jbd2: fix the way the b_modified flag is cleared
[linux-2.6.git] / fs / jbd2 / commit.c
index 10be512..988fbec 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * linux/fs/jbd/commit.c
+ * linux/fs/jbd2/commit.c
  *
  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
  *
 
 #include <linux/time.h>
 #include <linux/fs.h>
-#include <linux/jbd.h>
+#include <linux/jbd2.h>
 #include <linux/errno.h>
 #include <linux/slab.h>
 #include <linux/mm.h>
 #include <linux/pagemap.h>
-#include <linux/smp_lock.h>
+#include <linux/jiffies.h>
+#include <linux/crc32.h>
 
 /*
  * Default IO end handler for temporary BJ_IO buffer_heads.
@@ -93,45 +94,63 @@ static int inverted_lock(journal_t *journal, struct buffer_head *bh)
        return 1;
 }
 
-/* Done it all: now write the commit record.  We should have
+/*
+ * Done it all: now submit the commit record.  We should have
  * cleaned up our previous buffers by now, so if we are in abort
  * mode we can now just skip the rest of the journal write
  * entirely.
  *
  * Returns 1 if the journal needs to be aborted or 0 on success
  */
-static int journal_write_commit_record(journal_t *journal,
-                                       transaction_t *commit_transaction)
+static int journal_submit_commit_record(journal_t *journal,
+                                       transaction_t *commit_transaction,
+                                       struct buffer_head **cbh,
+                                       __u32 crc32_sum)
 {
        struct journal_head *descriptor;
+       struct commit_header *tmp;
        struct buffer_head *bh;
-       int i, ret;
+       int ret;
        int barrier_done = 0;
 
        if (is_journal_aborted(journal))
                return 0;
 
-       descriptor = journal_get_descriptor_buffer(journal);
+       descriptor = jbd2_journal_get_descriptor_buffer(journal);
        if (!descriptor)
                return 1;
 
        bh = jh2bh(descriptor);
 
-       /* AKPM: buglet - add `i' to tmp! */
-       for (i = 0; i < bh->b_size; i += 512) {
-               journal_header_t *tmp = (journal_header_t*)bh->b_data;
-               tmp->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
-               tmp->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
-               tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
+       tmp = (struct commit_header *)bh->b_data;
+       tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
+       tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
+       tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
+
+       if (JBD2_HAS_COMPAT_FEATURE(journal,
+                                   JBD2_FEATURE_COMPAT_CHECKSUM)) {
+               tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
+               tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
+               tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
        }
 
-       JBUFFER_TRACE(descriptor, "write commit block");
+       JBUFFER_TRACE(descriptor, "submit commit block");
+       lock_buffer(bh);
+       get_bh(bh);
        set_buffer_dirty(bh);
-       if (journal->j_flags & JFS_BARRIER) {
+       set_buffer_uptodate(bh);
+       bh->b_end_io = journal_end_buffer_io_sync;
+
+       if (journal->j_flags & JBD2_BARRIER &&
+               !JBD2_HAS_INCOMPAT_FEATURE(journal,
+                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
                set_buffer_ordered(bh);
                barrier_done = 1;
        }
-       ret = sync_dirty_buffer(bh);
+       ret = submit_bh(WRITE, bh);
+       if (barrier_done)
+               clear_buffer_ordered(bh);
+
        /* is it possible for another commit to fail at roughly
         * the same time as this one?  If so, we don't want to
         * trust the barrier flag in the super, but instead want
@@ -145,21 +164,78 @@ static int journal_write_commit_record(journal_t *journal,
                        "disabling barriers\n",
                        bdevname(journal->j_dev, b));
                spin_lock(&journal->j_state_lock);
-               journal->j_flags &= ~JFS_BARRIER;
+               journal->j_flags &= ~JBD2_BARRIER;
                spin_unlock(&journal->j_state_lock);
 
