[PATCH] ocfs2: fix hang in dlm lock resource mastery
authorKurt Hackel <kurt.hackel@oracle.com>
Fri, 3 Mar 2006 02:09:26 +0000 (18:09 -0800)
committerMark Fasheh <mark.fasheh@oracle.com>
Fri, 24 Mar 2006 22:58:25 +0000 (14:58 -0800)
fixes hangs in lock mastery related to refcounting on the mle structure

Signed-off-by: Kurt Hackel <kurt.hackel@oracle.com>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
fs/ocfs2/dlm/dlmmaster.c

index 847dd3cc4cf507f0bc80d4ba59bcb362609e0a2c..78ac3a00eb54235128f8fc7a162697efe5167cf8 100644 (file)
@@ -792,7 +792,15 @@ redo_request:
                        mlog_errno(ret);
                if (mle->master != O2NM_MAX_NODES) {
                        /* found a master ! */
-                       break;
+                       if (mle->master <= nodenum)
+                               break;
+                       /* if our master request has not reached the master
+                        * yet, keep going until it does.  this is how the
+                        * master will know that asserts are needed back to
+                        * the lower nodes. */
+                       mlog(0, "%s:%.*s: requests only up to %u but master "
+                            "is %u, keep going\n", dlm->name, namelen,
+                            lockid, nodenum, mle->master);
                }
        }
 
@@ -860,7 +868,19 @@ recheck:
        /* check if another node has already become the owner */
        spin_lock(&res->spinlock);
        if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
+               mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
+                    res->lockname.len, res->lockname.name, res->owner);
                spin_unlock(&res->spinlock);
+               /* this will cause the master to re-assert across
+                * the whole cluster, freeing up mles */
+               ret = dlm_do_master_request(mle, res->owner);
+               if (ret < 0) {
+                       /* give recovery a chance to run */
+                       mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
+                       msleep(500);
+                       goto recheck;
+               }
+               ret = 0;
                goto leave;
        }
        spin_unlock(&res->spinlock);
@@ -1244,13 +1264,14 @@ int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data)
 {
        u8 response = DLM_MASTER_RESP_MAYBE;
        struct dlm_ctxt *dlm = data;
-       struct dlm_lock_resource *res;
+       struct dlm_lock_resource *res = NULL;
        struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
        struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
        char *name;
        unsigned int namelen;
        int found, ret;
        int set_maybe;
+       int dispatch_assert = 0;
 
        if (!dlm_grab(dlm))
                return DLM_MASTER_RESP_NO;
@@ -1287,7 +1308,6 @@ way_up_top:
                }
 
                if (res->owner == dlm->node_num) {
-                       u32 flags = DLM_ASSERT_MASTER_MLE_CLEANUP;
                        spin_unlock(&res->spinlock);
                        // mlog(0, "this node is the master\n");
                        response = DLM_MASTER_RESP_YES;
@@ -1300,16 +1320,7 @@ way_up_top:
                         * caused all nodes up to this one to
                         * create mles.  this node now needs to
                         * go back and clean those up. */
-                       mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
-                            dlm->node_num, res->lockname.len, res->lockname.name);
-                       ret = dlm_dispatch_assert_master(dlm, res, 1,
-                                                        request->node_idx,
-                                                        flags);
-                       if (ret < 0) {
-                               mlog(ML_ERROR, "failed to dispatch assert "
-                                    "master work\n");
-                               response = DLM_MASTER_RESP_ERROR;
-                       }
+                       dispatch_assert = 1;
                        goto send_response;
                } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
                        spin_unlock(&res->spinlock);
@@ -1357,9 +1368,13 @@ way_up_top:
                        }
                } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
                        set_maybe = 0;
-                       if (tmpmle->master == dlm->node_num)
+                       if (tmpmle->master == dlm->node_num) {
                                response = DLM_MASTER_RESP_YES;
-                       else
+                               /* this node will be the owner.
+                                * go back and clean the mles on any
+                                * other nodes */
+                               dispatch_assert = 1;
+                       } else
                                response = DLM_MASTER_RESP_NO;
                } else {
                        // mlog(0, "this node is attempting to "
@@ -1398,8 +1413,8 @@ way_up_top:
                        mle = (struct dlm_master_list_entry *)
                                kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
                        if (!mle) {
-                               // bad bad bad... this sucks.
                                response = DLM_MASTER_RESP_ERROR;
+                               mlog_errno(-ENOMEM);
                                goto send_response;
                        }
                        spin_lock(&dlm->spinlock);
@@ -1418,25 +1433,19 @@ way_up_top:
                // mlog(0, "mle was found\n");
                set_maybe = 1;
                spin_lock(&tmpmle->spinlock);
+               if (tmpmle->master == dlm->node_num) {
+                       mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
+                       BUG();
+               }
                if (tmpmle->type == DLM_MLE_BLOCK)
                        response = DLM_MASTER_RESP_NO;
                else if (tmpmle->type == DLM_MLE_MIGRATION) {
                        mlog(0, "migration mle was found (%u->%u)\n",
                             tmpmle->master, tmpmle->new_master);
-                       if (tmpmle->master == dlm->node_num) {
-                               mlog(ML_ERROR, "no lockres, but migration mle "
-                                    "says that this node is master!\n");
-                               BUG();
-                       }
                        /* real master can respond on its own */
                        response = DLM_MASTER_RESP_NO;
-               } else {
-                       if (tmpmle->master == dlm->node_num) {
-                               response = DLM_MASTER_RESP_YES;
-                               set_maybe = 0;
-                       } else
-                               response = DLM_MASTER_RESP_MAYBE;
-               }
+               } else
+                       response = DLM_MASTER_RESP_MAYBE;
                if (set_maybe)
                        set_bit(request->node_idx, tmpmle->maybe_map);
                spin_unlock(&tmpmle->spinlock);
@@ -1449,6 +1458,24 @@ way_up_top:
                dlm_put_mle(tmpmle);
        }
 send_response:
+
+       if (dispatch_assert) {
+               if (response != DLM_MASTER_RESP_YES)
+                       mlog(ML_ERROR, "invalid response %d\n", response);
+               if (!res) {
+                       mlog(ML_ERROR, "bad lockres while trying to assert!\n");
+                       BUG();
+               }
+               mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
+                            dlm->node_num, res->lockname.len, res->lockname.name);
+               ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx, 
+                                                DLM_ASSERT_MASTER_MLE_CLEANUP);
+               if (ret < 0) {
+                       mlog(ML_ERROR, "failed to dispatch assert master work\n");
+                       response = DLM_MASTER_RESP_ERROR;
+               }
+       }
+
        dlm_put(dlm);
        return response;
 }
@@ -1471,8 +1498,11 @@ static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
        int to, tmpret;
        struct dlm_node_iter iter;
        int ret = 0;
+       int reassert;
 
