jbd2: fix error handling for checkpoint io
Hidehiro Kawai [Sat, 11 Oct 2008 00:29:13 +0000 (20:29 -0400)]
When a checkpointing IO fails, current JBD2 code doesn't check the
error and continue journaling.  This means latest metadata can be
lost from both the journal and filesystem.

This patch leaves the failed metadata blocks in the journal space
and aborts journaling in the case of jbd2_log_do_checkpoint().
To achieve this, we need to do:

1. don't remove the failed buffer from the checkpoint list where in
   the case of __try_to_free_cp_buf() because it may be released or
   overwritten by a later transaction
2. jbd2_log_do_checkpoint() is the last chance, remove the failed
   buffer from the checkpoint list and abort the journal
3. when checkpointing fails, don't update the journal super block to
   prevent the journaled contents from being cleaned.  For safety,
   don't update j_tail and j_tail_sequence either
4. when checkpointing fails, notify this error to the ext4 layer so
   that ext4 don't clear the needs_recovery flag, otherwise the
   journaled contents are ignored and cleaned in the recovery phase
5. if the recovery fails, keep the needs_recovery flag
6. prevent jbd2_cleanup_journal_tail() from being called between
   __jbd2_journal_drop_transaction() and jbd2_journal_abort()
   (a possible race issue between jbd2_log_do_checkpoint()s called by
   jbd2_journal_flush() and __jbd2_log_wait_for_space())

Signed-off-by: Hidehiro Kawai <hidehiro.kawai.ez@hitachi.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

fs/jbd2/checkpoint.c
fs/jbd2/journal.c
fs/jbd2/recovery.c
include/linux/jbd2.h

index 42895d3..9203c33 100644 (file)
@@ -94,7 +94,8 @@ static int __try_to_free_cp_buf(struct journal_head *jh)
        int ret = 0;
        struct buffer_head *bh = jh2bh(jh);
 
-       if (jh->b_jlist == BJ_None && !buffer_locked(bh) && !buffer_dirty(bh)) {
+       if (jh->b_jlist == BJ_None && !buffer_locked(bh) &&
+           !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
                JBUFFER_TRACE(jh, "remove from checkpoint list");
                ret = __jbd2_journal_remove_checkpoint(jh) + 1;
                jbd_unlock_bh_state(bh);
@@ -176,21 +177,25 @@ static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
  * buffers. Note that we take the buffers in the opposite ordering
  * from the one in which they were submitted for IO.
  *
+ * Return 0 on success, and return <0 if some buffers have failed
+ * to be written out.
+ *
  * Called with j_list_lock held.
  */
-static void __wait_cp_io(journal_t *journal, transaction_t *transaction)
+static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
 {
        struct journal_head *jh;
        struct buffer_head *bh;
        tid_t this_tid;
        int released = 0;
+       int ret = 0;
 
        this_tid = transaction->t_tid;
 restart:
        /* Did somebody clean up the transaction in the meanwhile? */
        if (journal->j_checkpoint_transactions != transaction ||
                        transaction->t_tid != this_tid)
-               return;
+               return ret;
        while (!released && transaction->t_checkpoint_io_list) {
                jh = transaction->t_checkpoint_io_list;
                bh = jh2bh(jh);
@@ -210,6 +215,9 @@ restart:
                        spin_lock(&journal->j_list_lock);
                        goto restart;
                }
+               if (unlikely(buffer_write_io_error(bh)))
+                       ret = -EIO;
+
                /*
                 * Now in whatever state the buffer currently is, we know that
                 * it has been written out and so we can drop it from the list
@@ -219,6 +227,8 @@ restart:
                jbd2_journal_remove_journal_head(bh);
                __brelse(bh);
        }
+
+       return ret;
 }
 