                /* And try again, without the barrier */
-               clear_buffer_ordered(bh);
                set_buffer_uptodate(bh);
                set_buffer_dirty(bh);
-               ret = sync_dirty_buffer(bh);
+               ret = submit_bh(WRITE, bh);
        }
-       put_bh(bh);             /* One for getblk() */
-       journal_put_journal_head(descriptor);
+       *cbh = bh;
+       return ret;
+}
 
-       return (ret == -EIO);
+/*
+ * This function along with journal_submit_commit_record
+ * allows to write the commit record asynchronously.
+ */
+static int journal_wait_on_commit_record(struct buffer_head *bh)
+{
+       int ret = 0;
+
+       clear_buffer_dirty(bh);
+       wait_on_buffer(bh);
+
+       if (unlikely(!buffer_uptodate(bh)))
+               ret = -EIO;
+       put_bh(bh);            /* One for getblk() */
+       jbd2_journal_put_journal_head(bh2jh(bh));
+
+       return ret;
 }
 
+/*
+ * Wait for all submitted IO to complete.
+ */
+static int journal_wait_on_locked_list(journal_t *journal,
+                                      transaction_t *commit_transaction)
+{
+       int ret = 0;
+       struct journal_head *jh;
+
+       while (commit_transaction->t_locked_list) {
+               struct buffer_head *bh;
+
+               jh = commit_transaction->t_locked_list->b_tprev;
+               bh = jh2bh(jh);
+               get_bh(bh);
+               if (buffer_locked(bh)) {
+                       spin_unlock(&journal->j_list_lock);
+                       wait_on_buffer(bh);
+                       if (unlikely(!buffer_uptodate(bh)))
+                               ret = -EIO;
+                       spin_lock(&journal->j_list_lock);
+               }
+               if (!inverted_lock(journal, bh)) {
+                       put_bh(bh);
+                       spin_lock(&journal->j_list_lock);
+                       continue;
+               }
+               if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
+                       __jbd2_journal_unfile_buffer(jh);
+                       jbd_unlock_bh_state(bh);
+                       jbd2_journal_remove_journal_head(bh);
+                       put_bh(bh);
+               } else {
+                       jbd_unlock_bh_state(bh);
+               }
+               put_bh(bh);
+               cond_resched_lock(&journal->j_list_lock);
+       }
+       return ret;
+  }
+
 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
 {
        int i;
@@ -239,7 +315,7 @@ write_out_data:
                if (locked && test_clear_buffer_dirty(bh)) {
                        BUFFER_TRACE(bh, "needs writeout, adding to array");
                        wbuf[bufs++] = bh;
-                       __journal_file_buffer(jh, commit_transaction,
+                       __jbd2_journal_file_buffer(jh, commit_transaction,
                                                BJ_Locked);
                        jbd_unlock_bh_state(bh);
                        if (bufs == journal->j_wbufsize) {
@@ -248,21 +324,25 @@ write_out_data:
                                bufs = 0;
                                goto write_out_data;
                        }
-               }
-               else {
+               } else if (!locked && buffer_locked(bh)) {
+                       __jbd2_journal_file_buffer(jh, commit_transaction,
+                                               BJ_Locked);
+                       jbd_unlock_bh_state(bh);
+                       put_bh(bh);
+               } else {
                        BUFFER_TRACE(bh, "writeout complete: unfile");
-                       __journal_unfile_buffer(jh);
+                       __jbd2_journal_unfile_buffer(jh);
                        jbd_unlock_bh_state(bh);
                        if (locked)
                                unlock_buffer(bh);
-                       journal_remove_journal_head(bh);
+                       jbd2_journal_remove_journal_head(bh);
                        /* Once for our safety reference, once for
-                        * journal_remove_journal_head() */
+                        * jbd2_journal_remove_journal_head() */
                        put_bh(bh);
                        put_bh(bh);
                }
 
-               if (lock_need_resched(&journal->j_list_lock)) {
+               if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
                        spin_unlock(&journal->j_list_lock);
                        goto write_out_data;
                }
@@ -271,21 +351,44 @@ write_out_data:
        journal_do_submit_data(wbuf, bufs);
 }
 
