]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/blobdiff - fs/jbd/checkpoint.c
jbd: test BH_Write_EIO to detect errors on metadata buffers
[linux-2.6-omap-h63xx.git] / fs / jbd / checkpoint.c
index d0685596e5a66df8bfd7bbf10f4e3d8835942cf4..fe852193324302c3d1272b6a0ebe6d7ee4fc1122 100644 (file)
@@ -1,6 +1,6 @@
 /*
- * linux/fs/checkpoint.c
- * 
+ * linux/fs/jbd/checkpoint.c
+ *
  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
  *
  * Copyright 1999 Red Hat Software --- All Rights Reserved
@@ -9,8 +9,8 @@
  * the terms of the GNU General Public License, version 2, or at your
  * option, any later version, incorporated herein by reference.
  *
- * Checkpoint routines for the generic filesystem journaling code.  
- * Part of the ext2fs journaling system.  
+ * Checkpoint routines for the generic filesystem journaling code.
+ * Part of the ext2fs journaling system.
  *
  * Checkpointing is the process of ensuring that a section of the log is
  * committed fully to disk, so that that portion of the log can be
@@ -93,7 +93,8 @@ 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 = __journal_remove_checkpoint(jh) + 1;
                jbd_unlock_bh_state(bh);
@@ -160,21 +161,25 @@ 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);
@@ -194,6 +199,9 @@ 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
@@ -203,6 +211,8 @@ restart:
                journal_remove_journal_head(bh);
                __brelse(bh);
        }
+
+       return ret;
 }
 
 #define NR_BATCH       64
@@ -226,7 +236,8 @@ __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
@@ -256,6 +267,9 @@ static int __process_buffer(journal_t *journal, struct journal_head *jh,
                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");
                __journal_remove_checkpoint(jh);
@@ -263,14 +277,13 @@ static int __process_buffer(journal_t *journal, struct journal_head *jh,
                jbd_unlock_bh_state(bh);
                journal_remove_journal_head(bh);
                __brelse(bh);
-               ret = 1;
        } else {
                /*
                 * Important: we are about to write the buffer, and
                 * possibly block, while still holding the journal lock.
                 * We cannot afford to let the transaction logic start
                 * messing around with this buffer before we write it to
-                * disk, as that would break recoverability.  
+                * disk, as that would break recoverability.
                 */
                BUFFER_TRACE(bh, "queue");
                get_bh(bh);
@@ -293,8 +306,9 @@ static int __process_buffer(journal_t *journal, struct journal_head *jh,
  * Perform an actual checkpoint. We take the first transaction on the
  * list of transactions to be checkpointed and send all its buffers
  * 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 log_do_checkpoint(journal_t *journal)
 {
@@ -304,10 +318,10 @@ int log_do_checkpoint(journal_t *journal)
 
        jbd_debug(1, "Start checkpoint\n");
 
-       /* 
+       /*
         * First thing: if there are any transactions in the log which
         * don't need checkpointing, just eliminate them from the
-        * journal straight away.  
+        * journal straight away.
         */
        result = cleanup_journal_tail(journal);
        jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
@@ -318,6 +332,7 @@ int 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;
@@ -334,7 +349,7 @@ 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;
@@ -347,7 +362,10 @@ restart:
                                break;
                        }
                        retry = __process_buffer(journal, jh, bhs,&batch_count);
-                       if (!retry && lock_need_resched(&journal->j_list_lock)){
+                       if (retry < 0 && !result)
+                               result = retry;
+                       if (!retry && (need_resched() ||
+                               spin_needbreak(&journal->j_list_lock))) {
                                spin_unlock(&journal->j_list_lock);
                                retry = 1;
                                break;
@@ -370,14 +388,18 @@ 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 = cleanup_journal_tail(journal);
        if (result < 0)
-               return result;
-       return 0;
+               journal_abort(journal, result);
+       else
+               result = cleanup_journal_tail(journal);
+
+       return (result < 0) ? result : 0;
 }
 
 /*
@@ -385,16 +407,17 @@ out:
  * we have already got rid of any since the last update of the log tail
  * in the journal superblock.  If so, we can instantly roll the
  * superblock forward to remove those transactions from the log.
- * 
+ *
  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
- * 
+ *
  * Called with the journal lock held.
  *
  * 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 cleanup_journal_tail(journal_t *journal)
@@ -403,9 +426,12 @@ int 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. 
-        * 
+        * the log block it starts at.
+        *
         * If the log is now empty, we need to work out which is the
         * next transaction ID we will write, and where it will
         * start. */
@@ -480,7 +506,7 @@ static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
        if (!jh)
                return 0;
 
-       last_jh = jh->b_cpprev;
+       last_jh = jh->b_cpprev;
        do {
                jh = next_jh;
                next_jh = jh->b_cpnext;
@@ -558,7 +584,7 @@ out:
        return ret;
 }
 
-/* 
+/*
  * journal_remove_checkpoint: called after a buffer has been committed
  * to disk (either by being write-back flushed to disk, or being
  * committed to the log).
@@ -602,15 +628,15 @@ int __journal_remove_checkpoint(struct journal_head *jh)
 
        /*
         * There is one special case to worry about: if we have just pulled the
-        * buffer off a committing transaction's forget list, then even if the
-        * checkpoint list is empty, the transaction obviously cannot be
-        * dropped!
+        * buffer off a running or committing transaction's checkpoing list,
+        * then even if the checkpoint list is empty, the transaction obviously
+        * cannot be dropped!
         *
-        * The locking here around j_committing_transaction is a bit sleazy.
+        * The locking here around t_state is a bit sleazy.
         * See the comment at the end of journal_commit_transaction().
         */
-       if (transaction == journal->j_committing_transaction) {
-               JBUFFER_TRACE(jh, "belongs to committing transaction");
+       if (transaction->t_state != T_FINISHED) {
+               JBUFFER_TRACE(jh, "belongs to running/committing transaction");
                goto out;
        }
 
@@ -636,7 +662,7 @@ out:
  * Called with the journal locked.
  * Called with j_list_lock held.
  */
-void __journal_insert_checkpoint(struct journal_head *jh, 
+void __journal_insert_checkpoint(struct journal_head *jh,
                               transaction_t *transaction)
 {
        JBUFFER_TRACE(jh, "entry");
@@ -658,7 +684,7 @@ void __journal_insert_checkpoint(struct journal_head *jh,
 
 /*
  * We've finished with this transaction structure: adios...
- * 
+ *
  * The transaction must have no links except for the checkpoint by this
  * point.
  *