   4
   5 Yuanhan has reported that when he is running fsync(2) heavy workload
   6 creating new files over ramdisk, significant amount of time is spent in
   7 __jbd2_journal_clean_checkpoint_list() trying to clean old transactions
   8 (but they cannot be cleaned up because flusher hasn't yet checkpointed
   9 those buffers). The workload can be generated by:
  10   fs_mark -d /fs/ram0/1 -D 2 -N 2560 -n 1000000 -L 1 -S 1 -s 4096
  11
  12 Reduce the amount of scanning by stopping to scan the transaction list
  13 once we find a transaction that cannot be checkpointed. Note that this
  14 way of cleaning is still enough to keep freeing space in the journal
  15 after fully checkpointed transactions.
  16
  17 Reported-and-tested-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
  18 Signed-off-by: Jan Kara <jack@suse.cz>
  19 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
  20 ---
  21  fs/jbd2/checkpoint.c | 32 ++++++++++++++++++--------------
  22  1 file changed, 18 insertions(+), 14 deletions(-)
  23
  24 diff --git a/fs/jbd2/checkpoint.c b/fs/jbd2/checkpoint.c
  25 index 7f34f4716165..e39b2d0e1079 100644
  26 --- a/fs/jbd2/checkpoint.c
  27 +++ b/fs/jbd2/checkpoint.c
  28 @@ -478,7 +478,6 @@ int jbd2_cleanup_journal_tail(journal_t *journal)
  29   * Find all the written-back checkpoint buffers in the given list and
  30   * release them.
  31   *
  32 - * Called with the journal locked.
  33   * Called with j_list_lock held.
  34   * Returns number of buffers reaped (for debug)
  35   */
  36 @@ -498,12 +497,12 @@ static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
  37                 jh = next_jh;
  38                 next_jh = jh->b_cpnext;
  39                 ret = __try_to_free_cp_buf(jh);
  40 -               if (ret) {
  41 -                       freed++;
  42 -                       if (ret == 2) {
  43 -                               *released = 1;
  44 -                               return freed;
  45 -                       }
  46 +               if (!ret)
  47 +                       return freed;
  48 +               freed++;
  49 +               if (ret == 2) {
  50 +                       *released = 1;
  51 +                       return freed;
  52                 }
  53                 /*
  54                  * This function only frees up some memory
  55 @@ -523,7 +522,6 @@ static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
  56   *
  57   * Find all the written-back checkpoint buffers in the journal and release them.
  58   *
  59 - * Called with the journal locked.
  60   * Called with j_list_lock held.
  61   * Returns number of buffers reaped (for debug)
  62   */
  63 @@ -531,7 +529,8 @@ static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
  64  int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
  65  {
  66         transaction_t *transaction, *last_transaction, *next_transaction;
  67 -       int ret = 0;
  68 +       int ret;
  69 +       int freed = 0;
  70         int released;
  71
  72         transaction = journal->j_checkpoint_transactions;
  73 @@ -543,17 +542,21 @@ int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
  74         do {
  75                 transaction = next_transaction;
  76                 next_transaction = transaction->t_cpnext;
  77 -               ret += journal_clean_one_cp_list(transaction->
  78 +               ret = journal_clean_one_cp_list(transaction->
  79                                 t_checkpoint_list, &released);
  80                 /*
  81                  * This function only frees up some memory if possible so we
  82                  * dont have an obligation to finish processing. Bail out if
  83                  * preemption requested:
  84                  */
  85 -               if (need_resched())
  86 +               if (need_resched()) {
  87 +                       freed += ret;
  88                         goto out;
  89 -               if (released)
  90 +               }
  91 +               if (released) {
  92 +                       freed += ret;
  93                         continue;
  94 +               }
  95                 /*
  96                  * It is essential that we are as careful as in the case of
  97                  * t_checkpoint_list with removing the buffer from the list as
  98 @@ -561,11 +564,12 @@ int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
  99                  */
 100                 ret += journal_clean_one_cp_list(transaction->
 101                                 t_checkpoint_io_list, &released);
 102 -               if (need_resched())
 103 +               freed += ret;
 104 +               if (need_resched() || !ret)
 105                         goto out;
 106         } while (transaction != last_transaction);
 107  out:
 108 -       return ret;
 109 +       return freed;
 110  }
 111
 112  /*
 113 --
 114 1.8.1.4
 115
 116 --
 117 To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
 118 the body of a message to majordomo@vger.kernel.org
 119 More majordomo info at  http://vger.kernel.org/majordomo-info.html
 120