+static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
+{
+       struct page *page = bh->b_page;
+       char *addr;
+       __u32 checksum;
+
+       addr = kmap_atomic(page, KM_USER0);
+       checksum = crc32_be(crc32_sum,
+               (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
+       kunmap_atomic(addr, KM_USER0);
+
+       return checksum;
+}
+
+static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
+                                  unsigned long long block)
+{
+       tag->t_blocknr = cpu_to_be32(block & (u32)~0);
+       if (tag_bytes > JBD2_TAG_SIZE32)
+               tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
+}
+
 /*
- * journal_commit_transaction
+ * jbd2_journal_commit_transaction
  *
  * The primary function for committing a transaction to the log.  This
  * function is called by the journal thread to begin a complete commit.
  */
-void journal_commit_transaction(journal_t *journal)
+void jbd2_journal_commit_transaction(journal_t *journal)
 {
+       struct transaction_stats_s stats;
        transaction_t *commit_transaction;
        struct journal_head *jh, *new_jh, *descriptor;
        struct buffer_head **wbuf = journal->j_wbuf;
        int bufs;
        int flags;
        int err;
-       unsigned long blocknr;
+       unsigned long long blocknr;
        char *tagp = NULL;
        journal_header_t *header;
        journal_block_tag_t *tag = NULL;
@@ -293,6 +396,9 @@ void journal_commit_transaction(journal_t *journal)
        int first_tag = 0;
        int tag_flag;
        int i;
+       int tag_bytes = journal_tag_bytes(journal);
+       struct buffer_head *cbh = NULL; /* For transactional checksums */
+       __u32 crc32_sum = ~0;
 
        /*
         * First job: lock down the current transaction and wait for
@@ -305,10 +411,10 @@ void journal_commit_transaction(journal_t *journal)
        spin_unlock(&journal->j_list_lock);
 #endif
 
-       /* Do we need to erase the effects of a prior journal_flush? */
-       if (journal->j_flags & JFS_FLUSHED) {
+       /* Do we need to erase the effects of a prior jbd2_journal_flush? */
+       if (journal->j_flags & JBD2_FLUSHED) {
                jbd_debug(3, "super block updated\n");
-               journal_update_superblock(journal, 1);
+               jbd2_journal_update_superblock(journal, 1);
        } else {
                jbd_debug(3, "superblock not updated\n");
        }
@@ -325,6 +431,11 @@ void journal_commit_transaction(journal_t *journal)
        spin_lock(&journal->j_state_lock);
        commit_transaction->t_state = T_LOCKED;
 
+       stats.u.run.rs_wait = commit_transaction->t_max_wait;
+       stats.u.run.rs_locked = jiffies;
+       stats.u.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
+                                               stats.u.run.rs_locked);
+
        spin_lock(&commit_transaction->t_handle_lock);
        while (commit_transaction->t_updates) {
                DEFINE_WAIT(wait);
@@ -350,7 +461,7 @@ void journal_commit_transaction(journal_t *journal)
         * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
         * that there are no such buffers: if a large filesystem
         * operation like a truncate needs to split itself over multiple
-        * transactions, then it may try to do a journal_restart() while
+        * transactions, then it may try to do a jbd2_journal_restart() while
         * there are still BJ_Reserved buffers outstanding.  These must
         * be released cleanly from the current transaction.
         *
@@ -358,25 +469,25 @@ void journal_commit_transaction(journal_t *journal)
         * again before modifying the buffer in the new transaction, but
         * we do not require it to remember exactly which old buffers it
         * has reserved.  This is consistent with the existing behaviour
-        * that multiple journal_get_write_access() calls to the same
+        * that multiple jbd2_journal_get_write_access() calls to the same
         * buffer are perfectly permissable.
         */
        while (commit_transaction->t_reserved_list) {
                jh = commit_transaction->t_reserved_list;
                JBUFFER_TRACE(jh, "reserved, unused: refile");
                /*
-                * A journal_get_undo_access()+journal_release_buffer() may
+                * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
                 * leave undo-committed data.
                 */
                if (jh->b_committed_data) {
                        struct buffer_head *bh = jh2bh(jh);
 