 #define NR_BATCH       64
@@ -242,7 +252,8 @@ __flush_batch(journal_t *journal, struct buffer_head **bhs, int *batch_count)
  * Try to flush one buffer from the checkpoint list to disk.
  *
  * Return 1 if something happened which requires us to abort the current
- * scan of the checkpoint list.
+ * scan of the checkpoint list.  Return <0 if the buffer has failed to
+ * be written out.
  *
  * Called with j_list_lock held and drops it if 1 is returned
  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
@@ -274,6 +285,9 @@ static int __process_buffer(journal_t *journal, struct journal_head *jh,
                jbd2_log_wait_commit(journal, tid);
                ret = 1;
        } else if (!buffer_dirty(bh)) {
+               ret = 1;
+               if (unlikely(buffer_write_io_error(bh)))
+                       ret = -EIO;
                J_ASSERT_JH(jh, !buffer_jbddirty(bh));
                BUFFER_TRACE(bh, "remove from checkpoint");
                __jbd2_journal_remove_checkpoint(jh);
@@ -281,7 +295,6 @@ static int __process_buffer(journal_t *journal, struct journal_head *jh,
                jbd_unlock_bh_state(bh);
                jbd2_journal_remove_journal_head(bh);
                __brelse(bh);
-               ret = 1;
        } else {
                /*
                 * Important: we are about to write the buffer, and
@@ -314,6 +327,7 @@ static int __process_buffer(journal_t *journal, struct journal_head *jh,
  * to disk. We submit larger chunks of data at once.
  *
  * The journal should be locked before calling this function.
+ * Called with j_checkpoint_mutex held.
  */
 int jbd2_log_do_checkpoint(journal_t *journal)
 {
@@ -339,6 +353,7 @@ int jbd2_log_do_checkpoint(journal_t *journal)
         * OK, we need to start writing disk blocks.  Take one transaction
         * and write it.
         */
+       result = 0;
        spin_lock(&journal->j_list_lock);
        if (!journal->j_checkpoint_transactions)
                goto out;
@@ -357,7 +372,7 @@ restart:
                int batch_count = 0;
                struct buffer_head *bhs[NR_BATCH];
                struct journal_head *jh;
-               int retry = 0;
+               int retry = 0, err;
 
                while (!retry && transaction->t_checkpoint_list) {
                        struct buffer_head *bh;
@@ -371,6 +386,8 @@ restart:
                        }
                        retry = __process_buffer(journal, jh, bhs, &batch_count,
                                                 transaction);
+                       if (retry < 0 && !result)
+                               result = retry;
                        if (!retry && (need_resched() ||
                                spin_needbreak(&journal->j_list_lock))) {
                                spin_unlock(&journal->j_list_lock);
@@ -395,14 +412,18 @@ restart:
                 * Now we have cleaned up the first transaction's checkpoint
                 * list. Let's clean up the second one
                 */
-               __wait_cp_io(journal, transaction);
+               err = __wait_cp_io(journal, transaction);
+               if (!result)
+                       result = err;
        }
 out:
        spin_unlock(&journal->j_list_lock);
-       result = jbd2_cleanup_journal_tail(journal);
        if (result < 0)
-               return result;
-       return 0;
+               jbd2_journal_abort(journal, result);
+       else
+               result = jbd2_cleanup_journal_tail(journal);
+
+       return (result < 0) ? result : 0;
 }
 
 /*
@@ -418,8 +439,9 @@ out:
  * This is the only part of the journaling code which really needs to be
  * aware of transaction aborts.  Checkpointing involves writing to the
  * main filesystem area rather than to the journal, so it can proceed
- * even in abort state, but we must not update the journal superblock if
- * we have an abort error outstanding.
+ * even in abort state, but we must not update the super block if
+ * checkpointing may have failed.  Otherwise, we would lose some metadata
+ * buffers which should be written-back to the filesystem.
  */
 
 int jbd2_cleanup_journal_tail(journal_t *journal)
@@ -428,6 +450,9 @@ int jbd2_cleanup_journal_tail(journal_t *journal)
        tid_t           first_tid;
        unsigned long   blocknr, freed;
 
