Branch:b1_6

The catlog should be initialized anyway even in recovery.
b=12086
i=nathan
i=adilger

diff --git a/lustre/ChangeLog b/lustre/ChangeLog
index 99c3a94..ae83eb2 100644 (file)
--- a/lustre/ChangeLog
+++ b/lustre/ChangeLog
@@ -275,6 +275,16 @@ Description: service threads may hog cpus when ther eis a lot of requests
 Details    : Insert cond_resched to give other threads a chance to use some of
              the cpu
 
+Severity   : normal
+Frequency  : rare
+Bugzilla   : 12086 
+Description: the cat log was not initialized in recovery 
+Details    : When mds(mgs) do recovery, the tgt_count might be zero, so the
+            unlink log on mds will not be initialized until mds post
+            recovery. And also in mds post recovery, the unlink log will 
+            initialization will be done asynchronausly, so there will be race
+            between add unlink log and unlink log initialization.
+
 --------------------------------------------------------------------------------
 
 2007-05-03  Cluster File Systems, Inc. <info@clusterfs.com>

diff --git a/lustre/mds/mds_lov.c b/lustre/mds/mds_lov.c
index a411b89..8696611 100644 (file)
--- a/lustre/mds/mds_lov.c
+++ b/lustre/mds/mds_lov.c
@@ -235,6 +235,13 @@ static int mds_lov_update_desc(struct obd_device *obd, struct obd_export *lov)
                "%d/%d\n", mds->mds_max_mdsize, mds->mds_max_cookiesize,
                stripes);
 
+        /* If we added a target we have to reconnect the llogs */
+        /* We only _need_ to do this at first add (idx), or the first time
+           after recovery.  However, it should now be safe to call anytime. */
+        mutex_down(&obd->obd_dev_sem);
+        llog_cat_initialize(obd, mds->mds_lov_desc.ld_tgt_count, NULL);
+        mutex_up(&obd->obd_dev_sem);
+
 out:
         OBD_FREE(ld, sizeof(*ld));
         RETURN(rc);
@@ -292,14 +299,6 @@ static int mds_lov_update_mds(struct obd_device *obd,
                       mds->mds_lov_objids[idx], idx);
         }
 
-        /* If we added a target we have to reconnect the llogs */
-        /* We only _need_ to do this at first add (idx), or the first time
-           after recovery.  However, it should now be safe to call anytime. */
-        CDEBUG(D_CONFIG, "reset llogs idx=%d\n", idx);
-        mutex_down(&obd->obd_dev_sem);
-        llog_cat_initialize(obd, mds->mds_lov_desc.ld_tgt_count, uuid);
-        mutex_up(&obd->obd_dev_sem);
-
         RETURN(rc);
 }
 

diff --git a/lustre/tests/replay-single.sh b/lustre/tests/replay-single.sh
index 32ffd8b..73f6cab 100755 (executable)
--- a/lustre/tests/replay-single.sh
+++ b/lustre/tests/replay-single.sh
@@ -1141,6 +1141,19 @@ test_59() {
 }
 run_test 59 "test log_commit_thread vs filter_destroy race"
 
+# race between add unlink llog vs cat log init in post_recovery (only for b1_6)
+# bug 12086: should no oops and No ctxt error for this test
+test_60() {
+    mkdir $DIR/$tdir
+    createmany -o $DIR/$tdir/$tfile-%d 200
+    replay_barrier mds
+    unlinkmany $DIR/$tdir/$tfile-%d 0 100
+    fail mds
+    unlinkmany $DIR/$tdir/$tfile-%d 100 100
+    local no_ctxt=`dmesg | grep "No ctxt"`
+    [ -z "$no_ctxt" ] || error "ctxt is not initialized in recovery" 
+}
+run_test 60 "test llog post recovery init vs llog unlink"
 
 equals_msg `basename $0`: test complete, cleaning up
 $CLEANUP