                        jbd_lock_bh_state(bh);
-                       jbd_slab_free(jh->b_committed_data, bh->b_size);
+                       jbd2_free(jh->b_committed_data, bh->b_size);
                        jh->b_committed_data = NULL;
                        jbd_unlock_bh_state(bh);
                }
-               journal_refile_buffer(journal, jh);
+               jbd2_journal_refile_buffer(journal, jh);
        }
 
        /*
@@ -385,7 +496,7 @@ void journal_commit_transaction(journal_t *journal)
         * frees some memory
         */
        spin_lock(&journal->j_list_lock);
-       __journal_clean_checkpoint_list(journal);
+       __jbd2_journal_clean_checkpoint_list(journal);
        spin_unlock(&journal->j_list_lock);
 
        jbd_debug (3, "JBD: commit phase 1\n");
@@ -393,7 +504,11 @@ void journal_commit_transaction(journal_t *journal)
        /*
         * Switch to a new revoke table.
         */
-       journal_switch_revoke_table(journal);
+       jbd2_journal_switch_revoke_table(journal);
+
+       stats.u.run.rs_flushing = jiffies;
+       stats.u.run.rs_locked = jbd2_time_diff(stats.u.run.rs_locked,
+                                              stats.u.run.rs_flushing);
 
        commit_transaction->t_state = T_FLUSH;
        journal->j_committing_transaction = commit_transaction;
@@ -405,22 +520,6 @@ void journal_commit_transaction(journal_t *journal)
        jbd_debug (3, "JBD: commit phase 2\n");
 
        /*
-        * First, drop modified flag: all accesses to the buffers
-        * will be tracked for a new trasaction only -bzzz
-        */
-       spin_lock(&journal->j_list_lock);
-       if (commit_transaction->t_buffers) {
-               new_jh = jh = commit_transaction->t_buffers->b_tnext;
-               do {
-                       J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||
-                                       new_jh->b_modified == 0);
-                       new_jh->b_modified = 0;
-                       new_jh = new_jh->b_tnext;
-               } while (new_jh != jh);
-       }
-       spin_unlock(&journal->j_list_lock);
-
-       /*
         * Now start flushing things to disk, in the order they appear
         * on the transaction lists.  Data blocks go first.
         */
@@ -428,44 +527,21 @@ void journal_commit_transaction(journal_t *journal)
        journal_submit_data_buffers(journal, commit_transaction);
 
        /*
-        * Wait for all previously submitted IO to complete.
+        * Wait for all previously submitted IO to complete if commit
+        * record is to be written synchronously.
         */
        spin_lock(&journal->j_list_lock);
-       while (commit_transaction->t_locked_list) {
-               struct buffer_head *bh;
+       if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
+               JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT))
+               err = journal_wait_on_locked_list(journal,
+                                               commit_transaction);
 
-               jh = commit_transaction->t_locked_list->b_tprev;
-               bh = jh2bh(jh);
-               get_bh(bh);
-               if (buffer_locked(bh)) {
-                       spin_unlock(&journal->j_list_lock);
-                       wait_on_buffer(bh);
-                       if (unlikely(!buffer_uptodate(bh)))
-                               err = -EIO;
-                       spin_lock(&journal->j_list_lock);
-               }
-               if (!inverted_lock(journal, bh)) {
-                       put_bh(bh);
-                       spin_lock(&journal->j_list_lock);
-                       continue;
-               }
-               if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
-                       __journal_unfile_buffer(jh);
-                       jbd_unlock_bh_state(bh);
-                       journal_remove_journal_head(bh);
-                       put_bh(bh);
-               } else {
-                       jbd_unlock_bh_state(bh);
-               }
-               put_bh(bh);
-               cond_resched_lock(&journal->j_list_lock);
-       }
        spin_unlock(&journal->j_list_lock);
 
        if (err)
-               __journal_abort_hard(journal);
+               jbd2_journal_abort(journal, err);
 
