fix possible deadlock in jbd-checkpoint-on-commit.

diff --git a/lustre/kernel_patches/patches/jbd-checkpoint-on-commit.patch b/lustre/kernel_patches/patches/jbd-checkpoint-on-commit.patch
index 86b2973..7541cfd 100644 (file)
--- a/lustre/kernel_patches/patches/jbd-checkpoint-on-commit.patch
+++ b/lustre/kernel_patches/patches/jbd-checkpoint-on-commit.patch
@@ -1,3 +1,66 @@
+Index: linux/fs/jbd/checkpoint.c
+===================================================================
+--- linux.orig/fs/jbd/checkpoint.c
++++ linux/fs/jbd/checkpoint.c
+@@ -170,6 +170,15 @@ static int __cleanup_transaction(journal
+                       spin_unlock(&journal->j_list_lock);
+                       jbd_unlock_bh_state(bh);
+                       log_start_commit(journal, tid);
++                      if (journal->j_task == current) {
++                              /*
++                               * Don't wait for commit if called from
++                               * kjournald to avoid deadlock.
++                               */
++                              ret = -EAGAIN;
++                              spin_lock(&journal->j_list_lock);
++                              break;
++                      }
+                       log_wait_commit(journal, tid);
+                       goto out_return_1;
+               }
+@@ -286,6 +295,7 @@ static int __flush_buffer(journal_t *jou
+ int log_do_checkpoint(journal_t *journal)
+ {
+       int result;
++      int result2;
+       int batch_count = 0;
+       struct buffer_head *bhs[NR_BATCH];
+ 
+@@ -314,7 +324,7 @@ int log_do_checkpoint(journal_t *journal
+               transaction_t *transaction;
+               struct journal_head *jh, *last_jh, *next_jh;
+               int drop_count = 0;
+-              int cleanup_ret, retry = 0;
++              int retry = 0;
+               tid_t this_tid;
+ 
+               transaction = journal->j_checkpoint_transactions;
+@@ -364,17 +374,19 @@ int log_do_checkpoint(journal_t *journal
+                * finding anything to write to disk.  We had better be
+                * able to make some progress or we are in trouble.
+                */
+-              cleanup_ret = __cleanup_transaction(journal, transaction);
+-              J_ASSERT(drop_count != 0 || cleanup_ret != 0);
++              result = __cleanup_transaction(journal, transaction);
++              J_ASSERT(drop_count != 0 || result != 0);
++              if (result == -EAGAIN)
++                      break;
+               if (journal->j_checkpoint_transactions != transaction)
+                       break;
+       }
+       spin_unlock(&journal->j_list_lock);
+-      result = cleanup_journal_tail(journal);
+-      if (result < 0)
+-              return result;
++      result2 = cleanup_journal_tail(journal);
++      if (result == 0)
++              result = result2;
+ 
+-      return 0;
++      return result;
+ }
+ 
+ /*
 Index: linux/fs/jbd/journal.c
 ===================================================================
 --- linux.orig/fs/jbd/journal.c
@@ -26,7 +89,7 @@ Index: linux/fs/jbd/journal.c
                spin_unlock(&journal->j_state_lock);
                del_timer_sync(journal->j_commit_timer);
                journal_commit_transaction(journal);
-@@ -168,6 +172,23 @@ loop:
+@@ -168,6 +172,27 @@ loop:
        }
  
        wake_up(&journal->j_wait_done_commit);
@@ -41,9 +104,13 @@ Index: linux/fs/jbd/journal.c
 +      spin_lock(&journal->j_state_lock);
 +      while (__log_space_left(journal) <
 +             journal->j_max_transaction_buffers + 2 * transaction_size) {
++              int result;
++
 +              spin_unlock(&journal->j_state_lock);
-+              log_do_checkpoint(journal);
++              result = log_do_checkpoint(journal);
 +              spin_lock(&journal->j_state_lock);
++              if (result < 0)
++                      break;
 +      }
 +      up(&journal->j_checkpoint_sem);
 +