+       if (is_journal_aborted(journal))
+               return 1;
+
        /* OK, work out the oldest transaction remaining in the log, and
         * the log block it starts at.
         *
index 01c3901..783de11 100644 (file)
@@ -1451,9 +1451,12 @@ recovery_error:
  *
  * Release a journal_t structure once it is no longer in use by the
  * journaled object.
+ * Return <0 if we couldn't clean up the journal.
  */
-void jbd2_journal_destroy(journal_t *journal)
+int jbd2_journal_destroy(journal_t *journal)
 {
+       int err = 0;
+
        /* Wait for the commit thread to wake up and die. */
        journal_kill_thread(journal);
 
@@ -1476,11 +1479,16 @@ void jbd2_journal_destroy(journal_t *journal)
        J_ASSERT(journal->j_checkpoint_transactions == NULL);
        spin_unlock(&journal->j_list_lock);
 
-       /* We can now mark the journal as empty. */
-       journal->j_tail = 0;
-       journal->j_tail_sequence = ++journal->j_transaction_sequence;
        if (journal->j_sb_buffer) {
-               jbd2_journal_update_superblock(journal, 1);
+               if (!is_journal_aborted(journal)) {
+                       /* We can now mark the journal as empty. */
+                       journal->j_tail = 0;
+                       journal->j_tail_sequence =
+                               ++journal->j_transaction_sequence;
+                       jbd2_journal_update_superblock(journal, 1);
+               } else {
+                       err = -EIO;
+               }
                brelse(journal->j_sb_buffer);
        }
 
@@ -1492,6 +1500,8 @@ void jbd2_journal_destroy(journal_t *journal)
                jbd2_journal_destroy_revoke(journal);
        kfree(journal->j_wbuf);
        kfree(journal);
+
+       return err;
 }
 
 
@@ -1717,10 +1727,16 @@ int jbd2_journal_flush(journal_t *journal)
        spin_lock(&journal->j_list_lock);
        while (!err && journal->j_checkpoint_transactions != NULL) {
                spin_unlock(&journal->j_list_lock);
+               mutex_lock(&journal->j_checkpoint_mutex);
                err = jbd2_log_do_checkpoint(journal);
+               mutex_unlock(&journal->j_checkpoint_mutex);
                spin_lock(&journal->j_list_lock);
        }
        spin_unlock(&journal->j_list_lock);
+
+       if (is_journal_aborted(journal))
+               return -EIO;
+
        jbd2_cleanup_journal_tail(journal);
 
        /* Finally, mark the journal as really needing no recovery.
@@ -1742,7 +1758,7 @@ int jbd2_journal_flush(journal_t *journal)
        J_ASSERT(journal->j_head == journal->j_tail);
        J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
        spin_unlock(&journal->j_state_lock);
-       return err;
+       return 0;
 }
 
 /**
index 058f50f..7306328 100644 (file)
@@ -225,7 +225,7 @@ do {                                                                        \
  */
 int jbd2_journal_recover(journal_t *journal)
 {
-       int                     err;
+       int                     err, err2;
        journal_superblock_t *  sb;
 
        struct recovery_info    info;
@@ -263,7 +263,10 @@ int jbd2_journal_recover(journal_t *journal)
        journal->j_transaction_sequence = ++info.end_transaction;
 
        jbd2_journal_clear_revoke(journal);
-       sync_blockdev(journal->j_fs_dev);
+       err2 = sync_blockdev(journal->j_fs_dev);
+       if (!err)
+               err = err2;
+
        return err;
 }
 
index 66c3499..c9e7d78 100644 (file)
@@ -1060,7 +1060,7 @@ extern void          jbd2_journal_clear_features
                   (journal_t *, unsigned long, unsigned long, unsigned long);
 extern int        jbd2_journal_create     (journal_t *);
 extern int        jbd2_journal_load       (journal_t *journal);
-extern void       jbd2_journal_destroy    (journal_t *);
+extern int        jbd2_journal_destroy    (journal_t *);
 extern int        jbd2_journal_recover    (journal_t *journal);
 extern int        jbd2_journal_wipe       (journal_t *, int);
 extern int        jbd2_journal_skip_recovery   (journal_t *);