-       journal_write_revoke_records(journal, commit_transaction);
+       jbd2_journal_write_revoke_records(journal, commit_transaction);
 
        jbd_debug(3, "JBD: commit phase 2\n");
 
@@ -486,6 +562,12 @@ void journal_commit_transaction(journal_t *journal)
         */
        commit_transaction->t_state = T_COMMIT;
 
+       stats.u.run.rs_logging = jiffies;
+       stats.u.run.rs_flushing = jbd2_time_diff(stats.u.run.rs_flushing,
+                                                stats.u.run.rs_logging);
+       stats.u.run.rs_blocks = commit_transaction->t_outstanding_credits;
+       stats.u.run.rs_blocks_logged = 0;
+
        descriptor = NULL;
        bufs = 0;
        while (commit_transaction->t_buffers) {
@@ -499,7 +581,7 @@ void journal_commit_transaction(journal_t *journal)
 
                if (is_journal_aborted(journal)) {
                        JBUFFER_TRACE(jh, "journal is aborting: refile");
-                       journal_refile_buffer(journal, jh);
+                       jbd2_journal_refile_buffer(journal, jh);
                        /* If that was the last one, we need to clean up
                         * any descriptor buffers which may have been
                         * already allocated, even if we are now
@@ -519,9 +601,9 @@ void journal_commit_transaction(journal_t *journal)
 
                        jbd_debug(4, "JBD: get descriptor\n");
 
-                       descriptor = journal_get_descriptor_buffer(journal);
+                       descriptor = jbd2_journal_get_descriptor_buffer(journal);
                        if (!descriptor) {
-                               __journal_abort_hard(journal);
+                               jbd2_journal_abort(journal, -EIO);
                                continue;
                        }
 
@@ -529,8 +611,8 @@ void journal_commit_transaction(journal_t *journal)
                        jbd_debug(4, "JBD: got buffer %llu (%p)\n",
                                (unsigned long long)bh->b_blocknr, bh->b_data);
                        header = (journal_header_t *)&bh->b_data[0];
-                       header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
-                       header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
+                       header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
+                       header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
                        header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 
                        tagp = &bh->b_data[sizeof(journal_header_t)];
@@ -543,25 +625,25 @@ void journal_commit_transaction(journal_t *journal)
                        /* Record it so that we can wait for IO
                            completion later */
                        BUFFER_TRACE(bh, "ph3: file as descriptor");
-                       journal_file_buffer(descriptor, commit_transaction,
+                       jbd2_journal_file_buffer(descriptor, commit_transaction,
                                        BJ_LogCtl);
                }
 
                /* Where is the buffer to be written? */
 
-               err = journal_next_log_block(journal, &blocknr);
+               err = jbd2_journal_next_log_block(journal, &blocknr);
                /* If the block mapping failed, just abandon the buffer
                   and repeat this loop: we'll fall into the
                   refile-on-abort condition above. */
                if (err) {
-                       __journal_abort_hard(journal);
+                       jbd2_journal_abort(journal, err);
                        continue;
                }
 
                /*
                 * start_this_handle() uses t_outstanding_credits to determine
                 * the free space in the log, but this counter is changed
-                * by journal_next_log_block() also.
+                * by jbd2_journal_next_log_block() also.
                 */
                commit_transaction->t_outstanding_credits--;
 