        BUG_ON(namelen > O2NM_MAX_NAME_LEN);
+again:
+       reassert = 0;
 
        /* note that if this nodemap is empty, it returns 0 */
        dlm_node_iter_init(nodemap, &iter);
@@ -1504,9 +1534,17 @@ static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
                             "got %d.\n", namelen, lockname, to, r);
                        dlm_dump_lock_resources(dlm);
                        BUG();
+               } else if (r == EAGAIN) {
+                       mlog(0, "%.*s: node %u create mles on other "
+                            "nodes and requests a re-assert\n", 
+                            namelen, lockname, to);
+                       reassert = 1;
                }
        }
 
+       if (reassert)
+               goto again;
+
        return ret;
 }
 
@@ -1528,6 +1566,8 @@ int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data)
        char *name;
        unsigned int namelen;
        u32 flags;
+       int master_request = 0;
+       int ret = 0;
 
        if (!dlm_grab(dlm))
                return 0;
@@ -1642,11 +1682,22 @@ ok:
        // mlog(0, "woo!  got an assert_master from node %u!\n",
        //           assert->node_idx);
        if (mle) {
-               int extra_ref;
+               int extra_ref = 0;
+               int nn = -1;
                
                spin_lock(&mle->spinlock);
-               extra_ref = !!(mle->type == DLM_MLE_BLOCK
-                              || mle->type == DLM_MLE_MIGRATION);
+               if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
+                       extra_ref = 1;
+               else {
+                       /* MASTER mle: if any bits set in the response map
+                        * then the calling node needs to re-assert to clear
+                        * up nodes that this node contacted */
+                       while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES, 
+                                                   nn+1)) < O2NM_MAX_NODES) {
+                               if (nn != dlm->node_num && nn != assert->node_idx)
+                                       master_request = 1;
+                       }
+               }
                mle->master = assert->node_idx;
                atomic_set(&mle->woken, 1);
                wake_up(&mle->wq);
@@ -1677,10 +1728,15 @@ ok:
        }
 
 done:
+       ret = 0;
        if (res)
                dlm_lockres_put(res);
        dlm_put(dlm);
-       return 0;
+       if (master_request) {
+               mlog(0, "need to tell master to reassert\n");
+               ret = EAGAIN;  // positive. negative would shoot down the node.
+       }
+       return ret;
 
 kill:
        /* kill the caller! */
@@ -1713,6 +1769,10 @@ int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
        item->u.am.request_from = request_from;
        item->u.am.flags = flags;
 
+       if (ignore_higher) 
+               mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len, 
+                    res->lockname.name);
+               
        spin_lock(&dlm->work_lock);
        list_add_tail(&item->list, &dlm->work_list);
        spin_unlock(&dlm->work_lock);