[PATCH] Fix reiserfs latencies caused by data=ordered
authorChris Mason <mason@suse.com>
Fri, 29 Sep 2006 08:59:56 +0000 (01:59 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Fri, 29 Sep 2006 16:18:11 +0000 (09:18 -0700)
ReiserFS does periodic cleanup of old transactions in order to limit the
length of time a journal replay may take after a crash.  Sometimes, writing
metadata from an old (already committed) transaction may require committing
a newer transaction, which also requires writing all data=ordered buffers.
This can cause very long stalls on journal_begin.

This patch makes sure new transactions will not need to be committed before
trying a periodic reclaim of an old transaction.  It is low risk because if
a bad decision is made, it just means a slightly longer journal replay
after a crash.

Signed-off-by: Chris Mason <mason@suse.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
fs/reiserfs/journal.c

index 9b3672d69367251b6109c6550c5e8e7f81986634..e6b5ccf23f152178eb2981423ae3218f1abab6c0 100644 (file)
@@ -1186,6 +1186,21 @@ static struct reiserfs_journal_list *find_newer_jl_for_cn(struct
        return NULL;
 }
 
+static int newer_jl_done(struct reiserfs_journal_cnode *cn)
+{
+       struct super_block *sb = cn->sb;
+       b_blocknr_t blocknr = cn->blocknr;
+
+       cn = cn->hprev;
+       while (cn) {
+               if (cn->sb == sb && cn->blocknr == blocknr && cn->jlist &&
+                   atomic_read(&cn->jlist->j_commit_left) != 0)
+                                   return 0;
+               cn = cn->hprev;
+       }
+       return 1;
+}
+
 static void remove_journal_hash(struct super_block *,
                                struct reiserfs_journal_cnode **,
                                struct reiserfs_journal_list *, unsigned long,
@@ -1604,6 +1619,31 @@ static int flush_journal_list(struct super_block *s,
        return err;
 }
 
+static int test_transaction(struct super_block *s,
+                            struct reiserfs_journal_list *jl)
+{
+       struct reiserfs_journal_cnode *cn;
+
+       if (jl->j_len == 0 || atomic_read(&jl->j_nonzerolen) == 0)
+               return 1;
+
+       cn = jl->j_realblock;
+       while (cn) {
+               /* if the blocknr == 0, this has been cleared from the hash,
+                ** skip it
+                */
+               if (cn->blocknr == 0) {
+                       goto next;
+               }
+               if (cn->bh && !newer_jl_done(cn))
+                       return 0;
+             next:
+               cn = cn->next;
+               cond_resched();
+       }
+       return 0;
+}
+
 static int write_one_transaction(struct super_block *s,
                                 struct reiserfs_journal_list *jl,
                                 struct buffer_chunk *chunk)
@@ -3433,16 +3473,6 @@ static void flush_async_commits(void *p)
                flush_commit_list(p_s_sb, jl, 1);
        }
        unlock_kernel();
-       /*
-        * this is a little racey, but there's no harm in missing
-        * the filemap_fdata_write
-        */
-       if (!atomic_read(&journal->j_async_throttle)
-           && !reiserfs_is_journal_aborted(journal)) {
-               atomic_inc(&journal->j_async_throttle);
-               filemap_fdatawrite(p_s_sb->s_bdev->bd_inode->i_mapping);
-               atomic_dec(&journal->j_async_throttle);
-       }
 }
 
 /*
@@ -3844,7 +3874,9 @@ static void flush_old_journal_lists(struct super_block *s)
                entry = journal->j_journal_list.next;
                jl = JOURNAL_LIST_ENTRY(entry);
                /* this check should always be run, to send old lists to disk */
-               if (jl->j_timestamp < (now - (JOURNAL_MAX_TRANS_AGE * 4))) {
+               if (jl->j_timestamp < (now - (JOURNAL_MAX_TRANS_AGE * 4)) &&
+                   atomic_read(&jl->j_commit_left) == 0 &&
+                   test_transaction(s, jl)) {
                        flush_used_journal_lists(s, jl);
                } else {
                        break;