@@ -576,13 +658,13 @@ void journal_commit_transaction(journal_t *journal)
 
                set_bit(BH_JWrite, &jh2bh(jh)->b_state);
                /*
-                * akpm: journal_write_metadata_buffer() sets
+                * akpm: jbd2_journal_write_metadata_buffer() sets
                 * new_bh->b_transaction to commit_transaction.
                 * We need to clean this up before we release new_bh
                 * (which is of type BJ_IO)
                 */
                JBUFFER_TRACE(jh, "ph3: write metadata");
-               flags = journal_write_metadata_buffer(commit_transaction,
+               flags = jbd2_journal_write_metadata_buffer(commit_transaction,
                                                      jh, &new_jh, blocknr);
                set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
                wbuf[bufs++] = jh2bh(new_jh);
@@ -592,15 +674,15 @@ void journal_commit_transaction(journal_t *journal)
 
                tag_flag = 0;
                if (flags & 1)
-                       tag_flag |= JFS_FLAG_ESCAPE;
+                       tag_flag |= JBD2_FLAG_ESCAPE;
                if (!first_tag)
-                       tag_flag |= JFS_FLAG_SAME_UUID;
+                       tag_flag |= JBD2_FLAG_SAME_UUID;
 
                tag = (journal_block_tag_t *) tagp;
-               tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
+               write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
                tag->t_flags = cpu_to_be32(tag_flag);
-               tagp += sizeof(journal_block_tag_t);
-               space_left -= sizeof(journal_block_tag_t);
+               tagp += tag_bytes;
+               space_left -= tag_bytes;
 
                if (first_tag) {
                        memcpy (tagp, journal->j_uuid, 16);
@@ -614,7 +696,7 @@ void journal_commit_transaction(journal_t *journal)
 
                if (bufs == journal->j_wbufsize ||
                    commit_transaction->t_buffers == NULL ||
-                   space_left < sizeof(journal_block_tag_t) + 16) {
+                   space_left < tag_bytes + 16) {
 
                        jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 
@@ -622,11 +704,20 @@ void journal_commit_transaction(journal_t *journal)
                            submitting the IOs.  "tag" still points to
                            the last tag we set up. */
 
-                       tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
+                       tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
 
 start_journal_io:
                        for (i = 0; i < bufs; i++) {
                                struct buffer_head *bh = wbuf[i];
+                               /*
+                                * Compute checksum.
+                                */
+                               if (JBD2_HAS_COMPAT_FEATURE(journal,
+                                       JBD2_FEATURE_COMPAT_CHECKSUM)) {
+                                       crc32_sum =
+                                           jbd2_checksum_data(crc32_sum, bh);
+                               }
+
                                lock_buffer(bh);
                                clear_buffer_dirty(bh);
                                set_buffer_uptodate(bh);
@@ -634,6 +725,7 @@ start_journal_io:
                                submit_bh(WRITE, bh);
                        }
                        cond_resched();
+                       stats.u.run.rs_blocks_logged += bufs;
 
                        /* Force a new descriptor to be generated next
                            time round the loop. */
@@ -642,6 +734,23 @@ start_journal_io:
                }
        }
 
+       /* Done it all: now write the commit record asynchronously. */
+
+       if (JBD2_HAS_INCOMPAT_FEATURE(journal,
+               JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
+               err = journal_submit_commit_record(journal, commit_transaction,
+                                                &cbh, crc32_sum);
+               if (err)
+                       __jbd2_journal_abort_hard(journal);
+
+               spin_lock(&journal->j_list_lock);
+               err = journal_wait_on_locked_list(journal,
+                                               commit_transaction);
+               spin_unlock(&journal->j_list_lock);
+               if (err)
+                       __jbd2_journal_abort_hard(journal);
+       }
+
        /* Lo and behold: we have just managed to send a transaction to
            the log.  Before we can commit it, wait for the IO so far to
            complete.  Control buffers being written are on the
@@ -678,14 +787,14 @@ wait_for_iobuf:
                clear_buffer_jwrite(bh);
 
                JBUFFER_TRACE(jh, "ph4: unfile after journal write");
-               journal_unfile_buffer(journal, jh);
+               jbd2_journal_unfile_buffer(journal, jh);
 
