ocfs2/dlm: Fix a race between migrate request and exit domain

Patch address a racing migrate request message and an exit domain message.
Instead of blocking exit domains for the duration of the migrate, we ignore
failure to deliver that message. This is because an exiting domain should
not have any active locks and thus has no role to play in the migration.

Signed-off-by: Sunil Mushran <sunil.mushran@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>

diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
index 44f87caf36834fda999768a06448953afcaa6ef4..92fd1d7d61266b64afdf6842986c0ee51e1d7e54 100644 (file)
--- a/fs/ocfs2/dlm/dlmmaster.c
+++ b/fs/ocfs2/dlm/dlmmaster.c
@@ -2949,7 +2949,7 @@ static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
                                  struct dlm_node_iter *iter)
 {
        struct dlm_migrate_request migrate;
-       int ret, status = 0;
+       int ret, skip, status = 0;
        int nodenum;
 
        memset(&migrate, 0, sizeof(migrate));
@@ -2966,12 +2966,27 @@ static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
                    nodenum == new_master)
                        continue;
 
+               /* We could race exit domain. If exited, skip. */
+               spin_lock(&dlm->spinlock);
+               skip = (!test_bit(nodenum, dlm->domain_map));
+               spin_unlock(&dlm->spinlock);
+               if (skip) {
+                       clear_bit(nodenum, iter->node_map);
+                       continue;
+               }
+
                ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
                                         &migrate, sizeof(migrate), nodenum,
                                         &status);
-               if (ret < 0)
-                       mlog_errno(ret);
-               else if (status < 0) {
+               if (ret < 0) {
+                       mlog(0, "migrate_request returned %d!\n", ret);
+                       if (!dlm_is_host_down(ret)) {
+                               mlog(ML_ERROR, "unhandled error=%d!\n", ret);
+                               BUG();
+                       }
+                       clear_bit(nodenum, iter->node_map);
+                       ret = 0;
+               } else if (status < 0) {
                        mlog(0, "migrate request (node %u) returned %d!\n",
                             nodenum, status);
                        ret = status;