                /*
                 * ->t_iobuf_list should contain only dummy buffer_heads
-                * which were created by journal_write_metadata_buffer().
+                * which were created by jbd2_journal_write_metadata_buffer().
                 */
                BUFFER_TRACE(bh, "dumping temporary bh");
-               journal_put_journal_head(jh);
+               jbd2_journal_put_journal_head(jh);
                __brelse(bh);
                J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
                free_buffer_head(bh);
@@ -702,7 +811,7 @@ wait_for_iobuf:
                    we finally commit, we can do any checkpointing
                    required. */
                JBUFFER_TRACE(jh, "file as BJ_Forget");
-               journal_file_buffer(jh, commit_transaction, BJ_Forget);
+               jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
                /* Wake up any transactions which were waiting for this
                   IO to complete */
                wake_up_bit(&bh->b_state, BH_Unshadow);
@@ -733,19 +842,26 @@ wait_for_iobuf:
 
                BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
                clear_buffer_jwrite(bh);
-               journal_unfile_buffer(journal, jh);
-               journal_put_journal_head(jh);
+               jbd2_journal_unfile_buffer(journal, jh);
+               jbd2_journal_put_journal_head(jh);
                __brelse(bh);           /* One for getblk */
                /* AKPM: bforget here */
        }
 
        jbd_debug(3, "JBD: commit phase 6\n");
 
-       if (journal_write_commit_record(journal, commit_transaction))
-               err = -EIO;
+       if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
+               JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
+               err = journal_submit_commit_record(journal, commit_transaction,
+                                               &cbh, crc32_sum);
+               if (err)
+                       __jbd2_journal_abort_hard(journal);
+       }
+       if (!err && !is_journal_aborted(journal))
+               err = journal_wait_on_commit_record(cbh);
 
        if (err)
-               __journal_abort_hard(journal);
+               jbd2_journal_abort(journal, err);
 
        /* End of a transaction!  Finally, we can do checkpoint
            processing: any buffers committed as a result of this
@@ -789,14 +905,14 @@ restart_loop:
                 * Otherwise, we can just throw away the frozen data now.
                 */
                if (jh->b_committed_data) {
-                       jbd_slab_free(jh->b_committed_data, bh->b_size);
+                       jbd2_free(jh->b_committed_data, bh->b_size);
                        jh->b_committed_data = NULL;
                        if (jh->b_frozen_data) {
                                jh->b_committed_data = jh->b_frozen_data;
                                jh->b_frozen_data = NULL;
                        }
                } else if (jh->b_frozen_data) {
-                       jbd_slab_free(jh->b_frozen_data, bh->b_size);
+                       jbd2_free(jh->b_frozen_data, bh->b_size);
                        jh->b_frozen_data = NULL;
                }
 
@@ -804,12 +920,13 @@ restart_loop:
                cp_transaction = jh->b_cp_transaction;
                if (cp_transaction) {
                        JBUFFER_TRACE(jh, "remove from old cp transaction");
-                       __journal_remove_checkpoint(jh);
+                       cp_transaction->t_chp_stats.cs_dropped++;
+                       __jbd2_journal_remove_checkpoint(jh);
                }
 
                /* Only re-checkpoint the buffer_head if it is marked
                 * dirty.  If the buffer was added to the BJ_Forget list
-                * by journal_forget, it may no longer be dirty and
+                * by jbd2_journal_forget, it may no longer be dirty and
                 * there's no point in keeping a checkpoint record for
                 * it. */
 
@@ -828,9 +945,9 @@ restart_loop:
 
                if (buffer_jbddirty(bh)) {
                        JBUFFER_TRACE(jh, "add to new checkpointing trans");
-                       __journal_insert_checkpoint(jh, commit_transaction);
+                       __jbd2_journal_insert_checkpoint(jh, commit_transaction);
                        JBUFFER_TRACE(jh, "refile for checkpoint writeback");
-                       __journal_refile_buffer(jh);
+                       __jbd2_journal_refile_buffer(jh);
                        jbd_unlock_bh_state(bh);
                } else {
                        J_ASSERT_BH(bh, !buffer_dirty(bh));
@@ -842,11 +959,11 @@ restart_loop:
                         * disk and before we process the buffer on BJ_Forget
                         * list. */
                        JBUFFER_TRACE(jh, "refile or unfile freed buffer");
-                       __journal_refile_buffer(jh);
+                       __jbd2_journal_refile_buffer(jh);
                        if (!jh->b_transaction) {
                                jbd_unlock_bh_state(bh);
                                 /* needs a brelse */
-                               journal_remove_journal_head(bh);
+                               jbd2_journal_remove_journal_head(bh);
                                release_buffer_page(bh);
                        } else
                                jbd_unlock_bh_state(bh);
@@ -855,10 +972,10 @@ restart_loop:
        }
        spin_unlock(&journal->j_list_lock);
        /*
-        * This is a bit sleazy.  We borrow j_list_lock to protect
-        * journal->j_committing_transaction in __journal_remove_checkpoint.
-        * Really, __journal_remove_checkpoint should be using j_state_lock but
-        * it's a bit hassle to hold that across __journal_remove_checkpoint
+        * This is a bit sleazy.  We use j_list_lock to protect transition
+        * of a transaction into T_FINISHED state and calling
+        * __jbd2_journal_drop_transaction(). Otherwise we could race with
+        * other checkpointing code processing the transaction...
         */
        spin_lock(&journal->j_state_lock);
        spin_lock(&journal->j_list_lock);
@@ -878,14 +995,45 @@ restart_loop:
 
        J_ASSERT(commit_transaction->t_state == T_COMMIT);
 
+       commit_transaction->t_start = jiffies;
+       stats.u.run.rs_logging = jbd2_time_diff(stats.u.run.rs_logging,
+                                               commit_transaction->t_start);
+
+       /*
+        * File the transaction for history
+        */
+       stats.ts_type = JBD2_STATS_RUN;
+       stats.ts_tid = commit_transaction->t_tid;
+       stats.u.run.rs_handle_count = commit_transaction->t_handle_count;
+       spin_lock(&journal->j_history_lock);
+       memcpy(journal->j_history + journal->j_history_cur, &stats,
+                       sizeof(stats));
+       if (++journal->j_history_cur == journal->j_history_max)
+               journal->j_history_cur = 0;
+
+       /*
+        * Calculate overall stats
+        */
+       journal->j_stats.ts_tid++;
+       journal->j_stats.u.run.rs_wait += stats.u.run.rs_wait;
+       journal->j_stats.u.run.rs_running += stats.u.run.rs_running;
+       journal->j_stats.u.run.rs_locked += stats.u.run.rs_locked;
+       journal->j_stats.u.run.rs_flushing += stats.u.run.rs_flushing;
+       journal->j_stats.u.run.rs_logging += stats.u.run.rs_logging;
+       journal->j_stats.u.run.rs_handle_count += stats.u.run.rs_handle_count;
+       journal->j_stats.u.run.rs_blocks += stats.u.run.rs_blocks;
+       journal->j_stats.u.run.rs_blocks_logged += stats.u.run.rs_blocks_logged;
+       spin_unlock(&journal->j_history_lock);
+
        commit_transaction->t_state = T_FINISHED;
        J_ASSERT(commit_transaction == journal->j_committing_transaction);
        journal->j_commit_sequence = commit_transaction->t_tid;
        journal->j_committing_transaction = NULL;
        spin_unlock(&journal->j_state_lock);
 
-       if (commit_transaction->t_checkpoint_list == NULL) {
-               __journal_drop_transaction(journal, commit_transaction);
+       if (commit_transaction->t_checkpoint_list == NULL &&
+           commit_transaction->t_checkpoint_io_list == NULL) {
+               __jbd2_journal_drop_transaction(journal, commit_transaction);
        } else {
                if (journal->j_checkpoint_transactions == NULL) {
                        journal->j_checkpoint_transactions = commit_transaction;